text
stringlengths
14
1.76M
# Cross-Lingual Named Entity Recognition Using Parallel Corpus: A New Approach Using XLM-RoBERTa Alignment Bing Li Microsoft <EMAIL_ADDRESS> &Yujie He Microsoft <EMAIL_ADDRESS> &Wenjin Xu Microsoft <EMAIL_ADDRESS> ###### Abstract We propose a novel approach for cross-lingual Named Entity Recognition (NER) zero-shot transfer using parallel corpora. We built an entity alignment model on top of XLM-RoBERTa to project the _entities_ detected on the English part of the parallel data to the target language sentences, whose accuracy surpasses all previous unsupervised models. With the alignment model we can get pseudo-labeled NER data set in the target language to train task-specific model. Unlike using translation methods, this approach benefits from natural fluency and nuances in target-language original corpus. We also propose a modified loss function similar to focal loss but assigns weights in the opposite direction to further improve the model training on noisy pseudo- labeled data set. We evaluated this proposed approach over 4 target languages on benchmark data sets and got competitive F1 scores compared to most recent SOTA models. We also gave extra discussions about the impact of parallel corpus size and domain on the final transfer performance. ## 1 Introduction Named entity recognition (NER) is a fundamental task in natural language processing, which seeks to classify words in a sentence to predefined semantic types. Due to its nature that the ground truth label exists at word level, supervised training of NER models often requires large amount of human annotation efforts. In real-world use cases where one needs to build multi- lingual models, the required human labor scales at least linearly with number of languages, or even worse for low resource languages. Cross-Lingual transfer on Natural Language Processing(NLP) tasks has been widely studied in recent years Conneau et al. (2018); Kim et al. (2017); Ni et al. (2017); Xie et al. (2018); Ni and Florian (2019); Wu and Dredze (2019); Bari et al. (2019); Jain et al. (2019), in particular zero-shot transfer which leverages the advances in high resource language such as English to benefit other low resource languages. In this paper, we focus on the cross-lingual transfer of NER task, and more specifically using parallel corpus and pretrained multilingual language models such as mBERT Devlin (2018) and XLM-RoBERTa (XLM-R) Lample and Conneau (2019); Conneau et al. (2020). Our motivations are threefold. (1) Parallel corpus is a great resource for transfer learning and is rich between many language pairs. Some recent research focus on using completely unsupervised machine translations (e.g. word alignment Conneau et al. (2017)) for cross-lingual NER, however inaccurate translations could harm the transfer performance. For example in the word-to-word translation approach, word ordering may not be well represented during translations, such gaps in translation quality may harm model performance in down stream tasks. (2) A method could still provide business value-add even if it only works for some major languages that have sufficient parallel corpus as long as it has satisfying performance. It is a common issue in industry practices where there is a heavily customized task that need to be extended into major markets but you do not want to annotate large amounts of data in other languages. (3) Previous attempts using parallel corpus are mostly heuristics and statistical-model based Jain et al. (2019); Xie et al. (2018); Ni et al. (2017). Recent breakthroughs in multilingual language models have not been applied to such scenarios yet. Our work bridges the gap and revisits this topic with new technologies. We propose a novel semi-supervised method for the cross-lingual NER transfer, bridged by parallel corpus. First we train an NER model on source-language data set - in this case English - assuming that we have labeled task-specific data. Second we label the English part of the parallel corpus with this model. Then, we project those recognized entities onto the target language, i.e. label the span of the same entity in target-language portion of the parallel corpus. In this step we will leverage the most recent XLM-R model Lample and Conneau (2019); Conneau et al. (2020), which makes a major distinction between our work and previous attempts. Lastly, we use this pseudo-labeled data to train the task-specific model in target language directly. For the last step we explored the option of continue training from a multilingual model fine- tuned on English NER data to maximize the benefits of model transfer. We also tried a series of methods to mitigate the noisy label issue in this semi- supervised approach. The main contributions of this paper are as follows: * • We leverage the powerful multilingual model XLM-R for entity alignment. It was trained in a supervised manner with easy-to-collect data, which is in sharp contrast to previous attempts that mainly rely on unsupervised methods and human engineered features. * • Pseudo-labeled data set typically contains lots of noise, we propose a novel loss function inspired by the focal loss Lin et al. (2017). Instead of using native focal loss, we went the opposite direction by weighting hard examples less as those are more likely to be noise. * • By leveraging existing natural parallel corpus we got competitive F1 scores of NER transfer on multiple languages. We also tested that the domain of parallel corpus is critical in an effective transfer. ## 2 Related Works There are different ways to conduct zero-shot multilingual transfer. In general, there are two categories, model-based transfer and data-based transfer. Model-based transfer often use source language to train an NER model with language independent features, then directly apply the model to target language for inference Wu and Dredze (2019); Wu et al. (2020). Data-based transfer focus on combining source language task specific model, translations, and entity projection to create weakly-supervised training data in target language. Some previous attempts includes using annotation projection on aligned parallel corpora, translations between a source and a target langauge Ni et al. (2017); Ehrmann et al. (2011), or to utilize Wikipedia hyperlink structure to obtain anchor text and context as weak labels Al-Rfou et al. (2015); Tsai et al. (2016). Different variants exist in annotation projection, e.g. Ni et al. Ni et al. (2017) used maximum entropy alignment model and data selection to project English annotated labels to parallel target language sentences. Some other work used bilingual mapping combined with lexical heuristics or used embedding approach to perform word-level translation with which naturally comes the annotation projection Mayhew et al. (2017); Xie et al. (2018); Jain et al. (2019). This kind of translation + projection approach is used not just in NER, but in other NLP tasks as well such as relation extraction Kumar (2015). There are obvious limitations to the translation + projection approach, word or phrase-based translation makes annotation projection easier, but sacrifices the native fluency and language nuances. In addition, orthographic and phonetic based features for entity matching may only be applicable to languages that are alike, and requires extensive human engineered features. To address these limitations, we proposed a novel approach which utilizes machine translation training data and combined with pretrained multilingual language model for entity alignment and projection. ## 3 Model Design We will demonstrate the entity alignment model component and the full training pipeline of our work in following sections. ### 3.1 Entity Alignment Model Translation from a source language to target language may break the word ordering therefore an alignment model is needed to project entities from source language sentence to target language, so that the labels from source language can be zero-shot transferred. In this work, we use XLM-R Lample and Conneau (2019); Conneau et al. (2020) series of models which introduced the translation language model(TLM) pretraining task for the first time. TLM trains the model to predict a masked word using information from both the context and the parallel sentence in another language, making the model acquire great cross-lingual and potentially alignment capability. Our alignment model is constructed by concatenating the English name of the entity and the target language sentence, as segment A and segment B of the input sequence respectively. For token level outputs of segment B, we predict 1 if it is inside the translated entity, 0 otherwise. This formulation transforms the entity alignment problem into a token classification task. An implicit assumption is that the entity name will still be a consecutive phrase after translation. The model structure is illustrated in Fig 1. Figure 1: Entity alignment model. The query entity on the left is ’Cologne’, and the German sentence on the right is ’Köln liegt in Deutschlands’, which is ’Cologne is located in Germany’ in English, and ’Köln’ is the German translation of ’Cologne’. The model predicts the word span aligned with the query entity. ### 3.2 Cross-Lingual Transfer Pipeline Fig 2 shows the whole training/evaluation pipeline which includes 5 stages: (1) Fine-tune pretrained language model on CoNLL2003 Sang and Meulder (2003) to obtain English NER model; (2) Infer labels for English sentences in the parallel corpus; (3) Run the entity alignment model from the previous subsection and find corresponding detected English entities in the target language, failed examples are filtered out during the alignment process; (4) Fine-tune the multilingual model with data generated from step (3); (5) Evaluate the new model on the target language test sets. Figure 2: Training Pipeline Diagram. Yellow pages represent English documents while light blue pages represent German documents. Step 1 and 5 used original CoNLL data for train and test respectively; step 2, 3, and 4 used machine translation data from OPUS website. Pretrained model is either mBert or XLM-R. The final model was first fine-tuned on English NER data set then fine-tuned on target language pseudo-labeled NER data set. ## 4 Experiments And Discussions ### 4.1 Parallel Corpus In our method, we leveraged the availability of large-scale parallel corpus to transfer the NER knowledge obtained in English to other languages. Existing parallel corpora is easier to obtain than annotated NER data. We used parallel corpus crawled from the OPUS website Tiedemann (2012). In our experiments, we used the following data sets: * • Ted2013: consists of volunteer transcriptions and translations from the TED web site and was created as training data resource for the International Workshop on Spoken Language Translation 2013. * • OpenSubtitles: a new collection of translated movie subtitles that contains 62 languages. Lison and Tiedemann (2016) * • WikiMatrix: Mined parallel sentences from Wikipedia in different languages. Only pairs with scores above 1.05 are used. Schwenk et al. (2019) * • UNPC: Manually translated United Nations documents from 1994 to 2014. Ziemski et al. (2016) * • Europarl: A parallel corpus extracted from the European Parliament web site. Koehn (2005) * • WMT-News: A parallel corpus of News Test Sets provided by WMT for training SMT that contains 18 languages.111http://www.statmt.org/wmt19/translation- task.html * • NewsCommentary: A parallel corpus of News Commentaries provided by WMT for training SMT, which contains 12 languages.222http://opus.nlpl.eu/News- Commentary.php * • JW300: Mined, parallel sentences from the magazines Awake! and Watchtower. Agić and Vulić (2019) In this work, we focus on 4 languages, German, Spanish, Dutch and Chinese. We randomly select data points from all data sets above with equal weights. There might be slight difference in data distribution between languages due to data availability and relevance. ### 4.2 Alignment Model Training The objective of alignment model is to find the entity from a foreign paragraph given its English name. We feed the English name and the paragraph as segment A and B into the XLM-R model Lample and Conneau (2019); Conneau et al. (2020). Unlike NER task, the alignment task has no requirement for label completeness, since we only need one entity to be labeled in one training example. The training data set can be created from Wikipedia documents where anchor text in hyperlinks naturally indicate the location of entities and one can get the English entity name via linking by Wikipedia entity Id. An alternative to get the English name for mentions in another language is through state-of-the-art translation system. We took the latter approach to make it simple and leveraged Microsoft’s Azure Cognitive Service to do the translation. During training, we also added negative examples with faked English entities which did not appear in the other language’s sentence. The intuition is to force model to focus on English entity (segment A) and its translation instead of doing pure NER and picking out any entity in other language’s sentence (segment B). We also added examples of noun phrases or nominal entities to make the model more robust. We generated a training set of 30K samples with 25$\%$ negatives for each language and trained a XLM-R-large model Conneau et al. (2020) for 3 epochs with batch size 64. The initial learning rate is 5e-5 and other hyperparameters were defaults from HuggingFace Transformers library for the token classification task. The precision/recall/F1 on the reserved test set reached 98$\%$. The model training was done on 2 Tesla V100 and took about 20 minutes. ### 4.3 Cross-Lingual Transfer | | DE | | | | ES | | | | NL | | | | ZH | ---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|--- Model | P | R | F1 | | P | R | F1 | | P | R | F1 | | P | R | F1 Bari et al. (2019) | - | - | 65.24 | | - | - | 75.9 | | - | - | 74.6 | | - | - | - Wu and Dredze (2019) | - | - | 71.1 | | - | - | 74.5 | | - | - | 79.5 | | - | - | - Moon et al. (2019) | - | - | 71.42 | | - | - | 75.67 | | - | - | 80.38 | | - | - | - Wu et al. (2020) | - | - | 73.16 | | - | - | 76.75 | | - | - | 80.44 | | - | - | - Wu et al. (2020) | - | - | 73.22 | | - | - | 76.94 | | - | - | 80.89 | | - | - | - Wu et al. (2020) | - | - | 74.82 | | - | - | 79.31 | | - | - | 82.90 | | - | - | - Our Models | | | | | | | | | | | | | | | mBERT zero-transfer | 67.6 | 77.4 | 72.1 | | 72.4 | 78.2 | 75.2 | | 77.8 | 79.3 | 78.6 | | 64.1 | 65.0 | 64.6 mBERT fine-tune | 73.1 | 76.2 | 74.6 | | 77.7 | 77.6 | 77.6 | | 80.5 | 76.7 | 78.6 | | 80.8 | 63.3 | 71.0 | | | (+2.5) | | | | (+2.4) | | | | (+0.0) | | | | (+6.4) XLM-R zero-transfer | 67.9 | 79.8 | 73.4 | | 79.8 | 81.9 | 80.8 | | 82.2 | 80.3 | 81.2 | | 68.7 | 65.5 | 67.1 XLM-R fine-tune | 75.5 | 78.4 | 76.9 | | 76.9 | 81.0 | 78.9 | | 81.2 | 78.4 | 79.7 | | 77.8 | 65.9 | 71.3 | | | (+3.5) | | | | (-1.9) | | | | (-1.5) | | | | (+4.2) Table 1: Cross-Lingual Transfer Results on German, Spanish, Dutch and Chinese: Experiments are done with both mBERT and XLM-RoBERTa model. For each of them we compared zero-transfer result (trained on CoNLL2003 English only) and fine- tune result using zero-transfer pseudo-labeled target language NER data. Test sets for German, Spanish, Dutch are from CoNLL2003 and CoNLL2002, and People’s Daily data set for Chinese. We used the CoNLL2003 Sang and Meulder (2003) and CoNLL2002333http://lcg- www.uia.ac.be/conll2002/ner/ data sets to test our cross-lingual transfer method for German, Spanish and Dutch. We ignored the training sets in those languages and only evaluated our model on test sets. For Chinese, we used People’s Daily444https://github.com/zjy-ucas/ChineseNER as the major evaluation set, and we also reported numbers on MSRA Levow (2020) and Weibo Peng and Dredze (2015) data sets in the next section. One notable difference for People’s Daily data set is that it only covers three entity types, LOC, ORG and PER, so we suppressed the MISC type from English during the transfer by training English NER model with ’MISC’ marked as ’O’. To enable cross-lingual transfer, we first trained an English teacher model using the CoNLL2003 EN training set with XLM-R-large as the base model. We trained with focal loss Lin et al. (2017) for 5 epochs. We then ran inference with this model on the English part of the parallel data. Finally, with the alignment model, we projected entity labels onto other languages. To ensure the quality of target language training data, we discarded examples if any English entity failed to map to tokens in the target language. We also discarded examples where there are overlapping target entities because it will cause conflicts in token labels. Furthermore, when one entity is mapped to multiple, we only keep the example if all the target mention phrases are the same. This is to accommodate the situation where same entity is mentioned more than once in one sentence. As the last step, we fine-tuned the multilingual model pre-trained on English data set with lower n(0, 3, 6, etc.) layers frozen, on the target language pseudo-labeled data. We used both mBERT Devlin et al. (2019); Devlin (2018) and XLM-RConneau et al. (2020) with about 40K training samples for 1 epoch. The results are shown in Table 1. All the inference, entity projection and model training experiments are done on 2 Tesla V100 32G gpus and the whole pipeline takes about 1-2 hours. All numbers are reported as an average of 5 random runs with the same settings. For loss function we used something similar to focal loss Lin et al. (2017) but with opposite weight assignment. The focal loss was designed to weight hard examples more. This intuition holds true only when training data is clean. In some scenarios such as the cross-lingual transfer task, the pseudo- labeled training data contains lots of noise propagated from upper-stream of the pipeline, in which case, those ’hard’ examples are more likely to be just errors or outliers and could hurt the training process. We went the opposite direction and lowered their weights instead, so that the model could focus on less noisy labels. More specifically, we added weight $(1+p_{t})^{\gamma}$ instead of $(1-p_{t})^{\gamma}$ on top of the regular cross entropy loss, and for the hyper-parameter $\gamma$ we experimented with values from 1 to 5 and the value of 4 it works best. From Table 1, we can see for mBERT model, fine-tuning with pseudo-labeled data has significant effects on all languages except NL. The largest improvement is in Chinese, 6.4$\%$ increase in F1 on top of the zero-transfer result, this number is 2.5$\%$ for German and 2.4$\%$ for Spanish. The same experiment with XLM-R model shows a different pattern, F1 increased 3.5$\%$ for German but dropped a little bit on Spanish and Dutch after fine-tuning. For Chinese, we see a comparable improvement with mBERT of 4.2$\%$. The negative result on Spanish and Dutch is probably because XLM-R has already had very good pretraining and language alignment for those European languages, which can be seen from the high zero-transfer numbers, therefore a relatively noisy data set did not bring much gain. On the contrary, Chinese is a relatively distant language from the perspective of linguistics so that the add-on value of task specific fine-tuning with natural data is larger. Another pattern we observed from Table 1 is that across all languages, the fine-tuning step with pseudo-labeled data is more beneficial to precision compared with recall. We observed a consistent improvement in precision but a small drop in recall in most cases. ## 5 Discussions ### 5.1 Impact of the amount of parallel data One advantage of the parallel corpora method is high data availability compared to the supervised approach. Therefore a natural question next is whether more data is beneficial for the cross-lingual transfer task. To answer this question, we did a series of experiments with varying number of training examples ranging from 5K to 200K, and the model F1 score increases with the amount of data at the beginning and plateaued around 40K. All the numbers displayed in Table 1 are reported for training on a generated data set of size around 40K (sentences). One possible explanation for the plateaued performance might be due to the propagated error in the pseudo-labeled data set. Domain mismatch may also limit the effectiveness of transfer between languages. More discussions on this topic in the next section. ### 5.2 Impact of the domain of parallel data Figure 3: F1 scores evaluated on three data sets using different domains’ parallel data. Blue column on the left is the result of zero-shot model transfer. To its right are F1 scores for 3 different domains and all domains combined. Learnings from machine translation community showed that the quality of neural machine translation models usually strongly depends on the domain they are trained on, and the performance of a model could drop significantly when evaluated on a different domain Koehn and Knowles (2017). Similar observation can be used to explain challenges of the NER cross-lingual transfer. In NER transfer, the first domain mismatch comes from the natural gap in entity distributions between different language corpus. Many entities only live inside the ecosystem of a specific group of languages and may not be translated naturally to others. The second domain mismatch is between the parallel corpora and the NER data set. The English model might not have good domain adaptation ability and could perform well on the CoNLL2003 data set but poorly on the parallel data set. Domain | PER | ORG | LOC | All ---|---|---|---|--- OpenSubtitles | 24,036 | 3,809 | 5,196 | 33,041 UN | 1,094 | 25,875 | 12,718 | 39,687 News | 10,977 | 9,568 | 28,168 | 48,713 All Domains Combined | 10,454 | 14,269 | 17,412 | 42,135 Table 2: Entity Count by type in the pseudo-labeled Chinese NER training data set. We listed multiple domains that were extracted from different parallel data source. And AllDomains is a combination of all resources. To study the impact of the domain of parallel data on transfer performance, we did an experiment on Chinese using parallel data from different domains. We picked three representative data sets from OPUS Tiedemann (2012), OpenSubtitles, UN(contains UN and UNPC), News(contains WMT and News- Commentary) and another one with all these three combined. OpenSubtitles is from movie subtitles and language style is informal and oral. UN is from united nation reports and language style is more formal and political flavored. News data is from newspaper and content is more diverse and closer to CoNLL data sets. We evaluated the F1 on three Chinese testsets, Weibo, MSRA and People’s Daily, where Weibo contains messages from social media, MSRA is more formal and political, and People’s Daily data set are newspaper articles. From Fig 3 we see OpenSubtitles performs best on Weibo but poorly on the other two. On the contrary UN performs worst on Weibo but better on the other two. News domain performs the best on People’s Daily, which is also consistent with one’s intuition because they are both from newspaper articles. All domains combined approach has a decent performance on all three testsets. Different domains of data have quite large gap in the density and distribution of entities, for example OpenSubtitles contains more sentences that does not have any entity. In the experiments above we did filtering to keep the same ratio of ’empty’ sentences among all domains. We also examined the difference in type distributions. In Tab 2, we calculated entity counts by type and domain. OpenSubtitles has very few ORG and LOC entities, whereas UN data has very few PER data. News and All domain data are more balanced. Figure 4: F1 score by type evaluated on People’s Daily data set. The same as Fig 3, we compare the results using different domains of parallel data for the NER transfer. In Fig 4, we show evaluation on People’s Daily by type. We want to understand how does parallel data domain impacts transfer performance for different entity types. News data has the best performance on all types, and Subtitles has very bad result on ORG. All these observations are consistent with the type distribution in Table 2. ## 6 Ablation Study To better understand the contribution of each stage in the training process, we conducted an ablation study on German data set with mBERT model. We compared 6 different settings: (1) approach proposed in this work, i.e. fine- tune the English model with pseudo-labeled data with the new loss, denoted as re-weighted (RW) loss; (2) zero-transfer is direct model transfer which was trained only on English NER data; (3) fine-tune the English model with pseudo- labeled data with regular cross-entropy (CE) loss; (4) Skip fine-tune on English and directly fine-tune mBERT on pseudo-labeled data. (5)(6) fine-tune the model directly with mixed English and pseudo-labeled data simultaneously with RW and CE losses respectively. | Test P | Test R | F1 ---|---|---|--- Sequential fine-tune with RW | 73.1 | 76.2 | 74.6 Zero-transfer | 67.6 | 77.4 | 72.1 Sequential fine-tune with CE | 71.2 | 74.3 | 72.7 Skip fine-tune on English | 71.0 | 73.6 | 72.3 Fine-tune on En/De mixed (CE) | 73.1 | 75.9 | 74.5 Fine-tune on En/De mixed (RW) | 65.5 | 42.6 | 51.6 Table 3: Ablation study results evaluated on CoNLL2002 German NER data. All experiments used the mBERT-base model. RW denotes the re-weighted loss we proposed in the paper; CE denotes the regular cross entropy loss. From Table 3, we see that both pretraining on English and fine-tuning with pseudo German data are essential to get the best score. The RW loss performed better in sequential fine-tuning than in simultaneous training with mixed English and German data, this is probably because noise portion in English training set is much smaller than in the German pseudo-labeled training set, using RW loss on English data failed to exploit the fine-grained information in some hard examples and results in insufficiently optimized model. Another observation is that training the mBERT with a combination of English and German using cross entropy loss, we can get almost the same score as our best model, which is trained with two stages. ## 7 Conclusion In this paper, we proposed a new method of doing NER cross-lingual transfer with parallel corpus. By leveraging the XLM-R model for entity projection, we are able to make the whole pipeline automatic and free from human engineered feature or data, so that it could be applied to any other language that has a rich resource of translation data without extra cost. This method also has the potential to be extended to other NLP tasks, such as question answering. In this paper, we thoroughly tested the new method in four languages, and it is most effective in Chinese. We also discussed the impact of parallel data domain on NER transfer performance and found that a combination of different domain parallel corpora yielded the best average results. We also verified the contribution of the pseudo-labeled parallel data by an ablation study. In the future we will further improve the alignment model precision and also explore alternative ways of transfer such as self teaching instead of direct fine- tuning. We are also interested to see how the propose approach generalizes to cross-lingual transfers for other NLP tasks. ## References * Ni et al. (2017) Jian Ni, Georgiana Dinu, and Radu Florian. 2017. Weakly supervised cross-lingual named entity recognition via effective annotation and representation projection. In _Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)_ , pages 1470–1480. Association for Computational Linguistics. * Kim et al. (2017) Joo-Kyung Kim, Young-Bum Kim, Ruhi Sarikaya, and Eric Fosler-Lussier. 2017. Cross-lingual transfer learning for POS tagging without cross-lingual resources. In _Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing_ , pages 2832–2838, Copenhagen, Denmark. Association for Computational Linguistics. * Conneau et al. (2018) Alexis Conneau, Ruty Rinott, Guillaume Lample, Adina Williams, Samuel Bowman, Holger Schwenk, and Veselin Stoyanov. 2018. XNLI: Evaluating cross-lingual sentence representations. In _Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing_ , pages 2475–2485, Brussels, Belgium. Association for Computational Linguistics. * Xie et al. (2018) Jiateng Xie, Zhilin Yang, Graham Neubig, Noah A. Smith, and Jaime Carbonell. 2018\. Neural cross-lingual named entity recognition with minimal resources. In _Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing_ , pages 369–379, Brussels, Belgium. Association for Computational Linguistics. * Ni and Florian (2019) Jian Ni and Radu Florian. 2019. Neural cross-lingual relation extraction based on bilingual word embedding mapping. In _Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)_ , pages 399–409, Hong Kong, China. Association for Computational Linguistics. * Lin et al. (2017) T.-Y. Lin, P. Goyal, R. Girshick, K. He, and P. Dollár. Focal loss for dense object detection. arXiv preprint arXiv:1708.02002, 2017. * Wu and Dredze (2019) Shijie Wu and Mark Dredze. 2019. Beto, bentz, becas: The surprising cross-lingual effectiveness of BERT. In _Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)_ , pages 833–844, Hong Kong, China. Association for Computational Linguistics. * Jain et al. (2019) Alankar Jain, Bhargavi Paranjape, and Zachary C. Lipton. Entity projection via machine translation for cross-lingual NER. In EMNLP, pages 1083–1092, 2019. * Bari et al. (2019) M Saiful Bari, Shafiq Joty, and Prathyusha Jwalapuram. Zero-resource cross-lingual named entity recognition. arXiv preprint arXiv:1911.09812, 2019. * Sang and Meulder (2003) Erik F. Tjong Kim Sang and Fien De Meulder. 2003. Introduction to the conll-2003 shared task: Language-independent named entity recognition. In _Proceedings of the Seventh Conference on Natural Language Learning, CoNLL 2003, Held in cooperation with HLT-NAACL 2003, Edmonton, Canada, May 31 - June 1, 2003_ , pages 142–147. * Wu and Dredze (2019) Shijie Wu and Mark Dredze. Beto, bentz, becas: The surprising cross-lingual effectiveness of BERT. arXiv preprint arXiv:1904.09077, 2019. * Devlin (2018) Jacob Devlin. 2018. Multilingual bert readme document. * Devlin et al. (2019) Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In _Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers)_ , pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. * Conneau et al. (2017) Alexis Conneau, Guillaume Lample, Marc’Aurelio Ranzato, Ludovic Denoyer, and Hervé Jégou. 2017. Word translation without parallel data. _arXiv preprint arXiv:1710.04087_. * Lample and Conneau (2019) Guillaume Lample and Alexis Conneau. 2019. Cross-lingual language model pretraining. _arXiv preprint arXiv:1901.07291_. * Conneau et al. (2020) Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. In _Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics_ , pages 8440–8451, Online. Association for Computational Linguistics. * Tiedemann (2012) Jörg Tiedemann. 2012. Parallel Data, Tools and Interfaces in OPUS. In _Proceedings of the Eight International Conference on Language Resources and Evaluation (LREC’12)_ , Istanbul, Turkey. European Language Resources Association (ELRA). * Lison and Tiedemann (2016) Pierre Lison and Jörg Tiedemann. 2016. OpenSubtitles2016: Extracting Large Parallel Corpora from Movie and TV Subtitles. In _Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC 2016)_ , Paris, France. European Language Resources Association (ELRA). * Schwenk et al. (2019) Holger Schwenk, Vishrav Chaudhary, Shuo Sun, Hongyu Gong, and Francisco Guzmán. 2019. WikiMatrix: Mining 135M Parallel Sentences in 1620 Language Pairs from Wikipedia. _arXiv preprint arXiv:11907.05791_. * Ziemski et al. (2016) Michał Ziemski, Marcin Junczys-Dowmunt, and Bruno Pouliquen. 2016. The United Nations Parallel Corpus v1.0. In _Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC 2016)_ , Paris, France. European Language Resources Association (ELRA). * Koehn (2005) Philipp Koehn. 2005. Europarl: A Parallel Corpus for Statistical Machine Translation. In _Conference Proceedings: the tenth Machine Translation Summit_ , pages 79–86, Phuket, Thailand. AAMT, AAMT. * Agić and Vulić (2019) Željko Agić and Ivan Vulić. 2019. JW300: A Wide-Coverage Parallel Corpus for Low-Resource Languages. In _Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics_ , pages 3204–3210, Florence, Italy. Association for Computational Linguistics. * Ehrmann et al. (2011) Maud Ehrmann, Marco Turchi, and Ralf Steinberger. 2011. Building a multilingual named entity-annotated corpus using annotation projection. In Proceedings of Recent Advances in Natural Language Processing. Association for Computational Linguistics, pages 118–124. http://aclweb.org/anthology/R11-1017. * Al-Rfou et al. (2015) Rami Al-Rfou, Vivek Kulkarni, Bryan Perozzi, and Steven Skiena. 2015. Polyglot-ner: Massive multilingual named entity recognition. In Proceedings of the 2015 SIAM International Conference on Data Mining. SIAM, Vancouver, British Columbia, Canada. https://doi.org/10.1137/1.9781611974010.66. * Tsai et al. (2016) Chen-Tse Tsai, Stephen Mayhew, and Dan Roth. 2016. Cross-lingual named entity recognition via wikification. In _CoNLL_ , pages 219–228. * Mayhew et al. (2017) Stephen Mayhew, Chen-Tse Tsai, and Dan Roth. 2017. Cheap translation for cross-lingual named entity recognition. In _EMNLP_ , pages 2526–2535. * Kumar (2015) Faruqui Kumar, S. 2015. Multilingual open relation extraction using cross-lingual In _Proceedings of NAACL-HLT_ , 1351–1356. * Wu et al. (2020) Qianhui Wu, Zijia Lin, Guoxin Wang, Hui Chen, Börje F Karlsson, Biqing Huang, and Chin-Yew Lin. Enhanced meta-learning for cross-lingual named entity recognition with minimal resources. In AAAI, 2020. * Moon et al. (2019) Taesun Moon, Parul Awasthy, Jian Ni, and Radu Florian. 2019. Towards lingua franca named entity recognition with bert. _arXiv preprint arXiv:1912.01389_. * Wu et al. (2020) Wu, Q.; Lin, Z.; Karlsson, B. F.; Lou, J.-G.; and Huang, B. 2020. Single-/Multi-Source Cross-Lingual NER via Teacher-Student Learning on Unlabeled Data in Target Language. In _Association for Computational Linguistics_. * Wu et al. (2020) Qianhui Wu and Zijia Lin and Börje F. Karlsson and Biqing Huang and Jian-Guang Lou. UniTrans: Unifying Model Transfer and Data Transfer for Cross-Lingual Named Entity Recognition with Unlabeled Data. In _IJCAI 2020_. * Levow (2020) Gina-Anne Levow. The Third International Chinese Language Processing Bakeoff: Word Segmentation and Named Entity Recognition. In Proceedings of the Fifth SIGHAN Workshop on Chinese Language Processing, 2006. . * Peng and Dredze (2015) Nanyun Peng and Mark Dredze. Named Entity Recognition for Chinese Social Media with Jointly Trained Embeddings. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, 2015. . * Koehn and Knowles (2017) Philipp Koehn and Rebecca Knowles. Six Challenges for Neural Machine Translation. In Proceedings of the First Workshop on Neural Machine Translation, pages 28–39. 2017. .
Proc R Soc A mathematical modeling, artificial intelligence, category theory John D. Foley # Operads for complex system design specification, analysis and synthesis John D. Foley1 Spencer Breiner2 Eswaran Subrahmanian2,3 and John M. Dusel1 1Metron, Inc., 1818 Library St., Reston, VA, USA 2US National Institute of Standards and Technology, Gaithersburg, MD, USA 3Carnegie Mellon University, Pittsburgh, PA, USA<EMAIL_ADDRESS> ###### Abstract As the complexity and heterogeneity of a system grows, the challenge of specifying, documenting and synthesizing correct, machine-readable designs increases dramatically. Separation of the system into manageable parts is needed to support analysis at various levels of granularity so that the system is maintainable and adaptable over its life cycle. In this paper, we argue that operads provide an effective knowledge representation to address these challenges. Formal documentation of a syntactically correct design is built up during design synthesis, guided by semantic reasoning about design effectiveness. Throughout, the ability to decompose the system into parts and reconstitute the whole is maintained. We describe recent progress in effective modeling under this paradigm and directions for future work to systematically address scalability challenges for complex system design. ###### keywords: complex systems, system design, automated reasoning, compositionality, applied category theory, operads ††journal: rspa ## 1 Introduction We solve complex problems by separating them into manageable parts [2, 86]. Human designers do this intuitively, but details can quickly overwhelm intuition. Multiple aspects of a problem may lead to distinct decompositions and complementary models of a system–e.g. competing considerations for cyberphysical systems [63, 87]–or simulation of behavior at many levels of fidelity–e.g. in modeling and simulation [88]–leading to a spectrum of models which are challenging to align. We argue that operads, formal tools developed to compose geometric and algebraic objects, are uniquely suited to separate complex systems into manageable parts and maintain alignment across complementary models. (1)(2)(3)SyntaxAbstractsystem designs$\mapsto$ComposingdesignsSemanticsComputationalmodels$\mapsto$Composingmodels Figure 1: Separating concerns with operads: (1) Composition separates subsystem designs (green boundaries –); (2) Functorial semantics separate abstract systems from computational model instances (red boundaries –); (3) Natural transformations separate and align (blue boundary –) complementary models ($\square$, $\diamond$ ). Operads provide three ways to separate concerns for complex systems: (1) designs for subsystems are separated into composable modules; (2) syntactic designs to compose systems are separated from the semantic data that model them; and (3) separate semantic models can be aligned to evaluate systems in different ways. The three relationships are illustrated in Figure 1. Hierarchical decomposition (Fig. 1, –) is nothing new. Both products and processes are broken down to solve problems from design to daily maintenance. Operads provide a precise language to manage complex modeling details that the intuitive–and highly beneficial–practice of decomposition uncovers, e.g., managing multiple, complementary decompositions and models. Operads separate the syntax to compose subsystems from the semantic data modeling them (Fig. 1, –). Syntax consists of abstract “operations” to design the parts and architecture of a system. Semantics define how to interpret and evaluate these abstract blueprints. Operad syntax is typically lightweight and declarative. Operations can often be represented both graphically and algebraically (Fig. 4), formalizing intuitive design diagrams. Operad semantics model specific aspects of a system and can range from fast to computationally expensive. The most powerful way operads separate is by aligning complementary models while maintaining compatibility with system decompositions (Fig. 1, –). Reconciling complementary models is a persistent and pervasive issue across domains [84, 26, 29, 66, 63, 27, 79, 69]. Historically, Eilenberg & Mac Lane [34] invented _natural transformations_ to align computational models of topological spaces. Operads use natural transformations to align hierarchical decompositions, which is particularly well-suited to system design. This paper articulates a uniform and systematic foundation for system design and analysis. In essence, the syntax of an operad defines _what can be_ put together, which is a prerequisite to decide _what should be_ put together. Interfaces define which designs are _syntactically_ feasible, but key _semantic_ information must be expressed to evaluate candidate designs. Formulating system models within operad theory enforces the intellectual hygiene required to make sure that different concerns stay separated while working together to solve complex design problems. We note five strengths of this foundation that result from the three ways operads separate a complex problem and key sections of the paper that provide illustrations. Expressive, unifying meta-language. A meta- or multi-modeling [18] language is needed to express and relate multiple representations. The key feature of operad-based meta-modeling is its focus on coherent mappings between models (Fig. 1, –, –), as opposed to a universal modeling context, like UML, OWL, etc., which is inevitably under or over expressive for particular use cases. Unification allows complementary models to work in concert, as we see in Sec. 5 for function and control. Network operads—originally developed to design systems—were applied to task behavior. This power to unify and express becomes especially important when reasoning across domains with largely independent modeling histories; compare, e.g., [87]. (22.2, 22.4, 3, 44.1, 5) Minimal data needed for specification. Data needed to set up each representation of a problem is minimal in two ways: (1) any framework must provide similar, generative data; and (2) each level only needs to specify data relevant to that level. Each representation is self-sufficient and can be crafted to efficiently address a limited portion of the full problem. The modeler can pick and choose relevant representations and extend the meta-model as needed. (4, 66.2) Efficient exploration of formal correct designs. An operad precisely defines how to iteratively construct designs or adapt designs by substituting different subsystems. Constructing syntactically invalid designs is not possible, restricting the relevant design space, and correctness is preserved when moving across models. Semantic reasoning guides synthesis–potentially at several levels of detail. This facilitates lazy evaluation: first syntactic correctness is guaranteed, then multitudes of coarse models are explored before committing to later, more expensive evaluations. The basic moves of iteration, substitution, and moving across levels constitute a rich framework for exploration. We obtain not only an effective design but also formal documentation of the models which justify this choice. (22.2–2.3, 6, 77.5) Separates representation from exploitation. Operads and algebras provide structure and representation for a problem. Exploitation of this structure and representation is a separate concern. As Herbert Simon noted during his Nobel Prize speech [85]: “…decision makers can satisfice either by finding optimum solutions for a simplified world, or by finding satisfactory solutions for a more realistic world.” This is an either-or proposition for a simple representation. By laying the problem across multiple semantic models, useful data structures for each model–e.g. logical, evolutionary or planning frameworks–can be exploited by algorithms that draw on operad-based iteration and substitution. (6, 77.5) Hierarchical analysis and synthesis. Operads naturally capture options for the hierarchical decomposition of a system, either within a semantic model to break up large scale problems or across models to gradually increase modeling fidelity. (22.1, 5,66.3, 77.1) ### 1.1 Contribution to design literature There are well-known examples of the successful exploitation of separation. For instance, electronic design automation (EDA) has had decades of success leveraging hierarchical separation of systems and components to achieve very large scale integration (VLSI) of complex electronic systems [16, 36, 82]. We do not argue that operads are needed for extremely modular domains. Instead, operads may help broaden the base of domains that benefit from separation and provide a means to integrate and unify treatment across domains. On the other hand, for highly integral domains the ability to separate in practically useful ways may be limited [89, 102]. The recent applications we present help illustrate where operads may prove useful in the near and longer term; see 77.3 for further discussion. Compared to related literature, this article is application driven and outward focused. Interest in applying operads and category theory to systems engineering has surged [21, 38, 53, 67, 71, 94] as part of a broader wave applying category theory to design databases, software, proteins, etc. [22, 31, 40, 46, 91, 92, 93]. While much of loc. cit. matches applications to existing theoretical tools, the present article describes recent _application driven_ advancements and overviews _specific methods_ developed to address challenges presented by domain problems. We introduce operads for design to a general scientific audience by explaining what the operads do relative to broadly applied techniques and how specific domain problems are modeled. Research directions are presented with an eye towards opening up interdisciplinary partnerships and continuing application driven investigations to build on recent insights. ### 1.2 Organization of the paper The present article captures an intermediate stage of technical maturity: operad-based design has shown its practicality by lowering barriers of entry for applied practitioners and demonstrating applied examples across many domains. However, it has not realized its full potential as an applied meta- language. Much of this recent progress is not focused solely on the analytic power of operads to separate concerns. Significant progress on explicit specification of domain models and techniques to automatically synthesize designs from basic building blocks has been made. Illustrative use cases and successful applications for design specification, analysis and synthesis organize the exposition. Section 2 introduces operads for design by analogy to other modeling approaches. Our main examples are introduced in Section 3. Section 4 describes how concrete domains can be specified with minimal combinatorial data, lowering barriers to apply operads. Section 5 concerns analysis of a system with operads. Automated synthesis is discussed in Section 6. Future research directions are outlined in Section 7, which includes a list of open problems. Ex. 33.1SpecificationSection 4Ex. 33.2 Analysis Section 5Ex. 33.3 Synthesis Section 6GeneratorsCompositionalityScalability Figure 2: Organization of the paper around applied examples introduced in Sec. 3. Notations. Throughout, we maintain the following notional conventions for: * • syntax operads (Fig. 1, left), capitalized calligraphy: $\mathcal{O}$ * • types (Fig. 1, edges on left), capitalized teletype: $\mathtt{X},\mathtt{Y},\mathtt{Z},\ldots$ * • operations (Fig. 1, nodes on left), uncapitalized teletype: $\mathtt{f},\mathtt{g},\mathtt{h},\ldots$ * • semantic contexts (Fig. 1, right), capitalized bold: $\mathbf{Sem},\mathbf{Set},\mathbf{Rel},\dots$ * • functors from syntax to semantics (Fig. 1, arrows across red), capitalized sans serif: $\mathsf{Model}\colon{\mathcal{O}}\to\mathbf{Sem}$; * • alignment of semantic models via natural transformations (Fig. 1, double arrow across blue), uncapitalized sans serif: $\mathsf{align}\colon\mathsf{Model_{1}}\Rightarrow\mathsf{Model_{2}}$; ## 2 Applying operads to design We introduce operads by an analogy, explaining what an operad is and motivating its usefulness for systems modeling and analysis. The theory [64, 70, 103] pulls together many different intuitions. Here we highlight four analogies or ‘views’ of an operad: hierarchical representations (tree view), strongly-typed programming languages (API 111Application Programming Interface view), algebraic equations (equational view) and system cartography (map- maker’s view). Each view motivates operad concepts; see Table 1. The paradigm of this paper is based on a typed operad, also known as a ‘colored operad’ [103] or ‘symmetric multicategory’ [64, 2.2.21]. A typed operad ${\mathcal{O}}$ has: * • A set $T$ of types. * • Sets of operations ${\mathcal{O}}(\mathtt{X_{1}},\ldots,\mathtt{X_{n}};\mathtt{Y})$ where $\mathtt{X_{i}},\mathtt{Y}\in T$ and we write $\mathtt{f}\colon\langle\mathtt{X_{i}}\rangle\to\mathtt{Y}$ to indicate that $\mathtt{f}\in{\mathcal{O}}(\mathtt{X_{1}},\ldots,\mathtt{X_{n}};\mathtt{Y})$. * • A specific way to compose any operation $\mathtt{f}\colon\langle\mathtt{Y_{i}}\rangle\to\mathtt{Z}$ with $\mathtt{g_{i}}\colon\langle\mathtt{X_{ij}}\rangle\to\mathtt{Y_{i}}$ whose output types match the inputs of $f$ to obtain a composite $\mathtt{f}\circ(\mathtt{g_{1}},\dots,\mathtt{g_{n}})=\mathtt{h}\colon\langle\mathtt{X_{ij}}\rangle\to\mathtt{Z}$. These data are subject to rules [103, 11.2] governing permutation of arguments and assuring that iterative composition is coherent, analogous to associativity for ordinary categories [68, I]. Table 1: The theory of operads draws on many familiar ideas, establishing a dictionary between contexts. Operads | Tree | API | Equational | Systems ---|---|---|---|--- Types | Edges | Data types | Variables | Boundaries Operations | Nodes | Methods | Operators | Architectures Composites | Trees | Scripts | Evaluation | Nesting Algebras | Labels | Implementations | Values | Models ### 2.1 The tree view Hierarchies are everywhere, from scientific and engineered systems to government, business and everyday life; they help to decompose complex problems into more manageable pieces. The fundamental constituent of an operad, called an _operation_ , represents a single step in a hierarchical decomposition. We can think of this as a single branching in a labeled tree, e.g.: $\mathtt{op}$$\mathtt{System}$$\mathtt{Sub_{1}}$$\mathtt{Sub_{2}}$ Formally, this represents an element $\mathtt{op}\in{\mathcal{O}}(\mathtt{Sub_{1}},\mathtt{Sub_{2}};\mathtt{System})$. More generally, we can form new operations—trees—by _composition_. Given further refinements for the two subsystems $\mathtt{Sub_{1}}$ and $\mathtt{Sub_{2}}$, by $\mathtt{op_{1}}$ and $\mathtt{op_{2}}$, respectively, we have three composites: $\begin{array}[]{ccc}\scalebox{0.85}{ \leavevmode\hbox to98.66pt{\vbox to100.79pt{\pgfpicture\makeatletter\hbox{\hskip 57.14551pt\lower-64.61296pt\hbox to0.0pt{\pgfsys@beginscope\pgfsys@invoke{ }\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\pgfsys@setlinewidth{0.4pt}\pgfsys@invoke{ }\nullfont\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\hbox to0.0pt{\pgfsys@beginscope\pgfsys@invoke{ }{{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{{{}}}{{}}{}{}{}{}{}{}{}{}{}{{}\pgfsys@moveto{10.75293pt}{0.0pt}\pgfsys@curveto{10.75293pt}{5.93875pt}{5.93875pt}{10.75293pt}{0.0pt}{10.75293pt}\pgfsys@curveto{-5.93875pt}{10.75293pt}{-10.75293pt}{5.93875pt}{-10.75293pt}{0.0pt}\pgfsys@curveto{-10.75293pt}{-5.93875pt}{-5.93875pt}{-10.75293pt}{0.0pt}{-10.75293pt}\pgfsys@curveto{5.93875pt}{-10.75293pt}{10.75293pt}{-5.93875pt}{10.75293pt}{0.0pt}\pgfsys@closepath\pgfsys@moveto{0.0pt}{0.0pt}\pgfsys@stroke\pgfsys@invoke{ } }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{-5.27779pt}{-1.18056pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{op}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{{{}}}{{}}{}{}{}{}{}{}{}{}{}{{}\pgfsys@moveto{-10.57571pt}{-28.45276pt}\pgfsys@curveto{-10.57571pt}{-22.50798pt}{-15.39479pt}{-17.6889pt}{-21.33957pt}{-17.6889pt}\pgfsys@curveto{-27.28435pt}{-17.6889pt}{-32.10342pt}{-22.50798pt}{-32.10342pt}{-28.45276pt}\pgfsys@curveto{-32.10342pt}{-34.39754pt}{-27.28435pt}{-39.21661pt}{-21.33957pt}{-39.21661pt}\pgfsys@curveto{-15.39479pt}{-39.21661pt}{-10.57571pt}{-34.39754pt}{-10.57571pt}{-28.45276pt}\pgfsys@closepath\pgfsys@moveto{-21.33957pt}{-28.45276pt}\pgfsys@stroke\pgfsys@invoke{ } }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{-28.01736pt}{-29.63332pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{op_{1}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{-15.72226pt}{26.00832pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{System}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{-53.8125pt}{-59.47551pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{Sub_{11}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{-18.24655pt}{-59.47551pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{Sub_{12}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{18.7194pt}{-59.47551pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{Sub_{2}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{0.0pt}{10.95293pt}\pgfsys@lineto{0.0pt}{20.53088pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{{}} {{{{{}}{}{}{}{}{{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{-6.57181pt}{-8.76236pt}\pgfsys@lineto{-14.7612pt}{-19.68166pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{4.8987pt}{-9.79639pt}\pgfsys@lineto{24.50134pt}{-48.99808pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{-27.9181pt}{-37.22353pt}\pgfsys@lineto{-36.7494pt}{-48.99808pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{-16.43616pt}{-38.2591pt}\pgfsys@lineto{-11.0663pt}{-48.99808pt}\pgfsys@stroke\pgfsys@invoke{ } \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope{}{}{}\hss}\pgfsys@discardpath\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\hss}}\lxSVG@closescope\endpgfpicture}} }&\scalebox{0.85}{ \leavevmode\hbox to112.89pt{\vbox to100.79pt{\pgfpicture\makeatletter\hbox{\hskip 48.63231pt\lower-64.61296pt\hbox to0.0pt{\pgfsys@beginscope\pgfsys@invoke{ }\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\pgfsys@setlinewidth{0.4pt}\pgfsys@invoke{ }\nullfont\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\hbox to0.0pt{\pgfsys@beginscope\pgfsys@invoke{ }{{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{{{}}}{{}}{}{}{}{}{}{}{}{}{}{{}\pgfsys@moveto{10.75293pt}{0.0pt}\pgfsys@curveto{10.75293pt}{5.93875pt}{5.93875pt}{10.75293pt}{0.0pt}{10.75293pt}\pgfsys@curveto{-5.93875pt}{10.75293pt}{-10.75293pt}{5.93875pt}{-10.75293pt}{0.0pt}\pgfsys@curveto{-10.75293pt}{-5.93875pt}{-5.93875pt}{-10.75293pt}{0.0pt}{-10.75293pt}\pgfsys@curveto{5.93875pt}{-10.75293pt}{10.75293pt}{-5.93875pt}{10.75293pt}{0.0pt}\pgfsys@closepath\pgfsys@moveto{0.0pt}{0.0pt}\pgfsys@stroke\pgfsys@invoke{ } }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{-5.27779pt}{-1.18056pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{op}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{{{}}}{{}}{}{}{}{}{}{}{}{}{}{{}\pgfsys@moveto{32.10342pt}{-28.45276pt}\pgfsys@curveto{32.10342pt}{-22.50798pt}{27.28435pt}{-17.6889pt}{21.33957pt}{-17.6889pt}\pgfsys@curveto{15.39479pt}{-17.6889pt}{10.57571pt}{-22.50798pt}{10.57571pt}{-28.45276pt}\pgfsys@curveto{10.57571pt}{-34.39754pt}{15.39479pt}{-39.21661pt}{21.33957pt}{-39.21661pt}\pgfsys@curveto{27.28435pt}{-39.21661pt}{32.10342pt}{-34.39754pt}{32.10342pt}{-28.45276pt}\pgfsys@closepath\pgfsys@moveto{21.33957pt}{-28.45276pt}\pgfsys@stroke\pgfsys@invoke{ } }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{14.66177pt}{-29.63332pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{op_{2}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{-15.72226pt}{26.00832pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{System}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{-18.24655pt}{-59.47551pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{Sub_{21}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{10.2062pt}{-59.47551pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{Sub_{22}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{38.65897pt}{-59.47551pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{Sub_{23}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{-45.2993pt}{-59.47551pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{Sub_{1}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{0.0pt}{10.95293pt}\pgfsys@lineto{0.0pt}{20.53088pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{{}} {{{{{}}{}{}{}{}{{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{6.57181pt}{-8.76236pt}\pgfsys@lineto{14.7612pt}{-19.68166pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{-5.80553pt}{-9.28781pt}\pgfsys@lineto{-30.62671pt}{-48.99808pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{29.09235pt}{-36.2052pt}\pgfsys@lineto{41.88611pt}{-48.99808pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{21.33957pt}{-39.41661pt}\pgfsys@lineto{21.33957pt}{-48.99808pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{13.58679pt}{-36.2052pt}\pgfsys@lineto{0.79303pt}{-48.99808pt}\pgfsys@stroke\pgfsys@invoke{ } \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope{}{}{}\hss}\pgfsys@discardpath\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\hss}}\lxSVG@closescope\endpgfpicture}} }&\scalebox{0.85}{ \leavevmode\hbox to149.86pt{\vbox to100.79pt{\pgfpicture\makeatletter\hbox{\hskip 71.37189pt\lower-64.61296pt\hbox to0.0pt{\pgfsys@beginscope\pgfsys@invoke{ }\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\pgfsys@setlinewidth{0.4pt}\pgfsys@invoke{ }\nullfont\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\hbox to0.0pt{\pgfsys@beginscope\pgfsys@invoke{ }{{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{{{}}}{{}}{}{}{}{}{}{}{}{}{}{{}\pgfsys@moveto{10.75293pt}{0.0pt}\pgfsys@curveto{10.75293pt}{5.93875pt}{5.93875pt}{10.75293pt}{0.0pt}{10.75293pt}\pgfsys@curveto{-5.93875pt}{10.75293pt}{-10.75293pt}{5.93875pt}{-10.75293pt}{0.0pt}\pgfsys@curveto{-10.75293pt}{-5.93875pt}{-5.93875pt}{-10.75293pt}{0.0pt}{-10.75293pt}\pgfsys@curveto{5.93875pt}{-10.75293pt}{10.75293pt}{-5.93875pt}{10.75293pt}{0.0pt}\pgfsys@closepath\pgfsys@moveto{0.0pt}{0.0pt}\pgfsys@stroke\pgfsys@invoke{ } }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{-5.27779pt}{-1.18056pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{op}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{{{}}}{{}}{}{}{}{}{}{}{}{}{}{{}\pgfsys@moveto{-24.8021pt}{-28.45276pt}\pgfsys@curveto{-24.8021pt}{-22.50798pt}{-29.62117pt}{-17.6889pt}{-35.56595pt}{-17.6889pt}\pgfsys@curveto{-41.51073pt}{-17.6889pt}{-46.3298pt}{-22.50798pt}{-46.3298pt}{-28.45276pt}\pgfsys@curveto{-46.3298pt}{-34.39754pt}{-41.51073pt}{-39.21661pt}{-35.56595pt}{-39.21661pt}\pgfsys@curveto{-29.62117pt}{-39.21661pt}{-24.8021pt}{-34.39754pt}{-24.8021pt}{-28.45276pt}\pgfsys@closepath\pgfsys@moveto{-35.56595pt}{-28.45276pt}\pgfsys@stroke\pgfsys@invoke{ } }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{-42.24374pt}{-29.63332pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{op_{1}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{-15.72226pt}{26.00832pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{System}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{-68.03888pt}{-59.47551pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{Sub_{11}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{-32.47293pt}{-59.47551pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{Sub_{12}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{0.0pt}{10.95293pt}\pgfsys@lineto{0.0pt}{20.53088pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{{}} {{{{{}}{}{}{}{}{{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{-8.5528pt}{-6.84222pt}\pgfsys@lineto{-27.00462pt}{-21.6037pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{-42.14449pt}{-37.22353pt}\pgfsys@lineto{-50.97578pt}{-48.99808pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{-30.66254pt}{-38.2591pt}\pgfsys@lineto{-25.29268pt}{-48.99808pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{{{}}}{{}}{}{}{}{}{}{}{}{}{}{{}\pgfsys@moveto{46.3298pt}{-28.45276pt}\pgfsys@curveto{46.3298pt}{-22.50798pt}{41.51073pt}{-17.6889pt}{35.56595pt}{-17.6889pt}\pgfsys@curveto{29.62117pt}{-17.6889pt}{24.8021pt}{-22.50798pt}{24.8021pt}{-28.45276pt}\pgfsys@curveto{24.8021pt}{-34.39754pt}{29.62117pt}{-39.21661pt}{35.56595pt}{-39.21661pt}\pgfsys@curveto{41.51073pt}{-39.21661pt}{46.3298pt}{-34.39754pt}{46.3298pt}{-28.45276pt}\pgfsys@closepath\pgfsys@moveto{35.56595pt}{-28.45276pt}\pgfsys@stroke\pgfsys@invoke{ } }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{28.88815pt}{-29.63332pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{op_{2}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{-4.02017pt}{-59.47551pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{Sub_{21}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{24.43259pt}{-59.47551pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{Sub_{22}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{52.88535pt}{-59.47551pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{Sub_{23}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{0.0pt}{10.95293pt}\pgfsys@lineto{0.0pt}{20.53088pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{{}} {{{{{}}{}{}{}{}{{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{8.5528pt}{-6.84222pt}\pgfsys@lineto{27.00462pt}{-21.6037pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{27.81317pt}{-36.2052pt}\pgfsys@lineto{15.01941pt}{-48.99808pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{35.56595pt}{-39.41661pt}\pgfsys@lineto{35.56595pt}{-48.99808pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{43.31873pt}{-36.2052pt}\pgfsys@lineto{56.11249pt}{-48.99808pt}\pgfsys@stroke\pgfsys@invoke{ } \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope{}{}{}\hss}\pgfsys@discardpath\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\hss}}\lxSVG@closescope\endpgfpicture}}}\end{array}$ (1) Together with the original operation, these represent four views of the same system at different levels of granularity; compare, e.g., [65, Fig. 2]. This reveals an important point: an operad provides a collection of interrelated models that fit together to represent a complex system. The relationship between models is constrained by the _principle of compositionality_ : the whole is determined by its parts _and_ their organization. Here, the whole is the root, the parts are the leaves, and each tree is an organizational structure. Formally, _associativity axioms_ , which generalize those of ordinary categories, enforce compositionality. For example, composing the left-hand tree above with $\mathtt{op_{2}}$ must give the same result as composing the center tree with $\mathtt{op_{1}}$. Both give the tree on the right, since they are built up from the same operations. In day-to-day modeling these axioms are mostly invisible, ensuring that everything “just works”, but the formal definitions [103, 11.2] provide explicit requirements and desiderata for modeling languages “under the hood”. Operads encourage principled approaches to emergence by emphasizing the organization of a system. Colloquially speaking, an emergent system is “more than the sum of its parts”; operations provide a means to describe these nonlinearities. This does not explain emergent phenomena, which requires detailed semantic modeling, but begins to break up the problem with separate (but related) representation of components and their interactions. The interplay between these elements can be complex and unexpected, even when the individual elements are quite simple.222For example, diffusion rates (components) and activation/inhibition (interactions) generate zebra’s stripes in Turing’s model of morphogenesis [99]. Compositional models may develop and exhibit emergence as interactions between components are organized, in much the same way as the systems they represent. ### 2.2 The API view For most applications, trees bear labels: fault trees, decision trees, syntax trees, dependency trees and file directories, to name a few. A tree’s labels indicate its semantics either explicitly with numbers and symbols or implicitly through naming and intention. In an operad, nodes identify operations while edges—called _types_ —restrict the space of valid compositions. This is in analogy to type checking in strongly-typed programming languages, where we can only compose operations when types match. In the API view, the operations are abstract method declarations: > def op(x1 : Sub1, x2 : Sub2) : System, > --- > def op1(y1 : Sub11, y2 : Sub12) : Sub1, > def op2(z1 : Sub21, z2 : Sub22, z3 : Sub23) : Sub2. Composites are essentially scripted methods defined in the API. For example, > def treeLeft(y1 : Sub11, y2 : Sub12, x2 : Sub2) : System > = op(op1(y1, y2), x2), is a script for left-most tree above. However, the compiler will complain with an invalid syntax error for any script where the types don’t match, say > def badTree(y1 : Sub11, y2 : Sub12, x2 : Sub2) : System > = op(x2 ,op1(y1, y2)), If an operad is an API—a collection of abstract types and methods—then an _operad algebra_ $\mathsf{A}$ is a concrete implementation. An algebra declares: 1) a set of instances for each type; 2) a function for each operation, taking instances as arguments and returning a single instance for the composite system. That is, $\mathsf{A}\colon{\mathcal{O}}\to\mathbf{Set}$ has: * • for each type $\mathtt{X}\in T$, a set $\mathsf{A}(\mathtt{X})$ of instances of type $\mathtt{X}$, and * • for each operation $\mathtt{f}\colon\langle\mathtt{X_{i}}\rangle\to\mathtt{Y}$, the function $\mathsf{A}(\mathtt{f})$ acts on input elements $a_{i}\in\mathsf{A}(\mathtt{X_{i}})$ to obtain a single output element $\mathsf{A}(\mathtt{f})(a_{1},\dots,a_{n})\in\mathsf{A}(\mathtt{Y}).$ Required coherence rules [103, 13.2] are analogous to the definition of a functor into $\mathbf{Set}$ [68, I.3]. For example, we might declare a state space for each subsystem, and a function to calculate the overall system state given subsystem states. Alternatively, we might assign key performance indicators (KPIs) for each level in a system and explicit formulae to aggregate them. The main thing to remember is: just as an abstract method has many implementations, an operad has many algebras. Just like an API, the operad provides a common syntax for a range of specific models, suited for specific purposes. Unlike a traditional API, an operad provides an explicit framework to express and reason about semantic relationships between _different_ implementations. These different implementations are linked by type-indexed mappings between instances called _algebra homomorphisms_. For example, we might like to extract KPIs from system state. The principle of compositionality places strong conditions on this extraction: the KPIs extracted from the overall system state must agree with the KPIs obtained by aggregating subsystem KPIs. That is, in terms of trees and in pseudocode: $\mathsf{KPI}(\mathtt{op})$$\mathsf{extr}(\mathtt{Sub_{1}})$$\mathsf{extr}(\mathtt{Sub_{2}})$$\mathsf{KPI}(\mathtt{System})$$\mathsf{State}(\mathtt{Sub_{1}})$$\mathsf{State}(\mathtt{Sub_{2}})$$\mathsf{KPI}(\mathtt{Sub_{1}})$$\mathsf{KPI}(\mathtt{Sub_{2}})$ | = | $\mathsf{State}(\mathtt{op})$$\mathsf{extr}(\mathtt{System})$ $\mathsf{KPI}(\mathtt{System})$$\mathsf{State}(\mathtt{Sub_{1}})$$\mathsf{State}(\mathtt{Sub_{2}})$$\mathsf{State}(\mathtt{System})$ ---|---|--- KPI(op)(extr(x1), extr(x2)) | == | extr(State(op)(x1, x2)). For any state instances for $\mathtt{Sub_{1}}$ and $\mathtt{Sub_{2}}$ at the base of the tree, the two computations must produce the same KPIs for the overall system at the top of the tree. Here $\mathsf{KPI}(\mathtt{op})$ and $\mathsf{State}(\mathtt{op})$ implement $\mathtt{op}$ in the two algebras, while $\mathsf{extr}(-)$ are _components_ of the algebra homomorphism to extract KPIs. Similar to associativity, these compositionality conditions guarantee that extracting KPIs “just works” when decomposing a system hierarchically. ### 2.3 The equational view We have just seen an equation between trees that represent implementations. Because an operad can be studied without reference to an implementation, we can also define equations between abstract trees. This observation leads to another view of an operad: as a system of equations. The first thing to note is that equations occur within the sets of operations ${\mathcal{O}}(\mathtt{X_{1}},\ldots,\mathtt{X_{n}};\mathtt{Y})$; an equation between two operations only makes sense if the input and output types match. Second, if one side of an equation $\mathtt{f}=\mathtt{f}^{\prime}$ occurs as a subtree in a larger operation $\mathtt{g}$, substitution generates a new equation $\mathtt{g}=\mathtt{g^{\prime}}$. Two trees are equal if and only if they are connected by a chain of such substitutions (and associativity equations). In general, deciding whether two trees are equal (the word problem) may be intractable. Third, we can often interpret composition of operations as a normal-form computation: $\begin{array}[]{ccccc}\begin{array}[]{c}\scalebox{0.85}{ \leavevmode\hbox to149.86pt{\vbox to100.79pt{\pgfpicture\makeatletter\hbox{\hskip 71.37189pt\lower-64.61296pt\hbox to0.0pt{\pgfsys@beginscope\pgfsys@invoke{ }\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\pgfsys@setlinewidth{0.4pt}\pgfsys@invoke{ }\nullfont\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\hbox to0.0pt{\pgfsys@beginscope\pgfsys@invoke{ }{{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{{{}}}{{}}{}{}{}{}{}{}{}{}{}{{}\pgfsys@moveto{10.75293pt}{0.0pt}\pgfsys@curveto{10.75293pt}{5.93875pt}{5.93875pt}{10.75293pt}{0.0pt}{10.75293pt}\pgfsys@curveto{-5.93875pt}{10.75293pt}{-10.75293pt}{5.93875pt}{-10.75293pt}{0.0pt}\pgfsys@curveto{-10.75293pt}{-5.93875pt}{-5.93875pt}{-10.75293pt}{0.0pt}{-10.75293pt}\pgfsys@curveto{5.93875pt}{-10.75293pt}{10.75293pt}{-5.93875pt}{10.75293pt}{0.0pt}\pgfsys@closepath\pgfsys@moveto{0.0pt}{0.0pt}\pgfsys@stroke\pgfsys@invoke{ } }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{-5.27779pt}{-1.18056pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{op}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{{{}}}{{}}{}{}{}{}{}{}{}{}{}{{}\pgfsys@moveto{-24.8021pt}{-28.45276pt}\pgfsys@curveto{-24.8021pt}{-22.50798pt}{-29.62117pt}{-17.6889pt}{-35.56595pt}{-17.6889pt}\pgfsys@curveto{-41.51073pt}{-17.6889pt}{-46.3298pt}{-22.50798pt}{-46.3298pt}{-28.45276pt}\pgfsys@curveto{-46.3298pt}{-34.39754pt}{-41.51073pt}{-39.21661pt}{-35.56595pt}{-39.21661pt}\pgfsys@curveto{-29.62117pt}{-39.21661pt}{-24.8021pt}{-34.39754pt}{-24.8021pt}{-28.45276pt}\pgfsys@closepath\pgfsys@moveto{-35.56595pt}{-28.45276pt}\pgfsys@stroke\pgfsys@invoke{ } }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{-42.24374pt}{-29.63332pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{op_{1}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{-15.72226pt}{26.00832pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{System}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{-68.03888pt}{-59.47551pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{Sub_{11}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{-32.47293pt}{-59.47551pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{Sub_{12}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{0.0pt}{10.95293pt}\pgfsys@lineto{0.0pt}{20.53088pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{{}} {{{{{}}{}{}{}{}{{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{-8.5528pt}{-6.84222pt}\pgfsys@lineto{-27.00462pt}{-21.6037pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{-42.14449pt}{-37.22353pt}\pgfsys@lineto{-50.97578pt}{-48.99808pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{-30.66254pt}{-38.2591pt}\pgfsys@lineto{-25.29268pt}{-48.99808pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{{{}}}{{}}{}{}{}{}{}{}{}{}{}{{}\pgfsys@moveto{46.3298pt}{-28.45276pt}\pgfsys@curveto{46.3298pt}{-22.50798pt}{41.51073pt}{-17.6889pt}{35.56595pt}{-17.6889pt}\pgfsys@curveto{29.62117pt}{-17.6889pt}{24.8021pt}{-22.50798pt}{24.8021pt}{-28.45276pt}\pgfsys@curveto{24.8021pt}{-34.39754pt}{29.62117pt}{-39.21661pt}{35.56595pt}{-39.21661pt}\pgfsys@curveto{41.51073pt}{-39.21661pt}{46.3298pt}{-34.39754pt}{46.3298pt}{-28.45276pt}\pgfsys@closepath\pgfsys@moveto{35.56595pt}{-28.45276pt}\pgfsys@stroke\pgfsys@invoke{ } }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{28.88815pt}{-29.63332pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{op_{2}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{-4.02017pt}{-59.47551pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{Sub_{21}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{24.43259pt}{-59.47551pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{Sub_{22}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{52.88535pt}{-59.47551pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{Sub_{23}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{0.0pt}{10.95293pt}\pgfsys@lineto{0.0pt}{20.53088pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{{}} {{{{{}}{}{}{}{}{{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{8.5528pt}{-6.84222pt}\pgfsys@lineto{27.00462pt}{-21.6037pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{27.81317pt}{-36.2052pt}\pgfsys@lineto{15.01941pt}{-48.99808pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{35.56595pt}{-39.41661pt}\pgfsys@lineto{35.56595pt}{-48.99808pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{43.31873pt}{-36.2052pt}\pgfsys@lineto{56.11249pt}{-48.99808pt}\pgfsys@stroke\pgfsys@invoke{ } \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope{}{}{}\hss}\pgfsys@discardpath\pgfsys@invoke{\lxSVG@closescope }\pgfsys<EMAIL_ADDRESS>\leavevmode\hbox to149.86pt{\vbox to100.79pt{\pgfpicture\makeatletter\hbox{\hskip 71.37189pt\lower-64.61296pt\hbox to0.0pt{\pgfsys@beginscope\pgfsys@invoke{ }\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\pgfsys@setlinewidth{0.4pt}\pgfsys@invoke{ }\nullfont\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\hbox to0.0pt{\pgfsys@beginscope\pgfsys@invoke{ }{{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{}}{} {}{}{}{}{}{}{}{}{}{{}\pgfsys@moveto{38.52913pt}{-14.22638pt}\pgfsys@curveto{38.52913pt}{-7.71782pt}{21.27931pt}{-2.44173pt}{0.0pt}{-2.44173pt}\pgfsys@curveto{-21.27931pt}{-2.44173pt}{-38.52913pt}{-7.71782pt}{-38.52913pt}{-14.22638pt}\pgfsys@curveto{-38.52913pt}{-20.73494pt}{-21.27931pt}{-26.01103pt}{0.0pt}{-26.01103pt}\pgfsys@curveto{21.27931pt}{-26.01103pt}{38.52913pt}{-20.73494pt}{38.52913pt}{-14.22638pt}\pgfsys@closepath\pgfsys@moveto{0.0pt}{-14.22638pt}\pgfsys@stroke\pgfsys@invoke{ } }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{-23.91118pt}{-16.72638pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{op}(\mathtt{op_{1}},\mathtt{op_{2}})$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{-15.72226pt}{26.00832pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{System}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{-68.03888pt}{-59.47551pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{Sub_{11}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{-32.47293pt}{-59.47551pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{Sub_{12}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{0.0pt}{-2.24173pt}\pgfsys@lineto{0.0pt}{20.53088pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{-14.7728pt}{-25.3049pt}\pgfsys@lineto{-46.36618pt}{-48.99808pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{-5.92259pt}{-26.07002pt}\pgfsys@lineto{-17.3873pt}{-48.99808pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{-4.02017pt}{-59.47551pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{Sub_{21}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{24.43259pt}{-59.47551pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{Sub_{22}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}\hbox{\hbox{{\pgfsys@beginscope\pgfsys@invoke{ }{{}{}{{ {}{}}}{ {}{}} {{}{{}}}{{}{}}{}{{}{}} { }{{{{}}\pgfsys@beginscope\pgfsys@invoke{ }\pgfsys@transformcm{1.0}{0.0}{0.0}{1.0}{52.88535pt}{-59.47551pt}\pgfsys@invoke{ }\hbox{{\definecolor{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@rgb@stroke{0}{0}{0}\pgfsys@invoke{ }\pgfsys@color@rgb@fill{0}{0}{0}\pgfsys@invoke{ }\hbox{{$\mathtt{Sub_{23}}$}} }}\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope}}} {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{0.0pt}{-2.24173pt}\pgfsys@lineto{0.0pt}{20.53088pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{1.99507pt}{-26.19511pt}\pgfsys@lineto{5.79585pt}{-48.99808pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{9.67163pt}{-25.83138pt}\pgfsys@lineto{28.97888pt}{-48.99808pt}\pgfsys@stroke\pgfsys@invoke{ } {{}}{}{ {}{}{}} {{{{{}}{ {}{}}{}{}{{}{}}}}}{}{{{{{}}{}{}{}{}{}{{}}}}}{{}}{}{}{}{}\pgfsys@moveto{16.30753pt}{-25.09679pt}\pgfsys@lineto{52.1619pt}{-48.99808pt}\pgfsys@stroke\pgfsys@invoke{ } \pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope{}{}{}\hss}\pgfsys@discardpath\pgfsys@invoke{\lxSVG@closescope }\pgfsys@endscope\hss}}\lxSVG@closescope\endpgfpicture}}}\end{array}\end{array}$ We then compare composed operations directly to decide equality. For example, there is an operad whose operations are matrices. Composition computes a normal form for a composite operation by block diagonals and matrix multiplication, $\begin{array}[]{ccccc}\begin{array}[]{l}op:n\times(m_{1}+m_{2})\\\ op_{1}:m_{1}\times(k_{11}+k_{12})\\\ op_{2}:m_{2}\times(k_{21}+k_{22}+k_{23})\\\ \end{array}&\longmapsto&\Big{(}\ op\ \Big{)}\cdot\begin{pmatrix}\ op_{1}\ &0\\\ \\\ 0&\ op_{2}\ \\\ \end{pmatrix}.\end{array}$ Operad axioms constrain composition. For example, the axiom mentioned in 22.1 corresponds to: $\Big{(}\ op\ \Big{)}\cdot\begin{pmatrix}op_{1}&0\\\ \\\ 0&I_{m_{2}}\\\ \end{pmatrix}\cdot\begin{pmatrix}I_{k_{11}}&0&0\\\ 0&I_{k_{12}}&0\\\ 0&0&op_{2}\\\ \end{pmatrix}=\Big{(}\ op\ \Big{)}\cdot\begin{pmatrix}I_{m_{1}}&0\\\ \\\ 0&op_{2}\\\ \end{pmatrix}\cdot\begin{pmatrix}op_{1}&0&0&0\\\ 0&I_{k_{21}}&0&0\\\ 0&0&I_{k_{22}}&0\\\ 0&0&0&I_{k_{23}}\\\ \end{pmatrix}.$ The key point is that any algebra that implements the operad must satisfy _all_ of the equations that it specifies. Type discipline controls which operations can compose; equations between operations control the resulting composites. Declaring equations between operations provides additional contracts for the API. For instance, any unary operation $\mathtt{f}\colon\mathtt{X}\to\mathtt{X}$ (a loop) generates an infinite sequence of composites $\mathsf{id}_{\mathtt{X}},\mathtt{f},\mathtt{f}^{2},\mathtt{f}^{3},\ldots$. Sometimes this is a feature of the problem at hand, but in other cases we can short-circuit the infinite regress with assumptions like idempotence ($\mathtt{f}^{2}=\mathtt{f}$) or cyclicity ($\mathtt{f}^{n}=\mathsf{id}_{\mathtt{X}}$) and ensure that algebras contain no infinite loops. ### 2.4 The systems view When we apply operads to study systems, we often think of an operation $\mathtt{f}\colon\langle\mathtt{X_{i}}\rangle\to Y$ as a system architecture. Intuitively $\mathtt{Y}$ is the system and the $\mathtt{X_{1}},\ldots,\mathtt{X_{n}}$ are the components, but this is a bit misleading. It is better to think of types as boundaries or interfaces, rather than systems. Instead, $\mathtt{f}$ is the system architecture, with component interfaces $\mathtt{X_{i}}$ and environmental interface $\mathtt{Y}$. Composition formalizes the familiar idea [65, Fig. 2] that one engineer’s system is the next’s component; it acts by nesting subsystem architectures within higher-level architectures. Once we establish a system architecture, we would like to use this structure to organize our data and analyses of the system. Moreover, according to the principle of compositionality, we should be able to construct a system-level analysis from an operation by analyzing the component-level inputs and synthesizing these descriptions according to the given operations. The process of extracting computations from operations is called _functorial semantics_ , in which a model is represented as a mapping $\mathsf{M}\colon\mathbf{Syntax}\longrightarrow\mathbf{Semantics}$. The syntax defines a system-specific architectural design. Semantics are universal and provide a computational context to interpret specific models. Matrices, probabilities, proofs, and dynamical equations all have their own rules for composition, corresponding to different semantic operads. The mapping $\mathsf{M}$ encodes, for each operation, the data, assumptions and design artifacts (e.g., geometric models) needed to construct the relevant computational representations for the architecture, its components and the environment. From this, the system model as a whole is determined by composition in the semantic context. The algebras ($\mathsf{State}$,$\mathsf{KPI}$) described in 22.2 are typical examples, with syntax ${\mathcal{O}}$ and taking semantic values in sets and functions. The mappings themselves, called _functors_ , map types and operations ($\mathtt{System}$, $\mathtt{op}$) to their semantic values, while preserving how composition builds up complex operations. The functorial perspective allows complementary models–e.g. system state vs. KPIs–to be attached to the same design. This includes varying the semantic context as well as modeling details; see Sec. 5 for examples of non- deterministic semantics. Though functorial models may be radically different, they describe the _same system_ , as reflected by the overlapping syntax. In many cases, relevant models are _not_ independent, like system state and KPIs. Natural transformations, like the extraction homomorphism in 22.2, provide a means to align ostensibly independent representations. Since models are mappings, we often visualize natural transformations as a two-dimensional cells: $\textstyle{{\mathcal{O}}\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{\mathsf{State}}$$\scriptstyle{\mathsf{KPI}}$ $\Rightarrow$ $\textstyle{\mathbf{Set}.}$ (2) Formal conditions guarantee that when moving from syntax to semantics [103, 13.2] or between representations [64, 2.3.5], reasoning about how systems decompose hierarchically “just works.” Since functors and higher cells assure coherence with hierarchical decomposition, we can use them to build up a desired model in stages, working backwards from simpler models: $\textstyle{\mathcal{O}_{2}\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{\textsf{Extension}_{1}\ \ }$$\scriptstyle{\mathsf{Model}_{2}\textrm{ is defined by}}$$\textstyle{\mathbf{Sem_{2}}\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{\hskip 9.04166pt\textsf{Reduction}_{2}}$$\textstyle{\mathcal{O}_{1}\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{\textsf{Model}_{1}\textrm{ is defined by}}$$\scriptstyle{\textsf{Extension}_{2}\ \ }$$\textstyle{\mathbf{Sem}_{1}\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{\hskip 9.04166pt\textsf{Reduction}_{1}}$$\textstyle{\mathcal{O}_{0}\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{\textsf{Model}_{0}}$$\textstyle{\mathbf{Sem}_{0}}$ This is a powerful technique for at least two reasons. First, complexity can be built up in stages by layering on details. Second, complex models built at later stages are partially validated through their coherence with simpler ones. The latter point is the foundation for lazy evaluation: many coarse models can be explored before ever constructing expensive models. Separating out the different roles within a model encourages efficiency and reuse. An architecture (operation) developed for one analysis can be repurposed with strong coherence between models (algebra instances) indexed by the same conceptual types. The syntax/semantics distinction also helps address some thornier meta-modeling issues. For example, syntactic types can distinguish conceptually distinct entities while still mapping to the same semantic entities. We obtain the flexibility of structural or duck typing in the semantics without sacrificing the type safety provided by the syntax. ## 3 Main examples Though operads are general tools [64, 70, 103], we focus on two classes of operads applied to system design: wiring diagram operads and network operads. These are complementary. Wiring diagrams provide a top-down view of the system, whereas network operads are bottom-up. This section introduces 3 examples that help ground the exposition as in Fig. 2. ### 3.1 Specification Network operads describe atomic types of systems and ways to link them together with operations. These features enable: (1) specification of atomic building blocks for a domain problem; and (2) bottom up synthesis of designs from atomic systems and links. A general theory of network operads [6, 7, 73, 74] was recently developed under the Defense Advanced Research Projects Agency (DARPA) Complex Adaptive System Composition and Design Environment (CASCADE) program. Minimal data can be used to specify a functor–called a network model [6, 4.2]–which constructs a network operad [6, 7.2] customized to a domain problem. | $\mathtt{Boat}$ | $\mathtt{Helo}$ | $\mathtt{UAV}$ | $\mathtt{QD}$ ---|---|---|---|--- $\mathtt{Cut}$ | 1 | 1 | 1 | 1 $\mathtt{Boat}$ | | | 1 | 1 $\mathtt{FW}$ | | | | 1 $\mathtt{FSAR}$ | | | | 1 $\mathtt{Helo}$ | | | | 1 (a) Examples of carrying relationships in ${\mathcal{O}}_{Sail}$ $\bullet$ $\mathtt{Cut}$$\bullet$ $\mathtt{Helo}$$\bullet$ $\mathtt{QD}$ (b) Operation $\mathtt{f}\in{{\mathcal{O}}}_{Sail}$ to specify carrying Figure 3: Which types are allowed to carry other types–indicated with $1$–specify an operad ${\mathcal{O}}_{Sail}$; $\mathtt{f}$ specifies that a $\mathtt{Helo}$ ($\bullet$) and a $\mathtt{QD}$ ($\bullet$) are carried by a $\mathtt{Cut}$ ($\bullet$) and another $\mathtt{QD}$ ($\bullet$) is carried on the $\mathtt{Helo}$ ($\bullet$). The first example illustrates designs of search and rescue (SAR) architectures. The domain problem was inspired by the 1979 Fastnet Race and the 1998 Sydney to Hobart Yacht Race and we refer to it as the sailboat problem. It illustrates how network operads facilitate the specification of a model with combinatorial data called a network template. For example, Fig. 3 shows the carrying relationships between different system types to model (e.g., a $\mathtt{Boat}$ can carry a $\mathtt{UAV}$ (Unmanned Aerial Vehicle) but a $\mathtt{Helo}$ cannot). This data specifies a network operad ${\mathcal{O}}_{Sail}$ whose: (1) objects are lists of atomic system types; (2) operations describe systems carrying other systems; and (3) composition combines carrying instructions. We discuss this example in greater detail in Sec. 4. Functional Decomposition | Control Decomposition ---|--- $\bullet$ $\mathtt{Cut}$$\bullet$ $\mathtt{Helo}$$\bullet$ $\mathtt{QD}$$\mathtt{Lab}$$\mathtt{Box}$$\mathtt{Bath}$$\mathtt{Chassis}$$\mathtt{Optics}$$\mathtt{Intfr}$$\mathtt{TempSys}$$\mathtt{LengthSys}$$\mathtt{LSI}$$\mathtt{f}\left\\{\rule{0.0pt}{32.86288pt}\right.$$\mathtt{l}\left\\{\rule{0.0pt}{35.56593pt}\right.$$\left.\rule{0.0pt}{35.56593pt}\right\\}\mathtt{t}$ | $\bullet$ $\mathtt{Cut}$$\bullet$ $\mathtt{Helo}$$\bullet$ $\mathtt{QD}$$\mathtt{Lab}$$\mathtt{Box}$$\mathtt{Optics}$$\mathtt{Intfr}$$\mathtt{Chassis}$$\mathtt{Bath}$$\mathtt{Sensors}$$\mathtt{Actuators}$$\mathtt{LSI}$$\mathtt{g}\left\\{\rule{0.0pt}{32.72049pt}\right.$$\mathtt{s}\left\\{\rule{0.0pt}{35.56593pt}\right.$$\left.\rule{0.0pt}{35.56593pt}\right\\}\mathtt{a}$ Operad Equation: $\mathtt{f}(\mathtt{l},\mathtt{t})=\mathtt{g}(\mathtt{s},\mathtt{a})$ $\bullet$ $\mathtt{Cut}$$\bullet$ $\mathtt{Helo}$$\bullet$ $\mathtt{QD}$$\mathtt{Optics}$$\mathtt{Chassis}$$\mathtt{Intfr}$$\mathtt{Lab}$$\mathtt{Box}$$\mathtt{Bath}$ $\mathtt{LSI}$$\mathtt{laser}$$\mathtt{intensity}$$\mathtt{intensity}$$\mathtt{focus}$$\mathtt{drive}$$\mathtt{fringe}$$\mathtt{heat}_{1}$$\mathtt{heat}_{2}$$\mathtt{setPt}$$\mathtt{H_{2}O}$$\mathtt{temp}$$\mathtt{TempSys}$$\mathtt{LengthSys}$$\mathtt{Sensors}$$\mathtt{Actuators}$ Figure 4: An equation in a wiring diagram operad operations expresses a common refinement of hierarchies. ### 3.2 Analysis A wiring diagram operad describes the interface each system exposes, making it clear what can be put together [90, 94, 104]. The designer has to specify precisely how information and physical quantities are shared among components, while respecting their interfaces. The operad facilitates top-down analysis of a design by capturing different ways to decompose a composite system. The second example analyzes a precision-measurement system called the Length Scale Interferometer (LSI) with wiring diagrams. It helps illustrate the qualitative features of operads over and above other modeling approaches and the potential to exploit their analytic power to separate concerns. Figure 4 illustrates joint analysis of the LSI to address different aspects of the design problem: functional roles of subsystems and control of the composite system. This analysis example supports these illustrations in Sec. 5. $\mathtt{UH60}$$\mathtt{UH60}$$\mathtt{UH60}$ $\mathtt{UH60}$$\mathtt{UH60}+\mathtt{HC130}$$\mathtt{UH60}+\mathtt{HC130}$$2\mathtt{UH60}$$2\mathtt{UH60}$$\bullet$ $\mathtt{Cut}$$\bullet$ $\mathtt{Helo}$$\bullet$ $\mathtt{QD}$$\;\;a\;\;$$\;\;b\;\;$$\;\;c\;\;$$\;\;d\;\;$$\;\phantom{\Big{|}}\tau_{1}\;$$\;\phantom{\Big{|}}\tau_{2}\;$$\;\phantom{\Big{|}}\tau_{3}\;$$\;\phantom{\Big{|}}\tau_{4}\;$$\bullet$ $\mathtt{UH60}$$\bullet$ $\mathtt{HC130}$ (a) Specification of primitive tasks $:=$ transitions $\mathtt{UH60}$$\mathtt{UH60}$$\mathtt{UH60}$ $\mathtt{UH60}$$\mathtt{UH60}+\mathtt{HC130}$$\mathtt{UH60}+\mathtt{HC130}$$2\mathtt{UH60}$$2\mathtt{UH60}$$\bullet$ $\mathtt{Cut}$$\bullet$ $\mathtt{Helo}$$\bullet$ $\mathtt{QD}$$\;\;a\;\;$$\;\;b\;\;$$\;\;c\;\;$$\;\;d\;\;$$\;\phantom{\Big{|}}\tau_{1}\;$$\;\phantom{\Big{|}}\tau_{2}\;$$\;\phantom{\Big{|}}\tau_{3}\;$$\;\phantom{\Big{|}}\tau_{4}\;$$\bullet$ $\mathtt{UH60}$$\bullet$ $\mathtt{HC130}$$\phantom{{|}}\tau_{1}$$\phantom{{|}}\tau_{2}$$\phantom{{|}}\tau_{4}$$(a,0)$$(b,1)$$(c,2)$$(c,2)$$((d,d),(4,4))$ (b) Coordinate tasks to compose Figure 5: Primitive operations are composed for two UH60s ($\bullet$) to rendezvous at $c$ and maneuver together to $d$. Each primitive operation is indexed by a transition; types and space-time points must match to compose. ### 3.3 Synthesis The third example describes the automated design of mission task plans for SAR using network operads. The SAR tasking example illustrates the expressive power of applying existing operads and their potential to streamline and automate design synthesis. Fig. 5(a) is analogous to Fig. 3, but whereas a sparse matrix specify an architecture problem, here a Petri net is used to model coordinated groups of agents. For the SAR tasking problem, much of the complexity results from agents’ need to coordinate in space and time–e.g. when a helicopter is refueled in the air, as in $\tau_{3}$ of Fig. 5(a). To facilitate coordination, the types of the network operad are systematically extended via a network model whose target categories add space and time dimensions; compare, e.g., [7]. In this way, task plans are constrained at the level of syntax to enforce these key coordination constraints; see, e.g., Fig. 5(b) where two UH60s at the same space-time point ($\bullet$ $(c,2)$) maneuver together to $d$. We describe automated synthesis for this example in Sec. 6. ## 4 Cookbook modeling of domain problems In this section we describe some techniques for constructing operads and their algebras, using an example-driven, cookbook-style approach. We emphasize recent developments for network operads and dive deeper into the SAR architecture problem. ### 4.1 Network models The theory of network models provides a general method to construct an operad ${\mathcal{O}}$ by mixing combinatorial and compositional structures. Note that this lives one level of abstraction _above_ operads; we are interested in _constructing_ a language to model systems–e.g. for a specific domain. This provides a powerful alternative to coming up with operads one-by-one. A general construction allows the applied practitioner to cook-up a domain- specific syntax to compose systems by specifying some combinatorial ingredients. The first step is to specify what the networks to be composed by ${\mathcal{O}}$ look like. Often this is some sort of graph, but what kind? Are nodes typed (e.g., colored)? Are edges symmetric or directed? Are loops or parallel edges allowed? What about $n$-way relationships for $n>2$ (hyper- edges)? We can mix, match and combine such combinatorial data to define different _network models_ , which specify the system types and kinds of relationships between them relevant to some domain problem. The network model describes the operations we need to compose the networks specific to the problem at hand. Three compositional structures which describe the algebra of operations. The _disjoint_ or _parallel_ structure combines two operations for networks with $m$ and $n$ nodes, respectively, into a single operation for networks with $m+n$ nodes. More restrictively, the _overlay_ or _in series_ structure superimposes two operations to design networks on $n$ nodes. The former structure combines separate operations to support modular development of designs; the latter supports an incremental design process, either on top of existing designs or from scratch. The last ingredient permutes nodes in a network, which assures coherence between different ordering of the nodes. This last structure is often straightforward to specify. If it is not, one should consider if symmetry is being respected in a natural way. We can distill the main idea behind overlay by asking, what happens when we add an edge to a network? It depends on the kind of network being composed by ${\mathcal{O}}$: In a simple graph | but in a labeled graph | and in a multigraph ---|---|--- | $\textstyle{x\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\textstyle{y}$ | | $\textstyle{x\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\textstyle{y}$ | | $\textstyle{x\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\textstyle{y}$ | \+ $\textstyle{x\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\textstyle{y}$ | | +$\textstyle{x\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\textstyle{y}$ | | \+ $\textstyle{x\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\textstyle{y}$ | $\textstyle{x\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\textstyle{y,}$ | | $\textstyle{x\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{2}$$\textstyle{y,}$ | | $\textstyle{x\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\textstyle{y.}$ These differences are controlled by a _monoid_ 333A set with a binary operation, usually written $\cdot$ unless the operation is commutative ($m+n=n+m$). A monoid is always associative, $\ell\cdot(m\cdot n)=(\ell\cdot m)\cdot n$ and has a unit $e$ satisfying $e\cdot m=m=m\cdot e$–e.g. multiplication of $n\times n$ matrices., which provides each $+$ shown. Above, the monoids are bitwise OR, addition, and maximum, respectively. As a further example, if edge addition is controlled by $\mathbb{Z}/2\mathbb{Z}$ then $+$ will have a toggling effect. Consider simple graphs. Given a set of nodes $\mathtt{n}$, write $U_{\mathtt{n}}$ for the set of all undirected pairs $i\neq j$ (a.k.a. simple edges $i\mathchar 45\relax\mathchar 45\relax j$), so that $|U_{\mathtt{n}}|={\binom{|\mathtt{n}|}{2}}$. Then we can represent a simple graph over $\mathtt{n}$ as a $U_{\mathtt{n}}$-indexed vector of bits $\langle b_{i\mathchar 45\relax\mathchar 45\relax j}\rangle$ describing which edges to ‘turn on’ for a design. Each bit defines whether or not to add an $i\mathchar 45\relax\mathchar 45\relax j$ edge to the network and the overlay compositional structure is given by the monoid $\mathsf{SG}(\mathtt{n}):=\mathbf{Bit}^{U_{\mathtt{n}}}$, whose $+$ is bitwise OR for the product over simple edges–i.e. adding $i\mathchar 45\relax\mathchar 45\relax j$ then adding $i\mathchar 45\relax\mathchar 45\relax j$ is the same as adding $i\mathchar 45\relax\mathchar 45\relax j$ a single time. The disjoint structure $\sqcup:\mathsf{SG}(\mathtt{m})\times\mathsf{SG}(\mathtt{n})\longrightarrow\mathsf{SG}(\mathtt{m+n})$ forms the disjoint sum of the graphs $g$ and $h$. Finally, permutations act by permuting the nodes of a simple graph. Together, these compositional structures define a network model $\mathsf{SG}\colon\mathcal{S}\to\mathbf{Mon}$ which determines how operations are composed in the constructed network operad; see, Fig. 6 or [6, 3.2, 7.2] for complete technical details. $\mathtt{UH60}$$\mathtt{UH60}$$\mathtt{UH60}$ $\mathtt{UH60}$$\mathtt{UH60}+\mathtt{HC130}$$\mathtt{UH60}+\mathtt{HC130}$$2\mathtt{UH60}$$2\mathtt{UH60}$$\bullet$ $\mathtt{Cut}$$\bullet$ $\mathtt{Helo}$$\bullet$ $\mathtt{QD}$$\;\;a\;\;$$\;\;b\;\;$$\;\;c\;\;$$\;\;d\;\;$$\;\phantom{\Big{|}}\tau_{1}\;$$\;\phantom{\Big{|}}\tau_{2}\;$$\;\phantom{\Big{|}}\tau_{3}\;$$\;\phantom{\Big{|}}\tau_{4}\;$$\bullet$ $\mathtt{UH60}$$\bullet$ $\mathtt{HC130}$$\phantom{{|}}\tau_{1}$$\phantom{{|}}\tau_{2}$$\phantom{{|}}\tau_{4}$$(a,0)$$(b,1)$$(c,2)$$(c,2)$$((d,d),(4,4))$$2$$2$$2$$2$$2$$4$$\sqcup$$\mapsto$$2$$2$$4$$4$$+$$\mapsto$$2$$2$$4$ Figure 6: Parallel ($\sqcup$) and in series ($+$) compositional structures define how to combine operations. This definition has an analogue for $\mathbb{N}$-weighted graphs, $\mathsf{LG}(\mathtt{n}):=(\mathbb{N},+)^{U_{\mathtt{n}}}$, with overlay given by sum of edge weights and another for multi-graphs, $\mathsf{MG}(\mathtt{n}):=(\mathbb{N},\max)^{U_{\mathtt{n}}}$, with overlay equivalent to union of multisets; see [6, 3.3, 3.4] for details. More generally, we can label edges with the elements of _any_ monoid. Many of these examples are strange—binary addition makes edges cancel when they add—but their formal construction is straightforward; see [6, Thm. 3.1]. Equivalently, we can view the undirected edges in $U_{\mathtt{n}}$ as generators, subject to certain idempotence and commutativity relations: $\mathsf{SG}(\mathtt{n}):=\langle e\in U_{\mathtt{n}}|e\cdot e=e,e\cdot e^{\prime}=e^{\prime}\cdot e\rangle.$ Here the idempotence relations come from $\mathbf{Bit}$ while the commutativity relations promote the single copies of $\mathbf{Bit}$ for each $i\mathchar 45\relax\mathchar 45\relax j$ to a well- defined network model. Similar tricks work for lots of other network templates; we just change the set of generators to allow for new relationships. For example, to allow self-loops, we add loop edge generators $L_{\mathtt{n}}=\mathtt{n}+U_{\mathtt{n}}$ to express relationships from a node $i$ to itself. Likewise, network operads for directed graphs can be constructed by using generators $D_{\mathtt{n}}=\mathtt{n}\times\mathtt{n}$, and one can also introduce higher-arity relationships. In all cases, the formal definition of a network model assures that all the combinatorial and compositional ingredients work well together; one precise statement of “working well together” is given in [6, 2.3]. Once a _network template_ —which expresses minimal data to declare the ingredients for a network model—is codified in a theorem as in [6, 3.1], it can be reused in a wide variety of domains to set up the specifics of composition. ### 4.2 Cooking with operads The prototype for network operads is a simple network operad, which models only one kind of thing, such as aircraft. The types of a simple network operad are natural numbers, which serve to indicate how many aircraft are in a design. Operations of the simple network operad are simple graphs on some number of vertices. For example, Fig. 6 above shows a simple network operad to describe a design for point-to-point communication between aircraft. Structural network operads extend this prototype in two directions: (1) a greater diversity of things-to-be-modeled is supported by an expanded collection of types; and (2) more sorts of links or relationships between things are expressed via operations. To illustrate the impact of network templates, suppose we are modeling heterogeneous system types with multiple kinds interactions. For simplicity we consider simple interactions, which can be undirected or directed. A _network template_ need only declare the _primitive_ ways system types can interact to define a network model–e.g. a list of tuples $\textrm{(directed : carrying, }\mathtt{Helo},\textrm{ }\mathtt{Cut}\textrm{)}$. This data is minimal in two ways: (1) _any_ framework must provide data to specify potentially valid interactions; and (2) this approach allows _only_ those interactions that make sense upon looking at the types of the systems involved. Thus, interactions must be syntactically correct when constructing system designs. Presently, we will consider an example from the DARPA CASCADE program: the sailboat problem introduced in 33.1. This SAR application problem was inspired by the 1979 Fastnet Race and the 1998 Sydney to Hobart Yacht Race, in which severe weather conditions resulted in many damaged vessels distributed over a large area. Both events were tragic, with 19 and 6 deaths, respectively, and remain beyond the scale of current search and rescue planning. Various larger assets—e.g. ships, airplanes, helicopters—could be based at ports and ferry smaller search and rescue units—e.g. small boats, quadcopters—to the search area. Specifically, there were 8 atomic types to model: $P=\\{\mathtt{Port},\mathtt{Cut},\mathtt{Boat},\mathtt{FW},\mathtt{FSAR},\mathtt{Helo},\mathtt{UAV},\mathtt{QD}\\}.$ The primary relationship to specify a structural design is various assets carrying another types, so only one kind of interaction is needed: carrying. This relationship is directed; e.g., a cutter ($\mathtt{Cut}$) can carry a helicopter ($\mathtt{Helo}$) but not the other way around. Specifying allowed relationships amounts to specifying pairs of type $(p,p^{\prime})\in P\times P$ such that type $p^{\prime}$ can carry type $p$; see Fig. 3 for examples. Fig. 3 data is extended to: (1) specify to that $\mathtt{Port}$ can carry all types other than $\mathtt{Port}$, $\mathtt{UAV}$ and $\mathtt{QD}$; (2) conform to an input file format to declare simple directed or undirected interactions, e.g, the JSON format in Fig. 7. {‘colors’ : [‘port’, ‘cut’, …, ‘qd’], ‘directed’ : { ‘carrying’: { ‘cut’: [‘port’], ‘boat’: [‘port’, ‘cut’], …, ‘qd’: [‘cut’, …, ‘helo’] } } } (a) Network template data to specify the operad ${{\mathcal{O}}}_{Sail}$ $\mathtt{UH60}$$\mathtt{UH60}$$\mathtt{UH60}$ $\mathtt{UH60}$$\mathtt{UH60}+\mathtt{HC130}$$\mathtt{UH60}+\mathtt{HC130}$$2\mathtt{UH60}$$2\mathtt{UH60}$$\bullet$ $\mathtt{Cut}$$\bullet$ $\mathtt{Helo}$$\bullet$ $\mathtt{QD}$$\;\;a\;\;$$\;\;b\;\;$$\;\;c\;\;$$\;\;d\;\;$$\;\phantom{\Big{|}}\tau_{1}\;$$\;\phantom{\Big{|}}\tau_{2}\;$$\;\phantom{\Big{|}}\tau_{3}\;$$\;\phantom{\Big{|}}\tau_{4}\;$$\bullet$ $\mathtt{UH60}$$\bullet$ $\mathtt{HC130}$$\phantom{{|}}\tau_{1}$$\phantom{{|}}\tau_{2}$$\phantom{{|}}\tau_{4}$$(a,0)$$(b,1)$$(c,2)$$(c,2)$$((d,d),(4,4))$$2$$2$$2$$2$$2$$4$$\sqcup$$\mapsto$$2$$2$$4$$4$$+$$\mapsto$$2$$2$$4$$\bullet$ $\mathtt{Cut}$$\bullet$ $\mathtt{Helo}$$\bullet$ $\mathtt{QD}$ (b) Example operation $\mathtt{f}\in{{\mathcal{O}}}_{Sail}$ Figure 7: After specifying ${{\mathcal{O}}}_{Sail}$, $\mathtt{f}$ places a $\mathtt{QD}$ ($\bullet$) on a $\mathtt{Cut}$ ($\bullet$) and another $\mathtt{QD}$ ($\bullet$) on a $\mathtt{Helo}$ ($\bullet$). If another type of system or kind of interaction is needed, then the file is appropriately extended. For example, we can include buoys by appending Buoy to the array of colors and augmenting the relationships in the carrying node. Or, we can model the undirected (symmetric) relationship of communication by including an entry such as ‘undirected’: {‘communication’: {‘port‘: [‘cut’, ...], ...}}. Moreover, modifications to network templates–such as ignoring (undirected : communication) or combining $\mathtt{QD}$ and $\mathtt{UAV}$ into a single type–naturally induce mappings between the associated operads [6, 5.8]. ### 4.3 Cooking with algebras Because all designs are generated from primitive operations to add edges, it is sufficient to define how primitive operations act in order to define an algebra. For the sailboat problem, semantics are oriented to enable the delivery of a high capacity for search—known in the literature as search effort [97, 3.1]—in a timely manner. Given key parameters for each asset–e.g. speed, endurance, search efficiency across kinds of target and conditions, parent platform, initial locations–and descriptions of the search environment–e.g. expected search distribution, its approximate evolution over time–the expected number of surviving crew members found by the system can be estimated [97, Ch. 3]. Among these data, the parent platform and initial locations vary within a scenario and the rest describe the semantics of a given scenario. In fact, we assume all platforms must trace their geographical location to one of a small number of base locations, so that the system responds from bases, but is organized to support rapid search. Once bases are selected, the decision problem is a choice of operation: what to bring (type of the composite system) and how to organize it (operation to carry atomic systems). Data for the operational context specifies a particular algebra; see, e.g., Table 2. Just as for the operad, this data is lightweight and configurable. Table 2: Example properties captured in algebra for sailboat problem including time on station (ToS), speed for search (S) and max speed (R), and sweep widths measuring search efficiency for target types person in water (PIW), crew in raft (CIR) and demasted sailboats (DS) adrift. Type | Cost ($) | ToS (hr) | Speed (kn) S R | Sweep Width (nmi) PIW CIR DS ---|---|---|---|--- $\mathtt{Cut}$ | 200M | $\infty$ | 11 28 | 0.5 4.7 8.5 $\mathtt{Boat}$ | 500K | 6 | 22 35 | 0.4 4.2 7.5 $\mathtt{FW}$ | 60M | 9 | 180 220 | 0.1 2.2 7.6 $\mathtt{FSAR}$ | 72M | 10 | 180 235 | 0.5 12.1 16.6 $\mathtt{Helo}$ | 9M | 4 | 90 180 | 0.5 1.5 4.8 $\mathtt{UAV}$ | 250K | 3 | 30 45 | 0.5 1.8 4.5 $\mathtt{QD}$ | 15K | 4 | 35 52 | 0.5 1.5 4.8 Related cookbook approaches. Though we emphasized network operads, the generators approach is often studied and lends itself to encoding such combinatorially data with a “template,” in a cookbook fashion. The generators approach to “wiring” has been developed into a theory of hypergraph categories [38, 41], which induce wiring diagram operads. Explicit presentations for various wiring diagram operads are given in [104]. Augmenting monoidal categories with combinatorially specified data has also been investigated, e.g. in [42]. ## 5 Functorial Systems Analysis In this section we demonstrate the use of functorial semantics in systems analysis. As in 22.4, a functor establishes a relationship between a syntactic or combinatorial model of a system (components, architecture) and some computational refinement of that description. This provides a means to consider a given system from different perspectives, and also to relate those viewpoints to one another. To drive the discussion, we will focus on the Length Scale Interferometer (LSI) and its wiring diagram model introduced in 33.2. ### 5.1 Wiring diagrams Operads can be applied to organize both qualitative and quantitative descriptions of hierarchical systems. Because operations can be built up iteratively from simpler ones to specify a complete design, different ways to build up a given design provide distinct avenues for analysis. Figure 4 shows a wiring diagram representation of a precision measurement instrument called the Length Scale Interferometer (LSI) designed and operated by the US National Institute of Standards and Technology (NIST). Object types are system or component boundaries; Fig. 4 has: 6 components, the exterior, and 4 interior boundaries. Each boundary has an interface specifying its possible interactions, which are implicit in Fig. 4, but define explicit types in the operad. An operation in this context represents one step in a hierarchical decomposition, as in 22.1. For example, the blue boxes in Fig. 4 represent a functional decomposition of the LSI into length-measurement and temperature- regulation subsystems: $\mathtt{f}\colon\mathtt{LengthSys},\mathtt{TempSys}\to\mathtt{LSI}$. These are coupled via (the index of refraction of) a $\mathtt{laser}$ interaction and linked to interactions at the system boundary. The operation $\mathtt{f}$ specifies the connections between blue and black boundaries. Composition in a wiring diagram operad is defined by nesting. For this functional decomposition, two further decompositions $\mathtt{l}$ and $\mathtt{t}$ describe the components and interactions within $\mathtt{LengthSys}$ and $\mathtt{TempSys}$, respectively. The wiring diagram in Fig. 4 is the composite $\mathtt{f}(\mathtt{l},\mathtt{t})$. This approach cleanly handles multiple decompositions. Here the red boxes define a second, control-theoretic decomposition $\mathtt{g}:\mathtt{Sensors},\mathtt{Actuators}\to\mathtt{LSI}$. Unsurprisingly, the system is tightly coupled from this viewpoint, with heat flow to maintain the desired temperature, mechanical action to modify the path of the laser, and a feedback loop to maintain the position of the optical focus based on measured intensity. The fact that these two viewpoints specify the _same_ system design is expressed by the equation: $\mathtt{f}(\mathtt{l},\mathtt{t})=\mathtt{g}(\mathtt{s},\mathtt{a})$; see 22.3 for related discussion. ### 5.2 A probabilistic functor Wiring diagrams can be applied to document, organize and validate a wide variety of system-specific analytic models. Each model is codified as an algebra, a functor from syntax to semantics (22.4). For the example of this section, all models have the same source (syntax), indicating that we are considering the same system, but the target semantics vary by application. We have already seen some functorial models: the algebras in 44.3. These can be interpreted as functors from the carrying operad ${\mathcal{O}}_{Sail}$ to the operad of sets and functions $\mathbf{Set}$. Though $\mathbf{Set}$ is the “default” target for operad algebras, there are many alternative semantic contexts tailored to different types of analysis. Here we target an operad of probabilities $\mathbf{Prob}$, providing a simple model of nondeterministic component failure. The data for the functor is shown in Table 3. Model data is indexed by operations444Types and operations, more generally, but the types carry no data in this simple example. in the domain, an operad $\mathcal{W}$ extracted from the wiring diagram in Fig. 4. The functor assigns each operation to a probability distribution that specifies the chance of a failure in each subsystem, assuming some error within the super-system. For example, the length measurement and temperature regulation subsystems are responsible for 40% and 60% of errors in the LSI, respectively. This defines a Bernoulli distribution $P_{\mathtt{f}}$. Similarly, the decomposition $\mathtt{t}$ of the temperature system defines a categorical distribution with 3 outcomes: $\mathtt{Box}$, $\mathtt{Bath}$ and $\mathtt{Lab}$. Relative probabilities compose by multiplication. This allows us to compute more complex distributions for nested diagrams. For the operation shown in Fig. 4, this indicates that the bath leads to nearly half of all errors ($60\%\times 80\%=48\%$) in the system. Operad equations must be preserved in the semantics. Since $\mathtt{f}(\mathtt{l},\mathtt{t})=\mathtt{g}(\mathtt{s},\mathtt{a})$, failure probabilities of source components don’t depend on whether we think of them in terms of functionality or control. For the bath, this relative failure probability is $\overbrace{60\%}^{P_{\mathtt{f}}}\times\overbrace{80\%}^{P_{\mathtt{t}}}=48\%=\overbrace{72\%}^{P_{\mathtt{g}}}\times\overbrace{66.7\%}^{P_{\mathtt{a}}},$ and five analogous equations hold for the other source components. Functorial semantics separates concerns: different operad algebras answer different questions. Here we considered _if_ a component will fail. The LSI example is developed further in [20, 4] by a second algebra describing _how_ a component might fail, with Boolean causal models to propagate failures. The two perspectives are complementary, and loc. cit. explores integrating them with algebra homomorphisms (22.4). Table 3: Failure probabilities form an operad algebra for LSI component failure. $P_{\mathtt{f}}$ | $ls$ | $\mapsto$ | 40% | $P_{\mathtt{g}}$ | $sn$ | $\mapsto$ | 28% ---|---|---|---|---|---|---|--- $ts$ | $\mapsto$ | 60% | $ac$ | $\mapsto$ | 72% $P_{\mathtt{l}}$ | $in$ | $\mapsto$ | 10% | $P_{\mathtt{s}}$ | $lb$ | $\mapsto$ | 21.4% $\mathtt{op}$ | $\mapsto$ | 30% | $bt$ | $\mapsto$ | 21.4% $ch$ | $\mapsto$ | 60% | $\mathtt{op}$ | $\mapsto$ | 42.9% $P_{\mathtt{t}}$ | $ba$ | $\mapsto$ | 80% | $in$ | $\mapsto$ | 14.3% $bx$ | $\mapsto$ | 10% | $P_{\mathtt{a}}$ | $ch$ | $\mapsto$ | 33.3% $lb$ | $\mapsto$ | 10% | $ba$ | $\mapsto$ | 66.7% ### 5.3 Interacting semantics Its toy-example simplicity aside, the formulation of a failure model $\mathcal{W}\to\mathbf{Prob}$, as in Table 3 is limited in at least two respects. First, it tells us _which_ components fail, but not _how_ or _why_. Second, the model is static, but system diagnosis is nearly always a dynamic process. We give a high-level sketch of an extended analysis to illustrate the integration of overlapping functorial models. The first step is to characterize some additional information about the types in $\mathcal{W}$ (i.e., system boundaries). We start with the dual notions of _requirements_ and _failure modes_. For example, in the temperature regulation subsystem of the LSI we have $\begin{array}[]{ccc}T_{\mathtt{laser}}\leq 20.02^{\circ}\textrm{C}&\leftrightarrow&T_{\mathtt{laser}}\textsf{ too high}\\\\[2.15277pt] 19.98^{\circ}\textrm{C}\leq T_{\mathtt{laser}}&\leftrightarrow&T_{\mathtt{laser}}\textsf{ too low}\\\ \vdots&&\vdots\end{array}$ Requirements at different levels of decomposition are linked by traceability relations. These subsystem requirements trace up to the measurement uncertainty for the LSI as a whole. Dually, an out-of-band temperature at the subsystem level can be traced back to a bad measurement in the $\mathtt{Box}$ enclosure, a short in the $\mathtt{Bath}$ heater or fluctuations in the $\mathtt{Lab}$ environment. Traceability is compositional: requirements decompose and failures bubble up. This defines an operad algebra555Many operads are defined from ordinary categories using a symmetric monoidal products [56, 7]. If a category carries more than one product, we use a superscript to indicate which is in use. The the disjoint union (+) corresponds to the disjunctive composition “a failure in one component _or_ another; soon we will use the Cartesian product $\times$ to consider the conjunctive relationship between “the state of one component _and_ the other”. $\mathsf{Req}:\mathcal{W}\to\mathbf{Rel}^{+}$. Functoriality expresses the composition of traceability requirement across levels. See [20, 5] discussion of how to link these relations with Table 3 data. For dynamics, we need _state_. We start with a state space for each interaction among components. For example, consider the $\mathtt{laser}$ interaction coupling $\mathtt{Chassis}$, $\mathtt{Intfr}$, and $\mathtt{Box}$. The most relevant features of the laser are its vacuum wavelength $\lambda_{0}$ and the ambient temperature, pressure and humidity (needed to correct for refraction). This corresponds to a four-dimensional state-space (or a subset thereof) $\mathsf{State}(\mathtt{laser})\cong\overbrace{[-273.15,\infty)]}^{T_{\mathtt{laser}}}\times\overbrace{[0,\infty)]}^{P_{\mathtt{laser}}}\times\overbrace{[0,1]}^{RH_{\mathtt{laser}}}\times\overbrace{[0,\infty)}^{\lambda_{0}}\subseteq\mathbb{R}^{4}.$ A larger product defines an _external state space_ at each system boundary $\begin{array}[]{rl}\mathsf{State}(\mathtt{TempSys})=&\mathsf{State}(\mathtt{laser})\times\mathsf{State}(\mathtt{temp})^{2}\times\mathsf{State}(\mathtt{setPt})\times\mathsf{State}(\mathtt{H_{2}O})\\\\[4.30554pt] \mathsf{State}(\mathtt{Box})=&\mathsf{State}(\mathtt{laser})\times\mathsf{State}({\mathtt{temp}})\times\mathsf{State}(\mathtt{heat})^{2}\\\ \vdots\\\ \end{array}$ Similarly, we can define an _internal state space_ for each operation by taking the product over all the interactions that appear in that diagram. We can decompose the internal state space in terms of either the system boundary or the components666Coupled variables are formalized through a partial product called the pullback, a common generalization of the Cartesian product, subset intersection and inverse image constructions.: $\begin{array}[]{rcl}\mathsf{State}(\mathtt{f})&\cong&\mathsf{State}(\mathtt{LSI})\times\overbrace{\mathsf{State}(\mathtt{laser})}^{\mathclap{\textrm{hidden variable}}}\\\\[4.30554pt] &\cong&\mathsf{State}(\mathtt{LengthSys})\times\mathclap{\underbrace{\underset{\mathsf{State}(\mathtt{laser})}{}}_{\textrm{coupled variable}}}\mathsf{State}(\mathtt{TempSys})\end{array}$ The projections from these (partial) products form a relation, and these compose to define a functor $\mathcal{W}\to\mathbf{Rel}^{\times}$: $\textstyle{\mathtt{LSI}}$$\textstyle{\mathtt{LengthSys},\ \mathtt{TempSys}\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{\mathtt{f}\in\mathcal{W}}$$\textstyle{\mathsf{State}(\mathtt{LSI})}$$\textstyle{{\overbrace{\mathsf{State}(\mathtt{f})}^{\in\mathbf{Rel}^{\times}}}\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{p_{0}}$$\scriptstyle{\langle p_{1},p_{2}\rangle}$$\textstyle{\mathsf{State}(\mathtt{LengthSys})\times\mathsf{State}(\mathtt{TempSys})}$$\textstyle{{\overbrace{\left\langle{\scriptscriptstyle\ldots},\mbox{$\tiny\begin{pmatrix}1\\\ 2\\\ 3\end{pmatrix}$},\mbox{$\tiny\begin{pmatrix}8\\\ 9\end{pmatrix}$},{\scriptscriptstyle\ldots}\right\rangle}^{p_{0}(s)}}}$$\textstyle{{\overbrace{\left\langle{\scriptscriptstyle\ldots},\underbrace{\mbox{$\tiny\begin{pmatrix}1\\\ 2\\\ 3\end{pmatrix}$}}_{\mathclap{\mathtt{fringe}}},\underbrace{\mbox{$\tiny\begin{pmatrix}4\\\ 5\\\ 6\\\ 7\end{pmatrix}$}}_{\mathclap{\mathtt{laser}}},\underbrace{\mbox{$\tiny\begin{pmatrix}8\\\ 9\end{pmatrix}$}}_{\mathclap{\mathtt{H_{2}O}}},{\scriptscriptstyle\ldots}\right\rangle}^{s\in\mathsf{State}(\mathtt{f})}}\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\textstyle{\left\langle{\overbrace{\left\langle{\scriptscriptstyle\ldots},\mbox{$\tiny\begin{pmatrix}1\\\ 2\\\ 3\end{pmatrix}$},\mbox{$\tiny\begin{pmatrix}4\\\ 5\\\ 6\\\ 7\end{pmatrix}$}\right\rangle}^{p_{1}(s)}},{\overbrace{\left\langle\mbox{$\tiny\begin{pmatrix}4\\\ 5\\\ 6\\\ 7\end{pmatrix}$},\mbox{$\tiny\begin{pmatrix}8\\\ 9\end{pmatrix}$},{\scriptscriptstyle\ldots}\right\rangle}^{p_{2}(s)}}\right\rangle}$ Each requirement $R\in\mathsf{Req}(\mathtt{X})$ defines a subset $|R|\subseteq\mathsf{State}(\mathtt{X})$, and a state is _valid_ if it satisfies all the requirements: $\mathsf{Val}(\mathtt{X})=\bigcap_{R}|R|$. Using pullbacks (inverse image) we can translate validity to internal state spaces in two different ways. External validity (left square) checks that a system satisfies its contracts; joint validity (right square) couples component requirements to define the allowed joint states. $\textstyle{\mathsf{Val}(\mathtt{LSI})\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\textstyle{\mathsf{XVal}(\mathtt{f})\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\textstyle{\mathsf{JVal}(\mathtt{f})\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\textstyle{\mathsf{Val}(\mathtt{LengthSys})\times\mathsf{Val}(\mathtt{TempSys})\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\textstyle{\mathsf{State}(\mathtt{LSI})}$$\textstyle{\mathsf{State}(\mathtt{f})\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{p_{0}}$$\textstyle{\mathsf{State}(\mathtt{f})\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{\langle p_{1},p_{2}\rangle}$$\textstyle{\mathsf{State}(\mathtt{LengthSys})\times\mathsf{State}(\mathtt{TempSys})}$ A requirement model is _sound_ if joint validity entails external validity, corresponding to the dashed arrow above. With some work, one can show that these diagrams form the operations in an operad of entailments $\mathbf{Ent}$; see [21, 6] for a similar construction. The intuition is quite clear: $\begin{array}[]{crcl}&\textrm{component reqs.}&\Rightarrow&\textrm{subsystem reqs.}\\\ +&\textrm{subsystem reqs.}&\Rightarrow&\textrm{system reqs.}\\\ \hline\cr&\textrm{component reqs.}&\Rightarrow&\textrm{system reqs.}\\\ \end{array}$ There is a functor $\mathsf{Context}:\mathbf{Ent}\to\mathbf{Rel}^{\times}$, which extracts the relation across the bottom row of each entailment. Noting that the $\mathsf{State}$ relations occur in the validity entailment, we can reformulate requirement specification as a _lifting problem_ (Fig. 8(a)): given functors $\mathsf{State}$ and $\mathsf{Context}$, find a factorization $\mathsf{Val}$ making the triangle commute. The second and third diagrams (Fig. 8(b)–8(c)) show how to extend the lifting problem with prior knowledge, in this case a top-level requirement and a known (e.g., off the shelf) component capability. $\textstyle{\mathbf{Ent}\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{\mathsf{Context}}$$\textstyle{\mathcal{W}\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{\mathsf{State}}$$\scriptstyle{\mathsf{Val}}$$\textstyle{\mathbf{Rel}^{\times}}$ (a) Free Specification $\textstyle{1\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{\mathtt{LSI}}$$\scriptstyle{\lambda_{0}\cdot\mathtt{fringe}=L_{\mathtt{drive}}}$$\textstyle{\mathbf{Ent}\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{\mathsf{Context}}$$\textstyle{\mathcal{W}\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{\mathsf{State}}$$\scriptstyle{\mathsf{Val}}$$\textstyle{\mathbf{Rel}^{\times}}$ (b) Top-down requirement $\textstyle{1\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{\mathtt{Intfr}}$$\scriptstyle{\varepsilon_{\mathtt{fringe}}\leq\frac{\lambda_{0}}{8}}$$\textstyle{\mathbf{Ent}\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{\mathsf{Context}}$$\textstyle{\mathcal{W}\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{\mathsf{State}}$$\scriptstyle{\mathsf{Val}}$$\textstyle{\mathbf{Rel}^{\times}}$ (c) Bottom-up requirement Figure 8: Requirement specification expressed as lifting problems. Finally we are ready to admit dynamics, but it turns out that we have already done most of the work. All that is needed is to modify the spaces attached to our interactions. In particular, we can distinguish between static and dynamic state variables; for the $\mathtt{laser}$, $T$, $P$ and $RH$ are dynamic while $\lambda_{0}$ is static. Now we replace the static values $T,P,RH\in\mathbb{R}$ by functions $T(t),P(t),RH(t)\in\mathbb{R}^{T}$, thought of as _trajectories_ through the state space over a timeline $t\in\tau$. For example, we have $\mathsf{Traj}(\mathtt{laser})\subseteq\overbrace{(\mathbb{R}^{\tau})^{3}}^{T,P,RH}\times\overbrace{\mathbb{R}}^{\lambda_{0}}.$ From this, we construct $\mathsf{Traj}:\mathcal{W}\to\mathbf{Rel}^{\times}$ using exactly the same recipe as above. Trajectories and states are related by a pair of algebra homomorphisms $and$const$.Thefirstpicksoutainstantaneousstateforeachpointintime,whilethesecondidentifiesconstantfunctions,whichdescribefixed- pointsofthedynamics:$$\begin{array}[]{ccc}\textrm{Global view}&&\textrm{Local view}\\\ \hline\cr\vbox{\lx@xy@svg{\hbox{\raise 0.0pt\hbox{\kern 8.1389pt\hbox{\ignorespaces\ignorespaces\ignorespaces\hbox{\vtop{\kern 0.0pt\offinterlineskip\halign{\entry@#!@&&\entry<EMAIL_ADDRESS>0.0pt\hbox{\hbox{\kern 0.0pt\raise 0.0pt\hbox{\hbox{\kern 3.0pt\raise 0.0pt\hbox{$\textstyle{}$}}}}}}}{\hbox{\kern 32.1389pt\raise 0.0pt\hbox{\hbox{\kern 0.0pt\raise 0.0pt\hbox{\hbox{\kern 3.0pt\raise 0.0pt\hbox{$\textstyle{\mathbf{Rel}^{\times}\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$}}}}}}}\ignorespaces\ignorespaces\ignorespaces{}\ignorespaces\ignorespaces{\hbox{\lx@xy@drawline@}}\ignorespaces\ignorespaces\ignorespaces{\hbox{\kern 65.37808pt\raise-8.41386pt\hbox{{}\hbox{\kern 0.0pt\raise 0.0pt\hbox{\hbox{\kern 3.0pt\hbox{\hbox{\kern 0.0pt\raise-1.75pt\hbox{$\scriptstyle{-\times\tau}$}}}\kern 3.0pt}}}}}}\ignorespaces{\hbox{\kern 81.0778pt\raise-20.50468pt\hbox{\hbox{\kern 0.0pt\raise 0.0pt\hbox{\lx@xy@tip{1}\lx@xy@tip{-1}}}}}}\ignorespaces\ignorespaces{\hbox{\lx@xy@drawline@}}\ignorespaces{\hbox{\lx@xy@drawline@}}\ignorespaces\ignorespaces\ignorespaces\ignorespaces{}\ignorespaces{\hbox{\kern 40.16394pt\raise-16.4968pt\hbox{\hbox{\kern 0.0pt\raise-2.50002pt\hbox{$\scriptstyle{\displaystyle\Downarrow}$}}}}}\ignorespaces\ignorespaces{\hbox{\kern 41.60835pt\raise-13.7472pt\hbox{{}\hbox{\kern 0.0pt\raise 0.0pt\hbox{\hbox{\kern 3.0pt\hbox{\hbox{\kern 0.0pt\raise 0.0pt\hbox{$\scriptstyle{}$}}}\kern 3.0pt}}}}}}\ignorespaces{\hbox{\kern 40.16394pt\raise-41.24158pt\hbox{\hbox{\kern 0.0pt\raise-2.50002pt\hbox{$\scriptstyle{\displaystyle\Downarrow}$}}}}}\ignorespaces\ignorespaces\ignorespaces{\hbox{\kern 44.60835pt\raise-38.49196pt\hbox{{}\hbox{\kern 0.0pt\raise 0.0pt\hbox{\hbox{\kern 3.0pt\hbox{\hbox{\kern 0.0pt\raise-2.15279pt\hbox{$\scriptstyle{\mathsf{const}}$}}}\kern 3.0pt}}}}}}\ignorespaces{}{\hbox{\kern 90.54726pt\raise 0.0pt\hbox{\hbox{\kern 0.0pt\raise 0.0pt\hbox{\hbox{\kern 3.0pt\raise 0.0pt\hbox{$\textstyle{}$}}}}}}}{\hbox{\kern-8.1389pt\raise-27.49438pt\hbox{\hbox{\kern 0.0pt\raise 0.0pt\hbox{\hbox{\kern 3.0pt\raise 0.0pt\hbox{$\textstyle{\mathcal{W}\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$}}}}}}}\ignorespaces\ignorespaces\ignorespaces{}\ignorespaces\ignorespaces{\hbox{\lx@xy@drawline@}}\ignorespaces\ignorespaces\ignorespaces{\hbox{\kern 6.74547pt\raise-7.67497pt\hbox{{}\hbox{\kern 0.0pt\raise 0.0pt\hbox{\hbox{\kern 3.0pt\hbox{\hbox{\kern 0.0pt\raise-1.71112pt\hbox{$\scriptstyle{\mathsf{Traj}}$}}}\kern 3.0pt}}}}}}\ignorespaces{\hbox{\kern 39.74214pt\raise-3.0pt\hbox{\hbox{\kern 0.0pt\raise 0.0pt\hbox{\lx@xy@tip{1}\lx@xy@tip{-1}}}}}}\ignorespaces\ignorespaces{\hbox{\lx@xy@drawline@}}\ignorespaces{\hbox{\lx@xy@drawline@}}\ignorespaces\ignorespaces\ignorespaces\ignorespaces{}{\hbox{\lx@xy@droprule}}\ignorespaces{\hbox{\kern 38.80139pt\raise-27.49438pt\hbox{\hbox{\kern 0.0pt\raise-2.39166pt\hbox{$\scriptstyle{\mathsf{State}}$}}}}}{\hbox{\lx@xy@droprule}}\ignorespaces{\hbox{\kern 81.0778pt\raise-27.49438pt\hbox{\hbox{\kern 0.0pt\raise 0.0pt\hbox{\lx@xy@tip{1}\lx@xy@tip{-1}}}}}}{\hbox{\lx@xy@droprule}}{\hbox{\lx@xy@droprule}}\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces{}{}{}{{}{}{}{}{}{}{}{}{}{}{}{}{}}{}\ignorespaces\hbox{\hbox{\kern 0.0pt\raise 0.0pt\hbox{{}{}{{}}{{}{}{}}{}}}}\ignorespaces{}\ignorespaces{}{}{}{{}{}}\ignorespaces\ignorespaces{\hbox{\kern 37.1528pt\raise-55.0943pt\hbox{{}\hbox{\kern 0.0pt\raise 0.0pt\hbox{\hbox{\kern 3.0pt\hbox{\hbox{\kern 0.0pt\raise-1.71112pt\hbox{$\scriptstyle{\mathsf{Traj}}$}}}\kern 3.0pt}}}}}}\ignorespaces{}{}{}{{}{}{}{}{}{}{}{}{}{}{}{}{}{}{}{}{}{}}{\hbox{\kern 90.16716pt\raise-30.49342pt\hbox{\hbox{\kern 0.0pt\raise 0.0pt\hbox{\lx@xy@tip{1}\lx@xy@tip{-1}}}}}}\ignorespaces\hbox{\hbox{\kern 0.0pt\raise 0.0pt\hbox{{}{}{}{{}}{{}{}{}\lx@xy@spline@}{}}}}\ignorespaces{}\ignorespaces\hbox{\hbox{\kern 0.0pt\raise 0.0pt\hbox{{}{}{{}}{{}{}{}}{}}}}\ignorespaces{}{\hbox{\kern 41.60835pt\raise-27.49438pt\hbox{\hbox{\kern 0.0pt\raise 0.0pt\hbox{\hbox{\kern 3.0pt\raise 0.0pt\hbox{$\textstyle{}$}}}}}}}{\hbox{\kern 81.0778pt\raise-27.49438pt\hbox{\hbox{\kern 0.0pt\raise 0.0pt\hbox{\hbox{\kern 3.0pt\raise 0.0pt\hbox{$\textstyle{\mathbf{Rel}^{\times}}$}}}}}}}\ignorespaces}}}}\ignorespaces}&&\vbox{\lx@xy@svg{\hbox{\raise 0.0pt\hbox{\kern 33.78299pt\hbox{\ignorespaces\ignorespaces\ignorespaces\hbox{\vtop{\kern 0.0pt\offinterlineskip\halign{\entry@#!@&&\entry<EMAIL_ADDRESS>0.0pt\hbox{\hbox{\kern 0.0pt\raise 0.0pt\hbox{\hbox{\kern 3.0pt\raise 0.0pt\hbox{$\textstyle{\mathsf{Traj}(\mathtt{LSI})\times T\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$}}}}}}}\ignorespaces\ignorespaces\ignorespaces\ignorespaces{}{\hbox{\lx@xy@droprule}}\ignorespaces\ignorespaces\ignorespaces{\hbox{\kern 0.0pt\raise-14.45831pt\hbox{{}\hbox{\kern 0.0pt\raise 0.0pt\hbox{\hbox{\kern 3.0pt\hbox{\hbox{\kern 0.0pt\raise-2.39166pt\hbox{$\scriptstyle{{\mathtt{LSI}}}$}}}\kern 3.0pt}}}}}}\ignorespaces{\hbox{\kern 0.0pt\raise-18.41663pt\hbox{\hbox{\kern 0.0pt\raise 0.0pt\hbox{\lx@xy@tip{1}\lx@xy@tip{-1}}}}}}{\hbox{\lx@xy@droprule}}{\hbox{\lx<EMAIL_ADDRESS>0.0pt\raise 0.0pt\hbox{\hbox{\kern 3.0pt\raise 0.0pt\hbox{$\textstyle{\mathsf{State}(\mathtt{LSI})\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$}}}}}}}\ignorespaces\ignorespaces\ignorespaces\ignorespaces{}{\hbox{\lx@xy@droprule}}\ignorespaces\ignorespaces\ignorespaces{\hbox{\kern 0.0pt\raise-43.37494pt\hbox{{}\hbox{\kern 0.0pt\raise 0.0pt\hbox{\hbox{\kern 3.0pt\hbox{\hbox{\kern 0.0pt\raise-1.46945pt\hbox{$\scriptstyle{\mathsf{const}_{\mathtt{LSI}}}$}}}\kern 3.0pt}}}}}}\ignorespaces{\hbox{\kern 0.0pt\raise-47.33325pt\hbox{\hbox{\kern 0.0pt\raise 0.0pt\hbox{\lx@xy@tip{1}\lx@xy@tip{-1}}}}}}{\hbox{\lx@xy@droprule}}{\hbox{\lx<EMAIL_ADDRESS>0.0pt\raise 0.0pt\hbox{\hbox{\kern 3.0pt\raise 0.0pt\hbox{$\textstyle{\mathsf{Traj}(\mathtt{LSI})}$}}}}}}}\ignorespaces}}}}\ignorespaces}\end{array}$$\par Theproblemisthatthestatespaceexplodes;functionspacesareverylarge.Nonetheless,allofthesystemintegrationlogicisidentical,andusingtheentailmentoperad$Ent$,wecanbuildinadditionalrestrictionstolimitthesearchspace.Inparticular,wecanrestrictattentiontothesubsetoffunctionsthatsatisfiesaparticulardifferentialequationorstate- transitionrelationship.Thisdrasticallylimitsthesetofvalidtrajectories,thoughtheresultingsetmaybedifficulttocharacterizeandthemethodsforexploringitwillvarybycontext.\par{\bf Relatedanalyticapplications.}Wiringdiagramshaveanestablishedappliedliteratureforsystemdesignproblems,see,e.g.,\cite[cite]{[\@@bibref{}{CyberWire, Seven, OpWire, CatSci, SpivakTan, OpenDyn}{}{}]}.Morebroadly,theanalyticstrengthofcategorytheorytoexpresscompositionalityandfunctionalsemanticsisexploredinnumerousrecentappliedworks --e.g.\ engineeringdiagrams\cite[cite]{[\@@bibref{}{ PaLin, Props, OpenPetriNets, OpenCM, CoyaThesis, Seven, DigitalCircuits}{}{}]},Markovprocesses\cite[cite]{[\@@bibref{}{CompMark, BioMark}{}{}]},databaseintegration\cite[cite]{[\@@bibref{}{BSW, Seven, CompPow, CatSci, SpivakKent, SpivakWis, MultiManu}{}{}]},behaviorallogic\cite[cite]{[\@@bibref{}{Seven, ToposSem, TempType, SheafEvent}{}{}]},naturallanguageprocessing\cite[cite]{[\@@bibref{}{FoundNLP, SentNPL, QuantNL}{}{}]},machinelearning\cite[cite]{[\@@bibref{}{Len, BackProp}{}{}]},cybersecurity\cite[cite]{[\@@bibref{}{YonedaHack, SemanticsCyber, CyberWire}{}{}]},quantumcomputation\cite[cite]{[\@@bibref{}{OptQuant, ReduceQuant, Quantomatic}{}{}]}andopengames\cite[cite]{[\@@bibref{}{GameGraph, GameMixed, CompGame}{}{}]}.\par$ ## 6 Automated synthesis with network operads An operad acting on an algebra provides a starting point to automatically generate and evaluate candidate designs. Formally correct designs (operations in some operad) combine basic systems (elements of some algebra of that operad) into a composite system. ### 6.1 Sailboat example Consider the sailboat problem introduced in 33.1 and revisited in 44.2–4.3. Network operads describe assets and ports carrying each other while algebra- based semantics guided the search for effective designs by capturing the impact of available search effort. To apply this model to automate design synthesis, algorithms explored designs within budget constraints based on costs in Table 2. Exploration iteratively composed up to budget constraints and operational limits on carrying777Though not used for this application, it turns of that degree limits–e.g. how many quadcopters a helicopter can carry–can be encoded directly into operad operations; the relevant mathematics was worked out in [73].. With these analytic models, greater sophistication was not needed; other combinatorial search algorithms–e.g. simulated annealing–are readily applied to large search spaces. The most effective designs could ferry a large number of low cost search and rescue units–e.g. quadcopters ($\mathtt{QD}$)–quickly to the scene–e.g. via helicopters ($\mathtt{Helo}$). ### 6.2 Tasking example Surprisingly, network operads—originally developed to design systems—can also be applied to “task" them: in other words, declare their behavior. An elegant example of this approach is given in [7] where “catalyst" agents enable behavioral options for a system. The SAR tasking problem. The sailboat problem is limited by search: once sailboat crew members are found, their recovery is relatively straightforward. In hostile environments, recovery of isolated personnel (IPs) can become very complex. The challenge is balancing the time criticality of recovery with the risk to the rescuers by judiciously orchestrating recovery teams888The recovery of downed airman Gene Hambleton, call sign Bat 21 Bravo, during the Vietnam War is a historical example of ill-fated SAR risk management. Hambleton’s eventual recovery cost 5 additional aircraft being shot down and 11 deaths; for comparison, a total 71 rescuers and 45 aircraft were lost to save 3,883 lives during Vietnam War SAR [23]. . Consider the potential challenges of a large scale earthquake during severe drought conditions which precipitates multiple wildfires over a large area. The 2020 Creek Fire near Fresno, CA required multiple mass rescue operations (MROs) to rescue over 100 people in each case by pulling in National Guard, Navy and Marine assets to serve as search and rescue units (SRUs) [52, 62]. Though MRO scenarios are actively considered by U.S. SAR organizations, the additional challenge of concurrent MROs distributed over a large area is not typically studied. In this SAR tasking example, multiple, geographically distributed IP groups compete for limited SRUs. The potential of coordinating multiple agent types—e.g., fire fighting airplanes together with helicopters—to jointly overcome environment risks is considered as well as aerial refueling options for SRUs to extend their range. Depending on available assets, recovery demands and risks, a mission plan may need to work around some key agent types–e.g. refueling assets–and maximize the impact of others–e.g. moving protective assets between recovery teams. Under CASCADE, tasking operations were built up from primitive tasks that coordinate multiple agent types to form a composite task plan. Novel concepts to coordinate teams of SRUs are readily modeled with full representation of the diversity of potential mission plan solutions. Network models for tasking. A network model for tasking defines atomic agent types $C$ and possible task plans for each list of agent types. Whereas a network model to design structure $\mathsf{\Gamma}\colon\mathcal{S}(C)\to\mathbf{Mon}$ has values that are possible graphical designs, a network model to task behavior $\mathsf{\Lambda}\colon\mathcal{S}(C)\to\mathbf{Cat}$ has values that are categories whose morphisms index possible task plans for the assembled types; compare, e.g., [7, Thm. 9]. Each morphism declares a sequence of tasks for each agent–many of which will be coordinated with other agents. If the system is comprised of only a single UH-60 helicopter, its possible tasks are captured in $\mathsf{\Lambda}(\mathtt{UH60})$. In this application, these tasks are paths in a graph describing ‘safe maneuvers.’ For unsafe maneuvers, UH-60s travel in pairs–or perhaps with escorts such as a HC-130 or CH-47 equipped with a Modular Airborne Fire Fighting System (MAFFS). Anything one UH-60 can do, so can two, but not vice versa. Thus there is a proper inclusion $\mathsf{\Lambda}(\mathtt{UH60})\times\mathsf{\Lambda}(\mathtt{UH60})\subsetneq\mathsf{\Lambda}(\mathtt{UH60}\otimes\mathtt{UH60})$. Similarly, $\mathsf{\Lambda}(\mathtt{UH60})\times\mathsf{\Lambda}(\mathtt{HC130})\subsetneq\mathsf{\Lambda}(\mathtt{UH60}\otimes\mathtt{HC130})$ since once both a UH-60 and HC-130 are present, a joint behavior of midair refueling of the UH-60 by the HC-130 becomes possible. Formally, these inclusions are lax structure maps–e.g. $\Phi_{(\mathtt{UH60},\mathtt{UH60})}\colon\mathsf{\Lambda}(\mathtt{UH60})\times\mathsf{\Lambda}(\mathtt{UH60})\to\mathsf{\Lambda}(\mathtt{UH60}\otimes\mathtt{UH60})$, specifies: given tasks for a single UH-60 (left coordinate) and tasks for another UH-60 (right coordinate), define the corresponding joint tasking of the pair. Here the joint tasking is: each UH-60 operates independently within the safe graph. On the other hand, tasks in $\mathsf{\Lambda}(\mathtt{UH60}\otimes\mathtt{UH60})$ to maneuver in unsafe regions can not be constructed from independent taskings of each UH-60. Such tasks must be set for some pair or other allowed team–e.g. a CH-47 teamed with an UH-60. $\mathtt{UH60}$$\mathtt{UH60}$$\mathtt{UH60}$ $\mathtt{UH60}$$\mathtt{UH60}+\mathtt{HC130}$$\mathtt{UH60}+\mathtt{HC130}$$2\mathtt{UH60}$$2\mathtt{UH60}$$\mathtt{UH60}$$\mathtt{UH60}$$\mathtt{UH60}$ $\mathtt{UH60}$$\mathtt{UH60}+\mathtt{HC130}$$\mathtt{UH60}+\mathtt{HC130}$$2\mathtt{UH60}$$2\mathtt{UH60}$$\bullet$ $\mathtt{Cut}$$\bullet$ $\mathtt{Helo}$$\bullet$ $\mathtt{QD}$$\;\;a\;\;$$\;\;b\;\;$$\;\;c\;\;$$\;\;d\;\;$$\;\phantom{\Big{|}}\tau_{1}\;$$\;\phantom{\Big{|}}\tau_{2}\;$$\;\phantom{\Big{|}}\tau_{3}\;$$\;\phantom{\Big{|}}\tau_{4}\;$$\bullet$ $\mathtt{UH60}$$\bullet$ $\mathtt{HC130}$$\phantom{{|}}\tau_{1}$$\phantom{{|}}\tau_{2}$$\phantom{{|}}\tau_{4}$$(a,0)$$(b,1)$$(c,2)$$(c,2)$$((d,d),(4,4))$$2$$2$$2$$2$$2$$4$$\sqcup$$\mapsto$$2$$2$$4$$4$$+$$\mapsto$$2$$2$$4$$\bullet$ $\mathtt{Cut}$$\bullet$ $\mathtt{Helo}$$\bullet$ $\mathtt{QD}$$\;\;a\;\;$$\;\;b\;\;$$\;\;c\;\;$$\;\;d\;\;$$\;\phantom{\Big{|}}\tau_{1}\;$$\;\phantom{\Big{|}}\tau_{2}\;$$\;\phantom{\Big{|}}\tau_{3}\;$$\;\phantom{\Big{|}}\tau_{4}\;$$\bullet$ $\mathtt{UH60}$$\bullet$ $\mathtt{HC130}$ (a) Four primitive tasks specified in a Petri net; arcs indicate types involved in each task. $\mathtt{UH60}$$\mathtt{UH60}$$\mathtt{UH60}$ $\mathtt{UH60}$$\mathtt{UH60}+\mathtt{HC130}$$\mathtt{UH60}+\mathtt{HC130}$$2\mathtt{UH60}$$2\mathtt{UH60}$$\mathtt{UH60}$$\mathtt{UH60}$$\mathtt{UH60}$ $\mathtt{UH60}$$\mathtt{UH60}+\mathtt{HC130}$$\mathtt{UH60}+\mathtt{HC130}$$2\mathtt{UH60}$$2\mathtt{UH60}$$\bullet$ $\mathtt{Cut}$$\bullet$ $\mathtt{Helo}$$\bullet$ $\mathtt{QD}$$\;\;a\;\;$$\;\;b\;\;$$\;\;c\;\;$$\;\;d\;\;$$\;\phantom{\Big{|}}\tau_{1}\;$$\;\phantom{\Big{|}}\tau_{2}\;$$\;\phantom{\Big{|}}\tau_{3}\;$$\;\phantom{\Big{|}}\tau_{4}\;$$\bullet$ $\mathtt{UH60}$$\bullet$ $\mathtt{HC130}$$\phantom{{|}}\tau_{1}$$\phantom{{|}}\tau_{2}$$\phantom{{|}}\tau_{4}$$(a,0)$$(b,1)$$(c,2)$$(c,2)$$((d,d),(4,4))$$2$$2$$2$$2$$2$$4$$\sqcup$$\mapsto$$2$$2$$4$$4$$+$$\mapsto$$2$$2$$4$$\bullet$ $\mathtt{Cut}$$\bullet$ $\mathtt{Helo}$$\bullet$ $\mathtt{QD}$$\;\;a\;\;$$\;\;b\;\;$$\;\;c\;\;$$\;\;d\;\;$$\;\phantom{\Big{|}}\tau_{1}\;$$\;\phantom{\Big{|}}\tau_{2}\;$$\;\phantom{\Big{|}}\tau_{3}\;$$\;\phantom{\Big{|}}\tau_{4}\;$$\bullet$ $\mathtt{UH60}$$\bullet$ $\mathtt{HC130}$ ${\mathsf{\Lambda}(\mathtt{UH60}\otimes\mathtt{UH60})}$${\mathsf{\Lambda}(\mathtt{UH60})}$${\mathsf{\Lambda}(\mathtt{UH60}\otimes\mathtt{HC130})}$${{\begin{array}[]{c}M(\mathtt{UH60}\otimes\mathtt{UH60})\\\ \scalebox{0.5}{\mbox{$\displaystyle\begin{bmatrix}-1&0&1&0&0&0&0&0\\\ 0&-1&1&0&0&0&0&0\\\ 0&0&0&0&-1&0&1&0\\\ 0&0&0&0&0&-1&1&0\\\ 0&0&-1&1&0&0&-1&1\end{bmatrix}$}}\\\ {}\hfil\\\ M^{s}(\mathtt{UH60}\otimes\mathtt{UH60})\\\ \scalebox{0.5}{\mbox{$\displaystyle\begin{bmatrix}1&0&0&0&0&0&0&0\\\ 0&1&0&0&0&0&0&0\\\ 0&0&0&0&1&0&0&0\\\ 0&0&0&0&0&1&0&0\\\ 0&0&1&0&0&0&1&0\end{bmatrix}$}}\end{array}}}$${{\begin{array}[]{c}M(\mathtt{UH60})\\\ \scalebox{1.0}{\mbox{$\displaystyle\begin{bmatrix}-1&0&1&0\\\ 0&-1&1&0\end{bmatrix}$}}\\\ {}\hfil\\\ M^{s}(\mathtt{UH60})\\\ \scalebox{1.0}{\mbox{$\displaystyle\begin{bmatrix}1&0&0&0\\\ 0&1&0&0\end{bmatrix}$}}\end{array}}}$${{\begin{array}[]{c}M(\mathtt{UH60}\otimes\mathtt{HC130})\\\ \scalebox{0.75}{\mbox{$\displaystyle\begin{bmatrix}-1&0&1&0&0&0&0&0\\\ 0&-1&1&0&0&0&0&0\\\ 0&0&0&0&0&0&0&0\\\ \end{bmatrix}$}}\\\ {}\hfil\\\ M^{s}(\mathtt{UH60}\otimes\mathtt{HC130})\\\ \scalebox{0.75}{\mbox{$\displaystyle\begin{bmatrix}1&0&0&0&0&0&0&0\\\ 0&1&0&0&0&0&0&0\\\ 0&0&1&0&0&0&1&0\\\ \end{bmatrix}$}}\end{array}}}$Network Model Constraint Matrices (b) More primitive tasks become possible as available agent types increase. Type update matrices $M(-)$ and target to source constraint matrices $M^{s}(-)$ translate type changing and matching, resp. Figure 9: Specified primitive tasks determine an operad ${\mathcal{O}}_{SAR}$ and a constraint program to explore operations. Applying the cookbook: operads. While the above discussion sketches how to specify a network model for tasking, which constructs a network operad [6], these precise details [37] need not concern the applied practitioner999That is, a Petri net specifies the network model $\mathsf{\Lambda}\colon\mathcal{S}(C)\to\mathbf{Cat}$ to task behavior. The construction of $\mathsf{\Lambda}$ [37] is similar to the construction described in [7, Thm. 9], but adapted to colored Petri nets whose transitions preserve the number of tokens of each color; see, e.g., Fig. 9(a). Compared to [7, Thm. 9], $C$ corresponds to token colors, rather than catalysts [7, Def. 6], and species index discrete coordination locations. Target categories encode allowed paths for each atomic agent type, (cont.) 9 (cont.), e.g., for Fig. 9(a) $\mathsf{\Lambda}(\mathtt{UH60})$ is (freely) generated by objects $\\{a,b,c,d\\}$ and morphisms $\tau_{1}\colon a\to c$ and $\tau_{2}\colon a\to c$, whereas $\mathsf{\Lambda}(\mathtt{HC130})$ is just by generated $\\{a,b,c,d\\}$ since no transition involves a single $\mathtt{HC130}$. By describing each target category as an appropriate subcategory of a product of path categories, the symmetric group action is given permuting coordinates, which allows the role of each atomic agent in a task to be specified.. It is sufficient to provide a Petri net as template, from which a network operad is constructed. Whereas a template to design structures defines the basic ways system types can interact, a template to task behavior defines the primitive tasks for agent types $C$, which are token colors in the Petri net. No specification of ‘staying put’ tasks are needed; these are implicit. All other primitive tasks are (sparsely) declared. For example, each edge of the ‘safe graph’ for a solo UH-60 declares: (1) a single agent of type $\mathtt{UH60}$ participates in this ‘traverse edge’ task; and (2) participation is possible if a $\mathtt{UH60}$ is available at the source of the edge. Likewise, each edge of the ‘unsafe graph’ for pairs of UH-60s should declare similar information for pairs, but what about operations to refuel an UH-60 with a HC-130? It turns out that transitions in a Petri net carry sufficient data [7, 37] and have a successful history of specifying generators for a monoidal category [9, 72, 8]. The Petri net Fig. 9(a) shows examples where, for simplicity, tasks to traverse edges are only shown in the left to right direction. This sparse declaration is readily extended–e.g. to add recovery focused CH-47s, which tested their operational limits to rescue as many as 46 people during the 2020 Creek Fire–$C$ and the set of transitions are augmented to encode the new options for primitive tasks. This specification of syntax is almost sufficient for the SAR tasking problem and would be for situations where only the sequence of tasks for each agent needs to be planned. When tasking SAR agents, _when_ tasks are performed is semantically important because where and how long air-based agents ‘stay put’ impacts success: (1) fuel burned varies dramatically for ground vs. air locations; (2) risk incurred varies dramatically for safe vs. unsafe locations. For comparison, in a ground-based domain without environmental costs, these considerations might be approximately invariant relative to the time tasks occur, and therefore, can be omitted from tasking syntax. Timing information creates little added burden for building a template–transitions declaring primitive tasks need only be given durations derivable from scenario data–and it is technically straightforward to add a time dimension to the network model. Constraints from syntax. A direct translation of primitive tasks to decision variables for a constraint program is possible. For syntax, the idea is very simple: enforce type matching constraints on composing operad morphisms. Here we will briefly indicate the original mixed integer linear program developed for SAR tasking; later this formulation was reworked to leverage the scheduling toolkit of the CPLEX optimization software package. To illustrate the concept, let us first consider the constraint program for an operad to plan tasks without time and then add the time dimension101010Simply increasingly dimensionality is not computationally wise–which was the point of exploring the CPLEX scheduling toolkit to address the time dimension–but this model still serves as a conceptual reference point.. Operad types are translated to boolean vectors $m_{j}$–whose entries capture individual agents at discrete coordination locations. Parallel composition of primitive operations is expressed with boolean vectors $\Sigma_{j}$ indexed over primitive tasks for specific agents. Type vectors $m_{j}$ indicate the coordination location each agent with value one; operation vectors $\Sigma_{j}$ indicate which tasks are planned in parallel. Assuming an operation with task vector $\Sigma_{j}$ and source vector $m_{j}$, the target is $m_{j+1}=m_{j}+M\Sigma_{j}$ where $M$ describes the relationship between source and target for primitive tasks. Rows of $M$ correspond to primitive tasks while columns correspond to individual agents. The target to source constraint for single step of in-series composition is $m_{j+1}\geq M^{s}\Sigma_{j+1}$ where $M^{s}$ has rows that give requirements for each primitive task. Here the LHS describes the target and the RHS describes the source. The inequality appears to allow for implicit identities for agents without tasking–e.g. if $\Sigma_{j}$ is a zero vector, then $m_{j+1}=m_{j}$. This constraint prevents an individual agent from being assigned conflicting tasks or ‘teleporting’ to begin a task. As seen in Fig. 9(b), additional agents: (1) enable more primitive tasks, indexed by Petri net transitions (top two rows); and (2) expand the type vector/matrix column dimension to account for new agent-location pairs and increase the matrix row dimension to account for new tasks (bottom two rows). For example, the first 4 rows of $M(\mathtt{UH60}\otimes\mathtt{UH60})$ correspond to the image of $\mathsf{\Lambda}(\mathtt{UH60})\times\mathsf{\Lambda}(\mathtt{UH60})$ in $\mathsf{\Lambda}(\mathtt{UH60}\otimes\mathtt{UH60})$. The last row corresponds to new task, $\tau_{4}$, for the available pair of UH-60s. During implementation, the constraints can be declared task-by-task/row-by-row to sparsely couple the involved agents. Once a limit on the number of steps of in series composition is set–i.e. a bound for the index $j$ is given–a finite constraint program is determined. Time is readily modeled discretely with tasks given integer durations. This corresponds to a more detailed network model, $\mathsf{\Lambda}_{t}$, whose types include a discrete time index; see Fig. 5(b) for example operations. Under these assumptions, one simply replaces the abstract steps of in series composition with a time index and decomposes $M$ and $\Sigma_{j}$ by the duration $d$ of primitive tasks: $m_{t}+\sum_{d=1}^{d_{\max}}M_{d}\Sigma_{t-d,d}=m_{t+1};~{}~{}~{}~{}m_{t+1}\geq\sum_{d=1}^{d_{\max}}M^{s}_{d}\Sigma_{t+1,d}$ so that $\Sigma_{t,d}$ describes tasks beginning at time $t$; the inequality allows for ‘waiting’ operations. One can also model tasks more coarsely–with $\mathsf{\Lambda}_{\bullet}\colon\mathbb{N}(C)\to\mathbf{Cat}$–to construct an operad to task counts of agents without individual identity. Then, the type vectors $m_{j}$ (resp., operation vectors $\Sigma_{j}$) have integer entries to express agent counts (resp., counts of planned tasks) with corresponding reductions in dimensionality. These three levels of network models $\mathtt{UH60}$$\mathtt{UH60}$$\mathtt{UH60}$ $\mathtt{UH60}$$\mathtt{UH60}+\mathtt{HC130}$$\mathtt{UH60}+\mathtt{HC130}$$2\mathtt{UH60}$$2\mathtt{UH60}$$\mathtt{UH60}$$\mathtt{UH60}$$\mathtt{UH60}$ $\mathtt{UH60}$$\mathtt{UH60}+\mathtt{HC130}$$\mathtt{UH60}+\mathtt{HC130}$$2\mathtt{UH60}$$2\mathtt{UH60}$$\bullet$ $\mathtt{Cut}$$\bullet$ $\mathtt{Helo}$$\bullet$ $\mathtt{QD}$$\;\;a\;\;$$\;\;b\;\;$$\;\;c\;\;$$\;\;d\;\;$$\;\phantom{\Big{|}}\tau_{1}\;$$\;\phantom{\Big{|}}\tau_{2}\;$$\;\phantom{\Big{|}}\tau_{3}\;$$\;\phantom{\Big{|}}\tau_{4}\;$$\bullet$ $\mathtt{UH60}$$\bullet$ $\mathtt{HC130}$$\phantom{{|}}\tau_{1}$$\phantom{{|}}\tau_{2}$$\phantom{{|}}\tau_{4}$$(a,0)$$(b,1)$$(c,2)$$(c,2)$$((d,d),(4,4))$$2$$2$$2$$2$$2$$4$$\sqcup$$\mapsto$$2$$2$$4$$4$$+$$\mapsto$$2$$2$$4$$\bullet$ $\mathtt{Cut}$$\bullet$ $\mathtt{Helo}$$\bullet$ $\mathtt{QD}$$\;\;a\;\;$$\;\;b\;\;$$\;\;c\;\;$$\;\;d\;\;$$\;\phantom{\Big{|}}\tau_{1}\;$$\;\phantom{\Big{|}}\tau_{2}\;$$\;\phantom{\Big{|}}\tau_{3}\;$$\;\phantom{\Big{|}}\tau_{4}\;$$\bullet$ $\mathtt{UH60}$$\bullet$ $\mathtt{HC130}$${\mathcal{S}(C)}$${\mathbf{Cat}}$${\mathbb{N}(C)}$$\scriptstyle{\mathsf{\Lambda}_{t}}$$\scriptstyle{\mathsf{\Lambda}}$$\scriptstyle{\mathsf{\Lambda}_{\bullet}}$ naturally induce morphisms of network operads111111Strictly speaking, the coarsest (lowest) level is not network model; its domain is a free commutative monoidal category. Nevertheless, a completely analogous construction produces a typed operad fitting into this diagram. [6, 6.18] and encode mappings of syntactic variables that preserve feasibility. In particular, the top two levels describe a precise mapping from task scheduling (highest) to task planning (middle). The lowest level $\mathsf{\Lambda}_{\bullet}$ forgets the individual identity of agents, providing a coarser level for planning. This very simple idea of enforcing type matching constraints is inherently natural121212I.e., operad morphisms push forward feasible assignments variables in the domain to feasible assignments in the codomain.. However, further research is needed to determine if this natural hierarchical structure can be exploited by algorithms–e.g. by branching over pre-images of solutions to coarser levels–perhaps for domains were operational constraints coming from algebras are merely a nuisance, as opposed to being a central challenge for SAR planning. For instance, a precise meta-model for planning and scheduling provides a common jumping off point to apply algorithms from those two disciplines. Applying the cookbook: algebras. Because the operad template defines generating operations, specifying algebras involves: (1) capturing the salient features of each agent type as its internal state; and (2) specifying how these states update under generating morphisms–including, for operads with time, the implicit ‘waiting’ operations. For the SAR tasking problem, the salient features are fuel level and cumulative probability of survival throughout the mission. Typical primitive operations will not increase these values; fuel is expended or some risk is incurred. The notable exception is refueling operations which return the fuel level of the receiver to maximum. By specifying the non-increasing rate for each agent–location pair, the action of ‘waiting’ operations are specified. In practice, these data are derivable from environmental data for a scenario so that end users can manipulate them indirectly. Operational constraints from algebras. Salient features of each agent type are captured as auxiliary variables determined by syntactic decision variables. The values of algebra variables are constrained of update equations–e.g. to update fuel levels for agents with $\max(f_{j}+F\Sigma_{j},f_{\max})=f_{j+1}$, where $f_{\max}$ specifies max fuel capacities. Having expressed the semantics for generating operations, one can enforce additional operational constraints–e.g. safe fuel levels: $f_{j+1}\geq f_{\min}.$ Extending the domain of application. As noted above, this sparse declaration of a tasking domain is readily extended–e.g. to add a new atomic type or new ways for agents to coordinate. Syntactically, this is amounts to new elements of $C$ or transitions to define primitive tasks. Semantics must capture the impact of primitive operations on state, which can be roughly estimated initially and later refined. This flexibility is especially useful for rapid prototyping of ‘what if’ options for asset types and behaviors, as the wildfire SAR tasking problem illustrates. Suppose, for example, that we wanted to model a joint SAR and fire fighting problem. Both domains are naturally expressed with network operads to task behavior. Even if the specification formats were independently developed: (1) each format must encode the essential combinatorial data for each domain; and (2) category theory provides a method to integrate domain data: construct a pushout. Analogous to taking the union of two sets along a common intersection, one identifies the part of the problem common to both domains–e.g. MAFFS-equipped HC-130s and their associated tasks appearing in both domains–to construct a cross-domain model $\mathtt{UH60}$$\mathtt{UH60}$$\mathtt{UH60}$ $\mathtt{UH60}$$\mathtt{UH60}+\mathtt{HC130}$$\mathtt{UH60}+\mathtt{HC130}$$2\mathtt{UH60}$$2\mathtt{UH60}$$\mathtt{UH60}$$\mathtt{UH60}$$\mathtt{UH60}$ $\mathtt{UH60}$$\mathtt{UH60}+\mathtt{HC130}$$\mathtt{UH60}+\mathtt{HC130}$$2\mathtt{UH60}$$2\mathtt{UH60}$$\bullet$ $\mathtt{Cut}$$\bullet$ $\mathtt{Helo}$$\bullet$ $\mathtt{QD}$$\;\;a\;\;$$\;\;b\;\;$$\;\;c\;\;$$\;\;d\;\;$$\;\phantom{\Big{|}}\tau_{1}\;$$\;\phantom{\Big{|}}\tau_{2}\;$$\;\phantom{\Big{|}}\tau_{3}\;$$\;\phantom{\Big{|}}\tau_{4}\;$$\bullet$ $\mathtt{UH60}$$\bullet$ $\mathtt{HC130}$$\phantom{{|}}\tau_{1}$$\phantom{{|}}\tau_{2}$$\phantom{{|}}\tau_{4}$$(a,0)$$(b,1)$$(c,2)$$(c,2)$$((d,d),(4,4))$$2$$2$$2$$2$$2$$4$$\sqcup$$\mapsto$$2$$2$$4$$4$$+$$\mapsto$$2$$2$$4$$\bullet$ $\mathtt{Cut}$$\bullet$ $\mathtt{Helo}$$\bullet$ $\mathtt{QD}$$\;\;a\;\;$$\;\;b\;\;$$\;\;c\;\;$$\;\;d\;\;$$\;\phantom{\Big{|}}\tau_{1}\;$$\;\phantom{\Big{|}}\tau_{2}\;$$\;\phantom{\Big{|}}\tau_{3}\;$$\;\phantom{\Big{|}}\tau_{4}\;$$\bullet$ $\mathtt{UH60}$$\bullet$ $\mathtt{HC130}$${\mathtt{Spec_{\cap}}}$${\mathtt{Spec_{SAR}}}$${\mathtt{Spec_{FF}}}$${\mathtt{Spec_{\cup}}.}$${\ulcorner}$ The arrows in this diagram account for translating the file format for the overlap into each domain-specific format and choosing a specific output format for cross-domain data. On the other hand, suppose that the machine readable representation of each domain was tightly coupled to algorithms–e.g. mathematical programming for SAR and planning framework for fire fighting. There is no artifact suitable for integrating these domains since expression was prematurely optimized. We describe a general workflow to separate specification from representation and exploitable data structures and algorithms in 77.5. ### 6.3 Other examples of automated synthesis Though network templates facilitate exploration from atoms, how to explore valid designs is a largely distinct concern from defining the space of designs, as discussed in 1. Novel search strategies via substitution For example, in the DARPA Fundamentals of Design (FUN Design) program, composition of designs employed a genetic algorithm (GA). FUN Design focused on generating novel conceptual designs for mechanical systems–e.g. catapults to launch a projectile. Formulating this problem with network operads followed the cookbook approach: there were atomic types of mechanical components and basic operations to link them. The operad-based representation provided guarantees of design feasibility and informed how to generalize the GA implementation details. Specifically, composition for atomic algebra elements defined genetic data; crossover produced child data to compose from atoms; and mutation modified parameters of input algebra elements. Crafting a crossover step is typically handled case- by-case while this strategy generalizes to other problems that mix combinatorial and continuously varying data, provided this data is packaged as an operad acting on an algebra. Guarantees of feasibility dramatically reduced the number unfit offspring evaluated by simulation against multiple fitness metrics. Moreover, computational gains from feasibility guarantees increase as the design population becomes more combinatorially complex. Integrated structure and behavior Large classes of engineering problems compose components to form an ‘optimized’ network–e.g. in chemical process synthesis, supply chains, and water purification networks [61, 75, 80, 105]. Given a set of inputs, outputs and available operations (process equipment with input and output specification), the goal is to identify the optimal State Equipment Networks (SEN) for behavioral flows of materials and energy. A given production target for outputs is evaluated in terms of multiple objectives such as environmental impact and cost. For example, the chemical industry considers the supply chain, production and distribution network problem [105] systematically as three superstructure optimization problems that can be composed to optimize enterprise level, multi-subsystem structures. Each sub-network structure is further optimized for low cost and other metrics including waste, environmental impact and energy costs. The operadic paradigm would provide a lens to generalize and refine existing techniques to jointly explore structure and behavior. CASCADE prototyped integrated composition of structure and behavior for distributed logistics applications. Here an explicit resupply plan to task agents was desired. Structural composition was needed to account for the resupply capacity for heterogeneous delivery vehicles and the positioning of distributed resupply depots. Probabilistic models estimated steady state resupply capacities of delivery fleet mixes to serve estimates of demand. First, positioning resupply locations applied hill climbing to minimize the expected disruption of delivery routes when returning to and departing from resupply locations. Second, this disruption estimate was used to adjust the resupply capacity estimate of each delivery asset type. Third, promising designs where evaluated using a heuristic task planning algorithm. At each stage, algorithms focused on finding satisficing solutions which allowed broad and rapid explorations of the design and tasking search space. Synthesis with applied operads and categories. Research activity to apply operads and monoidal categories to automated design synthesis is increasing. Wiring diagrams have been applied to automate protein design [46, 92] and collaborative design [40, Ch. 4] of physical systems employing practical semantic models and algorithms [24, 25, 106, 107]. Software tools are increasingly focused on scaling up computation, e.g. [30, 55, 58], as opposed to software to augment human calculation, as in [14, 60, 81], and managing complex domains with commercial-grade tools [22, 95, 76, 101]. Recent work to optimize quantum circuits [35, 59] leverages such developments. The use of wiring diagrams to improve computational efficiency via normal forms is explored in [77]. In the next section, we discuss research directions to develop the meta- modeling potential of applied operads to: (1) decompose a problem within a semantic model to divide and conquer; and (2) move between models to fill in details from coarse descriptions. We also discuss how the flow of representations used for SAR–network template, operad model of composition, exploitation data structures and algorithms–could be systematized into a reusable software framework. ## 7 Toward practical automated analysis and synthesis In this section, we describe lessons learned from practical experiences with applying operads to automated synthesis. We frame separation of concerns in the language of operads to describe strategies to work around issues raised by this experience. This gives not only a clean formulation of separation but also a principled means to integrate and exploit concerns. ### 7.1 Lessons from automated synthesis The direct, network template approach facilitates correct and transparent modeling for complex tasking problems. However, computational tractability is limited to small problems–relative to the demands of applications. More research is needed to develop efficient algorithms that break up the search into manageable parts, leveraging the power of operads to separate concerns. Under CASCADE, we experimented with the CPLEX scheduling toolkit to informally model across levels of abstraction and exploit domain specific information. In particular, generating options to plan, but not schedule, key maneuvers with traditional routing algorithms helped factor the problem effectively. These applied experiments were not systematized into a formal meta-modeling approach, although our prototype results were promising. Specification of these levels–as in Sec. 4–and controlling the navigation of levels using domain-specifics would be ideal. The FUN DESIGN genetic algorithm approach illustrates the potential operads have to: (1) generalize case-by-case methods131313In fact, applying genetic algorithms to explore network structures was inspired by the success of NeuroEvolution of Augmenting Topologies (NEAT) [96] to generate novel neural network architectures.; (2) separate concerns, in this case by leveraging the operad syntax for combinatorial crossover and algebra parameters for continuous mutation; and (3) guarantee correctness as complexity grows. Distributed logistics applications in CASCADE show the flexibility afforded by multiple stage exploration for more efficient search. ### 7.2 Formal separation of concerns We begin by distinguishing focus from filter, which are two ways operads separate. Focus selects _what_ we look at, while filter captures _how_ we look at it. These are questions of syntax and semantics, respectively. To be useful, the _what_ of our focus must align with the _how_ of the filter. Separation of focus occurs within the syntax operad of system maps. In 22.1, four trees correspond to different views on the same system. We can zoom into one part of the system while leaving other portions black-boxed at a high level. Varying the target type of an operation changes the scope for system composition, such as restricting attention to a subsystem. Filtering, on the other hand, is semantic; we choose which salient features to model and which to suppress, controlled by the semantic context used to ‘implement’ the operations. As described in 55.3, the default semantic context is $\mathbf{Set}$ where: (1) each type in the operad is mapped to a set of possible instances for that type; and (2) each operation is mapped to a function to compose instances. Instances or algebra elements for the sailboat problem (Sec. 4) describe the key features of structural system designs. For SAR tasking (Sec. 6), mission plan instances track the key internal states of agents–notably fuel and risk–throughout its execution. Section 5 illustrates alternative semantic contexts as such probability $\mathbf{Prob}$ or relations between sets $\mathbf{Rel}$. Focus and filter come together to solve particular problems. The analysis of the LSI system in Sec. 5 tightly focuses the syntax operad $\mathcal{W}$ to include only the types and operations from Fig. 4. Formally, this is accomplished by considering the image of the generating types and operations in the operad of port-graphs [20, 3]. This tight focus means semantics need only be defined for LSI components. In each SAR tasking problem of Sec. 6, an initial, source configuration of agent types is given, narrowing the focus of each problem. The SAR focus is much broader because an operation to define the mission plan must be constructed. Even so, semantics filter down to just the key features of the problem and how to update them when generating operations act. Functorial semantics, as realized by an operad algebra $\mathsf{A}\colon{\mathcal{O}}\to\mathbf{Sem}$, helps factor the overall problem model to facilitate its construction and exploitation. For example, we can construct the probabilistic failure model in Table 3 by normalizing historical failures. First we limit focus from all port-graphs $\mathcal{P}$ to $\mathcal{W}$ then semantics for counts in $\mathbb{N}^{+}$, an operad of counts and sums, are normalized to obtain probabilities in $\mathbf{Prob}$: $\mathtt{UH60}$$\mathtt{UH60}$$\mathtt{UH60}$ $\mathtt{UH60}$$\mathtt{UH60}+\mathtt{HC130}$$\mathtt{UH60}+\mathtt{HC130}$$2\mathtt{UH60}$$2\mathtt{UH60}$$\mathtt{UH60}$$\mathtt{UH60}$$\mathtt{UH60}$ $\mathtt{UH60}$$\mathtt{UH60}+\mathtt{HC130}$$\mathtt{UH60}+\mathtt{HC130}$$2\mathtt{UH60}$$2\mathtt{UH60}$$\bullet$ $\mathtt{Cut}$$\bullet$ $\mathtt{Helo}$$\bullet$ $\mathtt{QD}$$\;\;a\;\;$$\;\;b\;\;$$\;\;c\;\;$$\;\;d\;\;$$\;\phantom{\Big{|}}\tau_{1}\;$$\;\phantom{\Big{|}}\tau_{2}\;$$\;\phantom{\Big{|}}\tau_{3}\;$$\;\phantom{\Big{|}}\tau_{4}\;$$\bullet$ $\mathtt{UH60}$$\bullet$ $\mathtt{HC130}$$\phantom{{|}}\tau_{1}$$\phantom{{|}}\tau_{2}$$\phantom{{|}}\tau_{4}$$(a,0)$$(b,1)$$(c,2)$$(c,2)$$((d,d),(4,4))$$2$$2$$2$$2$$2$$4$$\sqcup$$\mapsto$$2$$2$$4$$4$$+$$\mapsto$$2$$2$$4$$\bullet$ $\mathtt{Cut}$$\bullet$ $\mathtt{Helo}$$\bullet$ $\mathtt{QD}$$\;\;a\;\;$$\;\;b\;\;$$\;\;c\;\;$$\;\;d\;\;$$\;\phantom{\Big{|}}\tau_{1}\;$$\;\phantom{\Big{|}}\tau_{2}\;$$\;\phantom{\Big{|}}\tau_{3}\;$$\;\phantom{\Big{|}}\tau_{4}\;$$\bullet$ $\mathtt{UH60}$$\bullet$ $\mathtt{HC130}$${\mathcal{W}}$${\mathbb{N}^{+}}$${\mathcal{P}}$${\mathbf{Prob}.}$$\scriptstyle{\mathsf{A}}$ The power to focus and filter is amplified because we are not limited by a single choice of how to filter. In addition to limiting focus with the source of an operad algebra, we can simplify filters. Such natural transformations between functors are ‘filters of filters’ that align different compositional models precisely–e.g. requirements over state (55.3) or timed scheduling over two levels of planning (66.2). In this first case the syntax operad $\mathcal{W}$ stays the same and semantics are linked by an algebra homomorphism (22.4). In the second case, both the operad and algebra must change to determine simpler semantics–e.g. to neglect the impact of waiting operations, which bound performance. Such precision supports automation to explore design space across semantic models and aligns the ability to focus within each model. By working backward relative to the construction process, we can lift partial solutions to gradually increase model fidelity–e.g. exploring schedules over effective plans. This gives a foundation for lazy evaluation during deep exploration of design space, which we revisit in 77.5. For a simple but rich example of these concepts working together, consider the functional decomposition $\mathtt{f}(\mathtt{l},\mathtt{t})$ in Fig. 4. We could model the length system $\mathtt{l}$ using rigid-body dynamics, the temperature system $\mathtt{t}$ as a lumped-element model, and super-system $\mathtt{f}$ as a computation (Edlén equation) that corrects the observed $\mathtt{fringe}$ count based on the measured temperature: $\textstyle{\\{\mathtt{l}\\}\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\textstyle{\mathcal{N}_{\mathrm{mech}}\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{\mathsf{Rigid}}$$\textstyle{\mathsf{Dyn}\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{\mathsf{Impl}}$$\textstyle{\mathcal{W}\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{\displaystyle\Uparrow}$$\scriptstyle{\displaystyle\Downarrow}$$\textstyle{\mathcal{N}_{\mathrm{comp}}\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$Edlén$\textstyle{\mathsf{Type}}$$\textstyle{\\{\mathtt{t}\\}\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\textstyle{\mathcal{N}_{\mathrm{therm}}\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{\mathsf{Lump}}$$\textstyle{\mathsf{Dyn}\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{\mathsf{Impl}}$ (3) The upper and lower paths construct implementations of dynamical models based on the aforementioned formalisms. The center path implements a correction on the data stream coming from the interferometer, based on a stream of temperature data. The two natural transformations indicate extraction of one representation, a stream of state values, from the implementation of the dynamical models. Composition in $\mathcal{W}$ then constructs the two data streams and applies the correction. A key strength of the operadic paradigm is its genericity: the same principles of model construction, integration and exploitation developed for measurement and SAR apply to all kinds of systems. In principle, we could use the same tools and methodology to assemble transistors into circuits, unit processes into chemical factories and genes into genomes. The syntax and semantics change with the application, but the functorial perspective remains the same. For the rest of this section, we describe research directions to realize such a general purpose vision to decompose a complex design problem into subproblems and support rapid, broad exploration of design space. ### 7.3 Recent advancements, future prospects and limits Progress driven by applications. Section 4 describes how cookbook-style approaches enable practitioners to put operads to work. Generative data define a domain and compositionality combines it into operads and algebras to separate concerns. Network operads [6, 7, 73] were developed in response to the demands of applications to construct operads from generative data. Section 5 describes rich design analysis by leveraging multiple decompositions of complex systems and working across levels of abstraction. Focusing on a specific applied problem–the LSI at NIST–provided further opportunities for analysis since model semantics need only be defined for the problem at hand; see also Eq. 3. Progress in streamlining automated synthesis from building blocks is recounted in Sec. 6 where the domain drives coordination requirements to task behavior. Prospects. If interactions between systems are well-understood (specification) and can be usefully modeled by compositional semantics (analysis), then automated design synthesis leveraging separation for scalability becomes possible. For instance, most references from the end of Sec. 5 correspond to domains that are studied with diagrams that indicate interactions and have associated compositional models. This allows intricate interactions to be modeled–compare, e.g. classical [82] vs. quantum [1, 35, 59] computing–while unlocking separation of concerns. Cookbook and focused approaches guide practitioners to seek out the minimal data needed for a domain problem–as in the examples presented–but operads for design requires compositional models. Limitations. We note three issues limiting when operad apply: (1) key interactions among systems and components are inputs; (2) not all design problems become tractable via decomposition and hierarchy; and (3) there is no guarantee of compositional semantics to exploit. For instance, though the interactions for the $n$-body problem are understood (1), this does not lend itself to decomposition (2) or exploitable compositional semantics (3). Whitney [102] notes that integral mechanical system design must address safety issues at high power levels due to challenging, long-range interactions. Some aspects of mechanical system design may yield to operad analysis–e.g., bond graphs [17] or other sufficiently “diagrammatic” models–but others may not. Both examples illustrate how overnumerous or long range interaction can lead to (2). Operads can work at the system rather than component level if system properties can be extracted into compositional models. However, operads do not provide a means to extract such properties or understand problems that are truly inseparable theoretically or practically. ### 7.4 Research directions for applied operads We now briefly overview research directions toward automated analysis and synthesis. Operad-based decomposition and adaptation. Decomposition, ways a complex operation can be broken down into simpler operations, is a dual concept to the composition of operations. Any subsystem designed by a simpler operation can be adapted: precisely which operations can be substituted is known, providing a general perspective to craft algorithms. To be practical, the analytic questions of _how_ to decompose and _when_ to adapt subsystems must be answered. One research direction applies the lens of operad composition to abstract and generalize existing algorithms that exploit decomposition–e.g. to: (1) generalize superstructure optimization techniques discussed in 66.3; extend (2) extend the crossover and mutation steps for the FUN DESIGN work 6.1, which are global in the sense that they manipulate full designs, to local steps which adapt parts of a design, perhaps driven by analysis to re-work specific subsystems; and (3) explore routing as a proxy for tasking planning, analyzing foundational algorithms like Ford-Fulkerson [44] and decomposition techniques such as contraction hierarchies [45]. An intriguing, but speculative, avenue is to attempt to learn how to decompose a system or select subsystems to adapt in a data-driven way, so that the operad syntax constrains otherwise lightly supervised learning. A theoretical direction is to seriously consider the dual role of decomposition, analogous to Hopf and Frobenius algebra [33], and attempt to gain deeper understanding of the interplay of composition and decomposition, eventually distilling any results into algorithms141414For example, Bellman’s principle of optimality is _decompositional_ –i.e. parts of an optimal solution are optimal. . Multiple levels of modeling. The LSI example shows how a system model can be analysed to address different considerations. This sets the stage to adapt a design–e.g. bolster functional risk points and improve control in back and forth fashion–until both considerations are acceptable. Applied demonstrations for SAR tasking suggest a multi-level framework: (1) encoding operational concepts; (2) planning options for key maneuvers; and (3) multistage planning and scheduling to support these maneuvers. $\mathtt{UH60}$$\mathtt{UH60}$$\mathtt{UH60}$ $\mathtt{UH60}$$\mathtt{UH60}+\mathtt{HC130}$$\mathtt{UH60}+\mathtt{HC130}$$2\mathtt{UH60}$$2\mathtt{UH60}$$\mathtt{UH60}$$\mathtt{UH60}$$\mathtt{UH60}$ $\mathtt{UH60}$$\mathtt{UH60}+\mathtt{HC130}$$\mathtt{UH60}+\mathtt{HC130}$$2\mathtt{UH60}$$2\mathtt{UH60}$$\bullet$ $\mathtt{Cut}$$\bullet$ $\mathtt{Helo}$$\bullet$ $\mathtt{QD}$$\;\;a\;\;$$\;\;b\;\;$$\;\;c\;\;$$\;\;d\;\;$$\;\phantom{\Big{|}}\tau_{1}\;$$\;\phantom{\Big{|}}\tau_{2}\;$$\;\phantom{\Big{|}}\tau_{3}\;$$\;\phantom{\Big{|}}\tau_{4}\;$$\bullet$ $\mathtt{UH60}$$\bullet$ $\mathtt{HC130}$$\phantom{{|}}\tau_{1}$$\phantom{{|}}\tau_{2}$$\phantom{{|}}\tau_{4}$$(a,0)$$(b,1)$$(c,2)$$(c,2)$$((d,d),(4,4))$$2$$2$$2$$2$$2$$4$$\sqcup$$\mapsto$$2$$2$$4$$4$$+$$\mapsto$$2$$2$$4$$\bullet$ $\mathtt{Cut}$$\bullet$ $\mathtt{Helo}$$\bullet$ $\mathtt{QD}$$\;\;a\;\;$$\;\;b\;\;$$\;\;c\;\;$$\;\;d\;\;$$\;\phantom{\Big{|}}\tau_{1}\;$$\;\phantom{\Big{|}}\tau_{2}\;$$\;\phantom{\Big{|}}\tau_{3}\;$$\;\phantom{\Big{|}}\tau_{4}\;$$\bullet$ $\mathtt{UH60}$$\bullet$ $\mathtt{HC130}$Templates {‘‘colors’’ : [‘‘port’’, ‘‘cut’’, …, ‘‘qd’’], ‘‘undirected’’ : { ‘‘comms’: { ‘‘port’’: [‘‘cut’’, …, ‘‘uav" ], ‘‘cut’’: [‘‘boat’’, …, ‘‘helo’’] … } } } {‘‘colors’’ : [‘‘port’’, ‘‘cut’’, …, ‘‘u’’], ‘‘directed’’ : { ‘‘carrying’’: { ‘‘cut’’: [‘‘port’’], …, ‘‘u’’: [‘‘cut’’, …, ‘‘helo’’] } } } Core Meta-Model ${{\mathcal{O}}(\mathsf{\Gamma}_{1})}$${\mathbf{Sem}}$${{\mathcal{O}}(\mathsf{\Gamma}_{2})}$$\scriptstyle{\mathsf{A_{1}}}$$\scriptstyle{\mathsf{A_{2}}}$ Exploitation librariesGradient AscentEvol. Algorithm⋮Planning⋮ Figure 10: A software framework to leverage a meta-model: templates define each level and how to move between, libraries exploit each level, and core meta-model facilitates control across levels. Unifying top-down and bottom-up points of view. We have laid out the analytic–exemplified by wiring diagrams–and synthetic–exemplified by network operads–points of view for complex systems. Even if the goal is practical automated synthesis, scalability issues promote analytic decomposition and abstraction to efficiently reason toward satisficing solutions. Two approaches to unification include: (1) create a combined syntax for analysis and synthesis, a ‘super operad’ combining both features; (2) act by an analytic operad on the synthetic syntax, extending composition of operations. While the former approach is arguably more unified, the later more clearly separates analysis and synthesis and may provide a constructive approach to former. ### 7.5 Functorial programming with operads At this point, experience implementing operads for design suggests a software framework. While conceptually simple, this sketch helps clarify the practical role of a precise meta-model. Rather than working directly with operads to form a core meta-modeling language, cf. [18], a workflow akin to popular frameworks for JavaScript development would put developers in the drivers seat: adopters focus on controlling the flow of data and contribute to an ecosystem of libraries for lower-level data processing. Achieving this requires work before and after the meta-model. First, transferable methods get an applied problem into operads (Fig. 10, left). As in Section 4, this data constructs operads and algebras to form the core meta-model. Core data feeds explicitly exploitable data structures and algorithms to analyze (Sec. 5) and automatically construct (Sec. 6) complex systems (Fig. 10, right). On far the left, end user tools convert intent to domain inputs. Rightmost, libraries access exploitation data structures and algorithms, including those exploiting the syntax and semantics separation or substitution and adaptation. At the center, the core meta-model guarantees that the scruffier ends of the framework exposed to end users and developers are correctly aligned and coherently navigated. This framework provides significant opportunities to separate concerns compared to other approaches. Foremost, the core model separates syntax from semantics. As noted in 1, applied methods tend to conflate syntax and semantics. For instance, aggregate programming [15] provides: 1) semantics for networked components with spatial and temporal extent; and (2) interactions are proximity-based. The former feature is powerful but limiting: by choosing a single kind of semantics, modeling is wedded to the scales it abstracts well. The individual component scale is not modeled, even syntactically, which would complicate any attempt to align with other models. The latter precludes syntactic declaration of interactions–e.g. to construct architectures not purely based on proximity–and the absolute clarity about what can be put together provided by the operad syntax. Relative to computational efforts to apply operads or monoidal categories, e.g. [30, 55], this sketch places greater emphasis on specification and exploitation: specification of a domain is possible without exposing the meta-model, algorithms searching within each model are treated as black boxes that produce valid designs. Separate specification greatly facilitates set up by experts in the domain, but not the meta-model. Separate exploitation encourages importing existing data structures and algorithms to exploit each model. ### 7.6 Open problems The software framework just sketched separates out the issues of practical specification, meta-modeling and fast data structures and algorithms. We organize our discussion of open problems around concrete steps to advance these issues. In our problem statements, “multiple” means at least three to assure demonstration of the genericity of the operadic paradigm. Practical specification. The overarching question is whether the minimal combinatorial data which can specify operads, their algebras and algebra homomorphisms in theory can be practically implemented in software. We propose the following problems to advance the state-of-the-art for network template specification of operads described in Sec. 4: 1. 1. Demonstrate a specification software package for operad algebras for multiple domains. 2. 2. Develop specification software for algebra homomorphisms to demonstrate correctly aligned navigation between multiple models for a single domain. 3. 3. Develop and implement composition of specifications to combine multiple parts of a domain problem or integrate multiple domains. This last point is inline with the discussion of extending a domain in 66.2 and motivates a need to reconcile independently developed specification formats. 1. 4. Demonstrate automatic translation across specification formats. Core meta-model. As a practical matter, state-of-the-art examples exercise general principles of the paradigm but do not leverage general purpose software to encode the meta-model. 1. 5. Develop and demonstrate reusable middleware to explicitly encode multiple semantic models and maps between them which (a) takes inputs from specification packages; and (b) serves as a platform to navigate models. We have seen rich examples of focused analysis with wiring diagrams in Sec. 5 and automated composition from building blocks in Sec. 6. Theoretically, there is the question of integrating the top-down and bottom-up perspectives: 1. 6. Develop unified foundations to integrate: (a) analytic and synthetic styles of operads; and (b) composition with decomposition. Potential starting points for these theoretical advancements are described in 77.4. Developing understanding of limitations overviewed in 77.3 requires engagement with a range of applications: 1. 7. Investigate limits of operads for design to: (a) identify domains or specific aspects of domains lacking minimal data; (b) demonstrate the failure of compositionality for potentially useful semantics; and (c) characterize complexity barriers due to integrality. Navigation of effective data structures and algorithms. Lastly, there is the question of whether coherent navigation of models can be made practical. This requires explicit control of data across models and fast data structures and algorithms within specific models. The general-purpose evolutionary algorithms discussed in 66.3 motivate: 1. 8. Develop reusable libraries that exploit (a) substitution of operations and instances to adapt designs and (b) separation of semantics from syntax. SAR tasking experience and prototype explorations for distributed logistics illustrate the need to exploit moving _across_ models: 1. 9. Develop and demonstrate general purpose strategies to exploit separation across models via hierarchical representation of model fidelity–e.g. example: (a) Structure over behavior; and (b) planning over scheduling. 2. 10. Quantify the impact of separation of concerns on: (a) computational complexity; and (b) practical computation time. For this last point, _isolating_ the impact of each way to separate concerns is of particular interest to lay groundwork to systematically analyze complex domain problems. Finally, there is question of demonstrating an end-to-end system to exploit the operadic, meta-modeling paradigm. 1. 12. Demonstrate systematic, high-level control of iteration, substitution and moving across multiple models to solve a complex domain problem. 2. 13. Develop high-level control framework–similar to JavaScript frameworks for UI–or programming language–similar to probabilistic programming–to systematically control iteration, substitution and movement across multiple models. ## 8 Conclusion Operads provide a powerful meta-language to unite complementary system models within a single framework. They express multiple options for decomposition and hierarchy for complex designs, both within and across models. Diverse concerns needed to solve the full design problem are coherently separated by functorial semantics, maintaining compositionality of subsystems. Each semantic model can trade-off precision and accuracy to achieve an elegant abstraction, while algorithms exploit the specifics of each model to analyze and synthesize designs. The basic moves of iteration, substitution and moving across multiple models form a rich framework to explore design space. The trade-off is that the technical infrastructure needed to fully exploit this paradigm is daunting. Recent progress has lowered barriers to specify domain models and streamline tool chains to automatically synthesize designs from basic building blocks. Key parts of relevant theory and its implementation in software have been prototyped for example applications. Further research is needed to integrate advancements in automatic specification and synthesis with the analytic power of operads to separate concerns. To help focus efforts, we described research directions and proposed some concrete open problems. This article does not present research with ethical considerations. This article has no additional data. All authors contributed to the development of the review, its text and provided feedback and detailed edits on the work as a whole. JF coordinated the effort and led the writing of Sec. 6. SB led Sec. 5 and co-led Sec. 2, 3 and 4 with JF. ES focused on connections to applications while JD focused on assuring accessible mathematics discussion. We declare that we have no competing interests. JF and JD were supported by the DARPA Complex Adaptive System Composition and Design Environment (CASCADE) project under Contract No. N66001-16-C-4048. The authors thank John Baez, Tony Falcone, Ben Long, Tom Mifflin, John Paschkewitz, Ram Sriram and Blake Pollard for helpful discussions and two anonymous referees for comments that significantly improved the presentation. Official contribution of the National Institute of Standards and Technology; not subject to copyright in the United States. Any commercial equipment, instruments, or materials identified in this paper are used to specify the experimental procedure adequately. Such identification is not intended to imply recommendation or endorsement by the National Institute of Standards and Technology, nor is it intended to imply that the materials or equipment identified are necessarily the best available for the purpose. ## References * [1] S. Abramsky and B. Coecke, Categorical quantum mechanics, Handbook of quantum logic and quantum structures, 2 (2009), 261–325. * [2] C. Alexander, Notes on the Synthesis of Form, Harvard University Press, (1964) * [3] J. C. Baez, B. Coya and F. Rebro, Props in network theory, Theor. Appl. Categ. 33 25 (2018), 727–-783. * [4] J. C. Baez and B. Fong, A compositional framework for passive linear networks, Theor. Appl. Categ. 33 38 (2018), 1158–-1222. * [5] J. C. Baez, B. Fong, and B. S. Pollard, A compositional framework for Markov processes, J. Math. Phys. 57 3 (2016), 033301. * [6] J. C. Baez, J. Foley, J. Moeller and B. Pollard, Network models, Theor. Appl. Categ. 35 20 (2020), 700–744. * [7] J. C. Baez, J. Foley and J. Moeller, Network models from Petri nets with catalysts, Compositionality 1 4 (2019). * [8] J. C. Baez, F. Genovese, J. Master, and M. Shulman, Categories of Nets, Preprint (2021). Available as arXiv:2101.04238. * [9] J. C. Baez and J. Master, Open Petri nets, Math. Struct. Comp. Sci. 30 3 (2020), 314–341. * [10] J. C. Baez, D. Weisbart and A. Yassine, Open systems in classical mechanics, Preprint (2021). Available as arXiv:1710.11392. * [11] G. Bakirtzis, F. Genovese, and C. H. Fleming, Yoneda Hacking: The Algebra of Attacker Actions, Preprint, (2021), available as arXiv:2103:00044. * [12] G. Bakirtzis, C. H. Fleming, and C. Vasilakopoulou, Categorical Semantics of Cyber-Physical Systems Theory, ACM Transactions on Cyber-Physical Systems (2021, in press), available as arXiv:2010.08003. * [13] G. Bakirtzis, C. Vasilakopoulou, and C. H. Fleming, Compositional cyber-physical systems modeling, Proceedings 3rd Annual International Applied Category Theory Conference 2020 (ACT 2020) * [14] K. Bar, A. Kissinger and J. Vicary, Globular: an online proof assistant for higher-dimensional rewriting, Log. Methods Comput. Sci. 14 1 (2018), 1–16. * [15] J. Beal, D. Pianini and M. Viroli, Aggregate programming for the internet of things, Computer 48 9 (2015), 22–30. * [16] R. K. Brayton, G. D. Hachtel, C. McMullen and A. Sangiovanni-Vincentelli, Logic minimization algorithms for VLSI synthesis, Vol. 2. Springer Science & Business Media, 1984. * [17] B. Coya, Circuits, Bond Graphs, and Signal-Flow Diagrams: A Categorical Perspective, PhD thesis, University of California–Riverside, 2018. * [18] A. Boronat, A. Knapps, J. Meseguer and M. Wirsing, What is a multi-modeling language? International Workshop on Algebraic Development Techniques, Springer, Berlin, Heidelberg (2008), 71–87. * [19] S. Breiner, B. Pollard and E. Subrahmanian, Workshop on Applied Category Theory: Bridging Theory and Practice. Special Publication (NIST SP) 1249 (2020). * [20] S. Breiner, B. Pollard, E. Subrahmanian and O. Marie-Rose, Modeling Hierarchical System with Operads, Proceedings of the 2019 Applied Category Theory Conference, (2020) 72–83. * [21] S. Breiner, R. D. Sriram and E. Subrahmanian, Compositional models for complex systems, Artificial Intelligence for the Internet of Everything, eds. Academic Press, Cambridge Massachusetts (2019) 241–270. * [22] K. S. Brown, D. I. Spivak and R. Wisnesky, Categorical data integration for computational science, Comput. Mater. Sci. 164 (2019), 127–132. * [23] S. Busboom, Bat 21: A Case Study, Carlisle Barracks, PA: U.S. Army War College, (1990). * [24] A. Censi, A mathematical theory of co-design, Preprint (2015). Available as arXiv:1512.08055. * [25] A. Censi, Uncertainty in Monotone Codesign Problems, IEEE Robotics and Automation Letters 2 3 (2017), 1556–1563. * [26] M. Chechik, S. Nejati and M. Sabetzadeh, A relationship-based approach to model integration, Innovations in Systems and Software Engineering 8 1 (2012), 3–18. * [27] N. Chungoora and R. I. Young, Semantic reconciliation across design and manufacturing knowledge models: A logic-based approach, Applied Ontology 6 4 (2011), 295–295. * [28] B. Coecke, M. Sadrzadeh and S. Clark, Mathematical foundations for a compositional distributional model of meaning, Linguistic Analysis 36 (2010), 345–-384. * [29] H. Choi, C. Crump, C. Duriez, A. Elmquist, G. Hager, D. Han, F. Hearl, J. Hodgins, A. Jain, F. Leve, and C. Li, On the use of simulation in robotics: Opportunities, challenges, and suggestions for moving forward, Proc. Natl. Acad. Sci. U.S.A. 118 1 (2021) * [30] G. de Felice, A. Toumi and B. Coecke, DisCoPy: Monoidal Categories in Python, Proceedings 3rd Annual International Applied Category Theory Conference 2020 (ACT 2020). * [31] Z. Diskin and T. Maibaum, Category theory and model-driven engineering: From formal semantics to design patterns and beyond. Model-Driven Engineering of Information Systems: Principles, Techniques, and Practice 173 (2014). * [32] E. Di Lavore, J. Hedges, and P. Sobociński, Compositional Modelling of Network Games, Computer Science Logic 2021, Leibniz International Proceedings in Informatics 183 (2021). * [33] P. Dusko, Monoidal computer I: Basic computability by string diagrams, Info. Comp. 226 (2013), 94–116. * [34] S. Eilenberg and S. Mac Lane, General Theory of Natural Equivalences, Trans. AMS 58 (1945), 231–294. * [35] A. Fagan and R. Duncan, Optimising Clifford Circuits with Quantomatic, Proceedings 15th International Conference on Quantum Physics and Logic (QPL 2018), (2019) 85–-105. * [36] A. Ferrari and A. Sangiovanni-Vincentelli, System design: Traditional concepts and new paradigms, Proceedings 1999 IEEE International Conference on Computer Design: VLSI in Computers and Processors, (1999) 2–-12. * [37] J. D. Foley, An example of exploring coordinated SoS behavior with an operad and algebra integrated with a constraint program, CASCADE tech report, 2018. * [38] B. Fong, The algebra of open and interconnected systems, DPhil thesis, University of Oxford, 2016. * [39] B. Fong and M. Johnson, Lenses and learners, Proceedings of the Eighth International Workshop on Bidirectional Transformations (Bx 2019), (2019). * [40] B. Fong and D. I. Spivak, An invitation to applied category theory: seven sketches in compositionality, Cambridge University Press, 2019. * [41] B. Fong and D. I. Spivak, Hypergraph categories, J. Pure Appl. Algebra 223 11 (2019), 4746–4777. * [42] B. Fong and D. I. Spivak, Supplying bells and whistles in symmetric monoidal categories, Preprint (2020). Available as arXiv:1908.02633. * [43] B. Fong, D. I. Spivak and R. Tuyéras, Backprop as functor: A compositional perspective on supervised learning, 2019 34th Annual ACM/IEEE Symposium on Logic in Computer Science (LICS), 1–13. * [44] L. R. Ford and D. R. Fulkerson, Maximal flow through a network, Can. J. of Math. 8 (1956), 399–404. * [45] R. Geisberger, P. Sanders, D. Schultes and D. Delling, Contraction hierarchies: Faster and simpler hierarchical routing in road networks, International Workshop on Experimental and Efficient Algorithms, Springer, Berlin, Heidelberg (2008), 319–333. * [46] T. Giesa, R. Jagadeesan, D. I. Spivak and M. J. Buehler, Matriarch: a python library for materials architecture, ACS biomaterials science & engineering, 1 10 (2015), 1009–1015. * [47] D. R. Ghica, A. Jung and A. Lopez, Diagrammatic Semantics for Digital Circuits, 26th EACSL Annual Conference on Computer Science Logic (CSL 2017), Leibniz International Proceedings in Informatics, (2017) 82 24:1–24:16. * [48] N. Ghani, C. Kupke, A. Lambert and F. N. Forsberg, Compositional Game Theory with Mixed Strategies: Probabilistic Open Games Using a Distributive Law, Proceedings of the 2019 Applied Category Theory Conference, (2020) 95–-105. * [49] N. Ghani, J. Winschel and P. Zahn, Compositional game theory, Proceedings of the 33rd Annual ACM/IEEE Symposium on Logic in Computer Science, (2018) 472–481. * [50] J. Girard, Linear logic, Theor. Comput. Sci. 50 1 (1987), 1–101. * [51] E. Grefenstette, M. Sadrzadeh, S. Clark, B. Coecke, and S. Pulman, Concrete sentence spaces for compositional distributional models of meaning, Computing meaning, Springer, Dordrecht (2018) 71-86. * [52] J. Guy, 142 Evacuated From Creek Fire by Military Helicopters; Body of Deceased Man Flown to Fresno, Fresno Bee, Sept. 8, 2020. * [53] K. Gürlebeck, D. Hofmann and D. Legatiuk, Categorical approach to modelling and to coupling of models, Math. Meth. Appl. Sci. 40 3 (2017), 523–534. * [54] J. Hedges and M. Sadrzadeh, A generalised quantifier theory of natural language in categorical compositional distributional semantics with bialgebras, Math. Struct. Comp. Sci. 29 6 (2019), 783–809. * [55] M. Halter, E. Patterson, A. Baas and J. Fairbanks, Compositional Scientific Computing with Catlab and SemanticModels, Preprint (2020). Available as arXiv:2005.04831. * [56] C. Hermida. Representable Multicategories, Advances in Mathematics 151 2 (2000), 164-225. * [57] P. Johnson-Freyd, J. Aytac, and G. Hulett , Topos Semantics for a Higher-Order Temporal Logic of Actions, Proceedings of the 2019 Applied Category Theory Conference, (2020) 161–-171. * [58] A. Kissinger and J. van de Wetering, PyZX: Large Scale Automated Diagrammatic Reasonin, Proceedings 16th International Conference on Quantum Physics and Logic (QPL 2019), (2020) 229–-241. * [59] A. Kissinger and J. van de Wetering, Reducing the number of non-Clifford gates in quantum circuits, Physical Review A, 102 2 (2020) 022406. * [60] A. Kissinger and V. Zamdzhiev, Reducing the number of non-Clifford gates in quantum circuits, International Conference on Automated Deduction, Springer, Cham (2015) 326–336. * [61] C. S. Khor, B. Chachuat and N. Shah, A superstructure optimization approach for water network synthesis with membrane separation-based regenerators, Computers & chemical engineering 42 (2012), 48–63. * [62] R. Kuwada, J. Guy and D. Cooper, Creek Fire roars toward mountain resort towns, after airlift rescues hundreds trapped by flames, Fresno Bee, Sept. 6, 2020. * [63] E. A. Lee, The past, present and future of cyber-physical systems: A focus on models. Sensors, 15 3 (2015), 4837–4869. * [64] T. Leinster, Higher operads, higher categories, London Math. Soc. Lec. Note Series, 298 (2003) * [65] N. Leveson, The Drawbacks in Using The Term ‘System of Systems’, Biomedical Instrumentation & Technology, (2013), 115–118. * [66] C. Lisciandra, and J. Korbmacher, Multiple models, one explanation, J. Econ. Methodol. , (2021), 1–21. * [67] M. A. Mabrok and M. J. Ryan, Category theory as a formal mathematical foundation for model-based systems engineering, Appl. Math. Inf. Sci. 11 (2017), 43–51. * [68] S. Mac Lane. Categories for the Working Mathematician, 2nd edition, Vol. 5. Springer, (1998) * [69] E. Marder and A. L. Taylor, Multiple models to capture the variability in biological neurons and networks, Nat. Neurosci. 14 2 (2011), 133–138. * [70] M. Markl, S. Shnider and J. D. Stasheff, Operads in Algebra, Topology and Physics, AMS, 2002 * [71] J. Master, E. Patterson, and A. Canedo, String Diagrams for Assembly Planning, DIAGRAMS 2020 11th International Conference on the Theory and Application of Diagrams (2020) * [72] J. Meseguer and U. Montanari, Petri nets are monoids, Inf. Comput. 88 (1990), 105–155. * [73] J. Moeller, Noncommutative network models, Math. Struct. Comp. Sci. 30 1 (2020), 14–32. * [74] J. Moeller and C. Vasilakopoulou, Monoidal Grothendieck Construction, Theor. Appl. Categ. 35 31 (2020), 1159–1207. * [75] S. M. Neiro and J. M. Pinto, Supply chain optimization of petroleum refinery complexes, Proceedings of the 4th International Conference on Foundations of Computer-Aided Process Operations, (2003), 59–72. * [76] J. S. Nolan, B. S. Pollard, S. Breiner, D. Anand, and E. Subrahmanian, Compositional models for power systems, Proceedings of the 2019 Applied Category Theory Conference, (2020) 72–83. * [77] S. M. Patterson, D. I. Spivak and D. Vagner, Wiring diagrams as normal forms for computing in symmetric monoidal categories, Proceedings of the 2020 Applied Category Theory Conference (ACT 2020) * [78] B. S. Pollard, Open Markov processes: A compositional perspective on non-equilibrium steady states in biologys, Entropy 18 (2016), 140. * [79] A. Quarteroni, Mathematical models in science and engineering, Not. Am. Math. Soc. 56 1 (2009), 10–19. * [80] R. Raman and I. E. Grossmann, Integration of logic and heuristic knowledge in MINLP optimization for process synthesis, Computers & chemical engineering 16 3 (1992), 155–171. * [81] D. Reutter and J. Vicary, High-level methods for homotopy construction in associative n-categories, 2019 34th Annual ACM/IEEE Symposium on Logic in Computer Science (LICS), IEEE (2019), 1–13. * [82] A. Sangiovanni-Vincentelli, The tides of EDA, IEEE Design & Test of Computers 20 6 (2003), 59–75. * [83] P. Schultz and D. I. Spivak, Temporal Type Theory, Springer International Publishing, 2019. * [84] G. Simon, T. Levendovszky, S. Neema, E. Jackson, T. Bapty, J. Porter and J. Sztipanovits, Foundation for model integration: Semantic backplane, International Design Engineering Technical Conferences and Computers and Information in Engineering Conference, Vol. 45011, (2012), 1077–1086. American Society of Mechanical Engineers. * [85] H. A. Simon, Rational decision making in business organizations, Am. Econ. Rev. 69 4 (1979), 493–513. * [86] H. A. Simon, The architecture of complexity, Facets of systems science, Springer, (1991), 457–476. * [87] M. Sirjani, E. A. Lee, and E. Khamespanah, Verification of Cyberphysical Systems, Mathematics, 8 7, (2020), 1068. * [88] J. A. Sokolowski , and C. M. Banks, Modeling and simulation fundamentals: theoretical underpinnings and practical domains, John Wiley & Sons, (2010) * [89] M. E. Sosa, S. D. Eppinger, and C. M. Rowles, Identifying modular and integrative systems and their impact on design team interactions, J. Mech. Des. 125 2 (2003) 240–252. * [90] D. I. Spivak, The operad of wiring diagrams: formalizing a graphical language for databases, recursion, and plug-and-play circuits, Preprint (2013). Available as arXiv:1305.0297. * [91] D. I. Spivak, Category Theory for the Sciences, MIT Press, Cambridge Massachusetts, 2014. * [92] D. I. Spivak, T. Giesa, E. Wood and M. J. Buehler, Category theoretic analysis of hierarchical protein materials and social networks, PloS one 6 9 (2011). * [93] D. I. Spivak and R. E. Kent, Ologs: a categorical framework for knowledge representation, PloS one 7 1 (2012). * [94] D. I. Spivak and J. Tan, Nesting of dynamical systems and mode dependent networks, J. Complex Networks 5 3 (2017), 389–408. * [95] D. I. Spivak and R. Wisnesky, Fast Left-Kan Extensions Using The Chase, Preprint (2020) Available at www.categoricaldata.net. * [96] K. O. Stanley and R. Miikkulainen, Evolving neural networks through augmenting topologies, Evolutionary Comp. 10 2 (2002), 99–127. * [97] L. D. Stone, J. O. Royset, and A. L. Wasburn. Optimal Search for Moving Targets, Vol. 237. Springer, 2016. * [98] M. E. Szabo, Algebra of proofs, Studies in logic and the foundations of mathematics, Vol. 88. North-Holland Publishing Company, 1978. * [99] A. M. Turing, The Chemical Basis of Morphogenesis, Philos. Trans. R. Soc. Lond., B, Biol. Sci. 237 641 (1952), 37–72. * [100] D. Vagner, D. I. Spivak, and E. Lerman, Algebras of open dynamical systems on the operad of wiring diagrams, Theor. Appl. Categ. 30 55 (2015), 1793–1822. * [101] R. Wisnesky S. Breiner, A. Jones, D. I. Spivak and E. Subrahmanian, Using category theory to facilitate multiple manufacturing service database integration, Journal of Computing and Information Science in Engineering 17 2 (2017) * [102] D. E. Whitney, Physical limits to modularity, (2002) * [103] D. Yau, Colored Operads, American Mathematical Society, Providence, Rhode Island, 2016. * [104] D. Yau, Operads of Wiring Diagrams, Vol. 2192. Springer, 2018. * [105] H. Yeomans and I. E. Grossmann, A systematic modeling framework of superstructure optimization in process synthesis, Computers & Chemical Engineering 23 6 (1999), 709–731. * [106] G. Zardini, N. Lanzetti, M. Salazar, A. Censi, E. Frazzoli, and M. Pavone, Towards a co-design framework for future mobility systems, Annual Meeting of the Transportation Research Board (2020). * [107] G. Zardini, D. Milojevic, A. Censi, and E. Frazzoli, Co-Design of Embodied Intelligence: A Structured Approach, Preprint, (2020), available as arXiv:2011:10756. * [108] G. Zardini, D. I. Spivak, A. Censi, and E. Frazzoli, A Compositional Sheaf-Theoretic Framework for Event-Based Systems (Extended Version), Preprint, (2020), available as arXiv:2005:04715.
# Exact and Approximate Heterogeneous Bayesian Decentralized Data Fusion Ofer Dagan, Nisar R. Ahmed Manuscript received Month day, 2021; revised Month day, 2021.The authors are with the Smead Aerospace Engineering Sciences Department, University of Colorado Boulder, Boulder, CO 80309 USA (e-mail: <EMAIL_ADDRESS>Nisar.Ahmed@colorado.edu). ###### Abstract In Bayesian peer-to-peer decentralized data fusion, the underlying distributions held locally by autonomous agents are frequently assumed to be over the same set of variables (homogeneous). This requires each agent to process and communicate the full global joint distribution, and thus leads to high computation and communication costs irrespective of relevancy to specific local objectives. This work studies a family of heterogeneous decentralized fusion problems, where the set of problems in which either the communicated or the processed distributions describe different, but overlapping, states of interest that are subsets of a larger full global joint state is considered. We exploit the conditional independence structure of such problems and provide a rigorous derivation for a family of exact and approximate heterogeneous conditionally factorized channel filter methods. We further extend existing methods for approximate conservative filtering and decentralized fusion in heterogeneous dynamic problems. Numerical examples show more than 99.5% potential communication reduction for heterogeneous channel filter fusion, and a multi-target tracking simulation shows that these methods provide consistent estimates. ###### Index Terms: Bayesian decentralized data fusion (DDF), distributed robot systems, multi- robot systems, sensor fusion. ## I Introduction Bayesian decentralized data fusion (DDF) has a wide range of applications, such as cooperative localization [1], multi-target tracking [2], multi-robot localization and mapping [3], and more. Decentralized data fusion, while generally less accurate compared to centralized data fusion, offers advantages in terms of scalability, flexibility and robustness. One of the challenges of decentralized data fusion stems from the difficulty of accounting for common data and dependencies between communicating agents and avoiding ‘rumor propagation’, where dependencies between data sources are incorrectly ignored. In decentralized multi-agent systems aiming at some joint mission, such as autonomous cooperative robot localization [1],[4], [5], or tracking [2], the requirement for each agent to recursively update and communicate a global full joint posterior probability distribution function (pdf) over an identical (homogeneous) set of states leads to large overhead in local processing and communication bandwidth. It is therefore desirable to enable processing, communication and fusion of a posterior pdf over a subset of different but overlapping states; we name such a process _heterogeneous fusion_. Consider for example the 30 robot cooperative localization scenario given in [1]. If each agent has 4 unknown position states, the full joint distribution has 120 variables, and requires processing a $120\times 120$ covariance matrix (assuming Gaussian distribution). This includes states of agents ‘far away’ from each other in the network, which has negligible effect on the local position estimate and might be considered ‘irrelevant’. But, if each agent includes in its estimate only immediate ‘relevant’ neighbors’ states which form a subset of the global joint distribution, then the local heterogeneous joint distribution shrinks, e.g. to 16 states for a 3-neighbor topology. This has a clear computation and communication gain over the homogeneous alternative. However it might lead to indirect correlations between variables not mutually monitored by both agents and result in an overconfident estimate. To the best of our knowledge, existing methods have yet to solve this problem and accurately account for the indirect correlations. DDF algorithms can be considered exact or approximate, depending on how they account for dependencies in the data shared between agents, to guarantee that every piece of data is counted not more than once. In exact methods, these dependencies are explicitly accounted for, either by pedigree-tracking, which can be cumbersome and impractical in large ad hoc networks [6], or by adding a _channel filter_ (CF), which requires enforcing a tree-structured communication topology [7, 8]. Approximate methods assume different levels of correlation between the communicating agents and fuse them in such a way that the common data is promised not to be double counted, thus ensuring conservativeness of the fused posterior. The most commonly used approximate method is covariance intersection (CI) [9], when agents share only the first two moments (mean and covariance) of their distributions, or the geometric mean density (GMD) for general pdfs fusion [10]. In CI, the fused result is a weighted average of the information vector and matrix describing the first and second moments of the underlying distributions, respectively, where the weight is optimized based on a predetermined cost function, e.g., determinant of the posterior fused covariance matrix. Both exact and approximate fusion methods usually assume that the distributions to be fused as well as the posterior distributions are homogeneous, i.e. describe the same full joint state random vector. However, these methods cannot be directly applied to heterogeneous fusion, where pdfs over overlapping parts of the full joint pdfs are fused. This paper builds upon the work presented in [11] and further develops a rigorous Bayesian probabilistic approach for fusion of heterogeneous pdfs in tree-structured networks. The key idea is to utilize conditional independence properties between subsets of variables (states) in order to lower the dimension of the local joint distributions to be fused. In [11] the applicability of the conditionally factorized channel filter (CF2) algorithms for dynamic systems is limited since the full time history has to be tracked for conditional independence. This paper improves and extends the theory and applicability of the work presented in [11] by making the following contributions: 1. 1. The applicability of the CF2 heterogeneous fusion algorithms are extended to dynamic problems by (i) enabling conservative filtering (ii) developing the information augmented state (iAS) version of the augmented state algorithm [12]. 2. 2. The definition and desiderata for a heterogeneous fusion rule are more formally and clearly defined, along with practical measures for evaluation. 3. 3. The theory developed in this paper is brought to practice with a detailed explanation of two of the CF2 algorithms and pseudo-code for the case of Gaussian distributions is presented. 4. 4. The scalability of the heterogeneous fusion methods is demonstrated by an extended example, calculating the communication and computation savings, and covering different problem sizes. 5. 5. New simulations of dynamic decentralized multi-target tracking scenarios are provided, to demonstrate conservative filtering and compare it to the full time history smoothing approach. The rest of this paper consists of three main parts. The first (Sec. II-V) discusses heterogeneous fusion in terms of general probability distributions (pdfs); Sec. II defines the heterogeneous decentralized fusion problem and reviews the state of the art; Sec. III derives a family of exact and approximate conditionally factorized CF methods; Conditional independence in dynamic systems, including augmented state (AS) density and conservative filtering for general pdfs, is discussed in Sec. IV, and Sec. V details the proposed fusion algorithm. In the second part (Sec. VI) Gaussian distributions are treated as special case to develop: a closed-form fusion CF2 rule; the information augmented state (iAS) method; and a conservative Kalman filtering method. In the third part (Sec. VI-F-VII), numerical examples demonstrate the potential communication and computation gains of the described methods (Sec. VI-F) via static and dynamic multi-agent multi-target tracking simulations (Sec. VII). Sec. VIII then draws conclusions and suggests avenues for additional work. ## II Problem Statement and Related Work To motivate the problem, consider a simple static target tracking problem with one target and two tracking agents as a running example (Fig. 1a). Both agents $i$ and $j$ track the position of the target, described by the random state vector $x$, and are assumed to have perfect knowledge of their own position, but unknown constant biases in the agent-target relative measurement vectors $y_{i,k}$ and $y_{j,k}$, described by the random state vectors $s_{i}$ and $s_{j}$, respectively, where $k$ indicates the time step. The agents can also take relative observations $m_{i,k}$ and $m_{j,k}$ to different landmarks to locally collect data on their measurement biases. As shown in Fig. 1a, the agents’ local biases $s_{i}$ and $s_{j}$ become coupled due to measurements $y_{i,k}$ and $y_{j,k}$ of the common target $x$. In the case of homogeneous information fusion, the two agents preform inference over and communicate the full joint pdfs describing all the random variables, including each other’s local biases. But in the heterogeneous fusion case, agents might hold a pdf over only a subset of the random variables, for example over the target ($x$) and the agent’s local bias (e.g., $s_{i}$), making the dependencies between the local biases hidden. Therefore an agent might not be aware of the existence of another local bias random state vector (e.g., $s_{j}$) held by another agent. These dependencies are key to the problem, and the main challenge in heterogeneous fusion compared to homogeneous fusion is to account for them during fusion, where they are not explicitly represented in the local posterior pdfs. (a)$s_{i}$$s_{j}$$x$$y_{i,k}$$y_{j,k}$$m_{i,k}$$m_{j,k}$$k=1:N$$k=1:N$(b)$s_{i}$$s_{j}$$x_{1}$$x_{2}$$x_{k}$$y_{i,1}$$y_{j,1}$$y_{i,2}$$y_{j,2}$$y_{i,k}$$y_{j,k}$$m_{i,k}$$m_{j,k}$$k=1:N$ Figure 1: (a) Static and (b) Partially dynamic Bayesian networks for two local random vectors $s_{i},s_{j}$ (local measurement biases) and one common random vector $x$ (target state). In (a), $s_{i}$ and $s_{j}$ are conditionally independent given the static state $x$; in (b) the full time history $x_{1:k}$ is required for conditional independence. ### II-A Problem Statement Assume a network of $n_{a}$ autonomous agents, performing recursive decentralized Bayesian updates to their prior pdf, with the goal of monitoring some global random state vector $\chi_{k}$ at time $k$. Each agent $i\in\\{1,...,n_{a}\\}$ is tasked with a local subset of states $\chi_{i,k}\subseteq\chi_{k}$, an $n_{i}$-dimensional vector of random variables at time $k$. An agent can update its local prior pdf for $\chi_{i,k}$, by (i) using Bayes’ rule to recursively update a posterior pdf for $\chi_{i,k}$ with independent sensor data $Y_{i,k}$ described by the conditional likelihood $p(Y_{i,k}|\chi_{i,k})$, and (ii) performing peer-to- peer fusion of external data relevant to $\chi_{i,k}$ from any neighboring agent $j\in N_{a}^{i}$, where $N_{a}^{i}$ is the set of agents communicating with $i$. The heterogeneous fusion question is now: what is a peer-to-peer fusion rule $\mathbb{F}$ which given the local prior distribution $p_{i}(\chi_{i,k}|Z^{-}_{i,k})$ and a distribution over ‘relevant’ random states from a neighboring agent $j$, $p_{j}(\chi_{j,k}^{r}|Z^{-}_{i,k})$, returns a local fused posterior combining the data from both agents, $p_{i}(\chi_{i,k}|Z_{i,k}^{+})$, $\begin{split}p_{i}(\chi_{i,k}|Z_{i,k}^{+})=\mathbb{F}(p_{i}(\chi_{i,k}|Z^{-}_{i,k}),p_{j}(\chi_{j,k}^{r}|Z^{-}_{j,k})).\end{split}$ (1) Where $\chi_{j,k}^{r}$ is the subset of random states at agent $j$ that are relevant to agent $i$ and assumed to be a non-empty set. $Z_{i,k}^{-}$ ($Z_{j,k}^{-}$) is the local data agent $i$ ($j$) has prior to fusion with agent $j$ ($i$), and $Z_{i,k}^{+}\equiv Z_{i,k}^{-}\cup Z_{j,k}^{-}$ is the combined data after fusion. Notice that while the motivation is heterogeneous fusion, the above statement is general and can be used for homogeneous and heterogeneous fusion. For instance, in the target tracking example, if $\chi_{i,k}=\chi_{j,k}=\chi_{j,k}^{r}=[x^{T},s_{i}^{T},s_{j}^{T}]^{T}$ then (1) simplifies to a homogeneous fusion problem, which can be solved using different exact or approximate fusion rules, as discussed in the introduction. However, if $\chi_{i,k}=[x^{T},s_{i}^{T}]^{T}$ and $\chi_{j,k}=[x^{T},s_{j}^{T}]^{T}$, then the ‘relevant’ random states in $j$ are $\chi_{j,k}^{r}=x$, and (1) describes heterogeneous fusion. In this case, note that $i$’s knowledge about $s_{i}$, described by the marginal pdf $p(s_{i}|Z_{i,k}^{+})$, should still be updated following fusion with respect to $x$ (likewise for $j$’s marginal pdf over $s_{j}$). This key distinction separates conventional homogeneous fusion from heterogeneous fusion: in the latter, agents seek to update their posterior pdf over their entire local random state vector, using data gained only from fusion over relevant subsets of local random state vectors. Heterogeneous fusion thus encompasses the set of problems where the set of relevant random states $\chi_{j,k}^{r}$ is a subset of either agent $i$’s random states $\chi_{j,k}^{r}\subset\chi_{i,k}$, or agent $j$’s random states $\chi_{j,k}^{r}\subset\chi_{j,k}$ or both. The goal of this paper is to develop the theory and understanding of the heterogeneous fusion problem defined above. To be able to track the information flow in the system, the following assumptions are made: 1. 1. The $n_{a}$ agents all communicate in a bilateral tree-structured network topology. This guarantees that information can only flow in one path between any two agents, thus avoiding loops. 2. 2. Full rate communication and sequential processing of incoming messages from other agents (where agent sends a message at time step $k$ before processing messages received at the same time step $k$). 3. 3. The random state vector definitions between neighboring agents in the network communication graph have at least one overlapping random state in their local random state vector. Further, if agent $l$ is more than one hop away from agent $i$ and has an overlapping random state with $i$, then all agents on the tree path from $i$ to $l$ also share that common random state. 4. 4. At time step $k=0$, if there is a common prior distribution over the common states of interest it is known to both agents, i.e., $p_{i}(\chi_{i,0}\cap\chi_{j,0})=p_{j}(\chi_{i,0}\cap\chi_{j,0})=p_{c}(\chi_{i,0}\cap\chi_{j,0})$ where here $p_{c}$ is the common prior pdf. If there is no common prior, then $p_{c}(\chi_{i,0}\cap\chi_{j,0})=1$. 5. 5. The global problem structure (representing the full joint distribution over $\chi_{k}$) for state estimation at the agents is such that it allows conditional independence between random states that are local to each agent given (conditioned on) the common random states between them. An example for these assumptions is given in Fig. 2. Shown is a 5-agent, 6-target tracking problem, where the local measurements to targets $x_{t}(t=1,..6)$ and landmarks to estimate local biases $s_{i}(i=1,..5)$ are denoted by the full black arrows. The tree structured communication topology is indicated by the dashed red arrows between the agents. It can be seen that due to the bilateral communication between any two agents there are no loops in the network and information can flow only in one path. Assumption 3 is also demonstrated, for example, if agents 3 and 5 both estimate target 5’s random states, agent 4 must estimate it as well. Finally, the conditional independence holds, as for example agent 1’s local random state vectors $x_{1}$ and $s_{1}$ are conditionally independent from agent 2’s local random state vectors $x_{3}$ and $s_{2}$ given the common target random state vector $x_{2}$ ($x_{1},s_{1}\perp x_{3},s_{2}|x_{2}$). Figure 2: Target tracking example. Full black arrows denote local measurements to targets $x_{t}$ and landmarks to estimate local biases $s_{i}$, red dashed arrows indicate bi-directional communication channel between agents. The tree structure topology can be seen, as there are no loops. While assumption 1-4 can be easily enforced, assuming conditional independence (assumption 5) is not trivial in dynamic problems. In [11] the full time history is augmented to maintain conditional independence, but the problem of regaining conditional independence, while keeping the local pdf consistent and conservative when filtering (marginalizing out past random states) remains unsolved. Thus, in addition to a fusion rule $\mathbb{F}$, a complementary conservative filtering operation is needed to regain conditional independence. ### II-B What is A Good Fusion Rule? In the above formulation of the problem it is left to define a good fusion rule $\mathbb{F}$ in (1) and how to evaluate it. Recently Lubold and Taylor [13] claim that a fusion rule should provide a posterior pdf which is conservative, i.e., overestimates the uncertainty relative to the true pdf. They suggest new definitions for conservativeness, but to the best of our knowledge it is not widely used. In the context of homogeneous fusion, common definitions use the terms consistent and conservative interchangeably [9], [14] and assume that the point estimate uncertainty can be described by mean and covariance. In the following, the intuition regarding conservativeness from [13] is combined with the common definitions of consistency from homogeneous fusion to define a ‘good’ heterogeneous fusion rule $\mathbb{F}$, firstly in terms of pdfs and then in the case of Bayesian point estimation. From the standpoint of pdf fusion, a good heterogeneous fusion rule $\mathbb{F}$ results in an updated local posterior pdf that: (i) overestimates the uncertainty relative to the true pdf, and (ii) is conservative over the agent’s random states of interest $\chi_{i,k}\subseteq\chi_{k}$ relative to the marginal pdf over $\chi_{i,k}$ of a consistent centralized pdf. Consistency here means that the fused result does not overestimate or underestimate the uncertainty. The centralized pdf refers to the posterior pdf over the full random state vector $\chi_{k}$ conditioned on all the available data from all the agents up and including time step $k$, $p(\chi_{k}|\bigcup_{i\in N_{a}}Z^{-}_{i,k})$. Since consistency and conservativeness are often defined by the first two moments of the pdf, i.e., their mean and covariance, the above definition can be further narrowed in the context of Bayesian point estimation. A good heterogeneous fusion rule $\mathbb{F}$ in this case is then one that when forming a point estimate from its resulting local posterior $p_{i}(\chi_{i,k}|Z_{i,k}^{+})$, for example by finding the minimum mean squared error (MMSE) estimate, the estimate: (i) overestimates the uncertainty relative to the true state error statistics, and (ii) is conservative relative to the marginal error estimate of a consistent centralized point estimator. For example, assume the means of the Gaussian random state vectors $\chi_{i}$ and $\chi$ are $\mu_{\chi_{i}}$ and $\mu_{\chi}$, and the covariances, describing the mean squared error, are $\Sigma_{\chi_{i}}=E[(\chi_{i}-\mu_{\chi_{i}})(\chi_{i}-\mu_{\chi_{i}})^{T}]$ and $\Sigma_{\chi}=E[(\chi-\mu_{\chi})(\chi-\mu_{\chi})^{T}]$, respectively, where $E[\cdot]$ is the expectation operator. The actual values are unknown, and the approximate estimate of them is given by $\bar{\mu}_{\chi_{i}}$, $\bar{\mu}_{\chi}$, and $\bar{\Sigma}_{\chi_{i}}$, $\bar{\Sigma}_{\chi}$. The definitions above then translate to the following: 1. (i) Overestimation of the uncertainty relative to the true error statistics implies that $\bar{\Sigma}_{\chi_{i}}-\Sigma_{\chi_{i}}\succeq 0$, i.e., the resulting matrix difference is positive semi-definite (PSD). 2. (ii) Conservativeness relative to the marginal estimate of the centralized estimator implies that $\bar{\Sigma}_{\chi_{i}}-\bar{\Sigma}_{\chi_{i}}^{cent}\succeq 0$, where $\bar{\Sigma}_{\chi_{i}}^{cent}$ is the marginal covariance over $\chi_{i}$, taken from the joint centralized covariance over $\chi$. The centralized estimate in this case can be considered consistent if, for example, it passes the NEES hypothesis test [15]. Note that since a consistent centralized estimate neither overestimates nor underestimates the uncertainty, a conservative (higher uncertainty) local estimate is expected to underestimate the uncertainty. Thus requiring the local estimate to be conservative relative to a consistent centralized estimate implicitly requires it to not overestimate the uncertainty of the true error statistics. ### II-C Related Work The idea of splitting a global state vector into $N$ subsets of $N_{a}$ vectors ($N_{a}\ll N$) has been posed to solve two different problems. The first problem, as presented for example in [16, 17, 18, 19], tries to reconstruct an MMSE _point estimate_ of the global state vector, by fusing $N_{a}$ local heterogeneous _point estimates_. The second problem, which is the focus of this paper, tries to find $N_{a}$ different local _posterior pdfs_ , given all the information locally available up to the current time. The problem of locally fusing data from a non-equal state vectors dates back to [20]. While this work offers a simple way for distributing into heterogeneous local state vectors, it assumes that the local pdfs are decoupled. In [21], Khan _et al._ relax that assumption, but ignore dependencies between states in the fusion step and restrict the state distribution to only states that are directly sensed by local sensors, where in practice, an agent might be ‘interested’ in a larger set of local states that are not all directly sensed. Similarly in cooperative localization [4], [5] it is often assumed that agents are neighbors only if they take relative measurement to each-other. Then, the local state is augmented with the other agent’s position states to process the measurement, often by assuming or approximating marginal independence. It can be shown that this approach represents a subset of the heterogeneous fusion problems considered in the paper. In [22], Chong and Mori use conditional independence in Bayesian networks to reduce the subset of states to be communicated. However, they assume hierarchical fusion and for dynamic systems only consider the case of deterministic state processes, omitting the important case of stochastic processes which is considered in this paper. The work by Whitacre _et al._[2], does not formally discuss conditional independence but introduces the idea of marginalizing over common states to fuse Gaussian distributions, when cross correlations of common states are known. However, they implicitly assume cross correlations among the conditioned, or unique, states are small. While this assumption might hold for their application, it does not offer a robust solution. Reference [23] uses CI with Gaussian distributions to relax the assumption of known correlations of the fused state. Reference [24] suggests a similar solution to [23] but restrict it to systems where all agents have the same set of common states. Although scalable and simple, such approaches do not account for dependencies between locally exclusive states. This work aims at gaining insight and understanding on how such issues should be addressed by exploiting the structure of the underlying dynamical system in such problems. ## III Conditionally Factorized Channel Filters - CF2 The approach presented here assumes a tree-structured communication network and starts with a general probabilistic channel filter (CF) formulation. Then, as seen in Fig. 3, different fusion problems of interest are discussed. Starting from the original homogeneous CF, where all agents keep and share posterior pdf of the full global joint random state vector, cases are then considered where an agent is only interested in and/or observes a subset of the global joint distribution. It is shown that by leveraging conditional independence, agents can communicate only marginal pdfs, leading to the _Factorized CF (F-CF)_ and the _Bi-directional factorized CF (BDF-CF)_ , depending on whether marginal pdf is sent in one or two directions, respectively. These extensions to the CF enable communication reduction by sending only new and relevant data while maintaining the accuracy of the original CF. Additionally, a branch of approximate CF methods is introduced which includes the _approximate BDF-CF_ and the _heterogeneous state CF (HS-CF)_. In these methods, agents communicate only marginal pdf over common subsets, leaving the conditional pdfs over the unique subset of variables local at each agent. The main difference between these two methods is the size of the local joint distributions, which influences the local processing requirements. In approximate BDF-CF, each agent processes its unobserved variables in order to get a rough estimate over less relevant random states. However, in the HS-CF, the state space for the local joint distribution is minimized to hold only locally relevant variables, which significantly reduces computation as shown in Sec. VI-F. Figure 3: Different CFs derived in this paper, where the first three blocks describe exact fusion and the dashed blocks approximate, with local relevant states $s=[s_{i}^{T},s_{j}^{T}]^{T}$ and common states of interest $x$ for fusion shown. ### III-A Decentralized Fusion and The Homogeneous CF Let the full random state vector $\chi$ be defined as $\chi=\begin{bmatrix}X\\\ S\\\ \end{bmatrix}\ \ \ \in\mathbb{R}^{(n_{X}+n_{S})\times 1},$ (2) where $X$ and $S$ are vectors with $n_{X}$ and $n_{S}$ elements, respectively. In the target tracking example (Fig. 2), $X=[x_{1}^{T},x_{2}^{T},...,x_{n_{t}}^{T}]^{T}$ represent the random state vectors of $n_{t}=6$ targets and $S=[s_{1}^{T},s_{2}^{T},...,s_{n_{a}}]^{T}$ represent the unknown random bias state vectors of the $n_{a}=5$ tracking agents. For now it is assumed that the random state vector $\chi$ is static and the time index notation is dropped (dynamic states will be revisited and considered later). The state vector $\chi$ is a multivariate random variable and can be described by the joint pdf, $p(\chi)=p(X,S)$. The goal is to find the fused estimated underlying distribution of the state $\chi$ given the joint data $Z^{+}_{k}=Z^{-}_{i,k}\cup Z^{-}_{j,k}$ at agents $i$ and $j$. Using a distributed variant of Bayes’ rule, [8] shows that the exact posterior pdf conditioned on joint data of $i$ and $j$ is given by: $p_{f}(\chi|Z_{k}^{+})=\frac{1}{C}\cdot\frac{p_{i}(\chi|Z_{i,k}^{-})p_{j}(\chi|Z_{j,k}^{-})}{p_{c}(\chi|Z_{i,k}^{-}\cap Z_{j,k}^{-})},$ (3) where $C$ is a normalizing constant and $p_{c}$ is the posterior pdf conditioned on the common data shared by agents $i$ and $j$ prior to the current fusion. In [7], Grime and Durrant-Whyte suggest that each agent add a filter on the communication channel (hence channel filter) with a neighboring agent in a tree network. The CF explicitly tracks $p_{c}$ between the two agents, thus allowing exact removal of the common data to avoid double counting. While this operation leads to exact fusion, i.e., equal to a centralized fusion center (assuming full rate communication), the original CF formulation assumes that both agents hold distributions over the full random state vector $\chi$. Thus communication and local processing (inference) is with respect to (w.r.t.) the full joint pdf $p(\chi)$. This approach is not scalable for large state- spaces, which motivates the extension of the original CF into a family of heterogeneous fusion methods that ‘break’ the joint pdf into smaller parts. Intuitively, if agent $i$ only ‘cares’ about a subset $(x,s_{i})$ and $j$ only ‘cares’ about $(x,s_{j})$, we would like to enable communication of data only regarding the common target $x$ as in the case of the heterogeneous state CF (Fig. 3, right hand block). In the following, the structure of the underlying estimation problem is exploited to conditionally factorize into relevant subsets of the global random state vector, thereby enabling reduced communication costs and allowing each agent to locally hold a smaller pdf, i.e., reducing the computational cost of inference. ### III-B Exact Factorized CF From the law of total probability, the joint pdf over $\chi$ can be conditionally factorized as $p(\chi)=p(X)\cdot p(S|X)=p(S)\cdot p(X|S).$ (4) Using this factorization and taking $X=x$, (3) can be expressed as $\begin{split}&p_{f}(\chi|Z_{k}^{+})=\\\ &\frac{1}{C}\cdot\frac{p_{i}(x|Z_{i,k}^{-})p_{j}(x|Z_{j,k}^{-})}{p_{c}(x|Z_{i,k}^{-}\cap Z_{j,k}^{-})}\cdot\frac{p_{i}(S|x,Z_{i,k}^{-})p_{j}(S|x,Z_{j,k}^{-})}{p_{c}(S|x,Z_{i,k}^{-}\cap Z_{j,k}^{-})}.\end{split}$ (5) In the following sections, this factorization of the fusion equation and insights regarding conditional independence are used to derive new CF variants. These differ in their assumptions and result in different communication and local computation benefits. #### III-B1 Factorized Channel Filter (F-CF) First, a set of problems is considered where both agents take measurements to a common random state vector $x$, but only one of them, for example agent $i$, takes measurements to collect data on a local random state vector $S=s_{i}$. It is further assumed that agent $j$ does not gain any data regarding $s_{i}$ from any other communicating agent. Thus $s_{i}$ and the data in $j$ that is not available to agent $i$, noted by $Z_{j\setminus i,k}^{-}$, are conditionally independent given $x$ ($s_{i}\perp Z_{j\setminus i,k}^{-}|x$). This leads to the following important conclusion: $p_{j}(s_{i}|x,Z_{j,k}^{-})=p_{c}(s_{i}|x,Z_{i,k}^{-}\cap Z_{j,k}^{-})=p_{f}(s_{i}|x,Z_{k-1}^{+}),$ (6) which means that the data agent $j$ has regarding $s_{i}$ is equal to the common data with $i$, and is exactly the data it had available after the previous fusion step ($k-1$). Eq. (5) can therefore be written as, $\begin{split}p_{f}(\chi|Z_{k}^{+})=\frac{1}{C}\cdot\frac{p_{i}(x|Z_{i,k}^{-})p_{j}(x|Z_{j,k}^{-})}{p_{c}(x|Z_{i,k}^{-}\cap Z_{j,k}^{-})}\cdot p_{i}(s_{i}|x,Z_{i,k}^{-}).\end{split}$ (7) The fusion equation above has the following intuitive interpretation: if agent $j$ does not gain any new local data regarding state $s_{i}$ (conditioned on $x$), it should only communicate the marginal pdf $p_{j}(x|Z_{j,k}^{-})$, describing the common random state vector $x$, to agent $i$. This variant of the CF is dubbed _Factorized CF (F-CF)_ , since the distribution is factorized into two contributions: 1) marginal pdf regarding $x$, based on data gathered and shared by the two agents, and 2) conditional pdf regarding the local random state vector $s_{i}$ based only on data from agent $i$. The transition from the original CF to the factorized CF is shown in Fig. 3, and is enabled by using the conditional independence of the problem to separate common and local data. This fusion rule provides a possible $\mathbb{F}$ sought in (1): here, the communicated distributions are not over the same set of random states, as $i$ sends the joint distribution $p_{i}(\chi|Z^{-}_{i,k})$ and $j$ sends only the marginal distribution $p_{j}(x|Z^{-}_{j,k})$. Note that while the posterior fused pdf is the same as can be achieved by the homogeneous fusion rule (3), the math behind this heterogeneous fusion rule is fundamentally different. In the latter, the conditional pdf $p_{i}(s_{i}|x,Z_{i,k}^{-})$, which is kept local at agent $i$ or is simply replaced in agent $j$, treats the common random state vector $x$ as a function parameter. Then, the fused marginal is recombined with the local (or replaced) state conditional pdf in the joint distribution via the law of total probability (4) at each agent. This weights the conditional pdf differently as a function of $x$ and changes the overall joint pdf, thus implicitly/indirectly updating $s_{i}$. #### III-B2 Bi-directional Channel Filter (BDF-CF) Consider now the case where both agents have different local random state vectors $s_{i}$ and $s_{j}$ ($S=[s_{i}^{T},s_{j}^{T}]^{T}$). The global random state vector is then $\chi=[x^{T},s_{i}^{T},s_{j}^{T}]^{T}$, and $\chi\sim p(\chi)=p(x,s_{i},s_{j})$. If the local random state vectors are independent given the common random state vector ($s_{i}\perp s_{j}|x$), then the following holds: $p(\chi)=p(x)\cdot p(s_{i}|x)\cdot p(s_{j}|x).$ (8) Using the above factorization (5) can be split further, $\begin{split}&p_{f}(\chi|Z^{+})=\frac{1}{C}\cdot\frac{p_{i}(x|Z_{i}^{-})p_{j}(x|Z_{j}^{-})}{p_{c}(x|Z_{i}^{-}\cap Z_{j}^{-})}\cdot\\\ &\frac{p_{i}(s_{i}|x,Z_{i}^{-})p_{j}(s_{i}|x,Z_{j}^{-})}{p_{c}(s_{i}|x,Z_{i}^{-}\cap Z_{j}^{-})}\cdot\frac{p_{i}(s_{j}|x,Z_{i}^{-})p_{j}(s_{j}|x,Z_{j}^{-})}{p_{c}(s_{j}|x,Z_{i}^{-}\cap Z_{j}^{-})}.\end{split}$ (9) As in the F-CF, assume that $s_{i}\perp Z_{j\setminus i,k}^{-}|x$ and $s_{j}\perp Z_{i\setminus j,k}^{-}|x$, so (9) can be simplified: $\begin{split}&p_{f}(\chi|Z^{+})=\\\ &\frac{1}{C}\cdot\frac{p_{i}(x|Z_{i}^{-})p_{j}(x|Z_{j}^{-})}{p_{c}(x|Z_{i}^{-}\cap Z_{j}^{-})}\cdot p_{i}(s_{i}|x,Z_{i}^{-})\cdot p_{j}(s_{j}|x,Z_{j}^{-}).\end{split}$ (10) This CF variant is dubbed the _Bi-directional factorized CF (BDF-CF)_. The symmetry between the two agents can be seen: the agents share marginal pdfs regarding the random state vector $x$, and then each agent shares its unique conditional pdf regarding the local states $s_{i}$ or $s_{j}$. As in the F-CF, this gives another fusion rule $\mathbb{F}$, only now $i$ sends its marginal pdf $p_{i}(x,s_{i}|Z_{i}^{-})=p_{i}(x|Z_{i}^{-})\cdot p_{i}(s_{i}|x,Z_{i}^{-})$ and $j$ sends its marginal pdf $p_{j}(x,s_{j}|Z_{j}^{-})$. The posterior fused pdfs are equal, $p_{i,f}(x,s_{i},s_{j}|Z^{+})=p_{j,f}(x,s_{i},s_{j}|Z^{+})$. Both the F-CF and the BDF-CF are a mathematically equivalent versions of the original CF (for static systems). Their advantage though, is the considerable reduction in communication requirements achieved by sending only new and relevant data (given that the above assumptions are met). While the BDF-CF (and F-CF) has the potential of saving communication costs, as shown later in section VI-F, it still requires that every agent holds a local pdf over the full global random state vector and communicates the conditional pdfs regarding its local random states. By sacrificing part of the ‘exactness’ of the CF over less relevant random states, significant reductions in both computation and communication requirements can be gained. This leads to a new family of approximate CF methods. ### III-C Approximate Factorized CF As noted above, mathematical equivalence to the original CF is achieved under the assumption that agents do not locally gain data regarding each other’s local states. This might still require considerable communication volume, for example, when $s_{i}$ are agent $i$’s local states for a navigation filter, which typically has 16 or more states [25]. In the following approximate heterogeneous fusion algorithms, computation is further reduced by only sending the marginal pdfs regarding common random states. #### III-C1 Approximate BDF-CF Consider the case where agents prioritize their random states of interest (e.g., $[x^{T},s_{i}^{T}]$ for agent $i$) over random states local to other agents (e.g. $s_{j}$), i.e., it is more important that their local pdf is accurate in portions of the states of interest than in portions relevant to other agents. The approximate fusion rule at $i$ can then be written as $\begin{split}p_{i,f}&(\chi|Z_{i}^{+})=\\\ &\frac{1}{C}\cdot\frac{p_{i}(x|Z_{i}^{-})p_{j}(x|Z_{j}^{-})}{p_{c}(x|Z_{i}^{-}\cap Z_{j}^{-})}\cdot p_{i}(s_{i}|x,Z_{i}^{-})\cdot p_{i}(s_{j}|x).\end{split}$ (11) A similar expression can be written for agent $j$ by switching $i$ with $j$ for the conditional pdfs over $s_{i}$ and $s_{j}$. The terms $p_{i}(s_{i}|x,Z_{i}^{-})$ and $p_{i}(s_{j}|x)$ represent the local conditional distributions agent $i$ holds regarding the states not mutually monitored by $i$ and $j$. Notice that conditioned on $x$, the data local to agent $i$ is independent from $s_{j}$, i.e., agent $i$ doesn’t gain any data directly influencing agent $j$’s local random state vector $s_{j}$. This version of the heterogeneous CF is named the _Approximate BDF-CF_ , as it approximates the posterior fused pdf that can be achieved by receiving the conditional pdf $p_{j}(s_{j}|x,Z_{j}^{-})$ from agent $j$, as done in the BDF- CF (10). The function formulation for $\mathbb{F}$ still holds, since this fusion rule leads to different fused posterior results for agents $i$ and $j$. Here the fundamental mathematical difference from homogeneous fusion is highlighted again. In this heterogeneous fusion rule, fusion is over the common random state vector alone, thus the local pdf is directly updated only by the marginal pdf and this marginal fused pdf is the same at both agents. However, when agent $i$ ($j$) merges back the updated marginal pdf into its local joint pdf via the law of total probability, the conditional probabilities over $s_{i}$ and $s_{j}$, which depend on local and non-equal sets of data ($Z_{i}^{-}\neq Z_{j}^{-}$), will scale each of agent’s joint pdf differently, leading to non-equal joint pdfs at agents $i$ and $j$. #### III-C2 Heterogeneous State Channel Filter (HS-CF) So far it has been assumed that all agents across the network hold a local pdf over the same global random state vector $\chi$, which includes all locally relevant random states. Thus, as the number of tasks in the network increases so does the local computation load. However, if each agent to only hold a pdf over its locally relevant subsets of states, the local computation then scales with the agent tasks and not the global network tasks (or number of agents). This motivates the last CF variant, the _Heterogeneous state CF (HS-CF)_. Here, each agent holds its own pdf over heterogeneous subsets of random states, e.g., $\chi_{i}=\begin{bmatrix}x^{T},s_{i}^{T}\end{bmatrix}^{T}$ and $\chi_{j}=\begin{bmatrix}x^{T},s_{j}^{T}\end{bmatrix}^{T}$ for agents $i$ and $j$, respectively. The fusion rule for each agent, over their locally relevant random states, can be written as $\begin{split}&p_{i,f}(\chi_{i}|Z_{i}^{+})=\frac{1}{C_{i}}\cdot\frac{p_{i}(x|Z_{i}^{-})p_{j}(x|Z_{j}^{-})}{p_{c}(x|Z_{i}^{-}\cap Z_{j}^{-})}\cdot p_{i}(s_{i}|x,Z_{i}^{-})\\\ &p_{j,f}(\chi_{j}|Z_{j}^{+})=\frac{1}{C_{j}}\cdot\frac{p_{i}(x|Z_{i}^{-})p_{j}(x|Z_{j}^{-})}{p_{c}(x|Z_{i}^{-}\cap Z_{j}^{-})}\cdot p_{j}(s_{j}|x,Z_{j}^{-}).\end{split}$ (12) HS-CF fusion gives another fusion rule $\mathbb{F}$ for the problem statement in (1), where agents fuse marginal pdfs regarding the common random state vector $x$, and then update the joint pdf by merging the fused marginal with the local relevant conditional pdf, as in the Approximate BDF-CF. Notice that while here, as opposed to all previous heterogeneous fusion rules developed above, the two pdfs are over different sets of random states, however, the marginal pdfs over the common random state vector $x$ held by both agents will still be equal, i.e, $p_{i,f}(x)=p_{j,f}(x)$. ### III-D Log-Likelihood Representation In multi-agent data fusion problems it is common to use the log of the pdf instead the pdf itself. In addition to the connection that these log- likelihoods have with formal definitions of quantities of information (e.g., Shanon information, Fisher information), their advantages for data fusion is twofold: first, fusion is done by summation and subtraction instead of multiplication and division. Second, when possible, it allows for the exposure of the sufficient statistics, describing the pdf, e.g., the mean and covariance for Gaussian distributions, as detailed in Sec. VI. In the following, the fusion equations for the BDF-CF and HS-CF are expressed in the log-likelihood representation, where the rest of the fusion rules can be similarly translated and are omitted for brevity. Taking natural logarithm of equation (10) and combining local log-likelihoods via the law of total probability, the BDF-CF can be written as $\begin{split}\log[p_{f}(\chi|Z^{+})]&=\log[p_{i}(x,s_{i}|Z^{-}_{i})]+\log[p_{j}(x,s_{j}|Z^{-}_{j})]-\\\ &\log[p_{c}(x|Z^{-}_{i}\cap Z_{j}^{-})]+\tilde{C}.\end{split}$ (13) This representation demonstrates the intuitive interpretation of the BDF-CF fusion rule. Here the fused posterior data $Z^{+}$ is built from the summation of the contributions of local data available at each agent as it affects the relevant subset of states, minus the contribution of the data they have in common, where the latter only directly contributes to the marginal of the common target $x$. Similarly, the HS-CF fusion rule (12) representation in log-likelihood is given by, $\begin{split}\log[p_{i,f}&(\chi_{i}|Z^{+}_{i})]=\log[p_{i}(x,s_{i}|Z^{-}_{i})]+\\\ &\log[p_{j}(x|Z^{-}_{j})]-\log[p_{c}(x|Z^{-}_{i}\cap Z_{j}^{-})]+\tilde{C}_{i}\\\ \log[p_{j,f}&(\chi_{j}|Z^{+}_{j})]=\log[p_{j}(x,s_{j}|Z^{-}_{j})]+\\\ &\log[p_{i}(x|Z^{-}_{i})]-\log[p_{c}(x|Z^{-}_{i}\cap Z_{j}^{-})]+\tilde{C}_{j}.\\\ \end{split}$ (14) In the rest of this paper, this representation is used to discuss the fusion algorithm for general pdfs (Sec. V) and to derive a closed form fusion rule, based on the sufficient statistics, for the special case where the underlying distributions are Gaussian (Sec. VI). ## IV Conditional Independence in Dynamic Systems For a dynamic or partially-dynamic Bayesian network, as in Fig. 1(b), it is generally not possible to claim conditional independence of local states (or local data) based on the filtered dynamic state, $s_{i}\not\perp s_{j}|x_{k}$, i.e., $s_{i}$ and $s_{j}$ are not conditionally independent given $x_{k}$ when it is successively marginalized over time. Since this conditional independence is required to allow heterogeneous fusion, there is a need to regain conditional independence in dynamic stochastic systems. There are two approaches to solve this problem: (i) by keeping a distribution over the full time history $p(x_{k:0},s_{i},s_{j}|Z_{k})$, where $x_{k:0}$ denotes all common dynamic states from $k=0$ until current time step $k$; (ii) by enforcing conditional independence after marginalization by disconnecting the dependencies between the relevant random states. In the following, these solutions for the case of general pdfs are discussed. Then, Sec. VI derives specific closed-form representations for Gaussian distributions. ### IV-A Augmented State The distribution over the full augmented state $\chi_{k:0}=[x_{k:0}^{T},s_{i}^{T},s_{j}^{T}]^{T}$, given the data $Z_{k}^{-}$ can be recursively updated using the following formula [26]: $p(\chi_{k:0}|Z_{k}^{-})=\frac{1}{C}\cdot p(\chi_{k-1:0}|Z_{k-1}^{+})p(x_{k}|x_{k-1})p(Y_{k}|\chi_{k}),$ (15) where here $Z_{k-1}^{+}$ is used to indicate an agent’s data after the previous fusion step, $k-1$, $Y_{k}$ for the local sensor data gained at the current time step $k$, and $Z_{k}^{-}$ is the data at time step $k$ prior to fusion. The augmented state approach leads to increase in the communication and computation requirements as the size of the state vector $\chi_{k:0}$ increases. However, agents need to only send messages on the time window the size of the network (as information propagates in the tree), which bounds the communication requirement. Furthermore, due to the Markovian property of the dynamic system, algorithms for efficient inference that factorize the distribution by taking advantage of its structure can be designed. For example, for Gaussian distributions, the information matrix structure is close to block-diagonal (see Sec. VI-C), i.e., elimination can be done efficiently. ### IV-B Conservative Filtering Full knowledge over past system states has been assumed thus far, which enables conditional independence between two agents local states and the derivation of a family of heterogeneous fusion algorithms named CF2. However, in many distributed fusion applications it is desirable to maintain only a limited time window of recent state history. Thus, marginalizing out past states into a smaller sliding window of recent time steps might be favored, as maintaining the full accumulated state densities results in rapid state dimension growth and yields computation and communication burden. While marginalization is rather trivial for homogeneous fusion problems, in heterogeneous fusion extra care must be taken to maintain conditional independence. Without loss of generality, for the rest of the paper, a small sliding window of only the current time step (as done in Kalman-filter (KF) for example) will be used. The main assumption is that local bias states $s_{i},s_{j}$ are conditionally independent given the target states $x_{k:1}$. Since recursive marginalization of the past target state $x_{k-1}$ results a coupling between $s_{i}$ and $s_{j}$ (Fig. 1(b)), it is desirable to enforce conditional independence after marginalization. A similar principle is also known in the graphSLAM literature as conservative sparsification [27, 28], where Gaussian distributions are discussed. Here, the problem for general distributions is first described in a more formal way, with consideration of an open research question regarding the term ‘conservative’ for general pdfs. Then, Sec. (VI-D) focuses on Gaussian distributions and details the solution for conservative sparse marginalization to enable conservative filtering. Given a joint distribution $p(x_{1},x_{2},s_{i},s_{j})$, described by the PGM of Fig. 1(b), marginalizing out $x_{1}$, as done in filtering, results in coupling of all the variables in its Markov blanket, $x_{2},s_{i}$ and $s_{j}$. Since conditional independence between $s_{i}$ and $s_{j}$ is a fundamental assumption in the basis of the proposed methods, it is necessary to retain it after marginalization. Thus, the goal is to approximate the dense distribution $p(x_{2},s_{i},s_{j})$ by a sparse distribution such that $\tilde{p}(x_{2},s_{i},s_{j})=\frac{1}{C}\cdot p(x_{2})p(s_{i}|x_{2})p(s_{j}|x_{2}).$ (16) For the pdf to be consistent, the approximation $\tilde{p}(x_{2},s_{i},s_{j})$ has to be conservative w.r.t. $p(x_{2},s_{i},s_{j})$, which, loosely speaking, means the approximate distribution $\tilde{p}$ overestimates the uncertainty of the true distribution $p$. See Sec. II-B for a more detailed discussion and definitions of consistency and conservativeness as treated in this paper. ## V Fusion Algorithm Decentralized fusion algorithms, in general, are built out of two main steps: sending out a message and fusion of incoming message. In the original (homogeneous) channel filter (CF) algorithm [7], a message contains the sufficient statistics of the Gaussian distribution, and fusion is done simply by adding the received information and subtracting the common information (here ‘information’ is used to mean information vector and information matrix), where both actions are over the same full state vector $\chi$. In the heterogeneous CF2, on the other hand, either the communicated or the local distributions are over different random state vectors. Thus there is a need to clarify how to locally construct and fuse messages to or from neighboring agents, respectively. First, different sets of random states are defined, allowing the separation of the full global random state vector $\chi$ into smaller subsets as can be seen in Fig. 4. Recall that $\chi_{i}$ ($\chi_{j}$) was previously defined as the subset of locally relevant random states at agent _i_ (_j_), and let $\chi_{c}^{ij}=\chi_{c}^{ji}$ be the set of common random states between agents _i_ and _j_. Now, in some cases, as in the BDF-CF, agents might pass through pdfs regarding random states in their random state vector, where the data corresponding to that those random states were not gained by a local observation, but from communication with a different neighboring agent. For example, in the target tracking example of Fig. 2, agent 2 will passes a pdf over target 1, $p(x_{1}|Z_{1})$, based on data collected by agent 1 and communicated to agent 2 in previous fusion step, rather by taking a direct local sensor measurement to the target. These sets of states are dubbed ‘passed through’ states and defined as $\chi_{\neg i}^{ij}$ for states that are to be passed from _i_ to _j_ but are not local to _i_ and similarly $\chi_{\neg j}^{ji}$. With these definitions the following holds, $\chi=\chi_{i}+\chi_{j}-\chi_{c}^{ij}+\chi_{\neg i}^{ij}+\chi_{\neg j}^{ji}.$ It is now possible to discuss the content of messages and the expression for local fusion for the BDF-CF and the HS-CF. Figure 4: Diagram presenting the division of the full random state vector $\chi$ into smaller subsets. In the BDF-CF, an agent _i_ holds a posterior distribution over the full global random state vector $\chi$. Assuming tree communication topology for the network of agents, agent $i$ needs to communicate to its neighboring agent _j_ a distribution over the set of local states $\chi_{i}$ and the “pass through” set $\chi_{\neg i}^{ij}$. On the other hand, in the HS-CF, agent _i_ ’s local distribution is only over the set of local relevant states $\chi_{i}$. Agent _i_ thus sends agent _j_ the following distributions, $\begin{split}&\text{BDF-CF:}\ \ p_{i}^{ij}(\chi_{i}\cup\chi_{\neg i}^{ij})\ \ \ \forall j\in N^{i}_{a}\\\ &\text{HS-CF:}\ \ p_{i}^{ij}(\chi_{c}^{ij})\ \ \ \ \ \ \ \ \ \ \ \forall j\in N^{i}_{a},\end{split}$ (17) where $N^{i}_{a}$ is the set of all agents in _i_ ’s neighborhood. The local fusion equations requires summing (and subtracting) log-likelihoods over different sets of random states, depending on the sending agents, as each pair of agents, in general, have different sets of random states in common, as can be seen bellow: $\begin{split}\text{BDF-CF:}\ \ \ \\\ \log[p_{i,f}(\chi)]&=\log[p_{i}(\chi_{i})]+\\\ &\sum_{j\in N^{i}_{a}}\log[p_{j}^{ji}(\chi_{j}\cup\chi_{\neg j}^{ji})]-\log[p_{c}^{ji}(\chi_{c}^{ji})]\\\ \text{HS-CF:}\ \ \ \ \\\ \log[p_{i,f}(\chi_{i}&)]=\log[p_{i}(\chi_{i})]+\\\ &\sum_{j\in N^{i}_{a}}\log[p_{j}^{ji}(\chi_{c}^{ji})]-\log[p_{c}^{ji}(\chi_{c}^{ji})].\end{split}$ (18) Notice that for a two agent, one common target problem, with the definitions: $\begin{split}&\chi=[x^{T},s_{i}^{T},s_{j}^{T}]^{T},\ \ \ \chi_{i}=[x^{T},s_{i}^{T}]^{T},\\\ &\chi_{j}=[x^{T},s_{j}^{T}]^{T},\ \ \ \ \ \ \chi_{c}^{ji}=x,\ \ \ \chi_{\neg j}^{ji}=\varnothing\end{split}$ using the equations (17) and (18) result in the BDF-CF and HS-CF equations given in (13) and (14), respectively. An example of the algorithm for a linear Gaussian system, where agents communicate and fuse sufficient statistics (information vector and information matrix) is given in the next section. ## VI Gaussians - A Closed Form Fusion Rule The goal here is to derive a closed-form fusion rule for the special case of linear-Gaussian distributions to demonstrate how it works in practice and get insight into the structure of the heterogeneous fusion problem. The information (canonical) form of the Gaussian distribution will be used to this end, as it is particularly convenient for deriving and describing key steps in data fusion processing, for example in the information filter, the original CF [7], the CI algorithm [9], and more. The use of the Gaussian information form has two main advantages: (i) multiplication and division over pdfs become summation and subtraction of the sufficient statistics (information vector and information matrix) obtained from the log pdfs; and (ii) it gives insight into the conditional independence structure of the problem, since zero off-diagonal terms in the resulting information matrices indicate conditional independence between corresponding random states, as discussed further and utilized for conservative filtering in Sec. VI-D. ### VI-A Preliminaries Assume the full joint distribution over the random state vector $\chi$, defined in (2), is a multivariate Gaussian with mean $\mu$ and covariance matrix $\Sigma$, $\mu=\begin{pmatrix}\mu_{X}\\\ \mu_{S}\\\ \end{pmatrix},\ \ \ \ \ \ \ \Sigma=\begin{pmatrix}\Sigma_{XX}&\Sigma_{XS}\\\ \Sigma_{SX}&\Sigma_{SS}\end{pmatrix}\\\ $ (19) where $X$ and $S$ denote two correlated subsets of the joint random state $\chi$. The information form of $\chi$ is given by $\zeta=\Sigma^{-1}\mu=\begin{pmatrix}\zeta_{X}\\\ \zeta_{S}\\\ \end{pmatrix},\ \ \ \ \ \ \ \Lambda=\Sigma^{-1}=\begin{pmatrix}\Lambda_{XX}&\Lambda_{XS}\\\ \Lambda_{SX}&\Lambda_{SS}\end{pmatrix}\\\ $ (20) The pdf in information form for the normally distributed state $\chi$, with information vector $\zeta$ and information matrix $\Lambda$ is [29]: $p(\chi;\zeta,\Lambda)=\frac{\exp(-\frac{1}{2}\zeta^{T}\Lambda^{-1}\zeta)}{\det(2\pi\Lambda^{-1})^{\frac{1}{2}}}\exp\big{(}{-\frac{1}{2}\chi^{T}\Lambda\chi+\zeta^{T}\chi\big{)}}.$ (21) This pdf can also be expressed using factorization (4), where the marginal and conditional distributions of a Gaussian are also Gaussian, $\begin{split}p(X)&=\mathcal{N}^{-1}(X;\bar{\zeta}_{X},\bar{\Lambda}_{XX})\\\ p(S|X)&=\mathcal{N}^{-1}(S;\zeta_{S|X},\Lambda_{S|X})\end{split}$ (22) where $\mathcal{N}^{-1}$ represents the information form of the Gaussian distribution $\mathcal{N}$, and $(\bar{\zeta}_{X},\bar{\Lambda}_{XX})$ and $(\zeta_{S|X},\Lambda_{S|X})$ are the sufficient statistics for the marginal and conditional pdfs in information form, respectively, defined as [30]: $\begin{split}\bar{\zeta}_{X}=\zeta_{X}-\Lambda_{XS}\Lambda^{-1}_{SS}\zeta_{S}&,\ \ \ \ \bar{\Lambda}_{XX}=\Lambda_{XX}-\Lambda_{XS}\Lambda^{-1}_{SS}\Lambda_{SX}\\\ \zeta_{S|X}=\zeta_{S}-\Lambda_{SX}X&,\ \ \ \ \Lambda_{S|X}=\Lambda_{SS}\end{split}$ (23) ### VI-B Fusion Starting with the original homogeneous fusion rule (3), by substituting linear Gaussian distributions, taking logs and differentiating once for the fused information vector ($\zeta^{f}=(\Sigma^{f})^{-1}\mu^{f}$) and twice for the fused information matrix ($\Lambda^{f}=(\Sigma^{f})^{-1}$), [8] to obtain the fusion equations, $\begin{split}\zeta^{f}=\zeta^{i}+\zeta^{j}-\zeta^{c}\ ,\ \ \ \ \ \Lambda^{f}=\Lambda^{i}+\Lambda^{j}-\Lambda^{c}.\end{split}$ (24) These equations are the basis of the original linear-Gaussian CF [7], which explicitly tracks the ‘common information’ vector and matrix $(\zeta_{c},\Lambda_{c})$, describing the pdf over $\chi$ conditioned on the common data between communicating pairs of agents $i$ and $j$ ($Z^{-}_{i}\cap Z^{-}_{j}$) in tree-structured communication networks. Define $\bar{\zeta}^{f}_{X}$ and $\bar{\Lambda}^{f}_{XX}$ to be the _fused_ marginal information vector and matrix, respectively, over the common target random state $X$ corresponding to $p_{f}(X)$, represented in information form. Without loss of generality, the fused marginal information vector and matrix can be achieved by using different fusion methods, exact and approximate (CI [9] for example). This paper restricts attention to the CF for exact fusion. Then using (24), the fused marginal information vector and matrix for Gaussian distributions is given by $\begin{split}\bar{\zeta}^{f}_{X}=\bar{\zeta}^{i}_{X}+\bar{\zeta}^{j}_{X}-\bar{\zeta}^{c}_{X},\ \ \ \bar{\Lambda}^{f}_{XX}=\bar{\Lambda}^{i}_{XX}+\bar{\Lambda}^{j}_{XX}-\bar{\Lambda}^{c}_{XX}\end{split}$ (25) The HS-CF fusion rule, eq. (14), for Gaussian pdfs can be represented by the simple closed form expression, $\begin{split}\zeta^{i,f}=&\left(\begin{array}[]{c}\bar{\zeta}^{f}_{X}\\\ \hdashline[2pt/2pt]0\end{array}\right)+\left(\begin{array}[]{c}\Lambda^{i}_{xs_{i}}(\Lambda^{i}_{s_{i}s_{i}})^{-1}\zeta^{i}_{s_{i}}\\\ \hdashline[2pt/2pt]\zeta^{i}_{s_{i}}\end{array}\right)\\\ \Lambda^{i,f}=&\left(\begin{array}[]{c;{2pt/2pt}c}\bar{\Lambda}^{f}_{xx}&0\\\ \hdashline[2pt/2pt]0&0\end{array}\right)+\left(\begin{array}[]{c;{2pt/2pt}c}\Lambda^{i}_{xs_{i}}(\Lambda^{i}_{s_{i}s_{i}})^{-1}\Lambda^{i}_{s_{i}x}&\Lambda^i_{xs_i}\\\ \hdashline[2pt/2pt]\Lambda^{i}_{s_{i}x}&\Lambda^i_{s_is_i}\end{array}\right),\end{split}$ (26) where $\bar{\zeta}^{f}_{x}$ and $\bar{\Lambda}^{f}_{xx}$ are given in (25) and it is assumed $X=x$ and $S=[s_{i}^{T},s_{j}^{T}]^{T}$. An equivalent expression for the fused information vector and matrix at agent $j$ is achieved by switching $i$ with $j$. It is important to note, as seen from (25), that the fused marginal pdf $(\bar{\zeta}^{f}_{x},\bar{\Lambda}^{f}_{xx})$ is the same for agents $i$ and $j$. However, the conditional part is kept local (the right part of (26)), which means that after fusion, the _local joint_ distributions in $i$ (w.r.t. $x$ and $s_{i}$) and $j$ (w.r.t. $x$ and $s_{j}$) are different. While agents only update the information vector and matrix of the marginal pdf, over the common random state $x$, the local joint distribution in moment representation (e.g., $\mu^{i,f},\Sigma^{i,f}$) will be updated, thus also updating the local states $s_{i}$ ($s_{j}$). ### VI-C The Information Augmented State For Gaussian distributions there are two similar but not equivalent filters: the _augmented_ state (AS), [12], and the _accumulated_ state density (ASD), [26], where they mostly differ in their retrodiction formulation. This is more involved in the ASD and not imperative to our solution, so the attention here is limited to the augmented state implementation [12]. Here the emphasis is on Gaussian state distributions; a solution of the ASD for general distributions is given in [26]. The augmented state for a sliding time window from time step $n$ to time step $k$ (denoted by subscript $k:n$) given in [12] uses covariance formulation for the prediction step, and information representation for the update step. However, since the algorithms developed in this paper work in Gaussian information space, it is advantageous to work with a full information filter formulation. The derivation of the information AS (_iAS_) is given in Appendix A, with the main results provided here, namely the prediction and update steps in information form. For a dynamic system, described by the discrete time equations $\begin{split}&x_{k}=F_{k}x_{k-1}+Gu_{k}+\omega_{k},\ \ \ \ \omega_{k}\sim\mathcal{N}(0,Q_{k})\\\ &y_{k}=H_{k}x_{k}+v_{k},\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ v_{k}\sim\mathcal{N}(0,R_{k}),\end{split}$ (27) where $F_{k}$ is the state transition matrix, $G$ is the control effect matrix and $H_{k}$ is the sensing matrix. $\omega_{k}$ and $v_{k}$ are the zero mean white Gaussian process and measurement noise, respectively. The predicted information vector and matrix for the time window $k:n$, given all information up to and including time step $k-1$ are given by $\begin{split}&\zeta_{k:n|k-1}=\begin{pmatrix}Q_{k}^{-1}Gu_{k}\\\ \zeta_{k-1:n|k-1}-\mathbf{F}^{T}Q_{k}^{-1}Gu_{k}\end{pmatrix}\\\ &\Lambda_{k:n|k-1}=\begin{pmatrix}Q_{k}^{-1}&-Q_{k}^{-1}\mathbf{F}\\\ -\mathbf{F}^{T}Q_{k}^{-1}&\Lambda_{k-1:n|k-1}+\mathbf{F}^{T}Q_{k}^{-1}\mathbf{F}\end{pmatrix},\\\ \end{split}$ (28) where $\mathbf{F}=\big{[}F_{k-1}\ \ 0_{m\times m(k-n-2)}\big{]}$ and $m$ is the size of the (not augmented) state vector. Notice the simplicity of the above expression and its interesting interpretation: the predicted conditional information matrix at the current time step, given previous time steps (upper left block), depends inversely on process noise alone and is not affected by the state dynamics. For completeness, the measurement update in Gaussian information space is [12] $\begin{split}&\zeta_{k:n|k}=\zeta_{k:n|k-1}+J_{k}i_{k}\\\ &\Lambda_{k:n|k}=\Lambda_{k:n|k-1}+J_{k}I_{k}J_{k}^{T},\end{split}$ (29) where $J_{k}=\big{[}I_{m}\ \ 0_{m\times m(k-n-1)}\big{]}^{T}$, $i_{k}=H_{k}^{T}R_{k}^{-1}y_{k}$ and $I_{k}=H_{k}^{T}R_{k}^{-1}H_{k}$. For linear-Gaussian problems, the above _iAS_ can be used locally at each agent, enabling conditional independence such that for $n=1$, $\begin{split}p(x_{k:1},s_{i},s_{j})&=\\\ p(&x_{k:1}|Z_{k})\cdot p(s_{i}|x_{k:1},Z_{k})\cdot p(s_{j}|x_{k:1},Z_{k}).\end{split}$ (30) It can be seen that the full information matrix and vector given in eq. (28) grows rapidly with the time step $k$, inducing high processing and communication costs. Note that the block tri-diagonal structure of the updated information matrix $\Lambda_{k:1|k-1}$, resulting from the Markov property of the system dynamics can be utilized to reduce computation burden. But this does not resolve the communication load problem which scales with the size of the tree network, allowing messages to propagate through from one end of the tree to the other. Instead, a filtering approach is taken to marginalize past states to process only a sliding window $k:n$ ($n>1$) while maintaining conditional independence. This requires conservative filtering, discussed in the next section. ### VI-D Conservative Filtering The approach of Vial _et al._[27] is adopted for the special case of Gaussian distributions. The sparse structure of the marginalized approximate information matrix is enforced by removing the links between $s_{i}$ and $s_{j}$. In other words, given a true dense Gaussian distribution $\mathcal{N}(\zeta_{tr},\Lambda_{tr})$, a sparse approximate distribution $\mathcal{N}(\zeta_{sp},\Lambda_{sp})$ is sought such that the mean is unchanged and the approximation is conservative in the PSD sense, $\begin{split}\Lambda_{tr}^{-1}\zeta_{tr}=\Lambda_{sp}^{-1}\zeta_{sp},\ \ \ \ \ \Lambda_{tr}-\Lambda_{sp}\succeq 0,\end{split}$ (31) where again the information form of the Gaussian distribution is used. Reference [27] minimizes the Kullback-Leibler Divergence (KLD) to find a lower bound on the dense true information matrix $\Lambda_{tr}$. Along similar lines, [31] suggests a method named ‘Uniform Pre-Transmission Eigenvalue-Based Scaling’ to conservatively approximate a covariance matrix $\Sigma$ by inflating a diagonal matrix $D$ built out of the diagonal entries of the full matrix $\Sigma$. To achieve a conservative approximation $D_{c}$, $D$ is inflated by multiplying by the largest eigenvalue of $Q=D^{-\frac{1}{2}}\Sigma D^{-\frac{1}{2}}$. This results in $D_{c}=\lambda_{max}D$ such that $D_{c}-\Sigma\succeq 0$. This method is generalized here to find a lower bound sparse information matrix $\Lambda_{sp}$ and regain conditional independence between $s_{i}$ and $s_{j}$. This new generalized method differs from the one suggested in [31] in two ways. Firstly, the approximation, $\Lambda_{sp}$, is allowed to be any information matrix achieved by setting any off-diagonal elements of the true dense information matrix $\Lambda_{tr}$ to zero, i.e., the resulting matrix is in general not diagonal or even block-diagonal. Note that since the information matrix (i.e. not the covariance) is changed, setting off-diagonal elements to zero directly controls the conditional independence structure of the underlying distribution. Specifically for the purpose of this paper, terms relating local random states (e.g., $s_{i}$ and $s_{j}$) in $\Lambda_{tr}$ are set to zero to regain conditional independence given common target states (e.g., $x_{k:n}$). The second change from the original method is that the information matrix is approximated, as opposed to the covariance matrix. This means that a _lower_ bound is sought and not an _upper_ bound, i.e. ‘information’ must be deflated, instead of uncertainty being inflated. This is achieved by choosing the _minimal_ eigenvalue of $\tilde{Q}=\Lambda_{sp}^{-\frac{1}{2}}\Lambda_{tr}\Lambda_{sp}^{-\frac{1}{2}}$, resulting in $\Lambda_{tr}-\lambda_{min}\Lambda_{sp}\succeq 0,$ (32) where $\lambda_{min}\Lambda_{sp}$ is the sought of conservative marginal sparse approximation of the dense information matrix $\Lambda_{tr}$. The new information vector is computed to such that (31) holds $\zeta_{sp}=(\lambda_{min}\Lambda_{sp})\Lambda_{tr}^{-1}\zeta_{tr}.$ (33) ### VI-E Closed-Form Algorithm This subsection provides a full summary description of the different steps each agent $i$ takes to locally process, communicate and fuse data with its neighbor $j$. While the steps described in the pseudo code in Algorithm 1 are general in the sense that it can be applied with any pdf, it assumes linear- Gaussian distributions and uses the theory developed above to provide equation references (green arrows) for the closed-form expressions of the different operations. Following Sec. V, the BDF-CF and HS-CF algorithms are used as an example to detail the algorithm from the perspective of one agent $i$ communicating with a neighbor $j$. Algorithm 1 BDF-CF / HS-CF algorithm 1:Define: $\chi_{i}$, $\chi_{j}$, $\chi_{c}^{ij}$ $\chi_{\neg i}^{ij}$, $\chi_{\neg j}^{ji}$, Priors $\triangleright$ Fig. 4 2:for All time steps do 3: Propagate state local states $\triangleright$ Eq. 28 4: Propagate common states in the CF 5: if BDF-CF then 6: Conservative filtering $\triangleright$ Sec. VI-D 7: else if HS-CF then 8: Marginalize out past state 9: end if 10: Measurement update $\triangleright$ Eq. 29 11: Send message $\triangleright$ Eq. 17 12: Fuse received message $\triangleright$ Eq. 18 13: Update common information 14:end for 15:return ### VI-F Calculation of Communication and Computation Savings To highlight the potential gain of the proposed CF2 methods with respect to communication and computation complexity and how they change with scale, three numerical examples (small, medium and large) of a multi-agent multi-target tracking problem are presented. Consider the problem introduced earlier of tracking $n_{t}$ ground targets by $n_{a}$ agents (trackers), where each agent computes a local KF estimate, i.e., the system dynamics are assumed to be linear with additive Gaussian white noise (27). Each agent $i$ has 6 unknown local position states described by the random vector $s_{i}$ and takes measurements to $n_{t}^{i}$ targets, each having 4 position/velocity states described by the random vector $x_{t}$ (e.g., east and north coordinates and velocities). The full state vector then has $6n_{a}+4n_{t}$ random states. Assume tree topology in ascending order, where each agent tracks 1/2/3 targets, corresponding to the small/medium/large examples, respectively, but has only one target in common with its neighbor. Now, using the same logic as before, assume that each agent is only concerned with the targets it takes measurements to and its own position states. Each agent has only 10/14/18 local ‘relevant’ random states for tracking 1/2/3 targets, respectively. Table I presents a comparison between the different channel filters for the three different scale problems. The baseline for comparison is the original (homogeneous) CF, with each agent estimating the full state vector. For the communication data requirement, double precision (8 bytes per element) is assumed. Since the matrices are symmetric covariances, agents only needs to communicate $n(n+1)/2$ upper diagonal elements, where $n$ is the number of random states. Each agent’s computation complexity is determined by the cost of inverting an $n\times n$ matrix. It can be seen from the table that even for the small scale problem, the communication data reduction is significant; the BDF-CF requires about 42.7% of the original CF, while the approximate BDF- CF and HS-CF requires only 9.2% as agents only communicate common targets states information vectors and matrices. These gains then increase with scale, for the medium and large problems the BDF-CF communication is about 33% of the original CF and for the approximate methods is less than 1%. Another important aspect is computation complexity. As seen from the table, while the BDF-CF methods require each agent to process the full random state vector, the HS-CF offers significant computational reduction. Since each agent only processes the locally relevant random states, the HS-CF scales with subset of states and not with number of agents and targets. In the medium and large scale problems, as the size of the full system random vector states increase, the HS-CF computation complexity is less than 1% of the other methods, which can be critical in terms of computing power for resource- constrained tracking platforms. TABLE I: Data communication requirements and computational complexity for different fusion methods, for different problem scales. | | Small | Medium | Large ---|---|---|---|--- | ($n_{a}$, $n_{t}$) | (2, 1) | (10, 11) | (25, 51) | $n_{t}^{i}$ | 1 | 2 | 3 CF | Data req. $[$KB$]$ | 2.4 | 801 | 24300 Complexity | $O(16^{3})$ | $O(104^{3})$ | $O(354^{3})$ BDF-CF | Data req. $[$%CF$]$ | 42.7 | 33.7 | 33.2 Complexity | $O(16^{3})$ | $O(104^{3})$ | $O(354^{3})$ Approx. BDF-CF | Data req. $[$%CF$]$ | 9.2 | 0.25 | 0.02 Complexity | $O(16^{3})$ | $O(104^{3})$ | $O(354^{3})$ HS-CF | Data req. $[$%CF$]$ | 9.2 | 0.25 | 0.02 Complexity | $O(10^{3})$ | $O(14^{3})$ | $O(18^{3})$ ## VII Simulation Studies Multi-agent multi-target tracking simulation scenarios where performed to compare and validate the proposed algorithms, where the focus is on the BDF-CF and HS-CF heterogeneous fusion algorithms. Since the dynamics and measurement models are assumed to be linear with Gaussian noise, Algorithm 1 is used together with the _iAS_ as the inference engine, i.e., agents estimate the sufficient statistics (information vector and matrix) of the random state vector. First, the algorithms are tested on a static target case, where conditional independence of the local states can be easily guaranteed. This is followed by a dynamic target test case with only two agents and one target, to validate and compare the augmented state (smoothing) and the conservative filtering approaches. Lastly, the conservative filtering approach is used for a more interesting 4-agents 5-target scenario. Results for all the different scenarios are based on Monte Carlo simulations and compare the new algorithms to an optimal centralized estimator. ### VII-A Example 1 - Static Case A chain network, consisting of five agents connected bilaterally in ascending order $(1\leftrightarrow 2\leftrightarrow 3\leftrightarrow 4\leftrightarrow 5)$,as depicted in (Fig. 2), attempts to estimate the position of six stationary targets in a $2D$ space. Assume each tracking agent $i\ =1,...,5$ has perfect self position knowledge, but with constant agent-target relative position measurement bias vector in the east and north directions $s_{i}=[b_{e,i},b_{n,i}]^{T}$. In every time step $k$, each agent takes two kinds of measurements: one for the target and one to collect data on the local sensor bias random vector, which can be transformed into the linear pseudo- measurements, $\displaystyle\begin{split}y^{t}_{i,k}&=x^{t}+s_{i}+v^{i,1}_{k},\ \ v^{i,1}_{k}\sim\mathcal{N}(0,R^{1}_{i}),\\\ m_{i,k}&=s_{i}+v^{i,2}_{k},\ \ v^{i,2}_{k}\sim\mathcal{N}(0,R^{2}_{i}),\end{split}$ (34) where $y^{t}_{i,k}$ is agent $i$ relative measurement to target $t$ at time step $k$ and $m_{i,k}$ is a measurement to a known landmark at time step $k$ for bias estimation. $x^{t}=[e^{t},n^{t}]^{T}$ is the east and north position of target $t\ =1,...,6$. The tracking assignments for each agent, along with the measurements noise error covariances for the relative target ($R^{1}_{i}$) and landmark ($R^{2}_{i}$) measurements are given in Table II and illustrated in Fig. 2. The relative target measurement noise characteristics for different targets measured by the same agent are taken to be equal. For example, agent $1$ takes noisy measurements to targets $1$ and $2$ with $1\ m^{2}$ and $10\ m^{2}$ variances in the east and north directions, respectively, and $3\ m^{2}$ in both directions for the landmark. Following the definitions from Sec. III, the full state vector includes 22 random states $\chi=[{x^{1}}^{T},...,{x^{6}}^{T},s_{1}^{T},...,s_{5}^{T}]^{T},$ (35) where for the HS-CF fusion, define the local random state vector at agent $i$ $\chi_{i}=[{X^{\mathcal{T}_{i}}}^{T},s_{i}^{T}]^{T}.$ (36) Here $\mathcal{T}_{i}$ is the set of targets observed by agent $i$, and $X^{\mathcal{T}_{i}}$ includes all target random state vectors $x^{t}$, s.t $t\in\mathcal{T}_{i}$. In other words, the local random state vector at each agent includes only locally relevant targets and the local biases. In the HS- CF, where two agents $i$ and $j$ only share the marginal statistics regarding common states, messages should only consist data regarding targets $t\in\mathcal{T}_{i}\cap\mathcal{T}_{j}$. For example, according to Table II and the network tree topology, for agents $1$ and $2$: $\mathcal{T}_{1}\cap\mathcal{T}_{2}=T_{2}$. The data communication requirements for this relatively small system were calculated; similar to the results from Sec. VI-F, the BDF-CF and the HS-CF requires about 38% and 2.6% of the original CF communication data requirements, respectively. TABLE II: Local platform target assignments and sensor measurement error covariances. Agent | Tracked Targets | $R_{i}^{1}[m^{2}]$ | $R_{i}^{2}[m^{2}]$ ---|---|---|--- 1 | $T_{1},T_{2}$ | diag([1,10]) | diag([3,3]) 2 | $T_{2},T_{3}$ | diag([3,3]) | diag([3,3]) 3 | $T_{3},T_{4},T_{5}$ | diag([4,4]) | diag([2,2]) 4 | $T_{4},T_{5}$ | diag([10,1]) | diag([4,4]) 5 | $T_{5},T_{6}$ | diag([2,2]) | diag([5,5]) The BDF-CF and the HS-CF performance was tested with 500 Monte Carlo simulations and compared to a centralized estimator. As mentioned before, in the BDF-CF each platform processes the full random state vector (35), while in the HS-CF each platform processes only the locally relevant random states (36). In the simulations fusion occurs in every time step. Figure 5: Example 1 (static) - NEES hypothesis test based on 500 Monte Carlo simulations, where the dashed lines show bounds for 95$\%$ confidence level. Shown are test results of agents 1 and 4 using for different fusion methods, the results indicate all methods produce a consistent estimate. Fig. 5 shows a NEES consistency test [15] results for agents 1 and 4 and a centralized estimator. Results are based on 500 Monte Carlo simulations with 95% confidence level. It can be seen that the MMSE estimates of both agents, with the BDF-CF and the HS-CF are consistent. Note that since the HS-CF only estimates a subset of 6 states out of the full 22 state vector, the consistency bounds are different for these methods. The second test to determine the performance of a fusion method is whether it is conservative relative to a centralized estimator (see Sec. II-B). To verify, the local covariance matrix must be checked to see whether it is pessimistic relative to the centralized covariance. One simple test is by computing the eigenvalues of $\bar{\Sigma}_{\chi_{i}}-\bar{\Sigma}_{\chi_{i}}^{cent}$, where $\bar{\Sigma}_{\chi_{i}}$ is the agent’s covariance and $\bar{\Sigma}_{\chi_{i}}^{cent}$ is the centralized marginal covariance over $\chi_{i}$. If the minimal eigenvalue is bigger or equal to zero, all eigenvalue are bigger or equal to zero and the MMSE estimate is conservative in the PSD sense. In the above simulations the minimal eigenvalues between all agents and all simulations were 0, for both the BDF-CF and the HS-CF, thus they are conservative in the PSD sense. ### VII-B Example 2 - Dynamic Case In dynamic systems, as discussed in Sec. IV, there is a challenge in maintaining conditional independence. Two ways to solution are suggested, the first using the _iAS_ , thus keeping a distribution over the full time history over target random states, which is costly in both communication and computation requirements. The second, more efficient solution, is to perform conservative filtering by enforcing conditional independence in the marginalization step and deflating the information matrix (Algorithm 1). Since this process loses information due to deflation, the BDF-CF becomes an approximate solution and is expected to be less accurate than the _iAS_ implementation. This is shown using a two agent, one (dynamic) target tracking simulation. Here the target follows a linear dynamics model with time-varying acceleration control, $x_{k+1}=Fx_{k}+Gu_{k}+\omega_{k},\ \ \omega_{k}\sim\mathcal{N}(0,0.08\cdot I_{n_{x}\times n_{x}}),$ (37) where $\begin{split}F=\begin{bmatrix}1&\Delta t&0&0\\\ 0&1&0&0\\\ 0&0&1&\Delta t\\\ 0&0&0&1\end{bmatrix},\quad G=\begin{bmatrix}\frac{1}{2}\Delta t^{2}&0\\\ \Delta t&0\\\ 0&\frac{1}{2}\Delta t^{2}\\\ 0&\Delta t\end{bmatrix}.\end{split}$ (38) The acceleration input in the east and north directions is given by $u_{k}=\begin{bmatrix}a_{e}\cdot\cos(d_{e}\cdot k\Delta t)\ \ a_{n}\cdot\sin(d_{n}\cdot k\Delta t)\end{bmatrix}^{T}$, where $a_{e}/a_{n}$ and $d_{e}/d_{n}$ define the east and north amplitude and frequency, respectively. The measurement model is as in the static example, given in (34) with noise parameters defined for agents 1 and 2 in Table II. Results in Fig. 6 show a comparison between the _iAS_ with the full time window ($k:1$) and filtering approaches (sliding window of size 1), using the BDF-CF and the HS-CF for fusion. The plots in figure (a) show the NEES consistency tests (75 simulations, 95% confidence level) for agent 1, where the centralized (filtering results presented) and BDF-CF in the upper plot and the HS-CF with its different bounds in the lower. The results are consistent, with pessimistic behaviour of the BDF-CF due to the conservative filtering approach. (b) Shows the root mean squared error (RMSE) results of the same simulation, with agent 1 in the upper plot and agent 2 in the lower. The _iAS_ has better performance, with smaller RMSE and $2\sigma$ bounds, which is to be expected due to its smoothing operation, but at the expense of much higher computation and communication load. The conservativeness of the fused estimate was checked again by computing the minimal eigenvalues across 75 simulations and the two agents. For the _iAS_ approach, the BDF-CF and the HS-CF had small negative minimal eigenvalues of $-0.002$ and $-0.0015$ respectively, thus slightly overconfident relative to the centralized. For the conservative filtering approach, the BDF-CF was conservative with minimal eigenvalue of $0.0008$ and the HS-CF was overconfident with minimal eigenvalue of $-0.26$. Figure 6: Results from a 75 Monte Carlo simulation of a 2 agent, 1 target dynamic target tracking scenario. Shown is a comparison between _iAS_ and filtering. (a) NEES test with the upper figure showing the BDF-CF compared to the centralized (with filtering) and the lower showing the HS-CF. Here the dashed black lines show bounds for 95$\%$ confidence level. (b) RMSE comparisons between the BDF-CF and the HS-CF for _iAS_ and filtering approaches for agent 1 (upper) and agent 2 (lower). Conservative filtering allows to test the algorithms on a more interesting dynamic simulation. As a test case, a simulation of a cooperative target tracking task with 4 agents and 5 dynamic targets (full random state vector of 28 states) was performed. The dynamic model details are the same as in the 2 agent, 1 target dynamic simulation above, with measurement parameters defined by the first 4 agents in Table II. The advantages of the BDF-CF and HS-CF regarding communication and computation costs are highlighted again, as the BDF-CF saves 58% in communication costs relative to the original CF, and the HS-CF saves 94.5% in communication and 87.5% in computation complexity. Results of 500 Monte Carlo simulation with filtering for agents 1 and 4 are presented in Fig. 7. The plots in (a) show the NEES consistency test with 95% confidence bound marked with black dashed lines. The upper plot shows the centralized (black circles) and the BDF-CF for agents 1 (blue squares) and 4 (red x) NEES for the 28-state vector. The lower plot shows the HS-CF results, which has a smaller 10-state random vector. (b) shows the corresponding RMSE for agent 1 (upper plot) and 4 (lower plot). Note that the RMSE results for the centralized estimate and BDF-CF, which hold distributions over the full 28-state vector, are marginalized and computed only over relevant local 10 agent random states for this comparison. It is seen from the NEES plots that, as expected, the centralized estimator produces a consistent MMSE estimate, and the BDF-CF overestimates the uncertainty due to the information matrix deflation (covariance inflation) in the conservative filtering step. The BDF-CF also produces a conservative MMSE estimate relative to the centralized in the PSD sense for all agents, since the minimum eigenvalue between the agents is positive ($3e-04$). The HS-CF is slightly overconfident for both the consistency test and the PSD test, with negative minimal eigenvalue of $-0.26$. However, the degree of non- conservativeness in the HS-CF will in general be highly problem- and topology dependent. Hence, the choice of whether to task agents with the full random state vector, with either homogeneous DDF methods (e.g., classical CF and conventional CI) or heterogeneous fusion with the BDF-CF, or to task them with only a subset of relevant random states using the HS-CF, will hinge on the desired trade-off in communication/computation complexity vs. resulting overconfidence in state MMSE estimates, provided that the HS-CF allows for stable convergence. The HS-CF overconfidence is attributed to inaccurate removal of implicit and hidden correlations due to marginalization in the filtering step (line 8 in Algorithm 1). Correctly accounting for these dependencies is not in the scope of this paper, but is the focus of ongoing work. Figure 7: Results from a 500 Monte Carlo simulation of a cooperative target tracking task, with filtering, consisting of 4 tracking agents and 5 dynamic targets. Presented are results for agents 1 and 4. (a) NEES consistency test results for the centralized and BDF-CF estimates of 28 random states (upper) and the HS-CF estimates over 10 random states (lower). (b) Solid lines show the RMSE over target and agent local states relevant to that agent, dashed line shows the $2\sigma$ confidence bounds on logarithmic scale, for agent 1 (upper) and 4 (lower). ## VIII Conclusions Heterogeneous fusion defines a key family of problems for Bayesian DDF, as it enables flexibility for large scale autonomous sensing networks. As shown in this work, separating the global joint distribution into smaller subsets of local distributions significantly reduces local agent computation and communication requirements. The analysis and derivations presented in this paper, while assuming tree structured networks for the purposes of exact fusion via channel filtering, offers a basis for developing and analyzing similar methods for more general heterogeneous problems involving exact or approximate fusion in more complex networked fusion settings. Probabilistic graphical models (PGMs) were used here to develop Bayesian DDF algorithms. PGMs provided insight into the origin of the coupling between random states not mutually tracked by two agents and enabled exploitation of the conditional independence structure embedded in these graphs. This led to a family of conditionally factorized channel filter (CF2) approaches for general probabilistic and Gaussian pdfs that were demonstrated on static and dynamic target tracking problems. The latter motivated the development and use of the information augmented state (_iAS_) filter to regain conditional independence, on the expense of increasing computation and communication costs. To overcome this problem a conservative filtering approach was demonstrated to maintain conditional independence over a small time window, without the need of the full time history. The DDF framework naturally enables sparse distributed estimation for high- dimensional state estimation and the conditionally factorized-CF approach represents a practical and theoretical shift in the state of the art, subject to usual provisos and limitations of DDF. From a practical standpoint, the CF strategies developed in this paper can already be used to improve scalability in a variety of decentralized Bayesian estimation problems, such as cooperative localization and navigation [25, 1], multi-agent SLAM [3] and terrain height mapping [32], where a height distribution is estimated on a grid map. In this case, for example, the BDF-CF can be used to reduce communication in the network by dividing the the map into several overlapping regions of interest, allowing agents to communicate only regarding those cells in which they have new data to contribute. This scales the communication with the number of locally observed grid cells instead of the entire map. Indeed, some works are already leveraging heterogeneous DDF ideas for robotics, [1, 2], despite the gap in theoretical guarantees and understanding on the full nature of the problem and its limitations. This paper makes progress by building theoretical foundations for future research and surfacing a discussion on the assumptions and definitions of homogeneous DDF, as they appear to be inadequate for real world robotics problems of heterogeneous fusion. Heterogeneous fusion, as defined in this paper, requires a careful revisit of the idea of ‘ideal’ centralized/decentralized Bayesian estimation as well as the definitions of consistency and conservativeness for general (non-Gaussian) pdfs and more specifically in the case of heterogeneous pdfs in dynamic systems. ## Appendix A Derivation of the iAS The augmented state for a sliding time window from time step $n$ to time step $k$ (denoted by subscript $k:n$) as shown in [12] is given by the following equations: _Prediction Step_ $X_{k:n|k-1}=\begin{pmatrix}F_{k-1}\chi_{k-1|k-1}+Gu_{k}\\\ X_{k-1:n|k-1}\end{pmatrix}$ (39) $P_{k:n|k-1}=\begin{pmatrix}P_{k|k-1}&\mathbf{F}P_{k-1:n|k-1}\\\ P_{k-1:n|k-1}\mathbf{F}^{T}&P_{k-1:n|k-1}\end{pmatrix},$ (40) where $\mathbf{F}=\big{[}F_{k-1}\ \ 0_{m\times m(k-n-2)}\big{]}$ and $m$ is the size of the (not augmented) state vector. _Update Step_ The measurement update in information space is as follows: $P_{k:n|k}^{-1}=P_{k:n|k-1}^{-1}+J_{k}I_{k}J_{k}^{T}$ (41) $P_{k:n|k}^{-1}X_{k:n|k}=P_{k:n|k-1}^{-1}X_{k:n|k-1}+J_{k}i_{k},$ (42) where $J_{k}=\big{[}I_{m}\ \ 0_{m\times m(k-n-1)}\big{]}^{T}$, $i_{k}=H_{k}^{T}R_{k}^{-1}z_{k}$ and $I_{k}=H_{k}^{T}R_{k}^{-1}H_{k}$. Since the algorithms developed in this paper work in log space, it is advantageous to work with an information filter, which is based on the log- likelihood of the Gaussian distribution. Thus, a transformation of the prediction step given in (39)-(40) from state space to the Gaussian information space is needed. First, define $P_{k:n|k-1}^{-1}$ to be the augmented predicted information matrix: $\begin{split}P_{k:n|k-1}^{-1}&=\begin{pmatrix}V_{11}&V_{12}\\\ V_{21}&V_{22}\end{pmatrix},\end{split}$ (43) where from the matrix inversion lemma: $\begin{split}V_{11}&=(P_{k|k-1}-\mathbf{F}P_{k-1:n|k-1}P_{k-1:n|k-1}^{-1}P_{k-1:n|k-1}\mathbf{F}^{T})^{-1}\\\ &=(P_{k|k-1}-\mathbf{F}P_{k-1:n|k-1}\mathbf{F}^{T})^{-1}.\end{split}$ (44) The expression $\mathbf{F}P_{k-1:n|k-1}\mathbf{F}^{T}$ has the dimension $m\times m$. From the definition of $\mathbf{F}$ above (44) can be simplified by noticing that $\mathbf{F}P_{k-1:n|k-1}\mathbf{F}^{T}=F_{k-1}P_{k-1|k-1}F_{k-1}^{T}$, i.e. it depends only on the previous time step and not the full time history. Eq. (44) is thus: $V_{11}=(P_{k|k-1}-F_{k-1}P_{k-1|k-1}F_{k-1}^{T})^{-1}.$ (45) Here $P_{k|k-1}^{-1}$ is the predicted information matrix at time step $k$, given in the literature by: $P_{k|k-1}^{-1}=(F_{k-1}P_{k-1|k-1}F_{k-1}^{T}+Q)^{-1}$, where $Q$ is the process noise covariance. Taking the inverse and plugging in $P_{k|k-1}$, (45) can be simplified to: $V_{11}=Q^{-1}.$ (46) Applying the matrix inversion lemma again the expressions for other terms are: $V_{12}=V_{21}^{T}=-V_{11}\mathbf{F}P_{k-1:n|k-1}P_{k-1:n|k-1}^{-1}=-V_{11}\mathbf{F},$ (47) $\begin{split}V_{22}=P_{k-1:n|k-1}^{-1}+\mathbf{F}^{T}V_{11}\mathbf{F}.\end{split}$ (48) The predicted information matrix is then given by: $\begin{split}P_{k:n|k-1}^{-1}=\begin{pmatrix}Q^{-1}&-Q^{-1}\Gamma^{T}\mathbf{F}\\\ -\mathbf{F}^{T}\Gamma Q^{-1}&P_{k-1:n|k-1}^{-1}+\mathbf{F}^{T}\Gamma Q^{-1}\Gamma^{T}\mathbf{F}\end{pmatrix}\\\ \end{split}$ (49) and the predicted information vector can now be derived: $\begin{split}P_{k:n|k-1}^{-1}&X_{k:n|k-1}=\\\ &\begin{pmatrix}Q^{-1}Gu_{k}\\\ P_{k-1:n|k-1}^{-1}X_{k-1:n|k-1}-\mathbf{F}^{T}\Gamma Q^{-1}Gu_{k}\end{pmatrix}.\end{split}$ (50) ## References * [1] I. Loefgren, N. Ahmed, E. Frew, C. Heckman, and S. Humbert, “Scalable event-triggered data fusion for autonomous cooperative swarm localization,” in _2019 22th International Conference on Information Fusion (FUSION)_ , Jul. 2019, pp. 1–8. * [2] W. W. Whitacre and M. E. Campbell, “Decentralized geolocation and bias estimation for uninhabited aerial vehicles with articulating cameras,” _Journal of Guidance, Control, and Dynamics (JGCD)_ , vol. 34, no. 2, pp. 564–573, Mar. 2011. [Online]. Available: http://arc.aiaa.org/doi/10.2514/1.49059 * [3] A. Cunningham, V. Indelman, and F. Dellaert, “DDF-SAM 2.0: consistent distributed smoothing and mapping,” in _2013 IEEE International Conference on Robotics and Automation (ICRA)_ , May 2013, pp. 5220–5227, iSSN: 1050-4729. * [4] H. Li and F. Nashashibi, “Cooperative multi-vehicle localization using split covariance intersection filter,” in _2012 IEEE Intelligent Vehicles Symposium_ , Jun. 2012, pp. 211–216, iSSN: 1931-0587. * [5] J. Zhu and S. S. Kia, “Cooperative localization under limited connectivity,” _IEEE Transactions on Robotics_ , vol. 35, no. 6, pp. 1523–1530, Dec. 2019\. * [6] T. W. Martin and K. C. Chang, “A distributed data fusion approach for mobile ad hoc networks,” in _2005 7th International Conference on Information Fusion (FUSION)_ , vol. 2, Jul. 2005, pp. 1062–1069. * [7] S. Grime and H. Durrant-Whyte, “Data fusion in decentralized sensor networks,” _Control Engineering Practice_ , vol. 2, no. 5, pp. 849–863, Oct. 1994. * [8] C. Y. Chong, E. Tse, and S. Mori, “Distributed estimation in networks,” in _1983 American Control Conference (ACC)_ , Jun. 1983, pp. 294–300. * [9] S. J. Julier and J. K. Uhlmann, “A non-divergent estimation algorithm in the presence of unknown correlations,” in _Proceedings of the 1997 American Control Conference (ACC)_ , vol. 4, Jun. 1997, pp. 2369–2373 vol.4. * [10] T. Bailey, S. Julier, and G. Agamennoni, “On conservative fusion of information with unknown non-Gaussian dependence,” in _2012 15th International Conference on Information Fusion (FUSION)_ , Jul. 2012, pp. 1876–1883. * [11] O. Dagan and N. R. Ahmed, “Heterogeneous decentralized fusion using conditionally factorized channel filters,” in _2020 IEEE International Conference on Multisensor Fusion and Integration for Intelligent Systems (MFI)_ , Sep. 2020, pp. 46–53. * [12] C.-Y. Chong, S. Mori, F. Govaers, and W. Koch, “Comparison of tracklet fusion and distributed Kalman filter for track fusion,” in _17th International Conference on Information Fusion (FUSION)_ , Jul. 2014, pp. 1–8. * [13] S. Lubold and C. N. Taylor, “Formal definitions of conservative PDFs,” _arXiv:1912.06780v2 [ math.ST]_ , May 2021. [Online]. Available: http://arxiv.org/abs/1912.06780v2 * [14] J. K. Uhlmann, “Covariance consistency methods for fault-tolerant distributed data fusion,” _Information Fusion_ , vol. 4, no. 3, pp. 201–215, Sep. 2003. [Online]. Available: https://linkinghub.elsevier.com/retrieve/pii/S1566253503000368 * [15] Y. Bar-Shalom, X. R. Li, and T. kirubarajan, “Linear estimation in static systems,” in _Estimation with Applications to Tracking and Navigation_. John Wiley & Sons, Ltd, 2001, pp. 121–177. * [16] L. Chen, P. Arambel, and R. Mehra, “Fusion under unknown correlation - covariance intersection as a special case,” in _Proceedings of the Fifth International Conference on Information Fusion. (FUSION)_ , vol. 2, Jul. 2002, pp. 905–912 vol.2. * [17] B. Noack, J. Sijs, and U. D. Hanebeck, “Fusion strategies for unequal state vectors in distributed Kalman filtering,” _IFAC Proceedings Volumes_ , vol. 47, no. 3, pp. 3262–3267, Jan. 2014. [Online]. Available: http://www.sciencedirect.com/science/article/pii/S1474667016421105 * [18] J. R. A. Klemets and M. Hovd, “Hierarchical decentralized state estimation with unknown correlation for multiple and partially overlapping state vectors,” in _2018 IEEE Conference on Control Technology and Applications (CCTA)_ , Aug. 2018, pp. 508–514. * [19] S. Radtke, B. Noack, and U. D. Hanebeck, “Distributed estimation with partially overlapping states based on deterministic sample-based fusion,” in _2019 18th European Control Conference (ECC)_ , Naples, Italy, Jun. 2019, pp. 1822–1829. [Online]. Available: https://ieeexplore.ieee.org/document/8795853/ * [20] T. M. Berg and H. F. Durrant-Whyte, “Model distribution in decentralized multi-sensor data fusion,” in _1991 American Control Conference (ACC)_ , Jun. 1991, pp. 2292–2293. * [21] U. A. Khan and J. M. F. Moura, “Distributed Kalman filters in sensor networks: bipartite fusion graphs,” in _2007 IEEE/SP 14th Workshop on Statistical Signal Processing_ , Aug. 2007, pp. 700–704. * [22] C.-Y. Chong and S. Mori, “Graphical models for nonlinear distributed estimation,” in _Proceedings of the 7th International Conference on Information Fusion (FUSION)_ , Stockholm, Sweden, 2004, pp. 614–621. * [23] N. R. Ahmed, W. W. Whitacre, S. Moon, and E. W. Frew, “Factorized covariance intersection for scalable partial state decentralized data fusion,” in _2016 19th International Conference on Information Fusion (FUSION)_ , Jul. 2016, pp. 1049–1056. * [24] V. Saini, A. A. Paranjape, and A. Maity, “Decentralized information filter with noncommon states,” _Journal of Guidance, Control, and Dynamics (JGCD)_ , vol. 42, no. 9, pp. 2042–2054, 2019. [Online]. Available: https://doi.org/10.2514/1.G003862 * [25] S. J. Dourmashkin, N. R. Ahmed, D. M. Akos, and W. W. Whitacre, “GPS-limited cooperative localization using scalable approximate decentralized data fusion,” in _2018 IEEE/ION Position, Location and Navigation Symposium (PLANS)_ , Apr. 2018, pp. 1473–1484. * [26] W. Koch and F. Govaers, “On accumulated state densities with applications to out-of-sequence measurement processing,” _IEEE Transactions on Aerospace and Electronic Systems_ , vol. 47, no. 4, pp. 2766–2778, Oct. 2011, conference Name: IEEE Transactions on Aerospace and Electronic Systems. * [27] J. Vial, H. Durrant-Whyte, and T. Bailey, “Conservative sparsification for efficient and consistent approximate estimation,” in _2011 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)_ , Sep. 2011, pp. 886–893, iSSN: 2153-0866. * [28] N. Carlevaris-Bianco and R. M. Eustice, “Conservative edge sparsification for graph SLAM node removal,” in _2014 IEEE International Conference on Robotics and Automation (ICRA)_ , May 2014, pp. 854–860, iSSN: 1050-4729. * [29] T. B. Schon and F. Lindsten, “Manipulating the multivariate Gaussian density,” Linkoeping University, Tech. Rep., Jan. 2011\. * [30] S. Thrun, W. Burgard, and D. Fox, “The GraphSLAM algorithm,” in _Probabilistic Robotics_. MIT Press, Aug. 2005, pp. 337–383, google-Books-ID: wjM3AgAAQBAJ. * [31] R. Forsling, Z. Sjanic, F. Gustafsson, and G. Hendeby, “Consistent distributed track fusion under communication constraints,” in _2019 22th International Conference on Information Fusion (FUSION)_ , Jul. 2019, pp. 1–8. * [32] J. R. Schoenberg and M. Campbell, “Distributed terrain estimation using a mixture-model based algorithm,” in _2009 12th International Conference on Information Fusion (FUSION)_ , Jul. 2009, pp. 960–967. | Ofer Dagan received the B.S. degree in aerospace engineering, in 2010, and the M.S. degree in mechanical engineering, in 2015, from the Technion - Israel Institute of Technology, Haifa, Israel. He is currently working toward the Ph.D. degree in aerospace engineering with the Ann and H.J. Smead Aerospace Engineering Sciences Department, University of Colorado Boulder, Boulder, CO, USA. From 2010 to 2018 he was a research engineer in the aerospace industry. His research interests include theory and algorithms for decentralized Bayesian reasoning in heterogeneous autonomous systems. ---|--- | Nisar R. Ahmed received the B.S. degree in engineering from Cooper Union, New York City, NY,USA, in 2006 and the Ph.D. degree in mechanical engineering from Cornell University, Ithaca, NY, USA, in 2012. He is an Associate Professor of Autonomous Systems and H. Joseph Smead Faculty Fellow with Ann and H.J. Smead Aerospace Engineering Sciences Department, University of Colorado Boulder, Boulder, CO, USA. He was also a Postdoctoral Research Associate with Cornell University until 2014. His research interests include the development of probabilistic models and algorithms for cooperative intelligence in mixed human–machine teams. ---|---
# Named Entity Recognition in the Style of Object Detection Bing Li Microsoft <EMAIL_ADDRESS> ###### Abstract In this work, we propose a two-stage method for named entity recognition (NER), especially for nested NER. We borrowed the idea from the two-stage Object Detection in computer vision and the way how they construct the loss function. First, a region proposal network generates region candidates and then a second-stage model discriminates and classifies the entity and makes the final prediction. We also designed a special loss function for the second- stage training that predicts the entityness and entity type at the same time. The model is built on top of pretrained BERT encoders, and we tried both BERT base and BERT large models. For experiments, we first applied it to flat NER tasks such as CoNLL2003 and OntoNotes 5.0 and got comparable results with traditional NER models using sequence labeling methodology. We then tested the model on the nested named entity recognition task ACE2005 and Genia, and got F1 score of 85.6$\%$ and 76.8$\%$ respectively. In terms of the second-stage training, we found that adding extra randomly selected regions plays an important role in improving the precision. We also did error profiling to better evaluate the performance of the model in different circumstances for potential improvements in the future. ## 1 Introduction Named entity recognition (NER) is commonly dealt with as a sequence labeling job and has been one of the most successfully tackled NLP tasks. Inspired by the similarity between NER and object detection in computer vision and the success of two-stage object detection methods, it’s natural to ask the question, can we borrow some ideas from there and simply copy the success. This work aims at pushing further our understanding in NER by applying a two- stage object-detection-like approach and evaluating its effects in precision, recall and different types of errors in variant circumstances. As a prototype for many other following object detection models, Faster R-CNN Ren et al. (2015) is a good example to illustrate how the two stages work. In the first stage, a region proposal network is responsible for generating candidate regions of interest from the input image and narrowing down the search space of locations for a more complex and comprehensive investigation process. The second stage work is then done by the regression and classification models, which look at finer feature maps, further adjust the size and position of the bounding box and tell whether there is an object and what type it belongs to. Admittedly a two-stage pipeline is often blamed for error propagation, however in the case of object detection it apparently brings much more than that. We can easily see the analogy here in NER. To do NER in a similar way, first we find candidate ranges of tokens that are more likely to be entities. Second, we scrutinize into each of them, make a judgment if this is a true entity or not, and then do entity classification and also regression if necessary. Even though the search space in 1-D NER problem is significantly smaller than in a 2-D image, the benefits of a two-stage approach are still obvious. Firstly, with the better division of labor, the two components can focus on their specialized tasks which are quite different. More specifically, the region proposal part takes care of the context information and entity signals regardless of entity type, while the second part covers more on entity integrity and type characteristics. Secondly, since the region prediction has been given by the first stage, the second part of the model can take the global information of all the tokens in the region, instead of looking separately as in sequence labeling. Although a single token vector can also encode context information as in BERT Devlin et al. (2018) and other similar LMs Peters et al. (2018a); Yang et al. (2019); Liu et al. (2019b); Vaswani et al. (2017); Radford et al. (2019a), having a global view of all relevant tokens is definitely an advantage and may provide more hints for the final decision. Thirdly, the model gets better interpretability, since it separates the concepts of entityness (how likely this mention is a qualified named entity) and entity classification, you get an interpretable probability for each entity prediction. Finally, another benefit of this method is that each entity prediction is independent, which makes it possible to be used in a nested entity recognition problem that can not be easily handled by sequence labeling approach, which can only predict one label for each token. This paper is structured as follows. In Section 2 we explain the model architecture, including two stages, region proposal and entity classification. Section 3 reviews the past related works in NER, especially region based NER that are most similar to our work. Section 4 explains the training process and evaluation results in detail on flat NER tasks. Section 5 shows the training and evaluation on nested NER tasks. In Section 6 we evaluate the importance of different parts of the model by ablation. And finally in Section 7 we show the error analysis of the new method and compared it with the traditional sequence labeling methodology. ## 2 Model Design We propose a two-stage method for NER. In the first stage, an entity region proposal network predicts highly suspected entity regions. In the second stage, another model takes the output candidates, extracts more features and then makes a prediction whether this candidate is a qualified entity or not and what type should be attached to this entity. Precision and recall would be evaluated end-to-end, in the same manner as traditional NER models. Figure 1: Region proposal network is made by adding a linear layer on top of the BERT output. The start prediction is independent from the end prediction, both of which use cross entropy loss over two classes. For the end position, we assign it to the first token of the word right immediate after the entity. We only predict for the first token of every word if the tokenizer outputs WordPieces or other subword tokens, and all other tokens will be masked (the mask is omitted from the diagram). ### 2.1 Entity Region Proposal For the first stage, we used a simple model similar to sequence labeling ones. But instead of predicting IOB-format entity labels Ramshaw and Marcus (1995), we predict $\langle$Start$\rangle$ and $\langle$End$\rangle$ labels. We appended another linear layer on top of the hidden state outputs from the BERT model Devlin et al. (2018), to predict the probability for a token to be $\langle$Start$\rangle$ and $\langle$End$\rangle$, we only consider the starting token in each word and mask out all trailing tokens within that word. And we used cross entropy over two classes for the prediction. The model structure is demonstrated in Figure 1. The goal of the first-stage model is high recall and acceptable precision, so we could tilt the weights a little bit towards the positive labels to favor better recall numbers. After extracting the start and end tokens, we then select pairs to form a complete region, with the simple rule that the length of the entity cannot exceed some limit. Admittedly, there is an apparent disadvantage here that is we discard longer candidates without even a try. But in fact, those longer ones only compose quite a small portion and are usually very difficult to get right anyway. In practice we have used 6 and 12 as the length limit for different datasets and can easily cover 98$\%$-99$\%$ of ground-truth entities. ### 2.2 Entity Discrimination and Classification Figure 2: The second-stage model is responsible for re-examining the entity region proposals generated by the first stage. The model trains with 4 tasks simultaneously. The entityness part and type classification part takes global view of all tokens within the range, while the two boundary losses only zoom into the two tokens across the boundary to make a double check of the exact range. The second stage is also using BERT Devlin et al. (2018) as the encoding model. The second stage has two main tasks, discrimination and classification, which defines the two major components in our loss function, the entityness loss and the type classification loss. Contrary to a typical object detection model, we don’t do regression for the bounding box. The model will only tell if this range is correct or not and if not, discard it directly. One reason is that we want to keep it as simple as possible and another reason is that the problem is much easier than the 2-D object detection, and our proposals are usually accurate enough. Compared to the sequence labeling method, our model focuses more on aggregate features across all tokens spanning the entity, which can be seen from the max pooling layer in both the entityness and classification components. To make the model more sensitive to boundary errors, especially when coming across long entities, we added another two losses, the start loss and the end loss, which predict if the boundaries are correct, so that the prediction won’t be dominated by the bulk part but also pay attention to the boundary. These start/end logits are also concatenated with the entityness feature to predict the entityness score. The total loss function can be written as below: L $\displaystyle=\alpha\left(\textbf{L}_{\text{start}}+\textbf{L}_{\text{end}}\right)+\beta\textbf{L}_{\text{entityness}}+\textbf{L}_{\text{type}}$ (1) $\displaystyle\textbf{L}_{\text{s}}$ $\displaystyle=-\frac{1}{N}\sum_{i=1}^{N}\left[{\mathbbm{1}}_{i}^{\text{s}}\log(p_{i}^{\text{s}})+(1-{\mathbbm{1}}_{i}^{\text{s}})\log(1-p_{i}^{\text{s}})\right],$ $\displaystyle\>\>\>\>\>\>\>\>\>\>\>\>\>\>\>\>\text{s}=\text{start, end, entityness}$ (2) $\displaystyle\textbf{L}_{\text{type}}$ $\displaystyle=-\frac{1}{N}\sum_{i=1}^{N}{\mathbbm{1}}_{i}^{\text{entity}}\left[\sum_{c\in\textrm{classes}}{\mathbbm{1}}_{i}^{\text{c}}\log(p_{i}^{\text{c}})\right]$ (3) In Equation 1, $\alpha$ and $\beta$ are hyperparameters that control the weights of the boundary loss and the total entityness loss, and we used 0.5 and 1.0 as default values. In Equation 2-3, $i$ iterates through from sample 1 to N where each sample can be seen as a tuple $(sentence,index_{start},index_{end})$. ${\mathbbm{1}}_{i}^{\text{entity}}$ is an indicator that the sample $i$ is an entity. ${\mathbbm{1}}_{i}^{\text{c}}$ indicates if the entity belongs to type c. Same idea for ${\mathbbm{1}}_{i}^{\text{s}}$. The start loss, end loss and entityness loss are all cross entropy loss over two classes. The type classification loss is only applied when the region proposal matches a true entity, and would be ignored otherwise. The model structure is illustrated in Figure 2. For the start and end loss, we used a model that calculate multi-head dot products between the two tokens across the boundary. The intuition is that it needs to catch the relation signal between the two sides. After dot products we put an extra fully connected layer to transform the feature vector to the logits of a two-class classification. All heads have independent weights. For the entityness and type classification loss, we used the same architecture with separate weights, that is a fully connected layer followed by a max pooling layer, and then another fully connected layer to get the final logits. And we use ReLU activation function after each linear layer. The same as region proposal, we only consider the starting token in each word and mask out all trailing tokens within that word. During inference, we look at the entityness probability first. Only if it’s above a specified threshold will we look at the classification results and output the most likely type. ## 3 Related Work Named entity recognition (NER) is a classic problem in NLP. The goal of NER is to extract named entities from free text and these entities can be classified into several categories, for example person (PER), location (LOC) and geo- political entity (GPE). Traditionally NER is tackled by sequence labeling method. Many different models are developed along this direction, such as CRFs in Lafferty et al. (2001); Sutton et al. (2007), LSTM in Hammerton (2003), LSTM-CRF in Lample et al. (2016) etc. More recently, people start using large- scale language models, such as BERT Devlin et al. (2018) and ELMo Peters et al. (2018a). Nested named entity recognition takes the overlapping between entities into consideration Kim et al. (2003). This cannot be easily done by traditional sequence labeling in that one can only assign one tag for each token. There have been many different approaches proposed for this problem. One important branch is the region-based method and our work can also be classified into this category. Finkel and Manning (2009) leveraged parsing trees to extract subsequences as candidates of entities. Xu et al. (2017); Sohrab and Miwa (2018) considers all subsequences of a sentence as candidates. A more recent paper by Lin et al. (2019b) developed a model that locates anchor word first and then searches for boundaries of the entity, with the assumption that nested entities have different anchor words. Another work that is very close to ours is done by Zheng et al. (2019). In their paper they also proposed a two stage method. Our work is different than theirs from several perspectives. We have an entityness prediction in the second stage like the objectness in object detection, and thus we don’t completely depend on the first stage to determine the region. And our model is built on BERT language model and finetunes all lower layers, while theirs is using LSTM plus pretrained word embedding. Another branch of researches is trying to design more expressive tagging schemas, some representative works are done by Lu and Roth (2015); Katiyar and Cardie (2018); Wang and Lu (2018). The current state of the art is Li et al. (2019a), where they viewed the NER problem as a question answering problem and naturally solved the nested issue. Their model showed impressive power in both flat and nested NER tests. A major difference between our model and theirs is that they predict a paring score for any start and end index pair, which makes the feature matrix and the computational complexity a big issue. We instead predict entityness only for very few candidates, and we don’t need to duplicate training examples for multiple queries, thus our training process takes much less time. The main contribution of our work is that we bring up the idea to use a high- recall and relatively low-precision first stage model to select regions and use a more complicated model to predict the entityness and classification at the same time, with a global view of all the tokens spanning the candidate entity. Besides that, our model is much simpler and more lightweight than other similar models designed for nested NER tasks, and both training and inference run as fast as the plain BERT sequence labeling model. Our core model architecture is nothing but a linear layer plus a max pooling layer, but gives pretty good performance, especially on the ACE2005 dataset. ## 4 Flat NER Experiments For the flat NER experiment, we used the CoNLL2003 Sang and Meulder (2003) and OntoNotes 5.0 Pradhan et al. (2013). CoNLL2003 is an English dataset with four types of named entities, namely Location, Organization, Person and Miscellaneous. And OntoNotes 5.0 is an English dataset containing text from many sources and including 18 types of named entity, | Dev Precision | Dev Recall | Test Precision | Test Recall ---|---|---|---|--- CoNLL2003 Region Proposal | | | | BERT base | 71.3 | 98.0 | 70.7 | 96.2 BERT large | 71.4 | 98.0 | 70.6 | 96.5 OntoNotes 5.0 Region Proposal | | | | BERT base (weight 0.5:0.5) | 69.3 | 90.8 | 69.3 | 89.6 BERT base (weight 0.3:0.7) | 67.8 | 92.7 | 67.4 | 92.2 BERT base (weight 0.2:0.8) | 66.3 | 93.9 | 65.8 | 93.7 BERT base (weight 0.1:0.9) | 63.8 | 95.1 | 63.1 | 95.5 Table 1: Region Proposal Model Results. Precision and recall numbers are region metrics regardless of entity type. A prediction is correct as long as the region predicted matches the start and end of the ground-truth entity. Region precision and recall are reported for both dev and test sets. | Dev Precision | Dev Recall | Dev F1 | Test Precision | Test Recall | Test F1 ---|---|---|---|---|---|--- CoNLL2003 | | | | | | BERT base | 95.1 | 95.1 | 95.1 | 91.9 | 91.7 | 91.8 BERT large | 96.1 | 95.4 | 95.8 | 92.0 | 91.6 | 91.8 OntoNotes 5.0 | | | | | | BERT base | 87.7 | 87.3 | 87.5 | 86.9 | 86.5 | 86.7 BERT large | 88.0 | 88.3 | 88.2 | 86.8 | 87.3 | 87.0 Table 2: Flat NER Results. Standard NER precision and recall are reported here for both Dev and Test sets. We only showed the best model for each combination of dataset and the size of BERT encoder. ### 4.1 First-Stage Training For region prediction, we used the default training parameters provided by HuggingFace Transformers Wolf et al. (2019) for token classification task, i.e. AdamW optimizer with learning rate=$5\times 10^{-5}$, $\beta_{1}$=$0.9$ and $\beta_{2}$=$0.999$, hidden state dropout probability 0.1 etc. We finetuned both BERT-base-cased and BERT-large-cased models for 3 epochs with batch size of 64. The regions with length equal or less than 6 were selected as candidates for the second stage. We chose the threshold 6 because most named entites are shorter than 6 (99.9$\%$ in CoNLL2003 and 99.2$\%$ in OntoNotes 5.0). Since we ignore entity type in the first stage model, the precision and recall are based only on region proposals regardless of type. We keep the model to be as simple as possible because the only goal is to get high recall in the first stage. For the OntoNotes 5.0 model, a default training gave a model with pretty low recall, only 89.6$\%$, so we changed the weights in the cross entropy loss to raise the recall with a little trade-off of precision. Therefore precision and recall were reported at different weights for OntoNotes. The training results can be found in Table 1. From the result, we can see that for CoNLL2003 Sang and Meulder (2003), base and large models gave pretty close p/r numbers, so we used BERT base in the following experiments. For OntoNotes dataset, we tried several different weights, it turned out that weights 0.4:0.6 and 0.3:0.7 both gave pretty good end-to-end results. | | Genia | | | ACE2005 | ---|---|---|---|---|---|--- Model | Precision($\%$) | Recall($\%$) | F1($\%$) | Precision($\%$) | Recall($\%$) | F1($\%$) ARN Lin et al. (2019b) | 75.8 | 73.9 | 74.8 | 76.2 | 73.6 | 74.9 Boundary-aware Neural Zheng et al. (2019) | 75.8 | 73.6 | 74.7 | - | - | - Merge-BERT Fisher and Vlachos (2019) | - | - | - | 82.7 | 82.1 | 82.4 Seq2seq-BERT Straková et al. (2019) | 80.1 | 76.6 | 78.3 | 83.5 | 85.2 | 84.3 Path-BERT Shibuya and Hovy (2019) | 77.81 | 76.94 | 77.36 | 83.83 | 84.87 | 84.34 BERT-MRC Li et al. (2019a) | 85.18 | 81.12 | 83.75 | 87.16 | 86.59 | 86.88 Our Model | | | | | | BERT Base Model | 76.6 | 75.1 | 75.9 | 82.8 | 84.9 | 83.8 BERT Large Model | 77.4 | 76.3 | 76.8 | 85.2 | 85.9 | 85.6 Table 3: Nested NER Results on Genia and ACE2005. ### 4.2 Second-Stage Training There are quite a few differences between our second-stage training and the traditional sequence labeling NER training, in the sense that our training is on region proposal level while sequence labeling is on sentence level. Each training example is now a combination of a sentence, a proposal start index and a proposal end index, and one sentence could emit multiple training examples. And the labels are no longer token labels, but labels designed for our specific loss, specifying if the start index is correct, if the end index is correct, if the region is corresponding to a true entity, and at last what the entity type is if all the previous answers are positive. The training samples are those region proposals output from the first model and we added more randomly selected negative regions to make it more robust, which turned out to be very important and will be explained with more details in the following paragraphs. Finally we evaluated the model performance in an end-to- end manner. Precision, recall and F1 score on CoNLL2003 and OntoNotes 5.0 can be found in Table 2. For the CoNLL2003 dataset Sang and Meulder (2003), the best F1 score we got with base and large models are both 91.8, which is comparable but a little lower than the reported sequence labeling BERT results (BERT-Tagger). We didn’t spend too much time on finetuning hyperparameters, it’s possible that there is still room for the model. Another possible reason could be in our model we only predict for one entity at a time and didn’t consider the interaction between entities when there are multiple in one sentence. For the OntoNotes 5.0 dataset we got F1 score 87.0. We tried a few tricks to improve the p/r number. The most effective one is to add more randomly selected regions as negative samples during the second stage. For each sentence, we generate one random region that has length in range [1, 6] and that doesn’t fall into the existing candidates or true entities. We then labeled them as wrong samples and fed into the training process together with other candidates. With more random negative samples, the model is more robust when there is a big gap between the train and test set distributions, especially when we have a very strong stage-one model with high precision, which could have a strong bias on the distribution of negative samples. By adding more negatives, we have almost 0.7$\%$ gain in the F1 score, which will be shown with more details in the following ablation study section 6. We also tried adding an extra loss to take into account the classification error for those regions that were not predicted completely correct but have a large overlap with one of the true entities. They were not exactly matching the ground true region, therefore we assigned a weight less than 1 (0.2 in our experiment). Intuitively, this solution is equivalent to adding more classification training data. However we didn’t see significant improvement after this change. We also explored changing hyperparameters, for example the numbers of channels in the second-stage model. In the default settings we set the number of heads in dot product to be 32, and the dimension of feature vector in both entityness and classification to be 64. We tried half or double these numbers but the model performance degraded in both cases. More details can be found in the next ablation study subsection 6. ## 5 Nested NER Experiments For nested NER experiments, we chose two mainstream datasets, the ACE2005 dataset Christopher Walker and Maeda (2006) and Genia dataset Kim et al. (2003). The ACE2005 dataset contains 7 entity categories. For comparison, we follow the data preprocessing procedure in Lu and Roth (2015); Wang and Lu (2018); Katiyar and Cardie (2018) by keeping files from bw, bn, nw and wl, and splitting these files randomly into train, dev and test sets by 8:1:1, respectively. For the Genia dataset, again we use the same preprocessing as Finkel and Manning (2009); Lu and Roth (2015) and we also consider 5 entity types - DNA, RNA, protein, cell line and cell type. Our training process for nested NER is basically the same as the previous section, since our model doesn’t differentiate between flat and nested and will simply treat all entity regions in the same way even if they are overlapped. Considering the different distribution of entity lengths, we increased the length limit for region candidates from 6 words to 12 for ACE2005 and 8 for Genia. For the region proposal model, we again trained on top of BERT-base-cased for 3 epochs, and in the cross entropy loss we used weights 0.1:0.9. For the second stage, we trained both BERT-base and BERT- large for 10 epochs, the results are reported in Table 3. With the BERT large model we got an average F1 score of 85.6 on ACE2005 and 76.8 on Genia. This is not as good as the current state-of-the-art result, but is quite competitive, and especially on ACE2005 dataset it’s better than all other models except the BERT-MRC model. Entity Type | Test Precision | Test Recall ---|---|--- DNA | 74.8 | 72.6 RNA | 90.6 | 82.1 cell line | 78.5 | 67.1 cell type | 73.6 | 72.2 protein | 78.5 | 79.7 Overall | 77.4 | 76.3 Table 4: Precision and Recall by Category. (Genia) Entity Type | Test Precision | Test Recall ---|---|--- PER | 88.6 | 89.9 LOC | 77.0 | 76.2 ORG | 74.8 | 77.8 GPE | 87.7 | 86.7 FAC | 82.0 | 77.8 VEH | 72.3 | 71.4 WEA | 75.0 | 73.8 Overall | 85.2 | 85.9 Table 5: Precision and Recall by Category. (ACE2005) A detailed P/R analysis by category for Genia and ACE2005 datasets are given in Table 4 and Table 5. ## 6 Ablation Study In this part, we did an ablation study to assess the contribution from each component of the new model. In the first experiment, we removed the start/end logits from the concatenated vector for entityness prediction and also removed the start/end loss from the total loss, to see if they are helpful to resolve boundary errors. Then, we evaluated the effects of the max pooling layer over the whole entity and used only the first token’s feature vector instead. In the following experiments we also tried removing random negative samples and reducing the model size by using less channels or dimensions for the start/end unit, entityness prediction and entity classification etc. The evaluation results on the CoNLL2003 test set is shown in Table 6. From the results, we can see that the start/end prediction only has limited effect for the final F1, but with a more closer look at the errors we found that with start/end prediction, the boundary errors dropped from 149 (out of 5711 test samples) to 127 while the type classification errors changed from 222 to 230. The start/end loss did change the pattern of errors and corrected some boundary mistakes. Removing the max pooling layer brought an F1 drop greater than 1.0$\%$ and removing random negative samples brought 0.7$\%$ drop. After reducing the number of channels to half of the default, we got a drop of 0.5$\%$ and when cutting further, keeping only 25$\%$ channels, the model performance degraded pretty fast to 65$\%$, indicating a strong underfit. ## 7 Error Analysis To have a deeper understanding what type of errors our model is making, we did a further error profiling and also compared it with the result of a standard sequence labeling model based on BERT. Inspired by the methodology from Object Detection research, We divided the NER errors into the following types: (1) the region is correct but the type classification is wrong; (2) one of the left/right boundaries of the region is wrong, but classification is correct; (3) one of the left/right boundaries of the region is wrong, and the classification is also wrong; (4) over-triggering: the predicted entity has no overlap with any ground-truth entity; (5) under-triggering: for a true entity, either no overlapping region is proposed or the entityness model predicts no entity; (6) another type of error is that both boundaries are wrong but the region has overlap with at least one of true entities, in the evaluation we found that this type of error occurs only once, so we ignored it in the pie plot. We displayed the error composition for the new two-stage model as well as a traditional sequence labeling BERT model side by side, as shown in Figure 3. | Test P | Test R | F1 ---|---|---|--- Original setting | 91.9 | 91.7 | 91.8 Remove start/end prediction | 90.4 | 91.9 | 91.7 Remove max pooling | 90.0 | 90.7 | 90.4 Remove random negatives | 90.6 | 91.6 | 91.1 Remove 50$\%$ channels | 91.2 | 91.3 | 91.3 Remove 75$\%$ channels | 65.0 | 65.4 | 65.2 Table 6: Ablation Study using CoNLL2003 Dataset. All experiments are using the BERT base model and training with the same epochs. P, R and F1 are reported on test set. Figure 3: Error Analysis. We have divided the entity recognition error to 5 classes, more details can be found in the text. We profiled the errors for traditional sequence labeling BERT model (the left panel) and also the two-stage model proposed in this paper (the right panel), to provide a deeper insight of the composition of model errors. We ran analysis on two models both trained from BERT-base and having similar precision and recall. As we can see, for both models the dominant part is the type error (region is correct). We could consider using larger and more complicated type prediction models since there is a lot of room in that direction reading from the plot. The new model made significantly less over- triggering errors, which means the precision of the entityness prediction is good. And two models have similar amounts of single-boundary errors and under- triggering errors. ## 8 Conclusion In this paper, we proposed a new two-stage model for named entity recognition. And many of the ideas are coming from the inspiration of object detection in computer vision. More specifically, with a coarse first-stage model to provide region proposals, we rely on a second-stage model to predict entityness and entity types at the same time. Through sufficient experiments in both flat and nested NER tasks, we found that it has better performance on nested NER, and we got F1 85.6 on ACE2005 dataset and F1 76.8 on Genia datase, better than many more complicated models. On flat NER tasks, it’s still a few points behind the current SOTA results. In the future we are going to improve the model further to see where is the real limit of two-stage region-based named entity recognition. ## References * Ren et al. (2015) S. Ren, K. He, R. Girshick, and J. Sun. 2015. Faster r-cnn: Towards real-time object detection with region proposal networks. arXiv preprint arXiv:1506.01497, 2015. * Devlin et al. (2018) Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. _arXiv preprint arXiv:1810.04805_. * Yang et al. (2019) Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. 2019. Xlnet: Generalized autoregressive pretraining for language understanding. In _NIPS_ , pages 5754–5764. * Liu et al. (2019b) Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019b. Roberta: A robustly optimized BERT pretraining approach. _CoRR_ , abs/1907.11692. * Vaswani et al. (2017) Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In _NIPS_ , pages 5998–6008. * Radford et al. (2019a) Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019a. Language models are unsupervised multitask learners. * Ramshaw and Marcus (1995) Lance Ramshaw and Mitch Marcus. 1995. Text chunking using transformation-based learning. In _Proceedings of Third Workshop on Very Large Corpora_ , pages 82–94. * Lafferty et al. (2001) John Lafferty, Andrew McCallum, and Fernando CN Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. * Sutton et al. (2007) Charles Sutton, Andrew McCallum, and Khashayar Rohanimanesh. 2007. Dynamic conditional random fields: Factorized probabilistic models for labeling and segmenting sequence data. _Journal of Machine Learning Research_ , 8(Mar):693–723. * Hammerton (2003) James Hammerton. 2003. Named entity recognition with long short-term memory. In _Proceedings of the seventh conference on Natural language learning at HLT-NAACL 2003-Volume 4_ , pages 172–175. Association for Computational Linguistics. * Lample et al. (2016) Guillaume Lample, Miguel Ballesteros, Sandeep Subramanian, Kazuya Kawakami, and Chris Dyer. 2016. Neural architectures for named entity recognition. _arXiv preprint arXiv:1603.01360_. * Kim et al. (2003) J-D Kim, Tomoko Ohta, Yuka Tateisi, and Jun’ichi Tsujii. 2003. Genia corpus—a semantically annotated corpus for bio-textmining. _Bioinformatics_ , 19(suppl_1):i180–i182. * Peters et al. (2018a) Matthew E Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018a. Deep contextualized word representations. _arXiv preprint arXiv:1802.05365_. * Finkel and Manning (2009) Jenny Rose Finkel and Christopher D Manning. 2009. Nested named entity recognition. In _Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing: Volume 1-Volume 1_ , pages 141–150. Association for Computational Linguistics. * Sang and Meulder (2003) Erik F. Tjong Kim Sang and Fien De Meulder. 2003. Introduction to the conll-2003 shared task: Language-independent named entity recognition. In _Proceedings of the Seventh Conference on Natural Language Learning, CoNLL 2003, Held in cooperation with HLT-NAACL 2003, Edmonton, Canada, May 31 - June 1, 2003_ , pages 142–147. * Pradhan et al. (2013) Sameer Pradhan, Alessandro Moschitti, Nianwen Xue, Hwee Tou Ng, Anders Björkelund, Olga Uryupina, Yuchen Zhang, and Zhi Zhong. 2013. Towards robust linguistic analysis using OntoNotes. In _Proceedings of the Seventeenth Conference on Computational Natural Language Learning_ , pages 143–152, Sofia, Bulgaria. Association for Computational Linguistics. * Christopher Walker and Maeda (2006) Julie Medero Christopher Walker, Stephanie Strassel and Kazuaki Maeda. 2006. Ace 2005 multilingual training corpus. _Linguistic Data Consortium, Philadelphia 57._ * Lu and Roth (2015) Wei Lu and Dan Roth. 2015. Joint mention extraction and classification with mention hypergraphs. In _Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing_ , pages 857–867. * Wang and Lu (2018) Bailin Wang and Wei Lu. 2018. Neural segmental hypergraphs for overlapping mention recognition. _arXiv preprint arXiv:1810.01817_. * Katiyar and Cardie (2018) Arzoo Katiyar and Claire Cardie. 2018. Nested named entity recognition revisited. In _Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers)_ , pages 861–871. * Lin et al. (2019b) Hongyu Lin, Yaojie Lu, Xianpei Han, and Le Sun. 2019b. Sequence-to-nuggets: Nested entity mention detection via anchor-region networks. In _Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers_ , pages 5182–5192. * Shibuya and Hovy (2019) Takashi Shibuya and Eduard H. Hovy. 2019. Nested named entity recognition via second-best sequence learning and decoding. _CoRR_ , abs/1909.02250. * Fisher and Vlachos (2019) Joseph Fisher and Andreas Vlachos. 2019. Merge and label: A novel neural network architecture for nested NER. In _Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers_ , pages 5840–5850. * Straková et al. (2019) Jana Straková, Milan Straka, and Jan Hajic. 2019. Neural architectures for nested NER through linearization. In _Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics_ , pages 5326–5331, Florence, Italy. Association for Computational Linguistics. * Zheng et al. (2019) Changmeng Zheng, Yi Cai, Jingyun Xu, Ho-fung Leung, and Guandong Xu. 2019. A boundary-aware neural model for nested named entity recognition. In _Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)_ , pages 357–366. * Li et al. (2019a) Xiaoya Li, Jingrong Feng, Yuxian Meng, Qinghong Han, Fei Wu, and Jiwei Li. 2019a. A unified mrc framework for named entity recognition. _arXiv preprint arXiv:1910.11476_. * Xu et al. (2017) Mingbin Xu, Hui Jiang, and Sedtawut Watcharawittayakul. 2017. A local detection approach for named entity recognition and mention detection. In _Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)_ , volume 1, pages 1237–1247. * Sohrab and Miwa (2018) Mohammad Golam Sohrab and Makoto Miwa. 2018. Deep exhaustive model for nested named entity recognition. In _Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing_ , pages 2843–2849, Brussels, Belgium. Association for Computational Linguistics. * Wolf et al. (2019) Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R’emi Louf, Morgan Funtowicz, and Jamie Brew. 2019. HuggingFace’s Transformers: State-of-the-art natural language processing. _arXiv preprint arXiv:1910.03771_.
figure # A Coding Theory Perspective on Multiplexed Molecular Profiling of Biological Tissues Luca D’Alessio1 Broad Institute, Cambridge, MA <EMAIL_ADDRESS>Litian Liu1 MIT, Cambridge, MA <EMAIL_ADDRESS>Ken Duffy Maynooth University, Ireland <EMAIL_ADDRESS>Yonina C. Eldar Muriel Médard Weizmann Institute of Science, Israel <EMAIL_ADDRESS>MIT, Cambridge, MA <EMAIL_ADDRESS>Mehrtash Babadi Luca D’Alessio and Mehrtash Babadi acknowledge funding and support from Data Sciences Platform (DSP), Broad Institute. Litian Liu acknowledges financial support from Klarman Family Foundation. Yonina Eldar acknowledges funding from NIMH grant 1RF1MH121289-0. The authors thank Samouil L. Farhi for beneficial discussions, and Aviv Regev for supporting this project. Broad Institute, Cambridge, MA <EMAIL_ADDRESS> ###### Abstract High-throughput and quantitative experimental technologies are experiencing rapid advances in the biological sciences. One important recent technique is multiplexed fluorescence in situ hybridization (mFISH), which enables the identification and localization of large numbers of individual strands of RNA within single cells. Core to that technology is a coding problem: with each RNA sequence of interest being a codeword, how to design a codebook of probes, and how to decode the resulting noisy measurements? Published work has relied on assumptions of uniformly distributed codewords and binary symmetric channels for decoding and to a lesser degree for code construction. Here we establish that both of these assumptions are inappropriate in the context of mFISH experiments and substantial decoding performance gains can be obtained by using more appropriate, less classical, assumptions. We propose a more appropriate asymmetric channel model that can be readily parameterized from data and use it to develop a maximum a posteriori (MAP) decoders. We show that false discovery rate for rare RNAs, which is the key experimental metric, is vastly improved with MAP decoders even when employed with the existing sub- optimal codebook. Using an evolutionary optimization methodology, we further show that by permuting the codebook to better align with the prior, which is an experimentally straightforward procedure, significant further improvements are possible. **footnotetext: These two authors contribute equally to the work. ## I Introduction In recent years, the field of single-cell biology has witnessed transformative advances in experimental and computational methods. Of particular interest is the recent advent of multiplexed fluorescence in situ (in-place) hybridization (mFISH) microscopy techniques that allow molecular profiling of hundreds of thousands of cells without disturbing their complex arrangement in space. This highly-informative data modality paves the way to transformative progress in many areas of biology, including understanding morphogenesis, tissue regeneration, and disease at molecular resolution. One of the major challenges in designing such experiments is the vastness of functional bio-molecules. For example, the human genome codes nearly 30k non- redundant types of RNA molecules, many of which translate to proteins with specific functions. Modern data-driven biology heavily relies on our ability to measure as many different types of functional molecules as possible. Clearly, a sequential imaging approach is impractical. Fortunately, a typical cell produces a rather sparse set of all molecules, and some of the most promising mFISH techniques exploit molecular sparsity in space together with coding ideas in order to multiplex the measurements into fewer imaging rounds [1, 2, 3, 4, 5]. In brief, the mFISH technique involves assigning binary codes to RNA molecules of interest, chemically synthesizing and “hybridizing” these codes to the molecules, and measuring them in space one bit at a time via sequential fluorescence microscopy. A more detailed account of one such pioneering technique known as MERFISH (“multiplexed error-robust fluorescence in situ hybridization”)[1] is given in Sec. I-A (also, cf. Fig. 1). An important part of the MERFISH protocol is the utilization of sparse codes with large minimum distance to allow error correction. Referred to as MHD4 codes [1, 2, 3], these 16-bit codes have minimum Hamming distance 4 and contain 4 ones and 12 zeros each. The bit imbalance is motivated by the empirically observed $\sim 2\times$ higher signal fallout rate compared to false alarm. There are only 140 such codes and therefore one is limited to measuring at most 140 distinct molecules. These codes are randomly assigned to the RNA molecules of interest. The decoding method in current use relies on quantization, Hamming error correction, and rejection of ambiguous sequences. We point out that the assumptions motivating the codebook construction and decoding, tacitly yet heavily, rely on source uniformity and to a certain extent on the binary symmetric channel paradigm, both of which are violated in the context molecular profiling. For channel coding in communication, source can be readily assumed as uniformly distributed thanks to compression in source coding and the separation theorem [6]. In molecular profiling, however, source compression is not applicable and the distribution of RNA molecules is extremely non-uniform. Moreover, fluorescence microscopy is established to be highly asymmetric in terms of fallout and false alarm. These violated assumptions become a source of potential problems when directly applying communication encoding and decoding paradigms. For example, the false discovery rate of rare molecules is found to be unacceptably high in replicate experiments [1, 2], which we later show to be a consequence of the assumed source uniformity. Accurate quantification of rare RNA molecules (e.g. transcription factors) is particularly important for data-driven biological discovery since rare molecules often signal rare events, transient cells states, etc. This motivates our primary goal in this paper: to incorporate the prior non-uniformity in the decoding process in a principled way in order to control false discovery rate of rare molecules. In practice, either accurate priors are known, can be estimated from the data, or can be measured cheaply and effortlessly (e.g. using bulk RNA sequencing [1]). This paper is organized as follows: we review the MERFISH protocol in Sec. I-A and propose a generative model for the data in Sec. II-A, along with a model fitting algorithm and a procedure to derive a more tractable binary asymmetric channel (BAC) formulation from the fitted model. The BAC framework allows us to evaluate the performance of different encoding and decoding schemes. We incorporate the prior non-uniformity into the decoding algorithm by developing a maximum a posteriori (MAP) decoder with a tunable rejection threshold in Sec. II-D. We show that the false discovery rate of rare RNAs, which is the key experimental metric, is vastly improved compared to the presently used MLE-based decoding method [1, 2, 3], even when employed with the existing sub- optimal MHD4 codebook. Finally, we take a first step in data-driven code construction in Sec. III. Using an evolutionary optimization methodology, we show that by permuting the codebook to better align with the prior, which is an experimentally straightforward procedure, significant further improvements are possible. We conclude the paper in Sec. IV with of follow up research directions. ### I-A A brief overview of the MERFISH protocol Figure 1: A schematic overview of a typical mFISH experiment. (a) codebook and probe design; (b) sequential imaging; (c) image processing and decoding. In this section, we briefly review the MERFISH protocol [1], recount different sources of noise and nuisance, and motivate a generative process for MERFISH data. Fig. 1 shows a schematic overview of the MERFISH technique. This protocol consists of four main steps: Step 1. A unique binary codeword of length $L=16$ is assigned to the RNA molecules of interest; Step 2. The specimen is stained with carefully designed short RNA sequences called encoding probes. The middle part of the encoding probes bind with high specificity to a single RNA type while their flanking heads and tails contain a subset of $L$ artificial sequences, $\\{R_{1},\ldots,R_{L}\\}$, called readout sequences. The choice of readout sequences reflects the intended binary codeword. For instance, if the code for a certain RNA type contains “1” at positions 1, 3, 5, and 15, the encoding probes are designed to have $R_{1},R_{3},R_{5}$, and $R_{15}$ flanking sequences (see Fig. 1a); Step 3. The prepared tissue undergoes $L$ rounds of imaging. Imaging round $l$ begins with attaching fluorescent readout probes for round l to the prepared tissue. These probes bind to the flanking readout sequences and contain a fluorescent dye that emits light upon excitation. The round ends with bleaching the dye. In effect, imaging round $l$ reveals the position of all RNA molecules having “1” in their binary codeword at position $l$. Step 4. Finally, the position of RNA molecules, which appear as bright spots, are identified using conventional image processing operations (see Fig. 2). The data is summarized as an $N\times L$ intensity matrix ($N$ being the number of identified spots) and is ultimately decoded according to the codebook. MERFISH measurements are affected by several independent sources of noise. These include factors that are intrinsic to individual molecules, such as (1) stochasticity in the hybridization of encoding and readout probes, (2) random walk of molecules between imaging rounds, and (3) CCD camera shot noise. These factors module the intensity measurements independently in each round and are largely uncorrelated across rounds. Extreme multiplexing (e.g. as in the seqFISH+ protocol [5]) further leads to interference noise due to signal mixing between nearby molecules. This nuisance, however, is rather negligible in the MERFISH protocol. ## II Methodology and Results ### II-A A generative model for mFISH data Figure 2: Extraction of isolated spots from MERFISH images (data from [2]). (a) local peak finding; (b) identification of isolated spots; (c) intensity series from 10 random spots (rows); the leftmost 16 columns show the intensity measurements; the last two column show the summed intensity and nearest- neighbor cross-correlations and are used for filtering of poorly localized spots. Figure 3: Modeling spot intensities as two-component Gaussian mixture for each data dimension (i.e. readout round and color channel). (a) model fitting (black and red lines) and empirical histograms (gray); the green lines indicate the quantization thresholds for the ensuing BAC approximation; (b) QQ-plots for each data dimension; the labels shown in the sub-panels indicate hybridization rounds $\\{1,\ldots,8\\}$ and color channels $\\{1,2\\}$. In this section, we present a simple generative model for MERFISH spot intensity data, fit the model to real data, and evaluate the goodness of fit. This model will serve as a foundation for developing a MAP decoder. Fig. 2 shows a typical example of MERFISH data from [2]. We formalize the data generating process as follows: let $\mathsf{C}\subset\\{0,1\\}^{L}$ be a set of codewords with cardinality $|\mathsf{C}|=K$ which are assigned to $G\leq K$ molecules, let $a:\tilde{\mathsf{C}}\rightarrow\\{1,\ldots,G\\}$ be the bijective code assignment map where $\tilde{\mathsf{C}}\subset\mathsf{C},|\tilde{\mathsf{C}}|=G$ is the set of used codes, and let $\boldsymbol{\pi}_{1:G}$ be the prior distribution of molecules. Setting aside interference effects, we model the fluorescence intensity series $\mathbf{I}_{1:L}\in[0,\infty)^{L}$ measured for an arbitrary molecule as follows: $\begin{split}g&\sim\mathrm{Categorical}(\boldsymbol{\pi}),\\\ \mathbf{c}&=a^{-1}(g),\\\ \log I_{l}\,|\,c_{l}&\sim\mathcal{N}(\mu_{l}[c_{l}],\sigma^{2}_{l}[c_{l}]).\end{split}$ (1) As discussed earlier, the intrinsic spot intensity noise (1) is multiplicative, (2) results from a multitude of independent sources, and (3) is uncorrelated across imaging rounds, motivating factorizing $\mathbf{I}_{1:L}\,|\,\mathbf{c}_{1:L}$ in $l$ and modeling each conditional as a Gaussian in the logarithmic space. The well-known heteroscedasticity of fluorescence noise is reflected in having two different $\sigma^{2}[c]$ for $c\in\\{0,1\\}$ for the two binary states. ### II-B Image processing and model fitting The most straightforward way to fit the generative model to empirical data is by observing that marginalizing the (discrete) molecule identity variable $g$ yields a two-component Gaussian mixture model (GMM) for $\log I_{l}$, with weights determined by the prior $\boldsymbol{\pi}$, codebook $\mathsf{C}$, and the assignment $a$. The model parameters $\\{\mu_{1:L}[0],\sigma^{2}_{1:L}[0],\mu_{1:L}[1],\sigma^{2}_{1:L}[1]\\}$ can be readily estimated by ML GMM fitting to each column of the spot intensity table (cf. Fig. 1c), which can be performed efficiently using the conventional EM algorithm. In order to decouple the intrinsic and extrinsic spot noise in the raw data, we censor the dataset to only spatially isolated molecules. In brief, we process the images as described in [1], subtract the background, identify the position of molecules by local peak-finding, censor dense regions (e.g. cell nuclei), and retain local peaks that are separated from one another at least by $\sim 5~{}\mathrm{px}$, which is a few multiples of the diffraction limit. We perform additional filtering based on the spot intensity pattern and nearest-neighbor Pearson correlation (cf. Fig. 2) and only retain peaks with a symmetric appearance. This procedure yields $\sim$ 250k spots in the dataset published in [2]. The obtained fits are shown in Fig. 3 along with QQ-plots that confirm a remarkably good fit to the empirical marginal histograms. ### II-C Quantization, channel model and estimation The generative model specified by Eq. (1) readily yields the posterior distribution $\mathrm{Pr}(g\,|\,\mathbf{I};\boldsymbol{\pi},\mathsf{C},a)$ and can form the basis of an intensity-based MAP decoder. To make the formulation more amenable for computational and theoretical investigation, as well as making a connection to the currently used decoding method, we derive an approximate binary asymmetric channel (BAC) model from Eq. (1) through quantization. The optimal quantization thresholds $\boldsymbol{\theta}_{1:L}$ are determined for each $l$ to be the point of equal responsibility between the two Gaussian components, i.e. $\sum_{g=1}^{G}\pi_{g}\,a^{-1}(g)[l]\,\mathcal{N}(\theta_{l}\,|\,\mu_{l}[1],\sigma_{l}^{2}[1])=\sum_{g=1}^{G}\pi_{g}\,[1-a^{-1}(g)[l]\,\mathcal{N}(\theta_{l}\,|\,\mu_{l}[0],\sigma_{l}^{2}[0])$, which amdits a closed-form solution. Here, $a$ and $\boldsymbol{\pi}$ correspond to the known code assignment and prior distribution of the data used for fitting. The fallout $p^{1\rightarrow 0}$ and false alarm $p^{0\rightarrow 1}$ rates are given by the integrated probability weights of the two Gaussian components below and above the threshold (cf. Fig. 3a), i.e. $p^{0\rightarrow 1}_{l}=\Phi[(\mu_{l}[0]-\theta_{l})/\sigma_{l}[0]]$ and $p^{1\rightarrow 0}_{l}=\Phi[(\theta_{l}-\mu_{l}[1])/\sigma_{l}[1]]$, where $\Phi(\cdot)$ is the CDF of the standard normal distribution. We find $p_{l}^{0\rightarrow 1}$ and $p_{l}^{1\rightarrow 0}$ to be $0.046$ and $0.102$ (mean in $l$), respectively, for the data given in Ref. [2], which is in agreement with the estimates reported therein. We, however, observed significant round-to-round variation in the channel parameters and as such, refrained from further simplifying the channel model to a single BAC for all imaging rounds $l$. We refer to the bundle of estimated BAC parameters as $\boldsymbol{\theta}_{\mathrm{BAC}}$. ### II-D Decoding: MAP and MLE decoders A gratifying property of the BAC approximation of Eq. (1) is allowing us to evaluate the performance of various decoding strategies without resorting to time-consuming simulations or further analytical approximations. In the BAC model, the likelihood of a binary sequence $\mathbf{x}_{1:L}\in\\{0,1\\}^{L}$ conditioned on the codeword $\mathbf{c}\in\mathsf{C}$ is given as: $\log\mathrm{Pr}(\mathbf{x}\,|\,\mathbf{c},\boldsymbol{\theta}_{\mathrm{BAC}})=\sum_{l=1}^{L}\sum_{i,j\in\\{0,1\\}}\,\delta_{c_{l},i}\,\delta_{x_{l},j}\,\log\,p_{l}^{i\rightarrow j},$ (2) where $\delta_{\cdot,\cdot}$ is the Kronecker’s delta function. We define the posterior Voronoi set for each codeword $\mathbf{c}\in\mathsf{C}$ as: $\mathsf{V}(\mathbf{c}\,|\,a,\boldsymbol{\omega},\mathsf{C},\boldsymbol{\theta}_{\mathrm{BAC}})=\big{\\{}\mathbf{x}\in\\{0,1\\}^{L}\,|\,\forall\mathbf{c}^{\prime}\in\mathsf{C},\mathbf{c}\neq\mathbf{c}^{\prime}:\\\ \omega_{a(\mathbf{c})}\,\mathrm{Pr}(\mathbf{x}\,|\,\mathbf{c},\boldsymbol{\theta}_{\mathrm{BAC}})>\omega_{a(\mathbf{c}^{\prime})}\,\mathrm{Pr}(\mathbf{x}\,|\,\mathbf{c}^{\prime},\boldsymbol{\theta}_{\mathrm{BAC}})\big{\\}},$ (3) where $\boldsymbol{\omega}_{1:G}$ is the prior distribution assumed by the decoder. The Voronoi sets are mutually exclusive by construction, can be obtained quickly for short codes by exhaustive enumeration, and determine the optimal codeword for an observed binary sequence. The MLE decoder corresponds to using a uniform prior, i.e. $\boldsymbol{\omega}\leftarrow\mathbf{1}/G$ whereas the MAP decoder corresponds to using the actual (non-uniform) prior governing the data $\boldsymbol{\omega}\leftarrow\boldsymbol{\pi}$. We additionally introduce a MAPq decoder, which is a MAP decoder obtained from depleting the Voronoi sets from binary sequences for which the posterior probability of the best candidate code is below a set threshold $q$. Intuitively, the MAPq decoder is a Bayesian decoder with reject option that trades precision gain for sensitivity loss by filtering dubious sequences from the Voronoi sets. The decoding algorithm introduced by Ref.[1, 2, 3] can be thought of as the MLE decoder with a rejection subspace given by $S_{\mathrm{rej}}=\\{\mathbf{x}\,|\,\exists\,\mathbf{c},\mathbf{c^{\prime}},\mathbf{c}\neq\mathbf{c}^{\prime}\in\mathsf{C}:d_{\mathrm{H}}(\mathbf{c},\mathbf{x})=d_{\mathrm{H}}(\mathbf{c}^{\prime},\mathbf{x})=d^{*}(\mathbf{x},\mathsf{C})\\}$ where $d_{\mathrm{H}}(\cdot,\cdot)$ is the Hamming distance and $d^{*}(\mathbf{x},\mathsf{C})=\inf_{\mathbf{c}\in\mathsf{C}}d_{\mathrm{H}}(\mathbf{c},\mathbf{x})$. We refer to this decoder as Moffitt (2016). We remark that the acceptance criterion of Moffitt (2016) is extremely stringent: for MHD4 codes, $|S_{\mathrm{acc}}|=9100$, which is only $\sim 13\%$ of all possible sequences (here, $S_{\mathrm{acc}}$ is the complement of $S_{\mathrm{rej}}$). In all cases, the confusion matrix $\mathcal{T}(\mathbf{c}\,|\,\mathbf{c}^{\prime})$, i.e. the probability that a molecule coded with $\mathbf{c}^{\prime}$ is decoded to $\mathbf{c}$, can be immediately calculated: $\mathcal{T}(\mathbf{c}\,|\,\mathbf{c}^{\prime};\boldsymbol{\pi},\boldsymbol{\omega},\boldsymbol{\theta}_{\mathrm{BAC}})=\sum_{\mathbf{x}\in\mathsf{V}(\mathbf{c}|\boldsymbol{\omega},\ldots)}\,\pi_{a(\mathbf{c}^{\prime})}\,\mathrm{Pr}(\mathbf{x}\,|\,\mathbf{c}^{\prime},\boldsymbol{\theta}_{\mathrm{BAC}})$ (4) from which the marginal true positive rates $\mathrm{TPR}_{1:G}$ and false discovery rates $\mathrm{FDR}_{1:G}$ can be readily calculated. ### II-E Comparing the performance of different decoders Figure 4: Comparing the performance of different decoding schemes for randomly assigned MHD4 codes. (a) and (b) correspond to prior distribution for RNA molecules selected in [2] and [3], respectively. The top panels show the rank- ordered prior distribution and the estimated Dirichlet concentration parameter $\alpha$; the middle and bottom panels show the marginal TPR and FDR for each molecule type conditioned on $S_{\mathrm{acc}}$ and $S_{\mathrm{rej}}$ subspaces (cf. Sec. II-D); markers are color-coded according to prior rank of their corresponding molecules. Shaded regions indicate 5-95 percentile range as a matter of random code assignment; (c) the effect of prior non-uniformity on the performance of MLE and MAP decoders for randomly assigned MHD4 codes. The top panels show the uniform mismatch rate. The bottom panels show the histogram of marginal FDRs vs. Dirichlet prior concentration $\alpha$ in grayscale. The orange lines and regions indicate the median and 5-95 percentile ranges; (d) performance of MAP decoders with reject at different acceptance thresholds compared to method in [2]. Developments in previous sections allow us to compare the performance of MLE, MAP, MAPq, and Moffitt (2016). We use the BAC parameters obtained from the data in [2], 16-bit MHD4 codes with random assignment, and two different previously estimated and published source priors with different degree of non- uniformity. As a first step, we compare the performance of our proposed MAP and MLE decoders separately inside $S_{\mathrm{acc}}$ and $S_{\mathrm{rej}}$, the acceptance and rejection subspaces of Moffitt (2016), in Fig. 4a, b (middle, bottom). The priors are shown on the top, including the estimated Dirichlet concentration $\alpha$. Marginal performance metrics for different molecules are color-coded according to their prior rank from red (most abundant) to blue (least abundant). The MLE decoder inside $S_{\mathrm{acc}}$ is equivalent to Moffitt (2016). Both decoders perform well in this subspace. While the MLE decoder is performing poorly inside $S_{\mathrm{rej}}$, providing a sound basis for rejection as in Moffitt (2016), the MAP decoder yields acceptable FDR, hinting that the $S_{\mathrm{acc}}$ is too stringent for the MAP decoder and better performance can be expected from MAPq. It also noticed that MAP decoder controls FDR much better than MLE inside $S_{\mathrm{acc}}$ for the more non-uniform prior. We explore this observation more systematically in panel c. We sample $\boldsymbol{\pi}$ from a symmetric Dirichlet distribution with concentration $\alpha$ and calculate the distribution of the marginal FDRs (bottom) as well as the uniform mismatch rate (top). We notice that as the prior gets more concentrated $\log\alpha\rightarrow-\infty$, the MAP decoder behaves progressively better whereas the MLE decoder degrades and exhibits a bi-modal behavior: extremely low (high) FDR on abundant (rare) codes. As the prior gets more uniform $\log\alpha\rightarrow+\infty$, MLE and MAP become indistinguishable. The green and red symbols show the biological priors used in panels a and b, respectively, together with their estimated $\alpha$, in agreement with the trend of the Dirichlet prior model. Finally, panel d compares the performance of the MAPq decoder at different rejection thresholds $q$ with Moffitt (2016). The prior used here is the same as in panel b. It is noticed that the MAPq decoder is remarkably effective at controlling FDR for all codes whereas Moffitt (2016) degrades in FDR for rare codes, as expected from the source uniformity assumption. This finding explains the reportedly lower correlation between rare molecules in replicate experiments [1, 2]. The smaller panels at the top of panel c show mean TPR, FDR, and rejection rate across all molecules. The MAP0.5 decoder has similar sensitivity to Moffitt (2016) while yielding $\sim 20\%$ lower FDR on average and remarkably $\sim 60\%$ lower 5-95 FDR percentile range, implying significant improvement in reducing the mean and variance of false positives for both abundant and rare molecules. ## III Data-driven code construction The results presented so far were obtained randomly assigning a fixed set of MHD4 codes. Constructing codes to better reflect channel asymmetry and prior non-uniformity is another attractive opportunity for improving the performance of mFISH protocols. Constructing application-specific codes for mFISH is outside the scope of the present paper and is a topic for future research. Here, we continue to thread on the theme of utilizing prior non-uniformity and show that optimizing the assignment of the even sub-optimal codes to molecules with respect to prior abundance can significantly reduce FDR. This is to be expected given the rather wide performance outcomes shown in Fig. 4 that result from random code assignment. Explicitly, we seek to optimize the scalarized metric $\overline{\mathrm{FDR}}(a,\boldsymbol{\pi})=G^{-1}\sum_{g=1}^{G}\mathrm{FDR}_{g}(a,\boldsymbol{\pi})$ over the assignment operator $a$ for a given prior $\boldsymbol{\pi}$ through an evolutionary optimization process. We start with a population of $N=5000$ random code assignments, mutate the population via pairwise permutations with a small probability of $0.05$ per molecule per assignment, and select the fittest $N$ offsprings using $\overline{\mathrm{FDR}}$ as the measure of fitness. We do not use a crossover operation here. We hypothesize that a relevant surrogate for the optimality of $\overline{\mathrm{FDR}}$ is the concordance between the Hamming distance $d_{H}$ and the prior distance $d_{\boldsymbol{\pi}}(\mathbf{c},\mathbf{c}^{\prime})\equiv|\pi_{a(\mathbf{c})}-\pi_{a(\mathbf{c}^{\prime})}|$. We investigate the emergence of this order by monitoring the following order parameter during the evolution: $\chi(a,\boldsymbol{\pi})\equiv\frac{1}{G}\sum_{g=1}^{G}\rho^{\mathrm{s}}\Big{[}d_{\mathrm{H}}\big{(}a^{-1}(g),\mathbf{C}_{a}\big{)},d_{\boldsymbol{\pi}}\big{(}a^{-1}(g),\mathbf{C}_{a}\big{)}\Big{]},$ (5) where $\rho^{s}[\cdot,\cdot]$ denotes the Spearman correlation and $\mathbf{C}_{a}$ is the ordered list of all codes used by $a$ over which the correlation is calculated. We refer to the population average of $\chi(a,\boldsymbol{\pi})$ as $\overline{\chi}$. We implement the evolutionary algorithm using the PyMOO package [7] and vectorize the calculation of Voronoi sets with GPU acceleration. Fig. 5 shows the results obtained by running the evolutionary optimization for three days (NVIDIA Testla P100 GPU, MHD4 codes, prior from [3]). Panel a shows the monotonic decline of $\overline{\mathrm{FDR}}$ to $\sim 75\%$ of its initial value (random assignment). This trend proceeds concurrently with a monotonic upturn in $\overline{\chi}$, providing evidence for the hypothesized matching order between $d_{\mathrm{H}}$ and $d_{\boldsymbol{\pi}}$. Panel b compares the performance metrics of the MAP decoder between the first and last population of code assignments. It is noticed that the optimized code assignment predominantly reduces $\overline{\mathrm{FDR}}$ of rare molecules, the mean FDR of which reduce to $\sim 50\%$ of randomly assigned codes. The possibility to reduce the FDR of rare molecules is a particularly favorable outcome in practice. Figure 5: Evolutionary optimization of code assignment for MHD4 codes (for channel model described in Fig. 3 and prior distribution from [3]). (a) bottom: mean FDR vs. generation; top: $d_{\mathrm{H}}-d_{\boldsymbol{\pi}}$ matching order parameter vs. generation (see Eq. 5); (b) the performance of MAP decoder for randomly assigned codes (squares) vs. optimized assignment (circles). ## IV Conclusion and Outlook In this paper, we reviewed multiplexed molecular profiling experiments from the perspective of coding theory, proposed a motivated generative model for the data, based on which we derived an approximate parallel BAC model for the system. We show that the exact MAP decoder of the BAC model vastly outperforms the decoding algorithm in current use in terms of controlling FDR of rare molecules, the key experimental metric. This is achieved by taking into account the non-uniformity of source prior, a “non-classical” aspect of multiplexed molecular profiling viewed as a noisy channel. We also took the first step in data-driven code construction and show that optimizing the assignment of existing sub-optimal codes is another effective method for reducing false positives. Attractive directions for follow up research include constructing application- specific codes to increase the throughput of the mFISH experiments, theoretical progress in understanding the optimal assignment of existing codes (e.g. by investigating the geometry of Voronoi sets), extending the generative model and the ensuing channel description to $q$-ary codes (e.g. as in seqFISH and seqFISH+ experimental protocols [4, 5]), and taking into account spatial interference and color channel cross-talk in the data generating process. ## References * [1] K. H. Chen, A. N. Boettiger, J. R. Moffitt, S. Wang, and X. Zhuang, “Spatially resolved, highly multiplexed rna profiling in single cells,” _Science_ , vol. 348, no. 6233, p. aaa6090, 2015. * [2] J. R. Moffitt, J. Hao, G. Wang, K. H. Chen, H. P. Babcock, and X. Zhuang, “High-throughput single-cell gene-expression profiling with multiplexed error-robust fluorescence in situ hybridization,” _Proceedings of the National Academy of Sciences_ , vol. 113, no. 39, pp. 11 046–11 051, 2016. * [3] J. R. Moffitt, J. Hao, D. Bambah-Mukku, T. Lu, C. Dulac, and X. Zhuang, “High-performance multiplexed fluorescence in situ hybridization in culture and tissue with matrix imprinting and clearing,” _Proceedings of the National Academy of Sciences_ , vol. 113, no. 50, pp. 14 456–14 461, 2016. * [4] S. Shah, E. Lubeck, W. Zhou, and L. Cai, “seqfish accurately detects transcripts in single cells and reveals robust spatial organization in the hippocampus,” _Neuron_ , vol. 94, no. 4, pp. 752–758, 2017. * [5] C.-H. L. Eng, M. Lawson, Q. Zhu, R. Dries, N. Koulena, Y. Takei, J. Yun, C. Cronin, C. Karp, G.-C. Yuan _et al._ , “Transcriptome-scale super-resolved imaging in tissues by rna seqfish+,” _Nature_ , vol. 568, no. 7751, pp. 235–239, 2019. * [6] C. E. Shannon, “A mathematical theory of communication,” _Bell system technical journal_ , vol. 27, no. 3, pp. 379–423, 1948. * [7] J. Blank and K. Deb, “pymoo: Multi-objective optimization in python,” 2020.
# Self-stabilizing Algorithm for Maximal Distance-2 Independent Set Badreddine Benreguia‡ Hamouma Moumen‡111Address for correspondence: H. Moumen, Department of Informatics, University of Batna 2, Fesdis 05078, Batna, Algeria. Soheila Bouam‡ Chafik Arar‡ ‡ University of Batna 2, 05000 Batna, Algeria <EMAIL_ADDRESS> ###### Abstract In graph theory, an independent set is a subset of nodes where there are no two adjacent nodes. The independent set is maximal if no node outside the independent set can join it. In network applications, maximal independent sets can be used as cluster heads in ad hoc and wireless sensor networks. In order to deal with any failure in networks, self-stabilizing algorithms have been proposed in the literature to calculate the maximal independent set under different hypothesis.In this paper, we propose a self-stabilizing algorithm to compute a maximal independent set where nodes of the independent set are far from each other at least with distance 3. We prove the correctness and the convergence of the proposed algorithm. Simulation tests show the ability of our algorithm to find a reduced number of nodes in large scale networks which allows strong control of networks. Keywords: Self-stabilizing algorithm, distributed system, network, independent set. ## 1 Introduction ### 1.1 Context of the study and motivation Self-stabilization is a fault tolerance approach that allows distributed systems to achieve a global correct configuration starting from an unknown initial configuration. Without external intervention, a self-stabilizing algorithm is able to correct the global configuration of the distributed system in a finite time. Various self-stabilizing distributed algorithms have been proposed in the literature using graph theory such as leader election, nodes coloring, domination problem, identifying the independent set, constructing the spanning tree. These algorithms have many benefits in the real-life applications, for example independent sets have been used as cluster heads in ad hoc and sensor networks [1, 2, 3, 12]. In graphs, independence is commonly defined as follow: let $G=(V,E)$ be a graph, where $V$ is the set of nodes and $E$ is the set of edges. An independent set is a subset of nodes $S\subset V$ such that there is no two adjacent nodes in $S$. The distance between any two nodes in $S$ is greater than 1. An independent set $S$ is said $maximal$, if there is no superset of $S$ that could be an independent set. In other words, there is no node outside the maximal independent set (MIS) that may join MIS. It is well known in graph literature that MIS is considered also as dominating set because every node out of MIS has at least a neighbor in MIS (every node outside MIS is dominated by a node of MIS). In this paper, we deal with a particular case of independent set. We call $S$ maximal distance-2 independent set (MD2IS), if of $S$ nodes are independent and the distance between any two nodes among them is strictly greater than 2. Figure 1 illustrates difference between MIS and MD2IS where green nodes are independent. Observe that in MIS (a), distance of 2 could be found between independent nodes. However, the distance between green nodes in MD2IS (b) is strictly greater than 2. Figure 1: (a) Maximal independent set. (b) Maximal distance-2 independent set. Nodes of MIS are used as servers (cluster heads) in ad hoc and wireless sensor networks to provide important services for other nodes. Each cluster head has to guarantee services for nodes connected to it, that are called members of the cluster. Cluster members represent nodes outside of MIS. A cluster head could serve its members by routing information, providing keys of encryption, giving names for members,… Figure 1(b) shows that elements of MD2SI could be used as cluster heads where members connected to the head could be within distance of 2. However, using MIS, members could not be located within distance more than 1. Obviously, MD2IS provides a more reduced number of clusters than MIS. The choice of the cluster heads is important in order to contribute in extending lifetime of wireless sensor and ad hoc networks. Using MD2IS rather than MIS as cluster heads could provide a good alternative in this sense especially that lifetime is the major problem of these networks. In addition to that and in order to deal with any possible failure, we use self-stabilizing algorithm that ensures reconstructing cluster heads, after the failure occurs, allowing the network still operational. Finding the maximal independent set (MIS) in graphs using self-stabilization paradigm was studied in literature for the first time by Shukla et al. in 1995 [16]. Authors have used a straightforward idea based on two rules: (1) a node $v$ joins the set $S$ (which is under construction) if $v$ has no neighbor in $S$, and (2) a node $v$ leaves the set $S$ if at least one of its neighbors is in $S$. Other variants of self-stabilizing algorithms constructing independent set have been introduced to deal with particular problems which try to minimize the algorithm complexity [18] or to be suitable for distributed daemon 222See section 2.1. [5, 8]. Reader can refer to the survey [6] for more details on MIS self-stabilizing algorithms. Other self-stabilizing algorithms have been proposed for independent sets imposing additional constraints besides to the independence condition. For example, [13] has presented an algorithm to discover the independent set where each node $u$ out of $S$ has at least a neighbor $v$ in $S$ such that $deg(v)>deg(u)$. In [2] authors propose a distributed self-stabilizing algorithm to find MIS using two-hop (distance-2) information in wireless sensor networks. ### 1.2 Related works [9, 10] has proposed a self-stabilizing algorithm to find the independent dominating set imposing a distance greater than $k$ between any two nodes of the independent set. Work [10] is an improvement of the memory management regarding the first one [9]. Every node outside the independent set is within distance $k$. [4] presented a self-stabilizing algorithm to compute a dominating set $S$ (which is not independent) where every node out of $S$ has to be distant from $S$ at most by $k$. Although the precedent algorithms have bounded complexity $O(n+k)$ in rounds, authors indicate that these algorithms might still never converge under the distributed daemon, since the daemon could ignore an enabled nodes. It is known in literature that: if the round complexity of a self-stabilizing algorithm is finite, this does not mean it converges [4]. Therefore, the computation of the convergence time still an open question [4, 9, 10] for independent (or dominating) set at distance $k\geq 2$. ### 1.3 Contribution In this paper, we propose a self-stabilizing algorithm to find maximal distance-2 independent set called MD2IS. We prove the correctness and the convergence of the presented algorithm. Using a serial central daemon, and starting from an arbitrary configuration, MD2IS reaches the correct configuration in a limit number of moves. The serial central daemon allows reaching the correct configuration in the worst case at $2n$ moves. Which means that MD2IS converges in $O(n)$ moves using a central daemon under expression distance-2 model. For distance-one model, our algorithm reaches the correct configuration in $O(nm)$ moves using distributed daemon, where $n$ is the nodes number and $m$ is the edges number. Proofs and simulation tests confirm the convergence of the proposed algorithm that provides smaller independent sets in large scale graphs. Having independent set more reduced is useful in many applications and allows more control on large scale networks. For example, the problem of locating an anonymous source of diffusion [14, 15] needs the placement of few nodes as observers [17]. A reduced number of nodes that occupy important locations in graphs, can be used as observers to detect the source of rumors in social networks. ### 1.4 Organization of the paper This paper is made up of five sections. Section 2 presents the model and the terminology used for self-stabilization. In section 3, we introduce the proposed self-stabilizing algorithm for finding a maximal independent set at distance-2. Simulation tests are conducted in section 4. Finally, section 5 concludes the paper. ## 2 Model and terminology Networks and distributed systems are modelled generally as an undirected graph $G=(V,E)$ where the units of processing represent the set of nodes $V$ and the links are the set of edges $E$. The neighborhood of a node $v\in V$ is defined as $N(v)=\\{u\in V:vu\in E\\}$. Usually, we say that two nodes $v,u$ are $adjacent$ if $u\in N(v)$. We define the neighborhood at distance-2 as $N(v)_{dist2}=N(v)\cup\\{e\in V:(\exists u\in N(v):e\in N(u))\\}$. It is known in the graph literature that the distance between two adjacent nodes is 1. Clearly, the $v$’s neighborhood at distance-2 gathers the set of nodes within distance of 1 and 2. A set $S\subset V$ is $independent$ if no two nodes of $S$ are adjacent. In other word, there is no two nodes in $S$ at distance 1. Generally, a set $S$ of nodes is $distance$-$k$ independent if every node in $S$ is distant at least $k+1$ to any other node of $S$ [9]. Consequently, a distance-2 independent set is a subset $S$ of $V$ where every two nodes of $S$ are at distance $>2$. Recall that in the usual case of MIS, each node in the graph is either independent or dominated $i.e.$ every node of MIS is independent and every node outside MIS is dominated [7]. In our case, every node $u$ out of MD2IS is either dominated by a node $v\in$ MD2IS where $dist(u,v)=1$ or is dominated by $v$ where $dist(u,v)=2$. Note that, a node $u$ out of MD2IS could be dominated by many nodes of MD2IS, for example it is possible to find a node $u$ dominated by $v1$ and $v2$ where $dist(u,v1)=1$ and $dist(u,v2)=2$. Definition : A $distance$-$2$ $independent$ $set$ of a graph $G(V,E)$ is a subset $S$ of $V$ such that the distance between any two nodes of $S$ is strictly greater than $2$. $S$ is $maximal$ if no superset of $S$ is also a distance-2 independent set. An algorithm is self-stabilizing if it can (1) reach a global correct configuration called $legitimate$ (2) during a finite time after it starts from an unknown configuration. When a self-stabilizing algorithm reaches the correct configuration it must stay inside the correct configuration (known as $closure$ condition). Hence, to show that an algorithm is self-stabilizing, it is sufficient to prove its $closure$ for the legitimate configuration and its $convergence$ to achieve the desired configuration in a finite time. In the uniform self-stabilizing system, all nodes execute the same code which is a set of rules having the form: if $guard$ then $statement$ (written as: $guard\longrightarrow statement$). In this case, nodes use the same local variables that describe their $state$. The guard is a (or a collection of) boolean expression. If a guard is evaluated to be true, the corresponding rule is said $enabled$ (or $priviliged$). We say that a node is $enabled$, if at least one of its rules is enabled. Executing the statement of the enabled rule by the node is called a $move$. An enabled node can make a move only if it is selected by a scheduler called a $daemon$. A move allows updating the local state (local variables) in order to make the node more legitimate with its neighborhood. ### 2.1 Daemon notion The execution of self-stabilizing algorithms is managed by a daemon (scheduler) that selects the enabled nodes to move from a configuration to another configuration. Two types of daemons are widely used in literature: central and distributed daemons. In the central daemons, only one enabled node is selected to be moved among all the enabled nodes. The central daemon, called also serial, allows enabled nodes executing a move (one by one) by a serial order. However, in the distributed daemons, a subset of nodes are selected among the set of privileged nodes to make a move simultaneously. The selected subset of nodes to be moved simultaneously forms a $round$. The distributed daemon is said $synchronous$ when all the enabled nodes are selected to move simultaneously. ### 2.2 Distance model Generally, most of the existing self-stabilizing algorithms, use the distance- one model wherein each node has a partial view on neighbors at distance one (through an access to the $state$ variable of the neighbors). However, there is other distance-2 models (like expression model) where every node has to get information of its neighborhood at distance-2. In the expression model which is particular case of distance-2 model, access to distance-2 is reached indirectly through access to $expression$ of neighbors. In the distance-2 model, hypothesis should be constructed carefully. For example, to the best of our knowledge, there is no algorithm of distance-2 that can operate under distributed daemon. The existing algorithms of distance-2 have been developed only under central daemon. ### 2.3 Transformers A common approach, known in literature [19, 20], allows converting a self- stabilizing algorithm $A$ operating under a given hypothesis to a new algorithm $A^{T}$, that operates under other hypothesis different from the first ones. However, the transformation guarantees that the two algorithms converge to the same legitimate configuration. Different kinds of transformers can be found in literature like distance transformers and daemon transformers. Generally, an overhead of the algorithm complexity is caused by the transformation which leads generally to a slowdown in the convergence of the transformed algorithms. In this paper, we use the transformer proposed by [19] that allows to transform any self-stabilizing algorithm operating under serial central daemon and expression distance-2 model to a self-stabilizing algorithm that operates under distributed daemon and distance-one model. ### 2.4 Our execution model In this paper, we develop a uniform self-stabilizing algorithm. In a first step, we suppose our algorithm operates under expression distance-2 model using central daemon. After that, we use transformer proposed by [19] that allows converting our algorithm to another algorithm that runs under distance- one model using distributed daemon. ## 3 Self-stabilizing algorithm MD2IS The proposed self-stabilizing algorithm presented in Algorithm 1, allows finding the maximal distance-2 independent set. Each node $v$ maintains a local variable $state$ and an expression $exp$. The $state$ variable could take one of the values $In$ or $Out$. Once the system reaches the legitimate configuration, all the nodes are disabled and the set $S=\\{v\in V:v.state=In\\}$ forms the maximal distance-2 independent set. However, in the illegitimate configuration, a serial moves of enabled nodes are executed until the global correct configuration is reached. At every move, an enabled node is selected randomly by the central daemon. Every node checks its state regarding its neighborhood using $state$ and $exp$. Expression $exp$ is used to calculate the number of the neighbors in $S$. Generally, the expression model allows discovering the neighborhood at distance-2. Once a node could read expressions of its neighbors, it will have an information about its neighbors at distance-2. In our algorithm, R1 shows that if a node $v$ out of $S$ reads the $state$ and the $exp$ of all its neighbors and finds that all the $state=Out$ and all the $exp=0$, hence all the neighbors at distance-2 are out, therefore node $v$ has to join the set $S$. Conversely, R2 illustrates that if a node $v\in S$ knows that there exists at least one of its neighbors $u$ such that $u.state=In$ or $u.exp>1$, thus $v$ has to leave $S$. Note that if $v\in S$ and $u.exp>1$ means that $u$ is dominated at least by two nodes: $v$ and another node. Observe also that if $v.state=IN$ and $u.exp=1$, this means that $u$ is dominated only by $v$ and it is impossible that $u.exp=0$ because there exists at least $v\in S$ as a neighbor of $u$. It is clear that R1 ensures the independence property because node $v$ will be independent at distance-2 by executing R1 leading to $v.state=In\wedge\forall u\in N(v):(u.state=Out\wedge u.exp=0)$. However, R2 guarantees that every node $v$ out of MD2IS is dominated at least by distance-2 because ($v.state=Out\wedge\exists u\in N(v):(u.state=In\vee u.exp>1)$). Figure 2 shows how our algorithm converges from an initial illegitimate configuration to a final legitimate configuration using a central serial daemon. Nodes outlined in red are (privileged) candidates to execute a move. The central daemon selects in every step only one privileged node to execute a move. Green nodes represents MD2IS which is reached after 5 moves by the sequence of rules: R2, R2, R2, R1, R1. Observe that in some cases, a move (for example transition from configuration c to configuration d) can make other nodes enabled in the next configuration. Figure 2: Convergence to the final legitimate MD2IS configuration. (1) $v.exp::|\\{u\in N(v):u.state=In\\}|$ (2) R1: $v.state=Out\wedge\forall u\in N(v):(u.state=Out\wedge u.exp=0)\longrightarrow v.state=In$ (3) R2: $v.state=In\wedge\exists u\in N(v):(u.state=In\vee u.exp>1)\longrightarrow v.state=Out$ Algorithm 1 Maximal Distance-2 Independent Dominating Set - MD2IS ### 3.1 Closure ###### Lemma 1 When all the nodes are disabled, the set $S=\\{v\in V,v.state=In\\}$ is maximal distance-2 independent set. Proof Suppose that the system is in a legitimate configuration. Since $R2$ is not enabled for nodes $v$ in $S$, the condition $(\exists u\in N(v):(u.state=In\vee u.exp>1)$ is false. Thus, $\forall u\in N(v):(u.state=Out\wedge u.exp\leq 1$). Therefore, at distance 1 from $v$, all the nodes are out of $S$. And, at distance-2 from $v$ there is no node in $S$ because $u.exp\leq 1$ which is exactly $=1$ ($u$ has at least a neighbor $v$ in $S$, Therefore, all the neighbors of $u$ are out $S$ except $v$. This implies that all the nodes in the neighborhood of $v$ at distance-2 are out of $S$). To show that $S$ is maximal, observe that if we want to add one node $v$ with $v.state=Out$ then $v$ will have not all the neighbors at distance-2 out of $S$ ($i.e.$ $v$ has at least a neighbor at distance-2 in $S$). Thus, the addition of this node will violate distance-2 independence. $\Box_{Lemma~{}\ref{lem1}}$ ### 3.2 Convergence ###### Lemma 2 If a node $v$ executes $R1$ becoming independent, it remains independent, and every node in its neighborhood at distance-2 still out of $S$ and cannot be enabled. Proof When a node $v$ executes $R1$, that means all the nodes $u$ at distance-2 from $v$ are out of $S$. It is clear that no one of $u$ could enabled $R1$ because there exists at least $v$ in $S$ as a node in the neighborhood of $u$ at distance-2 from $u$. Thus, $v$ will still in $S$ and all the neighbors at distance-2 still out of $S$. $\Box_{Lemma~{}\ref{lem2}}$ ###### Lemma 3 Any node of $V$ will be enabled at most twice by R2 then R1. Thus, MD2IS terminates in the worst case at $2n$ moves under the expression model using a central daemon. Proof Since Lemma 2 shows that every node executes R1, it cannot move again. It follows that any node could be enabled (in the worst case) by only R2 and then R1. Consequently, $2n$ moves is an upper bound to stabilize for $n$ nodes. $\Box_{Lemma~{}\ref{lem3}}$ ###### Theorem 1 MD2IS is a self-stabilizing algorithm that constructs Maximal Distance-2 Independent Set in $O(n)$ moves under expression model using a central daemon. Proof The proof follows from Lemma 1 and Lemma 3. $\Box_{Theorem~{}\ref{th1}}$ The last theorem shows that MD2IS stabilizes under the central daemon and expression model. Now, we use the transformer proposed by [19] that gives a corresponding self-stabilizing algorithm MD2ISD which operates under distributed daemon and distance-one model. ###### Theorem 2 MD2ISD converges into legitimate configuration in $O(nm)$ moves in the distance-one model under a distributed daemon. Where $m$ is the number of edges. Proof Using Theorem 1, the proof follows from Theorem 18 of [19], where $m$ is the number of edges. $\Box_{Theorem~{}\ref{th2}}$ ## 4 Simulation results In this section, simulation tests are carried out in order to evaluate MD2IS. We calculate the cardinality of the MD2IS with MIS, and then, we observe how the cardinality of MD2IS is related when the graph size grows. Although the comparison with MIS is unfair, there is no other algorithm that we could refer to evaluate MD2IS cardinality. The algorithms are written in Java using expression model for MD2IS. For MIS [5], we use the implementation of Lukasz Kuszner [11] in which we have generated arbitrary graphs with different density having sizes from 500 nodes to 20000 nodes. For each size of graphs, we have carried out 5 to 10 executions and then we have taken the average value. Figure 3: Cardinality of MD2IS and MIS according density using graphs of 1000 and 3000 nodes. Figure 3 shows the cardinality of MD2IS and MIS according graphs density. It is clear that MD2IS gives independent sets smaller than those produced by MIS. The density of the graphs has a clear impact on the cardinality of the independent sets. More the density grows, more the cardinality converges to be smaller. The important observation is that the cardinality of MD2IS will be close to 1 when the density becomes greater than 0.5. This is a rational result because for complete graphs (density=1), cardinality of MD2IS is 1. Figure 4: Cardinality of MD2IS and MIS according graphs size (Density=0.01). Figure 4 illustrates the cardinality of independent sets according size of graphs. Using constant density 0.01, curves show that MIS increases proportionally with the graph sizes. However, cardinality of MD2IS is inversely proportional to the graph sizes. For graphs of 10000 nodes, the cardinality of MIS is greater than 100 whereas the cardinality of MD2IS is less than 10 nodes. Graph size | MD2IS Cardinality | MIS Cardinality | MD2IS Convergence | MIS Convergence ---|---|---|---|--- 1000 | 601.8 (60.18%) | 687.8 (68.78%) | 400.2 | 415.6 1500 | 703.6 (46.91%) | 926.4 (61.76%) | 652.8 | 636.6 2000 | 758.2 (37.91%) | 1099.4 (54.97%) | 910.8 | 845.4 2500 | 794.4 (31.78%) | 1248.2 (49.93%) | 1229.2 | 1078.4 3000 | 786.6 (26.22%) | 1385.0 (46.17%) | 1570.8 | 1327.4 3500 | 776.4 (22.18%) | 1522.4 (43.50%) | 1877.2 | 1624.6 4000 | 762.2 (19.06%) | 1614.4 (40.36%) | 2202.8 | 1887.4 4500 | 748.8 (16.64%) | 1720.4 (38.23%) | 2483.6 | 2206.0 5000 | 724.2 (14.48%) | 1801.0 (36.02%) | 2785.8 | 2454.8 5500 | 711.0 (12.93%) | 1887.0 (34.31%) | 3048.4 | 2749.2 6000 | 692.2 (11.54%) | 1964.4 (32.74%) | 3335.6 | 3061.6 6500 | 662.8 (10.20%) | 2020.8 (31.09%) | 3551.6 | 3324.6 7000 | 642.6 (9.18%) | 2111.8 (30.17%) | 3825.2 | 3691,4 7500 | 633.8 (8.45%) | 2149.8 (28.66%) | 4098.8 | 3927.2 8000 | 604.6 (7.56%) | 2216.6 (27.71%) | 4333.4 | 4229.2 8500 | 601.8 (7.08%) | 2269.2 (26.70%) | 4586.4 | 4542.8 9000 | 577.6 (6.42%) | 2320.4 (25.78%) | 4836.0 | 4830.8 9500 | 552.8 (5.82%) | 2364.6 (24.89%) | 5097.2 | 5152.4 10000 | 548.4 (5.48%) | 2415.2 (24.15%) | 5350.8 | 5379.2 10500 | 527.4 (5.02%) | 2476.2 (23.58%) | 5645.0 | 5728.4 11000 | 501.6 (4.56%) | 2507.4 (22.79%) | 5817.6 | 6063.2 11500 | 498.2 (4.33%) | 2560.4 (22.26%) | 6091.0 | 6303.6 12000 | 481.2 (4.01%) | 2602.2 (21.69%) | 6300.6 | 6622.0 12500 | 477.8 (3.82%) | 2635.2 (21.08%) | 6555.8 | 6923.8 13000 | 454.6 (3.50%) | 2665.6 (20.50%) | 6790.0 | 7195.0 13500 | 448.0 (3.32%) | 2705.0 (20.04%) | 7057.6 | 7496.6 14000 | 435.2 (3.11%) | 2747.8 (19.63%) | 7251.6 | 7758.6 14500 | 427.4 (2.95%) | 2785.0 (19.21%) | 7512.0 | 8045.0 15000 | 416.2 (2.77%) | 2793.0 (18.62%) | 7798.4 | 8234.0 15500 | 408.2 (2.63%) | 2839.4 (18.32%) | 7988.8 | 8579.8 16000 | 398.8 (2.49%) | 2867.4 (17.92%) | 8293.6 | 8857.8 16500 | 392.0 (2.38%) | 2886.6 (17.49%) | 8496.2 | 9177.6 17000 | 383.4 (2.26%) | 2934.2 (17.26%) | 8798.4 | 9433.6 17500 | 380.2 (2.17%) | 2962.6 (16.93%) | 9043.4 | 9704.0 18000 | 367.6 (2.04%) | 2976.6 (16.54%) | 9249.4 | 9972.0 18500 | 361.0 (1.95%) | 3017.8 (16.31%) | 9481.0 | 10284.0 19000 | 355.2 (1.87%) | 3036.6 (15.98%) | 9748.6 | 10535.8 19500 | 353.0 (1.81%) | 3054.8 (15.67%) | 9998.8 | 10785.0 20000 | 340.4 (1.70%) | 3080.2 (15.40%) | 10207.8 | 11032.4 Table 1: Cardinality and time of convergence of MD2IS (on graphs of density=0.001). Further results of simulation are given in table 1 where the time of convergence is shown to be finite and variates proportionally to the graph size. Results confirm the lemma 3 that algorithm MD2IS converges at most in $2n$ moves although the simulations gives smaller number of moves that is close to $n/2$. In these tests, random graphs have been generated for small density = 0.001. We use this later value of density in order to be more close to real networks. For example, for a graph of 10000 nodes, a density of 0.001 will give an average degree = 10 for each node which models a user having 10 friends on social networks. The generated graphs have orders from 1000 nodes to 20000 nodes. We take the average value after carrying out 5 tests. ## 5 Conclusion In this paper we proposed a first self-stabilizing MD2IS algorithm that converges into the correct configuration in $O(n)$ moves using a central daemon under expression model. In the distance-one model, MD2ISD terminates in $O(nm)$ moves using a distributed daemon. The computation of the convergence time for independent sets of $k$ distance which still an open question is left for the future work. We plan also to evaluate the use of MD2IS as observers for the problem of locating sources propagating rumors on real graphs of social networks. ## References * [1] O. Arapoglu, V. K. Akram, and O. Dagdeviren. An energy-effcient, self-stabilizing and distributed algorithm for maximal independent set construction in wireless sensor networks. Computer Standards and Interfaces, 62-32; 251 42, 2019. * [2] O. Arapoglu and O. Dagdeviren. An Asynchronous Self-Stabilizing Maximal Independent Set Algorithm in Wireless Sensor Networks Using Two-Hop Information, International Symposium on Networks, Computers and Communications (ISNCC), Istanbul, Turkey, 2019, pp. 1-5. doi:10.1109/ISNCC.2019.8909189. * [3] D. Bein, A. K. Datta, C. R. Jagganagari, and V. Villain. A self-stabilizing link-cluster algorithm in mobile ad hoc networks. In 8th International Symposium on Parallel Architectures, Algorithms, and Networks, ISPAN, pages 436-441. IEEE Computer Society, 2005. * [4] A. K. Datta, S. Devismes, and L. L. Larmore. A silent self-stabilizing algorithm for the generalized minimal k-dominating set problem. Theoretical 259 Computer Science, 753:35-63, 2019. * [5] W. Goddard, S. T. Hedetniemi, D. P. Jacobs, and P. K. Srimani. Self-stabilizing protocols for maximal matching and maximal independent sets for ad hoc networks. In 5th IPDPS Workshop on Advances in Parallel and Distributed Computational Models, page 162. IEEE Computer Society, 2003. doi:10.1109/IPDPS.2003.1213302. * [6] N. Guellati and H. Kheddouci. A survey on self-stabilizing algorithms for independence, domination, coloring, and matching in graphs. Journal of Parallel and Distributed Computing, 70(4):406-415, 2010. * [7] S. Hedetniemi, S. Hedetniemi, D. Jacobs, and P. Srimani. Self-stabilizing algorithms for minimal dominating sets and maximal independent sets. Computer and Mathematics with Applications, 46(56):805-811, 2003. * [8] M. Ikeda, S. Kamei, and H. Kakugawa. A space-optimal self-stabilizing algorithm for the maximal independent set problem. In 3rd International Conference on Parallel and Distributed Computing, Applications and Technologies, pages 70-74, 2002. * [9] C. Johnen. Fast, silent self-stabilizing distance-k independent dominating set construction. Information Processing Letters, 114:551-555, 2014. * [10] C. Johnen. Memory efficient self-stabilizing distance-k independent dominating set construction. In Networked Systems - Third International Conference, pages 354-366. Springer International Publishing, 2015. * [11] L. Kuszner. Tools to develop and test self-stabilizing algorithms. http://kaims.eti.pg.gda.pl/ kuszner/self-stab/main.html, 2005. * [12] A. Larsson and P. Tsigas. A self-stabilizing (k,r)-clustering algorithm with multiple paths for wireless ad-hoc networks. In 31 International Conference on Distributed Computing Systems, pages 353-362. IEEE, 2011. * [13] B. Neggazi, N. Guellati, M. Haddad, and H. Kheddouci. Efficient self-stabilizing algorithm for independent strong dominating sets in arbitrary graphs. International Journal of Foundations of Computer Science, 26(6):751-768, 2015. * [14] P. C. Pinto, P. Thiran, and M. Vetterli. Locating the source of diffusion in large-scale networks. Physical Review Letters, 109(6):068702, 2012. * [15] E. Seo, P. Mohapatra, and T. Abdelzaher. Identifying rumors and their sources in social networks. In Proceedings of SPIE - The International Society for Optical Engineering, volume 8389, 2012. * [16] S. K. Shukla, D. J. Rosenkrantz, and S. S. Ravi. Observations on self-stabilizing graph algorithms for anonymous networks. In 2nd Workshop on Self-Stabilizing Systems, 1995. * [17] B. Spinelli, L. E. Celis, and P. Thiran. Observer placement for source localization: The effect of budgets and transmission variance. In 54th Annual Allerton Conference on Communication, Control, and Computing , pages 743-751, 2016. * [18] V. Turau. Linear self-stabilizing algorithms for the independent and dominating set problems using an unfair distributed scheduler. Information Processing Letters, 103(3):88-93, 2007. * [19] V. Turau. Efficient transformation of distance-2 self-stabilizing algorithms. Journal of Parallel and Distributed Computing, 72(4):603-612, 2012. * [20] W. Goddard, P. K. Srimani, Daemon conversions in distributed selfstabilizing algorithms, in: Ghosh S.K., Tokuyama T. (eds) WALCOM: Algorithms and Computation. WALCOM 2013. Lecture Notes in Computer Science, Springer, Berlin, Heidelberg, 2013, pp. 146–157.
# Constraints on Planets in Nearby Young Moving Groups Detectable by High- Contrast Imaging and Gaia Astrometry A. L. Wallace${}^{\href https://orcid.org/0000-0002-6591-5290 \,1}$, M. J. Ireland${}^{\href https://orcid.org/0000-0002-6194-043X\,1}$, C. Federrath${}^{\href https://orcid.org/0000-0002-0706-2306\,1}$ 1Research School of Astronomy & Astrophysics, Australian National University, Canberra, ACT 2611, Australia E-mail<EMAIL_ADDRESS><EMAIL_ADDRESS> (Accepted XXX. Received YYY; in original form ZZZ) ###### Abstract The formation of giant planets can be studied through direct imaging by observing planets both during and after formation. Giant planets are expected to form either by core accretion, which is typically associated with low initial entropy (cold-start models) or by gravitational instability, associated with high initial entropy of the gas (hot-start models). Thus, constraining the initial entropy can provide insight into a planet’s formation process and determines the resultant brightness evolution. In this study, we find that, by observing planets in nearby moving groups of known age both through direct imaging and astrometry with Gaia, it will be possible to constrain the initial entropy of giant planets. We simulate a set of planetary systems in stars in nearby moving groups identified by BANYAN $\Sigma$ and assume a model for planet distribution consistent with radial velocity detections. We find that Gaia should be able to detect approximately 25% of planets in nearby moving groups greater than $\sim 0.3\,M_{\text{J}}$. Using 5$\sigma$ contrast limits of current and future instruments, we calculate the flux uncertainty, and using models for the evolution of the planet brightness, we convert this to an initial entropy uncertainty. We find that future instruments such as METIS on E-ELT as well as GRAVITY and VIKiNG with VLTI should be able to constrain the entropy to within 0.5 $k_{B}$/baryon, which implies that these instruments should be able to distinguish between hot and cold-start models. ###### keywords: gaseous planets – formation – detection ††pubyear: 2015††pagerange: Constraints on Planets in Nearby Young Moving Groups Detectable by High-Contrast Imaging and Gaia Astrometry–Constraints on Planets in Nearby Young Moving Groups Detectable by High-Contrast Imaging and Gaia Astrometry ## 1 Introduction In order to observationally constrain the formation and evolution of planetary systems, it is necessary to observe planets during or shortly after formation. While there has been some success with discovering planets in the process of formation (e.g., the PDS 70 system; Keppler et al. (2018); Benisty et al. (2021)), recent direct imaging surveys of the nearest star-forming regions, despite discoveries of new brown dwarf companions, highlighted a low frequency of wide orbit planetary mass companions (Kraus & Ireland, 2011; Wallace et al., 2020). The low number of positive detections of planets in wide orbits combined with sensitivity limits has produced upper limits on the frequency of giant planets (Bowler & Nielsen, 2018) and constraints on formation models (Nielsen et al., 2019; Vigan et al., 2020). From predicted occurrence rates, it has also been determined that current instruments have insufficient sensitivity at the expected separations to detect planets around solar-type stars in nearby star-forming regions (Wallace & Ireland, 2019). However, there has been greater success with wide-separation planets around high-mass stars in young nearby moving groups such as $\beta$-Pictoris (Lagrange et al., 2009) and 51-Eridani (Macintosh et al., 2015). Nearby moving groups have been studied in detail over the years (Torres et al., 2008; Zuckerman et al., 2011; Rodriguez et al., 2013) and recently, precise proper motion and parallax measurements of nearby stars have allowed reasonably accurate determination of membership to these groups (Gagné et al., 2018; Schneider et al., 2019). There are at least 27 such associations within 150 pc with ages less than $\sim$800 Myr. The young ages and small distances of these systems make them ideal for young planet surveys (López-Santiago et al., 2006). The upcoming Gaia DR3 and subsequent data releases promise high-precision mass calculations for many giant, long-period exoplanets (Perryman et al., 2014). A recent study of HR 8799 has already delivered results (Brandt et al., 2021) and measured the mass of HR 8799 e. However, Gaia’s expected 5–10 year mission lifetime puts an upper limit on the semi-major axes of Gaia-detectable planets with non-degenerate solutions, which limits the possibilities for high- constrast imaging studies. In order to conduct high-contrast imaging studies of Gaia-detectable planets, we must observe young planets in nearby systems. Planets in the process of formation radiate with a luminosity proportional to their accretion rate and total mass (Zhu, 2015). The amount of energy radiated away during formation has an effect on the internal entropy of the planet. If the accretion shock radiates all accretion energy away, the planet will form with low entropy (cold-start). If none of the accretion energy is radiated away, the planet will have high entropy (hot-start) (Berardo et al., 2017). The values of internal entropy corresponding to hot and cold-starts depend on planetary mass (as shown in the ‘tuning fork’ diagram from Marley et al. (2007).) Hot-start planets are assumed to form quickly whereas cold-start planets gradually accrete gas through the accretion shock (Mordasini et al., 2012). Thus, the initial entropy of a planet can indicate its formation conditions. After formation, planets cool and fade over time, but the rate of cooling depends on their internal entropy (Spiegel & Burrows, 2012). A hot-start planet will be brighter than a cold-start planet shortly after formation. The brightness of a planet can then be used to determine the initial entropy. As planets age, the luminosity decreases at a rate dependant on initial entropy. The luminosity of hot-start planets decreases faster than cold-start planets, which means, as planets age, information about the initial entropy is lost. However, if planets of known mass are observed at young ages, it should be possible to infer the initial entropy based on the observed flux (Marleau & Cumming, 2014). If the mass and age of a planet is known with reasonable precision, the greatest uncertainty results from the observed flux measurement. This flux uncertainty depends on the sensitivity of our instruments. In this study, we consider current instrument such as NIRC2, NaCo, SPHERE and GRAVITY and future instruments such as JWST, VIKiNG (interferometric instrument using VLTI), MICADO on the VLT and METIS on the E-ELT. These instruments have observed and theoretical detection limits which we convert to a flux uncertainty. Using models linking flux, mass and age to initial entropy, we convert this to an entropy uncertainty to determine how well the initial entropy can be constrained. The rest of the paper is organised as follows. Section 2 summarises our stellar and planet sample and models for the evolution of planet luminosity and its dependence on entropy. Section 3 focuses on detection limits of astrometry and direct imaging and Section 4 presents the numbers of detectable planets in our sample by both methods. Section 5 explains how we can constrain the initial entropy of planets if we know the mass, magnitude and age. Our conclusions are presented in Section 6. ## 2 Stellar and Planetary Properties ### 2.1 Stellar Sample Our stellar sample comes from nearby young (<800 Myr) moving groups which are promising targets for planet surveys. The stars are initially selected from Gaia’s second data release (DR2) (Brown et al., 2018) and then sorted into moving groups using the BANYAN $\Sigma$ from Gagné et al. (2018). Our initial sample from Gaia DR2 included stars across the entire sky within 70 pc, brighter than a G-magnitude of 13 and temperature greater than 3800 K. Beyond $\sim$70 pc the majority of giant planets in $\sim$10 year orbits are not detectable and moving group knowledge is less complete. Stars cooler than 3800 K are no-longer considered solar-type in this paper as they have a measured smaller fraction of planets (Johnson et al., 2007). The magnitude cutoff excludes stars considered too faint for reliable adaptive optics. The moving group membership of each star is then determined by the BANYAN $\Sigma$ algorithm. This process uses each star’s celestial coordinates, distance and proper motion values and associated errors to determine the probability of membership to a particular moving group from a list of 27 possible groups. A star that does not belong to any group with more than 70 % probability is discarded from the sample. A partial sky map of our stellar sample, indicating group membership is shown in Figure 1. This map only includes targets with a membership probability greater than 95%. Figure 1: Stars in our sample belonging to moving groups determined by BANYAN $\Sigma$. Map is shown in celestial coordinates with East to the left. Although the BANYAN $\Sigma$ algorithm can sort stars into 27 moving groups, there are only 10 groups with members closer than 70 pc as shown in Figure 1. Our target stars are spread across the sky, but the majority is in the southern hemisphere. The Argus association has the highest number of targets, but the existence of this group has been controversial (Bell et al., 2015), as it was unclear whether it represented a single moving group. However, recent studies have suggested the association does indeed exist with an age of 40–50 Myr (Zuckerman, 2018) so we include it in this study. Our resultant sample contains 1760 stars across 10 moving groups. The distributions of distance and mass of the stars in our sample is shown in Figure 2. (a) Distribution of Stellar Distance (b) Distribution of Stellar Mass Figure 2: Distribution of star distance and mass in our sample. Note the distance cuts off at 70 pc as we excluded stars further away than this. Most of our targets are at distances of $\sim$40–60 pc and low mass (<0.6 M⊙). A colour-magnitude diagram is shown in Figure 3 which plots absolute G-magnitude against effective temperature. Figure 3: Colour-magnitude diagram of our targets. The temperature scale cuts off at 3800 K as we don’t consider stars cooler than this. As shown in Figures 2 and 3, the majority of our targets are cool, low-mass stars which, in principle should make it easier to detect planets by astrometry. ### 2.2 Planet Distribution For each star in our sample, we simulate a system of planets with properties sampled from a broken power law in mass $M$ and period $P$ taken from Fernandes et al. (2019), which has the functional form $\frac{d^{2}N}{d\mathrm{ln}M\,d\mathrm{ln}{P}}=CM^{\alpha}P^{\beta},$ (1) where $N$ is the number of planets per star and $C$ is a normalisation constant. We assume the mass power law is approximately consistent across all masses with $\alpha=-0.45$. Based on radial velocity detections, the period power law changes with distance from the star with $\beta>0$ at short periods and $\beta<0$ for long periods. This broken power law is also consistent with theoretical core formation expectations (Mordasini et al., 2009). The study presented in Fernandes et al. (2019) gives several different values, but in this study we use the symmetric distribution in which $\beta=0.63$ for periods less than 860 days and $\beta=-0.63$ for periods greater than 860 days. The constant $C$ is set such that the total number of planets is consistent with observations. In the symmetric distribution from Fernandes et al. (2019), assuming a Sun-like star, $C=0.067$. We also note that this distribution is consistent out to 4 au with the models based on the California Legacy Survey (Fulton et al., 2021). When simulating our planet samples, we apply the distribution from Equation 1 in terms of semi-major axis $a$. This changes the power-law index to 0.945 at small separations and -0.945 at wide separations (multiplying by a factor of 3/2.) The power law changes at a fixed period of 860 days, which corresponds to a semi-major axis of 1.77 au for a 1 M⊙ star and scales with $M_{\star}^{1/3}$. This distribution implies the majority of planets are low mass and at small separations. While there have been high-mass planets observed at wide separations around HR 8799, $\beta$-Pic and 51 Eri, these are around high-mass stars where there is known to be an excess of high-mass planets (Johnson et al., 2010). The total number of planets is assumed to increase with stellar mass and some studies (e.g. Bowler et al. (2009)) suggest a steep relationship. However, for simplicity, we assume the number of planets scales linearly with stellar mass, as implied by Mulders (2018), and the normalisation constant $C$ is simply multiplied by $M_{\star}/M_{\odot}$. We simulate planet masses over a range of 0.3–13 MJ and semi-major axes over a range of 1–10 au. Integrating the power law shown in Equation 1 over this range gives $N\sim 0.07$ planets per star. The symmetric planet distribution from Fernandes et al. (2019) is shown in Figure 4. Period is converted to semi-major axis by assuming a 1 M⊙ star. Figure 4: The differential mass and semi-major axis distribution of our simulated planets from Fernandes et al. (2019). Integrating this across our simulation range gives $N=0.08$ planets per star. ### 2.3 Planet Luminosity and Magnitude As giant planets accrete gas, the accretion energy is emitted as radiation with a luminosity proportional to the accretion rate (Zhu, 2015) and are at their brightest during the period of runaway accretion (Lissauer et al., 2009). After a planet has formed, the luminosity declines over time and is dependant on the planet’s mass and entropy (Spiegel & Burrows, 2012). The conditions of planet formation have an effect on the post-formation entropy and luminosity. If the accretion rate is slow, the planet has more time to radiate energy away, resulting in a lower internal entropy (cold- starts). Planets with low internal entropy are typically smaller and cooler and thus have lower post-formation luminosity than planets formed by hot- starts. The luminosity as a function of age is shown in Figure 5 for planets of varying mass and entropy of 9.5 $k_{B}$/baryon (blue curves) and 10.5 $k_{B}$/baryon (red curves) using hot and cold-start models from Spiegel & Burrows (2012). This is calculated from applying the Stefan-Boltzmann Law to the radius and temperature plots in Figure 5 of their paper and interpolating between their initial entropies. Figure 5: Evolution of planet luminosity for different planet masses and initial entropies. The blue curves represent initial entropy of 9.5 $k_{B}$/baryon and red represents 10.5 $k_{B}$/baryon. This is calculated using the temperature and radius curves from Figure 5 in Spiegel & Burrows (2012), applying the Stefan-Boltzmann Law and interpolating between entropies shown in their paper. The absolute magnitude evolution in near-infrared bands is also calculated in order to determine planet detectability. This is also calculated using models from Spiegel & Burrows (2012), taking the magnitudes from their Figure 7 and interpolating between initial entropies. Some example magnitude evolution curves are shown in Figure 6 for the K (2.2$\mu$m) and L’ (3.77$\mu$m) bands using the same initial entropies as Figure 5. (a) Absolute Magnitude in K Band (b) Absolute Magnitude in L’ Band Figure 6: Evolution of planet magnitude in the K band (panel a) and in the L’ band (panel b) for different planet masses and initial entropies. The blue curves represent initial entropy of 9.5 $k_{\rm B}$/baryon and red represents 10.5 $k_{\rm B}$/baryon. This is based on the curves in Figure 7 of Spiegel & Burrows (2012) and, as before, we interpolate between entropies. The cooling curves in Figures 5 and 6 demonstrate that hot-start planets are initially brighter, but the luminosity declines faster. This implies that, at old ages, hot and cold-start planets become less distinguishable as the luminosities become approximately equal. However, at the young ages ($\sim$30 Myr) we consider in this study, there is a noticeable difference between the two entropies for high-mass planets, which indicates we should be able to constrain the formation models of massive ($>2\,\mathrm{M_{J}}$) planets. ## 3 Planet Detectability ### 3.1 Detection by Gaia Astrometry The upcoming Gaia data releases are expected to have improved astrometry measurements of stars and the potential to discover exoplanets through this method. As a planet with $M_{p}\ll M_{\star}$ and its host star of mass $M_{\star}$ orbit their common centre of mass, the astrometric semi-major axis, also known as the astrometric signature, is given by: $\alpha=\left(\frac{M_{p}}{M_{\star}}\right)\left(\frac{a_{p}}{d_{\star}}\right),$ (2) where $a_{p}$ is the semi-major axis of the planet in au and $d_{\star}$ is the distance to the star in pc. The planet detection capability depends on the signal to noise ratio, given by $S/N=\frac{\alpha}{\sigma_{\rm{fov}}},$ (3) where $\sigma_{\rm{fov}}$ is the accuracy per field of view crossing. The study in Perryman et al. (2014) concluded that a detection threshold of $S/N>2$ provides a reasonable estimate of planet detection numbers. The study in Ranalli et al. (2018) simulated the performance of Gaia and determined there is a 50% chance of detecting planets in the 5 and 10 year missions with $S/N>2.3$ and $1.7$ respectively. For this study, we assume a planet is detectable if $S/N>2$. The value of $\sigma_{\rm{fov}}$ depends on the star’s G magnitude, but is approximately constant at 34.2 $\mu$as for stars brighter than magnitude 12. In this study, we use the values from Table 2 in Perryman et al. (2014) and assume planets are detectable if $\alpha>2\sigma_{\rm{fov}}$. Gaia’s nominal mission is for 5 years with a possible extension to 10 years and this also places a constraint on detection capabilities. It is assumed that planets with periods greater than $\sim$10 years will be poorly constrained. ### 3.2 Detection by Direct Imaging Direct imaging has had a lower yield for planet detection than other techniques due to the high contrast ratios between stars and planets. Unlike transit and radial velocity, this method is more sensitive to planets at wide separations, which are less abundant. Planets at wide separations are typically found using a combination of angular differential imaging (ADI) and reference star differential imaging (RDI) (Marois et al., 2006; Lafreniere et al., 2007) to remove instrumental artefacts combined with coronagraphy to block out the light of the central star. Direct imaging methods are less sensitive at small separations, but this is gradually improving with new analysis methods such as kernel phase (Martinache, 2010) and observational methods such as interferometry (Lacour et al., 2019). The detection capability of an instrument is limited by both the resolution and the maximum signal to noise ratio. This provides a ‘contrast limit’, which typically improves with distance from the star. The experimental and theoretical contrast limits for current and future instruments are shown in Figure 7, assuming a stellar apparent magnitude of 7 in the K and L’ bands, which is close to the average magnitude of our targets, and an average distance of 50 pc. These limits were determined by observations or, in the case of future instruments, simulated performance. The SPHERE limit is derived from the recent SHINE survey (Langlois et al., 2021) and the MICADO limits come from Perrot et al. (2018). The GRAVITY limit is based on the lower limits of the curves in Abuter et al. (2019), assuming a 1 hour integration time. The NaCo limits come from the study presented in Quanz et al. (2012) and the METIS limits are based on Carlomagno et al. (2020). The NIRC2 limit comes from our contrast limit through recent observations of Taurus (Wallace et al., 2020), which are consistent with vortex coronagraph reference star differential imaging limits (Xuan et al., 2018). The JWST limit is adapted from Carter et al. (2021). The VIKiNG limits for L’ come directly from the model assuming 1 hour integrations by Martinache & Ireland (2018), assuming that the companions must be resolved in the kernel null maps, and with a loss in contrast as the planets approach the edge of the telescope PSF (at separations of 0.5-0.8$\lambda$/D). The assumed contrast of $4\times 10^{-5}$ at 5-$\sigma$ assumes either 80 nm RMS fringe tracking errors and 10% intensity fluctuations, or 120 nm RMS fringe tracking errors and 2% intensity fluctuations. Note that the Hi-5 instrument (Defrère et al., 2018) operating in L’ is compatible with the VIKiNG architecture, but may require longer integration times for the assumed contrast, depending on the finally adopted architecture. The VIKiNG limits for K are based on a more optimistic, but theoretically possible set of assumptions. For observations with the UTs, fringe tracking up to 0.5" off-axis is assumed with an RMS fringe tracking RMS error of 30 nm. The contrast limit shown is the magnitude difference between the faintest detectable planet and its star in the K and L’ bands. (a) K band (b) L’ band Figure 7: Assumed limits for current and future instruments as a function of angular and physical separation (assuming an average distance of 50 pc.) See text for detailed assumptions. Contrasts of planets of various masses are shown for comparison, assuming an age of 50 Myr, initial entropy of 10 $k_{B}$/baryon and stellar magnitude of 7. Figure 7 demonstrates that interferometry with ViKING and GRAVITY as well as instruments on large telescopes, such as MICADO and METIS are the most sensitive at small ($<$10 AU) separations. According to the distribution shown in Figure 4, this is where planets are most likely to occur. The apparent background-limited magnitude limits of these instruments are shown in Table 1, which apply mostly to faint targets. We consider a planet to be detectable if it is brighter than both the apparent magnitude limit and has higher contrast than the limit shown in Figure 7. Table 1: Apparent magnitude limits of instruments considered in this paper. Note that for most targets, the contrast and not these background limits most influences detectability. Instrument | K Mag. Limit | L’ Mag. Limit ---|---|--- SPHERE | 22.5 | - GRAVITY | 19.5 | - MICADO | 29.5 | - VIKiNG (UTs) | 23.2 | 18.5 VIKiNG (ATs) | 19.9 | 15.3 NIRC2 | - | 17.94 NaCo | - | 18.55 METIS | - | 21.8 JWST | - | 23.8 ## 4 Simulated Detectable Planets For each star in our sample shown in Section 2.1, we simulate a set of planet systems using the power-law distribution shown in Section 2.2. The normalization constant $C$ is scaled linearly with stellar mass. This simulation was run 5000 times per star, but only a small percentage of these simulations produced planets. This is due to the integration of the planet distribution shown in Figure 4. For a 1 M⊙ star, only 8% of simulations produced a planet more massive than 0.3 MJ. For simplicity, we assume circular orbits. ### 4.1 Planets Detectable by Gaia Using our simulated sample of planets around stars in nearby moving groups, we calculated how many can be detected by Gaia applying the methods detailed in Section 3.1. We assumed Gaia can detect planets with periods shorter than 10 years provided the astrometric signature $\alpha>2\sigma_{\rm{fov}}$ (Perryman et al., 2014; Ranalli et al., 2018). The average number of planets per group and the average number of planets detectable by Gaia are shown in Table 2, as a result of running the simulation 100 times. Table 2: Total number of simulated planets in each group and number detectable by Gaia. There are $\sim$0.06 planets per star and Gaia can detect approximately $\sim$30% of these. | | | | Average Number | Average Number ---|---|---|---|---|--- Group Name | Age (Myr) | Average Distance (pc) | Number of Stars | of Planets | of Detectable Planets AB Doradus | 149${}^{+51}_{-19}$ | 43.19 | 367 | 19.05 | 6.79 Argus | 45$\pm$5 | 48.33 | 630 | 35.10 | 11.32 $\beta$ Pictoris | 22$\pm$6 | 39.60 | 149 | 7.82 | 2.91 Carina | 45${}^{+11}_{-7}$ | 49.22 | 26 | 1.41 | 0.42 Carina-Near | $\sim$200 | 34.28 | 148 | 7.50 | 3.23 Columba | 42${}^{+6}_{-4}$ | 47.47 | 79 | 4.84 | 1.43 Hyades | 750$\pm$100 | 46.67 | 239 | 14.97 | 4.47 Tucana-Horologium | 45$\pm$4 | 49.34 | 94 | 5.31 | 1.58 TW Hydrae | 10$\pm$3 | 53.68 | 21 | 1.00 | 0.32 Ursa Major | $\sim$414 | 25.30 | 7 | 0.64 | 0.25 We have 1760 stars in our sample across all groups shown in Table 2 and, on average, we simulate $\sim$ 98 giant planets ($M>0.3$ MJ) across all groups ($\sim$ 0.056 planets per star.) Our estimates suggest that Gaia is able to detect $\sim$ 33 planets, which is approximately a third of our sample. ### 4.2 Planets detectable by both Gaia and Direct Imaging Using planetary cooling curves shown in Figure 6 and the limits of current and future instruments shown in Figure 7, we determine how many of the 25 Gaia- detectable planets are also detectable by direct imaging. This requires an estimate of the age and initial entropy of each planet. The planet age is assumed to be equal to the age of the moving group, listed in Table 2, within observational uncertainties and assume a range of initial planet entropies from 8.5–11.5 $k_{B}$/baryon. The possible inclination distribution is taken into account by calculating the averaged projected separation. This is simply calculated by multiplying the simulated semi-major axis by 0.8 as shown in Equation 7 of Fischer & Marcy (1992). Figure 8 shows the number of detectable planets for different instruments in the K and L’ bands as a function of initial entropy. (a) K band (b) L’ band Figure 8: Total number of planets detectable by both Gaia astrometry and high contrast imaging across all moving groups. Our results suggest that VIKiNG and METIS should be able to detect more than 4 Gaia-detectable planets regardless of initial entropy, while MICADO should be able to detect hot-start planets, if a survey of nearby moving groups were conducted. Note that VIKiNG with the ATs can detect more planets than the UTs in the L’ band despite being less sensitive. This is simply due to the wider range of separations detectable by interferometry with the ATs in the L’ band as shown in Figure 7(b). ## 5 Constraining the Initial Entropy As explained in Section 2.3, the initial entropy of a planet is related to its formation conditions and has an effect on the brightness evolution. If the age and mass of a planet is known to reasonable precision, it should be possible to constrain the initial entropy of a directly imaged planet. ### 5.1 Dependence of Magnitude on Entropy As shown in Figure 6, the magnitude evolution depends on its initial entropy, but this dependence decreases with age. Figure 9 shows the absolute magnitude as a function of initial entropy for a variety of planet masses and ages using models from Spiegel & Burrows (2012). (a) K band (b) L’ band Figure 9: Absolute Magnitude as a function of initial entropy. Solid, dashed and dotted curves are for planet ages of 20 Myr, 40 Myr and 80 Myr respectively. The curves shown in Figure 9 flatten as the planets age, but also at higher entropy, particularly in the L’ band. This implies that, while planets with higher initial entropy will be brighter and easier to detect, it could be harder to constrain the initial entropy of these planets. In order to determine how well different instruments can constrain entropy, we calculate the entropy uncertainty. ### 5.2 Entropy Uncertainty The entropy uncertainty was calculated for a set of simulated planets detectable by Gaia as shown in Table 2. Each planet is assigned a mass and semi-major axis from our distribution, but the initial entropy is unknown. The likelihood of a particular entropy given our simulated data, $L(S|D)$, is derived from Bayes Theorem, $L(S|D)\propto P(D|S)P(S),$ (4) where $P(D|S)$ is the probability of the data for a given entropy and $P(S)$ is the prior probability of that entropy. The probability as a function of entropy is calculated for a range of entropies $S_{i}$ using a Gaussian distribution in planet flux, given by $P(D|S_{i})\propto e^{-\frac{(f-f_{i})^{2}}{2\sigma_{f}^{2}}},$ (5) where $f$ is the planet’s flux given an input entropy $S$ and $f_{i}$ is the flux for a given entropy $S_{i}$. The flux error, $\sigma_{f}$, is calculated from the contrast limits of various instruments shown in Figure 7. The 5 $\sigma$ contrast limits are converted to planet flux limits and divided by 5 to obtain the flux error. For simplicity, we assume all values of entropy are equally possible and use a flat distribution for $P(S)$. As an example, we consider a star of apparent magnitude 7 in the K and L’ bands at a distance of 40 pc and age of 40 Myr. Assuming a true entropy of 10 $k_{B}$/baryon, the likelihood as a function of modelled entropy is shown in Figure 10 using VIKiNG with the UTs in both K and L’ bands. (a) K band (b) L’ band Figure 10: Likelihood of entropy for given data in K and L’ bands, with VIKiNG using the UTs, assuming an input entropy of 10 $k_{B}$/baryon and a planet 80 mas from a star of apparent magnitude 7. The solid curves are planets at 10 au and the dashed curves are planets at 20 au. The likelihood curves in Figure 10 confirm that the initial entropy of a 2 MJ planet cannot be constrained, while the entropy of a 4 MJ can be broadly constrained in the L’ band, but not in the K band. This method was applied to a random sample of Gaia detectable planets. Each planet was assigned a set of initial entropies from 8.5–11.5 $k_{B}$/baryon and the likelihood function was calculated for each of these. From this, we calculated the entropy uncertainty as a function of input entropy. As shown in Figure 8, only GRAVITY, MICADO, METIS and VIKiNG will be able to detect at least 1 planet that is also detectable by Gaia. The majority of simulated planets have entropies that cannot be constrained. As a preliminary result, we only consider 1 planet that is detectable by Gaia and all of the instruments mentioned above. The result for a 3.6 MJ planet 2.4 AU from a 0.75 M⊙ star in the $\beta$-Pic moving group is shown in Figure 11. (a) K band (b) L’ band Figure 11: Entropy Uncertainty as a function of input entropy. The horizontal axis is a simulated entropy which translates to a theoretical brightness and the vertical axis is the width of the entropy likelihood function as shown in Figure 10. For VIKiNG with the UTs, the entropy uncertainty is below 0.5 $k_{\rm{B}}$/baryon for all reasonable values of initial entropy which is less than the difference between hot and cold-start models. This implies VIKiNG should be able to distinguish between the two models. The curves in Figure 11 indicate that VIKiNG with the UTs will be able to constrain the initial entropy within 0.5 $k_{\rm{B}}$/baryon for the majority of input entropies. GRAVITY and METIS can also constrain the entropy within this value for a ‘warm’-start planet (entropy of 9–10 $k_{B}$/baryon), which indicates these instruments should be able to distinguish between hot and cold-start models for this planet. We note that, unlike the other instruments, VIKiNG is still only at a preliminary design study level at this point. ## 6 Conclusions In this paper, we have examined the brightness evolution of giant planets and the dependence on initial entropy. Given that high-entropy planets are brighter than low-entropy planets of similar mass but this difference becomes less at old ages. If we observe planets at young ages, it should be possible to constrain initial entropy. Using the expected uncertainty for Gaia astrometry from Perryman et al. (2014), we determine that Gaia should be able to detect approximately 25 % of giant planets in nearby moving groups and calculate their mass. Combining this with the estimated ages of these moving groups, we can use the detected flux to determine the initial entropy. We performed this over a simulated sample of planets around existing stars in nearby moving groups, assuming a symmetric planet distribution from Fernandes et al. (2019). We used the measured and expected 5$\sigma$ contrast limits for current and future instruments to estimate the expected flux error on the planets in our simulated sample. We found that future instruments MICADO and METIS have the best contrast levels at wide angles, while interferometers GRAVITY and VIKiNG are best at small angles. However, we note that improvements to GRAVITY, known as GRAVITY+ are currently being implemented and, of the VIKiNG concepts, only the L’ concept has significant funding. Given the technological challenges of achieving the required 30 nm fringe tracking uncertainty for a VIKiNG K and that entropy is better constrained by VIKiNG L’, this paper does not provide a reason to prioritise a high performance Nuller for VLTI operating in the K filter. Overall, assuming Gaia can detect giant planets in nearby moving groups, we find that these future instruments should also be able to detect some planets, if we were to conduct a survey of a relatively small number of Gaia-detected planets in nearby moving groups. Using the instrumental flux error, combined with the estimated ages and masses from Gaia, we can constrain the formation entropy of directly imaged planets. We found that GRAVITY, METIS and VIKiNG should all be able to constrain the formation entropy of a super-Jupiter to within 0.5 $k_{B}$/baryon and from this, distingush between hot and cold-start formation models. ## Data Availability The data underlying this article are available from the corresponding author on reasonable request. ## Acknowledgements We thank the anonymous referee for their useful comments, which greatly improved this study and the organization of this paper. We thank Mark Krumholz for initiating the discussions of planet entropy, which led to this work. This research was supported by the Australian Government through the Australian Research Council’s Discovery Projects funding scheme (DP190101477). C. F. acknowledges funding provided by the Australian Research Council through Future Fellowship FT180100495, and the Australia-Germany Joint Research Cooperation Scheme (UA-DAAD). ## References * Abuter et al. (2019) Abuter R., et al., 2019, Eso Messenger * Bell et al. (2015) Bell C. P., Mamajek E. E., Naylor T., 2015, Monthly Notices of the Royal Astronomical Society, 454, 593 * Benisty et al. (2021) Benisty M., et al., 2021, The Astrophysical Journal Letters, 916, L2 * Berardo et al. (2017) Berardo D., Cumming A., Marleau G.-D., 2017, The Astrophysical Journal, 834, 149 * Bowler & Nielsen (2018) Bowler B. P., Nielsen E. L., 2018, Handbook of Exoplanets, pp 1–17 * Bowler et al. (2009) Bowler B. P., et al., 2009, The Astrophysical Journal, 709, 396 * Brandt et al. (2021) Brandt G. M., Brandt T. D., Dupuy T. J., Michalik D., Marleau G.-D., 2021, The Astrophysical Journal Letters, 915, L16 * Brown et al. (2018) Brown A., et al., 2018, Astronomy & astrophysics, 616, A1 * Carlomagno et al. (2020) Carlomagno B., et al., 2020, Journal of Astronomical Telescopes, Instruments, and Systems, 6, 035005 * Carter et al. (2021) Carter A. L., et al., 2021, Monthly Notices of the Royal Astronomical Society, 501, 1999 * Defrère et al. (2018) Defrère D., et al., 2018, Experimental Astronomy, 46, 475 * Fernandes et al. (2019) Fernandes R. B., Mulders G. D., Pascucci I., Mordasini C., Emsenhuber A., 2019, The Astrophysical Journal, 874, 81 * Fischer & Marcy (1992) Fischer D. A., Marcy G. W., 1992, The Astrophysical Journal, 396, 178 * Fulton et al. (2021) Fulton B. J., et al., 2021, arXiv preprint arXiv:2105.11584 * Gagné et al. (2018) Gagné J., et al., 2018, The Astrophysical Journal, 856, 23 * Johnson et al. (2007) Johnson J. A., Butler R. P., Marcy G. W., Fischer D. A., Vogt S. S., Wright J. T., Peek K. M., 2007, The Astrophysical Journal, 670, 833 * Johnson et al. (2010) Johnson J. A., Aller K. M., Howard A. W., Crepp J. R., 2010, Publications of the Astronomical Society of the Pacific, 122, 905 * Keppler et al. (2018) Keppler M., et al., 2018, Astronomy & Astrophysics, 617, A44 * Kraus & Ireland (2011) Kraus A. L., Ireland M. J., 2011, The Astrophysical Journal, 745, 5 * Lacour et al. (2019) Lacour S., et al., 2019, Astronomy & Astrophysics, 623, L11 * Lafreniere et al. (2007) Lafreniere D., Marois C., Doyon R., Nadeau D., Artigau É., 2007, The Astrophysical Journal, 660, 770 * Lagrange et al. (2009) Lagrange A.-M., et al., 2009, Astronomy & Astrophysics, 493, L21 * Langlois et al. (2021) Langlois M., et al., 2021, arXiv preprint arXiv:2103.03976 * Lissauer et al. (2009) Lissauer J. J., Hubickyj O., D’Angelo G., Bodenheimer P., 2009, Icarus, 199, 338 * López-Santiago et al. (2006) López-Santiago J., Montes D., Crespo-Chacón I., Fernández-Figueroa M., 2006, The Astrophysical Journal, 643, 1160 * Macintosh et al. (2015) Macintosh B., et al., 2015, Science, 350, 64 * Marleau & Cumming (2014) Marleau G.-D., Cumming A., 2014, Monthly Notices of the Royal Astronomical Society, 437, 1378 * Marley et al. (2007) Marley M. S., Fortney J. J., Hubickyj O., Bodenheimer P., Lissauer J. J., 2007, The Astrophysical Journal, 655, 541 * Marois et al. (2006) Marois C., Lafreniere D., Doyon R., Macintosh B., Nadeau D., 2006, The Astrophysical Journal, 641, 556 * Martinache (2010) Martinache F., 2010, The Astrophysical Journal, 724, 464 * Martinache & Ireland (2018) Martinache F., Ireland M. J., 2018, arXiv preprint arXiv:1802.06252 * Mordasini et al. (2009) Mordasini C., Alibert Y., Benz W., Naef D., 2009, Astronomy & Astrophysics, 501, 1161 * Mordasini et al. (2012) Mordasini C., Alibert Y., Klahr H., Henning T., 2012, A&A, 547, A111 * Mulders (2018) Mulders G. D., 2018, Planet Populations as a Function of Stellar Properties. Springer International Publishing, Cham, pp 1–26, doi:10.1007/978-3-319-30648-3_153-1, https://doi.org/10.1007/978-3-319-30648-3_153-1 * Nielsen et al. (2019) Nielsen E. L., et al., 2019, The Astronomical Journal, 158, 13 * Perrot et al. (2018) Perrot C., Baudoz P., Boccaletti A., Rousset G., Huby E., Clénet Y., Durand S., Davies R., 2018, arXiv preprint arXiv:1804.01371 * Perryman et al. (2014) Perryman M., Hartman J., Bakos G. Á., Lindegren L., 2014, The Astrophysical Journal, 797, 14 * Quanz et al. (2012) Quanz S. P., Crepp J. R., Janson M., Avenhaus H., Meyer M. R., Hillenbrand L. A., 2012, The Astrophysical Journal, 754, 127 * Ranalli et al. (2018) Ranalli P., Hobbs D., Lindegren L., 2018, Astronomy & Astrophysics, 614, A30 * Rodriguez et al. (2013) Rodriguez D. R., Zuckerman B., Kastner J. H., Bessell M., Faherty J. K., Murphy S. J., 2013, The Astrophysical Journal, 774, 101 * Schneider et al. (2019) Schneider A. C., Shkolnik E. L., Allers K. N., Kraus A. L., Liu M. C., Weinberger A. J., Flagg L., 2019, The Astronomical Journal, 157, 234 * Spiegel & Burrows (2012) Spiegel D. S., Burrows A., 2012, The Astrophysical Journal, 745, 174 * Torres et al. (2008) Torres C. A., Quast G. R., Melo C. H., Sterzik M. F., 2008, arXiv preprint arXiv:0808.3362 * Vigan et al. (2020) Vigan A., et al., 2020, arXiv preprint arXiv:2007.06573 * Wallace & Ireland (2019) Wallace A., Ireland M., 2019, Monthly Notices of the Royal Astronomical Society, 490, 502 * Wallace et al. (2020) Wallace A., et al., 2020, Monthly Notices of the Royal Astronomical Society, 498, 1382 * Xuan et al. (2018) Xuan W. J., et al., 2018, AJ, 156, 156 * Zhu (2015) Zhu Z., 2015, The Astrophysical Journal, 799, 16 * Zuckerman (2018) Zuckerman B., 2018, The Astrophysical Journal, 870, 27 * Zuckerman et al. (2011) Zuckerman B., Rhee J. H., Song I., Bessell M., 2011, The Astrophysical Journal, 732, 61
# Boosting segmentation performance across datasets using histogram specification with application to pelvic bone segmentation ###### Abstract Accurate segmentation of the pelvic CTs is crucial for the clinical diagnosis of pelvic bone diseases and for planning patient-specific hip surgeries. With the emergence and advancements of deep learning for digital healthcare, several methodologies have been proposed for such segmentation tasks. But in a low data scenario, the lack of abundant data needed to train a Deep Neural Network is a significant bottle-neck. In this work, we propose a methodology based on modulation of image tonal distributions and deep learning to boost the performance of networks trained on limited data. The strategy involves pre-processing of test data through histogram specification. This simple yet effective approach can be viewed as a style transfer methodology. The segmentation task uses a U-Net configuration with an EfficientNet-B0 backbone, optimized using an augmented BCE-IoU loss function. This configuration is validated on a total of 284 images taken from two publicly available CT datasets, TCIA (a cancer imaging archive) and the Visible Human Project. The average performance measures for the Dice coefficient and Intersection over Union are 95.7% and 91.9%, respectively, give strong evidence for the effectiveness of the approach, which is highly competitive with state-of-the- art methodologies. Index Terms— Pelvic bone segmentation, data pre-processing, histogram specification, U-Net, fine-tuning. ## 1 Introduction In recent years, due to the increase in the incidence of pelvic injuries from traffic-related accidents [1], pelvic bone diseases within the aging population, and sufficient access to computed tomography (CT) imaging, automated pelvic bone segmentation in CT has gained considerable prominence. The segmentation results assist physicians in the early detection of pelvic injury and help expedite surgical planning and reduce the complications caused by pelvic fractures [2]. In CT data, structures like the bone marrow and bone surface appear as dark and bright regions due to their low and high densities compared to the surrounding tissues. However, given the variations in image quality between different CT datasets, distinguishing bone structures from the image background becomes cumbersome and leads to erroneous segmentation outputs. These issues indicate a need for a novel solution to develop a simple yet effective methodology for the accurate segmentation of pelvic bones from varying CT data. Contribution of this paper: The key novelties of this work are as follows: 1. 1. introduction of an encoder-decoder network, trained on limited data, for high accuracy segmentation of pelvic bones 2. 2. boosting model performance on unseen data by employing histogram specification The exact details of the approach are deferred until Sec. 3.3. Fig. 1 illustrates the results of the proposed method. $\begin{array}[]{cc}\includegraphics[width=99.58464pt,height=99.58464pt]{images1/3.png}&\includegraphics[width=99.58464pt,height=99.58464pt]{images1/4.png}\\\ \mbox{TCIA}&\mbox{(a1)}\\\ \includegraphics[width=99.58464pt,height=99.58464pt]{images1/t1.png}&\includegraphics[width=99.58464pt,height=99.58464pt]{images1/t2.png}\\\ \mbox{VHBD}&\mbox{(b1)}\\\ \end{array}$ Fig. 1: (a1) and (b1) – illustrate the segmentation outputs, for input images from TCIA [3] and VHBD [4], respectively. ## 2 Prior Art Fig. 2: Workflow of U-Net architecture with pre-trained backbone, detailing pelvic bone segmentation. Recent literature has seen many applications for the segmentation of the pelvis from CT imaging data. Traditional methods such as thresholding and region growth [5], deformable surface model [6], and others, have been commonly used to perform bone segmentation. However, these approaches often suffer from low accuracy due to varying image properties such as intensity, contrast, and the inherent variations between the texture of the bone structures (bone marrow and surface boundary) and the surrounding tissues. To overcome these challenges, supervised methods such as statistical shape models (SSM) and atlas-based deep learning (DL) methods have made significant contributions to segmentation tasks. Wang et al. [7, 8] suggested using a multi-atlas segmentation with joint label fusion for detecting regions on interest from CT images. Yokota et al. [9] showcased a combination of hierarchical and conditional SSMs for the automated segmentation of diseased hips from CT data. Chu et al. [10] presented a multi-atlas based method for accurately segmenting femur and pelvis. Zeng et al. [11] proposed a supervised 3D U-Net with multi-level supervision for segmenting femur in 3D MRI. Chen et al. [12] showcased a 3D feature enhanced network for quickly segmenting femurs from CT data. Chang et al. [13] proposed patch-based refinement on top of a conditional random field model for fine segmentation of healthy and diseased hips. Liu et al. [14] used 3D U-Nets in two-stages (trained on approximately 270K images) with a signed distance function for producing bone fragments from image-stacks. In the following section, we discuss a new technique addressing accurate segmentation of the pelvis from CT images of varying qualities. ## 3 Proposed Methodology The efficacy of using Encoder-Decoder architectures for designing high accuracy segmentation models for biomedical applications has been showcased in recent literature [15, 11, 14]. We employ a similar architecture, with various encoder modules for feature extraction and a decoder module for semantic segmentation. The details of the encoder and decoder modules are explained in the following. ### 3.1 Encoder Module In simple terms, an encoder takes the input image and generates a high- dimensional feature vector aggregated over multiple levels. We deploy a choice of the following well-known architectures as the encoder module: #### 3.1.1 ResNet Residual networks (ResNet) introduced residual mappings to solve the vanishing gradient problem in deep neural networks [16]. ResNets are easy to optimize and gain accuracy even with deeper models. #### 3.1.2 Inception V3 Inception Networks are computationally efficient architectures, both in terms of the model parameters and their memory usage. Adapting the Inception network for different applications while ensuring that changes do not impede its computational efficiency is difficult. Inception V3 introduced various strategies for optimizing network with ease of model adaptation capabilities [17]. #### 3.1.3 EfficientNet Conventional methods make use of scaling to increase the accuracy of the models. The models are scaled by increasing the depth/width of the network or using higher resolution input images. EfficientNet results from a novel scaling method that uses a compound coefficient to uniformly scale the network across all dimensions [18]. ### 3.2 Decoder Module The decoder module is responsible for generating a semantic segmentation mask using the aggregated high-dimensional features extracted by the encoder module. We make use of the popular U-Net model specially designed for medical imaging as the decoding module [15]. ### 3.3 Histogram Specification Histogram specification, or histogram matching, is a traditional image processing technique [19] that matches the input image’s histogram to a reference histogram. Histogram specification involves computing the cumulative distribution function (CDF) of histograms from both the target and the reference, following which a transformation function is obtained by mapping each gray level $[0,255]$ from the target’s CDF (input) to the gray level in the reference CDF. In this work, we construct the reference histogram by averaging over histograms from every image in the training set. Using this technique as a pre-processing step for the test data serves an important purpose, as the distribution of the test data is converted to a similar form seen by the network during training. ## 4 Experimental Validation Table 1: An overview of the datasets used in this work. Dataset | Resolution | Train-set | Val-set | Test-set ---|---|---|---|--- # Images | (%) | (%) | (%) TCIA [3] | 512 x 512 | 407 | 58 | 117 582 | (70%) | (10%) | (20%) VHBD [4] | 512 x 512 | – | – | 167 167 | (100%) VHBD-2 [4] | 512 x 512 | 116 | 17 | 34 167 | (70%) | (10%) | (20%) ### 4.1 Datasets The input data preparation and label annotation were done using the tools from Image-J software. A summary of TCIA–cancer imaging archive [3] and VHBD–Visible human project [4] datasets, image resolution, the number of images used in this study, and the respective data-splits for training- validation-testing, are shown in Table 1. ### 4.2 Performance Measures To quantify the quality of segmentation, we compute standard performance measures for segmentation tasks commonly used in literature, specifically, the mean Dice coefficient (mDice) and mean Intersection over Union (mIoU) [20, 21]. For a given segmentation output ($A$) and the ground truth ($B$), the Dice coefficient is given by $\text{Dice}=\frac{2\,|A\;\cap\;B|}{|A|\;+\;|B|}$, which can be interpreted as a weighted average of the precision and recall, and $\text{IoU}=\frac{|A\;\cap\;B|}{|A\;\cup\;B|}\text{,}$ (also known as Jaccard index) is commonly used for comparing the similarity between sets ($A$) and ($B$), while penalizing their diversity. ### 4.3 Network Training The implementations used were based on the documentation from [22]. The models used [16, 17, 18] were pre-trained on the Imagenet [23] dataset to improve the generalization capability on unseen data and achieve faster convergence. For the base-model, we use ResNet-34 [16] as the encoder and a U-Net decoder. We initialize the base-model with random weights (rnwt) and train without any data-augmentation (noaug) on images from [3], using an Nvidia RTX 2070 GPU, and an ADAM optimizer with a learning rate of 0.001, momentum of 0.9 and a weight decay of 0.0001, for 40 epochs. We chose a 70% : 10% : 20% split of the data (shown in the first row of Table 1), where the 70% was utilized for training and the 10% of the data was utilized for validation. The remaining 20% for testing was completely unseen during training. About 50 passes of random image batches, of size eight, from the training set, were used in each epoch. The model was then validated on the 10% data to evaluate the performance based on binary-cross-entropy loss (bce) and record the corresponding weights. After training, the weights that gave the best performance on the validation set were selected for the base-model, which was then evaluated on the unseen test-sets, i.e., 20% of [3] and 100% of [4], respectively, whose performance is showcased in the first row of Table 2. Extending beyond the base-model, data augmentation (aug) was performed using horizontal and vertical flips, affine transforms, image intensity modulation and blurring, for increasing training data size and to help reduce over- fitting. In addition, we try to find the best overall segmentation performance and generalization capability to completely unseen data, through further extension of the base-model with different configurations, using the following: * • encoder modules using ResNet-34 [16], Inception V3 [17] and EfficientNet-B0 [18], initialized with Imagenet weights (imwt) for transfer learning * • re-configuration of input data, or not, to the pre-trained model’s format and its pre-processing functions (ppr), for extraction of better features * • loss functions like Dice loss (dice), IoU loss (iou) and combined bce-iou loss, in place of bce loss, for propagating strong gradients for better optimization and learning $\begin{array}[]{ccc}\includegraphics[width=71.13188pt,height=71.13188pt]{images1/inp.png}&\includegraphics[width=71.13188pt,height=71.13188pt]{images1/dl1.png}&\includegraphics[width=71.13188pt,height=71.13188pt]{images1/dl2.png}\\\ \mbox{Input}&\mbox{(a)}&\mbox{(b)}\\\ \end{array}$ Fig. 3: Pelvic bone segmentation on TCIA data using: (a) Base U-Net with random weight initialization for ResNet-34 encoder, with no data-augmentation, optimized using BCE loss (least performing); and (b) fine-tuned U-Net with Imagenet weight initialization for EfficientNet-B0 encoder, with data- augmentation and input reconfiguration, optimized using combined BCE-IoU loss (best performing), are overlaid onto the binary ground-truth; yellow - TP; black - TN; green - FP; red - FN. $\begin{array}[]{ccc}\includegraphics[width=71.13188pt,height=71.13188pt]{images1/a.png}&\includegraphics[width=71.13188pt,height=56.9055pt]{images1/b1.png}&\includegraphics[width=71.13188pt,height=71.13188pt]{images1/c.png}\\\ \mbox{TCIA}&\mbox{(a1)}&\mbox{(a2)}\\\ \includegraphics[width=71.13188pt,height=71.13188pt]{images1/d.png}&\includegraphics[width=71.13188pt,height=56.9055pt]{images1/e1.png}&\includegraphics[width=71.13188pt,height=71.13188pt]{images1/f.png}\\\ \mbox{VHBD (target)}&\mbox{(b1)}&\mbox{(b2)}\\\ \includegraphics[width=71.13188pt,height=71.13188pt]{images1/g.png}&\includegraphics[width=71.13188pt,height=56.9055pt]{images1/h1.png}&\includegraphics[width=71.13188pt,height=71.13188pt]{images1/i.png}\\\ \mbox{H-VHBD}&\mbox{(c1)}&\mbox{(c2)}\\\ \end{array}$ Fig. 4: Performance in segmentation with histogram specification: (a1-c1) show the respective histograms of the input images; (a2-c2) show the pelvic bone segmentations overlaid on the ground-truth; and (b2-c2) decisively show the improvement in segmentation from matching target’s histogram to the reference. yellow - TP; black - TN; green - FP; red - FN. Table 2: Performance comparison of different U-Net configurations for pelvic bone segmentation on unseen data from TCIA, VHBD, and H-VHBD, i.e., VHBD after histogram specification. U-Net Configurations | TCIA | VHBD | H-VHBD ---|---|---|--- mIoU | mDice | mIoU | mDice | mIoU | mDice Res34-rnwt-noaug-bce | 0.788 $\pm$ 0.033 | 0.867 $\pm$ 0.025 | 0.131 $\pm$ 0.033 | 0.186 $\pm$ 0.037 | 0.774 $\pm$ 0.021 | 0.865 $\pm$ 0.015 Res34-imwt-aug-bce | 0.919 $\pm$ 0.007 | 0.957 $\pm$ 0.004 | 0.746 $\pm$ 0.028 | 0.840 $\pm$ 0.021 | 0.900 $\pm$ 0.004 | 0.947 $\pm$ 0.002 Res34-imwt-aug-dice | 0.925 $\pm$ 0.007 | 0.960 $\pm$ 0.004 | 0.523 $\pm$ 0.039 | 0.645 $\pm$ 0.038 | 0.907 $\pm$ 0.005 | 0.951 $\pm$ 0.002 Res34-imwt-aug-bce-iou | 0.922 $\pm$ 0.007 | 0.959 $\pm$ 0.004 | 0.790 $\pm$ 0.018 | 0.877 $\pm$ 0.012 | 0.906 $\pm$ 0.004 | 0.950 $\pm$ 0.002 IncepV3-imwt-aug-bce | 0.876$\pm$ 0.012 | 0.932 $\pm$ 0.007 | 0.663 $\pm$ 0.026 | 0.784 $\pm$ 0.019 | 0.906 $\pm$ 0.006 | 0.950 $\pm$ 0.003 IncepV3-ppr-imwt-aug-bce | 0.921 $\pm$ 0.011 | 0.957 $\pm$ 0.007 | 0.808 $\pm$ 0.014 | 0.890 $\pm$ 0.009 | 0.913 $\pm$ 0.005 | 0.954 $\pm$ 0.002 EffiB0-imwt-aug-bce | 0.923 $\pm$ 0.007 | 0.959 $\pm$ 0.004 | 0.835 $\pm$ 0.015 | 0.906 $\pm$ 0.010 | 0.901 $\pm$ 0.006 | 0.947 $\pm$ 0.003 EffiB0-ppr-imwt-aug-bce-iou | 0.924 $\pm$ 0.008 | 0.960 $\pm$ 0.004 | 0.836 $\pm$ 0.011 | 0.909 $\pm$ 0.006 | 0.914 $\pm$ 0.005 | 0.955 $\pm$ 0.002 ABLATION STUDY (♣) | 0.913 $\pm$ 0.007 | 0.954 $\pm$ 0.004 | 0.679 $\pm$ 0.047 | 0.801 $\pm$ 0.032 | 0.873 $\pm$ 0.013 | 0.931 $\pm$ 0.007 * * Encoder module \- Res34, IncepV3, EffiB0 are ResNet-34, Inception Net-V3, EfficientNet-B0, respectively. * * Encoder Weights \- rnwt and imwt are random weights and Imagenet weights, respectively. * * Augmentation \- aug and noaug means training with and without data- augmentation, respectively. * * Loss \- bce, dice, iou are the Binary Cross Entropy loss, Dice Loss and IoU loss, respectively . * * ppr\- configure input to the pre-trained backbone’s format. * * Grey background\- indicates improvement due to histogram-specification based pre-processing. ### 4.4 Results The detailed comparisons of the different U-Net configurations’ segmentation performance on test-sets with 95% confidence intervals are shown in Table 2. The segmentation outputs from the least-performing (base-model) and best- performing (fine-tuned U-Net with Imagenet weight initialization for EfficientNet-B0 encoder [18], with data-augmentation and input re- configuration, optimized using combined BCE-IoU loss) DL models are showcased in Fig. 3 (a) &(b). The predicted outputs are overlaid onto the ground-truth and color-coded (yellow - TP; black - TN; green - FP; red - FN) for visualizing the quality of segmentation. The results shown in Fig. 4(b2) &(c2) illustrate the desired effect on segmentation due to histogram specification. The reduction in the number of pixels labeled as FPs & FNs, and improvement in number of TPs from the overlays decisively show the significance of pre- processing test-data, which clearly boosts the model’s segmentation performance. Furthermore, the comparitive results tabulated in the last two columns of Table 2 give strong evidence for the success of the proposed methodology on all the specified model configurations. On analyzing the data shown in Table 3, the proposed methodology’s overall performance on the test-sets surpassed several state-of-the-art techniques that were trained on similarly sized datasets, with the exception of Liu et al. [14] who performed training on approximately 270,000 images. Since data drives any model, the proposed methodology (trained only on 407 images) shows room for further improvement in segmentation under the availability of larger datasets. Table 3: Overall performance comparison for pelvic bone segmentation with state-of-the-art techniques. Methodology | Dataset | mIoU | mDice ---|---|---|--- | (# Images) | | Yokota et al. [9] | Private (100) | — | 0.928 Chu et al. [10] | Private (318) | — | 0.941 Chang et al. [13] | Private ($\sim$3420) | — | 0.949 Liu et al. [14] | DS$\ddagger$($\sim$63K) | — | 0.984 Proposed method | TCIA,VHBD (284) | 0.919 | 0.957 * (DS$\ddagger$) KITS19, CERVIX, ABDOMEN, MSD T10, COLONOG, CLINIC; Train:Test $\approx$ 270K: 63K; K = $10^{3}$ ### 4.5 Ablation Study Images from [3, 4], with the data splits shown in rows 1 and 3 of Table 1, are used for training. The best-model was trained on the joint data whose test- data performance is shown in Table 2 (♣). The results showed that training the model on joint data degrades the performance on both datasets. The data imbalance and the varying image tonal distributions play a significant role in influencing the segmentation performance. And by using the proposed methodology, the model overcomes data imbalance and generalizes well to unseen datasets, which boosts its overall segmentation performance. ## 5 Conclusion To sum up, in this work, we presented a novel methodology for the automated segmentation of pelvic bones from axial CT images. We addressed the unmet need for superior pelvic bone segmentation methodology for images with varying properties by using histogram specification. This simple yet powerful approach of pre-processing the test-data improved segmentation performance by a significant margin, with the quantitative results confirming its validity. Through our approach, the encoder-decoder configuration overcame a significant hurdle of varying intensity distributions in CT images, which led to superior segmentation quality. Moreover, after validating the results on publicly available TCIA and VHBD datasets, the proposed methodology has been shown to be highly competent with-respect-to existing state-of-the-art techniques. Through this study, we saw that, although deep learning has pushed the limits for image processing applications, traditional image processing techniques are not necessarily obsolete and that combining the two approaches can lead to superior performance in segmentation. ## References * [1] Rebecca B Naumann, Ann M Dellinger, Eduard Zaloshnja, Bruce A Lawrence, and Ted R Miller, “Incidence and total lifetime costs of motor vehicle–related fatal and nonfatal injury by road user type, united states, 2005,” Traffic injury prevention, vol. 11, no. 4, pp. 353–360, 2010. * [2] Hui Yu, Haijun Wang, Yao Shi, Ke Xu, Xuyao Yu, and Yuzhen Cao, “The segmentation of bones in pelvic ct images based on extraction of key frames,” BMC medical imaging, vol. 18, no. 1, pp. 18, 2018. * [3] Kenneth Clark, Bruce Vendt, Kirk Smith, John Freymann, Justin Kirby, Paul Koppel, Stephen Moore, Stanley Phillips, David Maffitt, Michael Pringle, et al., “The cancer imaging archive (tcia): maintaining and operating a public information repository,” Journal of digital imaging, vol. 26, no. 6, pp. 1045–1057, 2013\. * [4] M. J. Ackerman, “The visible human project,” Proceedings of the IEEE, vol. 86, no. 3, pp. 504–511, 1998. * [5] Phan TH Truc, Sungyoung Lee, and Tae-Seong Kim, “A density distance augmented chan-vese active contour for ct bone segmentation,” in 2008 30th Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE, 2008, pp. 482–485. * [6] Dagmar Kainmueller, Hans Lamecker, Stefan Zachow, and Hans-Christian Hege, “Coupling deformable models for multi-object segmentation,” in International Symposium on Biomedical Simulation. Springer, 2008, pp. 69–78. * [7] Hongzhi Wang, Jung W Suh, Sandhitsu R Das, John B Pluta, Caryne Craige, and Paul A Yushkevich, “Multi-atlas segmentation with joint label fusion,” IEEE transactions on pattern analysis and machine intelligence, vol. 35, no. 3, pp. 611–623, 2012. * [8] Hongzhi Wang, Mehdi Moradi, Yaniv Gur, Prasanth Prasanna, and Tanveer Syeda-Mahmood, “A multi-atlas approach to region of interest detection for medical image classification,” in International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2017, pp. 168–176. * [9] Futoshi Yokota, Toshiyuki Okada, Masaki Takao, Nobuhiko Sugano, Yukio Tada, Noriyuki Tomiyama, and Yoshinobu Sato, “Automated ct segmentation of diseased hip using hierarchical and conditional statistical shape models,” in International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2013, pp. 190–197. * [10] Chengwen Chu, Junjie Bai, Xiaodong Wu, and Guoyan Zheng, “Mascg: Multi-atlas segmentation constrained graph method for accurate segmentation of hip ct images,” Medical image analysis, vol. 26, no. 1, pp. 173–184, 2015. * [11] Guodong Zeng, Xin Yang, Jing Li, Lequan Yu, Pheng-Ann Heng, and Guoyan Zheng, “3d u-net with multi-level deep supervision: fully automatic segmentation of proximal femur in 3d mr images,” in International workshop on machine learning in medical imaging. Springer, 2017, pp. 274–282. * [12] Fang Chen, Jia Liu, Zhe Zhao, Mingyu Zhu, and Hongen Liao, “Three-dimensional feature-enhanced network for automatic femur segmentation,” IEEE journal of biomedical and health informatics, vol. 23, no. 1, pp. 243–252, 2017. * [13] Yong Chang, Yongfeng Yuan, Changyong Guo, Yadong Wang, Yuanzhi Cheng, and Shinichi Tamura, “Accurate pelvis and femur segmentation in hip ct with a novel patch-based refinement,” IEEE journal of biomedical and health informatics, vol. 23, no. 3, pp. 1192–1204, 2018. * [14] Pengbo Liu, Hu Han, Yuanqi Du, Heqin Zhu, Yinhao Li, Feng Gu, Honghu Xiao, Jun Li, Chunpeng Zhao, Li Xiao, et al., “Deep learning to segment pelvic bones: Large-scale ct datasets and baseline models,” arXiv preprint arXiv:2012.08721, 2020. * [15] Olaf Ronneberger, Philipp Fischer, and Thomas Brox, “U-net: Convolutional networks for biomedical image segmentation,” in International Conference on Medical image computing and computer-assisted intervention. Springer, 2015, pp. 234–241. * [16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778. * [17] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna, “Rethinking the inception architecture for computer vision,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 2818–2826. * [18] Mingxing Tan and Quoc V Le, “Efficientnet: Rethinking model scaling for convolutional neural networks,” arXiv preprint arXiv:1905.11946, 2019. * [19] Richard Szeliski, Computer vision: algorithms and applications, Springer Science & Business Media, 2010. * [20] William R Crum, Oscar Camara, and Derek LG Hill, “Generalized overlap measures for evaluation and validation in medical image analysis,” IEEE Trans. Med. Imag., vol. 25, no. 11, pp. 1451–1461, 2006\. * [21] Herng-Hua Chang, Audrey H Zhuang, Daniel J Valentino, and Woei-Chyn Chu, “Performance measure characterization for evaluating neuroimage segmentation algorithms,” Neuroimage, vol. 47, no. 1, pp. 122–135, 2009. * [22] Pavel Yakubovskiy, “Segmentation models,” https://segmentation-models.readthedocs.io/en/latest/, 2019. * [23] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in 2009 IEEE conference on computer vision and pattern recognition. Ieee, 2009, pp. 248–255.
∎ 11institutetext: A. De Santis, T. Giovannelli, S. Lucidi, M. Roma 22institutetext: Dipartimento di Ingegneria Informatica, Automatica e Gestionale “A. Ruberti” SAPIENZA, Università di Roma via Ariosto, 25 – 00185 Roma, Italy 22email<EMAIL_ADDRESS>33institutetext: M. Messedaglia 44institutetext: ACTOR Start up of SAPIENZA Università di Roma via Nizza 45, 00198 Roma, Italy. 44email<EMAIL_ADDRESS> # Determining the optimal piecewise constant approximation for the Nonhomogeneous Poisson Process rate of Emergency Department patient arrivals Alberto De Santis https://orcid.org/0000-0001-5175-4951 Tommaso Giovannelli https://orcid.org/0000-0002-1436-5348 Stefano Lucidi https://orcid.org/0000-0003-4356-7958 Mauro Messedaglia Massimo Roma https://orcid.org/0000-0002-9858-3616 ###### Abstract Modeling the arrival process to an Emergency Department (ED) is the first step of all studies dealing with the patient flow within the ED. Many of them focus on the increasing phenomenon of ED overcrowding, which is afflicting hospitals all over the world. Since Discrete Event Simulation models are often adopted with the aim to assess solutions for reducing the impact of this problem, proper nonstationary processes are taken into account to reproduce time- dependent arrivals. Accordingly, an accurate estimation of the unknown arrival rate is required to guarantee reliability of results. In this work, an integer nonlinear black-box optimization problem is solved to determine the best piecewise constant approximation of the time-varying arrival rate function, by finding the optimal partition of the 24 hours into a suitable number of non equally spaced intervals. The black-box constraints of the optimization problem make the feasible solutions satisfy proper statistical hypotheses; these ensure the validity of the nonhomogeneous Poisson assumption about the arrival process, commonly adopted in the literature, and prevent to mix overdispersed data for model estimation. The cost function includes a fit error term for the solution accuracy and a penalty term to select an adeguate degree of regularity of the optimal solution. To show the effectiveness of this methodology, real data from one of the largest Italian hospital EDs are used. ###### Keywords: Emergency Department Arrival process Non Homogeneous Poisson Process Black-Box Optimization ## 1 Introduction Statistical modelling for describing and predicting patient arrival to Emergency Departments (EDs) represent a basic tool of each study concerning ED patient load and crowding. Indeed, all the approaches adopted to this aim require an accurate model of the patient arrival process. Of course, such a process plays a key role in tackling the widespread phenomenon of overcrowding which afflicts EDs all over the world Ahalt et al. (2018); Bernstein et al. (2003); Daldoul et al. (2018); Hoot and Aronsky (2008); Hoot et al. (2007); J Reeder et al. (2003); Vanbrabant et al. (2020); Wang et al. (2015); Weiss et al. (2004, 2006). The two factors that have the most significant effect on overcrowding are both external and internal. The first concerns the patient arrival process; the second regards the patient flow within the ED. Therefore, both the aspects must be accurately considered for a reliable study on ED operation. Several modelling approaches for analyzing ED patient flow have been proposed in literature (see Wiler et al. (2011) for a survey). The main quantitative methods used are based on statistical analysis (time–series, regression) or on general analytic formulas (queuing theory). Simulation modelling (both Discrete Event and Agent Based Simulation) is currently one of the most widely used and flexible tool for studying the patient flow through an ED. In fact, it enables to perform an effective scenario analysis, aiming at determining bottlenecks (if any) and testing different ED settings. We refer to Salmon et al. (2018) for a recent survey on simulation modelling for ED operation. A step forward is Simulation–Based Optimization methodology which combines a simulation model with a black-box optimization algorithm, aiming at determining an optimal ED settings, based on suited objective function (representing some KPIs) to be maximized or minimized Ahmed and Alkhamis (2009); Guo et al. (2016, 2017). Modeling methodologies are generally based on assumptions that, in some cases, may represent serious limitations when applied to complex real–world cases, such as ED operation. In particular, when dealing with ED patient arrival stochastic modelling, due to the nonstationarity of the process, a standard assumption is the use of Nonhomogeneous Poisson Process (NHPP) Ahalt et al. (2018); Ahmed and Alkhamis (2009); Guo et al. (2017); Kim and Whitt (2014a); Kuo et al. (2016); Zeinali et al. (2015). We recall that a counting process $X(t)$ is a NHPP if 1) arrivals occur one at a time (no batch); 2) the process has independent increments; 3) increments have Poisson distribution, i.e. for each interval $[t_{1},t_{2}]$, $P\left(X(t_{1})-X(t_{2})=n\right)=e^{-m(t_{1},t_{2})}\frac{[m(t_{1},t_{2})]^{n}}{n!},$ where $m(t_{1},t_{2})={\int_{t_{1}}^{t_{2}}\lambda(s)ds}$ and $\lambda(t)$ is the arrival rate. Unlike the Poisson process (where $\lambda(t)=\lambda$), NHPP has nonstationary increments and this makes the use of NHPP suitable for modelling ED arrival process, which is usually strongly time–varying. Of course, appropriate statistical tests must be applied to available data to check if NHPP fits. This is usually performed by assuming that NHPP has a rate which can be considered approximately piecewise constant. Hence, Kolmogorov–Smirnov (KS) statistical test can be applied in separate and equally spaced intervals and usually the classical Conditional–Uniform (CU) property of the Poisson process is exploited Brown et al. (2005); Kim and Whitt (2014a, b). Unlike standard KS test, in the CU KS test the data are transformed before applying the test. More precisely, by CU property, the piecewise constant NHPP is transformed into a sequence of i.i.d. random variables uniformly distributed on $[0,1]$ so that it can be considered a (homogeneous) Poisson process in each interval. In this manner, the data from all the intervals can be merged in a single sequence of i.i.d. random variables uniformly distributed on $[0,1]$. This procedure, proposed in Brown et al. (2005), enables to remove nuisance parameters and to obtain independence from the rate of the Poisson process on each interval. Hence data from separate intervals (with different rates on each of them) and also from different days can been combined, avoiding common drawback due to large within-day and day-to-day variation of the ED patient arrival rate. Actually, Brown et al. in Brown et al. (2005) apply CU KS test after performing a further logarithmic data transformation. In Kim and Whitt (2014b, 2015), this approach has been extensively tested along with alternative data transformations proposed in early papers Durbin (1961) and Lewis (1965). However, Kim and Whitt in Kim and Whitt (2014a) observed that this procedure applied to ED patient arrival data is fair only if they are “analyzed carefully”. This is due to the fact that the following three issues must be seriously considered: 1) data rounding, 2) choice of the intervals, 3) overdispersion. In fact, the first issue may produce batch arrivals (zero length interarrival times) that are not included in a NHPP, so that unrounded data (or an unrounding procedure) must be considered. The second is a major issue in dealing with ED patient arrivals, since arrival rate can rapidly change so that the piecewise constant approximation is reasonable only if the interval are properly chosen. The third issue regards combining data from multiple days. Indeed, in studying ED patient arrival process, it is common to combine data from the same time slot from different weekdays, being this imperative when data from a single day are not sufficient for statistical testing. Data collected from EDs database usually show large variability over successive weeks mainly due to seasonal phenomena like flu season, holiday season, etc. However, this overdispersion phenomenon must be checked by using a dispersion test on the available data (e.g. Kathirgamatamby (1953)). In this work we propose a new modelling approach for ED patient arrival process based on a piecewise constant approximation of the arrival rate accomplished with non equally spaced intervals. This choice is suggested by the typical situation that occurs in EDs where the arrival rate is low and varying during the night hours, and it is higher and more stable in the day time, this is indeed what happens in the chosen case study. Therefore, to obtain an accurate representation of the arrival rate $\lambda(t)$ by a piecewise constant function $\lambda_{D}(t)$, a finer discretization of the time–domain is required during the night hours, as opposite to day time. For this reason, the proposed method finds the best partition of the 24 hours into intervals not necessarily equally spaced. As far as the authors are aware, the use of an optimization method for identifying stochastic processes characterizing the patient flow through an ED was already proposed in Guo et al. (2016), but that study aimed at determining the optimal service time distribution parameters (by using a meta–heuristic approach) and it did not involve ED arrival process. Therefore our approach represents the first attempt to adopt an optimization method for determining the best stochastic model for the ED process arrivals. In the previous work De Santis et al. (2020) a preliminary study was performed following the same approach. Here, with respect to De Santis et al. (2020), we propose a significantly enhanced statistical model which allows us to obtain better results on the case study we consider. In constructing a statistical model of the ED patient arrivals, a natural way to define a selection criterion is to evaluate the fit error between $\lambda(t)$ and its approximation $\lambda_{D}(t)$. However, the true arrival rate is unknown. In the approach we propose, as opposite to Kim and Whitt (2014a), no analytical model is assumed for $\lambda(t)$, but it is substituted by an “empirical arrival rate model” $\lambda_{F}(t)$ obtained by a sample approximation corresponding to the very fine uniform partition of the $24$ hours into intervals of $15$ minutes. In each of these intervals the average arrival rate values has been estimated from data obtained by collecting samples over the same day of the week, for all the weeks in some months using experimental data for the ED patient arrival times. Hence, any other $\lambda_{D}(t)$ corresponding to a grosser partition of the day must be compared to $\lambda_{F}(t)$. In other words, an optimization problem is solved to select the best day partition in non equally spaced intervals, determining a piecewise constant approximation of the arrival rate over the 24 hours with the best fit to the empirical model. Therefore, the objective function (to be minimized) of the optimization problem we formulate, comprises the fit error, namely the mean squared error. Moreover, an additional penalty term is included aiming at obtaining the overall regularity of the optimal approximation, being the latter measured by means of the sum of the squares of the jumps between the values in adjacent intervals. The rationale behind this term is to avoid optimal solutions with too rough behavior, namely few long intervals with high jumps. To make the result reliable, a number of constraints must be considered. First, the length of each interval of the partition can not be less than a fixed value (half an hour, one hour). Moreover, for each interval, * • the CU KS test must be satisfied to support the NHPP hypothesis; * • the dispersion test must be satisfied to ensure that data are not overdispersed, and could be considered as a realization of the same process (no week seasonal effects). The resulting problem is a black-box constrained optimization problem and to solve it we use a method belonging to the class of Derivative-Free Optimization. In particular we use the new algorithmic framework recently proposed in Liuzzi et al. (2020) which handles black-box problems with integer variables. We performed an extensive experimentation on data collected from the ED of a big hospital in Rome (Italy), also including some significant sensitivity analyses. The results obtained show that this approach enables to determine the number of intervals and their length such that an accurate approximation of the empirical arrival rate is achieved, ensuring the consistency between the NHPP hypothesis and the arrival data. The regularity of optimal piecewise constant approximation can be also finely tuned by proper weighing a penalty term in the objective function with respect to the fit error term. It is worth noting that the use a piecewise constant function for approximating the arrival rate function is usually required by the most common discrete event simulation software packages when implementing ED patient arrivals process as a NHPP. The paper is organized as follows. In Section 2, we briefly report information on the hospital ED under study. Section 3 describes the statistical model we propose. The optimization problem we consider is stated in Section 4 and the results of an extensive experimentation are reported in Section 5. Finally Section 6 includes some concluding remarks. ## 2 The case study under consideration The case study we consider concerns the ED of the Policlinico Umberto I, a very large hospital in Rome, Italy. It is the biggest ED in the Lazio region in terms of yearly patients arrivals (about 140,000 on the average). Thanks to the cooperation of the ED staff, we were able to collect data concerning the patient flow through the ED for the whole year 2018. In particular, for the purpose of this work, we focus on the patient arrivals data collected in the first $m$ weeks of the year. Both walk-in patients and patients transported by emergency medical service vehicles are considered. In Figure 1, the weekly hourly average arrival rate to the ED is shown for $m=13$, i.e. for data collected from the 1st of January to the 31st of March. Figure 1: Plot of the weekly average arrival rate for the first 13 weeks of the year. In particular, for each day of the week, the arrival rate is obtained by averaging the number of arrivals occurring in the same hourly time slot over the 13 weeks considered. We observe that, in accordance with the literature (see, i.e., Kim and Whitt (2014a)), the average arrival rates among the days of the week are significantly different. Therefore, since averaging over these days would lead to inaccurate results, the different days of the week must be considered separately. Figure 2 reports the hourly average arrival rate for each day of the week, again referring to $m=13$, i.e. to the first 13 weeks of the year. Figure 2: Plot of the comparison among hourly average arrival rate for each day of the week for the first 13 weeks of the year. By observing this figure, we expect that, being the shape of each rate similar, the approach proposed in this work allows to obtain similar partitions of the 24 hours on different days of the week. This enables to focus only on one arbitrary day of the week. Specifically, Tuesday is the day chosen to apply the methodology under study, since the shape of its arrival rate can be considered representative of the other days. In Figure 3, the plot of the hourly average arrival rate for the Tuesdays over the 13 considered weeks is reported, while Figure 4 shows mean and variance of the interarrival times occurred on the first Tuesday of the year 2018. Figure 3: Plot of the average hourly arrival rate for the Tuesdays over the 13 considered weeks of the year. Figure 4: Plot of the average (in solid green) and variance (in dashed red) of the interarrival times for the first Tuesday of year 2018. On the abscissa axis, 3-hours time slots are considered. From this latter figure, we observe that these two statistics have similar values within each 3-hours time slot and this is in accordance with the property of the Poisson probability distribution for which mean and variance coincide. ## 3 Statistical model The arrival process at EDs is usually characterized by a strong within-day variation both in the arrival rate and interarrival times: experimental data show rapid changes in the number of arrivals during the night hours, as opposite to a smoother profile at day time. As we already mentioned in the Introduction, for this reason the ED arrival process is usually modeled as a NHPP. No analytical model is available for the arrival rate $\lambda(t)$, and therefore a suitable representation of the unknown function is needed. A realistic representation can be obtained by averaging the number of arrivals observed in experimental data on suitable intervals over the 24 hours of the day, non necessarily equally spaced. Let $\\{T_{i}\\}$ denote a partition $P$ of the observation period $T=[0,24]$ (hours) in $N$ intervals, and let $\\{\lambda_{i}\\}$ be corresponding sample average rates. Then a piecewise constant approximation of $\lambda(t)$ is written as follows $\lambda_{D}(t)=\sum_{i=1}^{N}\lambda_{i}\,\textbf{1}_{T_{i}}(t),\quad t\in T$ (1) where $\textbf{1}_{T_{i}}(t)$ is 1 for $t\in T_{i}$ and 0 otherwise (the indicator function of set $T_{i}$). Any partiton $P$ gives rise to a different approximation $\lambda_{D}(t)$, depending on the number of intervals and their lengths. Therefore a criterion is needed to select the best partition $P^{\star}$ with some desirable features. First of all, we need to ensure that there is no overdispersion in the arrival data. We refer to the commonly used dispersion test proposed in Kathirgamatamby (1953) and reported in Kim and Whitt (2014a). If it is satisfied, then it is possible to combine arrivals for the same day of the week over different weeks. To this aim, for any partition $P$, let $\\{k_{i}^{r}\\}$ denote the number of arrivals in the $i$-th partition interval $T_{i}$ in the $r$-th week, $r=1,\ldots,m$. Consider the statistics $Ds_{i}=\displaystyle\frac{1}{\mu_{i}}\displaystyle\sum_{r=1}^{m}\left(k_{i}^{r}-\mu_{i}\right)^{2},\quad i=1,\ldots,N,$ where $\mu_{i}=\frac{1}{m}\sum_{r=1}^{m}k_{i}^{r}$ is the average number of arrivals in the given interval for the same day of the week over the considered $m$ weeks. Under the null hypothesis that the counts $\\{k_{i}^{r}\\}$ are a sample of $m$ independent Poisson random variables with the same mean count $\mu_{i}$ (no overdispersion), then $Ds_{i}$ is distributed as $\chi^{2}_{m-1}$, the chi-squared distribution with $m-1$ degrees of freedom. Therefore the null hyphotesis is accepted with $1-\alpha$ confidence level if $Ds_{i}\leq\chi^{2}_{m-1,\alpha},\quad i=1,\ldots,N,$ (2) where $\chi^{2}_{m-1,\alpha}$ is of course the $\alpha$ level critical value of the $\chi^{2}_{m-1}$ distribution. Furthermore, the partition is feasible if data are consistent with NHPP. Namely, if we denote by $k_{i}$ the number of arrivals in each interval $T_{i}=[a_{i},b_{i})$ obtained by considering data of the same weekday, in the same interval, over $m$ weeks, i.e. $k_{i}=\sum_{r=1}^{m}k_{i}^{r}$, $i=1,\ldots,N$, the partition is feasible if each $k_{i}$ has a Poisson distribution with rate $\lambda_{i}$ obtained as $\mu_{i}/(b_{i}-a_{i})$. To check the validity of the Poisson hypothesis, the CU KS test can be performed (see Brown et al. (2005); Kim and Whitt (2014a)). We prefer to use CU KS with respect to Lewis KS test since this latter is highly sensitive to rounding of the data and moreover CU KS test has more power against alternative hypotheses involving exponential interarrival times (see Kim and Whitt (2014b) for a detailed comparison between the effectiveness of the two tests). To perform CU KS test, for any interval $T_{i}=[a_{i},b_{i})$, let $t_{ij}$, $j=1,\ldots,k_{i},$ be the arrival times within the $i$-th interval obtained as union over the $m$ weeks of the arrival times in each $T_{i}$. Now consider the rescaled arrival times defined by $\tau_{ij}=\displaystyle\frac{t_{ij}-a_{i}}{b_{i}-a_{i}}$. The rescaled arrival times, conditionally to the value $k_{i}$, are a collection of i.i.d. random variables uniformly distributed over $[0,1]$. Hence, in any interval we compare the theoretical cumulative distribution function (cdf) $F(t)=t$ with the empirical cdf $F_{i}(t)=\frac{1}{k_{i}}\sum_{j=1}^{k_{i}}\textbf{1}_{\\{\tau_{ij}\leq t\\}},\qquad 0\leq t\leq 1.$ The test statistics is defined as follows $D_{i}=\sup_{0\leq t\leq 1}(|F_{i}(t)-t|).$ (3) The critical value for this test is denoted as $T(k_{i},\alpha)$ and its values can be found on the KS test critical values table. Accordingly, the Poisson hypothesis is accepted if $D_{i}\leq T(k_{i},\alpha),\quad i=1,\ldots,N.$ (4) This test has to be satisfied on each interval $T_{i}$ to qualify the partition $P$ given by $\\{T_{i}\\}$ as feasible, in the sense that CU KS test is satisfied, too. A further restriction is imposed on the feasible partitions. Given the experimental data, realistic partitions can not have a granularity too fine to avoid that some $k_{i}$ being too small may unduly determine the rejection of the CU KS test. To this aim the value of 1 hour was chosen as lower threshold value, taking into account the specific case study considered (see also Figure 3). Now let us evaluate the feasible partitions also in terms of the characteristics of function $\lambda_{D}(t)$. It would be amenable to define a fit error with respect to $\lambda(t)$, which unfortunately is unknown. The problem can be got around by considering a piecewise constant approximation $\lambda_{F}(t)$ over a very fine partition ${P}_{F}$ of $T$. A set of 96 equally space intervals of $15$ minutes was considered and the corresponding average rates $\lambda_{i}^{F}$ were estimated from data. The plot of $\lambda_{i}^{F}$ is reported in Figure 5. Figure 5: Plot of the daily average arrival rate $\lambda_{F}(t)$ with intervals of 15 minutes. The function $\lambda_{F}(t)$ can be considered as an empirical arrival rate model. Note that partition ${P}_{F}$ need not be feasible since it only serves to define the finest piecewise constant approximation of $\lambda(t)$. Therefore the following fit error can be defined $E(P)=\sum_{j=1}^{N}\sum_{i_{j}=1}^{N_{j}}(\lambda_{j}-\lambda_{i_{j}}^{F})^{2}$ (5) where $N_{j}$ is the number of intervals of $15$ minutes contained in $T_{j}$, and identified by the set of indexes $\\{i_{j}\\}\subset\\{1,\ldots,96\\}$. Finally it is also advisable to characterize the “smoothness” of any approximation $\lambda_{D}(t)$ to avoid very gross partitions with high jumps between adjacent intervals by means of the mean squared error $S(P)=\sum_{j=2}^{N}(\lambda_{j}-\lambda_{j-1})^{2}.$ (6) In the following Section 4 the model features illustrated above are organized in a proper optimization procedure that provides the selection of the best partition according to conflicting goals. The approach we propose enables to well address the major two issues raised in Kim and Whitt (2014a) (and reported in the Introduction) when dealing with modelling ED patient arrivals, namely the choice of the intervals and the overdispersion. As concerns the third issue, the data rounding, the arrival times in the data we collected are rounded to seconds (format hh:mm:ss), and actually occurrences of simultaneous arrivals which would cause zero interarrival times are not present. Therefore, we do not need any unrounding procedure. Anyhow, as already pointed out above, the CU KS test we use is not very sensitive to data rounding. ## 4 Statement of the optimization problem Any partition $P=\\{T_{i}\\}$ of $T=[0,24]$ is characterized by the boundary points $\\{x_{i}\\}$ of its intervals and by their number $N$. Let us introduce a vector of variables $x\in\mathbb{Z}^{25}$ such that $T_{i}=[x_{i},x_{i+1}),$ $i=1,\ldots,24$, with $x_{1}=0$ and $x_{25}=24$. Functions in (5) and (6) are indeed functions of $x$, and therefore will be denoted by $E(x)$ and $S(x)$, respectively. Therefore, the objective function that constitutes the selection criterion is given by $f(x)=E(x)+wS(x),$ (7) where $w>0$ is a parameter that controls the weight of the smoothness penalty term with respect to the fit error: the larger $w$, the smaller the difference between average arrival rates in adjacent intervals; this in turn implies that on a steep section of $\lambda_{F}(t)$ an increased number of shorter intervals is adopted to fill the gap with relatively small jumps. The set $\cal P$ of feasible partitions is defined as follows: $\begin{array}[]{l}{\cal P}=\Bigl{\\{}x\in{\mathbb{Z}}^{25}~{}|~{}x_{1}=0,\quad x_{25}=24,\quad x_{i+1}-x_{i}\geq\ell_{i},\quad g_{i}(x)\leq 0,\bigr{.}\cr\cr\bigl{.}\hskip 28.45274pth_{i}(x)\leq 0,\quad i=1,\ldots,N\Bigr{\\}}\end{array}$ (8) where $\displaystyle\ell_{i}$ $\displaystyle=$ $\displaystyle\begin{cases}0\quad\hbox{if}\quad x_{i}=x_{i+1},\\\ \ell\quad\hbox{otherwise},\\\ \end{cases}$ (9) $\displaystyle g_{i}(x)$ $\displaystyle=$ $\displaystyle\begin{cases}0\quad\hbox{if}\quad x_{i}=x_{i+1},\\\ D_{i}-T(k_{i},\alpha)\quad\hbox{otherwise},\\\ \end{cases}$ (10) $\displaystyle h_{i}(x)$ $\displaystyle=$ $\displaystyle\begin{cases}0\quad\hbox{if}\quad x_{i}=x_{i+1},\\\ Ds_{i}-\chi^{2}_{m-1,\alpha}\quad\hbox{otherwise},\\\ \end{cases}$ (11) $i=1,\ldots,N$. The value $\ell$ in (9) denotes the minimum interval length allowed and we assume $\ell\geq 1/4$. Of course, constraints $g_{i}(x)\leq 0$ represents the satisfaction of the CU KS test in (4), while constraints $h_{i}(x)\leq 0$ concern the dispersion test in (2). Therefore, the best piecewise constant approximation $\lambda_{D}^{\star}(t)$ of the time-varying arrival rate $\lambda(t)$ is obtained by solving the following black-box optimization problem: $\begin{split}\max~{}~{}&f(x)\\\ s.t.~{}~{}&x\in{\cal P}.\\\ \end{split}$ (12) We highlight that the idea to use as constraints of the optimization problem a test to validate the underlying statistical hypothesis on data along with a dispersion test is completely novel in the framework of modeling ED patient arrivals process. The only proposal which use a similar approach is in our previous paper De Santis et al. (2020). It is important to note that in (7) the objective function has not analytical structure with respect to the independent variables and it can only be computed by a data-driven procedure once the $x_{i}$’s values are given. The same is true for the constraints $g_{i}(x)$ and $h_{i}(x)$ in (8), too. Therefore the problem in hand is an integer nonlinear constrained black-box problem, and both the objective function and the constraints are relatively expensive to compute and this makes it difficult to efficiently solve. In fact, classical optimization methods either can not be applied (since based on the analytic knowledge of the functions involved) or they are not efficient especially when evaluating the functions at a given point is very computationally expensive. Therefore to tackle problem (12) we turned our attention to the class of Derivative-Free Optimization and black-box methods (see, e.g., Audet and Hare (2017); Conn et al. (2009); Larson et al. (2019)). More in particular, we adopt the algorithmic framework recently proposed in Liuzzi et al. (2020). It represents a novel strategy for solving black-box problems with integer variables and it is based on the use of suited search directions and a nonmonotone linesearch procedure. Moreover, it can handle generally-constrained problems by using a penalty approach. We refer to Liuzzi et al. (2020) for a detailed description and we only highlight that the results reported in Liuzzi et al. (2020) clearly show that this algorithm framework is particularly efficient in tackling black-box problems like the one in (12). In particular, the effectiveness of the adopted exploration strategy with respect to state-of-the-art methods for black-box is shown. This is due to the fact that the approach proposed in Liuzzi et al. (2020) combines computational efficiency with a high level of reliability. ## 5 Experimental results In this section we report the results of an extensive experimentation on data concerning the case study described in Section 2, namely the ED patient arrivals collected in the first $m$ weeks of year 2018. Different values of the number of weeks $m$ have been considered. Standard significance level $\alpha=0.05$ is used the CU KS and dispersion tests. As regards the optimization problem in hand the value of $\ell$ in (9) is set to $1$ hour. Moreover, it is important to note that different values of the weight $w$ in the objective function (7) lead to various piecewise constant approximations with a different fitting accuracy and degree of regularity. Therefore, we performed a careful tuning of this parameter, aiming at determining a value which represents a good trade-off between a small fit error and the smoothness of the approximation. As concerns the parameter values of the optimization algorithm used in our experimentation, we used the default ones (see Liuzzi et al. (2020)). The stopping criterion is based on the maximum number of function evaluations set to 5000. As starting point $x^{0}$ of the optimization algorithm we adopt the following $x^{0}_{i}=i-1,\qquad i=1,\ldots,25,$ (13) which corresponds to the case of 24 intervals of unitary length. This choice is a commonly used partition in most of the approaches proposed in literature (see e.g. Ahalt et al. (2018); Kim and Whitt (2014a)). Table LABEL:tab:CUKSTest-initial in the Appendix reports the results of CU KS and dispersion tests applied to the partition corresponding to the starting point $x^{0}$, considering $m=13$ weeks. In particular, in Table LABEL:tab:CUKSTest- initial for each one-hour slot the sample size $k_{i}$ is reported along with the $p$-value and the acceptance/rejection of the null hypothesis of the corresponding test. We observe that the arrivals are not overdispersed in any interval of the partition corresponding to $x^{0}$, i.e. all the constraints $h_{i}(x)\leq 0$ are satisfied and this allows us to combine data for the same day of the week over successive weeks. However, this partition is even unfeasible, i.e. $g_{i}(x)>0$ for some $i$; this corresponds to reject the statistical hypothesis on some $T_{i}$. Notwithstanding, even if the starting point is unfeasible, the optimization algorithm we use is able find a feasible solution which minimizes the objective function. As we already mentioned, the choice of a proper value for the weight $w$ in the objective function (7) is important and not straightforward. On the other hand, the number $m$ of the considered weeks also affects both the accuracy of the approximation, through the average rates estimated on each interval, and the consistency of the results, which is ensured by constraints (10) and (11). However, while $w$ is related to the statement of the optimization problem (12) and it can be arbitrarily chosen, the choice of $m$ is strictly connected to the available data. In (Kim and Whitt, 2014a, Section 4), the authors assert that, having 10 arrivals in the one-hour slot 9–10 a.m., it is necessary to combine data over 20 weeks in order to have a sufficient sample size (200 patient arrivals). However, being their approach based on equally- spaced intervals, one-hour slots are also adopted during off-peak hours, for instance during the night. This implies that the sample size corresponding to data combination over 20 weeks for these slots could no longer be sufficient to guarantee good results. This is clearly pointed out in Table LABEL:tab:CUKSTest-initial where the sample size $k_{i}$ corresponding to some of the one-hour night slots is very low considering $m=13$ weeks and it remains insufficient even if 26 weeks are considered (see subsequent Table LABEL:tab:Test-initial). The approach we propose overcomes this drawback since, for each choice of $m$, we determine the length of the intervals as solution of the optimization problem (12). Of course, there could be values of $m$ such that problem (12) has not feasible solutions, i.e. a partition such that the NHPP hypothesis holds and the results are consistent does not exists for such $m$. In order to deeper examine how the parameters $w$ and $m$ affect the optimal partition, we performed a sensitivity analysis, focusing first on the case with fixed $m$ and $w$ varying. In particular, we have chosen to focus on $m=13$ weeks, which enables to achieve an optimal solution by running the optimization algorithm without overly computational burden. Anyhow, we expect that no substantial changes in the conclusions would be obtained with different values of $m$ and this is confirmed by further experimentation whose results are not reported here for the sake of brevity. This analysis allows us to obtain several partitions that may be considered for a proper fine-tuning of $w$. In particular, we consider different values of $w$ within the set $\\{0,~{}0.1,~{}1,~{}10,~{}10^{3}\\}$. Table LABEL:tab:CUKSTest-optimal in the Appendix reports the optimal partitions obtained by solving problem (12) for these values of $w$. In particular, Table LABEL:tab:CUKSTest-optimal includes the intervals of the partition, the value of the sample size $k_{i}$ corresponding to each interval over $13$ weeks and the results of the CU KS and dispersion tests, namely the $p$-value and the acceptance/rejection of the null hypothesis of the corresponding test. In Figure 6, for graphical comparison, we report the plots of the empirical arrival rate model $\lambda_{F}(t)$ and its piecewise constant approximation $\lambda_{D}(t)$ corresponding to the optimal partitions obtained. Figure 6: Graphical comparison between the empirical arrival rate model $\lambda_{F}(t)$ (in green) and the piecewise constant approximation $\lambda_{D}(t)$ (in red) corresponding to the optimal partition obtained by solving problem (12) for different values of the parameter $w$. From top to bottom: $w=0,0.1,1,10,10^{3}$. Two effects can be clearly observed as $w$ increases: on the one hand, on steep sections of $\lambda_{F}(t)$, shorter intervals are adopted to reduce large gaps between adjacent intervals; on the other hand, when $\lambda_{F}(t)$ is approximately flat, a lower number of intervals may be sufficient to guarantee small gaps. This is confirmed by the two top plots in Figure 6 which correspond to $w=0$ and $w=0.1$. In fact, in the first plot ($w=0$) where only the fit error is included in the objective function and in the second one ($w=0.1$) where anyhow the fit error is the dominant term of the objective function, the optimal partition is composed by a relatively large number of intervals. In particular, in the partition corresponding to $w=0.1$, fewer intervals are adopted during the day time. As expected, a smaller number of intervals is attained when $w=1$, $w=10$ and $w=10^{3}$. Note that, since on the steep section corresponding to the time slot 7:00–10:00 a.m. the maximum number of allowed intervals (due to the lower threshold value of one hour given by the choice $\ell=1$ in (9)) is already used, the only way to decrease the smoothness term of the objective function is to enlarge the intervals during both the day and the night. It is worth noting that for $w=10^{3}$, the number of intervals increases if compared with the case $w=10$. This occurs to offset the increase in the fit error term due to the use of a smaller number of intervals on the flatter sections. As a consequence, the partition has an unexpected interval at the end of the day. We point out that for each value of $w$, the optimization algorithm finds an optimal partition (of course feasible with respect to all the constraints), despite some constraints related to the CU KS test are violated in the initial partition, i.e. the one corresponding to $x^{0}$ in (13), namely the standard assumption of one-hour slots usually adopted. This means that the used data are in accordance with the NHPP hypothesis and they are sufficient to appropriately define the piecewise constant approximation of the ED arrival rate. Conversely, when the optimization algorithm does not find a feasible partition, the CU KS test or the dispersion test related to some $T_{i}$ are never satisfied. This implies that the process is not conforming to the NHPP hypothesis or that the data are overdispersed. This is clearly highlighted by our subsequent experimentation where we set $w=1$, letting $m$ varying within the set $\\{5,9,17,22,26\\}$. First, in Table LABEL:tab:Test-initial in the Appendix we report the results of CU KS and dispersion tests applied to the partition corresponding to the starting point in $x^{0}$(13), for these different values of $m$. Once more, this table evidences that the use of equally-spaced intervals of one-hour length during the whole day can be inappropriate. As an example, see the results of the tests on the time slot 02:00–03:00. Moreover, note that, for all these values of $m$, the initial partition corresponding to the starting point $x_{0}$ is infeasible, except when $m=5$. Indeed, the constraints corresponding to CU KS and dispersion tests are violated for some $T_{i}$, meaning that the validity of the standard assumption of one-hour time slots strongly depends on the time period considered for using the collected data. To this aim, a strength of our approach is its ability to assist in the selection of a reasonable value for $m$. If there is no value of $m$ such that the optimization algorithm determines an optimal solution (due to unfeasibility), then it may be inappropriate to consider the ED arrival process in hand as NHPP. The subsequent Table LABEL:tab:Test-infeasible-final includes the optimal partitions obtained by solving problem (12) for the considered values of $m\in\\{5,9,17,22,26\\}$. Like the previous tables, Table LABEL:tab:Test- infeasible-final includes the intervals of the partition, the value of the sample size $k_{i}$ corresponding to each interval and the results of CU KS and dispersion tests. For all the considered values of $m$ the optimization algorithm determines an optimal solution with the only exception of $m=26$. In this latter case, the maximum number of function evaluations allowed to the optimization algorithm is not enough to compute an optimal solution: in fact, we obtain an unfeasible solution since the CU KS test related to the last interval of the day is not satisfied. This could be partially unexpected, since more accurate results should be obtained when considering a greater sample size. However, by adding the last four weeks (passing from $m=22$ to $m=26$) which corresponds to the month of June, the data become affected by a seasonal trend and the NHPP assumption is no longer valid. In Figure 7 we report a graphical comparison between the empirical arrival rate model $\lambda_{F}(t)$ and the piecewise constant approximation $\lambda_{D}(t)$ corresponding to the optimal partitions obtained for the considered values of $m$. We observe that the variability of $\lambda_{F}(t)$ reduces as the value of $m$ increases since averaging on more data leads to flattening the fluctuation. Despite these rapid oscillations and unlike the other considered values of $m$, for $m=5$ the empirical model $\lambda_{F}(t)$ shows a constant trend during both the night and day hours. This results in a piecewise constant approximation $\lambda_{D}(t)$ that is flat in all the time slots of the 24 hours of the day except the ones related to the morning hours, for which many intervals are used. In fact, to guarantee a good fitting error between $\lambda_{D}(t)$ and $\lambda_{F}(t)$, it would be necessary to use shorter intervals, but this is not allowed by the choice $\ell=1$ in the constraints (9). For the other considered values of $m$, the number of intervals increases, leading to partitions that improve the fitting error if compared with the case $m=5$. In particular, we observe that the piecewise constant approximation $\lambda_{D}(t)$ obtained for $m=22$ benefits from the lower fluctuations resulting from averaging more data. Therefore, as expected, using the maximum number of available data leads to the most accurate piecewise constant approximation. However, when considering too many data, seasonal phenomena could give rise to the rejection of the null hypothesis of the considered tests, as observed for the case $m=26$. Moreover, as highlighted at the end of Section 5 in Kim and Whitt (2014a), a tendency to reject the NHPP hypothesis (i.e. the null hypothesis of the CU KS test) may be encountered when the sample size is large. In fact, a larger sample size requires a stronger evidence of the null hypothesis in order for the test to be passed. Notwithstanding, our approach is able to overcome these drawbacks, providing us with an optimal strategy to identify the best way of using the collected data. Figure 7: Graphical comparison between the empirical arrival rate model $\lambda_{F}(t)$ (in green) and the piecewise constant approximation $\lambda_{D}(t)$ (in red) corresponding to the optimal partition obtained by solving problem (12) for different values of the parameter $m$. From top to bottom: $m=5,9,17,22,26$. ## 6 Conclusions In this work, we examined the arrival process to EDs by providing a novel methodology that is able to improve the reliability of the modelling approaches frequently used to deal with this complex system, i.e. the Discrete Event Simulation modelling. In accordance with the literature, we adopted the standard assumption of representing the ED arrival process as a NHPP, which is suitable for modelling strongly time-varying processes. In particular, the final goal of the proposed approach is to accurately estimate the unknown arrival rate, i.e. the time-dependent parameter of the NHPP, by using a reasonable piecewise constant approximation. To this aim, an integer nonlinear black-box optimization problem is solved to determine the optimal partition of the 24 hours into a suitable number of non equally spaced intervals. To guarantee the reliability of this estimation procedure, two types of statistical tests are considered as constraints for each interval of any candidate partition: the CU KS test must be satisfied to ensure the consistency between the NHPP hypothesis and the ED arrivals; the dispersion test must be satisfied to avoid the overdispersion of data. To the best of our knowledge, our methodology represents the first optimization-based approach adopted for determining the best stochastic model for the ED arrival process. The extensive experimentation we performed on data collected from an ED of a big hospital in Italy, shows that our approach is able to find a piecewise constant approximation which represents a good trade-off between a small fit error with the empirical arrival rate model and the smoothness of the approximation. This result is accomplished by the optimization algorithm, despite some constraints in the starting point, which corresponds to the commonly adopted partition composed by one-hour time slots, are violated. Moreover, some significant sensitivity analyses are performed to investigate the fine-tuning of the two parameters affecting the quality of the piecewise constant approximation: the weight of the smoothness of the approximation in the objective function (with respect to the fit error) and the number of weeks considered from the arrivals data. While the former can be arbitrarily chosen by a user according to the desired level of smoothness, the latter affects the accuracy of the arrival rate estimation. In general, the more weeks are considered, the more accurate is the arrival rate approximation, as long as the NHPP assumption still holds and the data do not become overdispersed. As regards future work, in order to deeper analyze the robustness of the proposed approach, we could use alternative statistical tests, such as the Lewis and the Log tests described in Kim and Whitt (2014a), in place of the CU KS test. Moreover, whenever Discrete Event Simulation modelling is the chosen methodology to study ED operation, a model calibration approach could be also used to determine the best value of the weight used in the objective function to penalize the “smoothness term”. In fact, the optimal value of this parameter could be obtained by minimizing the deviation between the simulation outputs and the corresponding key performance indicators computed through the data. This enables to obtain a representation of the ED arrival process that leads to an improved simulation model of the system under study. ## Appendix A Appendix In this Appendix we report the detailed results of the CU KS and dispersion tests related to the partitions considered throughout the paper. Table 1: Results of the CU KS and dispersion tests (with a significance level of 0.05) applied to each interval of the partition corresponding to the starting point $x^{0}$. The considered number of weeks is $m=13$. For each interval of each partition, the sample size of the dispersion test is $m$. $H_{0}$ denotes the null hypothesis of the corresponding test. | CU KS test | Dispersion test ---|---|--- Interval | $k_{i}$ | $p$-value | $H_{0}$ | $p$-value | $H_{0}$ 00:00 – 01:00 | $48$ | $0.836$ | accepted | $0.801$ | accepted 01:00 – 02:00 | $38$ | $0.950$ | accepted | $0.450$ | accepted 02:00 – 03:00 | $22$ | $0.027$ | rejected | $0.521$ | accepted 03:00 – 04:00 | $24$ | $0.752$ | accepted | $0.652$ | accepted 04:00 – 05:00 | $21$ | $0.668$ | accepted | $0.366$ | accepted 05:00 – 06:00 | $32$ | $0.312$ | accepted | $0.524$ | accepted 06:00 – 07:00 | $29$ | $0.634$ | accepted | $0.538$ | accepted 07:00 – 08:00 | $59$ | $0.424$ | accepted | $0.252$ | accepted 08:00 – 09:00 | $86$ | $0.393$ | accepted | $0.734$ | accepted 09:00 – 10:00 | $136$ | $0.635$ | accepted | $0.803$ | accepted 10:00 – 11:00 | $143$ | $0.039$ | rejected | $0.966$ | accepted 11:00 – 12:00 | $154$ | $0.325$ | accepted | $0.999$ | accepted 12:00 – 13:00 | $132$ | $0.858$ | accepted | $0.948$ | accepted 13:00 – 14:00 | $121$ | $0.738$ | accepted | $0.984$ | accepted 14:00 – 15:00 | $125$ | $0.885$ | accepted | $0.500$ | accepted 15:00 – 16:00 | $127$ | $0.928$ | accepted | $0.610$ | accepted 16:00 – 17:00 | $117$ | $0.479$ | accepted | $0.987$ | accepted 17:00 – 18:00 | $111$ | $0.769$ | accepted | $0.516$ | accepted 18:00 – 19:00 | $102$ | $0.458$ | accepted | $0.912$ | accepted 19:00 – 20:00 | $100$ | $0.095$ | accepted | $0.527$ | accepted 20:00 – 21:00 | $101$ | $0.656$ | accepted | $0.586$ | accepted 21:00 – 22:00 | $115$ | $0.763$ | accepted | $0.604$ | accepted 22:00 – 23:00 | $101$ | $0.916$ | accepted | $0.305$ | accepted 23:00 – 24:00 | $70$ | $0.864$ | accepted | $0.104$ | accepted Table 2: Results of the CU KS and dispersion tests (with a significance level of 0.05) applied to each interval of the optimal partition obtained by solving problem (12) for different values of the parameter $w$, with $m$ fixed to 13 weeks. From top to bottom: $w=0,0.1,1,10,10^{3}$. For each interval of each partition, the sample size of the dispersion test is equal to $m$. $H_{0}$ denotes the null hypothesis of the corresponding test. | CU KS test | Dispersion test ---|---|--- $w$ | Interval | $k_{i}$ | $p$-value | $H_{0}$ | $p$-value | $H_{0}$ | 00:00 – 01:00 | $48$ | $0.836$ | accepted | $0.801$ | accepted | 01:00 – 02:00 | $38$ | $0.950$ | accepted | $0.450$ | accepted | 02:00 – 05:00 | $67$ | $0.504$ | accepted | $0.100$ | accepted | 05:00 – 06:00 | $32$ | $0.312$ | accepted | $0.524$ | accepted | 06:00 – 07:00 | $29$ | $0.634$ | accepted | $0.538$ | accepted | 07:00 – 08:00 | $59$ | $0.424$ | accepted | $0.252$ | accepted | 08:00 – 09:00 | $86$ | $0.393$ | accepted | $0.734$ | accepted | 09:00 – 10:00 | $136$ | $0.635$ | accepted | $0.803$ | accepted 0 | 10:00 – 12:00 | $297$ | $0.433$ | accepted | $0.994$ | accepted | 12:00 – 13:00 | $132$ | $0.858$ | accepted | $0.948$ | accepted | 13:00 – 16:00 | $373$ | $0.958$ | accepted | $0.502$ | accepted | 16:00 – 17:00 | $117$ | $0.479$ | accepted | $0.987$ | accepted | 17:00 – 19:00 | $213$ | $0.999$ | accepted | $0.937$ | accepted | 19:00 – 20:00 | $100$ | $0.095$ | accepted | $0.527$ | accepted | 20:00 – 21:00 | $101$ | $0.656$ | accepted | $0.586$ | accepted | 21:00 – 22:00 | $115$ | $0.763$ | accepted | $0.604$ | accepted | 22:00 – 23:00 | $101$ | $0.916$ | accepted | $0.305$ | accepted | 23:00 – 24:00 | $70$ | $0.864$ | accepted | $0.104$ | accepted | 00:00 – 01:00 | $48$ | $0.836$ | accepted | $0.801$ | accepted | 01:00 – 02:00 | $38$ | $0.950$ | accepted | $0.450$ | accepted | 02:00 – 05:00 | $67$ | $0.504$ | accepted | $0.100$ | accepted | 05:00 – 06:00 | $32$ | $0.312$ | accepted | $0.524$ | accepted | 06:00 – 07:00 | $29$ | $0.634$ | accepted | $0.538$ | accepted | 07:00 – 08:00 | $59$ | $0.424$ | accepted | $0.252$ | accepted | 08:00 – 09:00 | $86$ | $0.393$ | accepted | $0.734$ | accepted | 09:00 – 10:00 | $136$ | $0.635$ | accepted | $0.803$ | accepted 0.1 | 10:00 – 12:00 | $297$ | $0.433$ | accepted | $0.994$ | accepted | 12:00 – 13:00 | $132$ | $0.858$ | accepted | $0.948$ | accepted | 13:00 – 16:00 | $373$ | $0.958$ | accepted | $0.502$ | accepted | 16:00 – 17:00 | $117$ | $0.479$ | accepted | $0.987$ | accepted | 17:00 – 18:00 | $111$ | $0.769$ | accepted | $0.516$ | accepted | 18:00 – 22:00 | $418$ | $0.660$ | accepted | $0.987$ | accepted | 22:00 – 23:00 | $101$ | $0.916$ | accepted | $0.305$ | accepted | 23:00 – 24:00 | $70$ | $0.864$ | accepted | $0.104$ | accepted | 00:00 – 02:00 | $86$ | $0.825$ | accepted | $0.709$ | accepted | 02:00 – 05:00 | $67$ | $0.504$ | accepted | $0.100$ | accepted | 05:00 – 07:00 | $61$ | $0.739$ | accepted | $0.313$ | accepted | 07:00 – 08:00 | $59$ | $0.424$ | accepted | $0.252$ | accepted 1 | 08:00 – 09:00 | $86$ | $0.393$ | accepted | $0.734$ | accepted | 09:00 – 15:00 | $811$ | $0.073$ | accepted | $0.955$ | accepted | 15:00 – 16:00 | $127$ | $0.928$ | accepted | $0.610$ | accepted | 16:00 – 17:00 | $117$ | $0.479$ | accepted | $0.987$ | accepted | 17:00 – 18:00 | $111$ | $0.769$ | accepted | $0.516$ | accepted | 18:00 – 24:00 | $589$ | $0.059$ | accepted | $0.922$ | accepted | 00:00 – 02:00 | $86$ | $0.825$ | accepted | $0.709$ | accepted | 02:00 – 05:00 | $67$ | $0.504$ | accepted | $0.100$ | accepted | 05:00 – 07:00 | $61$ | $0.739$ | accepted | $0.313$ | accepted | 07:00 – 08:00 | $59$ | $0.424$ | accepted | $0.252$ | accepted 10 | 08:00 – 09:00 | $86$ | $0.393$ | accepted | $0.734$ | accepted | 09:00 – 15:00 | $811$ | $0.073$ | accepted | $0.955$ | accepted | 15:00 – 16:00 | $127$ | $0.928$ | accepted | $0.610$ | accepted | 16:00 – 17:00 | $117$ | $0.479$ | accepted | $0.987$ | accepted | 17:00 – 24:00 | $700$ | $0.063$ | accepted | $0.720$ | accepted | 00:00 – 02:00 | $86$ | $0.825$ | accepted | $0.709$ | accepted | 02:00 – 06:00 | $99$ | $0.451$ | accepted | $0.162$ | accepted | 06:00 – 07:00 | $29$ | $0.634$ | accepted | $0.538$ | accepted | 07:00 – 08:00 | $59$ | $0.424$ | accepted | $0.252$ | accepted $10^{3}$ | 08:00 – 09:00 | $86$ | $0.393$ | accepted | $0.734$ | accepted | 09:00 – 15:00 | $811$ | $0.073$ | accepted | $0.955$ | accepted | 15:00 – 16:00 | $127$ | $0.928$ | accepted | $0.610$ | accepted | 16:00 – 17:00 | $117$ | $0.479$ | accepted | $0.987$ | accepted | 17:00 – 18:00 | $111$ | $0.769$ | accepted | $0.516$ | accepted | 18:00 – 22:00 | $418$ | $0.660$ | accepted | $0.987$ | accepted | 22:00 – 24:00 | $171$ | $0.053$ | accepted | $0.681$ | accepted Table 3: Results of the CU KS and dispersion tests (with a significance level of 0.05) applied to each interval of the partition corresponding to the starting point $x^{0}$. From top to bottom: $m=5,9,17,22,26$. For each interval of each partition, the sample size of the dispersion test is $m$. $H_{0}$ denotes the null hypothesis of the corresponding test. | CU KS test | Dispersion test ---|---|--- $m$ | Interval | $k_{i}$ | $p$-value | $H_{0}$ | $p$-value | $H_{0}$ | 00:00 – 01:00 | $20$ | $0.167$ | accepted | $0.240$ | accepted | 01:00 – 02:00 | $11$ | $0.616$ | accepted | $0.151$ | accepted | 02:00 – 03:00 | $7$ | $0.887$ | accepted | $0.160$ | accepted | 03:00 – 04:00 | $7$ | $0.892$ | accepted | $0.683$ | accepted | 04:00 – 05:00 | $12$ | $0.217$ | accepted | $0.856$ | accepted | 05:00 – 06:00 | $8$ | $0.426$ | accepted | $0.219$ | accepted | 06:00 – 07:00 | $15$ | $0.884$ | accepted | $0.504$ | accepted | 07:00 – 08:00 | $27$ | $0.820$ | accepted | $0.164$ | accepted | 08:00 – 09:00 | $35$ | $0.875$ | accepted | $0.534$ | accepted | 09:00 – 10:00 | $50$ | $0.378$ | accepted | $0.844$ | accepted | 10:00 – 11:00 | $48$ | $0.083$ | accepted | $0.884$ | accepted | 11:00 – 12:00 | $59$ | $0.484$ | accepted | $0.966$ | accepted 5 | 12:00 – 13:00 | $51$ | $0.594$ | accepted | $0.765$ | accepted | 13:00 – 14:00 | $47$ | $0.651$ | accepted | $0.689$ | accepted | 14:00 – 15:00 | $44$ | $0.817$ | accepted | $0.412$ | accepted | 15:00 – 16:00 | $45$ | $0.811$ | accepted | $0.168$ | accepted | 16:00 – 17:00 | $47$ | $0.679$ | accepted | $0.987$ | accepted | 17:00 – 18:00 | $49$ | $0.486$ | accepted | $0.534$ | accepted | 18:00 – 19:00 | $37$ | $0.731$ | accepted | $0.344$ | accepted | 19:00 – 20:00 | $35$ | $0.436$ | accepted | $0.839$ | accepted | 20:00 – 21:00 | $44$ | $0.904$ | accepted | $0.794$ | accepted | 21:00 – 22:00 | $43$ | $0.459$ | accepted | $0.693$ | accepted | 22:00 – 23:00 | $32$ | $0.967$ | accepted | $0.667$ | accepted | 23:00 – 24:00 | $31$ | $0.306$ | accepted | $0.552$ | accepted | 00:00 – 01:00 | $33$ | $0.106$ | accepted | $0.527$ | accepted | 01:00 – 02:00 | $22$ | $0.658$ | accepted | $0.488$ | accepted | 02:00 – 03:00 | $13$ | $0.031$ | rejected | $0.390$ | accepted | 03:00 – 04:00 | $14$ | $0.258$ | accepted | $0.857$ | accepted | 04:00 – 05:00 | $16$ | $0.441$ | accepted | $0.471$ | accepted | 05:00 – 06:00 | $22$ | $0.707$ | accepted | $0.335$ | accepted | 06:00 – 07:00 | $23$ | $0.580$ | accepted | $0.608$ | accepted | 07:00 – 08:00 | $48$ | $0.500$ | accepted | $0.484$ | accepted | 08:00 – 09:00 | $54$ | $0.338$ | accepted | $0.573$ | accepted | 09:00 – 10:00 | $97$ | $0.391$ | accepted | $0.886$ | accepted | 10:00 – 11:00 | $97$ | $0.149$ | accepted | $0.836$ | accepted | 11:00 – 12:00 | $108$ | $0.384$ | accepted | $0.999$ | accepted 9 | 12:00 – 13:00 | $95$ | $0.911$ | accepted | $0.821$ | accepted | 13:00 – 14:00 | $82$ | $0.733$ | accepted | $0.923$ | accepted | 14:00 – 15:00 | $75$ | $0.979$ | accepted | $0.753$ | accepted | 15:00 – 16:00 | $89$ | $0.909$ | accepted | $0.456$ | accepted | 16:00 – 17:00 | $82$ | $0.429$ | accepted | $0.923$ | accepted | 17:00 – 18:00 | $78$ | $0.804$ | accepted | $0.596$ | accepted | 18:00 – 19:00 | $69$ | $0.277$ | accepted | $0.734$ | accepted | 19:00 – 20:00 | $69$ | $0.218$ | accepted | $0.477$ | accepted | 20:00 – 21:00 | $72$ | $0.731$ | accepted | $0.731$ | accepted | 21:00 – 22:00 | $75$ | $0.449$ | accepted | $0.541$ | accepted | 22:00 – 23:00 | $60$ | $0.989$ | accepted | $0.681$ | accepted | 23:00 – 24:00 | $48$ | $0.521$ | accepted | $0.689$ | accepted | 00:00 – 01:00 | $73$ | $0.729$ | accepted | $0.472$ | accepted | 01:00 – 02:00 | $48$ | $0.708$ | accepted | $0.291$ | accepted | 02:00 – 03:00 | $39$ | $0.009$ | rejected | $0.010$ | rejected | 03:00 – 04:00 | $32$ | $0.203$ | accepted | $0.622$ | accepted | 04:00 – 05:00 | $28$ | $0.706$ | accepted | $0.652$ | accepted | 05:00 – 06:00 | $38$ | $0.125$ | accepted | $0.607$ | accepted | 06:00 – 07:00 | $35$ | $0.908$ | accepted | $0.327$ | accepted | 07:00 – 08:00 | $70$ | $0.788$ | accepted | $0.075$ | accepted | 08:00 – 09:00 | $121$ | $0.786$ | accepted | $0.577$ | accepted | 09:00 – 10:00 | $174$ | $0.421$ | accepted | $0.729$ | accepted | 10:00 – 11:00 | $186$ | $0.332$ | accepted | $0.939$ | accepted | 11:00 – 12:00 | $203$ | $0.474$ | accepted | $0.999$ | accepted 17 | 12:00 – 13:00 | $176$ | $0.698$ | accepted | $0.986$ | accepted | 13:00 – 14:00 | $164$ | $0.589$ | accepted | $0.992$ | accepted | 14:00 – 15:00 | $161$ | $0.983$ | accepted | $0.570$ | accepted | 15:00 – 16:00 | $168$ | $0.506$ | accepted | $0.815$ | accepted | 16:00 – 17:00 | $153$ | $0.361$ | accepted | $0.996$ | accepted | 17:00 – 18:00 | $149$ | $0.596$ | accepted | $0.528$ | accepted | 18:00 – 19:00 | $134$ | $0.761$ | accepted | $0.909$ | accepted | 19:00 – 20:00 | $140$ | $0.101$ | accepted | $0.637$ | accepted | 20:00 – 21:00 | $141$ | $0.709$ | accepted | $0.760$ | accepted | 21:00 – 22:00 | $153$ | $0.938$ | accepted | $0.855$ | accepted | 22:00 – 23:00 | $129$ | $0.887$ | accepted | $0.393$ | accepted | 23:00 – 24:00 | $94$ | $0.950$ | accepted | $0.296$ | accepted | 00:00 – 01:00 | $95$ | $0.509$ | accepted | $0.720$ | accepted | 01:00 – 02:00 | $70$ | $0.938$ | accepted | $0.529$ | accepted | 02:00 – 03:00 | $52$ | $0.008$ | rejected | $0.022$ | rejected | 03:00 – 04:00 | $36$ | $0.094$ | accepted | $0.507$ | accepted | 04:00 – 05:00 | $34$ | $0.536$ | accepted | $0.420$ | accepted | 05:00 – 06:00 | $46$ | $0.045$ | rejected | $0.703$ | accepted | 06:00 – 07:00 | $48$ | $0.833$ | accepted | $0.590$ | accepted | 07:00 – 08:00 | $83$ | $0.805$ | accepted | $0.062$ | accepted | 08:00 – 09:00 | $165$ | $0.576$ | accepted | $0.108$ | accepted | 09:00 – 10:00 | $219$ | $0.105$ | accepted | $0.737$ | accepted | 10:00 – 11:00 | $235$ | $0.282$ | accepted | $0.960$ | accepted | 11:00 – 12:00 | $274$ | $0.585$ | accepted | $0.962$ | accepted 22 | 12:00 – 13:00 | $233$ | $0.956$ | accepted | $0.984$ | accepted | 13:00 – 14:00 | $216$ | $0.515$ | accepted | $0.999$ | accepted | 14:00 – 15:00 | $207$ | $0.872$ | accepted | $0.789$ | accepted | 15:00 – 16:00 | $213$ | $0.841$ | accepted | $0.905$ | accepted | 16:00 – 17:00 | $204$ | $0.491$ | accepted | $0.999$ | accepted | 17:00 – 18:00 | $192$ | $0.534$ | accepted | $0.683$ | accepted | 18:00 – 19:00 | $173$ | $0.818$ | accepted | $0.968$ | accepted | 19:00 – 20:00 | $177$ | $0.072$ | accepted | $0.768$ | accepted | 20:00 – 21:00 | $181$ | $0.655$ | accepted | $0.681$ | accepted | 21:00 – 22:00 | $196$ | $0.977$ | accepted | $0.810$ | accepted | 22:00 – 23:00 | $167$ | $0.688$ | accepted | $0.412$ | accepted | 23:00 – 24:00 | $118$ | $0.963$ | accepted | $0.209$ | accepted | 00:00 – 01:00 | $112$ | $0.171$ | accepted | $0.679$ | accepted | 01:00 – 02:00 | $75$ | $0.933$ | accepted | $0.377$ | accepted | 02:00 – 03:00 | $67$ | $0.012$ | rejected | $0.053$ | accepted | 03:00 – 04:00 | $46$ | $0.458$ | accepted | $0.450$ | accepted | 04:00 – 05:00 | $38$ | $0.987$ | accepted | $0.465$ | accepted | 05:00 – 06:00 | $57$ | $0.308$ | accepted | $0.535$ | accepted | 06:00 – 07:00 | $56$ | $0.935$ | accepted | $0.739$ | accepted | 07:00 – 08:00 | $100$ | $0.882$ | accepted | $0.128$ | accepted | 08:00 – 09:00 | $198$ | $0.566$ | accepted | $0.142$ | accepted | 09:00 – 10:00 | $259$ | $0.341$ | accepted | $0.844$ | accepted | 10:00 – 11:00 | $289$ | $0.091$ | accepted | $0.942$ | accepted | 11:00 – 12:00 | $320$ | $0.725$ | accepted | $0.984$ | accepted 26 | 12:00 – 13:00 | $274$ | $0.915$ | accepted | $0.996$ | accepted | 13:00 – 14:00 | $257$ | $0.228$ | accepted | $0.999$ | accepted | 14:00 – 15:00 | $243$ | $0.872$ | accepted | $0.835$ | accepted | 15:00 – 16:00 | $242$ | $0.574$ | accepted | $0.892$ | accepted | 16:00 – 17:00 | $236$ | $0.630$ | accepted | $0.942$ | accepted | 17:00 – 18:00 | $231$ | $0.808$ | accepted | $0.753$ | accepted | 18:00 – 19:00 | $204$ | $0.682$ | accepted | $0.980$ | accepted | 19:00 – 20:00 | $209$ | $0.170$ | accepted | $0.830$ | accepted | 20:00 – 21:00 | $219$ | $0.610$ | accepted | $0.735$ | accepted | 21:00 – 22:00 | $237$ | $0.803$ | accepted | $0.905$ | accepted | 22:00 – 23:00 | $198$ | $0.614$ | accepted | $0.366$ | accepted | 23:00 – 24:00 | $147$ | $0.972$ | accepted | $0.032$ | accepted Table 4: Results of the CU KS and dispersion tests (with a significance level of 0.05) applied to each interval of the final (infeasible) partition obtained by solving problem 12 for different values of the parameter $m$, with $w$ fixed to 1. From top to bottom: $m=5,9,17,22,26$. For each interval of each partition, the sample size of the dispersion test is $m$. $H_{0}$ denotes the null hypothesis of the corresponding test. | CU KS test | Dispersion test ---|---|--- $m$ | Interval | $k_{i}$ | $p$-value | $H_{0}$ | $p$-value | $H_{0}$ | 00:00 – 06:00 | $65$ | $0.068$ | accepted | $0.472$ | accepted | 06:00 – 07:00 | $15$ | $0.884$ | accepted | $0.504$ | accepted | 07:00 – 08:00 | $27$ | $0.820$ | accepted | $0.164$ | accepted | 08:00 – 09:00 | $35$ | $0.875$ | accepted | $0.534$ | accepted | 09:00 – 10:00 | $50$ | $0.378$ | accepted | $0.844$ | accepted 5 | 10:00 – 12:00 | $107$ | $0.734$ | accepted | $0.938$ | accepted | 12:00 – 13:00 | $51$ | $0.594$ | accepted | $0.765$ | accepted | 13:00 – 14:00 | $47$ | $0.651$ | accepted | $0.689$ | accepted | 14:00 – 15:00 | $44$ | $0.817$ | accepted | $0.412$ | accepted | 15:00 – 24:00 | $363$ | $0.214$ | accepted | $0.568$ | accepted | 00:00 – 02:00 | $55$ | $0.249$ | accepted | $0.607$ | accepted | 02:00 – 04:00 | $27$ | $0.309$ | accepted | $0.501$ | accepted | 04:00 – 05:00 | $16$ | $0.441$ | accepted | $0.471$ | accepted | 05:00 – 06:00 | $22$ | $0.707$ | accepted | $0.335$ | accepted | 06:00 – 07:00 | $23$ | $0.580$ | accepted | $0.608$ | accepted | 07:00 – 08:00 | $48$ | $0.500$ | accepted | $0.484$ | accepted 9 | 08:00 – 09:00 | $54$ | $0.338$ | accepted | $0.573$ | accepted | 09:00 – 16:00 | $643$ | $0.060$ | accepted | $0.717$ | accepted | 16:00 – 17:00 | $82$ | $0.429$ | accepted | $0.923$ | accepted | 17:00 – 18:00 | $78$ | $0.804$ | accepted | $0.596$ | accepted | 18:00 – 22:00 | $285$ | $0.919$ | accepted | $0.989$ | accepted | 22:00 – 23:00 | $60$ | $0.989$ | accepted | $0.681$ | accepted | 23:00 – 24:00 | $48$ | $0.522$ | accepted | $0.689$ | accepted | 00:00 – 02:00 | $121$ | $0.094$ | accepted | $0.535$ | accepted | 02:00 – 05:00 | $99$ | $0.098$ | accepted | $0.067$ | accepted | 05:00 – 07:00 | $73$ | $0.650$ | accepted | $0.203$ | accepted | 07:00 – 08:00 | $70$ | $0.788$ | accepted | $0.075$ | accepted | 08:00 – 09:00 | $121$ | $0.786$ | accepted | $0.577$ | accepted | 09:00 – 10:00 | $174$ | $0.421$ | accepted | $0.729$ | accepted 17 | 10:00 – 14:00 | $729$ | $0.089$ | accepted | $0.995$ | accepted | 14:00 – 16:00 | $329$ | $0.982$ | accepted | $0.410$ | accepted | 16:00 – 17:00 | $153$ | $0.361$ | accepted | $0.996$ | accepted | 17:00 – 18:00 | $149$ | $0.596$ | accepted | $0.528$ | accepted | 18:00 – 22:00 | $568$ | $0.586$ | accepted | $0.926$ | accepted | 22:00 – 24:00 | $223$ | $0.071$ | accepted | $0.793$ | accepted | 00:00 – 02:00 | $165$ | $0.198$ | accepted | $0.743$ | accepted | 02:00 – 06:00 | $168$ | $0.117$ | accepted | $0.122$ | accepted | 06:00 – 07:00 | $48$ | $0.833$ | accepted | $0.590$ | accepted | 07:00 – 08:00 | $83$ | $0.805$ | accepted | $0.062$ | accepted | 08:00 – 09:00 | $165$ | $0.576$ | accepted | $0.108$ | accepted | 09:00 – 10:00 | $219$ | $0.105$ | accepted | $0.737$ | accepted 22 | 10:00 – 14:00 | $958$ | $0.097$ | accepted | $0.994$ | accepted | 14:00 – 16:00 | $420$ | $0.952$ | accepted | $0.561$ | accepted | 16:00 – 17:00 | $204$ | $0.491$ | accepted | $0.999$ | accepted | 17:00 – 18:00 | $192$ | $0.534$ | accepted | $0.683$ | accepted | 18:00 – 22:00 | $772$ | $0.436$ | accepted | $0.968$ | accepted | 22:00 – 23:00 | $167$ | $0.688$ | accepted | $0.412$ | accepted | 23:00 – 24:00 | $118$ | $0.963$ | accepted | $0.209$ | accepted | 00:00 – 01:00 | $112$ | $0.171$ | accepted | $0.679$ | accepted | 01:00 – 02:00 | $75$ | $0.933$ | accepted | $0.378$ | accepted | 02:00 – 06:00 | $208$ | $0.072$ | accepted | $0.080$ | accepted | 06:00 – 07:00 | $56$ | $0.935$ | accepted | $0.739$ | accepted | 07:00 – 08:00 | $100$ | $0.882$ | accepted | $0.128$ | accepted | 08:00 – 09:00 | $198$ | $0.566$ | accepted | $0.142$ | accepted | 09:00 – 10:00 | $259$ | $0.341$ | accepted | $0.844$ | accepted | 10:00 – 11:00 | $289$ | $0.091$ | accepted | $0.942$ | accepted 26 | 11:00 – 12:00 | $320$ | $0.725$ | accepted | $0.984$ | accepted | 12:00 – 13:00 | $274$ | $0.915$ | accepted | $0.996$ | accepted | 13:00 – 15:00 | $500$ | $0.439$ | accepted | $0.971$ | accepted | 15:00 – 16:00 | $242$ | $0.574$ | accepted | $0.892$ | accepted | 16:00 – 18:00 | $467$ | $0.895$ | accepted | $0.939$ | accepted | 18:00 – 21:00 | $632$ | $0.643$ | accepted | $0.950$ | accepted | 21:00 – 22:00 | $237$ | $0.803$ | accepted | $0.905$ | accepted | 22:00 – 24:00 | $345$ | $0.034$ | rejected | $0.440$ | accepted ## Conflict of interest The authors declare that they have no conflict of interest. ## References * Ahalt et al. (2018) Ahalt V, Argon N, Strickler J, Mehrotra A (2018) Comparison of Emergency Department crowding scores: a discrete-event simulation approach. Health Care Management Science 21:144–155 * Ahmed and Alkhamis (2009) Ahmed MA, Alkhamis TM (2009) Simulation optimization for an Emergency Department healthcare unit in Kuwait. European Journal of Operational Research 198(3):936 – 942 * Audet and Hare (2017) Audet C, Hare W (2017) Derivative-Free and Blackbox Optimization. Springer Series in Operations Research and Financial Engineering, Springer * Bernstein et al. (2003) Bernstein S, Verghese V, Leung W, T Lunney A, Perez I (2003) Development and validation of a new index to measure Emergency Department crowding. Academic emergency medicine: official journal of the Society for Academic Emergency Medicine 10:938–42 * Brown et al. (2005) Brown L, Gans N, Mandelbaum A, Sakov A, Shen H, Zeltyn S, Zhao L (2005) Statistical analysis of a telephone call center: A queueing-science perspective. Journal of the American Statistical Association 100:36–50 * Conn et al. (2009) Conn A, Scheinberg K, Vicente L (2009) Derivative-Free Optimization. SIAM * Daldoul et al. (2018) Daldoul D, Nouaouri I, Bouchriha H, Allaoui H (2018) A stochastic model to minimize patient waiting time in an Emergency Department. Operations Research for Health Care 18:16 – 25, EURO 2016–New Advances in Health Care Applications * De Santis et al. (2020) De Santis A, Giovannelli T, Lucidi S, Messedaglia M, Roma M (2020) An optimal non–uniform piecewise constant approximation for the patient arrival rate for a more efficient representation of the Emergency Departments arrival process. Technical Report 1–2020, Dipartimento di Ingegneria Informatica Automatica e Gestionale “A. Ruberti”, SAPIENZA Università di Roma * Durbin (1961) Durbin J (1961) Some methods for constructing exact tests. Biometrika 48:41–55 * Guo et al. (2016) Guo H, Goldsman D, Tsui KL, Zhou Y, Wong SY (2016) Using simulation and optimisation to characterise durations of Emergency Department service times with incomplete data. International Journal of Production Research 54(21):6494–6511 * Guo et al. (2017) Guo H, Gao S, Tsui K, Niu T (2017) Simulation optimization for medical staff configuration at Emergency Department in Hong Kong. IEEE Transactions on Automation Science and Engineering 14(4):1655–1665 * Hoot and Aronsky (2008) Hoot N, Aronsky D (2008) Systematic review of Emergency Department crowding: causes, effects, and solutions. Annals of Emergency Medicine 52(2):126–136 * Hoot et al. (2007) Hoot NR, Zhou CH, Jones I, Aronsky D (2007) Measuring and forecasting Emergency Department crowding in real time. Annals of emergency medicine 49 6:747–55 * J Reeder et al. (2003) J Reeder T, Burleson D, G Garrison H (2003) The overcrowded Emergency Department: A comparison of staff perceptions. Academic emergency medicine: official journal of the Society for Academic Emergency Medicine 10:1059–64 * Kathirgamatamby (1953) Kathirgamatamby N (1953) Note on the Poisson index of dispersion. Biometrika 40:225–228 * Kim and Whitt (2014a) Kim SH, Whitt W (2014a) Are call center and hospital arrivals well modeled by nonhomogeneous Poisson process ? Manufactory & Service Operations Management 16:464–480 * Kim and Whitt (2014b) Kim SH, Whitt W (2014b) Choosing arrival process models for service systems: Tests of a nonhomogeneous Poisson process. Naval Research Logistics 61:66–90 * Kim and Whitt (2015) Kim SH, Whitt W (2015) The power of alternative Kolmogorov–Smirnov tests based on transformations of the data. ACM Transactions on Modeling and Computer Simulation 25(4):1–22 * Kuo et al. (2016) Kuo YH, Rado O, Lupia B, Leung JMY, Graham CA (2016) Improving the efficiency of a hospital Emergency Department: a simulation study with indirectly imputed service-time distributions. Flexible Services and Manufacturing Journal 28(1):120–147 * Larson et al. (2019) Larson J, Menickelly M, Wild S (2019) Derivative-free optimization methods. Acta Numerica 28:287–404 * Lewis (1965) Lewis P (1965) Some results on tests for poisson processes. Biometrika 52:67–77 * Liuzzi et al. (2020) Liuzzi G, Lucidi S, Rinaldi F (2020) An algorithmic framework based on primitive directions and nonmonotone line searches for black-box problems with integer variables, Mathematical Programming Computation * Salmon et al. (2018) Salmon A, Rachuba S, Briscoe S, Pitt M (2018) A structured literature review of simulation modelling applied to Emergency Departments: Current patterns and emerging trends. Operations Research for Health Care 19:1–13 * Vanbrabant et al. (2020) Vanbrabant L, Braekers K, Ramaekers K (2020) Improving Emergency Department performance by revising the patient–physician assignment process. Flexible Services and Manufacturing Journal * Wang et al. (2015) Wang H, Robinson RD, Garrett JS, Bunch K, Huggins CA, Watson K, Daniels J, Banks B, D’Etienne JP, Zenarosa NR (2015) Use of the SONET score to evaluate high volume Emergency Department overcrowding: A prospective derivation and validation study. Emergency Medicine International 11:1–11 * Weiss et al. (2004) Weiss S, Derlet R, Arndahl J, Ernst A, Richards J, Fernández-Frankelton M, Schwab R, Stair T, Vicellio P, Levy D, Brautigan M, Johnson A, Nick T (2004) Estimating the degree of Emergency Department overcrowding in academic medical centers: Results of the national ed overcrowding study (NEDOCS). Academic Emergency Medicine 11(1):38–50 * Weiss et al. (2006) Weiss S, Ernst AA, Nick TG (2006) Comparison of the national Emergency Department overcrowding scale and the Emergency Department work index for quantifying Emergency Department crowding. Academic emergency medicine: official journal of the Society for Academic Emergency Medicine 13 5:513–8 * Wiler et al. (2011) Wiler J, Griffey R, Olsen T (2011) Review of modeling approaches for Emergency Department patient flow and crowding research. Academic Emergency Medicine 18:1371–1379 * Zeinali et al. (2015) Zeinali F, Mahootchi M, Sepehri M (2015) Resource planning in the Emergency Departments: a simulation-base metamodeling approach. Simulation Modelling Practice and Theory 53:123–138
Local saddles of RAAR algorithms] Local saddles of relaxed averaged alternating reflections algorithms on phase retrieval Pengwen Chen Applied mathematics, National Chung Hsing University, Taiwan Phase retrieval can be expressed as a non-convex constrained optimization problem to identify one phase minimizer one a torus. Many iterative transform techniques have been proposed to identify the minimizer, e.g., relaxed averaged alternating reflections(RAAR) algorithms. In this paper, we present one optimization viewpoint on the RAAR algorithm. RAAR algorithm is one alternating direction method of multipliers(ADMM) with one penalty parameter. Pairing with multipliers (dual vectors), phase vectors on the primal space are lifted to higher dimensional vectors, RAAR algorithm is one continuation algorithm, which searches for local saddles in the primal-dual space. The dual iteration approximates one gradient ascent flow, which drives the corresponding local minimizers in a positive-definite Hessian region. Altering penalty parameters, the RAAR avoids the stagnation of these corresponding local minimizers in the primal space and thus screens out many stationary points corresponding to non-local minimizers. Keywords: Phase retrieval, relaxed averaged alternating reflections, alternating direction method of multipliers, Nash equilibrium, local saddles § INTRODUCTION Phase retrieval has recently attracted attentions in the mathematics community (see one review [1] and references therein). The problem of phase retrieval is motivated by the inability of photo detectors to directly measure the phase of an electromagnetic wave at frequencies of THz (terahertz) and higher. The problem of phase retrieval aims to reconstruct an unknown object $x_0\in \IC^n$ from its magnitude measurement data $b=|A^* x_0|$, where $A\in \IC^{n\times N} $ represents some isometric matrix and $A^*$ represents the Hermitian adjoint of $A$. Introduce one non-convex $N$-dimensional torus associated with its normalized torus $ \cZ:=\left\{z\in \IC^N: |z|=b\right\},\; \cU:=\left\{u\in \IC^N: |u|=1\right\}. $ The whole problem is equivalent to reconstructing the missing phase information $u$ and the unknown object $x=x_0$ via solving the constrained least squares problem min_x∈^n, |u|=1 { b⊙u-A^* x^2: u∈^N}= min_z∈ A_z^2, $A_\bot\in \IC^{(N-n)\times N}$ is an isometric matrix with unitary matrix $[A^*, A_\bot^*]$, \[ [A^*, A_\bot^*][A^*, A_\bot^*]^*=A_\bot^* A_\bot+A^* A=I.\] Here, $b\odot u$ represents the component-wise multiplication between two vectors $b,u$, respectively. The isometric condition is not very restrictive in applications, since Fourier transforms are commonly applied in phase retrieval. Even for non-Fourier transforms, we can still obtain equivalent problems via a QR-factorization, see [2]. Let $\cU_*$ denote the set in $\cU$ consisting of all the local minimizers of (<ref>). A vector $z_*\in \cZ$ minimizes (<ref>) is called a global solution. In the noiseless measurement case, $A_\bot z_*=0$ or $z_*=A^* x_*=b\odot u_*$ for some $u_*\in \cU$ and some $x_*\in \IC^n$. Numerically, it is a nontrivial task to obtain a global minimizer on the non-convex torus. The error reduction method is one traditional method [3], which could produce a local solution of poor quality for (<ref>), if no proper initialization is taken. During last decades, researchers propose various spectral initialization algorithms to overcome this challenge[4, 5, 6, 7, 2, 8, 9, 10, 11, 12]. On the other hand, phase retrieval can be also tackled by another class of algorithms, including the well-known hybrid input-output algorithm(HIO)[13, 14], the hybrid projection–reflection method[15], Fourier Douglas-Rachford algorithm (FDR)[16], alternating direction methods[17] and relaxed averaged alternating reflections(RAAR) algorithms[18]. An important feature of these algorithms is the empirical ability to avoid local minima and converge to a global mini-mum for noise-free oversampled diffraction patterns. For instance, the empirical study of FDR indicates the disappearance of the stagnation at poor local solutions under sufficiently many random masks. A limit point of FDR is a global solution in (<ref>) and the limit point with appropriate spectral gap conditions reconstructs the phase retrieval solution [16]. Traditional convergence study on Douglas-Rachford splitting algorithm [19, 20] heavily relies on the convexity assumption. Noise-free measurement is a strict requirement for HIO and FDR, which motivates the proposal of relaxed averaged alternating reflections algorithm [18, 21]. Let $\cA,\cB$ denote the sets $Range(A^*)$ and $\cZ$, respectively. Let $P_\cA$ and $P_\cB$ denote the projector on $\cA$ and $\cB$, respectively. Let $R_\cA, R_\cB$ denote the reflectors corresponding to $\cA,\cB$. With one parameter $\beta\in (0,1)$ relaxing the original feasibility problem (the intersection of $\cA$ and $\cB$), the $\beta$-RAAR algorithm [18] is defined as the iterations $\left\{S^k(w): k=1,2,\ldots \right\}$ for some initialization $w\in \IC^N$, \begin{eqnarray} S(w)&=& \beta \cdot \frac{1}{2} (R_\cA R_\cB+I) w+(1-\beta) P_\cB w\\ &=& \frac{\beta}{2}\{(2A^*A-I) (2b\odot \frac{w}{|w|}-w)+w\}+(1-\beta )b\odot \frac{w}{|w|} \\ &=&{\beta} w+(1-2{\beta}) b\odot \frac{w}{|w|}+{\beta} A^* A (2b\odot \frac{w}{|w|}-w).\label{RAART} \end{eqnarray} Fourier Douglas-Rachford algorithm can be deemed as an extreme case of $\beta$-RAAR family with $\beta=1$. As RAAR converges to a fixed point $w$, we could retrieve the phase information $u=w/|w|$ for (<ref>). Any $u$ in $\cU_*$ yields a fixed point $w$. Empirically, RAAR fixed points can produce local solutions of high quality, if a large value is properly chosen for $\beta$, as reported in [17, 21]. In this work, we disclose the relation between RAAR and the local minimizers $z$ in (<ref>). As HIO can be reformulated as one alternating direction method of multipliers in [17], we identify RAAR as one ADMM with penalty parameter $1/\beta'=(1-\beta)/\beta$ applied to the constrained optimization problem in (<ref>), e.g., Theorem. <ref>. This perspective links $\beta$-RAAR with a small parameter $\beta$ to multiplier methods with large penalty ${\beta'}^{-1}$. It is known in optimization that convergence of a multiplier method relies on a sufficiently large penalty (e.g., see Prop. 2.7 in [22]). From this perspective, it is not surprising that the convergence of RAAR to its fixed point also requires a large penalty parameter. large penalty has been employed to ensure various ADMM iterations converging to stationary points [23, 24, 25]. For instance, ADMM [25] is applied to solve the minimization of nonconvex nonsmooth functions. Global convergence to a stationary point can be established, when sufficiently large penalty parameters are used. Saddle plays a fundamental role in the theory and the application of convex optimization[26], in particular, the convergence of ADMM, e.g., [27]. For the application on phase retrieval, Sun et al[28] conduct saddle analysis on a quatradic objective function of Gaussian measurements. The geometric analysis shows that with high probability the global solution is the one local minimizer, when $N/n$ is sufficiently large. Most of critical points are saddles at actually. We believe that saddle analysis is also one key ingredient in explaining the avoidance of undesired critical points for the Lagrangian of RAAR. To some extent, promising empirical performance of non-convex ADMM conveys the impression that saddles exist in the Lagrangian function, which is not evident in the context of phase retrieval. This is a motivation of the current study. Recently, researchers have been cognizant of the importance of saddle structure in non-convex optimization research. Analysis of critical points in non-concave-convex problems leads to many interesting results in various applications. For instance, Lee et al. used a dynamical system approach to show that many gradient-descent algorithms almost surely converge to local minimizers with random initialization, even though they can get stuck at critical points theoretically [29, 30, 31]. The terminology “local saddle" is a crucial concept in understanding the min-max algorithm employed in modern machine learning research, e.g., gradient descent-ascent algorithms in generative adversarial networks (GANs)[32] and multi-agent reinforcement learning[33]. With proper Hessian adjustment, [34] and [35] proposed novel saddle algorithms to escape undesired critical points and to reach local saddles of min-max problems almost surely with random initialization. Jin et al. [36] proposed one non-symmetric definition of local saddles to address one basic question, “ what is a proper definition of local optima for the local saddle?" Later, Dai and Zhang gave saddle analysis on the constraint minimization problems [37]. Our study starts with one characterization of all the fixed point of RAAR algorithms in Theorem <ref>. These fixed points are critical points of (<ref>). By varying $\beta$, some of the fixed points become “local saddles" of a concave-nonconvex function $F$, max_λmin_z {F(z,λ; β):=(β/2 A_(z-λ)^2-1/2 λ^2 ), z∈, λ∈^N}. To characterize RAAR iterates, we investigate saddles in (<ref>) lying in a high dimensional primal-dual space. Our study aims to answer a few intuitive questions, whether these local dimensional critical points on $\cZ$ in the primal space can be lift to local saddles of (<ref>) in a primal-dual space under some spectral gap condition, and how the ADMM iterates avoid or converge to these local saddles under a proper penalty parameter? The line of thought motivates the current study on local saddles of (<ref>). Unfortunately, the definition of local saddles in [36] can not be employed to analyze the RAAR convergence, since the objective function in phase retrieval shares phase invariance, see Remark <ref>. The main goal of the present work is to establish an optimization view to illustrate the convergence of RAAR, and show by analysis and numerics, under the framework for phase retrieval with coded diffraction patterns in  [38], RAAR has a basin of attraction at a local saddle $(z_*,\lambda_*)$. For noiseless measurement, $z_*=A^* x_0$ is a strictly local minimizer of (<ref>). In practice, numerical stagnation of RAAR on noiseless measurements disappears under sufficient large $\beta$ values. Specifically, Theorem <ref> shows that RAAR is actually one ADMM to solve the constrained problem in (<ref>). Based on this identification, Theorem <ref> show that each limit of RAAR iterates can be viewed as a “local saddle" of $\max\min F$ in (<ref>). The rest of the paper is organized as follows. In section <ref>, we examine the fixed point condition of RAAR algorithm. By identifying RAAR as ADMM, we disclose the concave-non-convex function for the dynamics of RAAR, which provides a continuation viewpoint on RAAR iteration. In section <ref>, we present one proper definition for local saddles and show the existence of local saddles for oversampled coded diffraction patterns. In section <ref>, we show the convergence of RAAR to a local saddle under a sufficiently large parameter. Last, we provide experiments to illustrate the behavior of RAAR, (i)comparison experiments between RAAR and Douglas Rachford splitting proposed in [39]; (ii) applications of RAAR on coded diffraction patterns. § RAAR ALGORITHMS §.§ Critical points The following gives the first order optimality of the problem in (<ref>). This is a special case of Prop. <ref> with $\lambda=0$. We skip the proof. Let $z_0=b\odot u_0$ be a local minimizer of the problem in (<ref>). Let $ K_{z_0}^\bot:=\Re(\diag(\bar u_0)A_\bot^* A_\bot \diag(u_0))$. Then the first-order optimality condition is q_0:=z_0^-1⊙(A_^* A_z_0)∈^N, and the second-order necessary condition is that for all $\xi\in \IR^N$, Once a local solution $z$ is obtained, the unknown object of phase retrieval in (<ref>) can be estimated by $x=Az$. On the other hand, using $I=A^*A+A_\bot^* A_\bot$, we can express the first order condition as u^-1⊙(A^*A(b⊙u))=b⊙(1-q_0)∈^N.Using $\xi=e_i$ canonical vectors of $\IR^N$, we have a componentwise lower bound on $1-q_0$ from (<ref>): $b^{-1}\odot (1-q_0)\ge \|A e_i\|^2\ge 0$. In general, there exists many local minimizers on $\cZ$, satisfying (<ref>) and (<ref>). §.§ Fixed point conditions of RAAR We begin with fixed point conditions of $\beta$-RAAR iterations in (<ref>). For each $\beta\in (0,1)$, introduce one auxiliary parameter $\beta'\in (0,\infty)$ defined by β=β' /1+β', i.e., β'=β/1-β. We shall show the reduction in the cardinality of fixed points under with a small penalty parameter. Consider the application of the $\beta$-RAAR algorithm on the constrained problem in (<ref>). Write $w\in \IC^N$ in polar form $w=u\odot |w|$. For each $\beta\in (0,1)$, let ${\beta'}={\beta}/(1-{\beta})$ and c:= (1-1-β/β) b+1-β/β |w|∈^N. Then $w$ is a fixed point of $\beta$-RAAR, if and only if $w$ satisfies the phase condition and the magnitude condition, |w|=β' c+ (1-β')b≥0, i.e., c≥(1-β'^-1) b. In particular, for ${\beta} \in [1/2, 1)$, we have $c\ge 0$ from (<ref>). Observe that the inequality in (<ref>) ensures the well-defined magnitude vector $|w|$. Hence, the fixed points are critical points of (<ref>). Rearranging (<ref>), we obtain the fixed point condition of RAAR, ((1-β)|w|-(1-2β)b)⊙w/|w|=β A^*A { (2b-|w|)⊙w/|w|}. Equivalently, taking the projections $A^*A$ and $I-A^*A$ on (<ref>) yield \begin{eqnarray}&& A^* A \left\{ (b-|w|) \odot \frac{w}{|w|}\right\}=0, \label{f1}\\ && (I-A^*A)\left\{ b\odot \frac{w}{|w|}\right\}= {\beta'}^{-1}\left\{(b-|w |)\odot \frac{w}{|w|}\right\}.\label{f2} \end{eqnarray} For the only-if part, let $w$ be a fixed point of RAAR with (<ref>,<ref>). With the definition of $c$, (<ref>) gives A^*A ((c-b)⊙u)=β'A^*A((|w|-b)⊙u)=0, and (<ref>) gives A_^* A_(c⊙w/|w|)=A_^* A_({ b-β'^-1 (b-|w|)}⊙w/|w|)=0,which implies $ c\odot u$ in the range of $A^*$. Together with (<ref>), we have (<ref>). Also, (<ref>) is the result of the non-negativeness of $|w|$ in (<ref>). To verify the if-part, we need to show that $w$ constructed from a phase vector $u\in \IC^N$ satisfying (<ref>) and a magnitude vector $|w|$ satisfying meets (<ref>,<ref>). From (<ref>) and (<ref>), we have (<ref>), i.e., A^*A((b-|w|)⊙u)=A^*A{ β'(b-c)⊙u}=β'{c⊙u-c⊙u}=0. With the aid of (<ref>, <ref>), the fixed point condition in (<ref>) is ensured by the computation: (I-A^*A){(b-β'^-1(b-|w|))⊙u} the condition in (<ref>) is identical to the first optimality condition in (<ref>). Hence, the fixed points must be critical points of (<ref>). Theorem <ref> indicates that each fixed point $w$ can be re-parameterized by $(u,\beta)$ satisfying (<ref>) and (<ref>). The condition in (<ref>) always hold for ${\beta'}$ sufficiently small. The first order optimality in (<ref>) yields that the phase condition in (<ref>) is actually the critical point condition of $u\in \cU_*$ in (<ref>). Fix one critical point $u\in \cU_*$ and let $c$ be the corresponding vector given from (<ref>). From Theorem <ref>, $w$ given from the polar form $w=u\odot |w|$ with (<ref>) is a fixed point of $\beta$-RAAR, if $\beta$ satisfies the condition in (<ref>). To further examine (<ref>), we parameterize the fixed point $w$ by $(u,\beta)$. Let $b^{-1}\odot K^\bot b$ denote the threshold vector, where \[ K:=\Re(\diag(\bar u)A^*A \diag(u)), \; K^\bot:=I-K. \] The fixed point condition in (<ref>) indicates that $(u,\beta_1)$ gives a fixed point of $\beta_1$-RAAR with any ${\beta}_1\in (0,{\beta})$. That is, the corresponding parameter $(\beta')^{-1}$ must exceed the threshold vector, Since $\beta'= \beta/(1-\beta)$ can be viewed as one penalty parameter in the associated Lagrangian in (<ref>), we call (<ref>) the penalty-threshold condition of RAAR fixed points. In general, the cardinality of RAAR fixed points decreases under a large parameter $\beta$. See Fig. <ref>. For ${\beta}=1$, RAAR reduces to FDR, whose fixed point $w$ satisfies $\|A_\bot (b\odot w/|w|)\|=0$ and thus $\|A(b\odot w/|w|)\|=\|b\|$. When phase retrieval has uniqueness property, $A(b\odot w/|w|)$ gives the reconstruction. On the other hand, for $\beta=1/2$, (<ref>) gives S(w)=A^*A (b⊙w/|w|)+1/2(I-A^*A) w. Suppose a RAAR initialization is chosen from the range of $A^* $. The second term in (<ref>) always vanishes and thus RAAR iterations reduce to alternating projection iterations(AP) in [2]. From this perspective, one can regard $\beta$-RAAR as one family of algorithms interpolating AP and FDR, varying $\beta$ from $1/2$ to $1$. Illustration of the penalty-threshold condition of RAAR fixed points associated with critical points $u_1,u_2, u_3, u_*\in \cU_*$ of (<ref>). The set of fixed points of $\beta$-RAAR is a collection of line segments parameterized by $(u,\beta)\in \cU_*\times (0,1)$. As $\beta$ gets larger, the cardinality of intersections (i.e., the fixed points) decreases. Critical points associated with $\beta=0.5$ are $u_1, u_3$ and $ u_*$. The global minimizer $u_*$ is the only associated critical point, if $\beta\in (0.9,1)$ is used. §.§ Alternative directions method of multipliers Next, we present one relation between RAAR and the alternating direction method of multipliers (ADMM). The alternating direction method of multipliers was originally introduced in the 1970s[40, 41] and can be regarded as an approximation of the augmented Lagrangian method, whose primal update step is replaced by one iteration of the alternating minimization. Although ADMM is classified as one first-order method, practically ADMM could produce a solution with modest accuracy within a reasonable amount of time. Due to the algorithm simplicity, nowadays this approach is popular in many applications, in particular, applications of nonsmooth optimalication. See [27, 42, 43] and the references Use the standard inner product \[ \langle x, y\rangle:=\Re(x^* y),\; \forall x,y\in\IC^N. \] To solve the problem in (<ref>), introduce one auxiliary variable $y\in \IC^N$ with one constraint $y=z$ and one associated multiplier $\lambda\in \IC^N$, and form the Lagrangian function with some parameter $\beta'>0$ in (<ref>), β'/2 A_y^2+< λ, y-z>+1/2y-z^2, y∈^N, z∈. Equivalently, we have the augmented Lagrangian function, L(y,z, λ):= 1/2 A_y^2+ β'^-1< λ, y-z>+1/2β'y-z^2, when we $1/\beta'$ as a penalty parameter. To solve $(y,z,\lambda)$, ADMM starts with some initialization $z_1\in \cZ$ and $ \lambda_1\in \IC^N$ and generates the sequence $\left\{(y_k, z_k, \lambda_k): k=1,2,3,\ldots\right\}$ with stepsize $s>0$, according to rules, \begin{eqnarray*} y_{k+1}=arg\min_y L(y,z_k, \lambda_k),\label{y1}\\ z_{k+1}=arg\min_{|z|=b} L(y_{k+1},z, \lambda_k)\label{z1}, \\ \lambda_{k+1}=\lambda_k+s\nabla_\lambda L(y_{k+1}, z_{k+1}, \lambda)\label{eq9} \end{eqnarray*} Introducing one projection operator on $\cZ$, $[w]_{\cZ}:=w/|w|\odot b$ for $w\in \IC^N$. Algebraic computation yields \begin{eqnarray} y_{k+1}=(I+{\beta'} A_\bot ^*A_\bot )^{-1} (z_k-\lambda_k)=(I-{\beta} A_\bot ^* A_\bot )(z_k-\lambda_k), \label{y} \\ \\ \lambda_{k+1}=\lambda_k+s(y_{k+1}-z_{k+1}). \end{eqnarray} From the $y$-update in (<ref>), one reconstruction $x$ for the unknown object in (<ref>) can be computed by x=Ay=A(I-βA_^* A_)(z-λ)=A(z-λ). Theorem <ref> indicates that RAAR is actually an ADMM with proper initialization applied to the problem in (<ref>). In general, the step size $s$ of the dual vector $\lambda$ should be chosen properly to ensure the convergence. The following shows the relation between RAAR and the ADMM with $s=1$. Hence, we shall focus $s=1$ in this paper. Consider one $\beta$-RAAR iteration $\{{ w_0,w_1,}\ldots, \}$ with nonzero initialization ${ w_0}\in \IC^N$. Let $ \lambda_1=A_\bot^* A_\bot w_0, \; z_1=[w_0]_\cZ. one ADMM sequence $\left\{(y_{k+1}, z_{k+1}, \lambda_{k+1}): k={ 1,2,}\ldots\right\}$ with dual step size $s=1$, according to (<ref>, <ref>, <ref>) with initialization \lambda_1, z_1$. Construct a sequence $\left\{w'_k: k=1,2,\ldots\right\}$ from $(y_k, z_k, \lambda_k)$, w'_k:=y_k+1+λ_k=(I-βA_^* A_) z_k+βA_^* A_λ_k. the sequence $\left\{w'_k: k=1,2,\ldots\right\}$ is exactly the $\beta$-RAAR sequence, i.e., $w'_k=w_k$ for all $k\ge 1$. Use induction. For $k=1$, we have \[ w_1'=(I-\beta A_\bot^* A_\bot) z_1+\beta A_\bot^* A_\bot\lambda_1= (I-\beta A_\bot^* A_\bot) [w_0]_\cZ+\beta A_\bot^* A_\bot w_0=w_1. \] Suppose $w_k'=w_k$ for $k\ge 1$. From (<ref>) and (<ref>), we have A_^* A_w'_k=A_^* A_((1-β) z_k+βλ_k), and $ z_{k+1}=[w_k']_\cZ$. From (<ref>) and (<ref>), A_^* A_λ_k+1=A_^* A_{ βλ_k +(1-β) z_k} -A_^* A_z_k+1. Together with (<ref>), (<ref>) and (<ref>), we complete the proof by the calculation, \begin{eqnarray} (I-\beta A_\bot^* A_\bot) z_{k+1}+\beta A_\bot^* A_\bot\lambda_{k+1} \\ &=&(I-\beta A_\bot^* A_\bot ) z_{k+1}+ \beta A_\bot^* A_\bot \left\{ \beta \lambda_k +(1-\beta) z_k\right\} -\beta A_\bot^* A_\bot z_{k+1}\\ &=&(I-2\beta A_\bot^* A_\bot ) [w'_{k}]_\cZ + \beta A^*_\bot A_\bot w'_k\\ &=&(I-2\beta A_\bot^* A_\bot ) [w_{k}]_\cZ + \beta A^*_\bot A_\bot w_k=w_{k+1}.\label{eq12} \end{eqnarray} Theorem  <ref> provides one max-min viewpoint to explore the dynamics of RAAR, which motivates the study in the next section. Indeed, after eliminating $y$ in $L$ in (<ref>) via (<ref>), we end up with a max-min problem of a concave-non-convex function $F$, F(z,λ; β):=(β/2 A_(z-λ)^2-1/2 λ^2 ), z∈, λ∈^N. We can convert RAAR convergence to its fixed points into the convergence to saddles of the function $F$ in (<ref>) by one primal-dual algorithm. For notation simplicity, we shall omit $\beta$ in the function $F(z,\lambda; \beta)$, i.e., write $F(z, \lambda)$, if no confusion occurs in the context. § DEFINITION OF LOCAL SADDLES FOR RAAR When the objective function $F$ of a max-min problem is not concave-convex, saddle points do not exist in general. For some smooth function $F$, a point $(\lambda,z)$ is said to be a local max-min point in [44, 35], if $z=z_*$ is a local minimizer and $\lambda=\lambda_*$ is a local maximizer, i.e., for $(z,\lambda)$ near $(z_*,\lambda_*)$. The standard analysis can give the first-order and second-order characterizations. The existence of a saddle $(z_*, \lambda_*)$ in fact ensures the minimax equality. Indeed, since $ \min_z F(z,\lambda)\le \min_z\max_\lambda F(z,\lambda) $, then $ \max_\lambda \min_z F(z,\lambda)\le \min_z\max_\lambda F(z,\lambda) $. Together with min_z max_λF(z,λ) ≤max_λF(z_*, λ)≤F(z_*,λ_*)≤min_z F(z,λ_*)≤max_λmin_z F(z,λ), we have that $ \min_z \max_\lambda F(z,\lambda) = \max_\lambda \min_z F(z,\lambda)$ for $(z,\lambda)$ holds near $(z_*,\lambda_*)$. In this paper, we shall adopt the idea on “local max-min" proposed in [36] to emphasize the non-symmetric role of $z,\lambda$, i.e., $(\lambda,z)$ is said to be a local max-min point, if for any $(\lambda, z)$ near $(\lambda_*, z_*)$ within a distance $\delta>0$, max_z' { F(z',λ): z'-z_*≤h(δ)} ≤F(z_*,λ_*)≤F(z,λ_*) for some continuous function $h: \IR\to \IR$ with $h(0)=0$. That is, the minimizer $z$ is driven by the dual vector $\lambda$ maximizing the objective function $F$. Since $F$ in (<ref>) is strictly concave in $\lambda$ for $\beta\in (0,1)$, according to Prop. 18, 19 and 20[36], the first-order condition is $\nabla_z F(z_*,\lambda_*)=0$ and $\nabla_\lambda F(z_*,\lambda_*)=0$, and the second-order necessary/sufficient condition can be reduced to $\nabla_{zz} F(z_*,\lambda_*)\succeq 0$ and $\nabla_{zz} F(z_*,\lambda_*)\succ 0$, respectively. In short, thanks to the $\lambda$-concavity in $F$, we end up with the identical characterization for local-max-min points in [44, 35]. For simplicity, we shall also call these local max-min points as local saddles. §.§ Local saddles From Theorem <ref>, the RAAR convergence of $w_k$ to $w_*$ can be regarded as the convergence to a local saddle $(\lambda_k, z_k)$ of $F$ in (<ref>). For each $\lambda_k$, $z_k$ is one approximate of the corresponding minimizer of $F(\lambda_k, z)$. From (<ref>), we have the following optimality of $F$ in $\lambda$ and $z$, respectively. Since $0<\beta<1$, the strict concavity of $F$ in $\lambda$ ensures the uniqueness of $\lambda$ for any vector $z\in \cZ$. We omit the proof of Prop. <ref>. Fix one $z\in \cZ$. The maximizer $\lambda$ of $ F(z,\lambda)$ in (<ref>) satisfies λ=-β' A_^* A_z. Hence, $\lambda_*=-\beta' A_\bot^* A_\bot z_*$ holds for a saddle point $(z_*,\lambda_*)$. Let $q(z, \lambda)\in \IC^N$ be a vector-valued function of $z$ and $ \lambda$, q(z,λ):=z^-1⊙A_^* A_( z- λ). Fix one $\lambda$ in (<ref>), and consider the $z$-minimization min_z∈ {F(z, λ):=β/2A_(z- λ)^2-1/2 λ^2}. When $z\in\cZ$ is a local minimizer of $F(z, \lambda)$, then q(z, λ)∈^N ξ^⊤(K_z ^-((q)))ξ≥0, ∀ξ∈^N. Let $u=z/|z|$ be the phase vector of a local minimizer $z$. Consider the perturbation $z\to z\odot \exp(i\theta)$ in $\cZ$ applied on $F(z, \lambda)$ with $\theta\in \IR^N$, $\|\theta\|$ near $0$. Variation of $F$ can be expressed as one function of tangent vectors $\xi:=b\odot \theta$ , H(\xi; z)=H(b\odot \theta; z):=\|A_\bot^* A_\bot (z \odot \exp(i\theta) - \lambda)\|^2. With $z=b\odot u$, the Taylor expansion \begin{eqnarray*} H(\xi; z )&=&H(0; z)+\{2\left< i (b\odot u)\odot \theta,A_\bot^* A_\bot (b\odot u- \lambda) \right>-\left< (b\odot u) \odot \theta^2,A_\bot^* A_\bot (b\odot u- \lambda) \right> \\ &&+\|A_\bot (b\odot u\odot \theta)\|^2\}+o(\|\theta\|^2) \\ &=&H(0; z)+2\left< i \xi,q\odot b\right> +\left\{-\left< \xi ^2, q \right>+\xi^\top K^\bot \xi\right\}+o(\|\xi\|^2) \end{eqnarray*} implies that the first-order condition for a local minimizer $z$ is q\in \IR^N, $ and the second-order condition is the positive semi-definite condition, < ξ, ( K^-(q))ξ> ≥0. Unfortunately, the above optimality conditions for $F$ in (<ref>) yields nonexistence of local saddles under the definition in (<ref>)! Consider a noisy case, $\|A_\bot z\|>0$ for all $z\in\cZ$. No Nash equilibrium of $F$ in (<ref>) can exist, since (<ref>) and (<ref>) cannot hold simultaneously at any stationary point of $F$. Indeed, suppose that $(z_*,\lambda_*)$ is a saddle point with $\|A_\bot z_*\|>0$. The $\lambda$-optimality in (<ref>) gives $\left< z_*, \lambda_*\right>=-\beta' \|A_\bot z_*\|^2<0$. On the other hand, as $\xi=b$, (<ref>) gives ξ^⊤(K^-((q)))ξ=(z^* A_^* A_λ)≥0.This inconsistency indicates that as $\lambda$ tends to $\lambda_*$, the corresponding local minimizer $z$ does not approach $z_*$ continuously. In the next subsection, we shall give a proper definition for the local Nash equilibrium applied on phase retrievel. Remark <ref> indicates this difficulty always exists in the non convex-concave optimization with phase invariance. Consider the problem \[ \min_{ \lambda}\max_{z\in \cZ} \left\{ -F(z, \lambda)\right\}, \] where $F(z,\lambda)=F(\alpha z,\alpha\lambda)$ holds for any complex unit $\alpha$. Suppose that $(z, \lambda)$ is a local Nash equilibrium, then F(αz, λ)≤F(z,λ)≤F(z,αλ) for any complex unit $\alpha$ near $1$ and $\lambda$ is a local maximizer. Then phase-invariance of $F$ implies F(αz, λ)=F( z,α̅λ)≤F(z,λ). Contradiction to (<ref>) always occurs, if the above inequality in (<ref>) is strict. §.§ Cross sections To alleviate the difficulty in Remark <ref>, we shall restrict the neighbourhood $\cU(z_*)$ by slicing the projected torus $A_\bot \cZ$ into cross sections, such that (<ref>) can hold locally at each critical point. Fix one $z_0\in \cZ$ and introduce the set $ \cZ'(z_0):=\left\{z_0\odot \exp(i\theta): \left< \theta, b^2 \right>=0, \theta\in \IR^N\right\}. the optimization problem min_z∈'(z_0) { A_z^2}. Note that one partition $\cZ=\cup_\alpha \left\{\cZ'(z_0\alpha): |\alpha|=1 \right\}$ indicates that for each $z\in \cZ$, $z\in \cZ'(\alpha z_0)$ holds for some complex unit $\alpha=\exp(i\rho)$, $\rho\in \IR$. let $1_N\in\IR^N$ be a vector whose entries are all $1$. Since $z, z_0$ both lie in $\cZ$, then z⊙z_0^-1=exp( i δ)=exp( i (θ+ρ1_N)) for some $\delta, \theta\in \IR^N$ and $\rho:=\|b\|^{-2}\left< \delta, b^2\right>$ with $0=\left< \theta, b^2\right>$. To proceed, we need a few notations. At each $z\in \cZ$, introduce a matrix $K^\bot_{z}$ and a tangent plane $\left\{iu\odot \xi: \xi\in \Xi,\; u=z/|z| \right\}$ with \[K^\bot_{z}:=\Re\left(\diag\left(\overline{\frac{z}{|z|}}\right) A_\bot^* A_\bot \diag\left(\frac{z}{|z|}\right)\right), \Xi:=\left\{ \xi: \xi\in \IR^N, \left<\xi, b\right>=0\right\}.\] We shall drop the dependence on $z$ to simplify the notation, if no confusion occurs. Since the feasible set in $z$ is different from the setting in Prop. <ref>, we must investigate again the local $z$-optimality in $F(z,\lambda)$. Fix one $\lambda\in \IC^N$. Consider the minimization problem min_z∈'(z_0) {F(z, λ):=β/2A_( z- λ)^2-1/2 λ^2 }. Suppose $z=b\odot u$ is a local minimizer in (<ref>). Then $\Im(q(z, \lambda))=\rho 1_N$, $\rho\in \IR$. The second-order necessary condition is that for all $\xi\in \Xi$, we have \[ \left< \xi, (K_z^\bot -q(z, \lambda))\xi\right>\ge 0. \] In addition, a second-order sufficient condition is that for all nonzero $\xi\in \Xi$, ξ^-2< ξ, (K_z^-q(z, λ))ξ>> 0. A local minimizer $z$ with (<ref>) is called a strictly local minimizer. Consider a perturbation $z\to z\odot \exp(i\theta)$ with $\theta\in \IR^N$. Use arguments similar to the proof of Prop. <ref>. Since the objective function in (<ref>) is continuously differentiable, with $\xi:=b\odot \theta$, we have < ξ, iu⊙A_^* A_(z- λ)>=0 for all $\xi$ with < ξ, b>= < θ, b^2>=0. Then (<ref>) gives for some multiplier $\rho\in\IR$, (u⊙A_^* A_(z- λ))=ρb, i.e., (q)=ρ1_N∈^N. Note that $\xi\in\Xi$ and thus we have the second-order conditions. Let $z_0=b\odot u_0$ be a local minimizer of the problem in (<ref>). Then the first-order condition is q_0:=z_0^-1⊙(A_^* A_z_0)∈^N, and the second-order necessary condition is that for all $\xi\in \Xi:=\{\xi\in \IR^N; \langle b,\xi\rangle=0\}$, Hence, $z_0$ is a strictly local minimizer on $\cZ$, if ξ^-2ξ^⊤(K_z_0^-(q_0))ξ> 0, ξ∈Ξ, ξ>0 and (<ref>) hold. This is the special case of Prop. <ref> with $\lambda=0$. Note that $q_0=q(z,0)$ and we have $q_0\in \IR^N$ from (<ref>) and ρ=⟨b^2, ρ1_N⟩=⟨b^2,( q(z,λ))⟩=(A_z^2)=0. Readers should notice different tangent spaces used in Prop. <ref> and Prop. <ref>. Finally, we show that $(z_*, \lambda_*)$ can be a local saddle under sufficient small $\beta$, if $z_*$ is a strictly local minimier. For the application of Fourier phase retrieval, Theorem <ref> shows the existence of a strictly local minimizer of (<ref>) based on the spectral gap condition of coded diffraction patterns. Let $z_*$ be a strictly local minimizer in (<ref>). Let $\lambda_*=-\beta' A_\bot^* A_\bot z_*$. Then we can find $\beta$ satisfying ξ^-2 < ξ, (K_z_*^-(b^-1⊙K_z_*^b))ξ>> βK^ξ^2≥0 for any $\xi\in\Xi$, such that $( z_*,\lambda_*)$ is a local saddle of max_λmin_z∈'(z_0) {F(z, λ):=β/2A_( z- λ)^2-1/2 λ^2 }. Since $z_*$ is a strictly local minimizer, Cor. <ref> gives $q_0:=z_*^{-1}\odot (A_\bot^* A_\bot z_*)\in \IR^N$ and $\|\xi\|^{-2}\left< \xi, (K^\bot-\diag(q_0)) \xi\right>>0$. With $(1+\beta')=(1-\beta)^{-1}$ and \[ q(z_*, \lambda_*)=z_*^{-1}\odot A_\bot^* A_\bot (z_*- \lambda_*) =(1+\beta') z_*^{-1}\odot (A_\bot^* A_\bot z_*)=(1+\beta') q_0\in \IR^N,\] we have the second-order sufficient condition (<ref>) for $\beta$ satisfying (<ref>), (1-β) < ξ, ( K^-(q(z_*, λ_*)))ξ>=< ξ, ( K^-(q_0))ξ>-βK^ξ^2. Next, we show that with probability one, the global solution $z=A^*x_0$ is a strictly local minimizer of (<ref>) in the following Fourier phase retrieval. The main tool is the following spectral gap theorem in [16]. Let $\Phi$ be the oversampled discrete Fourier transform. Let $x_0$ be a given rank $\ge 2$ object and at least one of $\mu_j$, $j=1,\ldots, l\ge 2$, be continuously and independently distributed on the unit circle. )be isometric with a proper choice of $c_0$ and $B:= A\diag(u_0)$, $u_0=|A^* x_0|^{-1}\odot (A^* x_0)$. Then with probability one, λ_2=max{ (B^* v) : v∈^n, vi x_0, v=1}<1. Let $\Phi$ be the oversampled discrete Fourier transform. Let $x_0$ be a given rank $\ge 2$ object and at least one of $\mu_j$, $j=1,\ldots, l\ge 2$, be continuously and independently distributed on the unit circle. )be isometric with a proper choice of $c_0$. Then with probability one, $z=A^* x_0$ is a strictly local minimizer of (<ref>), and min_ξ { ξ^-2⟨ξ, (K_z^-((q(z, 0))))ξ⟩: ξ∈^N, ⟨|z|,ξ⟩=0 } where $\lambda_2$ is given in (<ref>). Note that (<ref>) implies that $\Im(B^* v)=\lambda_2\xi$ holds for some unit vector $\xi\in \IR^N$. Then $\xi$ is one left singular vector of :=[(B^*), (B^*)]∈^N×2nwith singular value $\lambda_2$, while and the associated right singular vector is $[\Im(v)^\top , \Re(v)^\top]^\top$. Let $c=K|z|$ and $z=Ax_0$. The left and right singular vectors $\cB$ corresponding to singular value $1$ are $c\in \IR^N$ and $[\Im(ix_0)^\top, \Re(ix_0)]^\top $, respectively. Since $\cB^\top \cB=\Re(B^* B)=\Re(\diag(\overline{u} )A^*A\diag(u))=K_{z}$, then $\xi, c$ are eigenvectors of $K_{z}$. From theorem <ref>, with probability one we have 1>λ_2≥max_ξ { ⟨ξ, K_zξ⟩: ξ=1,⟨|z|,ξ⟩=0}. By definition in (<ref>), $q(z, 0)=b^{-1}\odot (K_{z}^\bot b)=1-b^{-1}\odot K_z b=0$. Finally, since $K^\bot_{z}=I-K_{z}$, (<ref>) gives (<ref>) and (<ref>). § RAAR CONVERGENCE ANALYSIS §.§ Convergence of RAAR We shall derive one inequality stated in (<ref>), which ensures the convergence of RAAR iterations $\left\{w_k: k=1,2,\ldots \right\}$ in Prop. <ref>. In Theorem. <ref>, we shall show that the condition in (<ref>) holds near local saddles under a sufficient large penalty $1/\beta'$. From the $\lambda$-iterations of ADMM in (<ref>), we have λ_k+1=λ_k+(y_k+1-z_k+1), and λ_k+ y_k+1=λ_k+1+ z_k+1. The $z$ iteration yields z_{k+1}=[w_{k}]_{\cZ}=[ z_{k+1}+\lambda_{k+1}]_{\cZ} and $z_{k+1}+\lambda_{k+1}$ shares the same phase with $z_{k+1}$. We have a lower bound, With $y_*=z_*$, $\lambda_*=-\beta'A_\bot^* A_\bot z_*$, and $y_{k+1}=(I-{\beta}A_\bot^* A_\bot) (z_k-\lambda_k)$, ${\beta'}={\beta}/(1-{\beta})$, introduce \begin{eqnarray}&&T(z_k,\lambda_k) :={\beta'} \|A_\bot (y_*-y_{k+1})\|^2+\|y_{k+1}-z_k\|^2\nonumber\\ {\beta} (1-{\beta}) \|A_\bot ((z_k-z_*)-(\lambda_k-\lambda_*))\|^2\nonumber\\ +\|A_\bot (-{\beta} (z_k-z_*)-(1-{\beta})(\lambda_k-\lambda_*))\|^2 +\|A(-\lambda_k)\|^2 \nonumber \\ &=&{\beta} \|A_\bot (z_k-z_*)\|^2+(1-{\beta})\|A_\bot (\lambda_k-\lambda_*)\|^2 \end{eqnarray} We shall derive a few inequalities from the optimal condition of $y$ and $z$ in (<ref>), respectively. Let $T$ be defined in (<ref>). Then +2<z_k-z_*, λ_k-λ_*> Let $C$ be the cost function, $C(y)={\beta'} \|A_\bot y\|^2/2$, for $ y\in \IC^N$. We shall prove 1/2 z_k-z_*^2≥< λ_k-λ_*, y_k+1-y_*>+1/2 y_k+1-y_*^2+1/2T(z_k,λ_k). To this end, we make two claims. First, the optimality of $y_{k+1}$ in (<ref>) indicates that for all $y\in \IC^N$, \begin{eqnarray} C(y)-C(y_{k+1})&=&-\frac{1}{2} \|y-z_k\|^2+\left< \lambda_k, (y_{k+1}-y)\right>+\frac{1}{2} \|y-y_{k+1}\|^2\\ \left(\frac{1}{2} \|y_{k+1}-z_k\|^2 +\frac{{\beta'}}{2} \|A_\bot (y-y_{k+1})\|^2\right).\label{Q2a} \end{eqnarray} Second, the optimality of $y_*$ in $C(y)$ indicates C(y)-C(y_*)+< (y-y_*), λ_*>≥0. To verify (<ref>), use the optimality $y_{k+1}=(I+{\beta'} {A_\bot^*} A_\bot )^{-1} (z_k-\lambda_k)$ in (<ref>), which gives $ \nabla_y L(y_{k+1}, z_k, \lambda_k)=0$. The quadratic convexity of $L$ in $y$ gives (<ref>), i.e., \begin{eqnarray*} &&L(y, z_k, \lambda_k)=C(y)+\left< \lambda_k, (y-z_k)\right>+\frac{1}{2} \|y-z_k\|^2\\ C(y_{k+1})+\left< \lambda_k, (y_{k+1}-z_k)\right>+\frac{1}{2} \|y_{k+1}-z_k\|^2+ (\frac{{\beta'}}{2} \|A_\bot (y-y_{k+1})\|^2 +\frac{1}{2} \|y-y_{k+1}\|^2 \end{eqnarray*} For (<ref>), with $\lambda_*=-{\beta'} A_\bot^* A_\bot z_*=-{\beta'} A_\bot^* A_\bot y_*$, Taylor's expansion of $C(y)$ at $y_*$ gives \begin{eqnarray} && C( y)-C(y_*)= {\beta'} \left< A_\bot^* A_\bot y_*, y-y_*\right>+ \frac{{\beta'}}{2} \|A_\bot(y_*-y)\|^2 \ge \left< -\lambda_*, y-y_*\right>. \end{eqnarray} With $y=y_*=z_*$ in (<ref>) and $y=y_{k+1}$ in (<ref>), (<ref>)+(<ref>) gives (<ref>). Next, from (<ref>), we have two identities, \begin{eqnarray}\label{sq1} \end{eqnarray} The difference of the squares of (<ref>) and (<ref>) gives \begin{eqnarray} - \|z_*-z_k\|^2+ \|z_*-y_{k+1}\|^2\nonumber \\ \|w_{k}-w_*\|^2-\|w_{k-1}-w_*\|^2 -2\left<w_{k}-w_{k-1}, \lambda_k-\lambda_* \right>\label{end:2} \\ \|w_{k}-w_*\|^2-\|w_{k-1}-w_*\|^2 +2\left<z_{k}-z_*, \lambda_k-\lambda_*\right>\nonumber\\ &&-2\left<\lambda_k -\lambda_*, y_{k+1}-z_*\right>, \label{end:1} \end{eqnarray} where the last equality is given by the difference of (<ref>) and (<ref>). The proof of (<ref>) is completed by (<ref>) and (<ref>). Note that for each fixed point $w_*:=z_*+\lambda_*$ of RAAR, $\alpha w_*$ is also a fixed point of RAAR with any complex unit $\alpha$. For $z,\lambda\in \IC^N$, let $\alpha $ be the corresponding global phase factor between $w$ and $w_*$, argmin_α{ w-αw_*: |α|=1}, w=z+λ, w_*:=z_*+λ_*.Suppose that there exists some constant $c_0>1$ such that the following inequality < αz_*-z, λ-αλ_*> holds for $(z,\lambda)=(z_k,\lambda_k)$ for $k\ge k_0$. Then any limit point $(z',\lambda')$ of RAAR satisfies \[ A_\bot (z'-\alpha z_*)=0,\; \lambda-\alpha \lambda_*=0 \; \textrm{ for some complex unit $\alpha$}. \] Recall $w_{k-1}=z_k+\lambda_k$ and $w_*=z_*+\lambda_*$. Let $\alpha_k$ be some global factor in $ \alpha_k:=arg\min_{|\alpha|=1}\|w_k-\alpha w_*\|.$ Summing  (<ref>) over $k=k_0,\ldots, k_1$, for some global phase factors $\alpha_{k_0},\ldots, \alpha_{k_1}$, \begin{eqnarray}&& \|w_{k_0-1}-\alpha_{k_0-1}w_* \|^2-\|w_{k_1-1} -\alpha_{k_1-1}w_*\|^2 \\ &=&\|z_{k_0}+\lambda_{k_0}-\alpha_{k_0-1}(z_*+\lambda_*) \|^2-\|z_{k_1}+\lambda_{k_1} -\alpha_{k_1-1} (z_*+\lambda_*)\|^2 \\ &\ge & \sum_{k=k_0}^{k_1-1} \left\{ \|z_{k}+\lambda_{k}-\alpha_{k-1}(z_*+\lambda_*) \|^2-\|z_{k+1}+\lambda_{k+1} -\alpha_{k}(z_*+\lambda_*)\|^2 \right\}\\ & \ge& \sum_{k=k_0}^{k_1-1} \left\{ T(z_k,\lambda_k)-2\left<\alpha_{k-1} z_*-z_{k}, \lambda_{k}-\alpha_{k-1}\lambda_*\right> \right\}\\ &\ge& (1-c_0^{-1}) \sum_{k=k_0}^{k_1-1} T(z_k,\lambda_k). \end{eqnarray} ( 1-c_0^-1) (k_1-k_0) {min_k=k_0,…, k_1-1 T(z_k,λ_k)}≤w_k_0-1-α_k_0-1w_* ^2. Let $k_1\to \infty$. Since the left-hand side is bounded above and $1-c_0^{-1}>0$, then \[ \liminf_{k\to \infty} T(z_k,\lambda_k)=0. \] Let $(z',\lambda')$ be a limiting point and $\alpha$ be the limiting phase factor. For ${\beta}\in (0,1)$, from (<ref>) Aλ'=0, A_(-β (z'-αz_*)-(1-β) (λ'-αλ_*))=0=A_((z'-αz_*)-(λ'-αλ_*)). The second part of (<ref>) gives $A_\bot \lambda'=\alpha A_\bot \lambda_*$. Thus $\lambda'=\alpha \lambda_*$ and $A_\bot z'=\alpha A_\bot z_*$. When (<ref>) holds eventually, then (<ref>) indicates that converges to $0$, i.e., (<ref>) indicates sub-linear convergence of RAAR, $O((k_0-k_1)^{-1})$. This is consistent with sub-linear convergence in FDR numerical experiments in [16]. §.§ Justification of (<ref>) from local saddles Next we shall verify that the convergence condition in (<ref>) holds,i.e., β A_(z-z_*)^2+(1-β)A_(λ-λ_*)^2 > 2 < z_*-z, λ-λ_*>. the positive definite condition (<ref>) holds at $z_*$ and $(z,\lambda)$ is close to $(z_*, \lambda_*)$. For the sake of simplicity, we shall omit the global factors $\alpha$ in front of $z$ and $\lambda$, if no confusion occurs. With ${\beta}\in (0,1)$, $T(z,\lambda)$ can quantize one distance between $(z,\lambda)$ and $(z_*, \lambda_*)$. That is, for $\epsilon>0$, from (<ref>), $T(z,\lambda)<\epsilon$ implies max{ β A_(z-z_*)^2,(1-β)A_(λ-λ_*)^2, Thus, $ \|A_\bot (z-z_*)\|^2\le \epsilon/{\beta},\; \|\lambda-\lambda_*\|^2\le \epsilon/(1-{\beta}). Let $z_*$ be a strictly local minimizer in (<ref>). Then we can find $\beta\in (0,1)$ satisfying (1- β) < ξ, K_z_*^ξ> > 2 < ξ, (b^-1⊙(K_z_*^b))ξ> for any unit vector $\xi\in \Xi$. Consider $\beta$-RAAR with this ${\beta}\in (0,1)$. Let \lambda_*=-{\beta'} A_\bot^* A_\bot z_*$ $w_*=z_*+\lambda_*$. Then there is some constant $\epsilon>0$, such that (<ref>) holds for all $(z,\lambda)$ with $ \|w-w_*\|^2<\epsilon$, where a proper complex unit is applied on $w_*$ according to (<ref>). First, the existence of $\beta$ for (<ref>) is ensured by Prop. <ref>. The RAAR iterations satisfying (<ref>,<ref>, <ref>,<ref>) indicate the decomposition $w=z+\lambda$ with $z\in \cZ$ and $\lambda\odot z^{-1}\in \IR^N$. Let $u=z/|z|$, $u_*=z_*/|z_*|$ and $q_0=b^{-1}\odot K^\bot b$. Then $ \bar u\odot \lambda\in \IR^N, \bar u_*\odot \lambda_*\in \IR^N$ \begin{equation} (z_*)^{-1}\odot \lambda_*=-\beta' b^{-1}\odot K^\bot b=-\beta' q_0. \end{equation} Using continuity arguments on (<ref>), we have with $u'=(z+ z_*)/|z+ z_*|$, $\xi=(-i)\bar u'\odot ( z_*-z)\in \IR^N$, \begin{equation} \label{eq86} \left< \xi, \bar u'\odot (A_\bot^* A_\bot (u'\odot \xi)) \right> \left< \xi, \left(\frac{\lambda_*}{z_*}+\frac{\lambda}{z}\right)\odot \xi\right> \end{equation} for $(\lambda, z)$ sufficiently close to $( \lambda_*, z_*)$. Observe that as $z\to z_*$, we have \left< \xi, b\right>=0. $ Note that $T(z,\lambda)$ has an upper bound $2(1-c_0^{-1})^{-1}\|w_{k_0-1}-w_*\|^2/2$ from (<ref>). According to Remark <ref>, (<ref>) holds, if $\|w_{k_0-1}-w_*\|<\epsilon$ holds for some $\epsilon$ sufficiently small. With (<ref>), algebraic computation gives (<ref>). Indeed, \begin{eqnarray*} &&2c_0 \left< \lambda-\lambda_*, z_*-z\right>\\ 2c_0 \left< \lambda\odot z^{-1},\bar z\odot ( z_*-z)\right>-2c_0 \left<\lambda_*\odot z_*^{-1},\bar z_* \odot (z_*-z) \right>\\ - c_0 \left<\lambda\odot z^{-1} + \lambda_*\odot z_*^{-1}, | (z-z_*)|^2\right> \\ &\le & \label{m2} \beta \left< \xi, \bar u'\odot (A_\bot^* A_\bot (u'\odot \xi))\right>= \beta\| A_\bot(z-z_*)\|^2\le T(z,\lambda). \end{eqnarray*} Thus, $\|w_{k_0}-w_*\|<\epsilon$ gives the closeness condition for the sequential vector $(z,\lambda)$. § NUMERICAL EXPERIMENTS §.§ Gaussian-DRS The $\beta$-RAAR algorithm is not the only algorithm, which can screen out some undesired local saddles via varying penalty parameters. Recently,[39] proposed Gaussian-Douglas-Rachford Splitting to solve phase retrieval via minimizing a loss function $\||z|-b\|^2$ subject to $z$ in the range of $A^*$. Let $x$ be the unknown object. Let $1_\cF(y)$ be the indicator function of the range $\cF$ of $A^*$. Then $A^* x \in \cF$. Similar to RAAR, the algorithm can be formulated as ADMM with a penalty parameter $\rho>0$ to reach a local max-min point of the Lagrangian function max_λmin_y,z∈^N {1/2 |z|-b^2+ < λ, z-y>+ρ/2 z-y^2+1_(y)}. The ADMM scheme consists of repeating the following three updates to reach a fixed point $(y,z,\lambda)$: * $y\leftarrow A^*A (z+\rho^{-1}\lambda)$; * $ z\leftarrow (1+\rho)^{-1}(\frac{w}{|w|}\odot b+\rho w)$ where $w=y-\rho^{-1}\lambda$; * $\lambda \leftarrow \lambda+\rho (z-y)$. Similar to the RAAR reconstruction in (<ref>), once a fixed point of this ADMM is obtained, the object $x$ can be computed by $x=Ay=A(z+\rho^{-1}\lambda)$ from the $y$-update. Introduce $P=A^*A$ and $P^\bot=I-P$. After eliminating $y$, the local max-min problem reduces to max_λmin_z {1/2|z|-b^2+ρ/2 {P^(z+λ/ρ)^2-λ/ρ^2}}. Algebraic computations on ADMM scheme yield the fixed-point condition of DRS. Denote $\mu:=\lambda/\rho$. Let $(z,\mu)$ be a fixed point of $\rho$-DRS. Then z+ρμ=[z-μ]_=[z]_, Pμ=0 and P^z=0. We skip the proof of (<ref>). Note that the condition implies that the vector $z$ shares the same phase vector $u:=z/|z|$ with $z-\mu$, and $[z]_\cZ$ has the $(P,P^\bot)$-decomposition $[z]_\cZ=z+\rho \mu$, where $P\mu=0$ and $P^\bot z=0$. Hence, $\mu/z\in \IR^N$ is a real vectors with bounds, $-\rho^{-1} \le z^{-1}\odot \mu\le 1$. Next we derive conditions for local saddles of $L$ in (<ref>). Let $(z,\mu)$ be a local saddle of $L$ in (<ref>). Then the first-order optimal condition is z+ρμ=[z]_, Pμ=0, P^z=0. When $\rho>0$, the concavity of $L$ in $\lambda$ is obvious. Let $K:=\Re(\diag(\bar u) P \diag(u))$ and $ u=z/|z|.$ The second-order necessary condition of $z$ in (<ref>) is The optimality of $\mu$ in (<ref>) is P^(z+μ)=μ, which implies P^z=0 and Pμ=0. From the derivative of $L$ with respect to $z$, the $z$-optimality is \[ \frac{z}{|z|}\odot (|z|-b)+\rho P^\bot (z+\mu)=0, \textrm{ i.e., } z-[z]_\cZ+\rho P^\bot (z+ \mu)=0. \] Together with (<ref>), we have z+\rho \mu=[z]_\cZ. Next, we derive the second-order necessary condition of $z$. Consider a perturbation $z\to z+\epsilon$ with $\epsilon\in \IC^N$. From |z+\epsilon|=|z| \left(1+\Re(\frac{\epsilon}{z}) +\frac{1}{2}\Im(\frac{\epsilon}{z})^2\right)+o(\epsilon^2), \begin{eqnarray} %&=&2\left< z, \epsilon\right> +\|\epsilon\|^2 -2\left< b, %|z| \left(\Re(\frac{\epsilon}{z}) +\frac{1}{2}\Im(\frac{\epsilon}{z})^2\right)\right> +o(\|\epsilon\|^2), &=&2\left< z-b\odot \frac{z}{|z|}, \epsilon\right> +\|\epsilon\|^2 -\left< \frac{b}{|z|}, \Im(\epsilon\odot \bar u)^2\right>+o(\|\epsilon\|^2),\label{eq155} \end{eqnarray} we have the second-order condition of $L$, ρ< ϵ, P^ϵ> +ϵ^2 -< b/|z|, (ϵ⊙u̅)^2> ≥0. Taking $\epsilon=ic\odot u$ for $c\in\IR^N$ yields c^2≥1/ρ+1 < b/|z|, c^2>+ρ/ρ+1 <c, Kc>. Next, we show that a local saddle $(z,\mu)$ is always a fixed point of DRS. If $(\mu, z)$ is a local max-min point in (<ref>), then $z$ is a fixed point of DRS. At each stationary point $z$ of DRS, we have $|z|=Kb$ and $[z]=b\odot u$ of DRS. Taking $c=e_i$ in (<ref>) yields \[ (\rho+1) Kb=(\rho+1)|z|\ge b, \; i.e., \; \bar u\odot (z-\mu)=Kb-\mu\odot \bar u\ge 0, \] which implies $[z-\mu]_\cZ=[z]_\cZ$. Together with the $(P, P^\bot)$-decomposition, $[z]_\cZ=z+\rho \mu$, we have the fixed point condition. From the second-order condition in (<ref>), we expect that DRS with smaller $\rho$ yields a stronger screening-out ability. The next remark illustrates the screening-out similarity between RAAR and DRS in the case $\rho$ close to $0$. Roughly, for $\rho$ close to $0$, a local saddle at some phase vector $u$ of $\beta$-RAAR would be a local saddle at the same phase vector $u$ of $\rho$-DRS, if $\rho$ and $\beta$ satisfy $\beta^{-1}=\rho+1$. Indeed, the second-order necessary condition for RAAR function in (<ref>) is given by = -β/1-βI-K+1/1-β(Kb/b)≽0. That is, for any nonzero $\xi\in\IR^N$, < Kb/b, ξ^2>≥βξ^2+(1-β) ξ^⊤Kξ. On the other hand, for DRS, replacing $c$ with $\pm (Kb/b)^{1/2}\odot \xi$ and $|z|=Kb$ in (<ref>) gives < Kb/b, ξ^2>≥1/ρ+1ξ^2+ρ/ρ+1 < (Kb/b)^1/2⊙ξ, K ((Kb/b)^1/2⊙ξ)>. Comparing with (<ref>), (<ref>) is almost identical to (<ref>) under $\beta^{-1}=\rho+1$, if $\rho$ is close to $0$ and we ignore the difference of the second terms of their right hand side. §.§.§ Simulation of Gaussian matrices We provide one simulation to present the screening-out effect for $\beta$-RAAR and $\rho$-DRS. Generate Gaussian matrices $A$ with size $n\times N$, $n=100$, $N/n=3, 3.5, 4, 4.5$ and $ 5$, respectively. For simplicity, generate noise-free data $b=|A^* x_0|$ from some $x_0$. Apply $\beta$-RAAR and $\rho$-DRS to reconstruct phase retrieval solutions under a set of parameters $\beta, \rho$, respectively. Here, we test ρ= 1,1/2, 1/3, 1/4, …, 1/10 and β=1/2, 2/3, 3/4, …, 10/11.Figures <ref> show the success rate of reaching a global solution for each parameter value (among $40$ trials with random initializations). For $\rho$ close to $0$ or $\beta$ close to $1$, $\beta$-RAAR with \[ \beta=\frac{\rho^{-1}}{1+\rho^{-1}} \] gives a similar empirical performance as $\rho$-DRS, although RAAR performs slightly better. For instance, as $\beta\ge 0.8$ or $\rho\le 1/4$, with success rates higher than $70\%$, $\beta$-RAAR and $\rho$-DRS algorithms both can reconstruct a global solution in the case with $N/n\ge 4$. These empirical results are consistent with the theoretical analysis in Remark <ref>. Right and left subfigures show the success rate under $\beta$-RAAR algorithm and $\rho$-DRS algorithm with $ \beta=\rho^{-1}(1+\rho^{-1})^{-1}$, respectively. §.§ Coded diffraction patterns The following experiments present convergence behavior of RAAR on coded diffraction patterns. $1\frac{1}{2}$ coded diffraction patterns with oversampling, i.e., one coded pattern and one uncoded pattern as used in [16]. For test images $x_0$, we use the Randomly Phased Phantom(RPP) $x_0=p\odot \mu_0$, where $\mu_0:=e^{i\phi}$ and $\phi$ are i.i.d. uniform random variables over $[0,2\pi]$. The size is $128\times 128$, including the margins. we randomize the original phantom $p$ (in the left of Fig. <ref>) in order to make its reconstruction more challenging. A random object such as RPP is more difficult to recover than a deterministic object. Phantom image $p$ without phase randomization (left); Images reprsents the null vector initialization $\Re(x_{null}\odot \overline{\mu_0})$ in the noiseless case(middle) and in the noisy case(right). Theorem <ref> states the existence and strictly local minimizer and Theorem  <ref> indicates that the existence of a local saddle replies on a sufficiently large penalty parameter. The following experiment validates RAAR convergence to the local saddle under proper selection on $\beta$. Empirically, $\beta$-RAAR with large $\beta$ can easily diverge, but $\beta$-RAAR with small $\beta$ can easily get stuck (not necessarily converged) near distinct critical points on $\cZ$. To demonstrate the effectiveness of RAAR, we shall make two adjustments on application of $\beta$-RAAR. First, to alleviate the stagnation at far critical solutions, we employ the null vector method[8], which is one spectral initialization, to generate an initialization of RAAR. See the middle and right subfigures in Fig. <ref> for the initialization. Second, to reach a local saddle within 600 RAAR iterations, we vary the parameter $\beta$ along a $\beta$-path, starting from some initial value and then decreases to $0.5$, shown in Fig. <ref>. Each path consists of to phases: (i) $\beta$ remains one constant value selecting from $0.95, 0.9, 0.8, 0.7$ and $ 0.6$ within the first $300$ iterations; (ii) $\beta$ decreases to $0.5$ piecewise linearly within the second $300$ iterations. The corresponding $\beta$ value used within $600$ RAAR iterations of five different $\beta$-paths. Conduct four experiments to examine $\beta$-RAAR along five $\beta$-paths: (a) Noiseless data, $b=|A^* x_0|$ with $A$ defined in (<ref>). Use the null vector method $x_{null}$ as one initial vector for $\beta$-RAAR, i.e., z_1=[ A^* x_null ]_, λ_1=A^* x_null-z_1. (b) Noiseless data. RAAR with random initialization. (c) Noisy data. RAAR with null vector initialization as in (<ref>). (d) Noisy data. RAAR with random initialization. In (c) and (d), the source of noise is the counting statistics [45], i.e., each entry of the squared measurement $b^2$ follows a Poisson distribution, b^2∼Poisson(κ|A^* x_0|^2), κ>0. In the RPP experiment, the parameter $\kappa$ is chosen so that the noise level is $\| b-|A^* x_0|\|/\|b\|\approx 0.18$. §.§.§ Performance metrics in the case (a,b) and (c,d) are reported in figure <ref> and figure <ref>, respectively. Each row shows the performance metrics of $\beta$-RAAR iterations $\{w_k\}$. Here, $z,\lambda$ are computed from $w$ according to {z_k+1:=[w_k]_, λ_k+1:=w_k-z_k+1}. From (<ref>), the reconstruction of the object $x$ is estimated by $A(z_{k+1}-\lambda_{k+1})$. * Residual: $\|A_\bot z\|/\|b\|$. * Norm of derivative: The Wirtinger derivative of the objective $F$ in the $\lambda$-direction is -{βA_^*(A_(z-λ)) +λ}. When RAAR converges to one local saddle, the derivative norm would be $0$. The norm \[ \ID_\lambda(z,\lambda):=\left (\|A_\bot ((1-\beta) \lambda+\beta z) \|^2+\|A\lambda\|^2\right )^{1/2} \] can be employed to examine the quality of convergence. (Empirically, the derivative in the $z$-direction $\ID_z$ has a behavior similar to the one of $\ID_\lambda$. For simplicity, we do not report $\ID_z$ here.) * Inequality ratio $\IT(z_k,\lambda_k)$. In the noiseless setting, $\IT(z,\lambda)$ is positive, as $(z,\lambda)$ approaches a local saddle $(z_*,\lambda_*)$ with $\|A_\bot z_*\|=0$. Hence, a positive ratio \[ \IT(z,\lambda):=1+(\beta \|A_\bot z\|^2+(1-\beta) \|A_\bot \lambda\|^2+\|A\lambda\|^2)^{-1}(2\langle z, \lambda\rangle) \] can be used as one indicator that RAAR iterates enter the attraction basin of $(z_*,\lambda_*)$. §.§.§ Results on noiseless measurement In Fig. <ref>, the left column and the right column show the metric performance in the cases (a) and (b). * The left column shows the result of the 600 RAAR iterations along five $\beta$ curves with null initialization, i.e., case (a). The null initialization is illustrated in the middle of Fig. <ref>. Based on the residual and derivative metrics, RAAR converges to the global solutions for all five $\beta$-paths. The $\IT$ become positive after $100$ iterations, which indicates the closeness of the null initialization to the attraction basin. In particular, $\IT$ reaches $1$ in the early iterations of the case $\beta=0.95$. * The right column shows the case (b). The initialization difference between case (a) and case (b) reflects the influence of undesired local saddles. We observe two distinct convergence behaviors. First, in the case of $\beta=0.6, \beta=0.7$ and $\beta=0.8$, based on the metrics of the derivative norm and the residual, RAAR fails to converge within the first $300$ iterations. As $\beta$ decreases in the second 300 iterations, the iterates tend to different local saddles. Second, for the $\beta$ paths starting with $\beta=0.9$ or $\beta=0.95$, RAAR successfully converge to global solutions. Their $\IT$-values are negative in the early 50 iterations, but quickly turn to be positive after 100 iterations. Fig. <ref> demonstrates the reconstruction. §.§.§ Results on noisy measurements The left column in Fig. <ref> demonstrates the metric performance of RAAR in the noisy case (c). The null initialization shown in the right of Fig. <ref>, is used to reduce the chance of getting stuck at far local saddles. The reconstructed objects after the first 300 RAAR iterations are shown in the top row of in Fig. <ref>. Even though these reconstructions are very similar to the RPP, the metric $\ID_\lambda$ indicates that these RAAR with $\beta\ge 0.7$ fail to converge within the first 300 iterations. Hence, we decrease $\beta$ in the second 300 iterations to obtain local saddles. Observe that the derivative norm in all cases decays to $0$. Actually, by examining the correlation of reconstructed objects after 600 RAAR iterations, we verify that these five reconstructed objects are identical up to a phase factor. The right column in Fig. <ref> shows the metric performance of the noisy case (d). Five $\beta$-RAAR tend to different residual values in the first 300 iterations. For large $\beta$, i.e., $\beta=0.95$, $\beta=0.9$ and $\beta=0.8$, RAAR produce rather successful reconstructed objects shown the bottom row of in Fig. <ref>. These RAAR do not converge within the first 300 iterations. Hence, we reduce the $\beta$ value during the second $300$ iterations. By examining the correlation of reconstructed objects, we verify that three final reconstructions are all identical to the final reconstruction in (c). For small $\beta$, RAAR could get stuck at poor solutions, e.g., $\beta=0.7$ and $\beta=0.6$. Indeed, after the second 300 iterations, these two RAAR converge to non-global local solutions with larger residual values. The above experiment results suggest that RAAR starting with large $\beta$ typically performs better than RAAR starting with small $\beta$ in the lack of spectral methods. In numerical simulations, we demonstrate the effectiveness of RAAR on coded diffraction patterns, where $\beta$ travels from a large value to $0.5$. Left and right columns show the residue metric and the derivative metric of RAAR in case(a) and case (b), respectively. Left and right columns show the residue metric and the derivative metric of RAAR in case(c) and case (d), respectively. The row from left to right show the reconstruction of 300 RAAR iterations in the case (b), corresponding to $\beta$-paths with $\beta=0.95$, $0.9$, $0.8$, $0.7$ and $0.6$, respectively. The columns from left to right show the reconstruction of 300 RAAR iterations in the case (c,d), corresponding to $\beta$-paths with $\beta=0.95$, $0.9$, $0.8$, $0.7$ and $0.6$, respectively. The top row is the case (c) and the bottom row is the case (d). §.§ Conclusion and outlook In this paper, we examine the RAAR convergence from a viewpoint of local saddles of a concave-non-convex max-min problem. We show that the global solution is a strictly local minimizer in oversampled coded diffraction patterns, which ensures the existence of local saddles. Convergence to each local saddle of the RAAR Lagrangian function requires a sufficient large penalty parameter, which explains the avoidance of undesired local solutions under RAAR with a moderate penalty parameter. ADMM is a popular algorithm in handling various constraints. The current paper does not introduce any further assumption on unknown objects, except for the condition $x\in \IC^n$ in (<ref>). Stable recovery from incomplete measurements is actually possible, provided that additional assumptions of unknown objects are used. For instance, when unknown objects can be characterized by piecewise-smooth functions with small total variation seminorm, the recovery can be obtained from incomplete Fourier measurements with the aid of total variation regularization [46]. Another interesting work [47] demonstrates the number of measurements ensuring stable recovery of a sparse object under independent measurement vectors. From the above perspective, one interesting future work is the saddle analysis of ADMM associated with these additional object assumptions. §.§ Acknowledgements The author would like to thank Albert Fannjiang for helpful discussions. [1] Y. Shechtman, Y. C. Eldar, O. Cohen, H. N. Chapman, J. Miao, and M. Segev. Phase retrieval with application to optical imaging: a contemporary Signal Processing Magazine, IEEE, 32(3):87–109, 2015. [2] P. Chen, A. Fannjiang, and G Liu. Phase retrieval with one or two coded diffraction patterns by alternating projection with the null initialization. Journal of Fourier Analysis and Applications, pages 1–40, [3] R. W. Gerchberg. A practical algorithm for the determination of phase from image and diffraction plane pictures. Optik, 35:237, 1972. [4] Praneeth Netrapalli, Prateek Jain, and Sujay Sanghavi. Phase retrieval using alternating minimization. arxiv:1306.0160, 2013. [5] Yuxin Chen and Emmanuel J. Candes. Solving random quadratic systems of equations is nearly as easy as solving linear systems. arxiv:1505.05114, 2015. [6] Emmanuel J. Candes, Xiaodong Li, and Mahdi Soltanolkotabi. Phase retrieval via wirtinger flow: Theory and algorithms. IEEE Transactions on Information Theory, 61(4):1985–2007, apr 2015. [7] Huishuai Zhang and Yingbin Liang. Reshaped wirtinger flow for solving quadratic system of equations. In D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 29, pages 2622–2630. Curran Associates, Inc., 2016. [8] P. Chen, A. Fannjiang, and G Liu. Phase retrieval by linear algebra. SIAM J. Matrix Anal. appl., 38(3):864–868, 2017. [9] Wangyu Luo, Wael Alghamdi, and Yue M. Lu. Optimal spectral initialization for signal recovery with applications to phase retrieval. arXiv:1811.04420, 2018. [10] Yue M. Lu and Gen Li. Phase transitions of spectral initialization for high-dimensional nonconvex estimation. arxiv:1702.06435, 2017. [11] Marco Mondelli and Andrea Montanari. Fundamental limits of weak recovery with applications to phase Foundations of Computational Mathematics, 19(3):703–773, Jun [12] John C Duchi and Feng Ruan. Solving (most) of a set of quadratic equalities: Composite optimization for robust phase retrieval. Information and Inference: A Journal of the IMA, 8(3):471–529, [13] J. R. Fienup. Phase retrieval algorithms: a comparison. Applied optics, 21(15):2758–2769, 1982. [14] J. R. Fienup. Phase retrieval algorithms: a personal tour. Applied Optics, 52(1):45–56, 2013. [15] Heinz H. Bauschke, Patrick L. Combettes, and D. Russell Luke. Hybrid projection–reflection method for phase retrieval. J. Opt. Soc. Am. A, 20(6):1025–1034, Jun 2003. [16] Pengwen Chen and Albert Fannjiang. Fourier phase retrieval with a single mask by Douglas–Rachford Applied and Computational Harmonic Analysis, 44(3):665 – 699, [17] Zaiwen Wen, Chao Yang, Xin Liu, and Stefano Marchesini. Alternating direction methods for classical and ptychographic phase Inverse Problems, 28(11):115010, oct 2012. [18] D Russell Luke. Relaxed averaged alternating reflections for diffraction imaging. Inverse Problems, 21(1):37–50, nov 2004. [19] Jonathan Eckstein and Dimitri P. Bertsekas. On the Douglas—Rachford splitting method and the proximal point algorithm for maximal monotone operators. Mathematical Programming, 55(1):293–318, Apr 1992. [20] Bingsheng He and Xiaoming Yuan. On the convergence rate of Douglas-Rachford operator splitting Mathematical Programming, 153(2):715–722, 2015. [21] Ji Li and Tie Zhou. On relaxed averaged alternating reflections (RAAR) algorithm for phase retrieval with structured illumination. Inverse Problems, 33(2):025012, jan 2017. [22] Dimitri P. Bertsekas. Constrained Optimization and Lagrange Multiplier Methods (Optimization and Neural Computation Series). Athena Scientific, 1 edition, 1996. [23] Mingyi Hong, Zhi-Quan Luo, and Meisam Razaviyayn. Convergence analysis of alternating direction method of multipliers for a family of nonconvex problems. SIAM Journal on Optimization, 26(1):337–364, 2016. [24] Guoyin Li and Ting Kei Pong. Global convergence of splitting methods for nonconvex composite SIAM Journal on Optimization, 25(4):2434–2460, 2015. [25] Yu Wang, Wotao Yin, and Jinshan Zeng. Global convergence of ADMM in nonconvex nonsmooth optimization. Journal of Scientific Computing, 78(1):29–63, 2019. [26] D.P. Bertsekas. Constrained Optimization and Lagrange Multiplier Methods. Computer science and applied mathematics. Elsevier Science, 2014. [27] Stephen Boyd, Neal Parikh, Eric Chu, Borja Peleato, and Jonathan Eckstein. Distributed optimization and statistical learning via the alternating direction method of multipliers. Found. Trends Mach. Learn., 3(1):1–122, January 2011. [28] J. Sun, Q. Qu, and J. Wright. A geometric analysis of phase retrieval. In 2016 IEEE International Symposium on Information Theory (ISIT), pages 2379–2383, July 2016. [29] Jason D. Lee, Max Simchowitz, Michael I. Jordan, and Benjamin Recht. Gradient descent only converges to minimizers. In Vitaly Feldman, Alexander Rakhlin, and Ohad Shamir, editors, 29th Annual Conference on Learning Theory, volume 49 of Proceedings of Machine Learning Research, pages 1246–1257, Columbia University, New York, New York, USA, 23–26 Jun 2016. PMLR. [30] Simon S Du, Chi Jin, Jason D Lee, Michael I Jordan, Aarti Singh, and Barnabas Gradient descent can take exponential time to escape saddle points. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 30, pages 1067–1077. Curran Associates, Inc., [31] Chi Jin, Rong Ge, Praneeth Netrapalli, Sham M Kakade, and Michael I Jordan. How to escape saddle points efficiently. In 34th International Conference on Machine Learning, ICML 2017, pages 2727–2752. International Machine Learning Society (IMLS), 2017. [32] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. [33] Shayegan Omidshafiei, Jason Pazis, Christopher Amato, Jonathan P How, and John Deep decentralized multi-task multi-agent reinforcement learning under partial observability. arXiv preprint arXiv:1703.06182, 2017. [34] Leonard Adolphs, Hadi Daneshmand, Aurelien Lucchi, and Thomas Hofmann. Local saddle point optimization: A curvature exploitation approach. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 486–495. PMLR, 2019. [35] Constantinos Daskalakis and Ioannis Panageas. The limit points of (optimistic) gradient descent in min-max Advances in Neural Information Processing Systems, 31:9236–9246, 2018. [36] Chi Jin, Praneeth Netrapalli, and Michael Jordan. What is local optimality in nonconvex-nonconcave minimax In International Conference on Machine Learning, pages 4880–4889. PMLR, 2020. [37] Yu-Hong Dai and Liwei Zhang. Optimality conditions for constrained minimax optimization. CSIAM Transactions on Applied Mathematics, 1(2):296–315, 2020. [38] Albert Fannjiang. Absolute uniqueness of phase retrieval with random illumination. Inverse Problems, 28(7):075008, jun 2012. [39] Albert Fannjiang and Zheqing Zhang. Fixed point analysis of douglas–rachford splitting for ptychography and phase retrieval. SIAM Journal on Imaging Sciences, 13(2):609–650, 2020. [40] R. Glowinski and A. Marroco. Sur l'approximation, par éléments finis d'ordre un, et la résolution, par pénalisation-dualité d'une classe de problèmes de dirichlet non linéaires. ESAIM: Mathematical Modelling and Numerical Analysis - Modélisation Mathématique et Analyse Numérique, 9(R2):41–76, 1975. [41] Daniel Gabay and Bertrand Mercier. A dual algorithm for the solution of nonlinear variational problems via finite element approximation. Computers Mathematics with Applications, 2(1):17–40, 1976. [42] Junfeng Yang, Yin Zhang, and Wotao Yin. An efficient tvl1 algorithm for deblurring multichannel images corrupted by impulsive noise. SIAM Journal on Scientific Computing, 31(4):2842–2865, 2009. [43] Bhabesh Deka and Sumit Datta. Compressed sensing magnetic resonance image reconstruction Springer Series on Bio-and Neurosystems, 2019. [44] A. Cherukuri, B. Gharesifard, and J. Cortés. Saddle-point dynamics: Conditions for asymptotic stability of saddle SIAM Journal on Control and Optimization, 55(1):486–511, 2017. [45] P Thibault and M Guizar-Sicairos. Maximum-likelihood refinement for coherent diffractive imaging. New Journal of Physics, 14(6):063004, jun 2012. [46] Huibin Chang, Yifei Lou, Michael K. Ng, and Tieyong Zeng. Phase retrieval from incomplete magnitude information via total variation regularization. SIAM J Sci Comput, 38(6):A3672–A3695, 2016. [47] Y. C. Eldar and S. Mendelson. Phase retrieval: Stability and recovery guarantees. Applied and Computational Harmonic Analysis, 36(3):473–494,
# Inexact gradient projection method with relative error tolerance A. A. Aguiar Instituto de Matemática e Estatística, Universidade Federal de Goiás, CEP 74001-970 - Goiânia, GO, Brazil, E-mails: <EMAIL_ADDRESS><EMAIL_ADDRESS><EMAIL_ADDRESS>The authors was supported in part by CNPq grants 305158/2014-7 and 302473/2017-3, FAPEG/PRONEM- 201710267000532 and CAPES. O. P. Ferreira 11footnotemark: 1 L. F. Prudente 11footnotemark: 1 ###### Abstract A gradient projection method with feasible inexact projections is proposed in the present paper. The inexact projection is performed using a relative error tolerance. Asymptotic convergence analysis and iteration-complexity bounds of the method employing constant and Armijo step sizes are presented. Numerical results are reported illustrating the potential advantages of considering inexact projections instead of exact ones in some medium scale instances of a least squares problem over the spectrohedron. Keywords: Gradient method, feasible inexact projection, constrained convex optimization. AMS subject classification: 49J52, 49M15, 65H10, 90C30. ## 1 Introduction In this paper, we address general constrained convex optimization problems of the form $\min\\{f(x):~{}x\in C\\},$ (1) where $C$ is a closed and convex subset of $\mathbb{R}^{n}$ and $f:\mathbb{R}^{n}\to{\mathbb{R}}$ is a continuously differentiable function. Denote by $f^{*}:=\inf_{x\in C}f(x)$ the optimal value of (1) and by $\Omega^{*}$ its solution set, which we will assume to be non-empty unless the contrary is explicitly stated. The Problem (1) is a basic optimization issue, it has appeared in several areas of science and engineering, including machine learning, control theory and signal processing, see for example [9, 10, 17, 36, 46]. In the present paper, we are interested in gradient-type algorithms to solve it. The gradient projection method (GPM) is the one of the most oldest method to solve Problem (1), its convergence properties go back to the works of Goldstein [22] and Levitin and Polyak [35]. After these works, many variants of it have appeared throughout the years, resulting in a wide literature on the subject, see, for example, [4, 5, 6, 16, 17, 25, 27, 28, 43, 51]. The GPM has attracted the attention of the scientific community working in optimization, mainly due to its simplicity and easy implementation. Besides, since this method uses only first order derivatives, it is often considered as a scalable solver for large-scale optimization problems, see [31, 38, 39, 41, 46, 48]. At each iteration, the classical GPM moves along the direction of the negative gradient, and then projects the iterate onto $C$ if it is infeasible. Although the feasible set of many important problems has an easy-to-handle structure, in general this set could be so complex that the exact projection can not be easily computed. It is well known that the mostly computational burden of each iteration of the GPM is in the solution of this subproblem. In fact, one drawback of methods that use exact projections is to solve a quadratic problem at each stage, which can lead to a substantial increase in the cost per iteration if the number of unknowns is large. In order to reduce the computational effort spent on projections, inexact procedures have been proposed, resulting in more efficient methods, see for example [8, 16, 43, 51]. Moreover, considering inexact schemes provides theoretical support for real computational implementations of exact methods. It is worth mentioning that throughout the years there has been an increase in the popularity of inexact methods due to the emergence of large-scale problems in compressed sensing, machine learning applications and data fitting, see for instance [21, 44, 45, 46]. Motivated by practical and theoretical reasons, the purpose of the present paper is to present a new inexact version of the GPM, which we call Gradient-InexP method (GInexPM). It consists of using a general inexact projection instead of the exact one used in the GPM. The inexact projection concept considered in the present paper is a variation of the one appeared in [49, Example 1], which is defined by using an approximated property of the exact projection. In particular, it accepts the exact projection which can be adopted when it is easily obtained (for instance, the exact projection onto a box constraint or Lorentz cone can be easily obtained; see [42, p. 520] and [20, Proposition 3.3], respectively). It is worth noting that our approach to compute the inexact projection has not being considered in the study of the classical gradient method, in particular it is different from the ones proposed in [8, 16, 21, 43, 45, 51]. The analyses of GInexPM will be made employing two diferent step sizes, namely, constant step size and Armijo’s step size along the feasible direction. We point out that these step sizes are discussed extensively in the literature on this subject, where many of our results were inspired, see for example [4, 6, 28, 29, 33, 40]. Contributions: The main novelty in our work is the use of relative error tolerances in the computation of the inexact projection, to analyze the convergence properties of GInexPM. Our numerical experiments showed that GInexPM outperformed the GPM on a set of least squares problem over the spectrohedron. From a theoretical point of view, under suitable assumptions, the classic results of GPM were obtained for GInexPM as well. More specifically, we have showed that all cluster points of the sequence generated by GInexPM with constant step size or Armijo’s step size are solutions of problem (1). Futhermore, under convexity of the objective function, this sequence converges to a solution, if any. In both cases, the analysis establishes convergence results without any compactness assumption. We have also studied iteration-complexity bounds of GInexPM for both constant step size and Armijo’s step size. The presented analysis establishes that the complexity bound $\mathcal{O}(1/\sqrt{k})$ is unveil for finding $\epsilon$-stationary points for Problem (1), and, under convexity on $f$, the rate to find a $\epsilon$-optimal functional value is $\mathcal{O}(1/k)$. Content of the paper: In section 2, some notations and basic results used throughout the paper is presented. In particular, section 2.1 is devoted to present the concept relative feasible inexact projection and some properties about this concept. In section 3, we describe GInexPM method using the constant step size. The results of convergence using constant step size, as well as, results of iteration-complexity bound are presented in the sections 3.1 and 3.2, respectively. The results related to Armijo’s step sizes is presented in section 4. In section 4.1, we present the asymptotic convergence analysis of GInexPM using Armijo’s step size, and an iteration-complexity bound is presented in section 4.2. Numerical experiments are provided in section 5. Finally, the last section presents some final considerations. ## 2 Preliminaries In this section, we introduce some notation and results used throughout our presentation. We denote ${\mathbb{N}}:=\\{0,1,2,\ldots\\}$, $\langle\cdot,\cdot\rangle$ is the usual inner product in $\mathbb{R}^{n}$ and $\|\cdot\|$ is the Euclidean norm. Let $f:{\mathbb{R}}^{n}\to{\mathbb{R}}$ be a differentiable function and $C\subseteq{\mathbb{R}}^{n}$. The gradient $\nabla f$ of $f$ is said to be Lipschitz continuous on $C$ with constant $L>0$ if $\|\nabla f(x)-\nabla f(y)\|\leq L\|x-y\|,\qquad\forall~{}x,y\in C.$ (2) Combining this definition with the fundamental theorem of calculus, we obtain the following result, for which the proof can found in [6, Proposition A.24]. ###### Lemma 1. Let $f:{\mathbb{R}}^{n}\to{\mathbb{R}}$ be a differentiable function with Lipschitz continuous gradient on $C\subseteq{\mathbb{R}}^{n}$ with constant $L>0$. Then, $f(y)-f(x)-\langle\nabla f(x),y-x\rangle\leq\frac{L}{2}\|x-y\|^{2}$, for all $x,y\in C$. Let $f:{\mathbb{R}}^{n}\to{\mathbb{R}}$ be a differentiable function and $C\subseteq{\mathbb{R}}^{n}$ be a convex set. The function $f$ is strongly convex on $C$, if there exists a constant $\mu\geq 0$ such that $f(y)-f(x)-\langle\nabla f(x),y-x\rangle\geq\frac{\mu}{2}\|x-y\|^{2},\qquad\forall~{}x,y\in C.$ (3) When $\mu=0$, $f$ is said to be convex. If $f(x)\leq f(y)$ implies $\langle\nabla f(y),x-y\rangle\leq 0$, for any $x,y\in C$, then $f$ is said to be quasiconvex. Moreover, $f$ is said to be pseudoconvex if $\langle\nabla f(y),x-y\rangle\geq 0$ implies $f(x)\leq f(y)$, for any $x,y\in C$, for more details, see [37]. Recall that the convexity of a function guarantees pseudoconvexity, which in turn guarantees quasiconvexity, see [37]. A point ${\bar{x}}\in C$ is said to be a stationary point for Problem (1) if $\langle\nabla f({\bar{x}}),x-{\bar{x}}\rangle\geq 0,\qquad\forall~{}x\in C.$ (4) We end this section with a useful concept in the analysis of the sequence generated by the gradient method, for more details, see [12]. ###### Definition 1. A sequence $(y^{k})_{k\in\mathbb{N}}$ in $\mathbb{R}^{n}$ is quasi-Fejér convergent to a set $W\subset{\mathbb{R}}^{n}$ if, for every $w\in W$, there exists a sequence $(\epsilon_{k})_{k\in\mathbb{N}}\subset\mathbb{R}$ such that $\epsilon_{k}\geq 0$, $\sum_{k\in\mathbb{N}}\epsilon_{k}<+\infty$, and $\|y_{k+1}-w\|^{2}\leq\|y_{k}-w\|^{2}+\epsilon_{k}$, for all $k\in\mathbb{N}$. The main property of a quasi-Fejér convergent sequence is stated in the next result, and its proof can be found in [12]. ###### Theorem 2. Let $(y^{k})_{k\in\mathbb{N}}$ be a sequence in $\mathbb{R}^{n}$. If $(y^{k})_{k\in\mathbb{N}}$ is quasi-Fejér convergent to a nomempty set $W\subset{\mathbb{R}}^{n}$, then $(y^{k})_{k\in\mathbb{N}}$ is bounded. Furthermore, if a cluster point $y$ of $(y^{k})_{k\in\mathbb{N}}$ belongs to $W$, then $\lim_{k\rightarrow\infty}y_{k}=y$. ### 2.1 Inexact projection In this section, we present the concept of feasible inexact projection onto a closed and convex set. This concept has already been used in [1, 13, 14]. We also present some new properties of the feasible inexact projection used throughout the paper. The definition of feasible inexact projection is as follows. ###### Definition 2. Let $C\subset{\mathbb{R}}^{n}$ be a closed convex set and $\varphi_{\gamma}:{\mathbb{R}}^{n}\times{\mathbb{R}}^{n}\times{\mathbb{R}}^{n}\to{\mathbb{R}}_{+}$ be a function satisfying the following condition $\varphi_{\gamma}(u,v,w)\leq\gamma_{1}\|v-u\|^{2}+\gamma_{2}\|w-v\|^{2}+\gamma_{3}\|w-u\|^{2},\qquad\forall~{}u,v,w\in{\mathbb{R}}^{n},$ (5) where $\gamma=(\gamma_{1},\gamma_{2},\gamma_{3})\in{\mathbb{R}}^{3}_{+}$ is a given forcing parameter. The feasible inexact projection mapping relative to $u\in C$ with error tolerance $\varphi_{\gamma}$, denoted by ${\cal P}_{C}(\varphi_{\gamma},u,\cdot):{\mathbb{R}}^{n}\rightrightarrows C$, is the set-valued mapping defined as follows ${\cal P}_{C}(\varphi_{\gamma},u,v):=\left\\{w\in C:~{}\big{\langle}v-w,y-w\big{\rangle}\leq\varphi_{\gamma}(u,v,w),\quad\forall~{}y\in C\right\\}.$ (6) Each point $w\in{\cal P}_{C}(\varphi_{\gamma},u,v)$ is called a feasible inexact projection of $v$ onto $C$ relative to $u$ with error tolerance $\varphi_{\gamma}$. The feasible inexact projection generalizes the concept usual projection. In the following, we present some remarks about this concept and some examples of functions satisfying (5). ###### Remark 1. Let $\gamma_{1}$, $\gamma_{2}$ and $\gamma_{3}$ be nonnegative forcing parameters, $C\subset{\mathbb{R}}^{n}$, $u\in C$ and $\varphi_{\gamma}$ be as in Definition 2. Therefore, for all $v\in{\mathbb{R}}^{n}$, it follows from (6) that ${\cal P}_{C}(\varphi_{0},u,v)$ is the exact projection of $v$ onto $C$; see [6, Proposition 2.1.3, p. 201]. Moreover, ${\cal P}_{C}(\varphi_{0},u,v)\in{\cal P}_{C}(\varphi_{\gamma},u,v)$ which implies that ${\cal P}_{C}(\varphi_{\gamma},u,v)\neq\varnothing$, for all $u\in C$ and $v\in{\mathbb{R}}^{n}$. Consequently, the set-valued mapping ${\cal P}_{C}(\varphi_{\gamma},u,\cdot)$ as stated in (6) is well-defined. Note that the following functions $\varphi_{1}(u,v,w)=\gamma_{1}\|v-u\|^{2}+\gamma_{2}\|w-v\|^{2}+\gamma_{3}\|w-u\|^{2}$, $\varphi_{2}(u,v,w)=\gamma_{1}\|v-u\|^{2}$, $\varphi_{3}(u,v,w)=\gamma_{2}\|w-v\|^{2}$, $\varphi_{4}(u,v,w)=\gamma_{3}\|w-u\|^{2}$, and $\varphi_{5}(u,v,w)=\gamma_{1}\gamma_{2}\gamma_{3}\|v-u\|^{2}\|w-v\|^{2}\|w-u\|^{2}$ satisfy (5). Item (a) of the next lemma is a variation of [14, Lemma 6]. By using item (a), we will derive an inequality that together with this item will play an important role in the remainder of this paper. ###### Lemma 3. Let $v\in{\mathbb{R}}^{n}$, $u\in C$, $\gamma=(\gamma_{1},\gamma_{2},\gamma_{3})\in{\mathbb{R}}^{3}_{+}$ and $w\in{\cal P}_{C}(\varphi_{\gamma},u,v)$. Then, there hold: * (a) $\displaystyle\|w-x\|^{2}\leq\|v-x\|^{2}+\frac{2\gamma_{1}+2\gamma_{3}}{1-2\gamma_{3}}\|v-u\|^{2}-\frac{1-2\gamma_{2}}{1-2\gamma_{3}}\|w-v\|^{2}$, for all $x\in C$ and $0\leq\gamma_{3}<1/2$; * (b) $\displaystyle\big{\langle}v-w,y-w\big{\rangle}\leq\frac{\gamma_{1}+\gamma_{2}}{1-2\gamma_{2}}\|v-u\|^{2}+\frac{\gamma_{3}-\gamma_{2}}{1-2\gamma_{2}}\|w-u\|^{2}$, for all $y\in C$ and $0\leq\gamma_{2}<1/2$. ###### Proof. Let $x\in C$ and $0\leq\gamma_{3}<1/2$. First note that $\|w-x\|^{2}=\|v-x\|^{2}-\|w-v\|^{2}+2\langle v-w,x-w\rangle$. Since $w\in{\cal P}_{C}(\varphi_{\gamma},u,v)$, combining the last equality with (5) and (6), we obtain $\|w-x\|^{2}\leq\|v-x\|^{2}-(1-2\gamma_{2})\|w-v\|^{2}+2\gamma_{1}\|v-u\|^{2}+2\gamma_{3}\|w-u\|^{2}.$ (7) On the other hand, we have $\|w-u\|^{2}=\|v-u\|^{2}-\|w-v\|^{2}+2\langle v-w,u-w\rangle$. Thus, due to $w\in{\cal P}_{C}(\varphi_{\gamma},u,v)$ and $u\in C$, using (5) and (6), and considering that $0\leq\gamma_{3}<1/2$, we have $\|w-u\|^{2}\leq\frac{1+2\gamma_{1}}{1-2\gamma_{3}}\|v-u\|^{2}-\frac{1-2\gamma_{2}}{1-2\gamma_{3}}\|w-v\|^{2}.$ Therefore, combining the last inequality with (7), we obtain the inequality of item $(a)$. For proving item $(b)$, take $y\in C$ and $0\leq\gamma_{2}<1/2$. Using (5) and (6), we have $\big{\langle}v-w,y-w\big{\rangle}\leq\gamma_{1}\|v-u\|^{2}+\gamma_{2}\|w-v\|^{2}+\gamma_{3}\|w-u\|^{2}.$ (8) Applying item $(a)$ with $x=u$, after some algebraic manipulations, we conclude that $\|w-v\|^{2}\leq\frac{1+2\gamma_{1}}{1-2\gamma_{2}}\|v-u\|^{2}-\frac{1-2\gamma_{3}}{1-2\gamma_{2}}\|w-u\|^{2}.$ The last inequality together (8) yield $\big{\langle}v-w,y-w\big{\rangle}\leq\left(\gamma_{1}+\frac{\gamma_{2}+2\gamma_{1}\gamma_{2}}{1-2\gamma_{2}}\right)\|v-u\|^{2}+\left(\gamma_{3}-\frac{\gamma_{2}-2\gamma_{2}\gamma_{3}}{1-2\gamma_{2}}\right)\|w-u\|^{2},$ which is equivalent to the inequality in $(b)$. ∎ ## 3 GInexPM employing the constant step size rule In this section, we describe the GInexPM with a feasible inexact projection for solving problem (1). The rule for choosing the step size will be the same used in [3, 6], namely, the constant step size rule. For that, we take a exogenous sequence of real numbers $(a_{k})_{k\in\mathbb{N}}$ satisfying $0\leq a_{k}\leq b_{k-1}-b_{k},\qquad k=0,1,\ldots,$ (9) for some given nonincreasing sequence of nonegative real numbers $(b_{k})_{k\in\mathbb{N}}$ converging to zero, with the notation $b_{-1}\in{\mathbb{R}}_{++}$ such that $b_{-1}>b_{0}$. ###### Remark 2. Condition (9) implies that $\sum_{k\in\mathbb{N}}a_{k}<b_{-1}$. Examples of sequences $(a_{k})_{k\in\mathbb{N}}$ and $(b_{k})_{k\in\mathbb{N}}$ satisfying (9) are obtained by taking $a_{k}:=b_{k-1}-b_{k}$ and, for a given $\bar{b}>0$: (i) $b_{-1}=3\bar{b}$, $b_{0}=2\bar{b}$, $b_{k}=\bar{b}/k$, for all $k=1,2,\ldots$; (ii) $b_{-1}=3\bar{b}$, $b_{0}=2\bar{b}$, $b_{k}=\bar{b}/\ln(k+1)$, for all $k=1,2,\ldots$. The conceptual GInexPM is formally stated as follows. Algorithm 1 GInexPM employing constant step size Step 0: Take $(a_{k})_{k\in\mathbb{N}}$, $(b_{k})_{k\in\mathbb{N}}$ satisfying (9) and an error tolerance function $\varphi_{\gamma}$. Let $x^{0}\in C$ and set $k=0$. Step 1: If $\nabla f(x^{k})=0$, then stop; otherwise, choose real numbers $\gamma_{1}^{k},\gamma_{2}^{k}$ and $\gamma_{3}^{k}$ such that $0\leq\gamma_{1}^{k}+\gamma_{2}^{k}\leq\frac{a_{k}}{\|\nabla f(x^{k})\|^{2}},\qquad 0\leq\gamma_{2}^{k}<\bar{\gamma_{2}}<\frac{1}{2},\qquad 0\leq\gamma_{3}^{k}<\bar{\gamma}<\frac{1}{2},$ (10) and a fixed step size $\alpha>0$ and define the next iterate $x^{k+1}$ as any feasible inexact projection of $z^{k}:=x^{k}-\alpha\nabla f(x^{k})$ onto $C$ relative to $x^{k}$ with forcing parameters $\gamma^{k}:=(\gamma^{k}_{1},\gamma^{k}_{2},\gamma^{k}_{3})$, i.e., $x^{k+1}\in{\cal P}_{C}\left(\varphi_{\gamma^{k}},x^{k},z^{k}\right).$ Step 2: Set $k\leftarrow k+1$, and go to Step 1. Let us describe the main features of the GInexPM. Firstly, we take exogenous sequences $(a_{k})_{k\in\mathbb{N}}$ and $(b_{k})_{k\in\mathbb{N}}$ satisfying (9) and an error tolerance function $\varphi_{\gamma}$. Then, we check if at the current iterate $x^{k}$ we have $\nabla f(x^{k})=0$, otherwise, we choose nonnegative forcing parameters $\gamma_{1}^{k}$, $\gamma_{2}^{k}$ and $\gamma_{3}^{k}$ satisfying (10). Set a fixed step size $\alpha>0$. By using some inner procedure, the next iterate $x^{k+1}$ is computed as any feasible inexact projection of $z^{k}=x^{k}-\alpha\nabla f(x^{k})$ onto the feasible set $C$ relative to $x^{k}$, i.e., $x^{k+1}\in{\cal P}_{C}(\varphi_{\gamma^{k}},x^{k},z^{k})$; an example of such procedure will be presented in section 5. Note that, if $\gamma_{1}^{k}\equiv 0$, $\gamma_{2}^{k}\equiv 0$ and $\gamma_{3}^{k}\equiv 0$, then ${\cal P}_{C}({\varphi_{0}},x^{k},z^{k})$ is the exact projection, see Remark 1, and our method corresponds to the usual projected gradient method proposed, for example, in [3, 6]. It is worth noting that $\gamma_{1}^{k}$ and $\gamma_{2}^{k}$ in (10) can be chosen as any nonnegative real numbers satisfying $0\leq(\gamma_{1}^{k}+\gamma_{2}^{k})\|\nabla f(x^{k})\|^{2}\leq a_{k}$, for prefixed sequences $(a_{k})_{k\in\mathbb{N}}$ and $(b_{k})_{k\in\mathbb{N}}$ satisfying (9). In this case, we have $\sum_{k\in\mathbb{N}}\big{[}(\gamma_{1}^{k}+\gamma_{2}^{k})\|\nabla f(x^{k})\|^{2}\big{]}<+\infty.$ (11) In the next sections, we will deal with the convergence analysis of the sequence $(x^{k})_{k\in\mathbb{N}}$ generated by GInexPM. ### 3.1 Asymptotic convergence analysis The aim of this section is to prove the main convergence results about the asymptotic behavior of the sequence $(x^{k})_{k\in\mathbb{N}}$ generated by Algorithm 1. We assume that the gradient of the objective function $f$ is Lipschitz continuous with constant $L>0$. Moreover, we also assume that $0<\alpha<\frac{1-2\bar{\gamma}}{L}.$ (12) For future references, it is convenient to define the following constants: $\nu:=\frac{1-\bar{\gamma_{2}}-\bar{\gamma}}{\alpha}-\frac{L}{2}>0,\qquad\rho:=\frac{\alpha}{1-2\bar{\gamma_{2}}}>0.$ (13) In the sequel, we state and prove our first result for the sequence $(x^{k})_{k\in\mathbb{N}}$. The obtained inequality is the counterpart of the one obtained, for example, in [3, Lemma 9.11, p. 176]. ###### Lemma 4. The following inequality holds: $f(x^{k+1})\leq f(x^{k})+\rho(\gamma_{1}^{k}+\gamma_{2}^{k})\|\nabla f(x^{k})\|^{2}-\nu\|x^{k+1}-x^{k}\|^{2},\qquad\forall~{}{k\in\mathbb{N}}.$ (14) ###### Proof. Since $\nabla f$ satisfies (2), applying Lemma 1 with $x=x^{k}$ and $y=x^{k+1}$, we obtain $f(x^{k+1})\leq f(x^{k})+\big{\langle}\nabla f(x^{k}),x^{k+1}-x^{k}\big{\rangle}+\frac{L}{2}\|x^{k+1}-x^{k}\|^{2}.$ Thus, after some algebraic manipulations, we have $f(x^{k+1})\leq f(x^{k})+\frac{1}{\alpha}\big{\langle}[x^{k}-\alpha\nabla f(x^{k})]-x^{k+1},x^{k}-x^{k+1}\big{\rangle}-\left(\frac{1}{\alpha}-\frac{L}{2}\right)\|x^{k+1}-x^{k}\|^{2}.$ (15) Since $x^{k+1}\in{\cal P}_{C}(\varphi_{\gamma^{k}},x^{k},z^{k})$ with $z^{k}=x^{k}-\alpha\nabla f(x^{k})$, applying item $(b)$ of Lemma 3 with $u=x^{k}$, $y=x^{k}$, $v=z^{k}$, $w=x^{k+1}$, and $\varphi_{\gamma}=\varphi_{\gamma_{k}}$, we have $\big{\langle}[x^{k}-\alpha\nabla f(x^{k})]-x^{k+1},x^{k}-x^{k+1}\big{\rangle}\leq\frac{\gamma_{1}^{k}+\gamma_{2}^{k}}{1-2\gamma_{2}^{k}}\alpha^{2}\|\nabla f(x^{k})\|^{2}+\frac{\gamma_{3}^{k}-\gamma_{2}^{k}}{1-2\gamma_{2}^{k}}\|x^{k+1}-x^{k}\|^{2}.$ Then, combining (15) with the latter inequality yields $f(x^{k+1})\leq f(x^{k})+\frac{\alpha(\gamma_{1}^{k}+\gamma_{2}^{k})}{1-2\gamma_{2}^{k}}\|\nabla f(x^{k})\|^{2}-\left[\frac{1-\gamma_{2}^{k}-\gamma_{3}^{k}}{\alpha(1-2\gamma_{2}^{k})}-\frac{L}{2}\right]\|x^{k+1}-x^{k}\|^{2}.$ Therefore, taking into account (10) and (13), we have (14), which concludes the proof. ∎ The next result is an immediate consequence of Lemma 4. ###### Corollary 5. The sequence $(f(x^{k})+\rho b_{k-1})_{k\in\mathbb{N}}$ is monotone non- increasing. In particular, $\inf_{k}(f(x^{k})+\rho b_{k-1})=\inf_{k}f(x^{k})$. ###### Proof. Combining (10) with (14), we have $f(x^{k+1})\leq f(x^{k})+\rho a_{k}$, for all ${k\in\mathbb{N}}$. Thus, taking into account that $(a_{k})_{k\in\mathbb{N}}$ satisfies (9), we obtain $f(x^{k+1})+\rho{b_{k}}\leq f(x^{k})+\rho{b_{k-1}}$, for all ${k\in\mathbb{N}}$, implying the first statement. Since $(b_{k})_{k\in\mathbb{N}}$ converges to zero, the second statement holds. ∎ Now, we are ready to state and prove a partial asymptotic convergence result on $(x^{k})_{k\in\mathbb{N}}$. ###### Theorem 6. Assume that $-\infty<f^{*}$. If ${\bar{x}}\in C$ is a cluster point of the sequence $(x^{k})_{k\in\mathbb{N}}$, then ${\bar{x}}$ is a stationary point for problem (1). ###### Proof. By (10), we have $(\gamma_{1}^{k}+\gamma_{2}^{k})\|\nabla f(x^{k})\|^{2}\leq a_{k}$, for all ${k\in\mathbb{N}}$. Thus, Lemma 4 implies that that $\nu\|x^{k+1}-x^{k}\|^{2}\leq[f(x^{k})-f(x^{k+1})]+\rho a_{k}$, for all ${k\in\mathbb{N}}$. Using (9), after some adjustments, we obtain $\|x^{k+1}-x^{k}\|^{2}\leq\left[f(x^{k})+\rho b_{k-1}\right]/{\nu}-\left[f(x^{k+1})+\rho b_{k}\right]/{\nu}$, for all $k\in\mathbb{N}$. Hence, due to $f^{*}\leq\inf_{k}f(x^{k})$, Corollary 5 implies $\sum_{\ell=0}^{k}\|x^{\ell+1}-x^{\ell}\|^{2}\leq[f(x^{0})+\rho b_{-1}]/{\nu}-f^{*}/{\nu}$. Thus, we conclude that $\lim_{k\to+\infty}\|x^{k+1}-x^{k}\|=0$. Let ${\bar{x}}$ be a cluster point of $(x^{k})_{k\in\mathbb{N}}$ and $(x^{k_{j}})_{j\in\mathbb{N}}$ a subsequence of $(x^{k})_{k\in\mathbb{N}}$ such that $\lim_{j\to+\infty}x^{k_{j}}=~{}\bar{x}$. Since, $\lim_{j\to+\infty}(x^{k_{j}+1}-x^{k_{j}})=0$, we have $\lim_{j\to+\infty}x^{k_{j}+1}={\bar{x}}$. On the other hand, due to $x^{k_{j}+1}\in{\cal P}_{C}(\varphi_{\gamma^{k_{j}}},x^{k_{j}},z^{k_{j}})$, where $z^{k_{j}}:=x^{k_{j}}-\alpha\nabla f(x^{k_{j}})$, applying item $(b)$ of Lemma 3 with $v=z^{k_{j}}$, $u=x^{k_{j}}$, $w=x^{k_{j}+1}$, and $\varphi_{\gamma}=\varphi_{\gamma^{k_{j}}},$ we obtain $\big{\langle}z^{k_{j}}-x^{k_{j}+1},y-x^{k_{j}+1}\big{\rangle}\leq\frac{\gamma_{1}^{k_{j}}+\gamma_{2}^{k_{j}}}{1-2\gamma_{2}^{k_{j}}}\alpha^{2}\|\nabla f(x^{k_{j}})\|^{2}+\frac{\gamma_{3}^{k_{j}}-\gamma_{2}^{k_{j}}}{1-2\gamma_{2}^{k_{j}}}\|x^{k_{j}+1}-x^{k_{j}}\|^{2},\qquad\forall~{}y\in C.$ Thus, taking limits on both sides of the last inequality, we conclude, by using (11) and continuity of $\nabla f$, that $\big{\langle}[{\bar{x}}-\alpha\nabla f({\bar{x}})]-{\bar{x}},y-{\bar{x}}\big{\rangle}\leq 0$, for all $y\in C$. Therefore, $\big{\langle}\nabla f({\bar{x}}),y-{\bar{x}}\big{\rangle}\geq 0$, for all $y\in C$, which implies that ${\bar{x}}\in C$ is a stationary point for problem (1). ∎ In the following lemma, we establish a basic inequality satisfied by $(x^{k})_{k\in\mathbb{N}}$. In particular, it will be useful to prove the full asymptotic convergence of $(x^{k})_{k\in\mathbb{N}}$ under quasiconvexity of $f$. ###### Lemma 7. For each $x\in C$ and ${k\in\mathbb{N}}$, there holds $\|x^{k+1}-x\|^{2}\leq\|x^{k}-x\|^{2}+2\alpha\rho(\gamma_{1}^{k}+\gamma_{2}^{k})\|\nabla f(x^{k})\|^{2}+2\alpha\big{[}f(x^{k})-f(x^{k+1})+\langle\nabla f(x^{k}),x-x^{k}\rangle\big{]}.$ ###### Proof. Let $x\in C$. By using $z_{k}=x^{k}-\alpha\nabla f(x^{k})$, after some algebraic manipulations, we have $\displaystyle\|x^{k+1}-x\|^{2}=\|x^{k}-x\|^{2}-\|x^{k+1}-x^{k}\|^{2}+2\big{\langle}z^{k}-x^{k+1},x-x^{k+1}\big{\rangle}+2\alpha\big{\langle}\nabla f(x^{k}),x-x^{k+1}\big{\rangle}.$ Since $x^{k+1}\in{\cal P}_{C}(\varphi_{\gamma^{k}},x^{k},z^{k})$, applying item $(b)$ of Lemma 3 with $y=x$, $v=z^{k}$, $u=x^{k}$, $w=x^{k+1}$, and $\varphi_{\gamma}=\varphi_{\gamma^{k}},$ we obtain $\big{\langle}z^{k}-x^{k+1},x-x^{k+1}\big{\rangle}\leq\frac{\gamma_{1}^{k}+\gamma_{2}^{k}}{1-2\gamma_{2}^{k}}\alpha^{2}\|\nabla f(x^{k})\|^{2}+\frac{\gamma_{3}^{k}-\gamma_{2}^{k}}{1-2\gamma_{2}^{k}}\|x^{k+1}-x^{k}\|^{2}.$ On the other hand, since $\nabla f$ satisfies (2), Lemma 1 with $x=x^{k}$ and $y=x^{k+1}$ yields $\displaystyle\big{\langle}\nabla f(x^{k}),x-x^{k+1}\big{\rangle}$ $\displaystyle=\big{\langle}\nabla f(x^{k}),x^{k}-x^{k+1}\big{\rangle}+\big{\langle}\nabla f(x^{k}),x-x^{k}\big{\rangle}$ $\displaystyle\leq f(x^{k})-f(x^{k+1})+\frac{L}{2}\|x^{k+1}-x^{k}\|^{2}+\big{\langle}\nabla f(x^{k}),x-x^{k}\big{\rangle}.$ Combining last two inequities with the above equality, we conclude that $\|x^{k+1}-x\|^{2}\leq\|x^{k}-x\|^{2}-\left[\frac{1-2\gamma_{3}^{k}}{1-2\gamma_{2}^{k}}-\alpha L\right]\|x^{k+1}-x^{k}\|^{2}\\\ +2\alpha^{2}\left(\frac{\gamma_{1}^{k}+\gamma_{2}^{k}}{1-2\gamma_{2}^{k}}\right)\|\nabla f(x^{k})\|^{2}+2\alpha\left[f(x^{k})-f(x^{k+1})+\big{\langle}\nabla f(x^{k}),x-x^{k}\big{\rangle}\right].$ (16) Taking into account (10), we have $0\leq 1-2\bar{\gamma_{2}}\leq 1-2\gamma_{2}^{k}\leq 1$ and $1-2\gamma_{3}^{k}\geq 1-2\bar{\gamma}\geq 0$. Hence, it follows from (12) and (13) that $\frac{\alpha}{1-2\gamma_{2}^{k}}\leq\frac{\alpha}{1-2\bar{\gamma_{2}}}=\rho,\qquad\qquad\left[\frac{1-2\gamma_{3}^{k}}{1-2\gamma_{2}^{k}}-\alpha L\right]\geq\left[1-2\gamma_{3}^{k}-\alpha L\right]\geq\left[1-2\bar{\gamma}-\alpha L\right]\geq 0.$ These inequalities, together with (16), imply the desired inequality, which concludes the proof. ∎ To proceed with the analysis of $(x^{k})_{k\in\mathbb{N}}$, we also need the following auxiliary set $T:=\left\\{x\in C:f(x)\leq\inf_{k}f(x^{k}),\quad{k\in\mathbb{N}}\right\\}.$ ###### Corollary 8. Assume that $f$ is a quasiconvex function. If $T\neq\varnothing$, then $(x^{k})_{k\in\mathbb{N}}$ converges to a stationary point for problem (1). ###### Proof. Let $x\in T$. Since $f$ is a quasiconvex function and $f(x)\leq f(x^{k})$ for all $k\in\mathbb{N}$, we have $\big{\langle}\nabla f(x^{k}),x-x^{k}\big{\rangle}\leq~{}0$, for all $k\in\mathbb{N}$. Thus, applying Lemma 7, we conclude that $\|x^{k+1}-x\|^{2}\leq\|x^{k}-x\|^{2}+2\alpha\rho(\gamma_{1}^{k}+\gamma_{2}^{k})\|\nabla f(x^{k})\|^{2}+\\\ 2\alpha\left[f(x^{k})-f(x^{k+1})\right],\qquad\forall~{}{k\in\mathbb{N}}.$ Thus, using the first condition in (10) and considering that $(a_{k})_{k\in\mathbb{N}}$ and $(b_{k})_{k\in\mathbb{N}}$ satisfy (9), the latter inequality implies $\|x^{k+1}-x\|^{2}\leq\|x^{k}-x\|^{2}+2\alpha\left([f(x^{k})+\rho b_{k-1}]-[f(x^{k+1})+\rho b_{k}]\right),\quad\forall~{}k\in\mathbb{N}.$ (17) On the other hand, performing a sum of (17) for $k=0,1,\ldots,N$ and using that $x\in T$, we obtain $\sum_{k=0}^{N}\left(\big{[}f(x^{k})+\rho b_{k-1}\big{]}-\big{[}f(x^{k+1})+\rho b_{k}\big{]}\right)\leq f(x^{0})-f(x)+\rho(b_{-1}-b_{N}),$ (18) for any $N\in\mathbb{N}$. Therefore, (17) and (18) imply that $(x^{k})_{k\in\mathbb{N}}$ is quasi-Fejér convergent to $T$. Since by assumption $T\neq\varnothing$, it follows from Theorem 2 that $(x^{k})_{k\in\mathbb{N}}$ is bounded. Let $\bar{x}$ be a cluster point of $(x^{k})_{k\in\mathbb{N}}$ and $(x^{k_{j}})_{j\in\mathbb{N}}$ a subsequence of $(x^{k})_{k\in\mathbb{N}}$ such that $\lim_{j\to+\infty}x^{k_{j}}=\bar{x}$. Considering that $f$ is continuous, it follows from Corollary 5 that $\inf_{k}f(x^{k})=\inf_{k}\left(f(x^{k})+\rho b_{k-1}\right)=\lim_{j\to+\infty}\left(f(x^{k_{j}})+\rho b_{k_{j}-1}\right)=\lim_{j\to+\infty}f(x^{k_{j}})=f(\bar{x}).$ Therefore $\bar{x}\in T$. Using again Theorem 2, we have that $(x^{k})_{k\in\mathbb{N}}$ converges to $\bar{x}$, and the conclusion is obtained from Theorem 6. ∎ In the following, we present an important result, when $(x^{k})_{k\in\mathbb{N}}$ has no cluster points. This result has already appeared in several papers studding gradient method with exact projetion, see for example [4, 30]; however, since its proof is very simple and concise, we include it here for the sake of completeness. ###### Lemma 9. If $f$ is a quasiconvex function and $(x^{k})_{k\in\mathbb{N}}$ has no cluster points then $\Omega^{*}=\varnothing$, $\lim_{k\to\infty}\|x^{k}\|=\infty$, and $\lim_{k\to\infty}f(x^{k})=\inf\\{f(x):x\in C\\}$. ###### Proof. Since $(x^{k})_{k\in\mathbb{N}}$ has no cluster points, then $\lim_{k\to\infty}\|x^{k}\|=\infty$. Assume that problem (1) has an optimum, say $\tilde{x}$, so $f(\tilde{x})\leq f(x^{k})$ for all $k$. Thus, $\tilde{x}\in T$. Using Corollary 8, we have that $(x^{k})_{k\in\mathbb{N}}$ is convergent, contradicting that $\lim_{k\to\infty}\|x^{k}\|=\infty$. Therefore, $\Omega^{*}=\varnothing$. Now, we claim that $\lim_{k\to\infty}f(x^{k})=\inf\\{f(x):x\in C\\}$. If $\lim_{k\to\infty}f(x^{k})=-\infty$, the claim holds. Let $f^{*}=\inf_{x\in C}f(x)$. By contradiction, suppose that $\lim_{k\to\infty}f(x^{k})>f^{*}$. Then, there exists $\tilde{x}\in C$ such that $f(\tilde{x})\leq f(x^{k})$ for all $k$. Using Corollary 8, we obtain that $(x^{k})_{k\in\mathbb{N}}$ is convergent, contradicting again $\lim_{k\to\infty}\|x^{k}\|=\infty$, which conclude the proof. ∎ Finally, we presented the main convergence result when $f$ is pseudoconvex, which is a version of [4, Corollary 3] for our algorithm, see also [29, Proposition 5]. ###### Theorem 10. Assume that $f$ is a pseudoconvex function. Then, $\Omega^{*}\neq\varnothing$ if and only if $(x^{k})_{k\in\mathbb{N}}$ has at least one cluster point. Moreover, $(x^{k})_{k\in\mathbb{N}}$ converges to an optimum point if $\Omega^{*}\neq\varnothing$; otherwise, $\lim_{k\to\infty}\|x^{k}\|=\infty$ and $\lim_{k\to\infty}f(x^{k})=\inf\\{f(x):x\in C\\}$. ###### Proof. Note that pseudoconvex functions are quasiconvex. First assume that $\Omega^{*}\neq\varnothing$. In this case, we have also $T\neq\varnothing$. Thus, using Corollary 8, we conclude that $(x^{k})_{k\in\mathbb{N}}$ converges to a stationary point of problem (1) and, in particular, $(x^{k})_{k\in\mathbb{N}}$ has a cluster point. Considering that $f$ is pseudoconvex, this point is also an optimum point. Reciprocally, let $\bar{x}$ be a cluster point of $(x^{k})_{k\in\mathbb{N}}$ and $(x^{k_{j}})_{j\in\mathbb{N}}$ a subsequence of $(x^{k})_{k\in\mathbb{N}}$ such that $\lim_{j\to+\infty}x^{k_{j}}=\bar{x}$. Since, by Corollary 5, $(f(x^{k})+\rho b_{k-1})_{k\in\mathbb{N}}$ is monotone non-increasing, using the continuity of $f$, we have $\inf_{k}f(x^{k})=\inf_{k}\left(f(x^{k})+\rho b_{k-1}\right)=\lim_{j\to+\infty}\left(f(x^{k_{j}})+\rho b_{k_{j}-1}\right)=\lim_{j\to+\infty}f(x^{k_{j}})=f(\bar{x}).$ Therefore $\bar{x}\in T$. From Corollary 8, we obtain that $(x^{k})_{k\in\mathbb{N}}$ converges to a stationary point $\tilde{x}$ of problem (1). Thus, by (4), we have $\langle\nabla f(\tilde{x}),x-\tilde{x}\rangle\geq 0$ for all $x\in C$, which by the pseudo- convexity of $f$ implies $f(x)\geq f(\tilde{x})$ for all $x\in C$. Therefore, $\bar{x}\in\Omega^{*}$ and $\Omega^{*}\neq\varnothing$. The last part of the theorem follows by combining the first one with Lemma 9. ∎ ### 3.2 Iteration-complexity bound In this section, we establish some iteration-complexity bounds for the sequence $(x^{k})_{k\in\mathbb{N}}$ generated by Algorithm 1. For that, we take $x^{*}\in\Omega^{*}\neq\varnothing$, set $f^{*}=f(x^{*})$, and define the constant $\eta:=f(x^{0})-f^{*}+\rho b_{-1}.$ In next result, we do not assume any assumption on the convexity of the objective function. ###### Theorem 11. Let $\nu>0$ as in (13). Then, for all $N\in\mathbb{N}$, there holds $\min\left\\{\|x^{k+1}-x^{k}\|:~{}\forall~{}{k\in\mathbb{N}}\right\\}\leq\frac{\sqrt{\eta/\nu}}{\sqrt{N+1}}.$ (19) ###### Proof. It follows from (10) and (14) that $f(x^{k+1})\leq f(x^{k})+\rho a_{k}-\nu\|x^{k+1}-x^{k}\|^{2}$, for all ${k\in\mathbb{N}}$. Hence using (9), we have $\|x^{k+1}-x^{k}\|^{2}\leq\frac{1}{\nu}\left[\big{(}f(x^{k})+\rho b_{k-1}\big{)}-\big{(}f(x^{k+1})+\rho b_{k}\big{)}\right]$, for all ${k\in\mathbb{N}}$. Thus, summing both sides for $k=0,1,\ldots,N$ and using that $f^{*}\leq f(x^{k})$ for all ${k\in\mathbb{N}}$ , we obtain $\sum_{k=0}^{N}\|x^{k+1}-x^{k}\|^{2}\leq\frac{1}{\nu}\big{[}f(x^{0})-f^{*}+\rho(b_{-1}-b_{N})\big{]}\leq\frac{1}{\nu}\big{[}f(x^{0})-f^{*}+\rho b_{-1})\big{]}=\eta/\nu.$ Therefore, $(N+1)\min\left\\{\|x^{k+1}-x^{k}\|^{2}:~{}k=0,1,\ldots,N\right\\}\leq\eta/\nu$, which implies (19). ∎ In the following, we present an iteration-complexity bound for the sequence $(x^{k})_{k\in\mathbb{N}}$, for finding $\epsilon$-stationary points of function $f$. ###### Theorem 12. For every $N\in\mathbb{N}$, there holds $\min_{k=0,1,\ldots,N}\big{\langle}\nabla f(x^{k}),x^{k}-x\big{\rangle}\leq\left[\frac{1}{2\alpha}\|x^{0}-x\|^{2}+\eta\right]\frac{1}{N+1},\qquad\forall~{}x\in C.$ As a consequence, given $\epsilon>0$, the maximum number of iterations $N$ necessary for Algorithm 1 to generate an iterate $x^{\ell}$ such that $\big{\langle}\nabla f(x^{\ell}),x-x^{\ell}\big{\rangle}>-\epsilon$, for all $x\in C$, is $N\geq[\frac{1}{2\alpha}\|x^{0}-x\|^{2}+\eta]/\epsilon-1$. ###### Proof. Using Lemma 7 and (10), we obtain $\big{\langle}\nabla f(x^{k}),x^{k}-x\big{\rangle}\leq\frac{1}{2\alpha}\left[\|x^{k}-x^{*}\|^{2}-\|x^{k+1}-x^{*}\|^{2}\right]+\left[\rho a_{k}+f(x^{k})-f(x^{k+1})\right],$ for all ${k\in\mathbb{N}}$. Since $(a_{k})_{k\in\mathbb{N}}$ and $(b_{k})_{k\in\mathbb{N}}$ satisfy (9), we conclude that $\big{\langle}\nabla f(x^{k}),x^{k}-x\big{\rangle}\leq\frac{1}{2\alpha}\left[\|x^{k}-x^{*}\|^{2}-\|x^{k+1}-x^{*}\|^{2}\right]+\left[\left(f(x^{k})+\rho b_{k-1}\right)-\left(f(x^{k+1})+\rho b_{k}\right)\right].$ Thus, summing both sides for $k=0,1,\ldots,N$ and using that $f^{*}\leq f(x^{k})$ for all ${k\in\mathbb{N}}$, we have $\sum_{k=0}^{N}\big{\langle}\nabla f(x^{k}),x^{k}-x\big{\rangle}\leq\frac{1}{2\alpha}\|x^{0}-x\|^{2}+\big{[}f(x^{0})-f^{*}+\rho(b_{-1}-b_{N})\big{]}=\frac{1}{2\alpha}\|x^{0}-x\|^{2}+\eta,$ which implies that $(N+1)\min\left\\{\big{\langle}\nabla f(x^{k}),x^{k}-x\big{\rangle}:~{}k=0,1,\ldots,N\right\\}\leq\frac{1}{2\alpha}\|x^{0}-x\|^{2}+\eta,$ obtaining the first statement of the theorem. The second statement follows trivially from the first one. ∎ #### 3.2.1 Iteration-complexity bound under convexity Next result presents an iteration-complexity bound for $(f(x^{k}))_{k\in\mathbb{N}}$ when $f$ is a convex function. Similar bound for unconstrained problems can be found in [40, Theorem 2.1.14]. ###### Theorem 13. Assume that $f$ is a convex function. Then, for every $N\in\mathbb{N}$, there holds $\min\left\\{f(x^{k})-f^{*}:~{}k=1\ldots,N\right\\}\leq\frac{\|x^{0}-x^{*}\|+2\alpha\rho b_{-1}}{2\alpha N}.$ ###### Proof. Since $f$ is convex, we have $2\alpha\langle\nabla f(x^{k}),x^{*}-x^{k}\rangle\leq 2\alpha\left[f(x^{*})-f(x^{k})\right]$. Thus, using (10) and Lemma 7 with $x=x^{*}$, we obtain $2\alpha[f(x^{k+1})-f^{*}]\leq\|x^{k}-x^{*}\|^{2}-\|x^{k+1}-x^{*}\|^{2}+2\alpha\rho a_{k}$, for all $k\in\mathbb{N}$. Performing the sum of the this inequality for $k=0,1,\ldots,N-1$, we have $2\alpha\sum_{k=0}^{N-1}\left[f(x^{k+1})-f^{*}\right]\leq\|x^{0}-x^{*}\|^{2}-\|x^{N}-x^{*}\|^{2}+2\alpha\rho\sum_{k=0}^{N-1}a_{k}.$ Since $\sum_{k\in\mathbb{N}}a_{k}<b_{-1}$, we have $2\alpha N\min\left\\{f(x^{k})-f^{*}:k=1\ldots,N\right\\}\leq\|x^{0}-x^{*}\|+2\alpha\rho b_{-1},$ which implies the desired inequality. ∎ #### 3.2.2 Iteration-complexity bound under strong convexity Our next goal is to show an iteration-complexity bound for $\left(f(x^{k})\right)_{k\in\mathbb{N}}$ when $f$ is strongly convex. For this purpose, we first present an inequality that is a variation of [11, Lemma 3.6]. ###### Lemma 14. Assume that $f$ is $\mu-$strongly convex. Then, for all $k\in\mathbb{N}$, there holds $f(x^{k+1})-f^{*}\leq\frac{1}{\alpha}\big{\langle}x^{k}-x^{k+1},x^{k}-x^{*}\big{\rangle}-\nu\|x^{k+1}-x^{k}\|^{2}+\frac{\gamma_{1}^{k}+\gamma_{2}^{k}}{1-2\gamma_{2}^{k}}\alpha\|\nabla f(x^{k})\|^{2}-\frac{\mu}{2}\|x^{k}-x^{*}\|^{2}.$ ###### Proof. Applying Lemma 1 with $x=x^{k}$ and $y=x^{k+1}$, and then using (3), we obtain $\displaystyle f(x^{k+1})-f^{*}$ $\displaystyle\leq\big{\langle}\nabla f(x^{k}),x^{k+1}-x^{k}\big{\rangle}+\frac{L}{2}\|x^{k}-x^{k+1}\|^{2}+\big{\langle}\nabla f(x^{k}),x^{k}-x^{*}\big{\rangle}-\frac{\mu}{2}\|x^{k}-x^{*}\|^{2}.$ $\displaystyle=\big{\langle}\nabla f(x^{k}),x^{k+1}-x^{*}\big{\rangle}+\frac{L}{2}\|x^{k}-x^{k+1}\|^{2}-\frac{\mu}{2}\|x^{k}-x^{*}\|^{2}.$ (20) On the order hand, due to $z^{k}=x^{k}-\alpha\nabla f(x^{k})$, after some algebraic manipulations, we have $\big{\langle}\nabla f(x^{k}),x^{k+1}-x^{*}\big{\rangle}=\frac{1}{\alpha}\big{\langle}x^{k}-x^{k+1},x^{k+1}-x^{*}\big{\rangle}+\frac{1}{\alpha}\big{\langle}z^{k}-x^{k+1},x^{*}-x^{k+1}\big{\rangle}.$ (21) Since $x^{k+1}\in{\cal P}_{C}(\varphi_{\gamma^{k}},x^{k},z^{k})$, applying item $(b)$ of Lemma 3 with $y=x^{*}$, $v=z^{k}$, $u=x^{k}$, $w=x^{k+1}$, and $\varphi_{\gamma}=\varphi_{\gamma^{k}}$, we obtain $\big{\langle}z^{k}-x^{k+1},x^{*}-x^{k+1}\big{\rangle}\leq\frac{\gamma_{1}^{k}+\gamma_{2}^{k}}{1-2\gamma_{2}^{k}}\|z^{k}-x^{k}\|^{2}+\frac{\gamma_{3}^{k}-\gamma_{2}^{k}}{1-2\gamma_{2}^{k}}\|x^{k+1}-x^{k}\|^{2}.$ (22) Taking into account that $\langle x^{k}-x^{k+1},x^{k+1}-x^{*}\rangle=\langle x^{k}-x^{k+1},x^{k}-x^{*}\rangle-\|x^{k}-x^{k+1}\|^{2}$ and $z^{k}-x^{k}=-\alpha\nabla f(x^{k})$, the combination of (21) and (22) yields $\big{\langle}\nabla f(x^{k}),x^{k+1}-x^{*}\big{\rangle}\leq\frac{1}{\alpha}\big{\langle}x^{k}-x^{k+1},x^{k}-x^{*}\big{\rangle}-\frac{1}{\alpha}\left(\frac{1-\gamma_{3}^{k}-\gamma_{2}^{k}}{1-2\gamma_{2}^{k}}\right)\|x^{k+1}-x^{k}\|^{2}\\\ +\frac{\gamma_{1}^{k}+\gamma_{2}^{k}}{1-2\gamma_{2}^{k}}\alpha\|\nabla f(x^{k})\|^{2}.$ Therefore, considering that $0\leq\gamma_{2}^{k}\leq\bar{\gamma_{2}}$ and $0\leq\gamma_{3}^{k}\leq\bar{\gamma}$, the desired inequality follows by using the first condition in (13) and (3.2.2). ∎ To proceed, we assume that the sequence $(x^{k})_{k\in\mathbb{N}}$ converges to a point $x^{*}\in\Omega^{*}$. Moreover, to establish the iteration- complexity bound for $(f(x^{k}))_{k\in\mathbb{N}}$, we also take $\gamma_{1}^{k}=\gamma_{2}^{k}=0,\qquad\forall~{}{k\in\mathbb{N}}.$ (23) ###### Theorem 15. Assume that $f$ is $\mu-$strongly convex on $\mathbb{R}^{n}$. Then, the following inequality holds $\|x^{k+1}-x^{*}\|^{2}\leq\left(1-\alpha\mu\right)\|x^{k}-x^{*}\|^{2}.$ (24) ###### Proof. Using Lemma 14, considering (23) and $f^{*}\leq f(x^{k})$ for all ${k\in\mathbb{N}}$, we have that $-2\big{\langle}x^{k}-x^{k+1},x^{k}-x^{*}\big{\rangle}\leq-\alpha\mu\|x^{k}-x^{*}\|^{2}-2\alpha\nu\|x^{k+1}-x^{k}\|^{2}$. Therefore, since $1-2\alpha\nu<0$ and taking into account that $\|x^{k+1}-x^{*}\|^{2}=\|x^{k}-x^{*}\|^{2}+\|x^{k+1}-x^{k}\|^{2}-2\big{\langle}x^{k}-x^{k+1},x^{k}-x^{*}\big{\rangle}$, we obtain (24). ∎ ###### Remark 3. Letting $\alpha=1/L$, (24) yields $\|x^{k+1}-x^{*}\|^{2}\leq\left(1-\mu/L\right)^{k+1}\|x^{0}-x^{*}\|^{2}$, which is closed related to [11, Theorem 3.10]. See also [40, Theorem 2.1.15], for the unconstrained case. ## 4 GInexPM employing Armijo’s step size rule The aim of this section is to present the GInexPM for solving problem (1) employing Armijo’s search. Our method is an inexact version of the projected gradient method proposed in [28], see also [4]. Let us remind the iteration of the projected gradient method: If the current iterate $x^{k}$ is a non- stationary point of problem (1), then set $z^{k}=x^{k}-\alpha_{k}\nabla f(x^{k})$, compute $w^{k}={\cal P}_{C}(z^{k})$ and define the next iterate as $x^{k+1}=x^{k}+\tau_{k}(w^{k}-x^{k})$, where ${\cal P}_{C}$ is the exact projection operator on $C$, $\alpha_{k}$ and $\tau_{k}$ are suitable positive constants. In this scheme, $d^{k}=w^{k}-x^{k}$ is a feasible descent direction for $f$ at $x^{k}$. Thus, an Armijo’s search is employed to compute $\tau_{k}$ so that it decreases the function $f$ at $x^{k+1}$. In the same way as in Algorithm 1, we propose to compute a feasible inexact projection instead of calculating the exact one. Hence, for guarantee that the feasible direction is also a descent direction, we need to use $\varphi_{\gamma}:{\mathbb{R}}^{n}\times{\mathbb{R}}^{n}\to{\mathbb{R}}_{+}$ satisfying $\varphi_{\gamma_{3}}(u,w)\leq\gamma_{3}\|w-u\|^{2},\qquad\forall~{}u,w\in{\mathbb{R}}^{n},$ as the error tolerance function, i.e., we take $\gamma_{1}=\gamma_{2}=0$ in Definition 2. Hence, the inexact projection ${\cal P}_{C}(\varphi_{\gamma},u,v)$ onto $C$ of $v\in{\mathbb{R}}^{n}$ relative to $u\in C$ with error tolerance $\varphi_{\gamma}(u,\cdot)$ becomes ${\cal P}_{C}(\varphi_{\gamma},u,v):=\left\\{w\in C:~{}\big{\langle}v-w,y-w\big{\rangle}\leq\varphi_{\gamma_{3}}(u,w),\quad\forall~{}y\in C\right\\}.$ (25) Also, we assume that the mapping $(\gamma_{3},u,w)\mapsto\varphi_{\gamma_{3}}(u,w)$ is continuous. In this case, the gradient algorithm with inexact projection employing Armijo’s step size rule is formally defined as follows. Algorithm 2 GInexPM employing Armijo search Step 0: Choose $\sigma\in(0,1)$, $\tau\in(0,1)$ and $0<\alpha_{\min}\leq\alpha_{\max}$. Let $x^{0}\in C$ and set $k=0$. Step 1: Choose an error tolerance function $\varphi_{\gamma}$, real numbers $\alpha_{k}$ and $\gamma_{3}^{k}$ such that $\alpha_{\min}\leq\alpha_{k}\leq\alpha_{\max},\qquad\qquad 0\leq\gamma_{3}^{k}\leq\bar{\gamma}<\frac{1}{2},$ (26) and take $w^{k}$ as any feasible inexact projection of $z^{k}:=x^{k}-\alpha_{k}\nabla f(x^{k})$ onto $C$ relative to $x^{k}$ with error tolerance $\varphi_{\gamma_{3}^{k}}(x^{k},w^{k})$, i.e., $w^{k}\in{\cal P}_{C}\left(\varphi_{\gamma_{3}^{k}},x^{k},z^{k}\right).$ (27) If $w^{k}=x^{k}$, then stop; otherwise, set $\tau_{k}:=\tau^{j_{k}}$, where $j_{k}:=\min\left\\{j\in\mathbb{N}:~{}f\big{(}x^{k}+\tau^{j}(w^{k}-x^{k})\big{)}\leq f(x^{k})+\sigma\tau^{j}\big{\langle}\nabla f(x^{k}),w^{k}-x^{k}\big{\rangle}\right\\},$ (28) and set the next iterate $x^{k+1}$ as $x^{k+1}=x^{k}+\tau_{k}(w^{k}-x^{k}).$ (29) Step 2: Set $k\leftarrow k+1$, and go to Step 1. Let us describe the main features of Algorithm 2. In Step 1, we check if $w^{k}=x^{k}$. In this case, as we will show, the current iterate $x^{k}$ is a solution of problem (1), otherwise, we choose $\alpha_{k}$ such that $\alpha_{\min}\leq\alpha_{k}\leq\alpha_{\max}$. Then, by using some inner procedure, we compute $w^{k}$ as any feasible inexact projection of $z^{k}=x_{k}-\alpha_{k}\nabla f(x_{k})$ onto the feasible set $C$ relative to $x^{k}$, i.e., $w^{k}\in{\cal P}_{C}(\varphi_{\gamma_{3}^{k}},x^{k},z^{k})$. Recall that, if $\gamma_{3}^{k}=0$, then ${\cal P}_{C}(\varphi_{0},x^{k},z^{k})$ is the exact projection, see Remark 1. Therefore, Algorithm 2 can be seen as inexact version of the algorithm considered in [4, 28]. In the remainder of this section, we study the asymptotic properties and iteration-complexity bounds related to Algorithm 2. We begin by presenting some import properties of the inexact projection (25). ###### Lemma 16. Let $x\in C$, $\alpha>0$, and $0\leq\gamma_{3}\leq\bar{\gamma}<1/2$. Take $w(\alpha)$ as any feasible inexact projection of $z(\alpha)=x-\alpha\nabla f(x)$ onto $C$ relative to $x$ with error tolerance $\varphi_{\gamma_{3}}(x,w(\alpha))$, i.e., $w(\alpha)\in{\cal P}_{C}(\varphi_{\gamma_{3}},x,z(\alpha))$. Then, there hold: * (i) $\big{\langle}\nabla f(x),w(\alpha)-x\big{\rangle}\leq\left(\dfrac{\gamma_{3}-1}{\alpha}\right)\|w(\alpha)-x\|^{2}$; * (ii) the point $x$ is stationary for problem (1) if and only if $x\in{\cal P}_{C}(\varphi_{\gamma_{3}},x,z(\alpha))$; * (iii) if $x$ is a nonstationary point for problem (1), then $\big{\langle}\nabla f(x),w(\alpha)-x\big{\rangle}<0$. Equivalently, if there exists ${\bar{\alpha}}>0$ such that $\big{\langle}\nabla f(x),w({\bar{\alpha}})-x\big{\rangle}\geq 0$, then $x$ is stationary for problem (1). ###### Proof. Since $w(\alpha)\in{\cal P}_{C}(\varphi_{\gamma_{3}},x,z(\alpha))$, applying item $(b)$ of Lemma 3 with $\gamma_{1}=\gamma_{2}=0$, $w=w(\alpha)$, $v=z(\alpha)$, $y=x$, and $u=x$, we obtain $\big{\langle}x-\alpha\nabla f(x)-w(\alpha),x-w(\alpha)\big{\rangle}\leq\gamma_{3}\|w(\alpha)-x\|^{2}$ which, after some algebraic manipulations, yields the inequality of item $(i)$. For proving item $(ii)$, we first assume that $x$ is stationary for problem (1). In this case, (4) implies that $\big{\langle}\nabla f(x),w(\alpha)-x\big{\rangle}\geq 0$. Thus, considering that $\alpha>0$ and $0\leq\gamma_{3}\leq\bar{\gamma}<1/2$, the last inequality together item $(i)$ implies that $w(\alpha)=x$. Therefore, $x\in{\cal P}_{C}(\varphi_{\gamma_{3}},x,z(\alpha))$. Reciprocally, if $x\in{\cal P}_{C}(\varphi_{\gamma_{3}},x,z(\alpha))$ then applying item $(b)$ of Lemma (3) with $\gamma_{1}=\gamma_{2}=0$, $w=x$, $v=z(\alpha)$, and $u=x$, we obtain $\big{\langle}x-\alpha\nabla f(x)-x,y-x\big{\rangle}\leq 0$, for all $y\in C$. Considering that $\alpha>0$, the last inequality is equivalently to $\big{\langle}\nabla f(x),y-x\big{\rangle}\geq 0$, for all $y\in C$. Thus, according to (4), we conclude that $x$ is stationary point for problem (1). Finally, for prove item $(iii)$, take $x$ a nonstationary point for problem (1). Thus item $(ii)$ implies that $x\notin{\cal P}_{C}(\varphi_{\gamma_{3}},x,z(\alpha))$ and taking into account that $w(\alpha)\in{\cal P}_{C}(\varphi_{\gamma_{3}},x,z(\alpha))$, we conclude that $x\neq w(\alpha)$. Therefore, due to $\alpha>0$ and $0<\gamma_{3}\leq\bar{\gamma}$, it follows from item $(i)$ that $\big{\langle}\nabla f(x),w(\alpha)-x\big{\rangle}<0$ and the first sentence is proved. Finally, note that the second statement is the contrapositive of the first sentence. ∎ The next result follows from item $(iii)$ of Lemma 16 and its proof will be omitted. ###### Proposition 17. Let $\sigma\in(0,1)$, $x\in C$, $\alpha>0$, and $0\leq\lambda<{\bar{\lambda}}$. Take $w(\alpha)$ as any feasible inexact projection of $z(\alpha)=x-\alpha\nabla f(x)$ onto $C$ relative to $x$ with error tolerance $\varphi_{\lambda}(x,w(\alpha))$, i.e., $w(\alpha)\in{\cal P}_{C}(\varphi_{\lambda},x,z(\alpha))$. If $x$ is a nonstationary point for problem (1), then there exists $\delta>0$ such that $f\big{(}x+\zeta[w(\alpha)-x]\big{)}<f(x)+\sigma\zeta\big{\langle}\nabla f(x),w(\alpha)-x\big{\rangle}$, for all $\zeta\in(0,\delta)$. In the following, we establish the well definition of Algorithm 2. ###### Proposition 18. The sequence $(x^{k})_{k\in\mathbb{N}}$ generated by Algorithm 2 is well defined and belongs to $C$. ###### Proof. Proceeding by induction, let $x^{0}\in C$, $\alpha_{\min}\leq\alpha_{0}\leq\alpha_{\max}$ and $0\leq\gamma_{3}^{0}<\bar{\gamma}$. Set $z^{0}=x^{0}-\alpha_{0}\nabla f(x^{0})$. Since $C$ is closed and convex, it follows from Remark 1 that ${\cal P}_{C}(\varphi_{\gamma_{3}^{0}},x^{0},z^{0})\neq\varnothing$. Thus, we can take $w^{0}\in P_{C}(\varphi_{\gamma_{3}^{k}},x^{0},z^{0})$. If Algorithm 2 does not stop, i.e., $w^{0}\neq x^{0}$, then it follows from item $(i)$ of Lemma 16 that $\langle\nabla f(x^{0}),w^{0}-x^{0}\rangle<0$. In this case, Propoposition 17 implies that it is possible to compute $\tau_{0}\in(0,1]$ satisfying (28), for $k=0$. Therefore, $x^{1}=x^{0}+\tau_{0}(w^{0}-x^{0})$ in (29) is well defined and, considering that $x^{0},w^{0}\in C$ and $\tau_{0}\in(0,1]$, we have $x^{1}\in C$. The induction step is completely analogous, implying that $(x^{k})_{k\in\mathbb{N}}$ is well defined and belongs to $C$. ∎ ### 4.1 Asymptotic convergence analysis The aim of this section is to study asymptotic convergence properties related to Algorithm 2. We begin by presenting a partial convergence result of the sequence $(x^{k})_{k\in\mathbb{N}}$ generated by Algorithm 2. ###### Proposition 19. Algorithm 2 finishes in a finite number of iterations at a stationary point of problem (1), or generates an infinite sequence $(x^{k})_{k\in\mathbb{N}}$ for which $\left(f(x^{k})\right)_{k\in\mathbb{N}}$ is a decreasing sequence and every cluster point of $(x^{k})_{k\in\mathbb{N}}$ is stationary for problem (1). ###### Proof. First we assume that $(x^{k})_{k\in\mathbb{N}}$ is finite. In this case, according to Step 1, there exists $k\in\mathbb{N}$ such that $x^{k}=w^{k}\in{\cal P}_{C}(\varphi_{\gamma_{3}^{k}},x^{k},z^{k})$, where $z^{k}=x^{k}-\alpha_{k}\nabla f(x^{k})$, $0\leq\gamma_{3}^{k}\leq\bar{\gamma}$ and $\alpha_{k}>0$. Therefore, applying the second statement of item $(ii)$ of Lemma 16 with $x=x^{k}$, $\alpha=\alpha_{k}$, and $\gamma_{3}=\gamma_{3}^{k}$, we conclude that $x^{k}$ is stationary for problem (1). Now, we assume that $(x^{k})_{k\in\mathbb{N}}$ is infinite. Thus, according to Step 1, $x^{k}\neq w^{k}$ for all $k=0,1,\ldots$. Consequently, applying item $(ii)$ of Lemma 16 with $x=x^{k}$, $\alpha=\alpha_{k}$, and $\gamma_{3}=\gamma_{3}^{k}$, we have that $x^{k}$ is nonstationary for problem (1). Hence, item $(iii)$ of Lemma 16 implies that $\big{\langle}\nabla f(x^{k}),w^{k}-x^{k}\big{\rangle}<0$, for all $k=0,1,\ldots$. Therefore, it follows from (28) and (29) that $0<-\sigma\tau_{k}\big{\langle}\nabla f(x^{k}),w^{k}-x^{k}\big{\rangle}\leq f(x^{k})-f(x^{k+1}),\qquad\forall~{}k\in\mathbb{N},$ (30) with implies that $f(x^{k+1})<f(x^{k})$, for all $k=0,1,\ldots$, and then $\left(f(x^{k})\right)_{k\in\mathbb{N}}$ is a decreasing sequence. Let ${\bar{x}}$ be a cluster point of $(x^{k})_{k\in\mathbb{N}}$ and $(x^{k_{j}})_{j\in\mathbb{N}}$ a subsequence of $(x^{k})_{k\in\mathbb{N}}$ such that $\lim_{j\to+\infty}x^{k_{j}}=\bar{x}$. Since $C$ is closed, by Proposition 18, we have $\bar{x}\in C$. Since $\left(f(x^{k})\right)_{k\in\mathbb{N}}$ is decreasing and $\lim_{j\to+\infty}f(x^{k_{j}})=f(\bar{x})$, we conclude that $\lim_{k\to+\infty}f(x^{k})=f(\bar{x})$. On the order hand, using the last condition in (26), we have $1/(1-2\gamma_{3}^{k})<1/(1-2\bar{\gamma})$, for all $k=0,1,\ldots$. Since $w^{k}\in{\cal P}_{C}(\varphi_{\gamma_{3}^{k}},x^{k},z^{k})$, where $z^{k}=x^{k}-\alpha_{k}\nabla f(x^{k})$, applying item $(a)$ of Lemma 3 with $x=x^{k}$, $u=x^{k}$, $v=z^{k}$, $w=w^{k}$, $\gamma_{1}=\gamma_{2}=0$, and $\gamma_{3}=\gamma_{3}^{k}$, we obtain $\|w^{k_{j}}-x^{k_{j}}\|^{2}\leq\frac{\alpha_{k_{j}}^{2}}{1-2\gamma_{3}^{k_{j}}}\|\nabla f(x^{k_{j}})\|^{2}<\frac{\alpha_{\max}^{2}}{1-2\bar{\gamma}}\|\nabla f(x^{k_{j}})\|^{2},\qquad\forall~{}j\in\mathbb{N}.$ Considering that $(x^{k_{j}})_{j\in\mathbb{N}}$ converges to ${\bar{x}}$ and $\nabla f$ is continuous, the last inequality implies that $(w^{k_{j}})_{j\in\mathbb{N}}\subset C$ is also bounded. Thus, we can assume without loss of generality that $\lim_{j\to+\infty}w^{k_{j}}=\bar{w}\in C$. Now, due to $\tau_{k}\in(0,1]$, for all $k=0,1,\ldots$, we can also assume without loss of generality that $\lim_{j\to+\infty}\tau_{k_{j}}=\bar{\tau}\in[0,1].$ Therefore, owing to $\lim_{j\to+\infty}f(x^{k})=f(\bar{x})$, taking limits in (30) along an appropriate subsequence, we obtain $\bar{\tau}\big{\langle}\nabla f(\bar{x}),\bar{w}-\bar{x}\big{\rangle}=0.$ We have two possibilities: $\bar{\tau}>0$ or $\bar{\tau}=0$. If $\bar{\tau}>0$, then $\big{\langle}\nabla f(\bar{x}),\bar{w}-\bar{x}\big{\rangle}=0.$ Now, we assume that $\lim_{j\to+\infty}\tau_{k_{j}}=\bar{\tau}=0$. In this case, for any fixed $q\in\mathbb{N}$, there exits $j$ such that $\tau_{k_{j}}<\tau^{q}$. Hence, Armijo’s condition (28) does not hold for $\tau^{q}$, i.e., $f\big{(}x^{k_{j}}+\tau^{q}(w^{k_{j}}-x^{k_{j}})\big{)}>f(x^{k_{j}})+\sigma\tau^{q}\big{\langle}\nabla f(x^{k_{j}}),w^{k_{j}}-x^{k_{j}}\big{\rangle}$, for all $j\in\mathbb{N}.$ Thus, taking limits as $j$ goes to $+\infty$, we obtain $f\big{(}\bar{x}+\tau^{q}(\bar{w}-\bar{x})\big{)}\geq f(\bar{x})+\sigma\tau^{q}\big{\langle}\nabla f(\bar{x}),\bar{w}-\bar{x}\big{\rangle},$ which is equivalent to $\frac{f\big{(}\bar{x}+\tau^{q}(\bar{w}-\bar{x})\big{)}-f(\bar{x})}{\tau^{q}}\geq\sigma\big{\langle}\nabla f(\bar{x}),\bar{w}-\bar{x}\big{\rangle}.$ Since this inequality holds for all $q\in\mathbb{N}$, taking limits as $q$ goes to $+\infty$, we conclude that $\langle\nabla f(\bar{x}),\bar{w}-\bar{x}\rangle\geq\sigma\big{\langle}\nabla f(\bar{x}),\bar{w}-\bar{x}\big{\rangle}$. Hence, due to $\sigma\in(0,1)$, we obtain $\big{\langle}\nabla f(\bar{x}),\bar{w}-\bar{x}\big{\rangle}\geq 0$. We recall that $\langle\nabla f(x^{k_{j}}),w^{k_{j}}-x^{k_{j}}\rangle<0$, for all $j=0,1,\ldots$, which taking limits as $j$ goes to $+\infty$ yields $\big{\langle}\nabla f(\bar{x}),\bar{w}-\bar{x}\big{\rangle}\leq 0$. Hence, $\big{\langle}\nabla f(\bar{x}),\bar{w}-\bar{x}\big{\rangle}=0$. Therefore, for any of two possibilities, $\bar{\tau}>0$ or $\bar{\tau}=0$, we have $\langle\nabla f(\bar{x}),\bar{w}-\bar{x}\rangle=0$. On the other hand, $w^{k_{j}}\in{\cal P}_{C}(\varphi_{\gamma_{3}^{k_{j}}},x^{k_{j}},z^{k_{j}}),$ where $z^{k_{j}}=x^{k_{j}}-\alpha_{k_{j}}\nabla f(x^{k_{j}})$, $0\leq\gamma_{3}^{k_{j}}\leq\bar{\gamma}$, and $\alpha_{k_{j}}>0$. Thus, it follows from (25) that $\big{\langle}z^{k_{j}}-w^{k_{j}},y-w^{k_{j}}\big{\rangle}\leq\varphi_{\gamma_{3}^{k_{j}}}(x^{k_{j}},w^{k_{j}}),\qquad y\in C,\quad\forall~{}j\in\mathbb{N}.$ (31) Moreover, since $\alpha_{k}\in[\alpha_{\min},\alpha_{\max}]$, for all $k=0,1,\ldots$, we also assume without loss of generality that $\lim_{j\to+\infty}\alpha_{k_{j}}=\bar{\alpha}\in[\alpha_{\min},\alpha_{\max}]$. Thus, taking limits in (31) and considering that the mapping $(\gamma_{3},u,w)\mapsto\varphi_{\gamma_{3}}(u,w)$ is continuous, $\lim_{j\to+\infty}x^{k_{j}}=\bar{x}\in C$, $\lim_{j\to+\infty}w^{k_{j}}=\bar{w}\in C$, and $\lim_{j\to+\infty}\tau_{k_{j}}=\bar{\tau}\in[0,1]$, we conclude that $\big{\langle}\bar{z}-\bar{w},y-\bar{w}\big{\rangle}\leq\varphi_{\bar{\gamma}}(\bar{x},\bar{w})$, for all $y\in C$, where $\bar{z}=\bar{x}-{\bar{\alpha}}\nabla f(\bar{x})$. Hence, it follows from (25) that $\bar{w}\in{\cal P}_{C}\left(\varphi_{\bar{\gamma}},{\bar{x}},{\bar{z}}\right)$, where $\bar{z}=\bar{x}-{\bar{\alpha}}\nabla f(\bar{x})$. Therefore, due to $\big{\langle}\nabla f(\bar{x}),\bar{w}-\bar{x}\big{\rangle}=0$, we can apply the second sentence in item $(iii)$ of Lemma 16 with $x=\bar{x}$, $z({\bar{\alpha}})=\bar{z}$, and $w({\bar{\alpha}})=\bar{w}$ to conclude that $\bar{x}$ is stationary for problem (1). ∎ Due to Proposition 19, from now on we assume that the sequence $(x^{k})_{k\in\mathbb{N}}$ generated by Algorithm 2 is infinite. The following result establishes a basic inequality satisfied by the iterates of Algorithm 2, which will be used to study its convergence properties. For simplify the notations we define the following constant: $\xi:=\dfrac{2\alpha_{\max}}{\sigma}>0.$ (32) ###### Lemma 20. For each $x\in C$, there holds $\|x^{k+1}-x\|^{2}\leq\|x^{k}-x\|^{2}+2\alpha_{k}\tau_{k}\big{\langle}\nabla f(x^{k}),x-x^{k}\big{\rangle}+\xi\left[f(x^{k})-f(x^{k+1})\right],\quad\forall~{}k\in\mathbb{N}.$ (33) ###### Proof. We know that $\|x^{k+1}-x\|^{2}=\|x^{k}-x\|^{2}+\|x^{k+1}-x^{k}\|^{2}-2\big{\langle}x^{k+1}-x^{k},x-x^{k}\big{\rangle}$, for all $x\in C$ and $k=0,1,\ldots$. Thus, using (29), we have $\|x^{k+1}-x\|^{2}=\|x^{k}-x\|^{2}+\tau_{k}^{2}\|w^{k}-x^{k}\|^{2}-2\tau_{k}\big{\langle}w^{k}-x^{k},x-x^{k}\big{\rangle},\qquad\forall~{}k\in\mathbb{N}.$ (34) On the other hand, by using (27) we have $w^{k}\in{\cal P}_{C}(\varphi_{\gamma_{3}^{k}},x^{k},z^{k})$ with $z^{k}=x^{k}-\alpha_{k}\nabla f(x^{k})$. Thus, applying item $(b)$ of Lemma 3 with $y=x$, $u=x^{k}$, $v=z^{k}$, $w=w^{k}$, $\gamma_{1}=\gamma_{2}=0$, $\gamma_{3}=\gamma_{3}^{k}$, and $\varphi_{\gamma_{3}}=\varphi_{\gamma_{3}^{k}}$, we obtain $\langle x^{k}-\alpha_{k}\nabla f(x^{k})-w^{k},x-w^{k}\rangle\leq\gamma_{3}^{k}\|w^{k}-x^{k}\|^{2}$, for all $k\in\mathbb{N}$. After some algebraic manipulations in the last inequality, we have $\big{\langle}w^{k}-x^{k},x-x^{k}\big{\rangle}\geq\alpha_{k}\big{\langle}\nabla f(x^{k}),w^{k}-x\big{\rangle}+(1-\gamma_{3}^{k})\|w^{k}-x^{k}\|^{2}.$ Combining the last inequality with (34), we conclude $\|x^{k+1}-x\|^{2}\leq\|x^{k}-x\|^{2}-\tau_{k}\big{[}2(1-\gamma_{3}^{k})-\tau_{k}\big{]}\|w^{k}-x^{k}\|^{2}+2\tau_{k}\alpha_{k}\big{\langle}\nabla f(x^{k}),x-w^{k}\big{\rangle}.$ (35) Since $0\leq\gamma_{3}^{k}<\bar{\gamma}<1/2$ and $\tau_{k}\in(0,1]$, we have $2(1-\gamma_{3}^{k})-\tau_{k}\geq 1-2\bar{\gamma}>0$. Thus, (35) becomes $\|x^{k+1}-x\|^{2}\leq\|x^{k}-x\|^{2}+2\tau_{k}\alpha_{k}\big{\langle}\nabla f(x^{k}),x-w^{k}\big{\rangle},\qquad\forall~{}k\in\mathbb{N}.$ Therefore, considering that $\big{\langle}\nabla f(x^{k}),x-w^{k}\big{\rangle}=\big{\langle}\nabla f(x^{k}),x-x^{k}\big{\rangle}+\big{\langle}\nabla f(x^{k}),x^{k}-w^{k}\big{\rangle}$ and taking into account (28), we conclude that $\|x^{k+1}-x\|^{2}\leq\|x^{k}-x\|^{2}+2\tau_{k}\alpha_{k}\big{\langle}\nabla f(x^{k}),x-x^{k}\big{\rangle}+\frac{2\alpha_{k}}{\sigma}\left[f(x^{k})-f(x^{k+1})\right],\qquad\forall~{}k\in\mathbb{N}.$ (36) Since $0<\alpha_{k}\leq\alpha_{\max}$, Proposition 19 implies $\alpha_{k}\big{[}f(x^{k})-f(x^{k+1})\big{]}<\alpha_{\max}\big{[}f(x^{k})-f(x^{k+1})\big{]}$. Therefore, the desired inequality (33) follows from (36) by using (32). ∎ For the sequence $(x^{k})_{k\in\mathbb{N}}$ generated by Algorithm 2, we define the following auxiliary set: $U:=\left\\{x\in C:f(x)\leq\inf_{k}f(x^{k}),\quad{k\in\mathbb{N}}\right\\}.$ Next, we analyze the behavior of the sequence $(x^{k})_{k\in\mathbb{N}}$ when $f$ is a quasiconvex function. ###### Corollary 21. Assume that $f$ is a quasiconvex function. If $U\neq\varnothing$, then $(x^{k})_{k\in\mathbb{N}}$ converges to a stationary point of problem (1). ###### Proof. Let $x\in U$. Thus, $f(x)\leq f(x^{k})$ for all $k\in\mathbb{N}$. Since $f$ is quasiconvex, we have $\big{\langle}\nabla f(x^{k}),x-x^{k}\big{\rangle}\leq 0$, for all $k\in\mathbb{N}$. Using Lemma 20, we obtain $\|x^{k+1}-x\|^{2}\leq\|x^{k}-x\|^{2}+\xi\left[f(x^{k})-f(x^{k+1})\right],\quad\forall~{}k\in\mathbb{N}.$ Defining $\epsilon_{k}=\xi\big{[}f(x^{k})-f(x^{k+1})\big{]}$, we have $\|x^{k+1}-x\|^{2}\leq\|x^{k}-x\|^{2}+\epsilon_{k}$, for all $k\in\mathbb{N}$. On the other hand, summing $\epsilon_{k}$ with $k=0,1,\ldots,N$, we have $\sum_{k=0}^{N}\epsilon_{k}\leq\xi\big{[}f(x^{0})-f(x)\big{]}<\infty$. Thus, it follows from Definition 1 that $(x^{k})_{k\in\mathbb{N}}$ is quasi-Fejér convergent to $U$. Since $U$ is nonempty, it follows from Theorem 2 that $(x^{k})_{k\in\mathbb{N}}$ is bounded, and therefore it has a cluster point. Let $\bar{x}$ be a cluster point of $(x^{k})_{k\in\mathbb{N}}$ and $(x^{k_{j}})_{j\in\mathbb{N}}$ be a subsequence of $(x^{k})_{k\in\mathbb{N}}$ such that $\lim_{j\to\infty}x^{k_{j}}=\bar{x}$. Considering that $f$ is continuous, we have $\lim_{j\to\infty}f(x^{k_{j}})=f(\bar{x})$. Hence, since, by Proposition 19, $\left(f(x^{k})\right)_{k\in\mathbb{N}}$ is decreasing, we obtain $\inf\\{f(x^{k}):k=0,1,2,\ldots\\}=\lim_{k\to\infty}f(x^{k})=f(\bar{x}).$ Therefore, $\bar{x}\in U$. It follows from Theorem 2 that $(x^{k})_{k\in\mathbb{N}}$ converges to $\bar{x}$ and the conclusion is obtained by using again Proposition 19. ∎ The next two results are similar to Lemma 9 and Theorem 10, respectively. For completeness reasons, we have included their proofs here. ###### Lemma 22. If $f$ is a quasiconvex function and $(x^{k})_{k\in\mathbb{N}}$ has no cluster points, then $\Omega^{*}=\varnothing$, $\lim_{k\to\infty}\|x^{k}\|=\infty$, and $\lim_{k\to\infty}f(x^{k})=\inf\\{f(x):x\in C\\}$. ###### Proof. Since $(x^{k})_{k\in\mathbb{N}}$ has no cluster points, then $\lim_{k\to\infty}\|x^{k}\|=\infty$. Assume that problem (1) has an optimum, say $\tilde{x}$, so $f(\tilde{x})\leq f(x^{k})$ for all $k$. Thus, $\tilde{x}\in U$. Using Corollary 21, we obtain that $(x^{k})_{k\in\mathbb{N}}$ is convergent, contradicting that $\lim_{k\to\infty}\|x^{k}\|=\infty$. Therefore, $\Omega^{*}=\varnothing$. Now, we claim that $\lim_{k\to\infty}f(x^{k})=\inf\\{f(x):x\in C\\}$. If $\lim_{k\to\infty}f(x^{k})=-\infty$, the claim holds. Let $f^{*}=\inf_{x\in C}f(x)$. By contradiction, suppose that $\lim_{k\to\infty}f(x^{k})>f^{*}$. Then, there exists $\tilde{x}\in C$ such that $f(\tilde{x})\leq f(x^{k})$ for all $k$. Using Corollary 21, we have that $(x^{k})_{k\in\mathbb{N}}$ is convergent, contradicting again $\lim_{k\to\infty}\|x^{k}\|=\infty$, which concludes the proof. ∎ ###### Theorem 23. Assume that $f$ is a pseudoconvex function. Then, $\Omega^{*}\neq\varnothing$ if and only if $(x^{k})_{k\in\mathbb{N}}$ has at least one cluster point. Moreover, $(x^{k})_{k\in\mathbb{N}}$ converges to an optimum point if $\Omega^{*}\neq\varnothing$; otherwise, $\lim_{k\to\infty}\|x^{k}\|=\infty$ and $\lim_{k\to\infty}f(x^{k})=\inf\\{f(x):x\in C\\}$. ###### Proof. Recall that pseudoconvex functions are also quasiconvex. Assume that $\Omega^{*}\neq\varnothing$. In this case, $U\neq\varnothing$. Using Corollary 21, we conclude that $(x^{k})_{k\in\mathbb{N}}$ converges to a stationary point of problem (1). Reciprocally, let $\bar{x}$ be a cluster point of $(x^{k})_{k\in\mathbb{N}}$ and $(x^{k_{j}})_{j\in\mathbb{N}}$ be a subsequence of $(x^{k})_{k\in\mathbb{N}}$ such that $\lim_{j\to+\infty}x^{k_{j}}=\bar{x}$. Since, from Proposition 19, $(f(x^{k}))_{k\in\mathbb{N}}$ is monotone non- increasing, by continuity of $f$, we conclude that $\inf_{k}f(x^{k})=f(\bar{x})$ and hence $\bar{x}\in U$. Using Corollary 21, we obtain that $(x^{k})_{k\in\mathbb{N}}$ converges to a stationary point $\tilde{x}$ of problem (1). Since $f$ is pseudoconvex, this point is also an optimal solution of problem (1). The last part of the theorem follows by combining the first one with Lemma 22. ∎ ### 4.2 Iteration-complexity bound This section is devoted to study iteration-complexity bounds for the sequence generated by Algorithm 2, similar results in the multiobjetive context can be found in [19]. For that we assume that $f^{*}>-\infty$ and the objective function $f$ has Lipschitz continuous gradient with constant $L\geq 0$, i.e., we assume that $\nabla f$ satisfies (2). Moreover, we also assume that $(x^{k})_{k\in\mathbb{N}}$ generated by Algorithm 1 converges to a point $x^{*}$, i.e, $\lim_{k\to\infty}x^{k}=x^{*}$. To simplify the notations, we set $\tau_{\min}:=\min\left\\{\frac{2\tau(1-\sigma)(1-\bar{\gamma})}{\alpha_{\max}L},~{}1\right\\}.$ (37) The next lemma is a version of [19, Lemma 3.1] for our specific context. ###### Lemma 24. The step size $\tau_{k}$ in Algorithm 2 satisfies $\tau_{k}\geq\tau_{\min}$. ###### Proof. If $\tau_{k}=1$, then the result trivially holds. Thus, assume that $\tau_{k}<1$. It follows from Armijo’s condition in (28) that $f(x^{k}+\frac{\tau_{k}}{\tau}(w^{k}-x^{k}))>f(x^{k})+\sigma\frac{\tau_{k}}{\tau}\big{\langle}\nabla f(x^{k}),w^{k}-x^{k}\big{\rangle}$. Now, using Lemma 1, we have $f\left(x^{k}+\frac{\tau_{k}}{\tau}(w^{k}-x^{k})\right)\leq f(x^{k})+\frac{\tau_{k}}{\tau}\big{\langle}\nabla f(x^{k}),w^{k}-x^{k}\big{\rangle}+\frac{1}{\tau^{2}}\frac{L}{2}\tau_{k}^{2}\|w^{k}-x^{k}\|^{2}$. Hence, combining the two previous inequalities, we obtain $\tau(1-\sigma)\big{\langle}\nabla f(x^{k}),w^{k}-x^{k}\big{\rangle}+\frac{L}{2}\tau_{k}\|w^{k}-x^{k}\|^{2}>0.$ (38) On the order hand, since $w^{k}\in{\cal P}_{C}(\varphi_{\gamma_{3}^{k}},x^{k},z^{k})$, where $z^{k}=x^{k}-\alpha_{k}\nabla f(x^{k})$, applying item $(i)$ of Lemma 16 with $x=x^{k}$, $w(\alpha)=w^{k}$, $z=z^{k}$, $\gamma_{1}=\gamma_{2}=0$, $\gamma_{3}=\gamma_{3}^{k}$, and $\varphi_{\gamma}=\varphi_{\gamma^{k}}$, we have $\big{\langle}\nabla f(x^{k}),w^{k}-x^{k}\big{\rangle}\leq\left(\frac{\gamma_{3}^{k}-1}{\alpha_{k}}\right)\|w^{k}-x^{k}\|^{2}.$ Combining the last inequality with (38) yields $\left[{\tau(1-\sigma)(\gamma_{3}^{k}-1)}/{\alpha_{k}}+{L}\tau_{k}/2\right]\|w^{k}-x^{k}\|^{2}>0.$ Hence, using (26), it follows that $\tau_{k}>\frac{2\tau(1-\sigma)(1-\gamma_{3}^{k})}{\alpha_{k}L}\geq\frac{2\tau(1-\sigma)(1-\bar{\gamma})}{\alpha_{\max}L}.$ Therefore, since $\tau_{k}$ is never larger than one, the result follows and the proof is concluded. ∎ It follows from item $(ii)$ of Lemma 16 that if $x^{k}\in{\cal P}_{C}(\varphi_{\gamma_{3}},x^{k},z^{k})$, then the point $x^{k}$ is stationary for problem (1). Since $w^{k}\in{\cal P}_{C}(\varphi_{\gamma_{3}},x^{k},z^{k})$, the quantity $\|w^{k}-x^{k}\|$ can be seen as a measure of stationarity of $x^{k}$. Next theorem presents an iteration-complexity bound for this quantity, see a similar result in [19, Theorem 3.1]. ###### Theorem 25. Let $\tau_{\min}$ be defined in (37). Then, for every $N\in\mathbb{N}$, the following inequality holds $\min\left\\{\|w^{k}-x^{k}\|:~{}k=0,1\ldots,N-1\right\\}\leq\sqrt{\frac{\alpha_{\max}\left[f(x^{0})-f^{*}\right]}{\sigma\tau_{\min}{\left(1-\bar{\gamma}\right)}}}\frac{1}{\sqrt{N}}.$ ###### Proof. From the definition of $\tau_{k}$ and condition (28), we have $f(x^{k+1})-f(x^{k})\leq\sigma\tau_{k}\big{\langle}\nabla f(x^{k}),w^{k}-x^{k}\big{\rangle}.$ (39) Since $w^{k}\in{\cal P}_{C}(\varphi_{\gamma_{3}^{k}},x^{k},z^{k})$, where $z^{k}=x^{k}-\alpha_{k}\nabla f(x^{k})$, applying item $(i)$ of Lemma 16 with $x=x^{k}$, $w(\alpha)=w^{k}$, $z=z^{k}$, $\gamma_{1}=\gamma_{2}=0$, $\gamma_{3}=\gamma_{3}^{k}$, and $\varphi_{\gamma}=\varphi_{\gamma^{k}}$, we obtain $\big{\langle}\nabla f(x^{k}),w^{k}-x^{k}\big{\rangle}\leq\left(\frac{\gamma_{3}^{k}-1}{\alpha_{k}}\right)\|w^{k}-x^{k}\|^{2}.$ (40) By (26), we have $(1-\gamma_{3}^{k})/\alpha_{k}\geq(1-\bar{\gamma})/\alpha_{\max}$. Thus, combining (39) with (40) and taking into account Lemma 24, it follows that $f(x^{k})-f(x^{k+1})\geq\sigma\tau_{k}\left(\frac{1-\bar{\gamma}}{\alpha_{\max}}\right)\|w^{k}-x^{k}\|^{2}\geq\sigma\tau_{\min}\left(\frac{1-\bar{\gamma}}{\alpha_{\max}}\right)\|w^{k}-x^{k}\|^{2}.$ Hence, performing the sum of the above inequality for $k=0,1,\ldots,N-1$, we have $\sum_{k=0}^{N-1}\|w^{k}-x^{k}\|^{2}\leq\frac{\alpha_{\max}\left[f(x^{0})-f(x^{N})\right]}{\sigma\tau_{\min}(1-\bar{\gamma})}\leq\frac{\alpha_{\max}\left[f(x^{0})-f^{*}\right]}{\sigma\tau_{\min}(1-\bar{\gamma})},$ which implies the desired inequality. ∎ #### 4.2.1 Iteration-complexity bound under convexity In this section, we present an iteration-complexity bound for the sequence $\left(f(x^{k})\right)_{k\in\mathbb{N}}$ when $f$ is convex. ###### Theorem 26. Let $f$ be a convex function on $C$. Then, for every $N\in\mathbb{N}$, there holds $\min\left\\{f(x^{k})-f^{*}:~{}k=0,1\ldots,N-1\right\\}\leq\frac{\|x^{0}-x^{*}\|+\xi\left[f(x^{0})-f^{*}\right]}{2\alpha_{\min}\tau_{\min}}\frac{1}{N}.$ ###### Proof. Using the first inequality in (26) and Lemma 24, we have $2\alpha_{\min}\tau_{\min}\leq 2\alpha_{k}\tau_{k}$, for all $k\in{\mathbb{N}}$. From the convexity of $f$, we have $\big{\langle}\nabla f(x^{k}),x^{*}-x^{k}\big{\rangle}\leq f^{*}-f(x^{k})$, for all $k\in{\mathbb{N}}$. Thus, applying Lemma 20 with $x=x^{*}$, after some algebraic manipulations, we conclude $2\alpha_{\min}\tau_{\min}\left[f(x^{k})-f^{*}\right]\leq\|x^{k}-x^{*}\|^{2}-\|x^{k+1}-x^{*}\|^{2}+\xi\left[f(x^{k})-f(x^{k+1})\right]\quad k=0,1,\ldots.$ Hence, performing the sum of the above inequality for $k=0,1,\ldots,N-1$, we obtain $2\alpha_{\min}\tau_{\min}\sum_{k=0}^{N-1}\left[f(x^{k})-f^{*}\right]\leq\|x^{0}-x^{*}\|^{2}-\|x^{N}-x^{*}\|^{2}+\xi\left[f(x^{0})-f(x^{N})\right].$ Therefore, $2\alpha_{\min}\tau_{\min}N\min\\{f(x^{k})-f^{*}:k=0,1\ldots,N-1\\}\leq\|x^{0}-x^{*}\|+\xi\left[f(x^{0})-f(x^{N})\right]$, which implies the desired inequality. ∎ ## 5 Numerical experiments In this section, we summing up the results of our preliminary numerical experiments in order to verify the practical behavior of the proposed algorithms. In particular, we will illustrate the potential advantages of considering inexact projections instead of exact ones in a problem of least squares over the spectrahedron. The codes are written in Matlab and are freely available at https://orizon.ime.ufg.br. All experiments were run on a macOS 10.15.7 with 3.7GHz Intel Core i5 processor and 8GB of RAM. Let $\mathbb{S}^{n}$ be the space of $n\times n$ symmetric real matrices and $\mathbb{S}^{n}_{+}$ be the cone of positive semidefinite matrices in $\mathbb{S}^{n}$. Given $A$ and $B$ two $n\times m$ matrices, with $m\geq n$, we consider the following problem: $\begin{array}[]{cl}\displaystyle\min_{X\in\mathbb{S}^{n}}&f(X):=\displaystyle\frac{1}{2}\|AX-B\|^{2}_{F}\\\ \mbox{s.t.}&\textrm{tr}(X)=1,\\\ &X\in\mathbb{S}^{n}_{+},\\\ \end{array}$ (41) where $X$ is the $n\times n$ matrix that we seek to find. Here, $\|\cdot\|_{F}$ denotes the Frobenius matrix norm $\|A\|_{F}=\sqrt{\langle A,A\rangle}$, where the inner product is given by $\langle A,B\rangle=\textrm{tr}(A^{T}B)$. Problem (41) and its variants appear in applications in different areas such as statistics, physics and economics [15, 18, 26, 50], and were considered, for example, in the numerical tests of [8, 24, 31, 32]. Following we briefly discuss how to compute projections onto the feasible region of (41). Define $C=\\{X\in\mathbb{R}^{n\times n}\mid\textrm{tr}(X)=1,\;X\in\mathbb{S}^{n}_{+}\\}$. Since $\mathbb{S}^{n}_{+}$ is a convex and closed set, $C$ is convex and compact. Formally the problem of projection a given vector $V\in\mathbb{R}^{n\times n}$ onto $C$ is stated as $\begin{array}[]{cl}\displaystyle\min_{W\in\mathbb{S}^{n}}&\displaystyle\frac{1}{2}\|W-V\|_{F}^{2}\\\ \mbox{s.t.}&W\in C.\\\ \end{array}$ (42) Since $\|W-V\|_{F}^{2}=\|W-V_{S}\|_{F}^{2}+\|V_{A}\|^{2}_{F}$ for any $W\in\mathbb{S}^{n}$, where $V_{S}$ and $V_{A}$ denote the symmetric and the antisymmetric part of $V$, respectively, it can be assumed, without loss of generality, that $V\in\mathbb{S}^{n}$. Let $W^{*}$ be the unique solution of (42). Given the eigen-decomposition $V=QDQ^{T}$, it is well known that $W^{*}=QP_{\Delta_{n}}(D)Q^{T}$, where $P_{\Delta_{n}}(D)$ denotes the diagonal matrix obtained by projecting the diagonal elements of $D$ onto the $n$-dimensional simplex $\Delta_{n}=\\{x\in\mathbb{R}^{n}\mid x_{1}+\ldots+x_{n}=1;\;x\geq 0\\}$, see, for example, [26]. This means that computing $W^{*}$ requires a priori the full eigen-decomposition of $V$, which can be computationally prohibitive for high-dimensional problems. This drawback will appear clearly in the results reported in section 5.2. Inexact projections can be obtained by adding the constraint $\textrm{rank}(W)\leq p$, for a given $1\leq p<n$, to Problem (42). Denoting by $W_{p}$ the solution of this latter problem, we have $W_{p}=\sum_{i=1}^{p}\lambda_{i}q_{i}q_{i}^{T}$, where the scalars $(\lambda_{1},\ldots,\lambda_{p})\in\mathbb{R}^{p}$ are obtained by projecting the $p$ largest eigenvalues of $V$ onto the $p$-dimensional simplex $\Delta_{p}$ and $q_{i}\in\mathbb{R}^{n}$ are the corresponding unit eigenvectors for all $i=1,\ldots,p$, see [2]. Therefore, inexact projections can be computed by means of an incomplete eigen- decomposition of $V$, resulting in computational savings. Note that if $\textrm{rank}(W^{*})\leq p$, then $W_{p}$ coincides with $W^{*}$. As far as we know, this approach was first proposed in [2] and was also used in [24, 23]. We implemented the inexact projection scheme discussed above by choosing parameter $p$ in an adaptive way and introducing a suitable error criterion. Let us formally describe the adopted scheme to find an inexact solution $\tilde{W}$ to Problem (42) relative to $U\in C$ with error tolerance mapping $\varphi_{\gamma}$ and real numbers $\gamma_{1}$, $\gamma_{2}$ and $\gamma_{3}$ satisfying the suitable conditions (10) or (26), depending on the main algorithm. Algorithm 3 Procedure to compute $\tilde{W}\in{\cal P}_{C}\left(\varphi_{\gamma},U,V\right)$ Input: Let $1\leq p<n$ be given. Step 1: Compute $(\lambda_{i}^{V},q_{i})_{i=1}^{p}$ (with $\|q_{i}\|=1$, $i=1,\ldots,p$) the $p$ largest eigenpairs of $V$, then set $W_{p}:=\sum_{i=1}^{p}\lambda_{i}q_{i}q_{i}^{T},$ where $(\lambda_{1},\ldots,\lambda_{p})\in\mathbb{R}^{p}$ are obtained by projecting $(\lambda_{1}^{V},\ldots,\lambda_{p}^{V})\in\mathbb{R}^{p}$ onto the $p$-dimensional simplex $\Delta_{p}$. Step 2: Compute $Y_{p}:=\displaystyle\arg\min_{Y\in C}\,\langle W_{p}-V,Y-W_{p}\rangle$. Step 3: If $\langle W_{p}-V,Y_{p}-W_{p}\rangle\geq-\varphi_{\gamma}(U,V,W_{p})$, then set $\tilde{W}:=W_{p}$ and return to the main algorithm. Step 4: Set $p\leftarrow p+1$ and go to Step 1. Output: $\tilde{W}:=W_{p}$. Some comments regarding Algorithm 3 are in order. First, in Step 1 we compute the rank-$p$ projection of $V$ onto $C$. The computational cost os this step is dominated by the cost of computing the $p$ leading eigenpairs of $V$, since the projection of $(\lambda_{1}^{V},\ldots,\lambda_{p}^{V})\in\mathbb{R}^{p}$ onto the $p$-dimensional simplex $\Delta_{p}$ can be easily done in $O(p\log p)$ time, see, for example, [2]. Second, the subproblem in Step 2 is solved by computing the largest eigenpair of $V-W_{p}$. Indeed, $Y_{p}=qq^{T}$, where $q\in\mathbb{R}^{n}$ is the unit eigenvector corresponding to the largest eigenvalue of $V-W_{p}$, see, also, [2]. In our implementations, we used the Matlab function eigs to compute eigenvalues/eigenvectors [47, 34]. Third, in Step 3 if the stopping criterion $\langle W_{p}-V,Y_{p}-W_{p}\rangle\geq-\varphi_{\gamma}(U,V,W_{p})$ is satisfied, then from Definition 2, we conclude that $\tilde{W}=W_{p}\in{\cal P}_{C}\left(\varphi_{\gamma},U,V\right)$, i.e., the output is a feasible inexact projection of $V\in\mathbb{S}^{n}$ relative to $U\in C$. Otherwise, we increase parameter $p$ and proceed to calculate a more accurate eigen- decomposition of $V$. Fourth, in the first iteration of the main algorithm, we set $p=1$ as the input for Algorithm 3. For the subsequent iterations, we used, in principle, the success value for $p$ from the previous outer iteration. Without attempting to go into details, seeking computational savings, in some iterations, we consider decreasing the input $p$ with respect to the previous successful one. Concerning the stopping criterion of the main algorithms, all runs were stopped at an iterate $X^{k}$ declaring convergence if $\frac{\|X^{\ell}-X^{\ell-1}\|_{F}}{\|X^{\ell-1}\|_{F}}\leq 10^{-4},$ for $\ell=k$ and $\ell=k-1$. This means that we stopped the execution of the main algorithms when the above convergence metric is satisfied for two consecutive iterations. We used a similar strategy as in [24] to generate the test instances of (41). Given the dimensions $n$ and $m$, with $m\geq n$, we randomly generate $A$ (a sparse matrix with density $10^{-4}$) with elements between $(-1,1)$. Also, for a give parameter $\omega>1$, we define $\bar{X}:=\sum_{i=1}^{\omega}g_{i}g_{i}^{T}$, where $g_{i}\in\mathbb{R}^{n}$ is a random vector with only two nonnull components with the following structure $g_{i}=(\cdots\cos(\theta)\cdots\sin(\theta)\cdots)^{T}\in\mathbb{R}^{n}$, and then set $B=A\bar{X}$. Since $\bar{X}\notin C$, this procedure generally results in nonzero residue problems. ### 5.1 Influence of the forcing parameter $\gamma$ We start the numerical experiments by checking the influence of the forcing parameter $\gamma^{k}=(\gamma_{1}^{k},\gamma_{2}^{k},\gamma_{3}^{k})$ in Algorithm 1. We implemented Algorithm 1 with: $(a_{k})_{k\in\mathbb{N}}$ and $(b_{k})_{k\in\mathbb{N}}$ given as in Remark 2(ii) with $\bar{b}=100$, and using $\varphi_{\gamma^{k}}(U,V,W)=\gamma_{1}^{k}\|V-W\|_{F}^{2}+\gamma_{2}^{k}\|W-V\|_{F}^{2}+\gamma_{3}^{k}\|W-U\|_{F}^{2},\quad\forall k=1,2,\ldots,$ (43) as the error tolerance function, see Definition 2. We also set $\bar{\gamma_{2}}=0.49995$. Concerning parameter $\bar{\gamma}$, we considered different values for it, as we will explain below. Given a particular $\bar{\gamma}<1/2$, we set $\alpha=0.9999\cdot\frac{1-2\bar{\gamma}}{L},$ (44) where the Lipschitz constant $L$, with respect to problem (41), is given by $L=\|A^{T}A\|_{F}$. The choice (44) for the fixed step size $\alpha$ trivially satisfies (12). Since parameter $\gamma_{3}^{k}$ and the step size $\alpha$ are closely related (see (44) and recall that $0\leq\gamma_{3}^{k}\leq\bar{\gamma}<1/2$), we first investigate the behavior of Algorithm 1 by varying only the strategy for $\gamma_{3}^{k}$. We set $\gamma_{2}^{k}=\min\left(\frac{1}{2}\frac{a_{k}}{\|\nabla f(X^{k})\|_{F}^{2}},\bar{\gamma_{2}}\right),\quad\gamma_{1}^{k}=\frac{a_{k}}{\|\nabla f(X^{k})\|_{F}^{2}}-\gamma_{2}^{k},\quad\mbox{and}\quad\gamma_{3}^{k}=\bar{\gamma},\quad\forall k=1,2,\ldots,$ (45) and considered some different values for $\bar{\gamma}$. Note that the forcing parameter $\gamma^{k}$ given by (45) satisfies (10). We used an instance of problem (41) with $n=2000$, $m=4000$, and $\omega=10$. The results for the starting point $X^{0}=(1/n)I$ and different choices for $\bar{\gamma}$ are in Table 1. In the table, “$f(X^{*})$” is the function value at the final iterate, “it” is the number of outer iterations, “time(s)” is the run time in seconds, and “$\alpha$” is the corresponding fixed step size given by (44). $\gamma_{3}^{k}=\bar{\gamma}$ | $f(X^{*})$ | it | time(s) | $\alpha$ ---|---|---|---|--- $0.0$ | 0.4899 | 107 | 28.9 | 0.0698 $0.1$ | 0.4899 | 129 | 36.6 | 0.0558 $0.2$ | 0.4899 | 162 | 43.5 | 0.0419 $0.3$ | 0.4899 | 223 | 59.6 | 0.0279 $0.4$ | 0.4899 | 375 | 101.0 | 0.0140 Table 1: Influence of parameter $\bar{\gamma}_{3}$ in the performance of Algorithm 1 with the forcing parameter $\gamma^{k}$ given as in (45) for an instance of problem (41) with $n=2000$, $m=4000$, and $\omega=10$. As can be seen in Table 1, Algorithm 1 performs better for lower values of $\bar{\gamma}_{3}$. This is undoubtedly due to the fact that the fixed step size $\alpha$ is inversely proportional to $\bar{\gamma}_{3}$, see (44) and the last column of the table. This result suggests that for the gradient method with constant step size, the best choice is to take $\gamma_{3}^{k}=0$ for all $k$, leaving the inexactness of the projections to be controlled only by the terms of $\gamma_{1}^{k}$ and $\gamma_{2}^{k}$ in (43). From an algorithmic point of view, the term corresponding to $\gamma_{3}^{k}$ in (43) involves the last two generated iterates and is often used in inexactness measures. Therefore, for projection algorithms that use a constant step size, at least under restrictions as in (12), the theory developed here presents practical alternatives for the formulation of such measures. Taking $\bar{\gamma}=0$, we consider different combinations of $\gamma_{1}^{k}$ and $\gamma_{2}^{k}$ such that $\gamma_{1}^{k}+\gamma_{2}^{k}=a_{k}/\|\nabla f(X^{k})\|_{F}^{2}$. Our experiments showed that Algorithm 1 presented no significant performance difference with these combinations. Therefore, in the experiments reported in section 5.2, we set for Algorithm 1 the forcing parameter $\gamma^{k}$ as in (45) with $\bar{\gamma}=0$. ### 5.2 Comparison with exact projection approaches In the present section, we compare the performance of Algorithms 1 and 2 with their exact counterparts. Algorithm 2 was implemented using the error tolerance function (43) with $\gamma^{k}=(0,0,0.49995),\quad\forall k=1,2,\ldots,$ and $\alpha_{k}:=\left\\{\begin{array}[]{ll}\displaystyle\min\left(\alpha_{\max},\max\left(\alpha_{\min},\langle S^{k},S^{k}\rangle/\langle S^{k},Y^{k}\rangle\right)\right),&\mbox{if}\;\langle S^{k},Y^{k}\rangle>0\\\ \alpha_{\max},&\mbox{otherwise},\end{array}\right.$ (46) where $S^{k}:=X^{k}-X^{k-1}$, $Y^{k}:=\nabla f(X^{k})-\nabla f(X^{k-1})$, $\alpha_{\min}=10^{-10}$, and $\alpha_{\max}=10^{10}$. We observe that (46) corresponds to the spectral choice for $\alpha_{k}$, see [8, 7]. In the exact versions, the projections are calculated exactly, that is, involving full eigen-decompositions. We considered some instances of problem (41) with different parameters $n$, $m$ and $\omega$ and using three starting points given by $X^{0}(\beta)=(1-\beta)(1/n)I+\beta e_{1}e_{1}^{T}$, where $e_{1}\in\mathbb{R}^{n}$ is the first canonical vector and $\beta\in\\{0.00,0.50,0.99\\}$. The results in Table 2 shows that, in relation to the CPU time, the inexact algorithms were notably more efficient (mainly in the larger instances) than the corresponding exact versions. In general, moderate values for the rank parameter $p$ in Algorithm 3 (typically, less than 10) were sufficient to compute the inexact projections, allowing significant computational savings with respect to the exact approaches. Finally, we observe that Algorithm 2 was much more efficient than Algorithm 1 on the chosen set of test problems. This was already expected, due to the simplicity of the objective function of (41). Note that Algorithm 1 does not require evaluations of the objective function (only gradient evaluations). Therefore, we hope that Algorithm 1 can be competitive in problems where the objective function is expensive to be computationally evaluated. | Algorithm 1 | Algorithm 2 ---|---|--- | Inexact | Exact | Inexact | Exact $n$ | $m$ | $\omega$ | $\beta$ | $f(X^{*})$ | it | time(s) | $f(X^{*})$ | it | time(s) | $f(X^{*})$ | it | time(s) | $f(X^{*})$ | it | time(s) | | | 0.00 | 0.4899 | 107 | 30.0 | 0.4899 | 108 | 78.3 | 0.4899 | 8 | 3.6 | 0.4899 | 8 | 6.8 | | 10 | 0.50 | 0.4899 | 108 | 36.7 | 0.4899 | 108 | 78.4 | 0.4899 | 9 | 4.0 | 0.4899 | 9 | 7.7 | | | 0.99 | 0.4899 | 110 | 29.2 | 0.4899 | 110 | 78.8 | 0.4899 | 9 | 4.0 | 0.4899 | 9 | 7.6 | | | 0.00 | 0.7887 | 141 | 47.6 | 0.7887 | 141 | 101.7 | 0.7887 | 9 | 4.2 | 0.7887 | 9 | 7.5 | | 20 | 0.50 | 0.7887 | 139 | 46.8 | 0.7887 | 139 | 100.4 | 0.7887 | 9 | 4.2 | 0.7887 | 9 | 7.6 2000 | 4000 | | 0.99 | 0.7887 | 136 | 36.1 | 0.7887 | 138 | 96.6 | 0.7887 | 10 | 4.8 | 0.7887 | 10 | 8.3 | | | 0.00 | 0.2139 | 189 | 124.3 | 0.2139 | 189 | 543.1 | 0.2139 | 8 | 9.6 | 0.2139 | 8 | 27.0 | | 10 | 0.50 | 0.2139 | 189 | 133.2 | 0.2139 | 189 | 541.5 | 0.2139 | 8 | 9.6 | 0.2139 | 9 | 30.0 | | | 0.99 | 0.2139 | 190 | 121.8 | 0.2139 | 190 | 537.1 | 0.2139 | 10 | 11.5 | 0.2139 | 9 | 30.2 | | | 0.00 | 0.9837 | 173 | 113.1 | 0.9837 | 172 | 491.6 | 0.9837 | 11 | 13.0 | 0.9837 | 12 | 39.4 | | 20 | 0.50 | 0.9837 | 170 | 109.9 | 0.9837 | 170 | 485.2 | 0.9837 | 11 | 12.9 | 0.9837 | 10 | 33.2 3000 | 6000 | | 0.99 | 0.9837 | 161 | 102.8 | 0.9837 | 164 | 466.3 | 0.9837 | 12 | 14.1 | 0.9837 | 10 | 33.3 | | | 0.00 | 1.0046 | 166 | 194.2 | 1.0046 | 165 | 1092.5 | 1.0046 | 10 | 22.8 | 1.0046 | 11 | 83.7 | | 10 | 0.50 | 1.0046 | 165 | 191.6 | 1.0046 | 165 | 1088.8 | 1.0046 | 12 | 26.4 | 1.0046 | 11 | 83.7 | | | 0.99 | 1.0046 | 169 | 193.4 | 1.0046 | 169 | 1113.0 | 1.0046 | 11 | 25.0 | 1.0046 | 12 | 91.2 | | | 0.00 | 3.0753 | 90 | 126.2 | 3.0753 | 89 | 585.9 | 3.0753 | 8 | 16.9 | 3.0753 | 8 | 63.6 | | 20 | 0.50 | 3.0753 | 89 | 122.8 | 3.0753 | 89 | 584.9 | 3.0753 | 9 | 18.5 | 3.0753 | 8 | 62.0 4000 | 8000 | | 0.99 | 3.0753 | 87 | 99.4 | 3.0753 | 86 | 566.0 | 3.0753 | 9 | 18.8 | 3.0753 | 9 | 67.3 | | | 0.00 | 0.7182 | 243 | 599.6 | 0.7182 | 243 | 3212.4 | 0.7181 | 10 | 39.7 | 0.7181 | 10 | 150.8 | | 10 | 0.50 | 0.7182 | 244 | 594.4 | 0.7182 | 244 | 3206.3 | 0.7181 | 9 | 36.3 | 0.7181 | 10 | 151.4 | | | 0.99 | 0.7182 | 246 | 571.2 | 0.7182 | 246 | 3227.1 | 0.7181 | 10 | 38.2 | 0.7181 | 10 | 150.0 | | | 0.00 | 2.7721 | 178 | 436.3 | 2.7721 | 178 | 2339.8 | 2.7721 | 8 | 29.8 | 2.7721 | 8 | 122.8 | | 20 | 0.50 | 2.7721 | 177 | 436.2 | 2.7721 | 177 | 2325.5 | 2.7721 | 8 | 30.5 | 2.7721 | 7 | 108.5 5000 | 10000 | | 0.99 | 2.7721 | 172 | 395.9 | 2.7721 | 172 | 2253.6 | 2.7721 | 8 | 29.6 | 2.7721 | 8 | 122.3 Table 2: Performance of the inexact and exact versions of Algorithms 1 and 2 in some instances of problem (41). ## 6 Conclusions In this paper, we proposed a new inexact version of the classical gradient projection method (GPM) denoted by Gradient-InexP method (GInexPM) for solving constrained convex optimization problems. As a way to compute an inexact projection the GInexPM uses a relative error tolerance. Two different strategies for choosing the step size were employed in the analyses of the method. The convergence analysis was carried out without any compactness assumption. In addition, we provided some iteration-complexity results related to GInexPM. Numerical results were reported illustrating potential advantages of considering inexact projections instead of exact ones. We expect that this paper will contribute to the development of research in this field of inexact projections, mainly to solve large-scale problems. ## References * [1] A. A. Aguiar, O. P. Ferreira, and L. F. Prudente. Subgradient method with feasible inexact projections for constrained convex optimization problems. page arXiv:2006.08770, June 2020, 2006.08770. * [2] Z. Allen-Zhu, E. Hazan, W. Hu, and Y. Li. Linear convergence of a Frank-Wolfe type algorithm over trace-norm balls. In Advances in Neural Information Processing Systems, pages 6191–6200, 2017. * [3] A. Beck. Introduction to nonlinear optimization, volume 19 of MOS-SIAM Series on Optimization. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA; Mathematical Optimization Society, Philadelphia, PA, 2014. Theory, algorithms, and applications with MATLAB. * [4] J. Y. Bello Cruz and L. R. Lucambio Pérez. Convergence of a projected gradient method variant for quasiconvex objectives. Nonlinear Anal., 73(9):2917–2922, 2010. * [5] D. P. Bertsekas. On the Goldstein-Levitin-Polyak gradient projection method. IEEE Trans. Automatic Control, AC-21(2):174–184, 1976. * [6] D. P. Bertsekas. Nonlinear programming. Athena Scientific Optimization and Computation Series. Athena Scientific, Belmont, MA, second edition, 1999. * [7] E. G. Birgin, J. M. Martínez, and M. Raydan. Nonmonotone spectral projected gradient methods on convex sets. SIAM Journal on Optimization, 10(4):1196–1211, 2000, https://doi.org/10.1137/S1052623497330963. * [8] E. G. Birgin, J. M. Martínez, and M. Raydan. Inexact spectral projected gradient methods on convex sets. IMA Journal of Numerical Analysis, 23(4):539–559, 2003. * [9] L. Bottou, F. E. Curtis, and J. Nocedal. Optimization methods for large-scale machine learning. SIAM Rev., 60(2):223–311, 2018. * [10] S. Boyd, L. El Ghaoui, E. Feron, and V. Balakrishnan. Linear matrix inequalities in system and control theory, volume 15 of SIAM Studies in Applied Mathematics. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 1994. * [11] S. Bubeck. Convex optimization: Algorithms and complexity. Foundations and Trends® in Machine Learning, 8(3-4):231–357, 2015\. * [12] R. Burachik, L. M. G. Drummond, A. N. Iusem, and B. F. Svaiter. Full convergence of the steepest descent method with inexact line searches. Optimization, 32(2):137–146, 1995. * [13] F. R. de Oliveira, O. P. Ferreira, and G. N. Silva. Newton’s method with feasible inexact projections for solving constrained generalized equations. Comput. Optim. Appl., 72(1):159–177, 2019. * [14] R. Díaz Millán, O. P. Ferreira, and L. F. Prudente. Alternating conditional gradient method for convex feasibility problems. arXiv e-prints, page arXiv:1912.04247, Dec 2019, 1912.04247. * [15] R. Escalante and M. Raydan. Dykstra’s algorithm for constrained least-squares rectangular matrix problems. Computers & Mathematics with Applications, 35(6):73 – 79, 1998\. * [16] J. Fan, L. Wang, and A. Yan. An inexact projected gradient method for sparsity-constrained quadratic measurements regression. Asia-Pac. J. Oper. Res., 36(2):1940008, 21, 2019. * [17] M. A. T. Figueiredo, R. D. Nowak, and S. J. Wright. Gradient projection for sparse reconstruction: Application to compressed sensing and other inverse problems. IEEE Journal of Selected Topics in Signal Processing, 1(4):586–597, Dec 2007. * [18] R. Fletcher. A nonlinear programming problem in statistics (educational testing). SIAM Journal on Scientific and Statistical Computing, 2(3):257–267, 1981, https://doi.org/10.1137/0902021. * [19] J. Fliege, A. I. F. Vaz, and L. N. Vicente. Complexity of gradient descent for multiobjective optimization. Optim. Methods Softw., 34(5):949–959, 2019. * [20] M. Fukushima, Z.-Q. Luo, and P. Tseng. Smoothing functions for second-order-cone complementarity problems. SIAM J. Optim., 12(2):436–460, 2001/02. * [21] M. Golbabaee and M. E. Davies. Inexact gradient projection and fast data driven compressed sensing. IEEE Transactions on Information Theory, 64(10):6707–6721, 2018\. * [22] A. A. Goldstein. Convex programming in Hilbert space. Bull. Amer. Math. Soc., 70:709–710, 1964. * [23] D. S. Gonçalves, M. L. Gonçalves, and F. R. Oliveira. Levenberg-marquardt methods with inexact projections for constrained nonlinear systems. arXiv preprint arXiv:1908.06118, 2019. * [24] D. S. Gonçalves, M. L. N. Gonçalves, and T. C. Menezes. Inexact variable metric method for convex-constrained optimization problems. Optimization-Online e-prints, 2020. * [25] P. Gong, K. Gai, and C. Zhang. Efficient euclidean projections via piecewise root finding and its application in gradient projection. Neurocomputing, 74(17):2754 – 2766, 2011. * [26] D. Gonçalves, M. Gomes-Ruggiero, and C. Lavor. A projected gradient method for optimization over density matrices. Optimization Methods and Software, 31(2):328–341, 2016, https://doi.org/10.1080/10556788.2015.1082105. * [27] M. L. N. Gonçalves, J. G. Melo, and R. D. C. Monteiro. Projection-free accelerated method for convex optimization. Optimization Methods and Software, 0(0):1–27, 2020. * [28] A. N. Iusem. On the convergence properties of the projected gradient method for convex optimization. Comput. Appl. Math., 22(1):37–52, 2003. * [29] A. N. Iusem and B. F. Svaiter. A proximal regularization of the steepest descent method. RAIRO Rech. Opér., 29(2):123–130, 1995. * [30] K. C. Kiwiel and K. Murty. Convergence of the steepest descent method for minimizing quasiconvex functions. J. Optim. Theory Appl., 89(1):221–226, 1996. * [31] G. Lan. The Complexity of Large-scale Convex Programming under a Linear Optimization Oracle. arXiv e-prints, page arXiv:1309.5550, Sep 2013, 1309.5550. * [32] G. Lan and Y. Zhou. Conditional gradient sliding for convex optimization. SIAM J. Optim., 26(2):1379–1409, 2016. * [33] C.-P. Lee and S. Wright. First-order algorithms converge faster than $o(1/k)$ on convex problems. In K. Chaudhuri and R. Salakhutdinov, editors, Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceedings of Machine Learning Research, pages 3754–3762, Long Beach, California, USA, 09–15 Jun 2019. PMLR. * [34] R. B. Lehoucq, D. C. Sorensen, and C. Yang. ARPACK Users’ Guide. Society for Industrial and Applied Mathematics, 1998, https://epubs.siam.org/doi/pdf/10.1137/1.9780898719628. * [35] E. Levitin and B. Polyak. Constrained minimization methods. USSR Computational Mathematics and Mathematical Physics, 6(5):1 – 50, 1966. * [36] G. Ma, Y. Hu, and H. Gao. An accelerated momentum based gradient projection method for image deblurring. In 2015 IEEE International Conference on Signal Processing, Communications and Computing (ICSPCC), pages 1–4, 2015. * [37] O. L. Mangasarian. Nonlinear programming, volume 10 of Classics in Applied Mathematics. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 1994. Corrected reprint of the 1969 original. * [38] J. J. Moré. Gradient projection techniques for large-scale optimization problems. In Proceedings of the 28th IEEE Conference on Decision and Control, Vol. 1–3 (Tampa, FL, 1989), pages 378–381. IEEE, New York, 1989. * [39] J. J. Moré. On the performance of algorithms for large-scale bound constrained problems. In Large-scale numerical optimization (Ithaca, NY, 1989), pages 32–45. SIAM, Philadelphia, PA, 1990. * [40] Y. Nesterov. Introductory lectures on convex optimization, volume 87 of Applied Optimization. Kluwer Academic Publishers, Boston, MA, 2004. A basic course. * [41] Y. Nesterov and A. Nemirovski. On first-order algorithms for $\ell_{1}$/nuclear norm minimization. Acta Numer., 22:509–575, 2013. * [42] J. Nocedal and S. J. Wright. Numerical optimization. Springer Series in Operations Research and Financial Engineering. Springer, New York, second edition, 2006. * [43] A. Patrascu and I. Necoara. On the convergence of inexact projection primal first-order methods for convex minimization. IEEE Trans. Automat. Control, 63(10):3317–3329, 2018. * [44] M. Schmidt, N. L. Roux, and F. Bach. Convergence rates of inexact proximal-gradient methods for convex optimization. In Proceedings of the 24th International Conference on Neural Information Processing Systems, NIPS’11, page 1458–1466, Red Hook, NY, USA, 2011. Curran Associates Inc. * [45] A. M.-C. So and Z. Zhou. Non-asymptotic convergence analysis of inexact gradient methods for machine learning without strong convexity. Optim. Methods Softw., 32(4):963–992, 2017. * [46] S. Sra, S. Nowozin, and S. Wright. Optimization for Machine Learning. Neural information processing series. MIT Press, 2012. * [47] G. W. Stewart. A krylov–schur algorithm for large eigenproblems. SIAM Journal on Matrix Analysis and Applications, 23(3):601–614, 2002, https://doi.org/10.1137/S0895479800371529. * [48] J. Tang, M. Golbabaee, and M. E. Davies. Gradient projection iterative sketch for large-scale constrained least-squares. In D. Precup and Y. W. Teh, editors, Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 3377–3386, International Convention Centre, Sydney, Australia, 06–11 Aug 2017. PMLR. * [49] S. Villa, S. Salzo, L. Baldassarre, and A. Verri. Accelerated and inexact forward-backward algorithms. SIAM J. Optim., 23(3):1607–1633, 2013. * [50] K. G. Woodgate. Least-squares solution of f = pg over positive semidefinite symmetric p. Linear Algebra and its Applications, 245:171 – 190, 1996. * [51] F. Zhang, H. Wang, J. Wang, and K. Yang. Inexact primal–dual gradient projection methods for nonlinear optimization on convex set. Optimization, 0(0):1–27, 2019.
# Global rigidity for ultra-differentiable quasiperiodic cocycles and its spectral applications Hongyu Cheng Chern Institute of Mathematics and LPMC, Nankai University, Tianjin 300071, China<EMAIL_ADDRESS>, Lingrui Ge Department of Mathematics, University of California Irvine, CA, 92697-3875, USA <EMAIL_ADDRESS>, Jiangong You Chern Institute of Mathematics and LPMC, Nankai University, Tianjin 300071, China<EMAIL_ADDRESS>and Qi Zhou Chern Institute of Mathematics and LPMC, Nankai University, Tianjin 300071, China<EMAIL_ADDRESS> ###### Abstract. For quasiperiodic Schrödinger operators with one-frequency analytic potentials, from dynamical systems side, it has been proved that the corresponding quasiperiodic Schrödinger cocycle is either rotations reducible or has positive Lyapunov exponent for all irrational frequency and almost every energy [2]. From spectral theory side, the “Schrödinger conjecture” [2] and the “Last’s intersection spectrum conjecture” have been verified [35]. The proofs of above results crucially depend on the analyticity of the potentials. People are curious about if the analyticity is essential for those problems, see open problems by Fayad-Krikorian [26, 39] and Jitomirskaya-Marx [35, 46]. In this paper, we prove the above mentioned results for ultra-differentiable potentials. ## 1\. Introduction and main results We consider smooth quasiperiodic $SL(2,{\mathbb{R}})$ cocycles $(\alpha,A):\mathbb{T}\times\mathbb{R}^{2}\rightarrow\mathbb{T}\times\mathbb{R}^{2},\ \ (\theta,w)\mapsto(\theta+\alpha,A(\theta)w),$ where $\alpha\in{\mathbb{R}}\setminus{\mathbb{Q}}$, $A\in C^{\infty}({\mathbb{T}},SL(2,{\mathbb{R}}))$ and $\mathbb{T}:=\mathbb{R}/\mathbb{Z}$. Typical examples are Schrödinger cocycles where $\begin{split}A(\theta)=S_{E}^{V}(\theta)=\left(\begin{array}[]{ll}E-V(\theta)&-1\\\ 1&0\end{array}\right),\end{split}$ which is equivalent to the eigenvalue equations of the one-dimensional quasiperiodic Schrödinger operator $H_{V,\alpha,\theta}$ defined by (1.1) $\begin{split}(H_{V,\alpha,\theta}u)_{n}=u_{n-1}+u_{n+1}+V(\theta+n\alpha)u_{n}.\end{split}$ Quasiperiodic Schrödinger operators describe the conductivity of electrons in a two-dimensional crystal layer subject to an external magnetic field of flux acting perpendicular to the lattice plane. Due to the rich backgrounds in quantum physics, quasiperiodic Schrödinger operators have been extensively studied [44]. It has been proved that the (almost) reducibility of the above Schrödinger cocycles is a powerful tool in the study of the spectral theory of quasiperiodic Schrödinger operators [51]. Recall that $(\alpha,A)$ is $C^{r}$($r$ could be $\infty$ or $\omega$) reducible, if there exist $B\in C^{r}({\mathbb{T}},PSL(2,{\mathbb{R}}))$ and $C\in SL(2,{\mathbb{R}})$ such that $B(\cdot+\alpha)A(\cdot)B(\cdot)^{-1}=C.$ We remark that the reducibility is too restrictive since even an ${\mathbb{R}}$-valued cocycle is in general not reducible if the frequency is very Liouvillean. The appropriate notion is $C^{r}$ rotations reducibility, which means, there exist $B\in C^{r}({\mathbb{T}},PSL(2,{\mathbb{R}}))$ and $C\in C^{r}({\mathbb{T}},SO(2,{\mathbb{R}}))$ such that $B(\cdot+\alpha)A(\cdot)B(\cdot)^{-1}=C(\cdot).$ By Kotani’s theory, for Lebesgue almost every $E\in\mathbb{R}$, the Schrödinger cocycles $(\alpha,S_{E}^{V})$ is either $L^{2}$ rotations reducible or has positive Lyapunov exponent 111 We refer to Section 2.1 for the definitions and basic results.. In many circumstances, especially for its dynamical and spectral applications, what’s important is the rigidity, i.e., whether $L^{2}$ conjugacy for analytic (resp. smooth) cocycles implies analytic (resp. smooth) conjugacy under some additional assumptions. In this paper, we are interested in the global rigidity results for smooth quasiperiodic Schrödinger cocycles and its spectral applications. ### 1.1. Global rigidity results for smooth quasiperiodic cocycles Based on the powerful method of renormalization, Avila-Krikorian [5] proved that if $\alpha$ is recurrent Diophantine, $V\in C^{\omega}({\mathbb{T}},{\mathbb{R}})$, then for Lebesgue almost every $E,$ the Schrödinger cocycle $(\alpha,S_{E}^{V})$ is either nonuniformly hyperbolic or $C^{\omega}$ reducible. Later, Fayad-Krikorian [26] proved that for all Diophantine $\alpha$ 222Here $\alpha$ is Diophantine (denote $\alpha\in{\rm DC}(v,\tau)$), if there exist $v>0$ and $\tau>1$ such that $\|n\alpha\|_{{\mathbb{Z}}}:=\inf_{j\in{\mathbb{Z}}}\left|n\alpha-j\right|>\frac{v}{|n|^{\tau}},\quad\forall\ n\in{\mathbb{Z}}\backslash\\{0\\}.$ We also denote ${\rm DC}:=\bigcup_{v>0,\,\tau>1}{\rm DC}(v,\tau)$ the union., and $V\in C^{\infty}({\mathbb{T}},{\mathbb{R}})$, then for Lebesgue almost every $E,$ the Schrödinger cocycle $(\alpha,S_{E}^{V})$ is either nonuniformly hyperbolic or $C^{\infty}$-reducible. Indeed, it was pointed out by Fayad-Krikorian [26], to extend the results of [26] to any irrational number is an interesting and important problem. The problem was later settled by Avila-Fayad-Krikorian [2] in the analytic case. More precisely, for all irrational $\alpha$ and any $V\in C^{\omega}({\mathbb{T}},{\mathbb{R}})$, Avila-Fayad-Krikorian [2] proved that, the Schrödinger cocycle $(\alpha,S_{E}^{V})$ is either $C^{\omega}$ rotations reducible or has positive Lyapunov exponent for Lebesgue almost every $E$. Around 2011’s, R. Krikorian [39] asked the fourth author whether the global dichotomy results of [2] hold in the $C^{\infty}$ topology. In the present paper, we address R. Krikorian’s question for a large class of $C^{\infty}$ cocycles. We first introduce the definition of $M$-ultra- differentiable functions. It is known that the derivatives of a $C^{\infty}$ function $f$ may grow as fast as you like and its regularity is characterized by the growth of $D^{s}f$. For a given sequence of positive real numbers $M=(M_{s})_{s\in\mathbb{N}},$ we say $f\in C^{\infty}({\mathbb{T}},{\mathbb{R}})$ is $M$-ultra-differentiable if there exists $r>0$ such that $\|D^{s}f\|_{C^{0}}\leq r^{-s}M_{s},$ here $r$ is also called the “width”. The real-analytic and $\nu$-Gevrey functions are two special cases corresponding to $M_{s}=s!$ and $M_{s}=(s!)^{\nu^{-1}},0<\nu<1$ respectively. Then our main result is the following: ###### Theorem 1.1. Let $\alpha\in{\mathbb{R}}\setminus\mathbb{Q}$, $V:\mathbb{T}\rightarrow{\mathbb{R}}$ be an M-ultra-differentiable function with $M=(M_{s})_{s\in\mathbb{N}}$ satisfying $\mathbf{(H1)}$: Log-convex: $M_{\ell}^{s-k}<M_{k}^{s-\ell}M_{s}^{\ell-k},\quad s>\ell>k,$ $\mathbf{(H2)}$: Sub-exponential growth: $\lim_{s\rightarrow\infty}s^{-1}\ln(M_{s+1}/M_{s})=0.$ Then for Lebesgue almost every $E\in\mathbb{R},$ either the Schrödinger cocycle $(\alpha,S_{E}^{V})$ is $C^{\infty}$ rotations reducible or it has positive Lyapunov exponent. ###### Remark 1.1. Theorem 1.1 proved the $C^{\infty}$ rigidity for a large class of $C^{\infty}$ quasiperiodic cocycles. The almost rigidity in $C^{\infty}$ topology was already proved by Fayad-Krikorian [26]. More precisely, they proved the cocycle either has positive Lyapunov exponent, or the cocycle is $C^{\infty}$ almost rotations reducible, which means the cocycle can be approximated by rotations reducible cocycles in the $C^{\infty}$ topology. We remark that the assumptions $\mathbf{(H1)}$ and $\mathbf{(H2)}$ are not restrictive. It is obvious that both analytic and Gevrey class functions satisfy $\mathbf{(H1)}$ and $\mathbf{(H2)}$. Indeed, the log-convexity condition $\mathbf{(H1)}$ is a very classical assumption in the literature, which guarantees the space of M-ultra-differentiable functions form a Banach algebra. The sub-exponential condition $\mathbf{(H2)}$ was first introduced by Bounemoura-Fejoz [13] to guarantee the ultra-differentiable functions have an analogue of the Cauchy estimates for analytic functions, which is one of the main ingredients in KAM theory. We remark that the commonly used condition in the literature is called moderate growth condition: $\sup_{s,\ell\in{\mathbb{N}}}\left(\frac{M_{s+\ell}}{M_{s}M_{\ell}}\right)^{\frac{1}{s+\ell}}<\infty,$ which is stronger than $\mathbf{(H2)},$ see [13] for details. Attached to the sequence $(M_{s})_{s\in{\mathbb{N}}}$, one can define $\Lambda:[0,\ \infty)\rightarrow[0,\ \infty)$ by $\Lambda(y):=\ln\big{(}\sup_{s\in\mathbb{N}}y^{s}M_{s}^{-1}\big{)}=\sup_{s\in\mathbb{N}}(s\ln y-\ln M_{s}),$ which in fact describes the decay rate of the Fourier coefficients for periodic functions. For $C^{\infty}$ smooth periodic functions, the growth of $\Lambda(y)$ is faster than $\ln(y^{s})$ for any $s\in{\mathbb{N}}$ as $y$ goes to infinity. Consequently, $C^{\infty}$ means $\lim_{y\rightarrow\infty}\Lambda(y)/\ln(y)=\infty.$ On the other hand, one can easily check that $M_{s}=\exp\\{s^{\delta(\delta-1)^{-1}}\\}$ satisfies $\mathbf{(H1)}$ and $\mathbf{(H2)}$ if and only if $\delta>2.$ Attached to this $M_{s}$, $\Lambda(y)=(\ln y)^{\delta}.$ Notice that $\Lambda(y)=y^{\nu},0<\nu<1$ for Gevrey functions. Thus the space of M-ultra-differentiable functions with $\mathbf{(H1)}$ and $\mathbf{(H2)}$ is much bigger than the space of Gevrey functions, and quite close to the whole space of $C^{\infty}$ functions. However, those $C^{\infty}$ functions with $\Lambda(y)\leq(\ln y)^{2}$ are not included. We don’t know it is essential or due to the shortage of our method. The proof of Theorem 1.1 is based on renormalization technique and local KAM result. For M-ultra-differentiable functions, to describe the smallness of perturbation, we define the $\|\cdot\|_{M,r}$-norm by $\displaystyle\|f\|_{M,r}=c\sup_{s\in\mathbb{N}}\big{(}(1+s)^{2}r^{s}\|D_{\theta}^{s}f(\theta)\|_{C^{0}}M_{s}^{-1}\big{)}<\infty,\ c=4\pi^{2}/3,$ and denote by $U^{M}_{r}(\mathbb{T},*)$ the set of all these $*$-valued functions ($*$ will usually denote ${\mathbb{R}}$, $sl(2,{\mathbb{R}})$ $SL(2,{\mathbb{R}})$). Then our precise KAM-type result is the following: ###### Theorem 1.2. Let $r>0$, $\alpha\in{\mathbb{R}}\setminus\mathbb{Q}$ and $A\in U^{M}_{r}(\mathbb{T},SL(2,\mathbb{R}))$ with $M$ satisfying $\mathbf{(H1)}$ and $\mathbf{(H2)}.$ Then for every $\tau>1$ and $\gamma>0,$ there exists $\varepsilon_{*}=\varepsilon_{*}(\gamma,\tau,r,M)>0,$ such that if $\|A-R\|_{M,r}\leq\varepsilon_{*}$ for some $R\in SO(2,{\mathbb{R}}),$ and $\rho(\alpha,A)=:\rho_{f}\in DC_{\alpha}(\gamma,\tau),\ i.e.,$ $\|k\alpha\pm 2\rho_{f}\|_{\mathbb{Z}}\geq\gamma\langle k\rangle^{-\tau},\ \forall k\in\mathbb{Z},\langle k\rangle=\max\\{1,|k|\\},$ then $(\alpha,A)$ is $C^{\infty}$ rotations reducible. We point out that, Theorem 1.2 is a semi-local result in the terminology of [27], i.e., the smallness of the perturbation $\varepsilon_{*}$ does not depend on the frequency $\alpha$. One should not expect that $\varepsilon_{*}$ is independent of $\rho_{f}$ (in terms of $\gamma,\tau$) as this is not true in the $C^{\infty}$ topology (or even Gevrey class) [7]. To this end, we mention another open problem of Fayad-Krikorian [27]: Is the semi-local version of the almost reducibility conjecture true for cocycles in quasi- analytic classes? In the analytic topology, it has been established in [30, 52]. The technical reason why we introduce $\mathbf{(H1)}$ and $\mathbf{(H2)}$ is the following: The proof of Theorem 1.2 is based on a non-standard KAM scheme developed in [30, 40]. The key idea is to prove that the homological equations (1.2) $\mathrm{e}^{2\mathrm{i}(2\pi\rho_{f}+\widetilde{g}(\cdot))}f(\cdot+\alpha)-f+h=0,$ has a smooth approximating solution, consult section 4.1 for more discussions. Here $\widetilde{g}(\cdot)$ comes from the perturbation, in order to ensure that (1.2) has a smooth approximating solution, we do need some kind of control for all derivatives $\|D^{s}\widetilde{g}(\cdot)\|_{C^{0}},s\in{\mathbb{N}}$ which is guaranteed by $\mathbf{(H2)}$. Next we give a short review of local reducibility results. The pioneering result of local reducibility was due to Dinaburg-Sinai [24], who proved that if $\alpha\in DC$, and $V$ is analytically small, then $(\alpha,S_{E}^{V})$ is reducible for majority of $E$. Eliasson [25] further proved for Lebesgue almost surely $E$, $(\alpha,S_{E}^{V})$ is reducible. Note these two results are perturbative, i.e. the smallness of $V$ depends on Diophantine constants of $\alpha$. For reducibility results in other topology, one can consult [20, 22, 12] and the references therein. If $\alpha$ is Liouvillean, based on “algebraic conjugacy trick” developed in [26], Avila-Fayad-Krikorian [2] proved that in the local regime, $(\alpha,S_{E}^{V})$ is reducible for majority of $E$, thus gives a generalization of Dinaburg-Sinai’s Theorem [24] to arbitrary one-frequency. The result was also proved for analytic quasiperiodic linear systems by Hou- You in [30]. Later, Zhou-Wang [55] generalized $SL(2,{\mathbb{R}})$ cocycles result [2] to $GL(d,{\mathbb{R}})$ cocycles by different method. Theorem 1.1 and Theorem 1.2 can be seen as a generalization of [2] from analytic functions to ultra-differentiable functions. ### 1.2. The spectral applications We point out that global rigidity results in the analytic topology [5, 2] have many important applications in the spectral theory of quasiperiodic Schrödigner operators. To name a few, it was used to verify the Schrödinger Conjecture [49] in the Liouvillean context [2], it also plays an essential role in solving “Last’s intersection spectrum conjecture” [35], Aubry-Andre- Jitomirskaya’s conjecture [11]. With Theorem 1.2, one can prove the first two conjectures also hold for quasiperiodic operators with M-ultra-differentiable potentials satisfying $\mathbf{(H1)}$ and $\mathbf{(H2)}$. #### 1.2.1. Schrödinger conjecture The Schrödinger conjecture [49] says, for general discrete Schrödinger operators over uniquely ergodic base dynamics, all eigenfunctions are bounded for almost every energy in the support of the absolutely continuous part of the spectral measure. This conjecture has recently been disproved by Avila [9]. However, it is still interesting to know, to what extend the conjecture is true. For example, the KAM scheme of [2] implies that the Schrödinger conjecture is true in the quasiperiodic case with analytic potentials, and this was the first time it was verified in a Liouvillean context. Indeed, as pointed by Jitomirskaya and Marx in [46] (page 2363 of [46]): addressing the Schrödinger conjecture for quasiperiodic operators with lower regularities of the potentials still remains an open problem. With Theorem 1.1, we can prove the Schrödinger conjecture with M-ultra- differentiable quasiperiodic potentials. ###### Corollary 1. Let $\alpha\in{\mathbb{R}}\setminus\mathbb{Q}$, $V:\mathbb{T}\rightarrow{\mathbb{R}}$ be a M-ultra-differentiable function satisfying $\mathbf{(H1)}$ and $\mathbf{(H2)}$. Then the Schrödinger conjecture is true. #### 1.2.2. Last’s intersection spectrum conjecture Denote $\displaystyle S_{-}(\beta)=\cap_{\theta\in{\mathbb{T}}}\Sigma_{ac}(\beta,\theta),$ where $\Sigma_{ac}(\beta,\theta)$ is the absolutely continuous spectrum of (quasi)periodic Schrödinger operator $H_{V,\beta,\theta}$ defined by (1.1). For any $\alpha\in{\mathbb{R}}\setminus{\mathbb{Q}},$ it can be approximated by a sequence of rational numbers $(p_{n}/q_{n}).$ It is well known that the rational frequency approximation is indispensable for numeric analysis, thus the existence of the limits $S_{-}(p_{n}/q_{n})$ as $p_{n}/q_{n}\rightarrow\alpha$ is crucial. A conjecture of Y. Last says, up to a set of zero Lebesgue measure, the absolutely continuous spectrum can be obtained asymptotically from $S_{-}(p_{n}/q_{n})$, the spectrum of periodic operators associated with the continued fraction expansion of $\alpha.$ Jitomirskaya-Marx [35] settled the “Last’s intersection spectrum conjecture” for analytic quasiperiodic Schrödinger operators. They also pointed out, in [35], that the analyticity of the potential $V$ is essential for the proof of their result, and, whether or not the analyticity can be relaxed without reducing the range of frequencies for which the statement holds is an interesting open problem (page 5 of [35]). In this work, we will give a positive answer to this problem for $\nu$-Gevrey potentials with $1/2<\nu\leq 1$. In the following, we say two sets $A\doteq B$ if $\chi_{A}=\chi_{B}\ $ Lebesgue almost everywhere. Moreover, we say $\lim_{n\rightarrow\infty}B_{n}\doteq B$ if $\lim_{n\rightarrow\infty}\chi_{B_{n}}=\chi_{B}\ $ Lebesgue almost everywhere. ###### Theorem 1.3. Let $\alpha\in{\mathbb{R}}\setminus\mathbb{Q}$ and $V:\mathbb{T}\rightarrow{\mathbb{R}}$ be a $\nu$-Gevrey function with $1/2<\nu\leq 1$, there is a sequence $p_{n}/q_{n}\rightarrow\alpha$ such that $\displaystyle\lim_{n\rightarrow\infty}S_{-}(p_{n}/q_{n})\doteq S_{-}(\alpha)=\Sigma_{ac}(\alpha).$ ###### Remark 1.2. As we will see, in fact we will prove (1.3) $\Sigma_{ac}(\alpha)\subset\liminf_{n\rightarrow\infty}S_{-}(p_{n}/q_{n})$ for all M-ultra-differentiable potential satisfying $\mathbf{(H1)}$ and $\mathbf{(H2)}$ (Theorem 6.1). Gevrey property only plays a role in proving (1.4) $\limsup_{n\rightarrow\infty}S_{-}(p_{n}/q_{n})\subset\Sigma_{ac}(\alpha)$ for Diophantine frequency (Theorem 6.2). We briefly explain why analyticity is crucial for the proof of [35]. On the one hand, the key of (1.3) is to prove that $E\in\Sigma_{ac}(\alpha)$ implies exponentially small variation (in $q_{n}$) of the approximating discriminants (“generalized Chambers’ formula”). For the analytic potential, Jitomirskaya- Marx [35] got this estimate as a corollary of Avila’s quantization of acceleration [8], which can be defined only for analytic cocycles. On the other hand, the proof of (1.4) was first obtained by Shamis [47] as a corollary of the continuity of Lyapunov exponent: i.e. the Lyapunov exponent $L(\beta+\cdot,\cdot)$: ${\mathbb{T}}\times C^{\omega}({\mathbb{T}},SL(2,\mathbb{C}))$ is jointly continuous for any irrational $\beta$ [16, 32, 36]. However, the Lyapunov exponent $L(\beta+\cdot,\cdot)$: ${\mathbb{T}}\times C^{\infty}({\mathbb{T}},SL(2,\mathbb{C}))$ is not continuous [50]. In fact, it was also pointed out by Jitomirskaya and Marx in [35] that analyticity should not be essential for their results, while one needs new methods in the non-analytic case. To generalize the result in [35] to ultra- differentiable potential, we have to overcome the difficulty caused by the non-analyticity of potential. One key issue is to prove the “generalized Chambers’ formula” in ultra-differentiable case. Instead of using Avila’s quantization of acceleration [8], we will use perturbative argument which avoids the analyticity, showing that if the cocycle is smoothly rotations reducible, then $q$-step transfer matrices grows sub-exponentially in $q$ 333As pointed out in footnote 5 of [35], this ideas was first pointed out by the fourth author after first preprint of [35].. To do this, we will use inverse renormalization and quantitative KAM result, to show if the cocycle is almost reducible in ultra-differentiable topology, then we have a good control of the growth of $q$-step transfer matrices. As for the proof of second inclusion, the key is to prove that Lyapunov exponent can be still continuous with respect to the rational approximation of the frequency for $\nu$-Gevrey potential $V$ with $1/2<\nu<1,$ if $\alpha$ is Diophantine, which is a generalization of the results in [16]. See Theorem 6.3 for details. Recently, Ge-Wang-You-Zhao [45] further constructed counter-examples for $\nu$-Gevrey potential $V$ with $0<\nu<1/2$, which shows Theorem 6.3 is optimal. Finally, we review some related results. For general ergodic discrete Schrödinger operators, the relation between the absolutely continuous spectrum and the spectrum of certain periodic approximates has been studied by Last in [41, 42], more precisely, [42] essentially proved that for $V\in C^{1}({\mathbb{T}})$ and a.e. $\alpha$, $\limsup_{n\rightarrow\infty}S_{-}(p_{n}/q_{n})\subset\Sigma_{ac}(\alpha)$ up to sets of zero Lebesgue measure. The conjecture is known for the almost Mathieu operator where $V(\theta)=2\lambda\cos\theta$ ([10, 43] for a.e. $\alpha$, $\lambda$ and [5, 31, 42, 35] extending to all $\alpha$). More recently, the conjecture was settled for a.e. $\alpha$ and sufficiently smooth potential by Zhao [54]. ### 1.3. The structure of this paper The paper is arranged as follows. In Section 2 we give some definitions and preliminaries. Before giving the proof of Theorem 1.2, we first derive condition $\mathbf{(A)}$ on Fourier coefficients from assumptions $\mathbf{(H1)}$ and $\mathbf{(H2)}$ on Taylor coefficients (Lemma 3.3) in Section 3. Then we prove Theorem 1.2 in Section 4 and prove Theorem 1.1 in Section 5. The proof of Theorem 1.3 is given in Section 6, which was based on Theorem 6.1 and Theorem 6.2. In Section 7 we give the proof of Generalized Chambers’ formula (Proposition 3), and in Section 8 we give the proof of the joint continuity of Lyapunov exponent (Theorem 6.3), these two results are bases of the proof of Theorem 6.1 and Theorem 6.2 respectively. ## 2\. Definitions and preliminaries ### 2.1. Quasiperiodic cocycles Given $A\in C^{0}({\mathbb{T}},SL(2,{\mathbb{R}}))$ and $\alpha\in{\mathbb{R}}\setminus{\mathbb{Q}}$, the iterates of $(\alpha,A)$ are of the form $(\alpha,A)^{n}=(n\alpha,A_{n})$, where $A_{n}(\cdot):=\left\\{\begin{array}[]{l l}A(\cdot+(n-1)\alpha)\cdots A(\cdot+\alpha)A(\cdot),&n\geq 0\\\\[2.84526pt] A^{-1}(\cdot+n\alpha)A^{-1}(\cdot+(n+1)\alpha)\cdots A^{-1}(\cdot-\alpha),&n<0\end{array}\right..$ Define the finite Lyapunov exponent as $L_{n}(\alpha,A)=\frac{1}{n}\int_{\mathbb{T}}\ln\|A_{n}(\theta)\|d\theta,$ then by Kingman’s subadditive ergodic theorem, the Lyapunov exponent of $(\alpha,A)$ is defined as $\begin{split}L(\alpha,A)=\lim_{n\rightarrow\infty}L_{n}(\alpha,A)=\inf_{n>0}L_{n}(\alpha,A)\geq 0.\end{split}$ The cocycle $(\alpha,A)$ is called uniformly hyperbolic if there exists a continuous splitting $E_{s}(\theta)\oplus E_{u}(\theta)=\mathbb{R}^{2},$ and $C>0,0<\lambda<1,$ such that for every $n\geq 1$ we have $\begin{split}\|A_{n}(\theta)w\|\leq C\lambda^{n}\|w\|,\ \forall w\in E_{s}(\theta),\\\ \|A_{-n}(\theta)w\|\leq C\lambda^{n}\|w\|,\ \forall w\in E_{u}(\theta).\end{split}$ Assume now $A\in C^{0}({\mathbb{T}},SL(2,{\mathbb{R}}))$ is homotopic to the identity, then there exist $\psi:{\mathbb{T}}\times{\mathbb{T}}\to{\mathbb{R}}$ and $u:{\mathbb{T}}\times{\mathbb{T}}\to{\mathbb{R}}^{+}$ such that $\begin{split}A(x)\cdot\left(\begin{matrix}\cos 2\pi y\\\ \sin 2\pi y\end{matrix}\right)=u(x,y)\left(\begin{matrix}\cos 2\pi(y+\psi(x,y))\\\ \sin 2\pi(y+\psi(x,y))\end{matrix}\right).\end{split}$ The function $\psi$ is called a lift of $A$. Let $\mu$ be any probability measure on ${\mathbb{T}}\times{\mathbb{T}}$ which is invariant by the continuous map $T:(x,y)\mapsto(x+\alpha,y+\psi(x,y))$, projecting over Lebesgue measure on the first coordinate (for instance, take $\mu$ as any accumulation point of $\frac{1}{n}\sum_{k=0}^{n-1}T_{*}^{k}\nu$ where $\nu$ is Lebesgue measure on ${\mathbb{T}}\times{\mathbb{T}}$). Then the number $\begin{split}\rho(\alpha,A)=\int\psi d\mu\mod{\mathbb{Z}}\end{split}$ does not depend on the choices of $\psi$ and $\mu$ and is called the fibered rotation number of $(\alpha,A)$, see [33] and [29]. It is immediate from the definition that (2.1) $|\rho(\alpha,A)-\rho|\leq\|A-R_{\rho}\|_{C^{0}}.$ ### 2.2. Continued fraction expansion Let $\alpha\in(0,1)$ be irrational. Define $a_{0}=0,\alpha_{0}=\alpha,$ and inductively for $k\geq 1$, $a_{k}=[\alpha_{k-1}^{-1}],\qquad\alpha_{k}=\alpha_{k-1}^{-1}-a_{k}=G(\alpha_{k-1})=\\{\alpha_{k-1}^{-1}\\},$ where $G(\cdot)$ is the Gauss map. Let $p_{0}=0,p_{1}=1,q_{0}=1,q_{1}=a_{1},$ then we define inductively $p_{k}=a_{k}p_{k-1}+p_{k-2}$, $q_{k}=a_{k}q_{k-1}+q_{k-2}.$ The sequence $(q_{n})$ is the denominators of best rational approximations of $\alpha$ since we have $\begin{split}\|k\alpha\|_{\mathbb{Z}}\geq\|q_{n-1}\alpha\|_{\mathbb{Z}},\quad\forall\,\,1\leq k<q_{n},\end{split}$ and $\begin{split}(q_{n}+q_{n+1})^{-1}<\|q_{n}\alpha\|_{\mathbb{Z}}\leq q_{n+1}^{-1}.\end{split}$ For sequence $(q_{n})$, we will fix a particular subsequence $(q_{n_{k}})$ of the denominators of the best rational approximations for $\alpha,$ which for simplicity will be denoted by $(Q_{k})$. Denote the sequences $(q_{n_{k}+1})$ and $(p_{n_{k}})$ by $(\overline{Q}_{k})$ and $(P_{k}),$ respectively. Next, we introduce the concept of CD bridge which was introduced in [2]. ###### Definition 1 (CD bridge,[2]). Let $0<\mathbb{A}\leq\mathbb{B}\leq\mathbb{C}$. We say that the pair of denominators $(q_{m},q_{n})$ forms a $CD(\mathbb{A},\mathbb{B},\mathbb{C})$ bridge if $\begin{split}&\bullet\,\,q_{i+1}\leq q_{i}^{\mathbb{A}},\quad i=m,\cdots,n-1,\\\ &\bullet\,\,q_{m}^{\mathbb{C}}\geq q_{n}\geq q_{m}^{\mathbb{B}}.\end{split}$ ###### Lemma 2.1. _[2]_ For any $\mathbb{A}\geq 1$, there exists a subsequence $(Q_{k})$ of $(q_{n})$ such that $Q_{0}=1$ and for each $k\geq 0,$ $Q_{k+1}\leq\overline{Q}_{k}^{\mathbb{A}^{4}}$, either $\overline{Q}_{k}\geq Q_{k}^{\mathbb{A}}$, or the pairs $(\overline{Q}_{k-1},Q_{k})$ and $(Q_{k},Q_{k+1})$ are both $CD(\mathbb{A},\mathbb{A},\mathbb{A}^{3})$ bridges. Set $\tau>1$ and $\mathbb{A}>\tau+23>24$, then for $\\{\overline{Q}_{n}\\}_{n\geq 0},$ the selected subsequence in Lemma 2.1, we have the following lemma. ###### Lemma 2.2. For $\\{\overline{Q}_{n}\\}_{n\geq 0},$ we have $\begin{split}\overline{Q}_{n+1}\geq\overline{Q}_{n}^{\mathbb{A}},\ \forall n\geq 0.\end{split}$ ###### Proof. $\mathbf{Case\ one:}$ $\overline{Q}_{n+1}\geq Q_{n+1}^{\mathbb{A}}.$ Obviously $\overline{Q}_{n+1}\geq Q_{n+1}^{\mathbb{A}}\geq\overline{Q}_{n}^{\mathbb{A}}.$ $\mathbf{Case\ two:}$ $\overline{Q}_{n+1}<Q_{n+1}^{\mathbb{A}}.$ In this case we know that $(\overline{Q}_{n},Q_{n+1})$ forms a CD $(\mathbb{A},\mathbb{A},\mathbb{A}^{3})$ bridge. Thus $Q_{n+1}\geq\overline{Q}_{n}^{\mathbb{A}},$ which implies $\overline{Q}_{n+1}\geq\overline{Q}_{n}^{\mathbb{A}}.$ ∎ ### 2.3. Renormalization In this subsection we give the notations and definitions about the renormalization which are given in [5, 26, 6]. #### 2.3.1. ${\mathbb{Z}}^{2}-$actions Consider the cocycle $(\alpha,A)\in(0,1)\setminus\mathbb{Q}\times U_{r}^{M}(\mathbb{T},SL(2,\mathbb{R}))$ and set $\beta_{n}=\Pi_{l=0}^{n}\alpha_{l}=(-1)^{n}(q_{n}\alpha- p_{n})=(q_{n+1}+\alpha_{n+1}q_{n})^{-1},$ where $\alpha_{n}=G^{n}(\alpha)$. Let $\Omega^{r}={\mathbb{R}}\times U_{r}^{M}({\mathbb{R}},SL(2,{\mathbb{R}}))$ be the subgroup of Diff$({\mathbb{R}}\times U_{r}^{M}({\mathbb{R}},SL(2,{\mathbb{R}})))$ made of skew-product diffeomorphisms $(\alpha,A)\in{\mathbb{R}}\times U_{r}^{M}({\mathbb{R}},SL(2,{\mathbb{R}})).$ A $U_{r}^{M}$ fibered ${\mathbb{Z}}^{2}-$action is a homomorphism $\Phi:{\mathbb{Z}}^{2}\rightarrow\Omega^{r}.$ We denote by $\Lambda^{r}$ the space of such actions, and denote $\Phi=(\Phi(1,0),\Phi(0,1))$ for short. Let $\Pi_{1}:{\mathbb{R}}\times U_{r}^{M}({\mathbb{R}},SL(2,{\mathbb{R}}))\rightarrow{\mathbb{R}},$ $\Pi_{2}:{\mathbb{R}}\times U_{r}^{M}({\mathbb{R}},SL(2,{\mathbb{R}}))\rightarrow U_{r}^{M}({\mathbb{R}},SL(2,{\mathbb{R}}))$ be the coordinate projections. Let also $\gamma_{n,m}^{\Phi}=\Pi_{1}\circ\Phi(n,m)$ and $A_{n,m}^{\Phi}=\Pi_{2}\circ\Phi(n,m).$ Two fibered ${\mathbb{Z}}^{2}$ actions $\Phi$, $\Phi^{\prime}$ are said to be conjugate if there exists a smooth map $B:{\mathbb{R}}\rightarrow SL(2,\mathbb{R})$ such that $\Phi^{\prime}(n,m)=(0,B)\circ\Phi(n,m)\circ(0,B)^{-1},\qquad\forall(n,m)\in{\mathbb{Z}}^{2}.$ That is $A_{n,m}^{\Phi^{\prime}}(\cdot)=B(\cdot+\gamma_{n,m}^{\Phi})A_{n,m}^{\Phi}(\cdot)B(\cdot)^{-1},\qquad\gamma_{n,m}^{\Phi^{\prime}}=\gamma_{n,m}^{\Phi}.$ We denote $\Phi^{\prime}=\mathrm{Conj_{B}}(\Phi)$ for short. We say that an action is normalized if $\Phi(1,0)=(1,Id),$ and in that case, if $\Phi(0,1)=(\alpha,A),$ the map $A\in U_{r}^{M}({\mathbb{R}},SL(2,{\mathbb{R}}))$ is clearly ${\mathbb{Z}}-$periodic. For any $M$-ultra-differentiable function $f:{\mathbb{R}}\rightarrow{\mathbb{R}}$ (not necessary periodic), one can also define $\displaystyle\|f\|_{r,T}=c\sup_{s\in\mathbb{N}}\big{(}(1+s)^{2}r^{s}\|D_{\theta}^{s}f(\theta)\|_{C^{0}([0,T])}M_{s}^{-1}\big{)}\ ,c=4\pi^{2}/3.$ If $f:{\mathbb{T}}\rightarrow{\mathbb{R}}$ is periodic, we also denote $\|f\|_{r,1}=\|f\|_{M,r}$. ###### Lemma 2.3. _(Lemma 2 of[26])_ If $\Phi\in\Lambda^{r}$ with $\gamma_{1,0}^{\Phi}=1,$ then there exists $B\in U_{r}^{M}(\mathbb{R},SL(2,\mathbb{R}))$ and a normalized action $\widetilde{\Phi}$ such that $\widetilde{\Phi}=\mathrm{Conj_{B}}(\Phi)$. Moreover, for any $T\in{\mathbb{R}}^{+}$, we have estimate $\displaystyle\|B-Id\|_{rK_{*}^{-1},1}$ $\displaystyle\leq$ $\displaystyle\|\Phi(1,0)-Id\|_{r,1},$ $\displaystyle\|B\|_{r(K_{*}T)^{-1},T}$ $\displaystyle\leq$ $\displaystyle\|\Phi(1,0)\|_{r,T}^{T+1},\quad\forall T\in{\mathbb{R}}^{+},$ where $K_{*}$ is an absolute constant. #### 2.3.2. Renormalization of actions Following [5, 26, 6], we introduce the scheme of renormalization of ${\mathbb{Z}}^{2}$ actions. Fixing $\lambda\neq 0.$ Define $M_{\lambda}:\Lambda^{r}\rightarrow\Lambda^{r}$ by $M_{\lambda}(\Phi)(n,m):=(\lambda^{-1}\gamma_{n,m}^{\Phi},A_{n,m}^{\Phi}(\lambda\cdot)).$ Let $\theta_{*}\in{\mathbb{R}}.$ Define $T_{\theta_{*}}:\Lambda^{r}\rightarrow\Lambda^{r}$ by $T_{\theta_{*}}(\Phi)(n,m):=(\gamma_{n,m}^{\Phi},A_{n,m}^{\Phi}(\cdot+\theta_{*})).$ Let $U\in GL(2,{\mathbb{R}}).$ Define $N_{U}:\Lambda^{r}\rightarrow\Lambda^{r}$ by $N_{U}(\Phi)(n,m):=\Phi(n^{\prime},m^{\prime}),\ \text{where}\ \Big{(}\begin{matrix}n^{\prime}\\\ m^{\prime}\end{matrix}\Big{)}=U^{-1}\Big{(}\begin{matrix}n\\\ m\end{matrix}\Big{)}.$ Let $\widetilde{Q}_{n}=\Big{(}\begin{matrix}q_{n},&p_{n}\\\ q_{n-1},&p_{n-1}\end{matrix}\Big{)},$ and define for $n\in{\mathbb{N}}$ and $\theta_{*}\in{\mathbb{R}}$ the renormalized actions $\mathcal{R}^{n}(\Phi):=M_{\beta_{n-1}}\circ N_{\widetilde{Q}_{n}}(\Phi),\ \mathcal{R}_{\theta_{*}}^{n}(\Phi):=T_{\theta_{*}}^{-1}\big{[}\mathcal{R}^{n}(T_{\theta_{*}}(\Phi))\big{]}.$ For any given cocycle $(\alpha,A)$ with $\alpha\in\mathbb{R}\setminus\mathbb{Q}$, we set $\Phi=((1,Id),(\alpha,A)).$ Then by the definitions of the operators above, we get $\begin{split}\mathcal{R}_{\theta_{*}}^{n}(\Phi)=((1,A^{(n,0)}),(\alpha_{n},A^{(n,1)})),\end{split}$ where $\begin{split}A^{(n,0)}(\theta)&=A_{(-1)^{n-1}q_{n-1}}(\theta_{*}+\beta_{n-1}(\theta-\theta_{*})),\\\ A^{(n,1)}(\theta)&=A_{(-1)^{n}q_{n}}(\theta_{*}+\beta_{n-1}(\theta-\theta_{*})).\end{split}$ Thus $A^{(n,0)}$ and $A^{(n,1)}$ are $\beta_{n-1}^{-1}-$periodic and can be regarded as cocycles over the dynamics on $\mathbb{R}$ given by $\theta\mapsto\theta+1$ and $\theta\mapsto\theta+\alpha_{n}$. It is easy to see that $A^{(n,1)}(\theta+1)A^{(n,0)}(\theta)=A^{(n,0)}(\theta+\alpha_{n})A^{(n,1)}(\theta),$ which expresses the commutation of the cocycles. Based on this fact, there exists $D_{n}$ which is a normalizing map such that $\displaystyle D_{n}(\theta+1)A^{(n,0)}(\theta)D_{n}(\theta)^{-1}$ $\displaystyle=$ $\displaystyle Id,$ $\displaystyle D_{n}(\theta+\alpha_{n})A^{(n,1)}(\theta)D_{n}(\theta)^{-1}$ $\displaystyle=$ $\displaystyle A^{(n)}(\theta),$ which satisfies $A^{(n)}(\theta+1)=A^{(n)}(\theta).$ Thus $A^{(n)}$ can be seen as an element of $C^{0}(\mathbb{T},SL(2,\mathbb{R})),$ and $(\alpha_{n},A^{(n)})$ is called a representative of the $n$-th renormalization of $(\alpha,A).$ #### 2.3.3. Convergence of the renormalized actions The following result on convergence of renormalized actions was essentially contained in [5, 26, 6], which deal with cocycles in $C^{\ell}$ setting with $\ell\in\mathbb{N}$ and $\ell=\infty,\omega.$ We will sketch the proof in the ultra-differentiable setting, just for completeness. ###### Proposition 1 ([5, 26, 6]). Suppose that $(\alpha,A)\in(0,1)\setminus\mathbb{Q}\times U_{r}(\mathbb{T},SL(2,\mathbb{R}))$. If $(\alpha,A)$ is $L^{2}$-conjugated to rotations and homotopic to the identity, then for almost every $\theta_{*}\in\mathbb{R},$ there exists $D_{n}\in U_{r/K_{*}^{2}}(\mathbb{R},SL(2,\mathbb{R}))$ with (2.2) $\|D_{n}\|_{r/(K_{*}^{2}T),T}\leq C^{q_{n-1}(T+1)},$ such that $\displaystyle\mathrm{Conj_{D_{n}}}(\mathcal{R}_{\theta_{*}}^{n}(\Phi))=((1,Id),(\alpha_{n},\ R_{\rho_{n}}\mathrm{e}^{F_{n}})),$ with $\|F_{n}\|_{r/K_{*}^{2},1}\rightarrow 0,$ where $K_{*}$ is an absolute constant defined in Lemma 2.3. ###### Proof. We first prove $\\{A^{(n,i)}(\theta)\\}_{n\geq 0},i=1,2,$ are precompact in $U_{r/K_{*}}^{M}.$ Indeed, Theorem 5.1 of [5] shows that for any $(\alpha,A)\in(0,1)\setminus\mathbb{Q}\times C^{s}(\mathbb{T},SL(2,\mathbb{R}))$, if it is $L^{2}$-conjugated to rotation, then for almost every $\theta_{*}\in\mathbb{R},$ there exists $K_{*}>0$ such that for every $d>0$ and for every $n>n_{0}(d),$ $\displaystyle\|\partial^{\ell}A^{(n,i)}(\theta)\|\leq K_{*}^{\ell+1}\|A(\theta)\|_{C^{s}},\ i=1,2,\ 0\leq\ell\leq s,\ |\theta-\theta_{*}|<d/n,$ which implies that $\displaystyle\|A^{(n,i)}(\theta)\|_{C^{s}}\leq 2K_{*}^{s+1}\|A(\theta)\|_{C^{s}},\ i=1,2,$ therefore, by the definition of norms of ultra-differentiable functions, we have $\displaystyle\|A^{(n,i)}(\theta)\|_{r/K_{*},1}\leq 2K_{*}\|A(\theta)\|_{r,1}<\infty,\ i=1,2.$ That is the sequences $\\{A^{(n,i)}(\theta)\\}_{n\geq 0},i=1,2,$ are uniformly bounded in $U_{r/K_{*}},$ which implies $\\{A^{(n,i)}(\theta)\\}_{n\geq 0},i=1,2,$ are precompact in $U_{r/K_{*}}^{M}.$ Assume $B\in L^{2}({\mathbb{T}},SL(2,{\mathbb{R}}))$ is the conjugation such that $B(\theta+\alpha)A(\theta)B(\theta)^{-1}\in SO(2,{\mathbb{R}}),$ consequently by Theorems 4.3, Theorem 4.4 in [6] we have $\mathcal{R}_{\theta_{*}}^{n}(\mathrm{Conj_{B(\theta_{*})}}(\Phi))=((1,\widetilde{C}_{n}^{(1)}(\theta)),(\alpha_{n},\widetilde{C}^{(2)}_{n}(\theta))$ with $\widetilde{C}_{n}^{(i)}=R_{\rho_{n}}\mathrm{e}^{U_{n}^{(i)}(\theta)},i=1,2,$ and $\displaystyle\|U_{n}^{(i)}(\theta)\|_{r/K_{*},1}\rightarrow 0,\ i=1,2,\text{if}\ n\rightarrow\infty,|\theta-\theta_{*}|\leq d/n,n\geq n_{0}(d).$ Using Lemma 2.3, there is a normalizing conjugation $\widetilde{D}_{n},$ which is closed to identity in $\|\cdot\|_{rK_{*}^{-2},1}-$topology such that $\widetilde{D}_{n}(\theta+1)\widetilde{C}_{n}^{(1)}(\theta)\widetilde{D}_{n}(\theta)^{-1}=Id.$ Denote $D_{n}=\widetilde{D}_{n}B,$ the action $\mathrm{Conj_{D_{n}}}(\mathcal{R}_{\theta_{*}}^{n}(\Phi))$ is of form $((1,Id),(\alpha_{n},\ R_{\rho_{n}}\mathrm{e}^{F_{n}}))$ with $\|F_{n}\|_{rK_{*}^{-2},1}\rightarrow 0.$ Moreover, for any $T\in{\mathbb{R}}^{+}$, by Lemma 2.3 we get $\displaystyle\|\widetilde{D}_{n}\|_{r/(K_{*}^{2}T),T}$ $\displaystyle\leq\|\widetilde{C}_{n}^{(1)}\|_{r/K_{*},T}^{T+1}\leq\|B(\theta_{*})\|^{2(T+1)}\|A^{(n,0)}\|_{r,T}^{T+1}$ $\displaystyle\leq\|B(\theta_{*})\|^{2(T+1)}\|A\|_{r,1}^{q_{n-1}(T+1)}\leq C^{q_{n-1}(T+1)},$ then (2.2) follows directly. ∎ ## 3\. Ultra-differentiable functions As we introduced, one way to define the modulus of ultra-differentiable functions is by the growth of $D^{s}f$. For periodic function $f\in C^{\infty}({\mathbb{T}},{\mathbb{R}}),$ an alternative way is to define the modulus of ultra-differentiability by the decay rate of its Fourier coefficient. Attached to the sequence $(M_{s})_{s\in{\mathbb{N}}}$, we can define $\Lambda:[0,\ \infty)\rightarrow[0,\ \infty)$ by (3.1) $\Lambda(y):=\ln\big{(}\sup_{s\in\mathbb{N}}y^{s}M_{s}^{-1}\big{)}=\sup_{s\in\mathbb{N}}(s\ln y-\ln M_{s}).$ This defines a function $\Lambda:[0,\infty)\rightarrow[0,\infty),$ which is continuous, constant equal to zero for $y\leq 1$ and strictly increasing for $y\geq 1$ (see [13, 19] or [48]). For any $f\in U_{r}^{M}(\mathbb{T},\mathbb{R}),$ write it as $f(\theta)=\sum_{k\in\mathbb{Z}}\widehat{f}(k)\mathrm{e}^{2\pi\mathrm{i}k\theta}$, one easily checks that $\Lambda$ controls the decay of the Fourier coefficients in the sense that (3.2) $|\widehat{f}(k)|\leq\|f\|_{M,r}\exp\\{-\Lambda(|2\pi k|r)\\},\ \forall k\in\mathbb{Z}.$ For periodic function, using $\Lambda$ is more natural and convenient. Now we derive some properties of $\Lambda$ for $f\in U_{r}^{M}(\mathbb{T},\mathbb{R})$ from $\mathbf{(H1)}$ and $\mathbf{(H2)},$ which will be the bases of our whole proof of the KAM scheme. ###### Lemma 3.1 (Proposition 10 of [13]). Let $f,g\in U_{r}^{M}(\mathbb{T},\mathbb{R})$ with $M$ satisfying $\mathbf{(H1)}.$ Then $f\cdot g\in U_{r}^{M}(\mathbb{T},\mathbb{R}),$ and we have $\displaystyle\|f\cdot g\|_{M,r}\leq\|f\|_{M,r}\|g\|_{M,r}.$ ###### Remark 3.1. As explained in [13], the role of the normalizing constant $c>0$ in the definition of $\|f\|_{M,r}$ is to ensure that $U_{r}^{M}(\mathbb{T},\mathbb{R})$ forms a standard Banach algebra with respect to multiplication. ###### Lemma 3.2 (Proposition 8 of [13]). Let $f\in U_{r}^{M}(\mathbb{T},\mathbb{R})$ with $M$ satisfying $\mathbf{(H2)}.$ Then $\partial f\in U_{r/2}^{M}(\mathbb{T},\mathbb{R})$ with $\displaystyle\|\partial f\|_{r/2}\leq C_{M}r^{-1}\|f\|_{r},$ where $\displaystyle C_{M}:=\sup_{s\in\mathbb{N}}\\{2^{-s}M_{s+1}M_{s}^{-1}\\}<\infty.$ ###### Lemma 3.3. $\mathbf{(H1)}$ and $\mathbf{(H2)}$ imply that there exists $\Gamma:[1,\ \infty)\rightarrow\mathbb{R}^{+}$ such that the following hold: $\displaystyle\mathbf{(A):}\left\\{\begin{array}[]{l}\mathrm{(I)}:\lim_{x\rightarrow\infty}\Gamma(x)=\infty,\\\ \mathrm{(II)}:\Gamma(x)\ln x\ \mathrm{is}\ \mathrm{non-decreasing},\\\ \mathrm{(III)}:\Lambda(y)-\Lambda(x)\geq(\ln y-\ln x)\Gamma(x)\ln x,\ \forall y>x\geq 1.\end{array}\right.$ ###### Proof. For any $x\geq 1,$ we select $s(x)\in\mathbb{N}$ as the one such that (3.3) $\begin{split}\Lambda(x)=\sup_{s\in\mathbb{N}}\ln(x^{s}M_{s}^{-1})=\ln(x^{s(x)}M_{s(x)}^{-1}).\end{split}$ ###### Claim 1. The function $s(x)\in\mathbb{N}$ is well-defined, non-decreasing with (3.4) $\lim_{x\rightarrow\infty}s(x)(\ln x)^{-1}=\infty.$ ###### Proof. It is quite standard $s(x)\in\mathbb{N}$ is well-defined, we first prove that $s(x)$ is non-decreasing. By the definition of $s(x)$, we have $\begin{split}x^{s(x)}M_{s(x)}^{-1}\geq x^{s(x)+1}M_{s(x)+1}^{-1},\ x^{s(x)}M_{s(x)}^{-1}\geq x^{s(x)-1}M_{s(x)-1}^{-1},\end{split}$ which implies (3.5) $\begin{split}M_{s(x)}/M_{s(x)-1}\leq x\leq M_{s(x)+1}/M_{s(x)}.\end{split}$ Assume that there exist $y>x\geq 1$ such that $s(y)<s(x).$ The fact that $s(\cdot)\in{\mathbb{N}}$ implies $s(y)+1\leq s(x).$ First, by (3.5) we get $\begin{split}y\leq M_{s(y)+1}/M_{s(y)},\quad x\geq M_{s(x)}/M_{s(x)-1}.\end{split}$ However, by $\mathbf{(H1)}$ we know that $\\{M_{\ell+1}/M_{\ell}\\}_{\ell\in\mathbb{N}}$ is increasing, which together with $s(y)+1\leq s(x),$ implies that $\begin{split}y\leq M_{s(y)+1}/M_{s(y)}\leq M_{s(x)}/M_{s(x)-1}\leq x,\end{split}$ this contradicts with the assumption $y>x.$ Thus $s(x)$ is non-decreasing. By (3.5), we have $\begin{split}s^{-1}(x)\ln(M_{s(x)}/M_{s(x)-1})\leq s^{-1}(x)\ln x\leq s^{-1}(x)\ln(M_{s(x)+1}/M_{s(x)}),\end{split}$ then (3.4) follows from the assumption $\mathbf{(H2)}$. ∎ Let $\Gamma(x)=s(x)(\ln x)^{-1}$. Then (3.4) implies $\mathrm{(I)}.$ Moreover, note $\Gamma(x)\ln x=s(x),$ which together with the fact $s(x)\in\mathbb{N}$ is non-decreasing, implies $\mathrm{(II)}.$ Now we prove $\mathrm{(III)}$. For any $y>x$, by the fact $s(x)\in\mathbb{N}$ is non-decreasing, we can distinguish the proof into two cases: $\textbf{Case 1}:s(y)=s(x).$ By the definitions of $\Lambda(x)$ and $s(x)$, we get $\begin{split}\Lambda(y)-\Lambda(x)=(\ln y-\ln x)s(x)=(\ln y-\ln x)\Gamma(x)\ln x.\end{split}$ $\mathbf{Case\ 2}:\ s(y)\geq s(x)+1.$ The inequality on the left hand of (3.5) and the fact that $\\{M_{s+1}/M_{s}\\}_{s\in\mathbb{N}}$ is increasing imply $\begin{split}y\geq M_{s(y)}/M_{s(y)-1}\geq M_{s(x)+1}/M_{s(x)},\end{split}$ that is $\ln y\geq\ln M_{s(x)+1}-\ln M_{s(x)}.$ Together with the definitions of $\Lambda(x)$ and $s(x)$, it yields $\begin{split}\Lambda(y)-\Lambda(x)&\geq\ln(y^{s(x)+1}M_{s(x)+1}^{-1})-\ln(x^{s(x)}M_{s(x)}^{-1})\\\ &=(\ln y-\ln x)s(x)+\ln y-(\ln M_{s(x)+1}-\ln M_{s(x)})\\\ &\geq(\ln y-\ln x)s(x)=(\ln y-\ln x)\Gamma(x)\ln x.\end{split}$ We thus finish the whole proof. ∎ ###### Remark 3.2. To give a heuristic understanding of the function $\Gamma$, we can assume that $\Lambda$ is differentiable, by (3.3), we can rewrite it as $\Gamma(x)=x\Lambda^{\prime}(x)(\ln x)^{-1}.$ Now if we fix $\Lambda(x)=(\ln x)^{\delta},$ then $\Gamma(x)=x\Lambda^{\prime}(x)(\ln x)^{-1}=\delta(\ln x)^{\delta-2}$ and $\mathrm{(I)}$ and $\mathrm{(II)}$ are equivalent to $\delta(\ln x)^{\delta-2}\to+\infty$ and $\delta(\ln x)^{\delta-1}$ is non-decreasing, which means $\delta>2.$ For the function $f\in C^{\infty}(\mathbb{T},\mathbb{R})$ and any $K\geq 1$, we define the truncation operator $\mathcal{T}_{K}$ and projection operator $\mathcal{R}_{K}$ as $\mathcal{T}_{K}f(\theta)=\sum_{k\in\mathbb{Z},|k|<K}\widehat{f}(k)\mathrm{e}^{2\pi\mathrm{i}k\theta},\quad\mathcal{R}_{K}f(\theta)=\sum_{k\in\mathbb{Z},|k|\geq K}\widehat{f}(k)\mathrm{e}^{2\pi\mathrm{i}k\theta}.$ We denote the average of $f(\theta)$ on $\mathbb{T}$ by $[f(\theta)]_{\theta}=\int_{\mathbb{T}}f(\theta)\mathrm{d}\theta=\widehat{f}(0).$ The norm of $\mathcal{R}_{K}f(\theta)$ in a shrunken regime has the following estimate for $C^{\infty}$ functions satisfying $\mathbf{(H1)}$ and $\mathbf{(H2)}.$ ###### Lemma 3.4. Under the assumptions $\mathbf{(A)},$ there exists $T_{1}=T_{1}(M)$, such that for any $f\in U_{r}^{M}(\mathbb{T},\mathbb{R})$, if $Kr\geq T_{1}$, then (3.6) $\displaystyle\|\mathcal{R}_{K}f\|_{M,r/2}\leq C(Kr^{2})^{-1}\|f\|_{M,r}\exp\\{-9^{-1}\Gamma(4Kr)\ln(4Kr)\\}.$ Particularly, (3.7) $\displaystyle\|\mathcal{R}_{K}f\|_{C^{0}}\leq(Kr^{2})^{-1}\|f\|_{M,r}\exp\\{-\Lambda(\pi Kr)\\}.$ ###### Proof. First by $\mathrm{(II)}$ and $\mathrm{(III)}$ in $\mathbf{(A)},$ for any $|k|\geq K,$ we have $\displaystyle\Lambda(|2\pi k|r)-\Lambda(|2\pi k|(7/8)r)$ $\displaystyle\geq$ $\displaystyle\\{\ln(|2\pi k|r)-\ln(|2\pi k|(7/8)r)\\}\Gamma(|2\pi k|(7/8)r)\ln(|2\pi k|(7/8)r)$ $\displaystyle\geq$ $\displaystyle\ln(8/7)\Gamma(|4k|r)\ln(|4k|r)>9^{-1}\Gamma(4Kr)\ln(4Kr).$ Moreover, by $\mathrm{(I)}$ in $\mathbf{(A)}$, we know there exists $T_{1}=T_{1}(M)$, such that if $Kr\geq T_{1}$ then $\Gamma(4|k|r)>18,\forall|k|\geq K,$ which implies $\displaystyle\Lambda(|2\pi k|(7/8)r)-\Lambda(|2\pi k|(3/4)r)>2\ln(|4k|r).$ Consequently, direct calculations show that (3.8) $\displaystyle\sum_{|k|\geq K}\exp\\{-\Lambda(|2\pi k|r)\\}|2\pi k(3r/4)|^{s}M_{s}^{-1}$ $\displaystyle\leq$ $\displaystyle\sum_{|k|\geq K}\exp\\{-\Lambda(|2\pi k|r)\\}\exp\\{\Lambda(|2\pi k|(3/4)r)\\}$ $\displaystyle\leq$ $\displaystyle\sup_{|k|\geq K}\exp\\{-\Lambda(|2\pi k|r)+\Lambda(|2\pi k|(7/8)r)\\}$ $\displaystyle\sum_{|k|\geq K}\exp\\{-\Lambda(|2\pi k|(7/8)r)+\Lambda(|2\pi k|(3/4)r)\\}$ $\displaystyle\leq$ $\displaystyle\exp\\{-9^{-1}\Gamma(4Kr)\ln(4Kr)\\}\sum_{|k|\geq K}|4kr|^{-2}$ $\displaystyle\leq$ $\displaystyle(4Kr^{2})^{-1}\exp\\{-9^{-1}\Gamma(4Kr)\ln(4Kr)\\}.$ Finally, by (3.2), we have $\displaystyle\|D_{\theta}^{s}\mathcal{R}_{K}f\|_{C^{0}}\leq\|f\|_{M,r}\sum_{|k|\geq K}\exp\\{-\Lambda(|2\pi k|r)\\}|2\pi k|^{s},$ then $\displaystyle\|\mathcal{R}_{K}f\|_{M,r/2}$ $\displaystyle=3^{-1}4\pi^{2}\sup_{s\in\mathbb{N}}\big{(}(r/2)^{s}(1+s)^{2}\|D_{\theta}^{s}\mathcal{R}_{K}f\|_{C^{0}}M_{s}^{-1}\big{)}$ $\displaystyle\leq\|f\|_{M,r}\sup_{s\in\mathbb{N}}3^{-1}4\pi^{2}(1+s)^{2}(2/3)^{s}$ $\displaystyle\sum_{|k|\geq K}\exp\\{-\Lambda(|2\pi k|r)\\}\\{|2\pi k|(3r/4)\\}^{s}M_{s}^{-1}$ $\displaystyle\leq C(Kr^{2})^{-1}\|f\|_{M,r}\exp\\{-9^{-1}\Gamma(4Kr)\ln(4Kr)\\},$ where the last inequality follows from (3.8). The conclusion (3.7) follows from similar computations, we thus omit the details. ∎ For the given function $f\in U_{r}^{M}(\mathbb{T},\mathbb{R}),$ we define the $\|\cdot\|_{\Lambda,r}$\- norm by (3.9) $\displaystyle\|f\|_{\Lambda,r}=\sum_{k\in\mathbb{Z}}|\widehat{f}(k)|\mathrm{e}^{\Lambda(|2\pi k|r)},$ with $\Lambda$ being the one defined by (3.1). With the help of Lemma 3.3, we can discuss the relationship between the spaces $\|\cdot\|_{M,r}$ and $\|\cdot\|_{\Lambda,r}.$ ###### Lemma 3.5. Under the assumptions $\mathbf{(H1)}$ and $\mathbf{(H2)}$, we have (3.10) $\displaystyle\|f\|_{M,r}$ $\displaystyle\leq$ $\displaystyle C\|f\|_{\Lambda,2r},$ (3.11) $\displaystyle\|f\|_{\Lambda,r/2}$ $\displaystyle\leq$ $\displaystyle(2\pi r)^{-1}(4+c_{M})\|f\|_{M,r},$ where $c_{M}$ is a constant that only depends on the sequence $M$. ###### Proof. First, (3.1) implies that for any $y>0$, any $s\in{\mathbb{N}}$, $\exp\\{-\Lambda(y)\\}y^{s}\leq M_{s},$ which yields $\displaystyle\exp\\{-\Lambda(yr)\\}y^{s}=\exp\\{-\Lambda(yr)\\}(yr)^{s}r^{-s}\leq M_{s}r^{-s},\forall yr>0.$ Then $\displaystyle\sup_{k\in\mathbb{Z}}\exp\\{-\Lambda(|2\pi k|2r)\\}|2\pi k|^{s}\leq M_{s}(2r)^{-s}.$ Thus for any $f\in U_{r}^{M}(\mathbb{T},\mathbb{R})$, for any $s\in\mathbb{N},$ we get $\displaystyle\|D_{\theta}^{s}f\|_{C^{0}}$ $\displaystyle\leq\sum_{k\in\mathbb{Z}}|\widehat{f}(k)||2\pi k|^{s}$ $\displaystyle\leq\sup_{k\in\mathbb{Z}}\exp\\{-\Lambda(|2\pi k|2r)\\}|2\pi k|^{s}\sum_{k\in\mathbb{Z}}|\widehat{f}(k)|\exp\\{\Lambda(|2\pi k|2r)\\}$ $\displaystyle\leq\|f\|_{\Lambda,2r}M_{s}(2r)^{-s},$ which implies that $\displaystyle\|f\|_{M,r}$ $\displaystyle=3^{-1}4\pi^{2}\sup_{s\in\mathbb{N}}\big{(}r^{s}(1+s)^{2}\|D_{\theta}^{s}f\|_{C^{0}}M_{s}^{-1}\big{)}$ $\displaystyle\leq\|f\|_{\Lambda,2r}\sup_{s\in\mathbb{N}}3^{-1}4\pi^{2}2^{-s}(1+s)^{2}<C\|f\|_{\Lambda,2r}.$ Now we turn to the inequality in (3.11). Easily, for $y\geq 1,$ we have $\begin{split}\exp\\{-\Lambda(2y)+\Lambda{(y)}\\}&=\inf_{s\in\mathbb{N}}\\{(2y)^{-s}M_{s}\\}y^{s(y)}M_{s(y)}^{-1}\\\ &\leq(2y)^{-(s(y)+2)}M_{s(y)+2}y^{s(y)}M_{s(y)}^{-1}\leq c_{M}(2y)^{-2},\end{split}$ where (by $\mathbf{(H2)}$) $\displaystyle c_{M}:=\sup_{s\in\mathbb{N}}\\{2^{-s}M_{s+2}M_{s}^{-1}\\}<\infty.$ Note $\Lambda(y)=0,$ if $y\leq 1$. Consequently, we have $\begin{split}\|f\|_{\Lambda,r/2}&=\sum_{k\in{\mathbb{Z}}}|\widehat{f}(k)|\exp\\{\Lambda{(|\pi k|r)}\\}\\\ &\leq\|f\|_{M,r}\sum_{k\in{\mathbb{Z}}}\exp\\{-\Lambda(|2\pi k|r)+\Lambda{(|\pi k|r)}\\}\\\ &=\|f\|_{M,r}\\{\sum_{|k|<(\pi r)^{-1}}+\sum_{|k|\geq(\pi r)^{-1}}\\}\exp\\{-\Lambda(|2\pi k|r)+\Lambda{(|\pi k|r)}\\}\\\ &\leq 2(\pi r)^{-1}\|f\|_{M,r}+c_{M}\|f\|_{M,r}\sum_{|k|\geq(\pi r)^{-1}}(|2\pi k|r)^{-2}\\\ &\leq 2(\pi r)^{-1}\|f\|_{M,r}+(2\pi r)^{-1}c_{M}\|f\|_{M,r}.\end{split}$ ∎ We have stated the above lemma only for $\|f\|_{\Lambda,r/2}$ in (3.11) as this is the only case we shall need; but clearly one could obtain an estimate for any $\|\partial^{s}f\|_{\Lambda,r/2},\ s\in{\mathbb{N}},$ by the similar discussions above. ## 4\. The inductive step ### 4.1. Sketch of the proof The proof of Theorem 1.2 is based on a non-standard KAM scheme which was first developed in [40]. Now let us briefly introduce the main idea of the proof. We start from the cocycle $(\alpha,R_{\rho_{f}}\mathrm{e}^{F_{n}})$ with $\|F_{n}\|$ of size $\varepsilon_{n}$, to conjugate it into $(\alpha,R_{\rho_{f}}\mathrm{e}^{F_{n+1}})$ with a smaller perturbation, a crucial ingredient is to solve the homological equations (4.1) $\displaystyle f_{1}(\cdot+\alpha)-f_{1}=-(g_{1}-[g_{1}]_{\theta}),$ $\mathrm{e}^{4\pi\mathrm{i}\rho_{f}}f_{2}(\cdot+\alpha)-f_{2}+g_{2}=0.$ However, if $\alpha$ is Liouvillean, (4.1) cann’t be solved at all, even if in the analytic category. This is essentially different from the classical KAM scheme. Therefore, we have to leave $g_{1}(\theta)$ (at least the resonant terms of $g_{1}(\theta)$) into the normal form. As a result, from the second step of iteration we need to consider the modified cocycle $(\alpha,R_{\rho_{f}+(2\pi)^{-1}g(\theta)}\mathrm{e}^{F_{n}(\theta)})$, thus the second equation in (4.1) is of the form $\mathrm{e}^{2\mathrm{i}(2\pi\rho_{f}+g(\theta))}f(\cdot+\alpha)-f+g_{2}=0.$ In order to get desired result, we distinguish the discussions into three steps. In the first step we eliminate the lower order terms of $g(\theta)\in U_{r}^{M}({\mathbb{T}},{\mathbb{R}})$ by solving the equation $v(\theta+\alpha)-v(\theta)=-(\mathcal{T}_{\overline{Q}_{n}}g-[g]_{\theta}).$ Although $\|g(\theta)\|$ is of size $\varepsilon_{0}$, $\|\mathrm{e}^{\mathrm{i}v}\|_{r}$ could be very large in Liouvillean frequency case. To control $\|\mathrm{e}^{\mathrm{i}v}\|_{r}$, the trick is to control $\|\mathrm{Im}v(\theta)\|$ at the cost of reducing the analytic radius greatly, which was first developed in analytic case in [53]. The key point here is that $v(\theta)$ is in fact a trigonometric polynomial, one can analytic continue $v(\theta)$ to become a real analytic function, and the “width” $r$ just plays the role of analytic radius. Therefore, one can shrink $r$ greatly in order to control $\|\mathrm{e}^{\mathrm{i}v}\|_{r}$ (Lemma 4.1). Consequently, the “width” will go to zero rapidly, and the convergence of the KAM iteration only works in the $C^{\infty}$ category. The second step is to make the perturbation much smaller by solving the homological equation $\mathrm{e}^{2\mathrm{i}(2\pi\rho_{f}+\widetilde{g}(\cdot))}f(\cdot+\alpha)-f+h=0,$ where $\|\widetilde{g}\|=O(\|F_{n}\|).$ By the method of diagonally dominant [40], we can solve its approximation equation and then to make the perturbation as small as we desire (Lemma 4.3). By these two steps, we can already get $C^{\infty}$ almost reducible result (Corollary 2). However, to get $C^{\infty}$ rotations reducible result, at the end of one KAM step we need to inverse the first step, such that the conjugation is close to the identity (Lemma 4.6). For simplicity, in the following parts we will shorten $U_{r}^{M}({\mathbb{T}},*)$ and $\|\cdot\|_{M,r}$ as $U_{r}({\mathbb{T}},*)$ and $\|\cdot\|_{r}$, also the letter $C$ denotes suitable (possibly different) large constant that do not depend on the iteration step. ### 4.2. Main iteration lemma For the functions $\Lambda(x)$ and $\Gamma(x)$ in Lemma 3.3, by $\mathrm{(I)}$ in $\mathbf{(A)}$ we know that there exists $\widetilde{T}\geq T_{1}$, where $T_{1}$ is defined in Lemma 3.4, such that for any $x\geq\widetilde{T}$ (4.2) $\displaystyle\Gamma(x)$ $\displaystyle\geq$ $\displaystyle 64\mathbb{A}^{8}\tau^{4},$ (4.3) $\displaystyle\Lambda(x)$ $\displaystyle\geq$ $\displaystyle\ln x.$ Denote (4.4) $\begin{split}T=\max\\{c_{M}^{3},\widetilde{T}^{3},\ (2^{-1}r)^{-12},(4\gamma^{-1})^{2\tau}\\},\end{split}$ where $c_{M}$ is the one in (3.11). Then for the $T$ defined above, we claim that there exists $n_{0}$ such that $Q_{n_{0}+1}\leq T^{\mathbb{A}^{4}}$ and $\overline{Q}_{n_{0}+1}\geq T.$ Indeed, let $m_{0}$ be such that $Q_{m_{0}}\leq T\leq Q_{m_{0}+1}.$ If $\overline{Q}_{m_{0}}\geq T$, then we set $n_{0}=m_{0}-1.$ Otherwise, if $\overline{Q}_{m_{0}}\leq T,$ by the definition of $(Q_{k})$, it then holds $Q_{m_{0}+1}\leq T^{\mathbb{A}^{4}}.$ By the selection, $\overline{Q}_{m_{0}+1}\geq T,$ then $n_{0}=m_{0}$ satisfy our needs. In the following we will shorten $n_{0}$ as $0,$ that is $\overline{Q}_{n}$ stands for $\overline{Q}_{n+n_{0}}.$ Without loss of generality we assume $0<r_{0}:=2^{-1}r\leq 1$. Set (4.5) $\begin{split}\widetilde{\varepsilon}_{0}=0,\qquad\varepsilon_{0}=T^{-8\mathbb{A}^{4}\tau^{2}},\end{split}$ then $\varepsilon_{0}$ just depends on $\gamma,\tau,r,M,$ but not on $\alpha.$ Once we have this, we can define the iterative parameters as following, for $n\geq 1$ (4.6) $\begin{split}\overline{r}_{n}=2\overline{Q}_{n}^{-2}r_{0},&\qquad r_{n}=\overline{Q}_{n-1}^{-2}r_{0}.\\\ \varepsilon_{n}=\varepsilon_{n-1}\overline{Q}_{n}^{-\Gamma^{\frac{1}{2}}(\overline{Q}_{n}^{\frac{1}{3}})},&\qquad\widetilde{\varepsilon}_{n}=C\sum_{l=0}^{n-1}\varepsilon_{l}.\end{split}$ To simplify the notations, for any $g\in C^{0}(\mathbb{T},\mathbb{R}),$ we denote $\begin{split}R_{g}:=\left(\begin{matrix}\cos 2\pi g\ \ -\sin 2\pi g\\\ \sin 2\pi g\ \ \cos 2\pi g\end{matrix}\right)=\mathrm{e}^{-2\pi gJ},\ J=\left(\begin{matrix}0\ \ \ \ 1\\\ -1\ \ \ 0\end{matrix}\right),\end{split}$ and set $\begin{split}\mathcal{F}_{r}(\rho_{f},\eta,\widetilde{\eta}):=\Big{\\{}(\alpha,\ R_{\rho_{f}+(2\pi)^{-1}g(\theta)}\mathrm{e}^{F(\theta)}):&\ \|g\|_{r}\leq\eta,\|F\|_{r}\leq\widetilde{\eta},\\\ &\rho_{f}=\rho(\alpha,\ R_{\rho_{f}+(2\pi)^{-1}g(\theta)}\mathrm{e}^{F(\theta)})\Big{\\}}.\end{split}$ Then the main inductive lemma is the following: ###### Proposition 2. Assume that $\rho_{f}\in DC_{\alpha}(\gamma,\tau)$, then for $n\geq 1,$ the cocycle (4.7) $\begin{split}(\alpha,\ R_{\rho_{f}+(2\pi)^{-1}g_{n}}\mathrm{e}^{F_{n}(\theta)})\in\mathcal{F}_{r_{n}}(\rho_{f},\widetilde{\varepsilon}_{n},\varepsilon_{n}),\end{split}$ with $\mathcal{R}_{\overline{Q}_{n}}g_{n}=0$ can be conjugated to (4.8) $\begin{split}(\alpha,\ R_{\rho_{f}+(2\pi)^{-1}g_{n+1}}\mathrm{e}^{F_{n+1}(\theta)})\in\mathcal{F}_{r_{n+1}}(\rho_{f},\widetilde{\varepsilon}_{n+1},\varepsilon_{n+1}),\end{split}$ with $\mathcal{R}_{\overline{Q}_{n+1}}g_{n+1}=0$ by the conjugation $\Phi_{n}$ with the estimate (4.9) $\begin{split}\|\Phi_{n}-I\|_{r_{n+1}}\leq C\varepsilon_{n}^{\frac{1}{2}}.\end{split}$ The construction of the conjugation in Proposition 2 is divided into three steps given in Lemma 4.1, Lemma 4.3 and Lemma 4.6 of the following. ###### Lemma 4.1. For $n\geq 1,$ the cocycle (4.10) $\begin{split}(\alpha,\ R_{\rho_{f}+(2\pi)^{-1}g_{n}}\mathrm{e}^{F_{n}(\theta)})\in\mathcal{F}_{r_{n}}(\rho_{f},\widetilde{\varepsilon}_{n},\varepsilon_{n}),\end{split}$ with $\mathcal{R}_{\overline{Q}_{n}}g_{n}=0$ can be conjugated to the cocycle (4.11) $\begin{split}(\alpha,\ R_{\rho_{f}}\mathrm{e}^{\widetilde{F}_{n}(\theta)})\in\mathcal{F}_{\overline{r}_{n}}(\rho_{f},0,C\varepsilon_{n}),\end{split}$ via the conjugation $(0,\mathrm{e}^{-v_{n}J})$ with $\|\mathrm{e}^{-v_{n}J}\|_{\overline{r}_{n}}\leq C.$ Before giving the proof of Lemma 4.1 we give an auxiliary lemma. To this end, for $f(\theta)=\sum_{k\in\mathbb{Z}}\widehat{f}(k)\mathrm{e}^{2\pi\mathrm{i}k\theta}\in U_{r}(\mathbb{T},\mathbb{R})$, we set $\vartheta=\theta+\mathrm{i}\widetilde{\theta},(\theta\in\mathbb{T},|\widetilde{\theta}|\leq r)$ $\displaystyle\widetilde{f}(\vartheta)=\sum_{k\in\mathbb{Z}}\widehat{f}(k)\mathrm{e}^{2\pi\mathrm{i}k(\theta+\mathrm{i}\widetilde{\theta})}.$ Then we, formally, define the analytic norm $\displaystyle\|\widetilde{f}\|_{r}^{*}=\sum_{k\in\mathbb{Z}}|\widehat{f}(k)|\sup_{|\widetilde{\theta}|\leq r,\theta\in\mathbb{T}}\big{|}\mathrm{e}^{2\pi\mathrm{i}k(\theta+\mathrm{i}\widetilde{\theta})}\big{|}=\sum_{k\in\mathbb{Z}}|\widehat{f}(k)|\mathrm{e}^{|2\pi k|r}.$ If $\mathrm{Im}\vartheta=\widetilde{\theta}=0,$ then $\widetilde{f}(\vartheta)=f(\theta),$ and if $0<|\mathrm{Im}\vartheta|\leq r,$ one has $\displaystyle\|f\|_{\Lambda,r}=\sum_{k\in\mathbb{Z}}|\widehat{f}(k)|\mathrm{e}^{\Lambda(|2\pi k|r)}\leq\|\widetilde{f}\|_{r}^{*}.$ In general, $\|\widetilde{f}\|_{r}^{*}=\infty$, however, if $f$ is a trigonometric polynomial, then $\widetilde{f}$ really defines a real analytic function in the strip $|\mathrm{Im}\vartheta|\leq r,$ motivated by this we have the following: ###### Lemma 4.2. Assume that $v$ is the solution of (4.12) $v(\theta+\alpha)-v(\theta)=-(\mathcal{T}_{\overline{Q}_{n}}g-[g]_{\theta}),$ where $g\in U_{r_{n}}(\mathbb{T},\mathbb{R})$ with $\|g(\theta)\|_{r_{n}}\leq C\varepsilon_{0}.$ Then $\|\mathrm{e}^{\mathrm{i}v(\theta)}\|_{\overline{r}_{n}}\leq C.$ ###### Proof. By comparing the Fourier coefficients of (4.12) we have $v(\theta)=\sum_{0<|k|<\overline{Q}_{n}}\widehat{v}(k)\mathrm{e}^{2\pi\mathrm{i}k\theta}=-\sum_{0<|k|<\overline{Q}_{n}}\widehat{g}(k)(\mathrm{e}^{2\pi\mathrm{i}k\alpha}-1)^{-1}\mathrm{e}^{2\pi\mathrm{i}k\theta}$ with estimate (4.13) $|\widehat{v}(k)|\leq\overline{Q}_{n}|\widehat{g}(k)|,\ 0<|k|<\overline{Q}_{n}.$ For $\theta\in\mathbb{T},$ by the fact $g(\theta)\in{\mathbb{R}}$, one has $v(\theta)\in\mathbb{R}.$ Thus for the function $\begin{split}\widetilde{v}(\vartheta)-v(\theta)=\sum_{0<|k|<\overline{Q}_{n}}\widehat{v}(k)\mathrm{e}^{2\pi\mathrm{i}k\theta}\big{(}\mathrm{e}^{-2\pi k\widetilde{\theta}}-1\big{)},\end{split}$ we have $\mathrm{Im}\widetilde{v}(\vartheta)=\mathrm{Im}(\widetilde{v}(\vartheta)-v(\theta)).$ Consequently, by (4.13), we have: $\begin{split}\|\mathrm{Im}\widetilde{v}(\vartheta)\|_{4\overline{Q}_{n}^{-2}}^{*}&\leq\|\widetilde{v}(\vartheta)-v(\theta)\|_{4\overline{Q}_{n}^{-2}}^{*}\\\ &=\sum_{0<|k|<\overline{Q}_{n}}|\widehat{v}(k)|\sup_{|\widetilde{\theta}|\leq 4\overline{Q}_{n}^{-2},\theta\in\mathbb{T}}\big{|}\mathrm{e}^{2\pi\mathrm{i}k\theta}\big{(}\mathrm{e}^{-2\pi k\widetilde{\theta}}-1\big{)}\big{|}\\\ &\leq\sum_{0<|k|<\overline{Q}_{n}}\overline{Q}_{n}^{-1}|\widehat{g}(k)|16\pi|k|\\\ &\leq 16\pi\overline{Q}_{n}^{-1}\widetilde{T}(\pi r_{n})^{-1}\sum_{|k|<\widetilde{T}(\pi r_{n})^{-1}}|\widehat{g}(k)|\\\ &\ +16(\overline{Q}_{n}r_{n})^{-1}\sum_{\widetilde{T}(\pi r_{n})^{-1}\leq|k|<\overline{Q}_{n}}|\widehat{g}(k)|\exp\\{\Lambda(|\pi k|r_{n})\\}\\\ &\leq 32\widetilde{T}(\overline{Q}_{n}r_{n})^{-1}\|g\|_{\Lambda,r_{n}/2},\end{split}$ where the third inequality follows by (4.3). By (3.11) of Lemma 3.5, one can further compute $\|\mathrm{Im}\widetilde{v}(\vartheta)\|_{4\overline{Q}_{n}^{-2}}^{*}\leq 32\widetilde{T}(2\pi\overline{Q}_{n})^{-1}(4+c_{M})\overline{Q}_{n-1}^{4}r_{0}^{2}\|g\|_{r_{n}}\leq\|g\|_{r_{n}},$ the last inequality follows by $\overline{Q}_{n}\geq\max\\{T,\overline{Q}_{n-1}^{\mathbb{A}}\\},n\geq 1$ (by Lemma 2.2 and choice of $\overline{Q}_{n_{0}}$). Therefore, by (3.10) of Lemma 3.5, we have $\begin{split}\|\mathrm{e}^{\mathrm{i}v(\theta)}\|_{\overline{r}_{n}}&\leq C\|\mathrm{e}^{\mathrm{i}v(\theta)}\|_{\Lambda,2\overline{r}_{n}}\leq C\|\mathrm{e}^{\mathrm{i}\widetilde{v}(\vartheta)}\|_{2\overline{r}_{n}}^{*}\leq C\|\mathrm{e}^{\mathrm{i}\widetilde{v}(\vartheta)}\|_{4\overline{Q}_{n}^{-2}}^{*}\\\ &\leq C\exp\\{\|\mathrm{Im}\widetilde{v}(\vartheta)\|_{4\overline{Q}_{n}^{-2}}^{*}\\}<C\exp\\{\|g\|_{r_{n}}\\}<C.\end{split}$ ∎ Proof of Lemma 4.1: Assume that $v_{n}$ is the solution of $v_{n}(\theta+\alpha)-v_{n}(\theta)=-(g_{n}(\theta)-\widehat{g}_{n}(0)).$ Note $\mathcal{R}_{\overline{Q}_{n}}g_{n}=0$, then by Lemma 4.2 we have $\begin{split}\|\mathrm{e}^{v_{n}J}\|_{\overline{r}_{n}}\leq C.\end{split}$ Direct computation shows that $(0,\mathrm{e}^{-v_{n}J})$ conjugates the cocycle (4.10) into $(\alpha,R_{\rho_{f}+(2\pi)^{-1}\widehat{g}(0)}\mathrm{e}^{\overline{F}_{n}}),$ with $\overline{F}_{n}=\mathrm{e}^{-v_{n}J}F_{n}(\theta)\mathrm{e}^{v_{n}J}.$ Thus by Lemma 3.1, we have (4.14) $\|\overline{F}_{n}\|_{\overline{r}_{n}}\leq\|\mathrm{e}^{-v_{n}J}\|_{\overline{r}_{n}}\|F_{n}\|_{\overline{r}_{n}}\|\mathrm{e}^{v_{n}J}\|_{\overline{r}_{n}}\leq C\varepsilon_{n}.$ On the other hand, since $\mathrm{e}^{-v_{n}J}$ is homotopic to the identity, the fibered rotation number remains unchanged, then by (2.1), we have $|(2\pi)^{-1}\widehat{g}_{n}(0)|\leq\|\overline{F}_{n}\|_{\overline{r}_{n}}\leq C\varepsilon_{n},$ which means (4.15) $|\widehat{g}_{n}(0)|\leq C\varepsilon_{n}.$ Also note if $B,D$ are small $sl(2,\mathbb{R})$ matrices, then there exists $E\in sl(2,\mathbb{R})$ such that $\mathrm{e}^{B}\mathrm{e}^{D}=\mathrm{e}^{B+D+E},$ where $E$ is a sum of terms at least 2 orders in $B,D.$ Consequently, by (4.14), (4.15) and Lemma 3.1, there exists $\widetilde{F}_{n}\in U_{\overline{r}_{n}}(\mathbb{T},sl(2,\mathbb{R}))$ such that $R_{\rho_{f}+(2\pi)^{-1}\widehat{g}(0)}\mathrm{e}^{\overline{F}_{n}}=R_{\rho_{f}}\mathrm{e}^{\widetilde{F}_{n}}$ with estimate $\|\widetilde{F}_{n}\|_{\overline{r}_{n}}\leq C\varepsilon_{n}.$ ∎ Once we get (4.11), we will further conjugate it to another cocycle with much smaller perturbation. We will give a lemma which can be applied to more general cocycles rather than just (4.11). ###### Lemma 4.3. Consider the cocycle $(\alpha,\ R_{\rho_{f}}\mathrm{e}^{\widetilde{F}(\theta)})$ with $\rho_{f}=\rho(\alpha,\ R_{\rho_{f}}\mathrm{e}^{\widetilde{F}(\theta)})\in DC_{\alpha}(\gamma,\tau)$ and (4.16) $\begin{split}\|\widetilde{F}\|_{\overline{r}_{n}}\leq 8^{-2}\gamma^{2}Q_{n+1}^{-2\tau^{2}},\ n\geq 0.\end{split}$ Then there is a conjugation map $\Psi_{n}\in U_{r_{n+1}}(\mathbb{T},SL(2,\mathbb{R}))$ with $\begin{split}\|\Psi_{n}-I\|_{r_{n+1}}\leq\|\widetilde{F}\|_{\overline{r}_{n}}^{\frac{1}{2}},\ n\geq 0,\end{split}$ such that $\Psi_{n}$ conjugates the cocycle $(\alpha,\ R_{\rho_{f}}\mathrm{e}^{\widetilde{F}(\theta)})$ into (4.17) $\begin{split}(\alpha,\ R_{\rho_{f}+(2\pi)^{-1}\widetilde{g}_{n}}\mathrm{e}^{G(\theta)})\in\mathcal{F}_{r_{n+1}}(\rho_{f},2\|\widetilde{F}\|_{\overline{r}_{n}},\epsilon),\end{split}$ with $\mathcal{R}_{\overline{Q}_{n+1}}\widetilde{g}_{n}=0$ and $n\geq 0,$ where $\epsilon=C^{-2}\|\widetilde{F}\|_{\overline{r}_{n}}\overline{Q}_{n+1}^{-\Gamma^{\frac{1}{2}}(\overline{Q}_{n+1}^{\frac{1}{3}})}.$ Before give the proof of Lemma 4.3 we give one important lemma, which is about the estimate of small divisors and serves as the fundamental ingredients of the proof. Although the proof is quite simple, it is the key observation that to obtaining semi-local results. ###### Lemma 4.4. For any $0<\gamma<1,\,\tau>1,$ assume that $\overline{Q}_{n+1}\geq T$ and $\rho\in DC_{\alpha}(\gamma,\tau)$, then for any $|k|\leq\overline{Q}_{n+1}^{\frac{1}{2}},$ we have (4.18) $\begin{split}\big{|}\mathrm{e}^{2\pi\mathrm{i}(k\alpha\pm 2\rho)}-1\big{|}\geq\gamma Q_{n+1}^{-\tau^{2}}.\end{split}$ ###### Proof. We just need to estimate $\big{|}\mathrm{e}^{2\pi\mathrm{i}(k\alpha+2\rho)}-1\big{|}$ since $\begin{split}\big{|}\mathrm{e}^{2\pi\mathrm{i}(k\alpha-2\rho)}-1\big{|}=\big{|}\mathrm{e}^{2\pi\mathrm{i}(-k\alpha+2\rho)}-1\big{|}.\end{split}$ $\textbf{Case\ 1.}\ \overline{Q}_{n+1}\leq Q_{n+1}^{2\tau}.$ Then our assumptions imply $\begin{split}\big{|}\mathrm{e}^{2\pi\mathrm{i}(k\alpha+2\rho)}-1\big{|}&=2|\sin\pi(k\alpha+2\rho)|>\big{\|}k\alpha+2\rho\big{\|}_{\mathbb{Z}}\geq\gamma\overline{Q}_{n+1}^{-2^{-1}\tau}\geq\gamma Q_{n+1}^{-\tau^{2}}.\end{split}$ $\textbf{Case\ 2.}\ \overline{Q}_{n+1}>Q_{n+1}^{2\tau}.$ Write $k$ as $k=\widetilde{k}+mQ_{n+1},\ m\in\mathbb{Z}$ with $|\widetilde{k}|<Q_{n+1}.$ Then we have $\begin{split}|m|\leq|k|/Q_{n+1}<\overline{Q}_{n+1}^{\frac{1}{2}}/Q_{n+1}.\end{split}$ Consequently, by the assumption that $\rho\in DC_{\alpha}(\gamma,\tau)$, one has $\begin{split}\big{|}\mathrm{e}^{2\pi\mathrm{i}(k\alpha+2\rho)}-1\big{|}&>\big{\|}\widetilde{k}\alpha+mQ_{n+1}\alpha+2\rho\big{\|}_{\mathbb{Z}}\\\ &\geq\big{\|}\widetilde{k}\alpha+2\rho\big{\|}_{\mathbb{Z}}-|m|\|Q_{n+1}\alpha\|_{\mathbb{Z}}\\\ &\geq\gamma Q_{n+1}^{-\tau}-|m|/\overline{Q}_{n+1}\geq\gamma Q_{n+1}^{-\tau}-\overline{Q}_{n+1}^{-\frac{1}{2}}Q_{n+1}^{-1}>2^{-1}\gamma Q_{n+1}^{-\tau},\end{split}$ where the last inequality is by $\begin{split}\overline{Q}_{n+1}^{\frac{1}{2}}=\overline{Q}_{n+1}^{\frac{1}{2\tau}}\overline{Q}_{n+1}^{\frac{\tau-1}{2\tau}}\geq 2\gamma^{-1}Q_{n+1}^{\tau-1},\end{split}$ which is guaranteed by $\overline{Q}_{n+1}\geq(2\gamma^{-1})^{2\tau}$ and $\overline{Q}_{n+1}>Q_{n+1}^{2\tau}.$ ∎ Set $su(1,1)$ be the space consisting of matrices of the form $\left(\begin{matrix}\mathrm{i}t\ \ \ \ \ v\\\ \overline{v}\ \ -\mathrm{i}t\end{matrix}\right)$ $(t\in\mathbb{R},\ v\in\mathbb{C}),$ we simply denote such a matrix by $\\{t,v\\}.$ Recall that $sl(2,\mathbb{R})$ is isomorphic to $su(1,1)$ by the rule $A\mapsto MAM^{-1},$ where $M=\frac{1}{1+\mathrm{i}}\left(\begin{matrix}1\ \ -\mathrm{i}\\\ 1\quad\ \ \mathrm{i}\end{matrix}\right),$ and a simple calculation yields $\begin{split}M\left(\begin{array}[]{l}x\ \qquad y+z\\\ y-z\quad\ -x\end{array}\right)M^{-1}=\left(\begin{array}[]{l}\mathrm{i}z\ \qquad x-\mathrm{i}y\\\ x+\mathrm{i}y\quad\ -\mathrm{i}z\end{array}\right),x,y,z\in\mathbb{R}.\end{split}$ Motivated by Lemma 4.4, we can define the following non-resonant and resonant spaces. $\begin{split}\mathcal{B}_{r}^{(nre)}&=\Big{\\{}\\{0,\mathcal{T}_{\overline{Q}_{n+1}^{\frac{1}{2}}}g(\theta)\\}:g\in U_{r}(\mathbb{T},\mathbb{C})\Big{\\}},\\\ \mathcal{B}_{r}^{(re)}&=\Big{\\{}\\{f(\theta),\mathcal{R}_{\overline{Q}_{n+1}^{\frac{1}{2}}}g(\theta)\\}:g\in U_{r}(\mathbb{T},\mathbb{C}),\ f\in U_{r}(\mathbb{T},\mathbb{R})\Big{\\}}.\end{split}$ It follows that $U_{r}(\mathbb{T},su(1,1))=\mathcal{B}_{r}^{(nre)}\oplus\mathcal{B}_{r}^{(re)}.$ In order to prove Lemma 4.3, we will need the following lemma: ###### Lemma 4.5. Assume that $A=\mathrm{diag}\\{\mathrm{e}^{-2\pi\mathrm{i}\rho},\mathrm{e}^{2\pi\mathrm{i}\rho}\\}$ and $g\in U_{r}(\mathbb{T},su(1,1))$. If $\rho\in DC_{\alpha}(\gamma,\tau),$ and $\|g\|_{r}\leq 8^{-2}\gamma^{2}Q_{n+1}^{-2\tau^{2}},\ n\geq 0,$ then there exist $Y\in\mathcal{B}_{r}^{(nre)}$ and $g^{(re)}\in\mathcal{B}_{r}^{(re)}$ such that $\begin{split}\mathrm{e}^{Y(\cdot+\alpha)}A\mathrm{e}^{g(\cdot)}\mathrm{e}^{-Y(\cdot)}=A\mathrm{e}^{g^{(re)}(\cdot)}\end{split}$ with $\|Y\|_{r}\leq\|g\|_{r}^{1/2}$ and $\|g^{(re)}\|_{r}\leq 2\|g\|_{r}.$ The proof of this lemma, which involves the homotopy method, is postponed to Appendix. Similar proofs appeared in [53, 23]. Proof of Lemma 4.3: Since $SL(2,{\mathbb{R}})$ is isomorphic to $SU(1,1)$, instead of $(\alpha,\ R_{\rho_{f}}\mathrm{e}^{\widetilde{F}(\theta)})$, we just consider $(\alpha,A\mathrm{e}^{W(\theta)}),$ where $A=MR_{\rho_{f}}M^{-1}=\mathrm{diag}\\{\mathrm{e}^{-2\pi\mathrm{i}\rho_{f}},\mathrm{e}^{2\pi\mathrm{i}\rho_{f}}\\}\in SU(1,1),W=M\widetilde{F}M^{-1}\in su(1,1).$ Since $\rho_{f}\in D_{\alpha}(\gamma,\tau)$ and $\|\widetilde{F}\|_{\overline{r}_{n}}\leq 8^{-2}\gamma^{2}Q_{n+1}^{-2\tau^{2}}$, by Lemma 4.5, there exist $Y\in\mathcal{B}_{\overline{r}_{n}}^{(nre)}$ and $W^{(re)}\in\mathcal{B}_{\overline{r}_{n}}^{(re)}$ such that $\mathrm{e}^{Y}$ conjugates $(\alpha,\ A\mathrm{e}^{W})$ to $(\alpha,\ A\mathrm{e}^{W^{(re)}})$ with (4.19) $\begin{split}\|Y\|_{\overline{r}_{n}}\leq\|\widetilde{F}\|_{\overline{r}_{n}}^{\frac{1}{2}},\ \|W^{(re)}\|_{\overline{r}_{n}}\leq 2\|\widetilde{F}\|_{\overline{r}_{n}}.\end{split}$ Denote $W^{(re)}(\theta)=\\{\widetilde{f}(\theta),\mathcal{R}_{\overline{Q}_{n+1}^{\frac{1}{2}}}\widetilde{g}(\theta)\\}\in\mathcal{B}_{\overline{r}_{n}}^{(re)}.$ Thus by (4.19) $\begin{split}\|\widetilde{f}(\theta)\|_{\overline{r}_{n}}\leq\|W^{(re)}(\theta)\|_{\overline{r}_{n}}\leq 2\|\widetilde{F}(\theta)\|_{\overline{r}_{n}}.\end{split}$ Note $\overline{Q}_{n+1}\geq\overline{Q}_{n}^{24}$ (by Lemma 2.2) and $\overline{Q}_{n+1}\geq T\geq r_{0}^{-12}$ (by (4.4)) we get $\begin{split}\overline{Q}_{n+1}^{\frac{1}{2}}>\overline{Q}_{n+1}^{\frac{1}{3}}>4^{-1}\overline{Q}_{n}^{4}r_{0}^{-2}=\overline{r}_{n}^{-2}\gg 1,\ n\geq 0,\end{split}$ which implies (4.20) $\begin{split}\overline{Q}_{n+1}^{\frac{1}{2}}\overline{r}_{n}>\overline{Q}_{n+1}^{\frac{1}{3}}\geq T^{\frac{1}{3}}\geq\widetilde{T}\geq T_{1}.\end{split}$ Set $P(\theta)=W^{(re)}(\theta)-\\{\mathcal{T}_{\overline{Q}_{n+1}}\widetilde{f}(\theta),0\\},$ thus $\mathcal{T}_{\overline{Q}_{n+1}^{\frac{1}{2}}}P(\theta)=0,$ then by (3.6) in Lemma 3.4, we get $\begin{split}\|P(\theta)\|_{\overline{r}_{n}/2}&\leq C(\overline{Q}_{n+1}^{\frac{1}{2}}\overline{r}_{n}^{2})^{-1}\|P(\theta)\|_{\overline{r}_{n}}\exp\\{-9^{-1}\Gamma(4\overline{Q}_{n+1}^{\frac{1}{2}}\overline{r}_{n})\ln(4\overline{Q}_{n+1}^{\frac{1}{2}}\overline{r}_{n})\\}\\\ &\leq 12C\|\widetilde{F}(\theta)\|_{\overline{r}_{n}}\exp\\{-9^{-1}\Gamma(\overline{Q}_{n+1}^{\frac{1}{3}})\ln(\overline{Q}_{n+1}^{\frac{1}{3}})\\}\\\ &\leq(2C^{2})^{-1}\|\widetilde{F}(\theta)\|_{\overline{r}_{n}}\overline{Q}_{n+1}^{-\Gamma^{\frac{1}{2}}(\overline{Q}_{n+1}^{\frac{1}{3}})},\end{split}$ where the second inequality is by (4.20) and the fact that $\Gamma(x)\ln x$ is non-decreasing, i.e., $\mathrm{(II)}$ in $\mathbf{(A)},$ and the last inequality is by (4.2), that is $\Gamma(\overline{Q}_{n+1}^{\frac{1}{3}})>64\mathbb{A}^{8}\tau^{4}.$ Note $\begin{split}A\mathrm{e}^{W^{(re)}}=A\mathrm{e}^{\\{\mathcal{T}_{\overline{Q}_{n+1}}\widetilde{f},0\\}}E,\ E=\mathrm{e}^{-\\{\mathcal{T}_{\overline{Q}_{n+1}}\widetilde{f},0\\}}\mathrm{e}^{W^{(re)}}.\end{split}$ Then by Lemma 3.1 we have $\begin{split}\|E-I\|_{\overline{r}_{n}/2}&\leq\mathrm{e}^{\|\mathcal{T}_{\overline{Q}_{n+1}}\widetilde{f}\|_{\overline{r}_{n}/2}}\|\mathrm{e}^{W^{(re)}}-\mathrm{e}^{\\{\mathcal{T}_{\overline{Q}_{n+1}}\widetilde{f},0\\}}\|_{\overline{r}_{n}/2}\\\ &=\mathrm{e}^{\|\mathcal{T}_{\overline{Q}_{n+1}}\widetilde{f}\|_{\overline{r}_{n}/2}}\|\mathrm{e}^{\\{\mathcal{T}_{\overline{Q}_{n+1}}\widetilde{f},0\\}+P}-\mathrm{e}^{\\{\mathcal{T}_{\overline{Q}_{n+1}}\widetilde{f},0\\}}\|_{\overline{r}_{n}/2}\\\ &\leq\mathrm{e}^{2\|\mathcal{T}_{\overline{Q}_{n+1}}\widetilde{f}\|_{\overline{r}_{n}/2}}\mathrm{e}^{\|P\|_{\overline{r}_{n}/2}}\|P\|_{\overline{r}_{n}/2}\\\ &\leq 2\|P\|_{\overline{r}_{n}/2}\leq C^{-2}\|\widetilde{F}\|_{\overline{r}_{n}}\overline{Q}_{n+1}^{-\Gamma^{\frac{1}{2}}(\overline{Q}_{n+1}^{\frac{1}{3}})}.\end{split}$ Thus by implicit function theorem, there exists $\widetilde{G}\in U_{\overline{r}_{n}/2}(\mathbb{T},su(1,1))$ such that $E=\mathrm{e}^{\widetilde{G}}$ with $\begin{split}\|\widetilde{G}\|_{\overline{r}_{n}/2}\leq\|E-I\|_{\overline{r}_{n}/2}<C^{-2}\|\widetilde{F}(\theta)\|_{\overline{r}_{n}}\overline{Q}_{n+1}^{-\Gamma^{\frac{1}{2}}(\overline{Q}_{n+1}^{\frac{1}{3}})}.\end{split}$ Now we go back to $SL(2,\mathbb{R}).$ Let $\Psi_{n}=\mathrm{e}^{M^{-1}YM}.$ Then $\|\Psi_{n}-I\|_{\overline{r}_{n}}\leq\|Y\|_{\overline{r}_{n}}.$ Moreover, $\Psi_{n}$ conjugates the cocycle $(\alpha,\ R_{\rho_{f}}\mathrm{e}^{\widetilde{F}(\theta)})$ to (4.17) with $G=M^{-1}\widetilde{G}M,\widetilde{g}_{n}(\theta)=-\mathcal{T}_{\overline{Q}_{n+1}}\widetilde{f}(\theta).$ Obviously, $\mathcal{R}_{\overline{Q}_{n+1}}\widetilde{g}_{n}=0.$∎ To ensure the composition of the conjugations is close to the identity, we do one more conjugation which is the inverse of transformation in Lemma 4.1: ###### Lemma 4.6. Assume that $v_{n}$ is the one defined in Lemma 4.1. Then for any $n\geq 1$, $(0,\mathrm{e}^{v_{n}(\theta)J})$ further conjugates the cocycle $\begin{split}(\alpha,\ R_{\rho_{f}+(2\pi)^{-1}\widetilde{g}_{n}}\mathrm{e}^{G(\theta)})\in\mathcal{F}_{r_{n+1}}(\rho_{f},C\varepsilon_{n},C^{-1}\varepsilon_{n+1}),\end{split}$ with $\mathcal{R}_{\overline{Q}_{n+1}}\widetilde{g}_{n}=0$, to the cocycle $\begin{split}(\alpha,R_{\rho_{f}+(2\pi)^{-1}g_{n+1}}\mathrm{e}^{F_{n+1}})\in\mathcal{F}_{r_{n+1}}(\rho_{f},\widetilde{\varepsilon}_{n+1},\ \varepsilon_{n+1})\end{split}$ with $\mathcal{R}_{\overline{Q}_{n+1}}g_{n+1}=0.$ ###### Proof. Since $v_{n}$ is the solution of $v_{n}(\theta+\alpha)-v_{n}(\theta)=-g_{n}(\theta)+\widehat{g}_{n}(0),$ then $(0,\mathrm{e}^{v_{n}(\theta)J})$ conjugates the cocycle $(\alpha,\ R_{\rho_{f}+(2\pi)^{-1}\widetilde{g}_{n}}\mathrm{e}^{G(\theta)})$ to $\begin{split}(\alpha,R_{\rho_{f}+(2\pi)^{-1}(\widetilde{g}_{n}+g_{n}-\widehat{g}_{n}(0))}\mathrm{e}^{F_{n+1}(\theta)}),\end{split}$ where $F_{n+1}=\mathrm{e}^{v_{n}(\theta)J}G\mathrm{e}^{-v_{n}(\theta)J}$. Let $g_{n+1}=\widetilde{g}_{n}+g_{n}-\widehat{g}_{n}(0),$ then by Lemma 3.1 and Lemma 4.1, we have estimates $\begin{split}\|g_{n+1}\|_{r_{n+1}}&\leq\|\widetilde{g}_{n}\|_{r_{n+1}}+\|g_{n}-\widehat{g}_{n}(0)\|_{r_{n}}\leq C\varepsilon_{n}+\widetilde{\varepsilon}_{n}=\widetilde{\varepsilon}_{n+1},\\\ \|F_{n+1}\|_{r_{n+1}}&\leq\|\mathrm{e}^{v_{n}J}\|_{r_{n+1}}^{2}\|G\|_{r_{n+1}}\leq\varepsilon_{n+1}.\end{split}$ Obviously, $\mathcal{R}_{\overline{Q}_{n+1}}g_{n+1}=0.$ Moreover, the fibered rotation number does not change since $\mathrm{e}^{v_{n}J}$ is homotopic to the identity. ∎ Now we are in the position to prove Proposition 2. First by Lemma 4.1, $(0,\mathrm{e}^{-v_{n}J})$ conjugates the cocycle (4.7) to $(\alpha,\ R_{\rho_{f}}\mathrm{e}^{\widetilde{F}_{n}(\theta)})\in\mathcal{F}_{\overline{r}_{n}}(\rho_{f},0,C\varepsilon_{n}).$ Moreover, by our definition of $\varepsilon_{n}$, one can easily check that $C\varepsilon_{n}=C\varepsilon_{n-1}\overline{Q}_{n}^{-\Gamma^{\frac{1}{2}}(\overline{Q}_{n}^{\frac{1}{3}})}\leq C\overline{Q}_{n}^{-8\mathbb{A}^{4}\tau^{2}}\leq 8^{-2}\gamma^{2}Q_{n+1}^{-2\tau^{2}},\ n\geq 1,$ the last inequality holds since, by Lemma 2.1, $\overline{Q}_{n}^{\mathbb{A}^{4}}\geq Q_{n+1}$ and $\overline{Q}_{n}\geq T\geq(4\gamma^{-1})^{2\tau},n\geq 1.$ That is (4.16) holds with $C\varepsilon_{n}$ in place of $\|\widetilde{F}_{n}\|_{\overline{r}_{n}}.$ Then by the assumption $\rho_{f}\in DC_{\alpha}(\gamma,\tau)$, one can apply Lemma 4.3, and there exists $\Psi_{n}\in U_{r_{n+1}}(\mathbb{T},SL(2,\mathbb{R}))$ with $\|\Psi_{n}-I\|_{r_{n+1}}\leq C\varepsilon_{n}^{\frac{1}{2}},$ which further conjugates the obtained cocycle into $\begin{split}(\alpha,\ R_{\rho_{f}+(2\pi)^{-1}\widetilde{g}_{n}}\mathrm{e}^{G(\theta)})&\in\mathcal{F}_{r_{n+1}}(\rho_{f},2C\varepsilon_{n},C^{-2}C\varepsilon_{n}\overline{Q}_{n+1}^{-\Gamma^{\frac{1}{2}}(\overline{Q}_{n+1}^{\frac{1}{3}})})\\\ &=\mathcal{F}_{r_{n+1}}(\rho_{f},C\varepsilon_{n},C^{-1}\varepsilon_{n+1}).\end{split}$ Finally, by Lemma 4.6, $(0,\mathrm{e}^{v_{n}(\theta)J})$ further conjugates the cocycle above to (4.8) with desired estimates. Let $\Phi_{n}=\mathrm{e}^{v_{n}(\theta)J}\Psi_{n}\mathrm{e}^{-v_{n}(\theta)J},$ then by Lemma 3.1 and Lemma 4.1, we have $\begin{split}\|\Phi_{n}-I\|_{r_{n+1}}&=\|\mathrm{e}^{v_{n}(\theta)J}(\Psi_{n}-I)\mathrm{e}^{-v_{n}(\theta)J}\|_{r_{n+1}}\\\ &\leq\|\mathrm{e}^{v_{n}(\theta)J}\|_{\overline{r}_{n}}^{2}\|\Psi_{n}-I\|_{r_{n+1}}<C\varepsilon_{n}^{\frac{1}{2}},\end{split}$ which finishes the whole proof.∎ ## 5\. Proof of Theorem 1.1 and Theorem 1.2 ### 5.1. Proof of Theorem 1.2 Set $A=R_{\varrho}\mathrm{e}^{F}.$ By the assumption $\rho(\alpha,\ R_{\varrho}\mathrm{e}^{F})=\rho_{f}$ and (2.1) one has $|\rho_{f}-\varrho|\leq 2\|F\|_{C^{0}},$ thus one can rewrite $(\alpha,R_{\varrho}\mathrm{e}^{F})$ as $(\alpha,R_{\rho_{f}}\mathrm{e}^{\widetilde{F}})$ with $\|\widetilde{F}\|_{r}\leq C\|F\|_{r}.$ Set $\varepsilon_{*}:=C^{-1}\varepsilon_{0},$ where $\varepsilon_{0}$ is the one defined by (4.5). Set $\overline{r}_{0}=r,$ by the selection of $\varepsilon_{0}$ and $Q_{1}\leq T^{\mathbb{A}^{4}},\ T\geq(4\gamma^{-1})^{2\tau},$ we get $\begin{split}\|\widetilde{F}\|_{\overline{r}_{0}}\leq C\varepsilon_{*}=\varepsilon_{0}=T^{-8\mathbb{A}^{4}\tau^{2}}\leq 8^{-2}\gamma^{2}Q_{1}^{-2\tau^{2}}.\end{split}$ Since we further assume $\rho_{f}\in DC_{\alpha}(\gamma,\tau)$, one can apply Lemma 4.3, then there exists $\Psi_{0}\in U_{r_{1}}(\mathbb{T},SL(2,\mathbb{R}))$ with $\displaystyle\|\Psi_{0}-I\|_{r_{1}}\leq C\varepsilon_{0}^{\frac{1}{2}},$ which conjugates the cocycle $(\alpha,R_{\rho_{f}}\mathrm{e}^{\widetilde{F}})$ into $\begin{split}(\alpha,\ R_{\rho_{f}+(2\pi)^{-1}\widetilde{g}_{0}}\mathrm{e}^{G_{0}})\in\mathcal{F}_{r_{1}}(\rho_{f},\widetilde{\varepsilon}_{1},\varepsilon_{1}).\end{split}$ We emphasize that in the first iteration step, we only apply Lemma 4.3, without applying Proposition 2, which is quite different from the rest steps. Now we set $\widetilde{g}_{0}=g_{1},\ G_{0}=F_{1},$ and $\Phi_{0}=\Psi_{0}.$ Then one can apply Proposition 2 inductively, and get a sequence of transformations $\\{\Phi_{n}\\}_{n\geq 0}$ with estimate $\|\Phi_{n}-I\|_{r_{n+1}}\leq C\varepsilon_{n}^{\frac{1}{2}}.$ Furthermore, let $\begin{split}\Phi^{(n)}=\Phi_{n-1}\circ\Phi_{n-2}\circ\cdots\circ\Phi_{0},\ \Phi=\lim_{n\rightarrow\infty}\Phi^{(n)},\end{split}$ then $\Phi^{(n)}$ conjugates the original cocycle $(\alpha,\ R_{\rho_{f}}\mathrm{e}^{\widetilde{F}})$ to $(\alpha,\ R_{\rho_{f}+(2\pi)^{-1}g_{n}}\mathrm{e}^{F_{n}(\theta)})$. Finally, let’s show the convergence of $\Phi^{(n)}.$ Let $\Phi=\lim_{n\rightarrow\infty}\Phi^{(n)},$ we will show $\Phi\in C^{\infty}({\mathbb{T}},SL(2,{\mathbb{R}}))$. Indeed, by the definition of $\|\cdot\|_{r}-$norm we have (5.1) $\begin{split}\|D_{\theta}^{j}f\|_{C^{0}}\leq\|f\|_{r}r^{-j}M_{j},\ \forall f\in U_{r}(\mathbb{T},SL(2,\mathbb{R})),\end{split}$ and by $\mathrm{(I)}$ of $\mathbf{(A)}$, for any $j\in{\mathbb{N}}$, there exists $n_{j}\in\mathbb{N},$ such that for any $n\geq n_{j}$, we have $CM_{j}\leq\overline{Q}_{n}^{j},$ and $\Gamma^{\frac{1}{2}}(\overline{Q}_{n}^{\frac{1}{3}})\geq 24j$. By (4.9) and standard computation, we get $\|\Phi^{(n+1)}-\Phi^{(n)}\|_{r_{n+1}}\leq C\varepsilon_{n}^{\frac{1}{2}},$ then by (5.1) we can further compute $\begin{split}\big{\|}D^{j}(\Phi^{(n+1)}-\Phi^{(n)})\big{\|}_{C^{0}}&\leq\|\Phi^{(n+1)}-\Phi^{(n)}\|_{r_{n+1}}M_{j}r_{n+1}^{-j}\leq C\varepsilon_{n}^{\frac{1}{2}}M_{j}r_{n+1}^{-j}\\\ &=\overline{Q}_{n}^{-\frac{1}{2}\Gamma^{\frac{1}{2}}(\overline{Q}_{n}^{\frac{1}{3}})}\varepsilon_{n-1}^{\frac{1}{2}}CM_{j}\overline{Q}_{n}^{2j}r_{0}^{-j}\\\ &<\overline{Q}_{n}^{-\frac{1}{2}\Gamma^{\frac{1}{2}}(\overline{Q}_{n}^{\frac{1}{3}})}\varepsilon_{n-1}^{\frac{1}{2}}\overline{Q}_{n}^{4j}<\overline{Q}_{n}^{-\frac{1}{3}\Gamma^{\frac{1}{2}}(\overline{Q}_{n}^{\frac{1}{3}})}\varepsilon_{n-1}^{\frac{1}{3}}=\varepsilon_{n}^{\frac{1}{3}},\end{split}$ which means $\Phi\in C^{\infty}({\mathbb{T}},SL(2,{\mathbb{R}}))$. Let $g_{\infty}=\lim_{n\rightarrow\infty}g_{n}$, then $\Phi$ conjugates the cocycle $(\alpha,R_{\rho_{f}}\mathrm{e}^{\widetilde{F}})$ to $(\alpha,R_{\rho_{f}+(2\pi)^{-1}g_{\infty}})$, where $g_{\infty}\in C^{\infty}({\mathbb{T}},{\mathbb{R}})$. ∎ Note the proof of the proposition 2 is separated into three steps and if we just manipulate the first two steps, we get the local almost reducibility. ###### Corollary 2. Under the assumptions of Theorem 1.2, there exists a sequence of $B_{\ell}\in U_{r_{\ell}}({\mathbb{T}},SL(2,{\mathbb{R}}))$ transforming $(\alpha,A)$ into $(\alpha,R_{\rho_{f}}\mathrm{e}^{F_{\ell}})$ with estimates (5.2) $\|B_{\ell}\|_{r_{\ell}}\leq C,\ \|F_{\ell}\|_{r_{\ell}}\leq\varepsilon_{\ell}.$ ###### Proof. Set $B_{\ell}=\Psi_{\ell-1}\mathrm{e}^{v_{\ell-1}J}\Phi^{(\ell-1)},$ where $\Psi_{\ell-1},v_{\ell-1}$ are the ones in Section 4 and $\Phi^{(\ell-1)}$ is the one defined above with $\ell-1$ in place of $n.$ Obviously, $B_{\ell}$ transforming $(\alpha,A)$ into $(\alpha,R_{\rho_{f}}\mathrm{e}^{F_{\ell}})$ and the estimates of $B_{\ell}$ and $F_{\ell}$ follow from the estimates of $\Psi_{\ell-1},v_{\ell-1}$ and $\Phi^{(\ell-1)}.$ ∎ ### 5.2. Proof of Theorem 1.1 The proof of Theorem 1.1 relies on the renormalization theory of one-frequency quasiperiodic $SL(2,{\mathbb{R}})$ cocycles. Recall for any $0<\gamma<1$ and $\tau>1,$ $DC_{\alpha}(\gamma,\tau)$ denotes the set of all $\rho$ such that $\displaystyle\|k\alpha\pm 2\rho\|_{\mathbb{Z}}\geq\gamma\langle k\rangle^{-\tau},\ \langle k\rangle=\max\\{1,|k|\\},\ \forall k\in\mathbb{Z}.$ Let $\mathcal{P}\subset[0,1/2)$ be the set of all $\rho$ such that there exist $0<\gamma<1$ and $\tau>1$ with $\rho\beta_{n-1}^{-1}\in DC_{\alpha_{n}}(\gamma,\tau)$ for infinitely many $n.$ By Borel-Cantelli lemma, $\mathcal{P}$ is full measure in $[0,1/2)$. We will fix the sequence $\\{n_{j}\\}_{j\in\mathbb{N}}$ such that $\beta_{n_{j}-1}^{-1}\rho_{f}\in DC_{\alpha_{n_{j}}}(\gamma,\tau).$ We also recall the following well-known Kotani’s theory [38]. ###### Theorem 5.1 ([38]). Let $\widetilde{\mathcal{P}}\subset[0,1/2)$ be any full measure subset. For every $V\in C^{\infty}({\mathbb{T}},{\mathbb{R}})$, for almost every $E\in{\mathbb{R}}$, we have * • either $(\alpha,S_{E}^{V})$ has a positive Lyapunov exponent, or * • $(\alpha,S_{E}^{V})$ is $L^{2}$-conjugated to an $\mathrm{SO}(2,{\mathbb{R}})$-valued cocycle and the fibered rotation number of $(\alpha,S_{E}^{V})$ belongs to $\widetilde{\mathcal{P}}.$ We start from $(\alpha,S_{E}^{V})$ which can be $L^{2}$-conjugated to an $\mathrm{SO}(2,{\mathbb{R}})$-valued cocycle. By definition of $\mathcal{P}$, if $\rho(\alpha,S_{E}^{V})=\rho_{f}$ belongs to $\mathcal{P}$, we can find $0<\gamma<1$ and $\tau>1$, and arbitrary large $j>0$, such that $\rho_{f}\beta_{n_{j}-1}^{-1}\in DC_{\alpha_{n_{j}}}(\gamma,\tau)$. Now Proposition 1 ensures that $\|F_{n_{j}}\|_{rK_{*}^{-2}}\rightarrow 0$, then we can further choose $j$ large enough, such that $\|F_{n_{j}}\|_{rK_{*}^{-2}}\leq\varepsilon_{*}(\gamma,\tau,rK_{*}^{-2},M),$ where $\varepsilon_{*}=\varepsilon_{*}(\gamma,\tau,r,M)>0$ is the one in Theorem 1.2. Since $(-1)^{n_{j}}\rho_{f}\beta_{n_{j}-1}^{-1}$ is just the rotation number of $(\alpha_{n_{j}},R_{\rho_{n_{j}}}\mathrm{e}^{F_{n_{j}}})$, by Theorem 1.2 we know that $(\alpha_{n_{j}},R_{\rho_{n_{j}}}\mathrm{e}^{F_{n_{j}}})$ is $C^{\infty}$ rotations reducible. Note $(\alpha_{n_{j}},R_{\rho_{n_{j}}}\mathrm{e}^{F_{n_{j}}})$ is rotations reducible (or reducible) implies $(\alpha,S_{E}^{V})$ is rotations reducible (or reducible) in the same regularity class (consult Proposition 4.2 of [6] for example), then Theorem 1.1 follows directly. ∎ ## 6\. Last’s intersection spectrum conjecture Consider the Schrödinger operator $H_{V,\beta,\theta}$ defined by (1.1) with ultra-differentiable potential $V\in U_{r}(\mathbb{T},\mathbb{R}),$ frequency $\beta\in\mathbb{T}$ and phase $\theta\in\mathbb{T}.$ For fixed $\theta,$ denote by $\sigma(\beta,\theta)$ and $\sigma_{ac}(\beta,\theta)$ the spectrum of $H_{V,\beta,\theta}$ and its absolutely continuous (ac)-component, respectively. It is well known that in the case $\beta=p/q,$ $\sigma(p/q,\theta)$ is purely absolutely continuous and consists of $q,$ possibly touching, bands. Moreover, in the case $\beta=\alpha$ is irrational, the spectrum and ac spectrum do not depend on $\theta:$ $\displaystyle\sigma(\alpha,\theta)=:\Sigma(\alpha),\ \sigma_{ac}(\alpha,\theta)=:\Sigma_{ac}(\alpha),\ \forall\theta\in\mathbb{T}.$ In order to treat rational and irrational frequencies on the same footing, similar to Avron et al. [10], given $\beta\in\mathbb{T},$ we introduce the sets $\displaystyle S_{+}(\beta):=\bigcup_{\theta\in\mathbb{T}}\sigma(\beta,\theta)=\Sigma(\beta),$ and $\displaystyle S_{-}(\beta):=\bigcap_{\theta\in\mathbb{T}}\sigma_{ac}(\beta,\theta)=\Sigma_{ac}(\beta).$ Note that it was proved in [35] that $S_{+}(\alpha)=\Sigma(\alpha)=\lim\limits_{n\rightarrow\infty}S_{+}(p_{n}/q_{n}).$ Theorem 1.3 follows immediately from the following Theorem 6.1 and Theorem 6.2, while the key arguments are “generalized Chambers’ formula” (Proposition 3) and continuity of Lyapunov exponent (Theorem 6.3). ### 6.1. Generalized Chambers’ formula ###### Theorem 6.1. Let $\alpha\in{\mathbb{R}}\setminus\mathbb{Q}$, $V:\mathbb{T}\rightarrow{\mathbb{R}}$ be an M-ultra-differentiable function satisfying $\mathbf{(H1)}$ and $\mathbf{(H2)}$, then we have $\displaystyle S_{-}(\alpha)=\Sigma_{ac}(\alpha)\subset\liminf_{n\rightarrow\infty}S_{-}(p_{n}/q_{n}).$ The proof of Theorem 6.1 depends on the following generalized Chambers’ formula. To state this, recall that for each $\theta\in\mathbb{T},$ $H_{V,p/q,\theta}$ is a periodic operator whose spectrum, $\sigma(p/q,\theta),$ is given in terms of the discriminant by $\displaystyle\sigma(p/q,\theta)=t_{p/q}(\cdot,\theta)^{-1}[-2,2],$ where (6.1) $\displaystyle t_{p/q}(E,\theta)=\mathrm{tr}\\{\Pi_{s=q-1}^{0}S^{V}_{E}(\theta+sp/q)\\},$ which is called as the discriminant of $H_{V,p/q,\theta}$, here $``\mathrm{tr}"$ stands for the trace. In general, this discriminant is a polynomial of degree $q$ in $E$ and $q^{-1}-$periodic in $\theta$, whence one may write (6.2) $\displaystyle t_{p/q}(E,\theta)=\sum_{k\in\mathbb{Z}}a_{q,k}(E)\mathrm{e}^{2\pi\mathrm{i}qk\theta}.$ For the almost Mathieu operator, the potential $V=2\lambda\cos 2\pi\theta$ is in fact a trigonometric polynomial of degree 1. Thus in the formula (6.2) only the Fourier coefficients with $k=0,\pm 1$ survive, resulting the celebrated Chamber formula [21, 18, 15] $\displaystyle t_{p/q}(E,\theta)=a_{q,0}(E)+2\lambda^{q}\cos(2\pi q\theta).$ Note the classical Chamber’s formula holds for any $\lambda$. In particular, it shows that phase variations of the discriminant for the subcritical almost Mathieu operator (thus has absolutely continuous spectrum) are exponentially small in q. Now for any $C^{\infty}$ potential $V:\mathbb{T}\rightarrow{\mathbb{R}}$ satisfying $\mathbf{(H1)}$ and $\mathbf{(H2)}$, $E\in\Sigma_{ac}(\alpha)$, we will show that the difference between the determines $t_{p/q}(E,\theta)$ of rational approximates of $\alpha$, and its phase-average $a_{q,0}(E)$, is in fact sub-exponentially small in $q$: ###### Proposition 3. Let $\alpha\in{\mathbb{R}}\setminus\mathbb{Q}$, $V:\mathbb{T}\rightarrow{\mathbb{R}}$ be an M-ultra-differentiable function satisfying $\mathbf{(H1)}$ and $\mathbf{(H2)}$, then for almost every $E\in\Sigma_{ac}(\alpha)$, there exist $n_{*}=n(V,\alpha,E)\in\mathbb{N}$, $c=c(E)$ such that (6.3) $\displaystyle\|t_{p_{n}/q_{n}}(E,\theta)-a_{q_{n},0}(E)\|_{C^{0}}\leq 4\exp\\{-\Lambda(cq_{n})\\}$ whenever $n\geq n_{*}.$ ###### Remark 6.1. Indeed, one can select $c(E)$, such that $cq_{n}>q_{n}^{\frac{3}{4}}.$ If $V$ is analytic, Jitomirskaya-Marx (Proposition 3.1 in [35]) proved that $\|t_{p_{n}/q_{n}}(E,\theta)-a_{q_{n},0}(E)\|_{C^{0}}$ is exponentially small in $q$. Their proof depends on Avila’s quantization of acceleration [8], key of his global theory. While our proof is a perturbation argument, completely different from theirs. Proof of Theorem 6.1. We will first prove Theorem 6.1 assuming Proposition 3 and postpone the proof of Proposition 3 to Section 7. We point out the ideas of the proof was essentially given by Avila and sketched in [35]. We give the full proof here for completeness. Let $\mathcal{K}\subset[0,1/2)$ be the set of all $\rho$ such that $\inf_{p\in{\mathbb{Z}}}|q_{n}\rho-p|\geq n^{-2},\quad\text{eventually.}$ A simple Borel Cantelli argument shows $|\mathcal{K}|=1/2$. For any $\beta\in{\mathbb{T}}$, we denote by $N(\beta,E)$ the integrated density of states (IDS). Note the set $\mathcal{P}\subset[0,1/2)$ we defined in Section 5.2 is also full of measure, $i.e.,$ $|\mathcal{P}|=1/2,$ thus $\mathcal{P}\doteq\mathcal{K}.$ Moreover, Theorem 1.1 actually implies that for almost every $E\in\Sigma_{ac}(\alpha)\doteq\mathcal{P}$, the cocycle $(\alpha,S_{E}^{V})$ is rotations reducible. Thus by Theorem 6.1 in [3], if $E\in\Sigma_{ac}(\alpha)\doteq\mathcal{P}\doteq\mathcal{K}$, $N(\beta,E)$ is Lipschitz in $\alpha,$ i.e., there exists some $\Gamma(E)$ (6.4) $|N(\alpha,E)-N(p_{n}/q_{n},E)|<q_{n}^{-2}\Gamma(E).$ Since $N(\alpha,E)\in\mathcal{K}$, then by (6.4), for $n$ sufficiently large, we have (6.5) $p-1+\frac{1}{2q_{n}}<q_{n}N(p_{n}/q_{n},E)<p-\frac{1}{2q_{n}},$ for some $1\leq p\leq q_{n}$. On the other hand, it was calculated in [1] that if $E$ belongs to the $k$-th band of $S_{+}(p_{n}/q_{n})$, we have (6.6) $q_{n}N\big{(}p_{n}/q_{n},E\big{)}=k-1+2(-1)^{q_{n}+k-1}\int_{\mathbb{T}}\rho\big{(}p_{n}/q_{n},E,\theta\big{)}d\theta+\frac{1-(-1)^{q_{n}-k+1}}{2},$ where (6.7) $\rho\Big{(}p_{n}/q_{n},E,\theta\Big{)}=\left\\{\begin{aligned} &0&t_{p_{n}/q_{n}}(E,\theta)>2,\\\ &(2\pi)^{-1}\arccos(2^{-1}t_{p_{n}/q_{n}}(E,\theta))&|t_{p_{n}/q_{n}}(E,\theta)|\leq 2,\\\ &1/2&t_{p_{n}/q_{n}}(E,\theta)<-2.\end{aligned}\right.$ Then (6.5) and (6.6) imply that (6.8) $2\big{|}\cos\big{(}2\pi\int_{\mathbb{T}}\rho\big{(}p_{n}/q_{n},E,\theta\big{)}d\theta\big{)}\big{|}<2-\frac{1}{q_{n}^{2}}.$ Since $\rho\big{(}p_{n}/q_{n},E,\theta\big{)}$ is continuous in $\theta$, (6.8) and (6.7) imply that there exists $\tilde{\theta}\in{\mathbb{T}}$, such that $|t_{p_{n}/q_{n}}(E,\tilde{\theta})|=2\big{|}\cos\big{(}2\pi\rho(p_{n}/q_{n},E,\tilde{\theta})\big{)}\big{|}<2-\frac{1}{q_{n}^{2}}.$ Then by (6.3) in Proposition 3 we have $\displaystyle\left|a_{q_{n},0}(E)\right|\leq 2-\frac{1}{q_{n}^{2}}+4\exp\\{-\Lambda(cq_{n})\\}\leq 2-\frac{1}{2q_{n}^{2}}.$ By (6.3) again, for any $\theta\in{\mathbb{T}}$, we have $\left|t_{p_{n}/q_{n}}(E,\theta)\right|\leq 2,$ which means $E\in S_{-}(p_{n}/q_{n})$.∎ ### 6.2. Continuity of the Lyapunov exponent Theorem 6.1 proves $\Sigma_{ac}(\alpha)\subset\lim_{n\rightarrow\infty}S_{-}(p_{n}/q_{n})$ for any M-ultra-differentiable potentials satisfying $\mathbf{(H1)}$ and $\mathbf{(H2)},$ however, when we come to the inverse inclusion, we can only prove the result for $\nu$-Gevrey potentials with $1/2<\nu<1.$ It is interesting to extend the conclusion below to the cocycle with ultra- differential potentials, even with $C^{\infty}$ potentials. ###### Theorem 6.2. Let $V:\mathbb{T}\rightarrow{\mathbb{R}}$ be a $\nu$-Gevrey function with $1/2<\nu<1$, and assume that $\alpha\in{\mathbb{R}}\backslash{\mathbb{Q}}$. Then there is a sequence $p_{n}/q_{n}\rightarrow\alpha$, such that $\displaystyle\limsup_{n\rightarrow\infty}S_{-}(p_{n}/q_{n})\subset S_{-}(\alpha)=\Sigma_{ac}(\alpha).$ ###### Remark 6.2. The sequence $p_{n}/q_{n}$ will be the full sequence of continued fraction approximations in the case $\alpha$ is Diophantine, and an appropriate subsequence of it otherwise. For practical purposes of making conclusions about $S_{-}(\alpha)$ based on the information on $S_{-}(p_{n}/q_{n})$, it is sufficient to have convergence along a subsequence. However, in the latter case, the potential can be any stationary bounded ergodic one. The proof of Theorem 6.2 depends on the continuity of the Lyapunov exponent for the more general Gevrey cocycles. For a Gevrey (possibly matrix valued) function $f$, we let $\displaystyle\|f\|_{\nu,r}=\sum_{k\in\mathbb{Z}}|\widehat{f}(k)|\mathrm{e}^{|2\pi k|^{\nu}r},\ 0<\nu<1.$ We denote by $G_{r}^{\nu}({\mathbb{T}},*)$ the set of all these $*$-valued functions ($*$ will usually denote ${\mathbb{R}}$, $SL(2,{\mathbb{R}})$.). If we set $r=\widetilde{r}^{\nu},$ we get $\|f\|_{\nu,r}=\|f\|_{\Lambda_{\nu},\widetilde{r}},$ where $\|\cdot\|_{\Lambda_{\nu},r}-$norm is the one defined by (3.9) with $\Lambda_{\nu}(x)=x^{\nu}.$ To simplify the notation, we introduce $\|\cdot\|_{\nu,r}.$ Note the function $\Lambda_{\nu}(x)=x^{\nu},0<\nu<1,$ satisfies the subadditivity, thus $G_{r}^{\nu}({\mathbb{T}},*),0<\nu<1,$ is a Banach algebra. ###### Theorem 6.3. Let $\rho>0,2^{-1}<\nu<1.$ Consider the cocycle $(\alpha,A)\in(0,1)\setminus{\mathbb{Q}}\times G_{\rho}^{\nu}({\mathbb{T}},SL(2,{\mathbb{R}}))$ with $\alpha\in DC.$ Then $L(\alpha,A)$ is jointly continuous in the sense that $\lim\limits_{n\rightarrow\infty}L(p_{n}/q_{n},A_{n})=L(\alpha,A),$ where $p_{n}/q_{n}$ is the continued fraction expansion of $\alpha$, and $A_{n}\in G_{\rho}^{\nu}({\mathbb{T}},SL(2,{\mathbb{R}}))$ with $A_{n}\rightarrow A$ under the topology derived by $\|\cdot\|_{\nu,\rho}-$norm. The full joint continuity was first proved by Bourgain-Jitomirskaya [16] for analytic cocycles, which also plays a fundamental role in establishing the global theory of Schödinger operator [8]. However, due to lack of analyticity, it’s very difficult to generalize the above result to all irrational $\alpha$. The main reason is that in our large deviation theorem estimates, there is an upper bound for $N$ (see the assumptions in Proposition 5), thus we cannot deal with the extremely Liouvillean frequency. It’s an interesting open question whether one can prove the continuity of the Lyapunov exponent for Liouvillean frequency and non-analytic potentials. Proof of Theorem 6.2. We will first prove Theorem 6.2 and left the proof of Theorem 6.3 to Section 8. We separate the proof into two cases. $\mathbf{Case~{}I:}$ $\alpha\in DC(v,10)$. We can assume that $L(\alpha,E)>0$, then by Theorem 6.3, for any $p_{n}/q_{n}$ sufficiently close to $\alpha$, we have $L(p_{n}/q_{n},E)>0$. This implies that $E\notin\sigma(p_{n}/q_{n},\theta)$ for some $\theta\in{\mathbb{T}}$, hence $E\notin S_{-}(p_{n}/q_{n})$. $\mathbf{Case~{}II:}$ $\alpha\notin DC(v,10)$, we define a sequence $\\{V^{\theta}_{m}(n)\\}_{m=1}^{\infty}$ periodic potentials by: $\displaystyle V_{m}^{\theta}(n)=V(\theta+n\alpha),\ \ n=1,2,\cdots,m,$ $\displaystyle V_{m}^{\theta}(n+m)=V_{m}^{\theta}(n),$ such that $V_{m}^{\theta}$ is obtained from $V_{\omega}$ by “cutting” a finite piece of length $m$, and then repeating it. We denote by $\sigma_{m}(\theta)$ the spectrum of the periodic Schrödinger operators $\displaystyle Hu=u_{n+1}+u_{n-1}+V_{m}^{\theta}(n)u_{n}.$ By Theorem 1 in [42], for a.e. $\theta\in{\mathbb{T}}$, (6.9) $\displaystyle\limsup_{m\rightarrow\infty}\sigma_{m}(\theta)\subset\Sigma_{ac}(\alpha).$ We define $A^{\theta}_{q_{n}}(\xi)=\begin{pmatrix}V_{q_{n}}^{\theta}(1)&1&&&e^{-i\xi m}\\\ 1&V_{q_{n}}^{\theta}(2)&1&&\\\ &1&\ddots&\ddots&\\\ &&\ddots&\ddots&1\\\ e^{i\xi m}&&&1&V_{q_{n}}^{\theta}(q_{n})\end{pmatrix},$ $\widetilde{A}^{\theta}_{q_{n}}(\xi)=\begin{pmatrix}V(\theta+p_{n}/q_{n})&1&&&e^{-i\xi m}\\\ 1&V(\theta+2p_{n}/q_{n})&1&&\\\ &1&\ddots&\ddots&\\\ &&\ddots&\ddots&1\\\ e^{i\xi m}&&&1&V(\theta+q_{n}p_{n}/q_{n})\end{pmatrix}.$ It is standard that $\sigma_{q_{n}}(\theta)=\bigcup_{\xi}\text{Spec}(A_{q_{n}}^{\theta}(\xi)),\ \ \sigma(p_{n}/q_{n},\theta)=\bigcup_{\xi}\text{Spec}(\widetilde{A}_{q_{n}}^{\theta}(\xi)),$ where $\text{Spec}(A)$ denotes the sets of all eigenvalues of $A$. We need the following perturbation theory of matrices. ###### Proposition 4 (Corollary 12.2 of [14]). Let $A$ and $B$ be normal with $\|A-B\|=\varepsilon$. Then within a distance of $\varepsilon$ of every eigenvalue of $A$ there is at least one eigenvalue of $B$ and vice versa. Fix $\theta_{0}$ such that (6.9) holds with $\theta_{0}$ in place of $\theta.$ Notice for any $E_{0}\in\sigma_{q_{n}}(\theta_{0})$, there exists $\xi_{n}$ such that $E_{0}\in\text{Spec}(A_{q_{n}}^{\theta_{0}}(\xi_{n}))$. Applying Proposition 4 to $A_{q_{n}}^{\theta_{0}}(\xi_{n})$ and $\widetilde{A}_{q_{n}}^{\theta_{0}}(\xi_{n})$, there exists $E^{\prime}_{0}\in\text{Spec}(\widetilde{A}_{q_{n}}^{\theta_{0}}(\xi_{n}))$ such that $|E_{0}-E_{0}^{\prime}|\leq\big{\|}A_{q_{n}}^{\theta_{0}}(\xi_{n})-\widetilde{A}_{q_{n}}^{\theta_{0}}(\xi_{n})\big{\|}\leq C(V)\sum\limits_{j=1}^{q_{n}}\big{|}j(\alpha-p_{n}/q_{n})\big{|}\leq C(V)q_{n}^{2}\big{|}\alpha-p_{n}/q_{n}\big{|}.$ Since $E_{0}^{\prime}\in\sigma(p_{n}/q_{n},\theta_{0})$, it follows that (6.10) $\big{|}\sigma_{q_{n}}(\theta_{0})-\sigma\big{(}p_{n}/q_{n},\theta_{0}\big{)}\big{|}_{H}\leq C(V)q_{n}^{2}\big{|}\alpha-p_{n}/q_{n}\big{|},$ where $|A-B|_{H}$ denotes the Hausdorff distance of two sets. We denote $\sigma_{q_{n}}(\theta_{0})=\bigcup_{i=1}^{q_{n}^{\prime}}[a_{n,i},b_{n,i}]$, $q_{n}^{\prime}\leq q_{n}$. (6.10) implies that $\sigma\big{(}p_{n}/q_{n},\theta_{0}\big{)}\subset\bigcup_{i=1}^{q_{n}^{\prime}}\Big{[}a_{n,i}-C(V)q_{n}^{2}\big{|}\alpha- p_{n}/q_{n}\big{|},b_{n,i}+C(V)q_{n}^{2}\big{|}\alpha- p_{n}/q_{n}\big{|}\Big{]}.$ It follows that (6.11) $\displaystyle\big{|}\sigma(p_{n}/q_{n},\theta_{0})\backslash\sigma_{q_{n}}(\theta_{0})\big{|}\leq C(V)q_{n}^{3}|\alpha-p_{n}/q_{n}|.$ Since $\alpha\notin DC(v,10)$, by (6.11), there exists a subsequence $p_{n}/q_{n}$ such that $\displaystyle\limsup_{n\rightarrow\infty}\sigma(p_{n}/q_{n},\theta_{0})\subset\limsup_{n\rightarrow\infty}\sigma_{q_{n}}(\theta_{0})\subset\Sigma_{ac}(\alpha).$ Moreover, notice that $\displaystyle\limsup_{n\rightarrow\infty}S_{-}(p_{n}/q_{n})\subset\limsup_{n\rightarrow\infty}\sigma(p_{n}/q_{n},\theta_{0}),$ hence $\displaystyle\limsup_{n\rightarrow\infty}S_{-}(p_{n}/q_{n})\subset\Sigma_{ac}(\alpha).$ ## 7\. Proof of Proposition 3 Suppose that $V:\mathbb{T}\rightarrow{\mathbb{R}}$ is an M-ultra- differentiable function satisfying $\mathbf{(H1)}$ and $\mathbf{(H2)}$, then for almost every $E\in\Sigma_{ac}(\alpha)$, by Theorem 1.1, $(\alpha,S_{E}^{V})$ is $C^{\infty}$ rotations reducible. However, this is not enough for us to conclude $\displaystyle\|t_{p_{n}/q_{n}}(E,\theta)-a_{q_{n},0}(E)\|_{C^{0}}\leq 4\exp\\{-\Lambda(cq_{n})\\}.$ Since even we assume $(\alpha,S_{E}^{V})=(\alpha,R_{\psi(\theta)})$ with $\psi(\theta)\in C^{\infty}(\mathbb{T},\mathbb{R}),$ which only gives $\displaystyle\|t_{p_{n}/q_{n}}(E,\theta)-a_{q_{n},0}(E)\|_{C^{0}}\leq cq_{n}^{-\infty}.$ The idea is that our KAM scheme not only gives $C^{\infty}$ rotations reducibility, but also almost reducibility in the ultra-differentiable topology (Corollary 2), however, this only works for cocycles which are close to constant. Coupled with the renormalization argument, we will show that if the cocycle is $L^{2}$-conjugated to rotations and $\rho(\alpha,A)\in\mathcal{P},$ then $(\alpha,A)$ is also almost reducibility in the ultra-differentiable topology (Lemma 7.3). Consequently, Proposition 3 follows from the perturbation arguments. ### 7.1. Global almost reducibility To get desired quantitative estimates, the main method is the inverse renormalization which was first developed in [26]. We introduce the notation $\Psi:=\mathcal{J}\mathcal{R}_{\theta_{*}}^{n}(\Psi^{\prime})$ if $\Psi^{\prime}=\mathcal{R}_{\theta_{*}}^{n}(\Psi).$ It is easy to check that $\mathcal{J}\mathcal{R}_{\theta_{*}}^{n}(\Phi)=T_{\theta_{*}}^{-1}\circ N_{\widetilde{Q}_{n}^{-1}}\circ M_{\beta_{n-1}^{-1}}\circ T_{\theta_{*}}(\Phi).$ In our setting, one of the bases is the following: ###### Lemma 7.1. Let $\Phi_{1}=((1,Id),(\alpha_{n},R_{\rho_{n}}\mathrm{e}^{F(\theta)})),$ $\Phi_{2}=((1,Id),(\alpha_{n},R_{\rho_{n}}))$ where $F\in U_{r}({\mathbb{T}},sl(2,{\mathbb{R}}))$ with estimate $\|F\|_{r,1}\leq q_{n-1}^{-2}.$ Then, (7.1) $\displaystyle\|\mathcal{J}\mathcal{R}_{\theta_{*}}^{n}(\Phi_{1})-\mathcal{J}\mathcal{R}_{\theta_{*}}^{n}(\Phi_{2})\|_{\beta_{n-1}r,1}<2q_{n-1}\|F(\theta)\|_{r,1}.$ ###### Proof. Direct computation shows that $\begin{split}M_{\beta_{n-1}^{-1}}(\Phi_{1})&=((\beta_{n-1},Id),(\beta_{n-1}\alpha_{n},R_{\rho_{n}}\mathrm{e}^{F(\beta_{n-1}^{-1}\theta)})),\\\ M_{\beta_{n-1}^{-1}}(\Phi_{2})&=((\beta_{n-1},Id),(\beta_{n-1}\alpha_{n},R_{\rho_{n}})),\end{split}$ and consequently $\begin{split}N_{\widetilde{Q}_{n}^{-1}}\circ M_{\beta_{n-1}^{-1}}(\Phi_{1})&=((1,R_{\rho_{n}q_{n-1}}\mathrm{e}^{F_{1}(\theta)}),(\alpha,R_{\rho_{n}p_{n-1}}\mathrm{e}^{F_{2}(\theta)})),\\\ N_{\widetilde{Q}_{n}^{-1}}\circ M_{\beta_{n-1}^{-1}}(\Phi_{2})&=((1,R_{\rho_{n}q_{n-1}}),(\alpha,R_{\rho_{n}p_{n-1}})).\end{split}$ Note for any $\lambda\neq 0$, $\|D_{\theta}^{s}M_{\lambda}(\Phi)\|_{C^{0}([0,T])}=\lambda^{s}\|D_{\theta}^{s}\Phi\|_{C^{0}([0,T])}$, which gives (7.2) $\|M_{\lambda}(\Phi)\|_{\lambda^{-1}r,T}=\|\Phi\|_{r,\lambda T}.$ Thus the main task is to estimate of the norm under iteration of the cocycles. To do this, we need the following simple observation: ###### Lemma 7.2. _(Lemma 4.3 of[55]) _ Let $F_{i}\in U_{r}({\mathbb{R}},sl(2,{\mathbb{R}})),i=1,\cdots,j.$ Then it holds that $R_{\rho}\mathrm{e}^{F_{j}(\theta)}R_{\rho}\mathrm{e}^{F_{j-1}(\theta)}\cdots R_{\rho}\mathrm{e}^{F_{1}(\theta)}=R_{j\rho}\mathrm{e}^{\widetilde{F}(\theta)},$ with estimate $\|\widetilde{F}\|_{r,T}\leq\sum_{i=1}^{j}\|F_{i}\|_{r,T}.$ By (7.2) and Lemma 7.2, we have $\|F_{1}\|_{\beta_{n-1}r,1}\leq q_{n-1}\|F\|_{r,1},\qquad\|F_{2}\|_{\beta_{n-1}r,1}\leq p_{n-1}\|F\|_{r,1}.$ Then (7.1) follows directly. ∎ ###### Lemma 7.3. Assume that $(\alpha,A)$ is $L^{2}$-conjugated to rotations and homotopic to the identity. If $\rho_{f}=\rho(\alpha,A)\in\mathcal{P}.$ Then there exist $B_{j,\ell},F_{j,\ell}\in U_{\widetilde{r}_{\ell}}(\mathbb{T},SL(2,\mathbb{R}))$ such that (7.3) $\displaystyle B_{j,\ell}(\theta+\alpha)A(\theta)B_{j,\ell}(\theta)^{-1}=R_{\rho_{f}}\mathrm{e}^{F_{j,\ell}(\theta)},\quad\ell>\widehat{n},$ where $\widetilde{r}_{\ell}=r\beta_{n_{j}-1}/2K_{*}^{3}\overline{Q}_{\ell-1}^{2},$ and $\widehat{n}\in{\mathbb{N}}$ is the smallest one such that $\overline{Q}_{\widehat{n}}\geq C^{q_{n_{j}}^{2}}.$ Moreover, we have estimate $\|F_{j,\ell}\|_{\widetilde{r}_{\ell}}\leq\varepsilon_{\ell}^{\frac{1}{2}},\qquad\|B_{j,\ell}\|_{\widetilde{r}_{\ell}}<4C^{3q_{n_{j}-1}q_{n_{j}}}.$ ###### Proof. Since $(\alpha,A)$ is $L^{2}$-conjugated to rotations and homotopic to the identity, by Proposition 1 there exists $D_{n}\in U_{rK_{*}^{-2}}(\mathbb{R},SL(2,\mathbb{R}))$ with (7.4) $\|D_{n}\|_{r(K_{*}^{2}T)^{-1},T}\leq C^{q_{n-1}(T+1)},$ such that $\displaystyle\mathrm{Conj_{D_{n}}}(\mathcal{R}_{\theta_{*}}^{n}(\Phi))=((1,Id),(\alpha_{n},\ R_{\rho_{n}}\mathrm{e}^{F_{n}})),$ with $\|F_{n}\|_{rK_{*}^{-2},1}\rightarrow 0.$ Since $\rho_{f}=\rho(\alpha,A)\in\mathcal{P},$ which means $\rho_{f}\beta_{n_{j}-1}^{-1}\in DC_{\alpha_{n_{j}}}(\gamma,\tau)$ for infinitely many $n_{j}$, then we can further choose $j$ large enough, such that $\|F_{n_{j}}\|_{rK_{*}^{-2}}\leq\varepsilon_{*}(\gamma,\tau,rK_{*}^{-2},M),$ where $\varepsilon_{*}=\varepsilon_{*}(\gamma,\tau,r,M)>0$ is the one in Theorem 1.2. In the following, we will write $D_{j}$ for $D_{n_{j}}$ for short, and denote $T_{n}=\beta_{n-1}^{-1}=q_{n}+\alpha_{n}q_{n-1}$, $r_{\ell}=2^{-1}rK_{*}^{-2}\overline{Q}_{\ell-1}^{-2}$, then $\widetilde{r}_{\ell}=(K_{*}T_{n_{j}})^{-1}r_{\ell}$. Now we apply Corollary 2 to the action $\mathrm{Conj_{D_{j}}}(\mathcal{R}_{\theta_{*}}^{n_{j}}(\Phi))$ and denote $Z_{j,\ell}=B_{\ell}D_{j},$ then (7.5) $\displaystyle\mathrm{Conj_{Z_{j,\ell}}}(\mathcal{R}_{\theta_{*}}^{n_{j}}(\Phi))=\widetilde{\Phi}_{1}^{(j,\ell)}:=((1,Id),(\alpha_{n_{j}},\ R_{\rho_{n_{j}}}\mathrm{e}^{F_{j,\ell}})),$ which, together with (5.2) and (7.4) and the fact $r_{\ell}\ll rK_{*}^{-2}T_{n_{j}}^{-1}$, implies (7.6) $\displaystyle\|F_{j,\ell}\|_{r_{\ell},1}\leq\varepsilon_{\ell},\ \|Z_{j,\ell}\|_{r_{\ell},T_{n_{j}}}\leq\|D_{j}\|_{r_{\ell},T_{n_{j}}}\|B_{\ell}\|_{r_{\ell},1}\leq C^{3q_{n_{j}-1}q_{n_{j}}}.$ Once we have these, we can set $\widetilde{\Phi}_{2}^{(j,\ell)}=((1,Id),(\alpha_{n_{j}},\ R_{\rho_{n_{j}}}))$ and define $\Phi_{2,j,\ell}$ by (7.7) $\displaystyle\mathrm{Conj_{Z_{j,\ell}}}(\mathcal{R}_{\theta_{*}}^{n_{j}}(\Phi_{2,j,\ell}))=\widetilde{\Phi}_{2}^{(j,\ell)}.$ For given $G\in SL(2,\mathbb{R}),$ $T_{\theta_{*}}$ and $N_{U}$ commute with $\mathrm{Conj_{G}}$ while $M_{\lambda}\circ\mathrm{Conj_{G}}=\mathrm{Conj_{G(\lambda\cdot)}}\circ M_{\lambda},$ then by (7.5) and (7.7), we get (7.8) $\displaystyle\left\\{\begin{array}[]{l l}\Phi=\mathcal{J}\mathcal{R}_{\theta_{*}}^{n_{j}}\big{(}\mathrm{Conj_{Z_{j,\ell}^{-1}}}(\widetilde{\Phi}_{1}^{(j,\ell)})\big{)}=\mathrm{Conj_{Z_{j,\ell}^{-1}(\beta_{n_{j}-1}^{-1}\cdot)}}\big{(}\widetilde{\Phi}_{1,j,\ell}\big{)},\\\ \Phi_{2,j,\ell}=\mathcal{J}\mathcal{R}_{\theta_{*}}^{n_{j}}\big{(}\mathrm{Conj_{Z_{j,\ell}^{-1}}}(\widetilde{\Phi}_{2}^{(j,\ell)})\big{)}=\mathrm{Conj_{Z_{(j,\ell)}^{-1}(\beta_{n_{j}-1}^{-1}\cdot)}}\big{(}\widetilde{\Phi}_{2,j,\ell}\big{)},\end{array}\right.$ where (7.9) $\displaystyle\left\\{\begin{array}[]{l l}\widetilde{\Phi}_{1,j,\ell}&=\mathcal{J}\mathcal{R}_{\theta_{*}}^{n_{j}}(\widetilde{\Phi}_{1}^{(j,\ell)}),\\\ \widetilde{\Phi}_{2,j,\ell}&=\mathcal{J}\mathcal{R}_{\theta_{*}}^{n_{j}}(\widetilde{\Phi}_{2}^{(j,\ell)})=((1,R_{\rho_{n_{j}}q_{n_{j}-1}}),(\alpha,R_{\rho_{n_{j}}p_{n_{j}-1}})).\end{array}\right.$ Set $\widetilde{Z}_{j}(\theta)=R_{-\rho_{n_{j}}q_{n_{j}-1}\theta},$ then $\displaystyle\mathrm{Conj_{\widetilde{Z}_{j}}}\big{(}\widetilde{\Phi}_{2,j,\ell}\big{)}=((1,Id),(\alpha,R_{\rho_{f}})):=\Phi_{**},$ which, together with (7.8) and (7.9), implies (7.10) $\displaystyle\left\\{\begin{array}[]{l l}\Phi=\mathrm{Conj_{Z_{j,\ell}^{-1}(\beta_{n_{j}-1}^{-1}\cdot)\widetilde{Z}_{j}^{-1}}}(\Phi_{*}),\\\ \Phi_{2,j,\ell}=\mathrm{Conj_{Z_{j,\ell}^{-1}(\beta_{n_{j}-1}^{-1}\cdot)\widetilde{Z}_{j}^{-1}}}(\Phi_{**}),\end{array}\right.$ where $\Phi_{*}=\mathrm{Conj_{\widetilde{Z}_{j}}}\big{(}\widetilde{\Phi}_{1,j,\ell}\big{)}.$ Moreover, by our selection $\overline{Q}_{\ell}>\overline{Q}_{\widehat{n}}\geq C^{q_{n_{j}}^{2}}.$ (7.11) $\|\widetilde{Z}_{j}\|_{r_{\ell},1}\leq 2,$ $\|F_{j,\ell}\|_{r_{\ell},1}\leq\varepsilon_{\ell}\ll q_{n_{j}}^{-6}.$ In the following, we will give the estimate of the distance of $\Phi$ and $\Phi_{2,j,\ell}.$ First, we apply Lemma 7.1 with $\widetilde{\Phi}_{i}^{(j,\ell)}$ in place of $\Phi_{i},$ and $n_{j}$ in place of $n$ respectively, then by (7.1) $\displaystyle\|\widetilde{\Phi}_{1,j,\ell}-\widetilde{\Phi}_{2,j,\ell}\|_{T_{n_{j}}^{-1}r_{\ell},1}<\|F_{j,\ell}\|_{r_{\ell},1}^{\frac{3}{4}}\leq\varepsilon_{\ell}^{\frac{3}{4}}.$ Finally, by (7.6) and (7.8) and the inequality above we get (7.12) $\displaystyle\|\Phi-\Phi_{2,j,\ell}\|_{T_{n_{j}}^{-1}r_{\ell},1}$ $\displaystyle\leq$ $\displaystyle\|Z_{j,l}\|_{r_{\ell},T_{n_{j}}}^{2}\|\widetilde{\Phi}_{1,j,\ell}-\widetilde{\Phi}_{2,j,\ell}\|_{T_{n_{j}}^{-1}r_{\ell},1}$ $\displaystyle\leq$ $\displaystyle C^{6q_{n_{j}-1}q_{n_{j}}}\varepsilon_{\ell}^{\frac{3}{4}}.$ Notice that $\Phi_{2,j,\ell}$ may not be normalized, however, by (7.12) we know that (7.13) $\displaystyle\|\Phi_{2,j,\ell}(1,0)-Id\|_{T_{n_{j}}^{-1}r_{\ell},1}$ $\displaystyle=$ $\displaystyle\|\Phi_{2,j,\ell}(1,0)-\Phi(1,0)\|_{T_{n_{j}}^{-1}r_{\ell},1}$ $\displaystyle\leq$ $\displaystyle C^{6q_{n_{j}-1}q_{n_{j}}}\varepsilon_{\ell}^{\frac{3}{4}}.$ Thus by Lemma 2.3, there exists a conjugation $\widetilde{B}_{j,\ell}\in U_{\widetilde{r}_{\ell}}({\mathbb{R}},SL(2,{\mathbb{R}}))$ such that $\overline{\Phi}_{j,\ell}=\mathrm{Conj_{\widetilde{B}_{j,\ell}}}(\Phi_{2,j,\ell})$ is a normalized action. Moreover, we have estimate (7.14) $\displaystyle\|\widetilde{B}_{j,\ell}-Id\|_{\widetilde{r}_{\ell},1}\leq\|\Phi_{2,j,\ell}(1,0)-Id\|_{T_{n_{j}}^{-1}r_{\ell},1}\leq C^{6q_{n_{j}}q_{n_{j}-1}}\varepsilon_{\ell}^{\frac{3}{4}}.$ Since $\overline{\Phi}_{j,\ell}(0,1)=\widetilde{B}_{j,\ell}(\theta+\alpha)\Phi_{2,j,\ell}(0,1)\widetilde{B}_{j,\ell}(\theta)^{-1},$ by (7.13), (7.14) we have $\displaystyle\|\overline{\Phi}_{j,\ell}-\Phi_{2,j,\ell}\|_{\widetilde{r}_{\ell},1}\leq 2C^{6q_{n_{j}}q_{n_{j}-1}}\varepsilon_{\ell}^{\frac{3}{4}}.$ The inequality above, together with (7.12), yields (7.15) $\|\Phi-\overline{\Phi}_{j,\ell}\|_{\widetilde{r}_{\ell},1}\leq 3C^{6q_{n_{j}}q_{n_{j}-1}}\varepsilon_{\ell}^{\frac{3}{4}}.$ Set $B_{j,\ell}(\cdot)=\widetilde{Z}_{j}(\cdot)Z_{j,\ell}(\beta_{n_{j}-1}^{-1}\cdot)\widetilde{B}_{j,\ell}^{-1}(\cdot),$ then by (7.10) we get (7.16) $\displaystyle\mathrm{Conj_{B_{j,\ell}}}(\overline{\Phi}_{j,\ell})=\mathrm{Conj_{\widetilde{Z}_{j}Z_{j,\ell}(\beta_{n_{j}-1}^{-1}\cdot)}}(\Phi_{2,j,\ell})=\Phi_{**}.$ Thus $B_{j,\ell}$ is $1-$periodic since both $\overline{\Phi}_{j,\ell}$ and $\Phi_{**}$ are normalized. Moreover, (7.6), (7.11) and (7.14) imply (7.17) $\displaystyle\|B_{j,\ell}\|_{\widetilde{r}_{\ell},1}\leq\|\widetilde{Z}_{j}\|_{\widetilde{r}_{\ell},1}\|Z_{j,\ell}\|_{\widetilde{r}_{\ell},T_{n_{j}}}\|\widetilde{B}_{j,\ell}\|_{\widetilde{r}_{\ell},1}\leq 4C^{3q_{n_{j}-1}q_{n_{j}}}.$ By (7.15)-(7.17) we get, $\displaystyle\|(0,$ $\displaystyle B_{j,\ell}(\cdot+\alpha))\circ(\alpha,A)\circ(0,B_{j,\ell})^{-1}-(\alpha,R_{\rho_{f}})\|_{\widetilde{r}_{\ell},1}$ $\displaystyle\leq\|\mathrm{Conj_{B_{j,\ell}}}(\Phi)-\mathrm{Conj_{B_{j,\ell}}}(\overline{\Phi}_{j,\ell})\|_{\widetilde{r}_{\ell},1}\leq C^{13q_{n_{j}-1}q_{n_{j}}}\varepsilon_{\ell}^{\frac{3}{4}}\leq\varepsilon_{\ell}^{\frac{1}{2}},$ where the last inequality follows from our selection $\overline{Q}_{\ell}>\overline{Q}_{\widehat{n}}\geq C^{q_{n_{j}}^{2}}$ and definition of $\varepsilon_{\ell}$. Thus, by implicit function theorem, there exists a unique $F_{j,\ell}\in U_{\widetilde{r}_{\ell}}({\mathbb{T}},sl(2,{\mathbb{R}})),$ such that $\displaystyle B_{j,\ell}(\theta+\alpha)A(\theta)B_{j,\ell}(\theta)^{-1}=R_{\rho_{f}}\mathrm{e}^{F_{j,\ell}(\theta)}$ with $\|F_{j,\ell}\|_{\widetilde{r}_{\ell}}\leq\varepsilon_{\ell}^{\frac{1}{2}}.$ ∎ ### 7.2. Proof of Proposition 3 First we construct the desired sequence. For the sequence $(q_{\ell})_{\ell\in\mathbb{N}}$ and subsequence $(Q_{\ell})_{\ell\in\mathbb{N}}$ constructed in Lemma 2.1, first set $n_{1}\geq n_{0}$ to be the smallest integer such that (7.18) $\max\\{16C_{M}C^{6q_{n_{j}}^{2}}(2+2\|V\|_{r}),16r^{-1}q_{n_{j}}K_{*}^{3}\\}\leq\overline{Q}_{n_{1}},$ where $n_{0}$ is the one in section 4. Then we set $n_{*}$ be the smallest integer number such that (7.19) $\displaystyle q_{n_{*}}>\overline{Q}_{n_{1}+1}^{2\mathbb{A}^{4}\tau^{2}}.$ Then, for any fixed $n$ with $n\geq n_{*},$ we set $\ell\in\mathbb{N}$ be the smallest integer number such that $\overline{Q}_{\ell}^{2\mathbb{A}^{4}\tau^{2}}\geq q_{n}.$ That is (7.20) $\displaystyle\overline{Q}_{\ell-1}^{2\mathbb{A}^{4}\tau^{2}}<q_{n}\leq\overline{Q}_{\ell}^{2\mathbb{A}^{4}\tau^{2}}.$ Thus by (7.19) and (7.20) we get $\ell-1\geq n_{1}\geq\widehat{n}$, where $\widehat{n}$ is the one defined in Lemma 7.3. By our construction, for almost every $E\in\Sigma_{ac}(\alpha)$, $(\alpha,S_{E}^{V})$ is $L^{2}$-conjugated to rotations, and $\rho_{f}=\rho(\alpha,S_{E}^{V})\in\mathcal{P}.$ Then by (7.3) in Lemma 7.3, there exist $B_{j,\ell},F_{j,\ell}\in U_{\widetilde{r}_{\ell}}(\mathbb{T},SL(2,\mathbb{R}))$ such that (7.21) $\displaystyle B_{j,\ell}(\theta+\alpha)S_{E}^{V}(\theta)B_{j,\ell}(\theta)^{-1}=R_{\rho_{f}}\mathrm{e}^{F_{j,\ell}},$ with estimate $\|F_{j,\ell}\|_{\widetilde{r}_{\ell}}\leq\varepsilon_{\ell}^{\frac{1}{2}},\qquad\|B_{j,\ell}\|_{\widetilde{r}_{\ell}}<4C^{3q_{n_{j}-1}q_{n_{j}}}.$ We shorten $B_{j,\ell}$ and $F_{j,\ell}$ as $B$ and $F,$ respectively. By (7.21) we get (7.22) $\displaystyle B(\theta+p_{n}/q_{n})S_{E}^{V}(\theta)B(\theta)^{-1}=R_{\rho_{f}}+f(\theta),$ where $\displaystyle f(\theta)$ $\displaystyle=$ $\displaystyle R_{\rho_{f}}(\mathrm{e}^{F(\theta)}-I)+\\{B(\theta+p_{n}/q_{n})-B(\theta+\alpha)\\}S_{E}^{V}(\theta)B(\theta)^{-1}$ $\displaystyle=$ $\displaystyle(\textrm{I})+(\textrm{II}).$ Note for any $E\in\Sigma_{ac}(\alpha)\subset\Sigma(\alpha)$, we have $|E|<2+\|V\|_{r}$, thus by Cauchy’s estimate (Lemma 3.2), we get (7.23) $\displaystyle\|\textrm{II}\|_{\widetilde{r}_{\ell}/2}$ $\displaystyle\leq$ $\displaystyle|\alpha-\frac{p_{n}}{q_{n}}|\|\partial B\|_{\widetilde{r}_{\ell}/2}\|S_{E}^{V}\|_{\widetilde{r}_{\ell}}\|B^{-1}\|_{\widetilde{r}_{\ell}}$ $\displaystyle\leq$ $\displaystyle C_{M}\widetilde{r}_{\ell}^{-1}q_{n}^{-2}\|B^{-1}\|_{\widetilde{r}_{\ell}}\|B\|_{\widetilde{r}_{\ell}}(2+2\|V\|_{r})$ $\displaystyle\leq$ $\displaystyle 16C_{M}C^{6q_{n_{j}-1}q_{n_{j}}}(2+2\|V\|_{r})\widetilde{r}_{\ell}^{-1}q_{n}^{-2}.$ Since $B$ is 1-periodic, by (7.22), we have $\displaystyle B(\theta$ $\displaystyle+q_{n}p_{n}/q_{n})\Pi_{s=q_{n}-1}^{0}S_{E}^{V}(\theta+sp_{n}/q_{n})B(\theta)^{-1}$ $\displaystyle=\Pi_{s=q_{n}-1}^{0}B(\theta+(s+1)p_{n}/q_{n})S_{E}^{V}(\theta+sp_{n}/q_{n})B(\theta+sp_{n}/q_{n})^{-1}$ $\displaystyle=\Pi_{s=q_{n}-1}^{0}\\{R_{\rho_{f}}+f(\theta+sp_{n}/q_{n})\\}.$ As a consequence, $\displaystyle\mathrm{tr}\Pi_{s=q_{n}-1}^{0}S_{E}^{V}(\theta+sp_{n}/q_{n})=\mathrm{tr}\Pi_{s=q_{n}-1}^{0}\\{R_{\rho_{f}}+f(\theta+sp_{n}/q_{n})\\},$ which, together with (6.1) and (6.2), implies (7.24) $\displaystyle t_{p_{n}/q_{n}}(E,\theta)=\mathrm{tr}\Pi_{s=q_{n}-1}^{0}\\{R_{\rho_{f}}+f(\theta+sp_{n}/q_{n})\\}=\sum_{k\in\mathbb{Z}}a_{q_{n},k}(E)\mathrm{e}^{2\pi\mathrm{i}kq_{n}\theta}.$ The first equality in (7.24) implies (7.25) $\displaystyle\|t_{p_{n}/q_{n}}(E,\theta)\|_{\widetilde{r}_{\ell}/2}\leq 2\\{1+\|f\|_{\widetilde{r}_{\ell}/2}\\}^{q_{n}}.$ In the following we will give the estimate of $\|f\|_{\widetilde{r}_{\ell}/2},$ indeed, by (7.18) and $\ell-1\geq n_{1}$, we have (7.26) $\widetilde{r}_{\ell}^{-1}=2r^{-1}\beta_{n_{j}-1}^{-1}K_{*}^{3}\overline{Q}_{\ell-1}^{2}<4^{-1}\overline{Q}_{\ell-1}^{3}.$ Again, by (4.6), (7.18), (7.20) and (7.23), we have $\displaystyle\|f\|_{\widetilde{r}_{\ell}/2}\leq\|\textrm{I}\|_{\widetilde{r}_{\ell}/2}+\|\textrm{II}\|_{\widetilde{r}_{\ell}/2}\leq\frac{1}{2}\overline{Q}_{\ell}^{-4\mathbb{A}^{4}\tau^{2}}+\frac{1}{2}\overline{Q}_{\ell-1}^{4}q_{n}^{-2}\leq q_{n}^{-2(1-\mathbb{A}^{-4}\tau^{-2})}.$ Consequently, by (7.25) (7.27) $\displaystyle\|t_{p_{n}/q_{n}}(E,\theta)\|_{\widetilde{r}_{\ell}/2}\leq 2\\{1+q_{n}^{-2(1-\mathbb{A}^{-4}\tau^{-2})}\\}^{q_{n}}<4.$ On the other hand, by (7.26), we have $\displaystyle q_{n}2^{-1}\widetilde{r}_{\ell}>q_{n}\overline{Q}_{\ell-1}^{-3}>q_{n}\overline{Q}_{\ell-1}^{-4}>q_{n}^{1-2\mathbb{A}^{-4}\tau^{-2}}>T_{1}.$ Moreover, the second equality in (7.24) implies $\displaystyle t_{p_{n}/q_{n}}(E,\theta)-a_{q_{n},0}(E)=\mathcal{R}_{q_{n}}t_{p_{n}/q_{n}}(\theta,E).$ Thus by Lemma 3.4 and (7.27), we have $\displaystyle\|t_{p_{n}/q_{n}}(E,\theta)-a_{q_{n},0}(E)\|_{C^{0}}$ $\displaystyle\leq\|t_{p_{n}/q_{n}}(E,\theta)\|_{2^{-1}\widetilde{r}_{\ell}}\exp\\{-\Lambda(\pi q_{n}2^{-1}\widetilde{r}_{\ell})\\}$ $\displaystyle\leq 4\exp\\{-\Lambda(q_{n}^{1-2\mathbb{A}^{-4}\tau^{-2}})\\},$ the last inequality follows from the fact that $\Lambda(\cdot)$ is non- decreasing on $\mathbb{R}^{+}.$ ∎ ## 8\. Proof of Theorem 6.3 In this section we give the proof of Theorem 6.3 which is based on the large deviation theorem and avalanche principle. Notice that the cocycle in Theorem 6.3 is $\nu$-Gevrey with $1/2<\nu<1$, we will follow the method in [37] to approximate the Gevrey cocycle by its truncated cocycle which is analytic in the certain strip. For the continuity argument, our scheme is in the spirit of [16] with some modifications. Compared to the result in [37] with $1/2<\nu<1$, our large deviation theorem also works for more general cocycles (other than Schrödinger coycle) and rational frequencies, which is an analogue of Bourgain-Jitomirskaya [16]. More concretely, if we truncate the cocycle $A(\alpha,\theta)$ to $\widetilde{A}(\alpha,\theta)$, and denote $\widetilde{A}_{N}(\alpha,\theta)$ to be the transfer matrix, then $\det\widetilde{A}_{N}(\alpha,\theta)$ depends on $\theta$, is not constant anymore (of course not identical to 1), thus we have to prove that the subharmonic extension of $N^{-1}\ln\|\widetilde{A}_{N}(\alpha,\theta)\|$ is bounded. This boundedness will enable us to give an enhanced version of the large deviation bound shown in [16]. For more results and methods to prove the continuity of the Lyapunov exponents, we refer readers to [4, 32, 36, 34]. ### 8.1. Large deviation theorem. In this subsection we give a large deviation theorem for the $\nu$-Gevrey cocycle with $1/2<\nu<1.$ Let $A_{N}(\alpha,\theta)$ and $L_{N}(\alpha,A)$ be the associated transfer matrix and finite Lyapunov exponent of the cocycle $(\alpha,A)$. Then we have the following: ###### Proposition 5. Let $\rho>0$, $\frac{1}{2}<\nu<1,$ $0<\kappa<1.$ Assume that $A\in G_{\rho}^{\nu}(\mathbb{T},SL(2,\mathbb{R}))$, and $\displaystyle\big{|}\alpha-\frac{a}{q}\big{|}<\frac{1}{q^{2}},\ \ (a,q)=1.$ Then there exist $c,C_{i}(\kappa)>0,i=1,2$, $\sigma_{1}>\sigma>1>\gamma>0$ and $q_{0}(\kappa,\rho,\nu)\in{\mathbb{N}}^{+}$ such that for $q\geq q_{0}$, $C_{1}(\kappa)q^{\sigma}<N<C_{2}(\kappa)q^{\sigma_{1}}$, $\displaystyle mes\Big{\\{}\theta:\big{|}\frac{1}{N}\ln\|A_{N}(\alpha,\theta)\|-L_{N}(\alpha,A)\big{|}>\kappa\Big{\\}}<\mathrm{e}^{-cq^{\gamma}}.$ #### 8.1.1. Averages of shifts of subharmonic functions. Let $u=u(\theta)$ be a function on ${\mathbb{T}}$ having a subharmonic extension on the strip $[|\mathrm{Im}\vartheta|\leq\rho]$, and $\alpha\in{\mathbb{T}}.$ We prove that the mean of $u$ is close to the Fejér average of $u(\theta)$ for $\theta$ outside a small set (here being ‘close’ or ‘small’ is expressed in terms of the number of shifts considered). Consider the Fejér kernel of order p: (8.1) $K_{R}^{p}(t)=\big{(}\frac{1}{R}\sum\limits_{j=0}^{R-1}\mathrm{e}^{2\pi\mathrm{i}jt}\big{)}^{p},$ then we have $\begin{split}\big{|}K_{R}^{p}(t)\big{|}=\frac{1}{R^{p}}\big{|}\frac{1-\mathrm{e}^{2\pi\mathrm{i}Rt}}{1-\mathrm{e}^{2\pi\mathrm{i}t}}\big{|}^{p}\leq\frac{1}{R^{p}\|t\|_{{\mathbb{Z}}}^{p}}.\end{split}$ Notice also $\big{|}K_{R}^{p}(t)\big{|}\leq 1$, we have (8.2) $\begin{split}|K_{R}^{p}(t)|\leq\min\big{\\{}1,\frac{1}{R^{p}\|t\|_{\mathbb{Z}}^{p}}\big{\\}}\leq\frac{2}{1+R^{p}\|t\|_{\mathbb{Z}}^{p}}.\end{split}$ We can rewrite (8.1) as $\begin{split}K_{R}^{p}(t)=\frac{1}{R^{p}}\sum\limits_{j=0}^{p(R-1)}c_{R}^{p}(j)\mathrm{e}^{2\pi\mathrm{i}jt},\end{split}$ where $c^{p}_{R}(j)$ are positive integers so that $\begin{split}\frac{1}{R^{p}}\sum\limits_{j=0}^{p(R-1)}c^{p}_{R}(j)=1.\end{split}$ Notice that if $p=1$ then $K_{R}^{1}(t)=\frac{1}{R}\sum\limits_{j=0}^{R-1}\mathrm{e}^{2\pi\mathrm{i}jt}$, thus $c^{1}_{R}(j)=1$ for all $j.$ ###### Proposition 6. Let $\rho>0.$ Assume that $u:{\mathbb{T}}\rightarrow{\mathbb{R}}$ has a bounded subharmonic extension to the strip $[|\mathrm{Im}\vartheta|\leq\rho]$ and $\|u\|_{C^{0}}\leq S.$ If $\displaystyle\big{|}\alpha-\frac{a}{q}\big{|}<\frac{1}{q^{2}},\ \ (a,q)=1,$ and $\sigma>1,0<\varsigma<\sigma^{-1},\varsigma(1-\sigma^{-1})^{-1}<p<(\sigma-1)^{-1},$ then $\displaystyle mes\Big{\\{}\theta:\Big{|}\frac{1}{R^{p}}\sum\limits_{j=0}^{p(R-1)}c_{R}^{p}(j)u(\theta+j\alpha)-[u(\theta)]_{\theta}\Big{|}>\varsigma_{2}R^{-\varsigma_{1}}\Big{\\}}<\frac{R^{2\varsigma_{1}}}{2^{8}\exp\\{R^{\varsigma_{3}}\\}},$ provided $R=q^{\sigma}\geq q(\varsigma,p,\sigma)\ (\varsigma_{1}=p(1-\sigma^{-1}),\varsigma_{2}=2^{p+5}S\rho^{-1},\varsigma_{3}=\frac{1+p}{\sigma}-p).$ ###### Proof. The proof is divided into the following 3 steps. 1\. Shift of Fejér average. Since $u$ is subharmonic in the strip $[|\mathrm{Im}\vartheta|\leq\rho]$, then from Corollary 4.7 in [17], we get (8.3) $\displaystyle|\widehat{u}(k)|\leq\frac{S}{\rho|k|}.$ Consider the Fejér average of $u_{N}(\theta)$, and notice that $\displaystyle u(\theta+j\alpha)=[u(\theta)]_{\theta}+\sum\limits_{k\neq 0}\widehat{u}(k)\mathrm{e}^{2\pi\mathrm{i}k(\theta+j\alpha)},$ thus, by shortening $K_{R}^{p}(\cdot)$ as $K_{R}(\cdot),$ we get (8.4) $\displaystyle\frac{1}{R^{p}}\sum\limits_{j=0}^{p(R-1)}c_{R}^{p}(j)u(\theta+j\alpha)-[u(\theta)]_{\theta}$ $\displaystyle=$ $\displaystyle\sum\limits_{k\neq 0}\widehat{u}(k)\Big{(}\frac{1}{R^{p}}\sum\limits_{j=0}^{p(R-1)}c_{R}^{p}(j)\mathrm{e}^{2\pi\mathrm{i}jk\alpha}\Big{)}\mathrm{e}^{2\pi\mathrm{i}k\theta}$ $\displaystyle=$ $\displaystyle\sum\limits_{k\neq 0}\widehat{u}(k)\cdot K_{R}(k\alpha)\mathrm{e}^{2\pi\mathrm{i}k\theta}:=w(\theta)=\mathcal{T}_{K}w(\theta)+\mathcal{R}_{K}w(\theta),$ where $K>q$ is a large constant that will be determined later. 2\. Estimate of the $w(\theta)$. In the following, we will give the estimate of $w(\theta)$. Let $I_{\ell}=[\frac{q}{4}\ell,\frac{q}{4}(\ell+1)),$ then we write $\mathcal{T}_{K}w(\theta)$ as $\begin{split}\mathcal{T}_{K}w(\theta)=\sum\limits_{\ell=0}^{[4Kq^{-1}]+1}\sum\limits_{k\in I_{\ell}}\widehat{u}(k)\cdot K_{R}(k\alpha)\mathrm{e}^{2\pi\mathrm{i}k\theta}.\end{split}$ Note $\big{|}\alpha-\frac{a}{q}\big{|}<\frac{1}{q^{2}},$ it follows that for $|k|\leq\frac{q}{2}$ with $k\neq 0$, we have $|k\alpha-\frac{ka}{q}|<\frac{1}{2q},$ hence $\|k\alpha\|_{{\mathbb{Z}}}>\frac{1}{2q}$. Let $\alpha_{1},\cdots,\alpha_{q/4}$ be the decreasing rearrangement of $\\{\|k\alpha\|_{{\mathbb{Z}}}^{-1}\\}_{0<|k|\leq\frac{q}{4}}$. Then we have $\alpha_{i}\leq\frac{2q}{i}$. Moreover, for any interval of length $q/4$, same is true for $\\{\|k\alpha\|_{{\mathbb{Z}}}^{-1}\\}_{|k|\in I}$ if we exclude at most one value of $k$. By (8.2) and (8.3), we have (8.5) $\begin{split}\sum\limits_{0<|k|<\frac{q}{4}}\big{|}\widehat{u}(k)K_{R}(k\alpha)\big{|}\leq\sum\limits_{0<|k|<\frac{q}{4}}\frac{S\|k\alpha\|_{{\mathbb{Z}}}^{-p}}{|k|\rho R^{p}}\leq\sum\limits_{1\leq i<\frac{q}{4}}\frac{2S(2q/i)^{p}}{\rho R^{p}}\leq\frac{2^{p+3}S}{\rho}\left(\frac{q}{R}\right)^{p},\end{split}$ and for each $\ell\geq 1,$ we have $\displaystyle\sum\limits_{|k|\in I_{\ell}}\big{|}\widehat{u}(k)K_{R}(k\alpha)\big{|}\leq\frac{2S}{\frac{q}{4}\rho\ell}\Big{(}1+\sum\limits_{1\leq i<\frac{q}{4}}\frac{(2q/i)^{p}}{R^{p}}\Big{)}\leq\frac{8S}{\rho q\ell}\big{(}1+c(q/R)^{p}\big{)}.$ Thus we have $\displaystyle\Big{|}\sum\limits_{\ell=1}^{[4Kq^{-1}]+1}\sum\limits_{|k|\in I_{\ell}}\widehat{u}(k)K_{R}(k\alpha)\mathrm{e}^{2\pi\mathrm{i}k\theta}\Big{|}$ $\displaystyle\leq\sum\limits_{\ell=1}^{[4Kq^{-1}]+1}\frac{8S}{\rho q\ell}\big{(}1+c(q/R)^{p}\big{)}$ $\displaystyle\leq\frac{8S}{\rho q}\big{(}1+c(q/R)^{p}\big{)}\ln[4Kq^{-1}+1]$ (8.6) $\displaystyle\leq\frac{8S}{\rho q}\big{(}1+c(q/R)^{p}\big{)}\ln K.$ On the other hand, again by (8.3), we have (8.7) $\begin{split}\|\mathcal{R}_{K}w(\theta)\|^{2}_{\ell^{2}}\leq\sum\limits_{|k|\geq K}\frac{S^{2}}{(\rho|k|)^{2}}\leq\frac{S^{2}}{\rho^{2}}K^{-1}.\end{split}$ 3\. Choose approximate $K,p$. Now we can finish the proof of Proposition 6. By (8.4)-(8.1.1) , we have $\displaystyle\big{|}[u(\theta)]_{\theta}$ $\displaystyle-\frac{1}{R^{p}}\sum\limits_{j=0}^{p(R-1)}c_{R}^{p}(j)u(\theta+j\alpha)\big{|}$ $\displaystyle\leq\frac{2^{p+3}}{\rho}\left(\frac{q}{R}\right)^{p}+\frac{8S}{\rho q}\Big{(}1+c\left(\frac{q}{R}\right)^{p}\Big{)}\ln K+|\mathcal{R}_{K}w(\theta)|.$ Take $q=R^{\sigma^{-1}},\sigma>1,$ $K=\exp\\{R^{\sigma^{-1}-p(1-\sigma^{-1})}\\}$ and $0<\varsigma<\sigma^{-1},\varsigma(1-\sigma^{-1})^{-1}<p<(\sigma-1)^{-1}.$ Once we fixed the parameters above, we have, $\displaystyle\Big{|}\frac{2^{p+3}S}{\rho}\left(\frac{q}{R}\right)^{p}\Big{|}=2^{p+3}S\rho^{-1}R^{-\varsigma_{1}},$ $\displaystyle\Big{|}\frac{8S}{\rho q}\Big{(}1+c\Big{(}\frac{q}{R}\Big{)}^{p}\Big{)}\ln K\Big{|}$ $\displaystyle\leq 16\rho^{-1}Sq^{-1}\ln K=16S\rho^{-1}R^{-\varsigma_{1}},$ where $\varsigma_{1}=p(1-\sigma^{-1})$. By Chebyshev’s inequality and (8.7), one has $\displaystyle mes\Big{\\{}\theta:|\mathcal{R}_{K}w(\theta)|$ $\displaystyle>2^{4}S\rho^{-1}R^{-\varsigma_{1}}\Big{\\}}\leq(2^{4}S\rho^{-1}R^{-\varsigma_{1}})^{-2}\|\mathcal{R}_{K}w\|^{2}_{\ell^{2}}$ $\displaystyle\leq 2^{-8}R^{2\varsigma_{1}}\exp\\{-R^{\varsigma_{3}}\\},$ where $\varsigma_{3}=(1+p)\sigma^{-1}-p$. By the above argument, the desired result follows directly. ∎ #### 8.1.2. Trigonometric polynomial approximations. Since $A\in G_{\rho}^{\nu}(\mathbb{T},SL(2,\mathbb{R}))$, then we can write $A(\theta)=\sum_{k\in\mathbb{Z}}\widehat{A}(k)\mathrm{e}^{2\pi\mathrm{i}k\theta}$ with estimate (8.8) $\begin{split}|\widehat{A}(k)|\leq\|A\|_{\nu,\rho}\mathrm{e}^{-\rho|2\pi k|^{\nu}},\ \ \forall k\in{\mathbb{Z}}.\end{split}$ For any $N>0,$ denote $\widetilde{N}=N^{b\nu^{-1}},$ where $b=\delta(\nu^{-1}-1)^{-1}$ and $\delta\in(0,1)$ will be fixed later. Once we have this, we can consider the truncated cocycle $\widetilde{A}(\theta):=\mathcal{T}_{\widetilde{N}}A(\theta),$ denote by $\widetilde{A}_{N}(\alpha,\theta)$ and $\widetilde{L}_{N}(\alpha,\widetilde{A})$ the associated transfer matrix and finite Lyapunov exponent by substituting $\widetilde{A}(\theta)$ for $A(\theta)$. Then we have the following lemma. ###### Lemma 8.1. There exists $N(\rho,\nu,\|A\|_{\nu,\rho})\in{\mathbb{N}}$, $c=c(\rho,\nu,\|A\|_{\nu,\rho})$ such that if $N\geq N(\rho,\nu,\|A\|_{\nu,\rho})$, then we have the following estimates: (8.9) $\|A(\theta)-\widetilde{A}(\theta)\|\leq\mathrm{e}^{-c\widetilde{N}^{\nu}}=\mathrm{e}^{-cN^{b}},$ $\big{|}N^{-1}\ln\|A_{N}(\alpha,\theta)\|-N^{-1}\ln\|\widetilde{A}_{N}(\alpha,\theta)\|\big{|}\leq\mathrm{e}^{-\frac{c}{2}N^{b}},$ $\big{|}L_{N}(\alpha,A)-\widetilde{L}_{N}(\alpha,\widetilde{A})\big{|}<\mathrm{e}^{-\frac{c}{2}N^{b}}.$ ###### Proof. The estimate (8.9) follows directly from (8.8). Moreover, by telescoping argument, for $N\geq N(\rho,\nu,\|A\|_{\nu,\rho})$ which is large enough, we have $\begin{split}\big{\|}A_{N}(\alpha,\theta)-\widetilde{A}_{N}(\alpha,\theta)\big{\|}\leq(\|A\|_{\nu,\rho}+1)^{N}\mathrm{e}^{-cN^{b}}\leq\mathrm{e}^{-\frac{c}{2}N^{b}}.\end{split}$ It follows that $\begin{split}\Big{|}\frac{1}{N}\ln\|A_{N}(\alpha,\theta)\|-\frac{1}{N}\ln\|\widetilde{A}_{N}(\alpha,\theta)\|\Big{|}\leq\frac{1}{N}\big{\|}A_{N}(\theta)-\widetilde{A}_{N}(\theta)\big{\|}<\mathrm{e}^{-\frac{c}{2}N^{b}}.\end{split}$ By averaging, one thus has $\big{|}L_{N}(\alpha,A)-\widetilde{L}_{N}(\alpha,\widetilde{A})\big{|}<\mathrm{e}^{-\frac{c}{2}N^{b}}.$ ∎ Since $\widetilde{A}(\theta)$ is a trigonometric polynomial, then one can analytic continue $\widetilde{A}(\theta)$ to become an analytic function. Indeed, let $\rho_{N}=\frac{\rho}{4\pi}\widetilde{N}^{\nu-1}=\frac{\rho}{4\pi}N^{-b(\nu^{-1}-1)}:=\frac{\rho}{4\pi}N^{-\delta},$ and set $\vartheta=\theta+\mathrm{i}\widetilde{\theta}$, then $\widetilde{A}(\vartheta)$ is analytic in the strip $|\widetilde{\theta}|\leq\rho_{N}$: (8.10) $\begin{split}\|\widetilde{A}\|_{\rho_{N}}^{*}=\sum\limits_{|k|<\widetilde{N}}|\widehat{A}(k)|\mathrm{e}^{|2\pi k\rho_{N}|}\leq\|A\|_{\nu,\rho}\sum\limits_{|k|<\widetilde{N}}\mathrm{e}^{-\rho|2\pi k|^{\nu}}\mathrm{e}^{|2\pi k|\rho_{N}}:=\mathrm{e}^{C_{1}}<\infty.\end{split}$ For $\vartheta=\theta+\mathrm{i}\widetilde{\theta}$ with $|\widetilde{\theta}|\leq\rho_{N}$, set (8.11) $\begin{split}\tilde{u}_{N}(\vartheta):=\frac{1}{N}\ln\|\widetilde{A}_{N}(\vartheta)\|.\end{split}$ In the following lemma, we will prove that $|\tilde{u}_{N}(\vartheta)|$ is indeed a bounded subharmonic function in the strip $[|\mathrm{Im}\vartheta|<\rho_{N}].$ ###### Lemma 8.2. We have the estimate (8.12) $\begin{split}\sup_{\theta\in{\mathbb{T}}}\sup_{|\widetilde{\theta}|\leq\rho_{N}}|\tilde{u}_{N}(\vartheta)|\leq\max\\{\ln 2,C_{1}\\},\end{split}$ where $C_{1}$ is the one in (8.10). ###### Proof. We will prove that the analytic continuation $\widetilde{A}(\vartheta)$ in the strip $[|\mathrm{Im}\vartheta|<\rho_{N}]$ is not singular, which implies that $|\tilde{u}_{N}(\vartheta)|$ is a bounded subharmonic function. We first give a estimate about $\|\widetilde{A}(\vartheta)-\widetilde{A}(\theta)\|$ as follows: $\displaystyle\|\widetilde{A}(\vartheta)-\widetilde{A}(\theta)\|$ $\displaystyle\leq$ $\displaystyle\sum_{0<|k|<\widetilde{N}}|\widehat{\tilde{A}}(k)|\sup_{\theta\in{\mathbb{T}},|\widetilde{\theta}|\leq\rho_{N}}|\mathrm{e}^{2\pi\mathrm{i}k\theta}(\mathrm{e}^{-2\pi k\widetilde{\theta}}-1)|$ $\displaystyle=$ $\displaystyle\sum_{0<|k|\leq N^{\delta/2}}|\widehat{\tilde{A}}(k)|\mathrm{e}^{|2\pi k|\rho_{N}}(1-\mathrm{e}^{-|2\pi k\rho_{N}|})$ $\displaystyle+$ $\displaystyle\sum_{N^{\delta/2}<|k|<N^{b\nu^{-1}}}|\widehat{A}(k)|\mathrm{e}^{|2\pi k|^{\nu}\rho}(\mathrm{e}^{|2\pi k|\rho_{N}}-1)\mathrm{e}^{-|2\pi k|^{\nu}\rho}.$ To estimate the first term, note if $0<|k|\leq N^{\delta/2}$, then one has $|2\pi k|\rho_{N}\leq\rho N^{-\delta/2}/2\ll 1,$ which implies $\begin{split}1-\mathrm{e}^{-|2\pi k\rho_{N}|}\leq 2|2\pi k\rho_{N}|\leq\rho N^{-\delta/2}.\end{split}$ To estimate the second term, note for all $k$ with $N^{\delta/2}<|k|<N^{b\nu^{-1}},$ one has $\begin{split}|2\pi k|^{\nu}\rho/2-|2\pi k|\rho_{N}=2^{-1}\rho(|2\pi k|^{\nu}-|k|N^{-\delta})\geq 0,\end{split}$ which implies $\begin{split}(\mathrm{e}^{|2\pi k|\rho_{N}}-1)\mathrm{e}^{-|2\pi k|^{\nu}\rho/2}<2,\ \forall N^{\delta/2}<|k|<N^{b\nu^{-1}}.\end{split}$ Consequently, one has (8.13) $\displaystyle\|\widetilde{A}(\vartheta)-\widetilde{A}(\theta)\|$ $\displaystyle\leq$ $\displaystyle\rho N^{-\delta/2}\|\tilde{A}\|_{\rho_{N}}^{*}+2\|A\|_{\nu,\rho}\mathrm{e}^{-|2\pi N^{\delta/2}|^{\nu}\rho/2}$ $\displaystyle<$ $\displaystyle 2\rho N^{-\delta/2}\mathrm{e}^{C_{1}}.$ To see this, one only needs to check that $f(x)=x^{\nu}-(2\pi)^{-1}xN^{-\delta}>0$ on the interval $[2\pi,2\pi N^{b\nu^{-1}}].$ Now we give the estimate $|\det{\widetilde{A}(\theta)}|.$ First, the inequality in (8.13) implies (8.14) $\displaystyle|\det{\widetilde{A}(\vartheta)}-\det{\widetilde{A}(\theta)}|\leq 4\|\widetilde{A}(\vartheta)\|\|\widetilde{A}(\vartheta)-\widetilde{A}(\theta)\|\leq 8\rho\mathrm{e}^{2C_{1}}N^{-\delta/2}\ll 1.$ Moreover, note $A\in SL(2,{\mathbb{R}})$, then by Lemma 8.1, we have (8.15) $\begin{split}|1-\det{\widetilde{A}(\theta)}|\leq 8\|A(\theta)\|\|A(\theta)-\tilde{A}(\theta)\|\leq C\mathrm{e}^{-cN^{b}},\end{split}$ that is $|\det{\widetilde{A}(\theta)}|\geq 1/2,\ \forall\theta\in{\mathbb{T}},$ which, together with (8.14), yields (8.16) $\displaystyle|\det{\widetilde{A}(\vartheta)}|\geq 1/4,\forall(\theta,\widetilde{\theta})\in{\mathbb{T}}\times[-\rho_{N},\rho_{N}].$ Once we get the inequality in (8.16), we are ready to estimate $|\tilde{u}_{N}(\vartheta)|.$ Indeed, $\displaystyle\|\widetilde{A}_{N}(\vartheta)\|^{2}$ $\displaystyle\geq|\det{\widetilde{A}_{N}(\vartheta)}|=|\Pi_{\ell=0}^{N-1}\det{\widetilde{A}(\vartheta+\ell\alpha)}|$ $\displaystyle=\Pi_{\ell=0}^{N-1}|\det{\widetilde{A}(\vartheta+\ell\alpha)}|\geq 4^{-N},$ which yields $2^{-N}\leq\|\widetilde{A}_{N}(\vartheta)\|\leq\mathrm{e}^{NC_{1}},$ or $\displaystyle\sup_{\theta\in{\mathbb{T}}}\sup_{|\widetilde{\theta}|\leq\rho_{N}}|\tilde{u}_{N}(\vartheta)|\leq\max\\{\ln 2,C_{1}\\}.$ ∎ #### 8.1.3. Proof of Proposition 5 In this section we will give the proof of Proposition 5 by applying Proposition 6. First by Lemma 8.1, we have $\displaystyle|\frac{1}{N}\ln\|A_{N}(\theta)\|-L_{N}(\alpha,A)|$ $\displaystyle\leq|N^{-1}\ln\|A_{N}(\theta)\|-\tilde{u}_{N}(\theta)|$ $\displaystyle+|\tilde{u}_{N}(\theta)-[\tilde{u}_{N}(\theta)]_{\theta}|+|[\tilde{u}_{N}(\theta)]_{\theta}-L_{N}(\alpha,A)|$ (8.17) $\displaystyle\leq|\tilde{u}_{N}(\theta)-[\tilde{u}_{N}(\theta)]_{\theta}|+2\mathrm{e}^{-\frac{c}{2}N^{b}}.$ Thus we only need to estimate $|\tilde{u}_{N}(\theta)-[\tilde{u}_{N}(\theta)]_{\theta}|,$ which is controlled by the sum (8.18) $\begin{split}|\tilde{u}_{N}(\theta)-F_{R,p}[\tilde{u}_{N}](\theta)|+|F_{R,p}[\tilde{u}_{N}](\theta)-[\tilde{u}_{N}(\theta)]_{\theta}|,\end{split}$ where $F_{R,p}[u](\theta)=\frac{1}{R^{p}}\sum\limits_{j=0}^{p(R-1)}c_{R}^{p}(j)\tilde{u}_{N}(\theta+j\alpha).$ In the following, we will give the estimates of the two terms above. First we would like to bound $|\tilde{u}_{N}(\theta)-\tilde{u}_{N}(\theta+\alpha)|$. For the function $\tilde{u}_{N}(\theta)$ defined by (8.11), the inequality in (8.12) and $C_{1}>\ln 2$ imply $\tilde{u}_{N}(\vartheta)$ is a bounded subharmonic function on $[|\mathrm{Im}\vartheta|<\rho_{N}].$ It follows that (8.19) $\begin{split}\|\tilde{u}_{N}(\vartheta)\|\leq C_{1}=2^{-1}C_{2},\ C_{2}=2C_{1}.\end{split}$ That is this function $\tilde{u}_{N}(\theta)$ satisfies the hypotheses in Proposition 6. Consequently, $\displaystyle\Big{|}\tilde{u}_{N}(\theta)-\tilde{u}_{N}(\theta+\alpha)\Big{|}$ $\displaystyle=\frac{1}{N}\Big{|}\ln\|\widetilde{A}_{N}(\theta)\|-\ln\|\widetilde{A}_{N}(\theta+\alpha)\|\Big{|}$ $\displaystyle=\Big{|}\frac{1}{N}\ln\frac{\|\widetilde{A}_{N}(\theta)\|}{\|\widetilde{A}_{N}(\theta+\alpha)\|}\Big{|}$ $\displaystyle=\Big{|}\frac{1}{N}\ln\frac{\|\widetilde{A}(\theta+(N-1)\alpha)\cdots\widetilde{A}(\theta+\alpha)\widetilde{A}(\theta)\|}{\|\widetilde{A}(\theta+N\alpha)\widetilde{A}(\theta+(N-1)\alpha)\cdots\widetilde{A}(\theta+\alpha)\|}\Big{|}$ $\displaystyle\leq\frac{1}{N}\ln\big{(}\|\widetilde{A}(\theta+N\alpha)^{-1}\|\cdot\|\widetilde{A}(\theta)\|\big{)}.$ Thus we only need to estimate $\|\widetilde{A}(\theta+N\alpha)^{-1}\|$. Indeed, (8.15) implies $\displaystyle\|\widetilde{A}(\theta)^{-1}\|$ $\displaystyle\leq\|1/\det{\widetilde{A}(\theta)}\|\|\widetilde{A}(\theta)\|$ $\displaystyle\leq\\{1+2\|I-\det{\widetilde{A}(\theta)}\|\\}\|\widetilde{A}(\theta)\|\leq 2\mathrm{e}^{C_{1}}\leq\mathrm{e}^{2C_{1}}.$ Once we have this we get (8.20) $\Big{|}\tilde{u}_{N}(\theta)-\tilde{u}_{N}(\theta+\alpha)\Big{|}\leq\frac{3C_{2}}{2N}.$ For the fixed $\nu$ with $1/2<\nu<1,$ there exists $\delta\in(0,1)$ such that $\nu^{-1}<1+\delta.$ Once we fix $\nu$ and $\delta$ by this way, we set $b=\delta(\nu^{-1}-1)^{-1}>1.$ Then we choose $\sigma$ and $\varsigma$ in Proposition 6 as $1<\sigma<\min\\{2,\delta^{-1}\\}$, $\varsigma=\delta.$ That is the parameters $\sigma$ and $p>1$ in Proposition 6 satisfy $\displaystyle\delta=b(1/\nu-1)<\frac{1}{\sigma}<1,\ \ \frac{\delta\sigma}{\sigma-1}<p<\frac{1}{\sigma-1}.$ Take $\gamma=1+p(1-\sigma),\sigma_{1}=\frac{p}{\delta}(\sigma-1)$. It is obvious that $\displaystyle 1>\gamma=1+p(1-\sigma)>0,\ \ \sigma-\sigma_{1}=\frac{\delta\sigma-p(\sigma-1)}{\delta}<0.$ Notice that $1<\sigma<\delta^{-1}$ and $\nu^{-1}-1<\delta<1,$ thus $\delta,\sigma\rightarrow 1$ as $\nu\rightarrow 1/2,$ which imply $\sigma_{1}\rightarrow\sigma,p\rightarrow\infty$ and $\gamma\rightarrow 0$ as $\nu\rightarrow 1/2.$ For $q,R$ with $R=q^{\sigma}$ (as Proposition 6), set $\frac{9pC_{2}}{\kappa}q^{\sigma}<N<\big{(}\frac{\kappa\rho}{2^{p+6}\pi C_{2}}\big{)}^{\frac{1}{\delta}}q^{\sigma_{1}}$ and $K$ as the one in Proposition 6. Now we give the estimate of the first term in (8.18). More concretely, $\displaystyle\Big{|}F_{R,p}[\tilde{u}_{N}](\theta)-\tilde{u}_{N}(\theta)\Big{|}$ $\displaystyle\leq\frac{1}{R^{p}}\sum\limits_{j=0}^{p(R-1)}|\tilde{u}_{N}(\theta+j\alpha)-\tilde{u}_{N}(\theta)|c_{R}^{p}(j)$ $\displaystyle\leq\frac{1}{R^{p}}\sum\limits_{j=0}^{p(R-1)}c_{R}^{p}(j)\frac{j3C_{2}}{2N}<\frac{3p(R-1)C_{2}}{2N}$ (8.21) $\displaystyle\leq 3p(R-1)C_{2}\frac{\kappa}{18pC_{2}}q^{-\sigma}\leq\frac{\kappa}{6},$ where the third inequality is by (8.20). We will apply Proposition 6 to get the estimate of the second term in (8.18). More concretely, let $\varsigma_{i},i=1,2,3$ be the ones in Proposition 6 with $2^{-1}C_{2}$ and $\rho_{N}$ in place of $S$ and $\rho,$ respectively and note $N<\big{(}\frac{\kappa\rho}{2^{p+6}\pi C_{2}}\big{)}^{\frac{1}{\delta}}q^{\sigma_{1}}$ we get $\displaystyle\varsigma_{2}R^{-\varsigma_{1}}=2^{p+5}2^{-1}C_{2}4\pi\rho^{-1}N^{\delta}q^{-p(\sigma-1)}<\kappa.$ Thus by Proposition 6 we know that there is a set such that for all $\theta$ outside this set we have (8.22) $\displaystyle\Big{|}F_{R,p}[\tilde{u}_{N}](\theta)-[\tilde{u}_{N}(\theta)]_{\theta}\Big{|}\leq\varsigma_{2}R^{-\varsigma_{1}}<\kappa.$ Moreover, the measure of this set is less than (8.23) $\displaystyle 2^{-8}R^{2\varsigma_{1}}\exp\\{-R^{\varsigma_{3}}\\}=2^{-8}q^{2p(\sigma-1)}\exp\\{-q^{\gamma}\\}<\exp\\{-2^{-1}q^{\gamma}\\}.$ Set $C_{1}(\kappa)=\frac{9pC_{2}}{\kappa},C_{2}(\kappa)=\big{(}\frac{\kappa\rho}{2^{p+6}\pi C_{2}}\big{)}^{\frac{1}{\delta}}$, $c=\frac{1}{2}$ and $q\geq q_{0},$ with $q_{0}$ depending on $\kappa,\rho,\nu$ (by Lemma 8.1) and sufficiently large, then by (8.1.3)-(8.23) we finish the proof of Proposition 5. ### 8.2. Application of avalanche principle. ###### Proposition 7 ([28, 17]). Let $A_{1}$, $A_{2}$, $\cdots$, $A_{n}$ be a sequence in $SL(2,{\mathbb{R}})$ satisfying the conditions $\displaystyle\min\limits_{1\leq j\leq n}$ $\displaystyle\|A_{j}\|\geq\mu>n,$ $\displaystyle\max\limits_{1\leq j\leq n}$ $\displaystyle\left|\ln\|A_{j}\|+\ln\|A_{j+1}\|-\ln\|A_{j+1}A_{j}\|\right|<\frac{1}{2}\ln\mu.$ Then there exists a constant $C_{A}<\infty$ such that $\displaystyle\Big{|}\ln\|\prod\limits_{j=1}^{n}A_{j}\|+\sum\limits_{j=2}^{n-1}\ln\|A_{j}\|-\sum\limits_{j=1}^{n-1}\ln\|A_{j+1}A_{j}\|\Big{|}<C_{A}\frac{n}{\mu}.$ Following the ideas in [16], in case of positive Lyapunov exponent, the large deviation theorem provides us a possibility to apply avalanche principal (Proposition 7) to $A(\theta+jN\alpha)$ for $\theta$ in a set of large measure and therefore pass on to a lager scale. ###### Lemma 8.3. Assume that $|\alpha-\frac{a}{q}|<\frac{1}{q^{2}}$, $(a,q)=1$. Let $C_{1}(\kappa)q^{\sigma}<N<C_{2}(\kappa)q^{\sigma_{1}}$ and $q\geq q_{0}(\kappa)$ be the same as Proposition 5. Assume that $L_{N}(\alpha,A)>100\kappa>0$ and $L_{2N}(\alpha,A)>\frac{19}{20}L_{N}(\alpha,A)$. Then for $N^{\prime}$ such that $m=N^{\prime}N^{-1}$ satisfies $\mathrm{e}^{\frac{c}{2}q^{\gamma/4}}<m<\mathrm{e}^{\frac{c}{2}q^{\gamma}}$, we have $\displaystyle\Big{|}L_{N^{\prime}}(\alpha,A)+L_{N}(\alpha,A)-2L_{2N}(\alpha,A)\Big{|}<C\mathrm{e}^{-\frac{c}{2}q^{\gamma/4}},$ where $c$ is the one from the large deviation bound of Proposition 5. ###### Proof. We use the avalanche principal (Proposition 7) on $A_{j}^{N}(\theta)=A_{N}(\theta+jN\alpha)$ with $\theta$ being restricted to the set $\Omega\subset{\mathbb{T}}$, defined by $2m$ conditions: $\displaystyle\Big{|}\frac{1}{N}\ln\|A_{j}^{N}(\theta)\|-L_{N}(\alpha,A)\Big{|}\leq\kappa,\ \ 1\leq j\leq m,$ $\displaystyle\Big{|}\frac{1}{2N}\ln\|A_{j}^{2N}(\theta)\|-L_{2N}(\alpha,A)\Big{|}\leq\kappa,\ \ 1\leq j\leq m.$ By Proposition 5, we have for any $j$ $\displaystyle mes\Big{\\{}\theta:\big{|}\frac{1}{N}\ln\|A_{j}^{N}(\theta)\|-L_{N}(\alpha,A)\big{|}>\kappa\Big{\\}}<\mathrm{e}^{-cq^{\gamma}},$ $\displaystyle mes\Big{\\{}\theta:\big{|}\frac{1}{2N}\ln\|A_{j}^{2N}(\theta)\|-L_{2N}(\alpha,A)\big{|}>\kappa\Big{\\}}<\mathrm{e}^{-cq^{\gamma}}.$ Thus we have (8.24) $\displaystyle mes({\mathbb{T}}\setminus\Omega)<2m\mathrm{e}^{-cq^{\gamma}}.$ For each $A_{j}^{N}(\theta)$ with $\theta\in\Omega$, $\displaystyle\mathrm{e}^{N(L_{N}(\alpha,A)-\kappa)}<\|A_{j}^{N}(\theta)\|<\mathrm{e}^{N(L_{N}(\alpha,A)+\kappa)}.$ Note that since $L_{N}(\alpha,A)>100\kappa$, then $\displaystyle\|A_{j}^{N}(\theta)\|>\mathrm{e}^{\frac{99}{100}NL_{N}(\alpha,A)}:=\mu.$ For large enough $q$, and hence $N$ by hypothesis, we have $\mu>2m$ (since $\sigma>1>\gamma$). Also for $j<m$, by the fact that $A_{j+1}^{N}(\theta)A_{j}^{N}(\theta)=A_{j}^{2N}(\theta)$, we have $\displaystyle\big{|}\ln\|A_{j}^{N}(\theta)\|$ $\displaystyle+\ln\|A_{j+1}^{N}(\theta)\|-\ln\|A_{j+1}^{N}(\theta)A_{j}^{N}(\theta)\|\big{|}$ $\displaystyle<4N\kappa+2N|L_{N}(\alpha,A)-L_{2N}(\alpha,A)|$ $\displaystyle<\frac{1}{25}NL_{N}(\alpha,A)+2N(\frac{1}{20}L_{N}(\alpha,A))<\frac{1}{2}\ln\mu,$ where the second inequality follows by $L_{2N}(\alpha,A)>\frac{19}{20}L_{N}(\alpha,A)$. Thus, we can apply the avalanche principal (Proposition 7) for $\theta\in\Omega$ to obtain $\displaystyle\Big{|}\ln\|\prod\limits_{j=1}^{m}A_{j}^{N}(\theta)\|$ $\displaystyle+\sum\limits_{j=2}^{m-1}\ln\|A_{j}^{N}(\theta)\|-\sum\limits_{j=1}^{m-1}\ln\|A_{j+1}^{N}(\theta)A_{j}^{N}(\theta)\|\Big{|}$ $\displaystyle<C_{A}m/\mu<m\mathrm{e}^{-\frac{1}{2}NL_{N}(\alpha,A)}.$ Integrating on $\Omega$, we get $\displaystyle\Big{|}\int_{\Omega}\ln\|A_{N^{\prime}}(\theta)\|d\theta$ $\displaystyle+\sum\limits_{j=2}^{m-1}\int_{\Omega}\ln\|A_{N}(\theta+jN\alpha)\|d\theta$ $\displaystyle-\sum\limits_{j=1}^{m-1}\int_{\Omega}\ln\|A_{2N}(\theta+jN\alpha)\|d\theta\Big{|}<m\mathrm{e}^{-\frac{1}{2}NL_{N}(\alpha,A)},$ therefore, recalling (8.24) and the assumption $N>C_{1}(\kappa)q^{\sigma}$, we have $\displaystyle\Big{|}L_{N^{\prime}}(\alpha,A)$ $\displaystyle+\frac{m-2}{m}L_{N}(\alpha,A)-\frac{2(m-1)}{m}L_{2N}(\alpha,A)\Big{|}$ $\displaystyle<mN^{\prime-1}\mathrm{e}^{-\frac{1}{2}NL_{N}(\alpha,A)}+C\mathrm{e}^{-\frac{c}{2}q^{\gamma}}<C\mathrm{e}^{-\frac{c}{2}q^{\gamma}}.$ It follows that $\displaystyle|L_{N^{\prime}}(\alpha,A)$ $\displaystyle+L_{N}(\alpha,A)-2L_{2N}(\alpha,A)|$ $\displaystyle<C\mathrm{e}^{-\frac{c}{2}q^{\gamma}}+2m^{-1}|L_{N}(\alpha,A)-L_{2N}(\alpha,A)|$ $\displaystyle<C\mathrm{e}^{-\frac{c}{2}q^{\gamma}}+L_{N}(\alpha,A)(10m)^{-1}<C\mathrm{e}^{-\frac{c}{2}q^{\gamma/4}},$ where the second inequality is by $L_{2N}(\alpha,A)>\frac{19}{20}L_{N}(\alpha,A)$ and the last inequality is by $m>\mathrm{e}^{\frac{c}{2}q^{\gamma/4}}$. ∎ Actually, the condition “$L_{2N}(\alpha,A)>\frac{19}{20}L_{N}(\alpha,A)$” is not necessary if $q$ is sufficiently large and $L(\alpha,A)>0$. ###### Lemma 8.4. Assume that $L(\alpha,A)>100\kappa>0,$ then there exists $N_{0}\in{\mathbb{N}}$ with $C_{1}(\kappa)q_{0}^{\sigma}<N_{0}<C_{2}(\kappa)q_{0}^{\sigma_{1}}$, $q_{0}$ is the one defined in Proposition 5 such that (8.25) $\displaystyle L_{2N_{0}}(\alpha,A)>\frac{99}{100}L_{N_{0}}(\alpha,A).$ ###### Proof. Note that for any $n$, by subadditivity, we have $\displaystyle 100\kappa<L(\alpha,A)=\inf L_{n}(\alpha,A)\leq L_{2n}(\alpha,A)\leq L_{n}(\alpha,A)\leq C_{1},$ where $C_{1}$ is the one in (8.10). Set $j_{0}=\big{[}(\ln(100/99))^{-1}\ln(C_{1}/100\kappa)\big{]},$ that is (8.26) $\displaystyle(99/100)^{j_{0}+1}C_{1}<100\kappa\leq(99/100)^{j_{0}}C_{1}.$ Consider the sequence $\\{L_{2^{j}N}(\alpha,A)\\}$ where $N=[C_{1}(\kappa)q_{0}^{\sigma}]+1,$ $j\in{\mathbb{N}}$. If $\displaystyle L_{2^{j+1}N}(\alpha,A)\leq(99/100)L_{2^{j}N}(\alpha,A)$ hold for all $0\leq j\leq j_{0},$ then $\displaystyle 100\kappa<L_{2^{j_{0}+1}N}(\alpha,A)$ $\displaystyle\leq(99/100)^{j_{0}+1}L_{N}(\alpha,A)\leq(99/100)^{j_{0}+1}C_{1}<100\kappa,$ where the last inequality is by first inequality in (8.26). Thus there exists $j_{*}^{\prime}s$ with $0\leq j_{*}\leq j_{0}$ such that $\displaystyle L_{2^{j_{*}+1}N}(\alpha,A)>(99/100)L_{2^{j_{*}}N}(\alpha,A).$ Moreover, since $j_{0}$ is fixed, we can set $q_{0}$ large enough such that $\displaystyle 2^{j_{0}}<2^{-1}C_{1}(\kappa)^{-1}C_{2}(\kappa)q_{0}^{\sigma_{1}-\sigma}.$ Set $N_{0}=2^{j_{*}}N.$ Thus we have the estimates $\displaystyle C_{1}(\kappa)q_{0}^{\sigma}\leq N\leq N_{0}\leq 2^{j_{0}}N\leq C_{2}(\kappa)q_{0}^{\sigma_{1}}$ with $\displaystyle L_{2N_{0}}(\alpha,A)>(99/100)L_{N_{0}}(\alpha,A).$ ∎ ### 8.3. Inductive argument. Once one has Lemma 8.3, one can follow the induction arguments developed in [16]. However, in our case there is an upper bound of $N$ in the large deviation theorem. Thus we can only deal with Diophantine frequencies and their continued fraction expansions. Moreover, we need to deal with the Diophantine frequencies and their continued fraction expansions seperately. Let $p_{n}/q_{n}$ be the continued fraction expansion of $\alpha$. To apply Lemma 8.3 inductively, we first fix $\alpha\in DC(v,\tau),$ and inductively choose the following sequences: (8.27) $\displaystyle q_{0}=\tilde{q}_{0}<N_{0}<\tilde{q}_{1}<N_{1}<\cdots<N_{s}<\tilde{q}_{s+1}<N_{s+1}<\cdots,$ where $\tilde{p}_{i}/\tilde{q}_{i}$ is a subsequence of the continued fraction expansion of $\alpha$ with (8.28) $\text{$\tilde{q}_{s+1}$ is the smallest $q_{j}$ such that $\tilde{q}_{s+1}>\mathrm{e}^{\tilde{q}_{s}^{\gamma/2}}$},\ s\geq 0,$ (8.29) $\displaystyle C_{1}(\kappa)\tilde{q}_{s}^{\sigma}<N_{s}<C_{2}(\kappa)\tilde{q}_{s}^{\sigma_{1}},\ \ \tilde{q}_{s}|N_{s},\ s\geq 0,$ (8.30) $\displaystyle N_{s+1}=m_{s+1}N_{s},\ \ \mathrm{e}^{\frac{c}{2}\tilde{q}_{s}^{\gamma/4}}<m_{s+1}<2m_{s+1}<\mathrm{e}^{\frac{c}{2}\tilde{q}_{s}^{\gamma}},\ s\geq 0.$ Actually, we can inductively select a sequence $\\{N_{s}\\}$ such that (8.28)-(8.30) hold, indeed the starting case $s=0$ follows from Lemma 8.4. First by the selection of $\tilde{q}_{s}$ and the Diophantine condition of $\alpha$, one can check that $\mathrm{e}^{\tilde{q}_{s}^{\gamma/2}}<\tilde{q}_{s+1}<\mathrm{e}^{2\tau\tilde{q}_{s}^{\gamma/2}}$. Take $N_{s+1}=N_{s}m_{s+1}$ with $m_{s+1}:=\tilde{q}_{s+1}([\tilde{q}_{s+1}^{\sigma-1}]+1).$ It’s easy to check that $\displaystyle C_{1}(\kappa)\tilde{q}_{s+1}^{\sigma}<C_{1}(\kappa)\tilde{q}_{s}^{\sigma}\tilde{q}_{s+1}^{\sigma}<N_{s+1}<2C_{2}(\kappa)\tilde{q}_{s}^{\sigma_{1}}\tilde{q}_{s+1}^{\sigma}<C_{2}(\kappa)\tilde{q}_{s+1}^{\sigma_{1}},$ $\displaystyle\mathrm{e}^{\frac{c}{2}\tilde{q}_{s}^{\gamma/4}}<\mathrm{e}^{\sigma\tilde{q}_{s}^{\gamma/2}}<\tilde{q}_{s+1}^{\sigma}<m_{s+1}<2m_{s+1}\leq 4\tilde{q}_{s+1}^{\sigma}<4\mathrm{e}^{2\sigma\tau\tilde{q}_{s}^{\gamma/2}}<\mathrm{e}^{\frac{c}{2}\tilde{q}_{s}^{\gamma}}.$ Thus such a choice of $N_{s}$ satisfies all estimates in (8.28)-(8.30) if $\tilde{q}_{0}$ is sufficiently large. With the help of such sequence, we can prove the following: ###### Lemma 8.5. Assume that $\alpha\in DC(v,\tau)$ and $L(\alpha,A)>100\kappa>0$. There exist $c^{\prime\prime}>0$ and $C_{1}(\kappa)\tilde{q}_{0}^{\sigma}<N_{0}<C_{2}(\kappa)\tilde{q}_{0}^{\sigma_{1}}$ such that $\displaystyle|L(\alpha,A)+L_{N_{0}}(\alpha,A)-2L_{2N_{0}}(\alpha,A)|<\mathrm{e}^{-c^{\prime\prime}\tilde{q}_{0}^{\gamma/4}}.$ ###### Proof. Let $c/100<c_{3}<c_{2}<c_{1}<c/2$, $2C_{1}<C<\infty$, and $\tilde{q}_{-1}=0$. We use induction to show that the sequences $\\{N_{s}\\}$ and $\\{\tilde{q}_{s}\\},$ defined by (8.27), additionally satisfy, for $s\geq 0$, (8.31) $\displaystyle|L_{N_{s+1}}(\alpha,A)+L_{N_{s}}(\alpha,A)-2L_{2N_{s}}(\alpha,A)|<C\mathrm{e}^{-c_{1}\tilde{q}_{s}^{\gamma/4}},$ (8.32) $\displaystyle|L_{2N_{s+1}}(\alpha,A)-L_{N_{s+1}}(\alpha,A)|<C\mathrm{e}^{-c_{2}\tilde{q}_{s}^{\gamma/4}},$ (8.33) $\displaystyle|L_{N{s+1}}(\alpha,A)-L_{N_{s}}(\alpha,A)|<C\mathrm{e}^{-c_{3}\tilde{q}_{s-1}^{\gamma/4}}.$ We first check the case $s=0$. Fix $N_{1}$ satisfying (8.30). We will show $\displaystyle|L_{N_{1}}(\alpha,A)+L_{N_{0}}(\alpha,A)-2L_{2N_{0}}(\alpha,A)|<C\mathrm{e}^{-c_{1}\tilde{q}_{0}^{\gamma/4}},$ $\displaystyle|L_{2N_{1}}(\alpha,A)-L_{N_{1}}(\alpha,A)|<C\mathrm{e}^{-c_{2}\tilde{q}_{0}^{\gamma/4}},$ $\displaystyle|L_{N_{1}}(\alpha,A)-L_{N_{0}}(\alpha,A)|<C\mathrm{e}^{-c_{3}\tilde{q}_{-1}^{\gamma/4}}=C.$ In this case, the last inequality holds automatically since $\tilde{q}_{-1}=0$, one only needs to check the first two inequalities. By (8.25) and (8.29),(8.30) with $s=0$ we know that the conditions in Lemma 8.3 are all satisfied with $N^{\prime}=N_{1}$, $N=N_{0}$ and $q=\tilde{q}_{0}$. Therefore, by Lemma 8.3, we have $\displaystyle|L_{N_{1}}(\alpha,A)+L_{N_{0}}(\alpha,A)-2L_{2N_{0}}(\alpha,A)|<C\mathrm{e}^{-\frac{c}{2}\tilde{q}_{0}^{\gamma/4}}<C\mathrm{e}^{-c_{1}\tilde{q}_{0}^{\gamma/4}}.$ On the other hand, (8.30) ensures one can also apply Lemma 8.3 to $N^{\prime}=2N_{1}$, thus we have $\displaystyle|L_{2N_{1}}(\alpha,A)+L_{N_{0}}(\alpha,A)-2L_{2N_{0}}(\alpha,A)|<C\mathrm{e}^{-c_{1}\tilde{q}_{0}^{\gamma/4}}.$ It follows that $\displaystyle|L_{2N_{1}}(\alpha,A)-L_{N_{1}}(\alpha,A)|<2C\mathrm{e}^{-c_{1}\tilde{q}_{0}^{\gamma/4}}<C\mathrm{e}^{-c_{2}\tilde{q}_{0}^{\gamma/4}},$ and we have completed the initial case $s=0.$ For $j\geq 1,$ assume that (8.31)-(8.33) hold for all $s$ with $s\leq j-1.$ Now we consider the case $s=j.$ Fix $N_{j+1}$ satisfying (8.30). By induction we have $\displaystyle|L_{2N_{j}}(\alpha,A)-L_{N_{j}}(\alpha,A)|<C\mathrm{e}^{-c_{2}\tilde{q}_{j-1}^{\gamma/4}}\leq C\mathrm{e}^{-c_{2}\tilde{q}_{0}^{\gamma/4}}.$ This implies $L_{2N_{j}}(\alpha,A)>(19/20)L_{N_{j}}(\alpha,A),$ which together with (8.29), implies $N_{j}$ satisfies the two conditions of $N$ in Lemma 8.3 with $\tilde{q}_{j}$ in place of $q.$ Moreover, by (8.30), $m_{j+1}=N_{j+1}N_{j}^{-1}$ satisfies the estimate of $m$ in Lemma 8.3 with $\tilde{q}_{j}$ in place of $q.$ Thus by Lemma 8.3, with $N^{\prime}=N_{j+1},N=N_{j}$ and $q=\tilde{q}_{j}$ we get $\displaystyle|L_{N_{j+1}}(\alpha,A)+L_{N_{j}}(\alpha,A)-2L_{2N_{j}}(\alpha,A)|<C\mathrm{e}^{-\frac{c}{2}\tilde{q}_{j}^{\gamma/4}}<C\mathrm{e}^{-c_{1}\tilde{q}_{j}^{\gamma/4}}.$ Similarly, (8.30) ensures one can also apply Lemma 8.3 to $N^{\prime}=2N_{j+1}$, and we have $\displaystyle|L_{2N_{j+1}}(\alpha,A)+L_{N_{j}}(\alpha,A)-2L_{2N_{j}}(\alpha,A)|<C\mathrm{e}^{-c_{1}\tilde{q}_{j}^{\gamma/4}}.$ Thus $\displaystyle|L_{2N_{j+1}}(\alpha,A)-L_{N_{j+1}}(\alpha,A)|<2C\mathrm{e}^{-c_{1}\tilde{q}_{j}^{\gamma/4}}<C\mathrm{e}^{-c_{2}\tilde{q}_{j}^{\gamma/4}},$ $\displaystyle|L_{N_{j+1}}(\alpha,A)-L_{N_{j}}(\alpha,A)|\leq$ $\displaystyle|L_{N_{j+1}}(\alpha,A)+L_{N_{j}}(\alpha,A)-2L_{2N_{j}}(\alpha,A)|$ $\displaystyle+|2L_{2N_{j}}(\alpha,A)-2L_{N_{j}}(\alpha,A)|$ $\displaystyle<$ $\displaystyle C\mathrm{e}^{-c_{1}\tilde{q}_{j}^{\gamma/4}}+2C\mathrm{e}^{-c_{2}\tilde{q}_{j-1}^{\gamma/4}}<C\mathrm{e}^{-c_{3}\tilde{q}_{j-1}^{\gamma/4}}.$ That is the estimates in (8.31)-(8.33) hold for all $s\in{\mathbb{N}}.$ As a consequence of (8.31) with $s=0$ and (8.33) $\displaystyle|L(\alpha,A)$ $\displaystyle+L_{N_{0}}(\alpha,A)-2L_{2N_{0}}(\alpha,A)|$ $\displaystyle\leq|L_{N_{1}}(\alpha,A)+L_{N_{0}}(\alpha,A)-2L_{2N_{0}}(\alpha,A)|$ $\displaystyle\ +\sum_{s\geq 1}|L_{N_{s+1}}(\alpha,A)-L_{N_{s}}(\alpha,A)|<\mathrm{e}^{-c^{\prime\prime}\tilde{q}_{0}^{\gamma/4}}.$ ∎ For the rational frequency case, we will first estimate the difference between $L(p_{j}/q_{j},A_{j})$ and $L_{n}(p_{j}/q_{j},A_{j})$ with $n$ much larger than $q_{j}$ (Lemma 8.6), and then use avalanche principle estimate to get estimate of $L_{n}(p_{j}/q_{j},A_{j})$ (Lemma 8.7). ###### Lemma 8.6. Consider the cocycle $(p/q,A)\in{\mathbb{Q}}\times C^{0}({\mathbb{T}},SL(2,{\mathbb{R}}))$ with $p,q\in{\mathbb{N}},(p,q)=1,$ and $\|A\|\leq\mathrm{e}^{C_{1}}.$ Set $n=mq+r,m\in{\mathbb{N}},0\leq r<q,$ then $\displaystyle L_{n}(p/q,A)\leq L(p/q,A)+2n^{-1}(\ln m+qC_{1}).$ ###### Proof. Set $A_{q}:=A_{q}(\theta)=A(\theta+(q-1)p/q)A(\theta+(q-2)p/q)\cdots A(\theta).$ For the matrix $A_{q}$ there exists a unitary $U$ such that $A_{q}=U\begin{pmatrix}\lambda&\psi\\\ 0&\lambda^{-1}\end{pmatrix}U^{-1}.$ Then for $m\in{\mathbb{N}},$ we have $A_{q}^{m}=U\begin{pmatrix}\lambda^{m}&R_{m}(\lambda,\psi)\\\ 0&\lambda^{-m}\end{pmatrix}U^{-1},m\geq 2$ with $R_{m}(\lambda,\psi)=\left\\{\begin{aligned} &\sum_{l=1}^{k}\\{\lambda^{2l-1}+\lambda^{-(2l-1)}\\}\psi&m=2k,k\geq 1,\\\ &\psi+\sum_{l=1}^{k}\\{\lambda^{2l}+\lambda^{-2l}\\}\psi&m=2k+1,k\geq 1.\end{aligned}\right.$ Thus $\|A_{q}^{m}\|\leq\|\lambda^{m}\|+\|R(\lambda,\psi)\|\leq\|\lambda\|^{m}(1+m\|\psi\|)\leq\rho(A_{q})^{m}(1+m\exp\\{qC_{1}\\}),$ where $\rho(A)$ stands for the spectrum radius. Note, for $n=mq+r,$ $A_{n}(\theta)=A_{r}(\theta)A_{q}^{m}(\theta),$ then by the inequality above we get $\displaystyle L_{n}(p/q,A)$ $\displaystyle=$ $\displaystyle\frac{1}{n}\int_{{\mathbb{T}}}\ln\|A_{n}(\theta)\|d\theta\leq\frac{1}{n}\int_{{\mathbb{T}}}\ln\|A_{q}^{m}(\theta)\|d\theta+\frac{rC_{1}}{n}$ $\displaystyle\leq$ $\displaystyle\frac{1}{n}\Big{\\{}\int_{{\mathbb{T}}}\ln\rho(A_{q})^{m}d\theta+\int_{{\mathbb{T}}}\ln(1+m\exp\\{qC_{1}\\})d\theta\Big{\\}}+\frac{qC_{1}}{n}$ $\displaystyle\leq$ $\displaystyle L(p/q,A)+2n^{-1}(\ln m+qC_{1}).$ ∎ ###### Lemma 8.7. Assume that $\alpha\in DC(v,\tau)$ and $\\{p_{j}/q_{j}\\}$ is the sequence of continued fraction expansion of $\alpha,$ and $A_{j}\rightarrow A$ under the topology derived by $\|\cdot\|_{\nu,\rho}$-norm. Then there exists a $j_{1}$ such that for $j\geq j_{1},$ we have $\displaystyle|L(p_{j}/q_{j},A_{j})+L_{N_{0}}(p_{j}/q_{j},A_{j})-2L_{2N_{0}}(p_{j}/q_{j},A_{j})|<2\mathrm{e}^{-c^{\prime\prime}\tilde{q}_{0}^{\gamma/4}}.$ ###### Remark 8.1. Here $N_{0}$ and $c^{\prime\prime}$ are the ones in Lemma 8.5. ###### Proof. For the fixed $N_{0},$ note $L_{2N_{0}}(\cdot_{1},\cdot_{2})$ and $L_{N_{0}}(\cdot_{1},\cdot_{2})$ are continuous in both variables and $L_{2N_{0}}(\alpha,A)>(99/100)L_{N_{0}}(\alpha,A)$, $L_{N_{0}}(\alpha,A)>100\kappa>0$, then there exists $j_{1}\in{\mathbb{Z}}^{+}$, such that if $j>j_{1}$, we have (8.34) $\displaystyle L_{N_{0}}(p_{j}/q_{j},A_{j})$ $\displaystyle>$ $\displaystyle 99\kappa$ (8.35) $\displaystyle L_{2N_{0}}(p_{j}/q_{j},A_{j})$ $\displaystyle>$ $\displaystyle(49/50)L_{N_{0}}(p_{j}/q_{j},A_{j}).$ For the fixed $p_{j}/q_{j}$ and the sequence $\\{\tilde{q}_{\ell}\\}$ defined by (8.27), there exists $s_{j}\in{\mathbb{N}}$ such that $\tilde{q}_{s_{j}}\leq q_{j}<\tilde{q}_{s_{j}+1}.$ Then we define the same sequences $\\{\tilde{q}_{\ell}\\}_{\ell=0}^{s_{j}}$ and $\\{N_{\ell}\\}_{\ell=0}^{s_{j}+1}$ for $p_{j}/q_{j}$ as $\alpha$ such that (8.28)-(8.30) hold. Following Lemma 8.5, we will inductively show that (8.36) $\displaystyle|L_{N_{\ell+1}}(p_{j}/q_{j},A_{j})+L_{N_{\ell}}(p_{j}/q_{j},A_{j})-2L_{2N_{\ell}}(p_{j}/q_{j},A_{j})|<C\mathrm{e}^{-c_{1}\tilde{q}_{\ell}^{\gamma/4}},$ (8.37) $\displaystyle|L_{2N_{\ell+1}}(p_{j}/q_{j},A_{j})-L_{N_{\ell+1}}(p_{j}/q_{j},A_{j})|<C\mathrm{e}^{-c_{2}\tilde{q}_{\ell}^{\gamma/4}},$ (8.38) $\displaystyle|L_{N_{\ell+1}}(p_{j}/q_{j},A_{j})-L_{N_{\ell}}(p_{j}/q_{j},A_{j})|\leq C\mathrm{e}^{-c_{3}\tilde{q}_{s_{\ell-1}}^{\gamma/4}}.$ To give the induction, the key is to apply Lemma 8.3, and verify that (8.39) $\displaystyle|p_{j}/q_{j}-\tilde{p}_{\ell}/\tilde{q}_{\ell}|$ $\displaystyle<$ $\displaystyle\tilde{q}_{\ell}^{-2},$ (8.40) $\displaystyle L_{2N_{\ell}}(p_{j}/q_{j},A_{j})$ $\displaystyle>$ $\displaystyle(19/20)L_{N_{\ell}}(p_{j}/q_{j},A_{j}),$ (8.41) $\displaystyle L_{N_{\ell}}(p_{j}/q_{j},A_{j})$ $\displaystyle>$ $\displaystyle 90\kappa.$ Indeed, by the property of continued fraction expansion, (8.39) holds for any $0\leq\ell\leq s_{j}$, and (8.40) follows from (8.35) and (8.37). On the other hand, if $\ell=0$, by (8.34), (8.35) and (8.36), we have $\displaystyle|L_{N_{1}}($ $\displaystyle p_{j}/q_{j},A_{j})-L_{N_{0}}(p_{j}/q_{j},A_{j})|$ $\displaystyle\leq\Big{|}L_{N_{1}}(p_{j}/q_{j},A_{j})+L_{N_{0}}(p_{j}/q_{j},A_{j})-2L_{2N_{0}}(p_{j}/q_{j},A_{j})\Big{|}$ $\displaystyle\ +2|L_{2N_{0}}(p_{j}/q_{j},A_{j})-L_{N_{0}}(p_{j}/q_{j},A_{j})|$ $\displaystyle\leq C\mathrm{e}^{-c_{1}\tilde{q}_{0}^{\gamma/4}}+2L_{N_{0}}(p_{j}/q_{j},A_{j})/50,$ which implies that (8.42) $\displaystyle L_{N_{1}}(p_{j}/q_{j},A_{j})\geq 48L_{N_{0}}(p_{j}/q_{j},A_{j})/50-C\mathrm{e}^{-c_{1}\tilde{q}_{0}^{\gamma/4}}>95\kappa.$ As for (8.41), in case $\ell\geq 1$, by (8.42) and (8.38), one has $\displaystyle L_{N_{\ell+1}}(p_{j}/q_{j},A_{j})$ $\displaystyle\geq$ $\displaystyle L_{N_{1}}(p_{j}/q_{j},A_{j})-\sum_{k=1}^{\ell}|L_{N_{k+1}}(p_{j}/q_{j},A_{j})-L_{N_{k}}(p_{j}/q_{j},A_{j})|$ $\displaystyle>$ $\displaystyle 95\kappa-\sum_{k=0}^{\ell-1}C\mathrm{e}^{-c_{3}\tilde{q}_{i}^{\gamma/4}}>90\kappa.$ Therefore, the iteration can be conducted $s_{j}$ times, and we obtain (8.43) $\displaystyle|L_{N_{s_{j}+1}}(p_{j}/q_{j},A_{j})+L_{N_{0}}(p_{j}/q_{j},A_{j})-2L_{2N_{0}}(p_{j}/q_{j},A_{j})|<\mathrm{e}^{-c^{\prime\prime}\tilde{q}_{0}^{\gamma/4}}.$ Moreover, by Lemma 8.6, we get $\displaystyle L_{N_{s_{j}+1}}(p_{j}/q_{j},A_{j})\leq L(p_{j}/q_{j},A_{j})+5C_{1}C_{1}(\kappa)^{-1}\tilde{q}_{s_{j}}^{-(\sigma-1)}.$ The inequality above, together with (8.43) yields $\displaystyle|L(p_{j}/q_{j},A_{j})+L_{N_{0}}(p_{j}/q_{j},A_{j})-2L_{2N_{0}}(p_{j}/q_{j},A_{j})|<2\mathrm{e}^{-c^{\prime\prime}\tilde{q}_{0}^{\gamma/4}}.$ ∎ ### 8.4. Proof of Theorem 6.3. Assume $\alpha\in DC(v,\tau)$ and $p_{n}/q_{n}$ be the continued fraction expansion of $\alpha$. Notice that since for each $N$, $L_{N}(\alpha,A)$ is a continuous function in both variables, thus, $L(\alpha,A)=\inf L_{N}(\alpha,A)$ is upper semi-continuous, then in the case $L(\alpha,A)=0$ we get $\displaystyle 0\leq\liminf_{n\rightarrow\infty}L(p_{n}/q_{n},A_{n})\leq\limsup_{n\rightarrow\infty}L(p_{n}/q_{n},A_{n})\leq L(\alpha,A)=0,$ that is $\lim_{n\rightarrow\infty}L(p_{n}/q_{n},A_{n})=0.$ Therefore we may assume $L(\alpha,A)>100\kappa>0$. Take $j>j_{1}$ and $C_{1}(\kappa)q_{0}^{\sigma}<N_{0}<C_{2}(\kappa)q_{0}^{\sigma_{1}}$, by Lemma 8.5 and Lemma 8.7, we have $\displaystyle|L(\alpha,A)+L_{N_{0}}(\alpha,A)-2L_{2N_{0}}(\alpha,A)|<\mathrm{e}^{-c^{\prime\prime}\tilde{q}_{0}^{\gamma/4}},$ and $\displaystyle|L(p_{j}/q_{j},A_{j})+L_{N_{0}}(p_{j}/q_{j},A_{j})-2L_{2N_{0}}(p_{j}/q_{j},A_{j})|<2\mathrm{e}^{-c^{\prime\prime}\tilde{q}_{0}^{\gamma/4}}.$ Hence, one can estimate $\displaystyle|L(\alpha,A)$ $\displaystyle-L(p_{j}/q_{j},A_{j})|\leq|L_{N_{0}}(p_{j}/q_{j},A_{j})-L_{N_{0}}(\alpha,A)|$ $\displaystyle\ +2|L_{2N_{0}}(p_{j}/q_{j},A_{j})-L_{2N_{0}}(\alpha,A)|+3\mathrm{e}^{-c^{\prime\prime}\tilde{q}_{0}^{\gamma/4}}$ $\displaystyle\leq C(\kappa)^{N_{0}}\\{|p_{j}/q_{j}-\alpha|+\|A-A_{j}\|_{\nu,\rho}\\}+3\mathrm{e}^{-c^{\prime\prime}\tilde{q}_{0}^{\gamma/4}},$ it follows that $\displaystyle\limsup\limits_{j\rightarrow\infty}|L(\alpha,A)-L(p_{j}/q_{j},A_{j}|\leq 4\mathrm{e}^{-c^{\prime\prime}\tilde{q}_{0}^{\gamma/4}},$ let $\tilde{q}_{0}\rightarrow\infty$, we get the result. ## Appendix: Proof of Lemma 4.5 Define $B_{r}(\delta)=\\{Y\in\mathcal{B}_{r}^{(nre)}:\|Y\|_{r}\leq\delta\\}$ and set $\varepsilon=8^{-2}\gamma^{2}Q_{n+1}^{-2\tau^{2}}.$ Then we define the nonlinear functional $\begin{split}\mathcal{F}:B_{r}(\varepsilon^{1/2})\rightarrow\mathcal{B}_{r}^{(nre)}\end{split}$ by (8.44) $\begin{split}\mathcal{F}(Y)=\mathbb{P}_{nre}\ln(\mathrm{e}^{A^{-1}Y(\theta+\alpha)A}\mathrm{e}^{g}\mathrm{e}^{-Y}),\end{split}$ where $\mathbb{P}_{nre}$ denotes the standard projections from $\mathcal{B}_{r}$ to $\mathcal{B}_{r}^{(nre)}.$ We will find a solution of functional equation (8.45) $\begin{split}\mathcal{F}(Y_{t})=(1-t)\mathcal{F}(Y_{0}),\ Y_{0}=0.\end{split}$ The derivative of $\mathcal{F}$ at $Y\in B_{r}(\varepsilon^{1/2})$ along $Y^{\prime}\in\mathcal{B}_{r}^{(nre)}$ is given by (8.46) $\begin{split}D\mathcal{F}(Y)Y^{\prime}=\mathbb{P}_{nre}\big{\\{}A^{-1}Y^{\prime}(\theta+\alpha)A-Y^{\prime}+O(\|A\|^{2}g)Y^{\prime}+P[A,Y,Y^{\prime},g](\theta)\big{\\}},\end{split}$ where $\begin{split}P[A,Y,Y^{\prime},g](\theta)&=O(A^{-1}Y(\theta+\alpha)A)A^{-1}Y^{\prime}(\theta+\alpha)A+2^{-1}[Y^{\prime\prime\prime},F+H]+\cdots\\\ &-O(Y)Y^{\prime}+2^{-1}[F+H^{\prime},-Y^{\prime\prime}]+\cdots-O(\|A\|^{2}g)Y^{\prime},\\\ O(\|A\|^{2}g)Y^{\prime}&=O(g)A^{-1}Y^{\prime}(\theta+\alpha)A+O(g)Y^{\prime},\end{split}$ with $P[A,Y,Y^{\prime},0](\theta)=0,$ $\begin{split}Y^{\prime\prime\prime}(\theta+\alpha)&=A^{-1}Y^{\prime}(\theta+\alpha)A+O(A^{-1}Y(\theta+\alpha)A)A^{-1}Y^{\prime}(\theta+\alpha)A,\\\ Y^{\prime\prime}(\theta)&=Y^{\prime}(\theta)+O(Y(\theta))Y^{\prime}(\theta),\\\ F(\theta)&=A^{-1}Y(\theta+\alpha)A+g(\theta)-Y(\theta),\end{split}$ and $H,H^{\prime}$ being sums of terms at least 2 orders in $A^{-1}Y(\theta+\alpha)A,g(\theta),-Y(\theta).$ Moreover, the first $``\cdots"$ denotes the sum of terms which are at least 2 orders in $F+H$ but only 1 order in $Y^{\prime\prime\prime}.$ The second $``\cdots"$ denotes the sum of terms which are at least 2 orders in $F+H^{\prime}$ but only 1 order in $Y^{\prime\prime}.$ We give a estimate about the operator $D\mathcal{F}(Y)^{-1}$. ###### Proposition 8. For the fixed $Y\in\mathcal{B}_{r}(\varepsilon^{1/2}),$ $D\mathcal{F}(Y)$ (defined by (8.46)) is a linear map from $\mathcal{B}_{r}^{(nre)}$ to $\mathcal{B}_{r}^{(nre)}$ with estimate (8.47) $\begin{split}\|D\mathcal{F}(Y)^{-1}\|\leq 2^{-1}\varepsilon^{-1/2}.\end{split}$ ###### Proof. For the fixed $Y\in\mathcal{B}_{r}(\varepsilon^{1/2}),$ obviously, the operator $D\mathcal{F}(Y)$ defined by (8.46) is a linear map from $\mathcal{B}_{r}^{(nre)}$ to $\mathcal{B}_{r}^{(nre)}$. In the following we prove the estimate in (8.47). To this end, we consider the operator $D\mathcal{F}(0)$ given by $\begin{split}D\mathcal{F}(0)Y^{\prime}=A^{-1}Y^{\prime}(\theta+\alpha)A-Y^{\prime}+\mathbb{P}_{nre}O(\|A\|^{2}g)Y^{\prime},Y^{\prime}\in\mathcal{B}_{r}^{(nre)}.\end{split}$ Note the operator $D\mathcal{F}(0)$ is a linear map mapping $\mathcal{B}_{r}^{(nre)}$ to $\mathcal{B}_{r}^{(nre)}$. Next we give the estimate about $D\mathcal{F}(0)^{-1}.$ Note $\overline{Q}_{n+1}\geq T>(2\gamma^{-1})^{2\tau},n\geq 0$ ((4.4)), then by (4.18) in Lemma 4.4 we get $\begin{split}\|k\alpha\pm 2\rho_{f}\|_{\mathbb{Z}}\geq\gamma Q_{n+1}^{-\tau^{2}}=8\varepsilon^{\frac{1}{2}},|k|<\overline{Q}_{n+1}^{\frac{1}{2}}.\end{split}$ By the inequality above, one can check, for $Y^{\prime}\in\mathcal{B}_{r}^{(nre)},$ $\begin{split}A^{-1}&Y^{\prime}(\cdot+\alpha)A-Y^{\prime}\in\mathcal{B}_{r}^{(nre)},\\\ \|A^{-1}&Y^{\prime}(\cdot+\alpha)A-Y^{\prime}\|_{r}\geq 8\|A\|^{2}\varepsilon^{\frac{1}{2}}\|Y^{\prime}\|_{r}.\end{split}$ Moreover, by Lemma 3.1 (Banach algebra property) we get $\|O(\|A\|^{2}g)\|_{r}\leq 2\|A\|^{2}\varepsilon.$ Note $(8\varepsilon^{1/2})^{-1}2\varepsilon<1,$ then (8.48) $\begin{split}\|D\mathcal{F}(0)^{-1}\|_{r}\leq 2(8\varepsilon^{1/2})^{-1}=4^{-1}\varepsilon^{-1/2}.\end{split}$ Once we get (8.48), we will turn to $\|D\mathcal{F}(Y)^{-1}\|.$ The calculations below also depends on Lemma 3.1, we omit the reference about this lemma. Note $\\{D\mathcal{F}(Y)-D\mathcal{F}(0)\\}Y^{\prime}=\mathbb{P}_{nre}\big{\\{}P(A,Y,Y^{\prime}g)-P(A,0,Y^{\prime}g)\\},$ we get (8.49) $\begin{split}\sup_{\|Y\|_{r}\leq\varepsilon^{\frac{1}{2}},\|g\|_{r}\leq\varepsilon}\|D\mathcal{F}(Y)-D\mathcal{F}(0)\|\leq 2\varepsilon^{\frac{1}{2}}.\end{split}$ (8.48) and (8.49) yield $\begin{split}\|D\mathcal{F}(0)^{-1}(D\mathcal{F}(Y)-D\mathcal{F}(0))\|\leq 4^{-1}\varepsilon^{-1/2}2\varepsilon^{\frac{1}{2}}=2^{-1}.\end{split}$ Finally, note $\begin{split}D\mathcal{F}(Y)^{-1}=\\{1+D\mathcal{F}(0)^{-1}(D\mathcal{F}(Y)-D\mathcal{F}(0))\\}^{-1}D\mathcal{F}(0)^{-1},\end{split}$ then by the inequality above we know that $D\mathcal{F}(Y)$ is invertible with $\begin{split}\|D\mathcal{F}(Y)^{-1}\|\leq 2\|D\mathcal{F}(0)^{-1}\|\leq 2^{-1}\varepsilon^{-1/2}.\end{split}$ ∎ Now we turn to functional equation (8.45). Formally, we get (8.50) $\begin{split}Y_{t}=-\int_{0}^{t}D\mathcal{F}(Y_{s})^{-1}\mathcal{F}(Y_{0})ds=-\int_{0}^{t}D\mathcal{F}(Y_{s})^{-1}\mathbb{P}_{nre}gds,0\leq t\leq 1.\end{split}$ Moreover, by (8.47) $\begin{split}\|Y_{t}\|_{r}\leq\sup_{Y\in\mathcal{B}_{r}(\varepsilon^{1/2}),\|g\|_{r}\leq\varepsilon}\|D\mathcal{F}(Y)^{-1}\|\|g\|_{r}\leq 2^{-1}\varepsilon^{-1/2}\varepsilon<\varepsilon^{1/2},0\leq t\leq 1.\end{split}$ Therefore, the solution of (8.45) exists in $\mathcal{B}_{r}(\varepsilon^{1/2})$ and is given by (8.50). For $Y_{t},\ 0\leq t\leq 1,$ given above, we know that $\mathcal{F}(Y_{1})=0,$ that is (by (8.44)) $\begin{split}\mathbb{P}_{nre}\ln(\mathrm{e}^{A^{-1}Y_{1}(\theta+\alpha)A}\mathrm{e}^{g}\mathrm{e}^{-Y_{1}})=0,\end{split}$ which implies that there exists a matrix $g^{(re)}\in\mathcal{B}_{r}^{(re)}$ such that $\begin{split}\ln(\mathrm{e}^{A^{-1}Y_{1}(\theta+\alpha)A}\mathrm{e}^{g}\mathrm{e}^{-Y_{1}})=g^{(re)}.\end{split}$ That is $\begin{split}\mathrm{e}^{Y_{1}(\theta+\alpha)}A\mathrm{e}^{g}\mathrm{e}^{-Y_{1}}=A\mathrm{e}^{g^{(re)}}.\end{split}$ By standard calculations we get the estimate $\|g^{(re)}\|\leq 2\varepsilon.$ This $Y_{1},$ with the estimate $\|Y_{1}\|_{r}<\varepsilon^{\frac{1}{2}},$ is the one we want. ## Acknowledgments J. You and Q. Zhou were supported by National Key R&D Program of China (2020YFA0713300) and Nankai Zhide Foundation. H. Cheng was supported by NSFC grant (12001294). She would like to thank Y. Pan for useful discussion. L. Ge was partially supported by NSF DMS-1901462 and AMS-Simons Travel Grant 2020–2022. J. You was also partially supported by NSFC grant (11871286). Q. Zhou was also partially supported by NSFC grant (12071232), The Science Fund for Distinguished Young Scholars of Tianjin (No. 19JCJQJC61300). ## References * AD [08] A. Avila and D. Damanik. Absolute continuity of the integrated density of states for the almost Mathieu operator with non-critical coupling. Invent. Math., 172(2):439–453, 2008. * AFK [11] A. Avila, B. Fayad, and R. Krikorian. A KAM scheme for ${\rm SL}(2,\mathbb{R})$ cocycles with Liouvillean frequencies. Geom. Funct. Anal., 21(5):1001–1019, 2011. * AJ [09] A. Avila and S. Jitomirskaya. The Ten Martini Problem. Ann. of Math. (2), 170(1):303–342, 2009. * AJS [14] A. Avila, S. Jitomirskaya, and C. Sadel. Complex one-frequency cocycles. J. Eur. Math. Soc. (JEMS), 16(9):1915–1935, 2014. * AK [06] A. Avila and R. Krikorian. Reducibility or nonuniform hyperbolicity for quasiperiodic Schrödinger cocycles. Ann. of Math. (2), 164(3):911–940, 2006. * AK [15] A. Avila and R. Krikorian. Monotonic cocycles. Invent. Math., 202(1):271–331, 2015. * AK [16] A. Avila and R. Krikorian. Some remarks on local and semi-local results for schrödinger cocycles. 2016\. * [8] A. Avila. Global theory of one-frequency Schrödinger operators. Acta Math., 215(1):1–54, 2015. * [9] A. Avila. On the Kotani-Last and Schrödinger conjectures. J. Amer. Math. Soc., 28(2):579–616, 2015. * AvMS [90] J. Avron, P. H. M. van Mouche, and B. Simon. On the measure of the spectrum for the almost Mathieu operator. Comm. Math. Phys., 132(1):103–118, 1990. * AYZ [17] A. Avila, J. You, and Q. Zhou. Sharp phase transitions for the almost Mathieu operator. Duke Math. J., 166(14):2697–2718, 2017. * BCL [21] A. Bounemoura, C. Chavaudret, and S. Liang. Reducibility of ultra-differentiable quasi-periodic cocycles under an adapted arithmetic condition. 2021\. * [13] A. Bounemoura and J. Fejoz. Hamiltonian perturbation theory for ultra-differentiable functions. Mem. Amer. Math. Soc., to appear. * Bha [87] R. Bhatia. Perturbation bounds for matrix eigenvalues, volume 162 of Pitman Research Notes in Mathematics Series. Longman Scientific & Technical, Harlow; John Wiley & Sons, Inc., New York, 1987. * BHJ [19] Simon Becker, Rui Han, and Svetlana Jitomirskaya. Cantor spectrum of graphene in magnetic fields. Invent. Math., 218(3):979–1041, 2019. * BJ [02] J. Bourgain and S. Jitomirskaya. Continuity of the Lyapunov exponent for quasiperiodic operators with analytic potential. volume 108, pages 1203–1218. 2002. Dedicated to David Ruelle and Yasha Sinai on the occasion of their 65th birthdays. * Bou [05] J. Bourgain. Green’s function estimates for lattice Schrödinger operators and applications, volume 158 of Annals of Mathematics Studies. Princeton University Press, Princeton, NJ, 2005. * BS [82] J. Béllissard and B. Simon. Cantor spectrum for the almost Mathieu equation. J. Functional Analysis, 48(3):408–419, 1982. * CC [94] J. Chaumat and A.-M. Chollet. Surjectivité de l’application restriction à un compact dans des classes de fonctions ultradifférentiables. Math. Ann., 298(1):7–40, 1994. * CCYZ [19] A. Cai, C. Chavaudret, J. You, and Q. Zhou. Sharp Hölder continuity of the Lyapunov exponent of finitely differentiable quasi-periodic cocycles. Math. Z., 291(3-4):931–958, 2019. * Cha [65] W. G. Chambers. Linear-network model for magnetic breakdown in two dimensions. Phys. Rev. A, 140(1A):135–143, 1965. * Cha [13] C. Chavaudret. Strong almost reducibility for analytic and Gevrey quasi-periodic cocycles. Bull. Soc. Math. France, 141(1):47–106, 2013. * Dia [06] J.L. Dias. A normal form theorem for Brjuno skew systems through renormalization. J. Differential Equations, 230(1):1–23, 2006. * DS [75] E. I. Dinaburg and Ja. G. Sinaĭ. The one-dimensional Schrödinger equation with quasiperiodic potential. Funkcional. Anal. i Priložen., 9(4):8–21, 1975. * Eli [92] L. H. Eliasson. Floquet solutions for the $1$-dimensional quasi-periodic Schrödinger equation. Comm. Math. Phys., 146(3):447–482, 1992. * FK [09] B. Fayad and R. Krikorian. Rigidity results for quasiperiodic ${\rm SL}(2,\mathbb{R})$-cocycles. J. Mod. Dyn., 3(4):497–510, 2009. * FK [18] B. Fayad and R. Krikorian. Some questions around quasi-periodic dynamics. In Proceedings of the International Congress of Mathematicians—Rio de Janeiro 2018. Vol. III., pages 1909–1932. World Sci. Publ., Hackensack, NJ, 2018. * GS [01] M. Goldstein and W. Schlag. Hölder continuity of the integrated density of states for quasi-periodic Schrödinger equations and averages of shifts of subharmonic functions. Ann. of Math. (2), 154(1):155–203, 2001. * Her [83] M. R. Herman. Une methode pour minorer les exposants de Lyapounov et quelques exemples montrant le caractère local dun theoreme d${}^{\prime}{A}$rnold et de Moser sur le tore de dimension $2$. Comment. Math. Helv., 58(3):453–502, 1983. * HY [12] X. Hou and J. You. Almost reducibility and non-perturbative reducibility of quasi-periodic linear systems. Invent. Math., 190(1):209–260, 2012. * JK [02] S. Jitomirskaya and I. V. Krasovsky. Continuity of the measure of the spectrum for discrete quasiperiodic operators. Math. Res. Lett., 9(4):413–421, 2002. * JKS [09] S. Jitomirskaya, D. A. Koslover, and M.S. Schulteis. Continuity of the Lyapunov exponent for analytic quasiperiodic cocycles. Ergodic Theory Dynam. Systems, 29(6):1881–1905, 2009. * JM [82] R. Johnson. and J. Moser. The rotation number for almost periodic potentials. Comm. Math. Phys., 84(3):403–438, 1982. * JM [11] S. Jitomirskaya and C. A. Marx. Continuity of the Lyapunov exponent for analytic quasi-periodic cocycles with singularities. J. Fixed Point Theory Appl., 10(1):129–146, 2011. * [35] S. Jitomirskaya and C. A. Marx. Analytic quasi-periodic Schrödinger operators and rational frequency approximants. Geom. Funct. Anal., 22(5):1407–1443, 2012. * [36] S. Jitomirskaya and C. A. Marx. Analytic quasi-perodic cocycles with singularities and the Lyapunov exponent of extended Harper’s model. Comm. Math. Phys., 316(1):237–267, 2012. * Kle [05] S. Klein. Anderson localization for the discrete one-dimensional quasi-periodic Schrödinger operator with potential defined by a Gevrey-class function. J. Funct. Anal., 218(2):255–292, 2005. * Kot [84] S. Kotani. Ljapunov indices determine absolutely continuous spectra of stationary random one-dimensional Schrödinger operators. In Stochastic analysis (Katata/Kyoto, 1982), volume 32 of North-Holland Math. Library, pages 225–247. North-Holland, Amsterdam, 1984\. * [39] R. Krikorian. Priviate communications. * KWYZ [18] R. Krikorian, J. Wang, J. You, and Q. Zhou. Linearization of quasiperiodically forced circle flows beyond Brjuno condition. Comm. Math. Phys., 358(1):81–100, 2018. * Las [92] Y. Last. On the measure of gaps and spectra for discrete $1$D Schrödinger operators. Comm. Math. Phys., 149(2):347–360, 1992. * Las [93] Y. Last. A relation between a.c. spectrum of ergodic Jacobi matrices and the spectra of periodic approximants. Comm. Math. Phys., 151(1):183–192, 1993. * Las [94] Y. Last. Zero measure spectrum for the almost Mathieu operator. Comm. Math. Phys., 164(2):421–432, 1994. * Las [05] Yoram Last. Spectral theory of Sturm-Liouville operators on infinite intervals: a review of recent developments. In Sturm-Liouville theory, pages 99–120. Birkhäuser, Basel, 2005. * LWJZ [20] L.Ge, Y. Wang, J.You, and X. Zhao. Transition space for the continuity of the lyapunov exponent of quasiperiodic schrödinger cocycles. 2020\. * MJ [17] C. A. Marx and S. Jitomirskaya. Dynamics and spectral theory of quasi-periodic Schrödinger-type operators. Ergodic Theory Dynam. Systems, 37(8):2353–2393, 2017. * Sha [11] Mira Shamis. Some connections between almost periodic and periodic discrete Schrödinger operators with analytic potentials. J. Spectr. Theory, 1(3):349–362, 2011. * Thi [03] V. Thilliez. Division by flat ultradifferentiable functions and sectorial extensions. Results Math., 44(1-2):169–188, 2003. * VPMG [93] S. A. Molchanov V. P. Maslov and A. Y. Gordon. Behavior of generalized eigenfunctions at infinity and the Schrödinger conjecture. Russian J. Math. Phys., 1(1):71–104, 1993. * WY [13] Y. Wang and J. You. Examples of discontinuity of Lyapunov exponent in smooth quasiperiodic cocycles. Duke Math. J., 162(13):2363–2412, 2013. * You [18] J. You. Quantitative almost reducibility and its applications. In Proceedings of the International Congress of Mathematicians—Rio de Janeiro 2018. Vol. III., pages 2113–2135. World Sci. Publ., Hackensack, NJ, 2018. * YZ [13] J. You and Q. Zhou. Embedding of analytic quasi-periodic cocycles into analytic quasi-periodic linear systems and its applications. Comm. Math. Phys., 323(3):975–1005, 2013. * YZ [14] J. You and Q. Zhou. Phase transition and semi-global reducibility. Comm. Math. Phys., 330(3):1095–1113, 2014. * Zha [19] X. Zhao. On Last’s intersection spectrum conjecture. Nonlinearity, 32(7):2352–2364, 2019. * ZW [12] Q. Zhou and J. Wang. Reducibility results for quasiperiodic cocycles with Liouvillean frequency. J. Dynam. Differential Equations, 24(1):61–83, 2012.
# Optimal Oracles for Point-to-Set Principles D. M. Stull Department of Computer Science, Iowa State University Ames, IA 50011, USA <EMAIL_ADDRESS> ###### Abstract The point-to-set principle [14] characterizes the Hausdorff dimension of a subset $E\subseteq\mathbb{R}^{n}$ by the effective (or algorithmic) dimension of its individual points. This characterization has been used to prove several results in classical, i.e., without any computability requirements, analysis. Recent work has shown that algorithmic techniques can be fruitfully applied to Marstrand’s projection theorem, a fundamental result in fractal geometry. In this paper, we introduce an extension of point-to-set principle - the notion of optimal oracles for subsets $E\subseteq\mathbb{R}^{n}$. One of the primary motivations of this definition is that, if $E$ has optimal oracles, then the conclusion of Marstrand’s projection theorem holds for $E$. We show that every analytic set has optimal oracles. We also prove that if the Hausdorff and packing dimensions of $E$ agree, then $E$ has optimal oracles. Moreover, we show that the existence of sufficiently nice outer measures on $E$ implies the existence of optimal Hausdorff oracles. In particular, the existence of exact gauge functions for a set $E$ is sufficient for the existence of optimal Hausdorff oracles, and is therefore sufficient for Marstrand’s theorem. Thus, the existence of optimal oracles extends the currently known sufficient conditions for Marstrand’s theorem to hold. Under certain assumptions, every set has optimal oracles. However, assuming the axiom of choice and the continuum hypothesis, we construct sets which do not have optimal oracles. This construction naturally leads to a generalization of Davies theorem on projections. ## 1 Introduction Effective, i.e., algorithmic, dimensions were introduced [12, 1] to study the randomness of points in Euclidean space. The effective dimension, $\dim(x)$ and effective strong dimension, $\operatorname{Dim}(x)$, are real values which measure the asymptotic density of information of an individual point $x$. The connection between effective dimensions and the classical Hausdorff and packing dimension is given by the point-to-set principle of J. Lutz and N. Lutz [14]: For any $E\subseteq\mathbb{R}^{n}$, $\displaystyle\dim_{H}(E)$ $\displaystyle=\min\limits_{A\subseteq\mathbb{N}}\sup_{x\in E}\dim^{A}(x),\text{ and}$ (1) $\displaystyle\dim_{P}(E)$ $\displaystyle=\min\limits_{A\subseteq\mathbb{N}}\sup_{x\in E}\operatorname{Dim}^{A}(x)\,.$ (2) Call an oracle $A$ satisfying (1) a Hausdorff oracle for $E$. Similarly, we call an oracle $A$ satisfying (2) a packing oracle for $E$. Thus, the point- to-set principle shows that the classical notion of Hausdorff or packing dimension is completely characterized by the effective dimension of its individual points, relative to a Hausdorff or packing oracle, respectively. Recent work as shown that algorithmic dimensions are not only useful in effective settings, but, via the point-to-set principle, can be used to solve problems in geometric measure theory [15, 17, 18, 29]. It is important to note that the point-to-set principle allows one to use algorithmic techniques to prove theorems whose statements have seemingly nothing to do with computability theory. In this paper, we focus on the connection between algorithmic dimension and Marstrand’s projection theorem. Marstrand, in his landmark paper [20], was the first to study how the dimension of a set is changed when projected onto a line. He showed that, for any analytic set $E\in\mathbb{R}^{2}$, for almost every angle $\theta\in[0,\pi)$, $\dim_{H}(p_{\theta}\,E)=\min\\{\dim_{H}(E),1\\},$ (3) where $p_{\theta}(x,y)=x\cos\theta+y\sin\theta$111This result was later generalized to $\mathbb{R}^{n}$, for arbitrary $n$, as well as extended to hyperspaces of dimension $m$, for any $1\leq m\leq n$ (see e.g. [21, 22, 23]).. The study of projections has since become a central theme in fractal geometry (see [8] or [24] for a more detailed survey of this development). Marstrand’s theorem begs the question of whether the analytic requirement on $E$ can be dropped. It is known that, without further conditions, it cannot. Davies [5] showed that, assuming the axiom of choice and the continuum hypothesis, there are non-analytic sets for which Marstrands conclusion fails. However, the problem of classifying the sets for which Marstrands theorem does hold is still open. Recently, Lutz and Stull [19] used the point-to-set principle to prove that the projection theorem holds for sets for which the Hausdorff and packing dimensions agree222Orponen [28] has recently given another proof of Lutz and Stull’s result using more classical tools.. This expanded the reach of Marstrand’s theorem, as this assumption is incomparable with analyticity. In this paper, we give the broadest known sufficient condition (which makes essential use of computability theory) for Marstrand’s theorem. In particular, we introduce the notion of optimal Hausdorff oracles for a set $E\subseteq\mathbb{R}^{n}$. We prove that Marstrand’s theorem holds for every set $E$ which has optimal Hausdorff oracles. An optimal Hausdorff oracle for a set $E$ is a Hausdorff oracle which minimizes the algorithmic complexity of ”most“333By most, we mean a subset of $E$ of the same Hausdorff dimension as $E$ points in $E$. It is not immediately clear that any set $E$ has optimal oracles. Nevertheless, we show that two natural classes of sets $E\subseteq\mathbb{R}^{n}$ do have optimal oracles. We show that every analytic, and therefore Borel, set has optimal oracles. We also prove that every set whose Hausdorff and packing dimensions agree has optimal Hausdorff oracles. Thus, we show that the existence of optimal oracles encapsulates the known conditions sufficient for Marstrand’s theorem to hold. Moreover, we show that the existence of sufficiently nice outer measures on $E$ implies the existence of optimal Hausdorff oracles. In particular, the existence of exact gauge functions (Section 2.1) for a set $E$ is sufficient for the existence of optimal Hausdorff oracles for $E$, and is therefore sufficient for Marstrand’s theorem. Thus, the existence of optimal Hausdorff oracles is weaker than the previously known conditions for Marstrand’s theorem to hold. We also show that the notion of optimal oracles gives insight to sets for which Marstrand’s theorem does not hold. Assuming the axiom of choice and the continuum hypothesis, we construct sets which do not have optimal oracles. This construction, with minor adjustments, proves a generalization of Davies theorem proving the existence of sets for which (3) does not hold. In addition, the inherently algorithmic aspect of the construction might be useful for proving set-theoretic properties of exceptional sets for Marstrand’s theorem. Finally, we define optimal packing oracles for a set. We show that every analytic set $E$ has optimal packing oracles. We also show that every $E$ whose Hausdorff and packing dimensions agree have optimal packing oracles. Assuming the axiom of choice and the continuum hypothesis, we show that there are sets with optimal packing oracles without optimal Hausdorff oracles (and vice-versa). The structure of the paper is as follows. In Section 2.1 we review the concepts of measure theory needed, and the (classical) definition of Hausdorff dimension. In Section 2.2 we review algorithmic information theory, including the formal definitions of effective dimensions. We then introduce and study the notion of optimal oracles in Section 3. In particular, we give a general condition for the existence of optimal oracles in Section 3.1. We use this condition to prove that analytic sets have optimal oracles in Section 3.2. We conclude in Section 3.3 with an example, assuming the axiom of choice and the continuum hypothesis, of a set without optimal oracles. The connection between Marstrands projection theorem and optimal oracles is explored in Section 4. In this section, we prove that Marstrands theorem holds for every set with optimal oracles. In Section 4.1, we use the construction of a set without optimal oracles to give a new, algorithmic, proof of Davies theorem. Finally, in Sectino 5, we define and investigate the notion of optimal packing oracles. ## 2 Preliminaries ### 2.1 Outer Measures and Classical Dimension A set function $\mu:\mathcal{P}(\mathbb{R}^{n})\to[0,\infty]$ is called an outer measure on $\mathbb{R}^{n}$ if 1. 1. $\mu(\emptyset)=0$, 2. 2. if $A\subseteq B$ then $\mu(A)\leq\mu(B)$, and 3. 3. for any sequence $A_{1},A_{2},\ldots$ of subsets, $\mu(\bigcup_{i}A_{i})\leq\sum_{i}\mu(A_{i})$. If $\mu$ is an outer measure, we say that a subset $A$ is $\mu$-measurable if $\mu(A\cap B)+\mu(B-A)=\mu(B)$, for every subset $B\subseteq\mathbb{R}^{n}$. An outer measure $\mu$ is called a metric outer measure if every Borel subset is $\mu$-measurable and $\mu(A\cup B)=\mu(A)+\mu(B)$, for every pair of subsets $A,B$ which have positive Hausdorff distance. That is, $\inf\\{\|x-y\|\,|\,x\in A,y\in B\\}>0$. An important example of a metric outer measure is the $s$-dimensional Hausdorff measure. For every $E\subseteq[0,1)^{n}$, define the $s$-dimensional Hausdorff content at precision $r$ by $h^{s}_{r}(E)=\inf\left\\{\sum_{i}d(Q_{i})^{s}\,|\,\bigcup_{i}Q_{i}\text{ covers }E\text{ and }d(Q_{i})\leq 2^{-r}\right\\}$, where $d(Q)$ is the diameter of ball $Q$. We define the $s$-dimensional Hausdorff measure of $E$ by $\mathcal{H}^{s}(E)=\lim\limits_{r\to\infty}h^{s}_{r}(E)$. ###### Remark. It is well-known that $\mathcal{H}^{s}$ is a metric outer measure for every $s$. The Hausdorff dimension of a set $E$ is then defined by $\dim_{H}(E)=\inf\limits_{s}\\{\mathcal{H}^{s}(E)=\infty\\}=\sup\limits_{s}\\{\mathcal{H}^{s}(E)=0\\}$. Another important metric outer measure, which gives rise to the packing dimension of a set, is the $s$-dimensional packing measure. For every $E\subseteq[0,1)^{n}$, define the $s$-dimensional packing pre-measure by $p^{s}(E)=\limsup\limits_{\delta\to 0}\left\\{\sum\limits_{i\in\mathbb{N}}d(B_{i})^{s}\,|\,\\{B_{i}\\}\text{ is a set of disjoint balls and }B_{i}\in C(E,\delta)\right\\}$, where $C(E,\delta)$ is the set of all closed balls with diameter at most $\delta$ with centers in $E$. We define the $s$-dimensional packing measure of $E$ by $\mathcal{P}^{s}(E)=inf\left\\{\sum\limits_{j}p^{s}(E_{j})\,|\,E\subseteq\bigcup E_{j}\right\\}$, where the infimum is taken over all countable covers of $E$. For every $s$, the $s$-dimensional packing measure is a metric outer measure. The packing dimension of a set $E$ is then defined by $\dim_{P}(E)=\inf\limits_{s}\\{\mathcal{P}^{s}(E)=0\\}=\sup\limits_{s}\\{\mathcal{P}^{s}(E)=\infty\\}$. In order to prove that every analytic set has optimal oracles, we will make use of the following facts of geometric measure theory (see, e.g., [7], [2]). ###### Theorem 1. The following are true. 1. 1. Suppose $E\subseteq\mathbb{R}^{n}$ is compact and satisfies $\mathcal{H}^{s}(E)>0$. Then there is a compact subset $F\subseteq E$ such that $0<\mathcal{H}^{s}(F)<\infty$. 2. 2. Every analytic set $E\subseteq\mathbb{R}^{n}$ has a $\Sigma^{0}_{2}$ subset $F\subseteq E$ such that $\dim_{H}(F)=\dim_{H}(E)$. 3. 3. Suppose $E\subseteq\mathbb{R}^{n}$ is compact and satisfies $\mathcal{P}^{s}(E)>0$. Then there is a compact subset $F\subseteq E$ such that $0<\mathcal{P}^{s}(F)<\infty$. 4. 4. Every analytic set $E\subseteq\mathbb{R}^{n}$ has a $\Sigma^{0}_{2}$ subset $F\subseteq E$ such that $\dim_{P}(F)=\dim_{P}(E)$. It is possible to generalize the definition of Hausdorff measure using gauge functions. A function $\phi:[0,\infty)\to[0,\infty)$ is a gauge function if $\phi$ is monotonically increasing, strictly increasing for $t>0$ and continuous. If $\phi$ is a gauge, define the $phi$-Hausdorff content at precision $r$ by $h^{\phi}_{r}(E)=\inf\left\\{\sum_{i}\phi(d(Q_{i}))\,|\,\bigcup_{i}Q_{i}\text{ covers }E\text{ and }d(Q_{i})\leq 2^{-r}\right\\}$, where $d(Q)$ is the diameter of ball $Q$. We define the $phi$-Hausdorff measure of $E$ by $\mathcal{H}^{\phi}(E)=\lim\limits_{r\to\infty}h^{\phi}_{r}(E)$. Thus we recover the $s$-dimensional Hausdorff measure when $\phi(t)=t^{s}$. Gauged Hausdorff measures give fine-grained information about the size of a set. There are sets $E$ which Hausdorff dimension $s$, but $\mathcal{H}^{s}(E)=0$ or $\mathcal{H}^{s}(E)=\infty$. However, it is sometimes possible to find an appropriate gauge so that $0<\mathcal{H}^{\phi}(E)<\infty$. When $0<\mathcal{H}^{\phi}(E)<\infty$, we say that $\phi$ is an exact gauge for $E$. ###### Example. For almost every Brownian path $X$ in $\mathbb{R}^{2}$, $\mathcal{H}^{2}(X)=0$, but $0<\mathcal{H}^{\phi}(X)<\infty$, where $\phi(t)=t^{2}\log\frac{1}{t}\log\log\frac{1}{t}$. For two outer measures $\mu$ and $\nu$, $\mu$ is said to be absolutely continuous with respect to $\nu$, denoted $\mu\ll\nu$, if $\mu(A)=0$ for every set $A$ for which $\nu(A)=0$. ###### Example. For every $s$, let $\phi_{s}(t)=t^{s}\log\frac{1}{t}$. Then $\mathcal{H}^{s}\ll\mathcal{H}^{\phi_{s}}$. ###### Example. For every $s$, let $\phi_{s}(t)=\frac{t^{s}}{\log\frac{1}{t}}$. Then $\mathcal{H}^{\phi_{s}}\ll\mathcal{H}^{s}$. ### 2.2 Algorithmic Information Theory The _conditional Kolmogorov complexity_ of a binary string $\sigma\in\\{0,1\\}^{*}$ given binary string $\tau\in\\{0,1\\}^{*}$ is $K(\sigma|\tau)=\min_{\pi\in\\{0,1\\}^{*}}\left\\{\ell(\pi):U(\pi,\tau)=\sigma\right\\}\,,$ where $U$ is a fixed universal prefix-free Turing machine and $\ell(\pi)$ is the length of $\pi$. The _Kolmogorov complexity_ of $\sigma$ is $K(\sigma)=K(\sigma|\lambda)$, where $\lambda$ is the empty string. An important fact is that the choice of universal machine affects the Kolmogorov complexity by at most an additive constant (which, especially for our purposes, can be safely ignored). See [11, 27, 6] for a more comprehensive overview of Kolmogorov complexity. We can naturally extend these definitions to Euclidean spaces by introducing “precision” parameters [16, 14]. Let $x\in\mathbb{R}^{m}$, and $r,s\in\mathbb{N}$. The _Kolmogorov complexity of $x$ at precision $r$_ is $K_{r}(x)=\min\left\\{K(p)\,:\,p\in B_{2^{-r}}(x)\cap\mathbb{Q}^{m}\right\\}\,.$ The _conditional Kolmogorov complexity of $x$ at precision $r$ given $q\in\mathbb{Q}^{m}$_ is $\hat{K}_{r}(x|q)=\min\left\\{K(p\,|\,q)\,:\,p\in B_{2^{-r}}(x)\cap\mathbb{Q}^{m}\right\\}\,.$ The _conditional Kolmogorov complexity of $x$ at precision $r$ given $y\in\mathbb{R}^{n}$ at precision $s$_ is $K_{r,s}(x|y)=\max\big{\\{}\hat{K}_{r}(x|q)\,:\,q\in B_{2^{-s}}(y)\cap\mathbb{Q}^{n}\big{\\}}\,.$ We typically abbreviate $K_{r,r}(x|y)$ by $K_{r}(x|y)$. The _effective Hausdorff dimension_ and _effective packing dimension_ 444Although effective Hausdorff was originally defined by J. Lutz [13] using martingales, it was later shown by Mayordomo [25] that the definition used here is equivalent. For more details on the history of connections between Hausdorff dimension and Kolmogorov complexity, see [6, 26]. of a point $x\in\mathbb{R}^{n}$ are $\dim(x)=\liminf_{r\to\infty}\frac{K_{r}(x)}{r}\quad\text{and}\quad\operatorname{Dim}(x)=\limsup_{r\to\infty}\frac{K_{r}(x)}{r}\,.$ By letting the underlying fixed prefix-free Turing machine $U$ be a universal _oracle_ machine, we may _relativize_ the definition in this section to an arbitrary oracle set $A\subseteq\mathbb{N}$. The definitions of $K^{A}_{r}(x)$, $\dim^{A}(x)$, $\operatorname{Dim}^{A}(x)$, etc. are then all identical to their unrelativized versions, except that $U$ is given oracle access to $A$. Note that taking oracles as subsets of the naturals is quite general. We can, and frequently do, encode a point $y$ into an oracle, and consider the complexity of a point relative to $y$. In these cases, we typically forgo explicitly referring to this encoding, and write e.g. $K^{y}_{r}(x)$. We can also join two oracles $A,B\subseteq\mathbb{N}$ using any computable bijection $f:\mathbb{N}\times\mathbb{N}\to\mathbb{N}$. We denote the join of $A$ and $B$ by $(A,B)$. We can generalize this procedure to join any countable sequence of oracles. As mentioned in the introduction, the connection between effective dimensions and the classical Hausdorff and packing dimensions is given by the point-to- set principle introduced by J. Lutz and N. Lutz [14]. ###### Theorem 2 (Point-to-set principle). Let $n\in\mathbb{N}$ and $E\subseteq\mathbb{R}^{n}$. Then $\displaystyle\dim_{H}(E)$ $\displaystyle=\min\limits_{A\subseteq\mathbb{N}}\sup_{x\in E}\dim^{A}(x),\text{ and}$ $\displaystyle\dim_{P}(E)$ $\displaystyle=\min\limits_{A\subseteq\mathbb{N}}\sup_{x\in E}\operatorname{Dim}^{A}(x)\,.$ An oracle testifying to the the first equality is called a Hausdorff oracle for E. Similarly, an oracle testifying to the the second equality is called a packing oracle for E. ## 3 Optimal Hausdorff Oracles For any set $E$, there are infinitely many Hausdorff oracles for $E$. A natural question is whether there is a Hausdorff oracle which minimizes the complexity of every point in $E$. Unfortunately, it is, in general, not possible for a single oracle to maximally reduce every point. We introduce the notion of optimal Hausdorff oracles by weakening the condition to a single point. ###### Definition 3. Let $E\subseteq\mathbb{R}^{n}$ and $A\subseteq\mathbb{N}$. We say that $A$ is Hausdorff optimal for $E$ if the following conditions are satisfied. 1. 1. $A$ is a Hausdorff oracle for $E$. 2. 2. For every $B\subseteq\mathbb{N}$ and every $\epsilon>0$ there is a point $x\in E$ such that $\dim^{A,B}(x)\geq\dim_{H}(E)-\epsilon$ and for almost every $r\in\mathbb{N}$ $K^{A,B}_{r}(x)\geq K^{A}_{r}(x)-\epsilon r$. Note that the second condition only guarantees the existence of one point whose complexity is unaffected by the addtional information in $B$. However, we can show that this implies the seemingly stronger condition that “most” points are unaffected. For $B\subseteq\mathbb{N}$, $\epsilon>0$ define the set $N(A,B,\epsilon)=\\{x\in E\,|\,(\forall^{\infty}r)\,K^{A,B}_{r}(x)\geq K^{A}_{r}(x)-\epsilon r\\}$ ###### Proposition 4. Let $E\subseteq\mathbb{R}^{n}$ be a set such that $\dim_{H}(E)>0$ and let $A$ be an oracle. Then $A$ is a Hausdorff optimal oracle for $E$ if and only if $A$ is a Hausdorff oracle and $\dim_{H}(N(A,B,\epsilon))=\dim_{H}(E)$ for every $B\subseteq\mathbb{N}$ and $\epsilon>0$. ###### Proof. For the forward direction, let $A$ be a optimal Hausdorff oracle for $E$. Then by the first condition of the definition, $A$ is a Hausdorff oracle. Let $B\subseteq\mathbb{N}$ and $\epsilon>0$. Let $C$ be a Hausdorff oracle for $N(A,B,\epsilon)$. For the sake of contradiction, suppose that $\dim_{H}(N(A,B,\epsilon))<\dim_{H}(E)-\gamma$, for some $\gamma>0$. We will, without loss of generality, assume that $\gamma<\epsilon$. Then, by the point to set principle, for every $x\in N(A,B,\epsilon)$, $\displaystyle\dim^{A,(B,C)}(x)$ $\displaystyle\leq\dim^{C}(x)$ $\displaystyle\leq\dim_{H}(N(A,B,\epsilon))$ $\displaystyle<\dim_{H}(E)-\gamma.$ Since, $A$ is an optimal Hausdorff oracle for $E$, there is a point $x\in E$ such that $\dim^{A,(B,C)}(x)\geq\dim_{H}(E)-\gamma$ and for almost every $r\in\mathbb{N}$ $K^{A,(B,C)}_{r}(x)\geq K^{A}_{r}(x)-\gamma r$. By our previous discussion, any such point $x$ cannot be in $N(A,B,\epsilon)$. However, if $x\notin N(A,B,\epsilon)$, then for infinitely many $r$, $K^{A,(B,C)}_{r}(x)<K^{A}_{r}(x)-\epsilon r$. Thus, no such $x$ exists, contradicting the fact that $A$ is Hausdorff optimal. For the backward direction, let $A$ be an oracle satisfying the hypothesis. Then $A$ is a Hausdorff oracle for $E$ and the first condition of optimal Hausdorff oracles is satisfied. Let $B\subseteq\mathbb{N}$ and $\epsilon>0$. By our hypothesis and the point-to-set principle, $\displaystyle\dim_{H}(E)$ $\displaystyle=\dim_{H}(N(A,B,\epsilon))$ $\displaystyle\leq\sup\limits_{x\in N(A,B,\epsilon)}\dim^{A,B}(x).$ Therefore, there is certainly a point $x\in E$ such that $\dim^{A,B}(x)\geq\dim_{H}(E)-\epsilon$ and $K^{A,B}_{r}(x)\geq K^{A}_{r}(x)-\epsilon r$, for almost every $r\in\mathbb{N}$. ∎ A simple, but useful, result is if $B$ is an oracle obtained by adding additional information to an optimal Hausdorff oracle, then $B$ is also optimal. ###### Lemma 5. Let $E\subseteq\mathbb{R}^{n}$. If $A$ is an optimal Hausdorff oracle for $E$, then the join $C=(A,B)$ is Hausdorff optimal for $E$ for every oracle $B$. ###### Proof. Let $A$ be an optimal Hausdorff oracle for $E$. By the point-to-set principle (Theorem 2), $\displaystyle\dim_{H}(E)$ $\displaystyle=\sup\limits_{x\in E}\dim^{A}(x)$ $\displaystyle\geq\sup\limits_{x\in E}\dim^{C}(x)$ $\displaystyle\geq\dim_{H}(E).$ Hence, the oracle $C=(A,B)$ is a Hausdorff oracle for $E$. Let $B^{\prime}\subseteq\mathbb{N}$ be an oracle, and let $\epsilon>0$. Let $x\in E$ be a point such that $\dim^{A,(B,B^{\prime})}(x)\geq\dim_{H}(E)-\epsilon/2,$ (4) and $K_{r}^{A,(B,B^{\prime})}(x)\geq K^{A}_{r}(x)-\epsilon r/2,$ (5) for almost every precision $r$. Note that such a point exists since $A$ is optimal for $E$. For all sufficiently large $r$, $\displaystyle K^{(A,B),B^{\prime}}_{r}(x)$ $\displaystyle=K^{A,(B,B^{\prime})}_{r}(x)$ $\displaystyle\geq K^{A}_{r}(x)-\epsilon r/2$ $\displaystyle\geq K^{A,B}_{r}(x)-\epsilon r/2$ $\displaystyle=K^{C}_{r}(x)-\epsilon r/2.$ Therefore, $C=(A,B)$ is optimal for $E$. ∎ We now give some basic closure properties of the class of sets with optimal Hausdorff oracles. ###### Observation 6. Let $F\subseteq E$. If $\dim_{H}(F)=\dim_{H}(E)$ and $F$ has an optimal Hausdorff oracle, then $E$ has an optimal Hausdorff oracle. We can also show that having optimal Hausdorff oracles is closed under countable unions. ###### Proposition 7. Let $E_{1},E_{2},\ldots$ be a countable sequence of sets and let $E=\cup_{n}E_{n}$. If every set $E_{n}$ has an optimal Hausdorff oracle, then $E$ has an optimal Hausdorff oracle. ###### Proof. We first note that $\dim_{H}(E)=\sup_{n}\dim_{H}(E_{n})$. For each $n$, let $A_{n}$ be an optimal Hausdorff oracle for $E_{n}$. Let $A$ be the join of $A_{1},A_{2},\ldots$. Let $B$ be a Hausdorff oracle for $E$. Note that, by Lemma 5, for every $n$, since $A_{n}$ is an optimal Hausdorff oracle for $E_{n}$, $(A,B)$ is optimal for $E_{n}$. We now claim that $(A,B)$ is an optimal Hausdorff oracle for $E$. Theorem 2 shows that item (1) of the definition of optimal Hausdorff oracles is satisfied. For item (2), let $C\subseteq\mathbb{N}$ be an oracle, and let $\epsilon>0$. Let $n$ be a number such that $\dim_{H}(E_{n})>\dim_{H}(E)-\epsilon$. Since $(A,B)$ is Hausdorff optimal for $E_{N}$, there is a point $x\in E_{n}$ such that 1. (i) $\dim^{(A,B),C}(x)\geq\dim_{H}(E_{n})-\epsilon\geq\dim_{H}(E)-\epsilon$, and 2. (ii) for almost every $r$, $K^{(A,B),C}_{r}(x)\geq K^{(A,B)}_{r}(x)-\epsilon r$. Therefore, item (2) of the definition of optimal Hausdorff oracles is satisfied, and so $(A,B)$ is Hausdorff optimal for $E$. ∎ ### 3.1 Outer Measures and Optimal Oracles In this section we give a sufficient condition for a set to have optimal Hausdorff oracles. Specifically, we prove that if $\dim_{H}(E)=s$, and there is a metric outer measure, absolutely continuous with respect to $\mathcal{H}^{s}$, such that $0<\mu(E)<\infty$, then $E$ has optimal Hausdorff oracles. Although stated in this general form, the main application of this result (in Section 3.2) is for the case $\mu=\mathcal{H}^{s}$. For every $r\in\mathbb{N}$, let $\mathcal{Q}^{n}_{r}$ be the set of all dyadic cubes at precision $r$, i.e., cubes of the form $Q=[m_{1}2^{-r},(m_{1}+1)2^{-r})\times\ldots\times[m_{n}2^{-r},(m_{n}+1)2^{-r})$, where $0\leq m_{1},\ldots,m_{n}\leq 2^{r}$. For each $r$, we refer to the $2^{nr}$ cubes in $\mathcal{Q}_{r}$ as $Q_{r,1},\ldots,Q_{r,2^{nr}}$. We can identify each dyadic cube $Q_{r,i}$ with the unique dyadic rational $d_{r,i}$ at the center of $Q_{r,i}$. We now associate, to each metric outer measure, a discrete semimeasure on the dyadic rationals $\mathbb{D}$. Recall that discrete semimeasure on $\mathbb{D}^{n}$ is a function $p:\mathbb{D}^{n}\to[0,1]$ which satisfies $\Sigma_{r,i}p(d_{r,i})<\infty$. Let $E\subseteq\mathbb{R}^{n}$ and $\mu$ be a metric outer measure such that $0<\mu(E)<\infty$. Define the function $p_{\mu}:\mathbb{D}^{n}\rightarrow[0,1]$ by $p_{\mu,E}(d_{r,i})=\frac{\mu(E\cap Q_{r,i})}{r^{2}\mu(E)}$. ###### Observation 8. Let $\mu$ be a metric outer measure and $E\subseteq\mathbb{R}^{n}$ such that $0<\mu(E)<\infty$. Then for every $r$, every dyadic cube $Q\in\mathcal{Q}_{r}$, and all $r^{\prime}>r$, $\mu(E\cap Q)=\sum\limits_{\begin{subarray}{c}Q^{\prime}\subset Q\\\ Q^{\prime}\in\mathcal{Q}_{r^{\prime}}\end{subarray}}\mu(E\cap Q^{\prime})$. ###### Proposition 9. Let $E\subseteq\mathbb{R}^{n}$ and $\mu$ be a metric outer measure such that $0<\mu(E)<\infty$. Relative to some oracle $A$, the function $p_{\mu,E}$ is a lower semi-computable discrete semimeasure. ###### Proof. We can encode the real numbers $p_{\mu,E}(d)$ into an oracle $A$, relative to which $p_{\mu,E}$ is clearly computable. To see that $p_{\mu,E}$ is indeed a discrete semimeasure, by Observation 8, $\displaystyle\sum\limits_{r,i}p_{\mu,E}(d_{r,i})$ $\displaystyle=\sum\limits_{r}\sum\limits_{i=1}^{2^{2r}}\frac{\mu(E\cap Q_{r,i})}{r^{2}\mu(E)}$ $\displaystyle=\sum\limits_{r}\frac{1}{r^{2}\mu(E)}\sum\limits_{i=1}^{2^{2r}}\mu(E\cap Q_{r,i})$ $\displaystyle=\sum\limits_{r}\frac{\mu(E)}{r^{2}\mu(E)}$ $\displaystyle<\infty.$ ∎ In order to connect the existence of such an outer measure $\mu$ to the existence of optimal oracles, we need to relate the semimeasure $p_{\mu}$ and Kolmogorov complexity. We achieve this using a fundamental result in algorithmic information theory. Levin’s optimal lower semicomputable subprobability measure, relative to an oracle $A$, on the dyadic rationals $\mathbb{D}$ is defined by $\mathbf{m}^{A}(d)=\sum\limits_{\pi\,:\,U^{A}(\pi)=d}2^{-|\pi|}$. ###### Lemma 10. Let $E\subseteq\mathbb{R}^{n}$ and $\mu$ be a metric outer measure such that $0<\mu(E)<\infty$. Let $A$ be an oracle relative to which $p_{\mu,E}$ is lower semi-computable. Then is a constant $\alpha>0$ such that $\mathbf{m}^{A}(d)\geq\alpha p_{\mu,E}(d)$, for every $d\in\mathbb{D}^{n}$. ###### Proof. Case and Lutz [3], generalizing Levin’s coding theorem [9, 10], showed that there is a constant $c$ such that $\mathbf{m}^{A}(d_{r,i})\leq 2^{-K^{A}(d_{r,i})+K^{A}(r)+c}$, for every $r\in\mathbb{N}$ and $d_{r,i}\in\mathbb{D}^{n}$. The optimality of $\mathbf{m}^{A}$ ensures that, for every lower semicomputable (relative to $A$) discrete semimeasure $\nu$ on $\mathbb{D}^{n}$, $\mathbf{m}^{A}(d_{r,i})\geq\alpha\nu(d_{r,i})$. ∎ The results of this section have dealt with the dyadic rationals. However, we ultimately deal with the Kolmogorov complexity of Euclidean points. A result of Case and Lutz [3] relates the Kolmogorov complexity of Euclidean points with the complexity of dyadic rationals. ###### Lemma 11 ([3]). Let $x\in[0,1)^{n}$, $A\subseteq\mathbb{N}$, and $r\in\mathbb{N}$. Let $Q_{r,i}$ be the (unique) dyadic cube at precision $r$ containing $x$. Then $K^{A}_{r}(x)=K^{A}(d_{r,i})-O(\log r)$. ###### Lemma 12. Let $E\subseteq\mathbb{R}^{n}$ and $\mu$ be a metric outer measure such that $0<\mu(E)<\infty$. Let $A$ be an oracle relative to which $p_{\mu,E}$ is lower semi-computable. Then, for every oracle $B\subseteq\mathbb{N}$ and every $\epsilon>0$, the set $N=\\{x\in E\,|\,(\exists^{\infty})\;K^{A,B}_{r}(x)<K^{A}_{r}(x)-\epsilon r\\}$ has $\mu$-measure zero. ###### Proof. Let $B\subseteq\mathbb{N}$ and $\epsilon>0$. For every $R\in\mathbb{N}$, there is a set $\mathcal{C}_{R}$ of dyadic cubes satisfying the following. * • The cubes in $\mathcal{C}_{R}$ cover $N$. * • Every $Q_{r,i}$ in $\mathcal{C}_{R}$ satisfies $r\geq R$. * • For every $Q_{r,i}\in\mathcal{C}_{R}$, $K^{A,B}(d_{r,i})<K^{A}(d_{r,i})-\epsilon r+O(\log r)$. Note that the last item follows from our definition of $N$ by Lemma 11. Since the family of cubes in $\mathcal{C}_{R}$ covers $N$, by the subadditive property of $\mu$, $\sum\limits_{Q_{r,i}\in\mathcal{C}_{R}}\mu(E\cap Q_{r,i})\geq\mu(N)$. Thus, for every $R$, by Lemma 10 and Kraft’s inequality, $\displaystyle 1$ $\displaystyle\geq\sum\limits_{Q_{r,i}\in\mathcal{C}_{R}}2^{-K^{A,B}(d_{r,i})}$ $\displaystyle\geq\sum\limits_{Q_{r,i}\in\mathcal{C}_{R}}2^{\epsilon r-K^{A}(d_{r,i})}$ $\displaystyle\geq\sum\limits_{Q_{r,i}\in\mathcal{C}_{R}}2^{\epsilon r}\mathbf{m}^{A}(d_{r,i})$ $\displaystyle\geq\sum\limits_{Q_{r,i}\in\mathcal{C}_{R}}2^{\epsilon r-K^{A}(r)+c}\alpha p_{\mu,E}(d_{r,i})$ $\displaystyle\geq\sum\limits_{Q_{r,i}\in\mathcal{C}_{R}}2^{\epsilon r/2}p_{\mu,E}(d_{r,i})$ $\displaystyle\geq\sum\limits_{Q_{r,i}\in\mathcal{C}_{R}}2^{\epsilon r/2}\frac{\mu(E\cap Q_{r,i})}{r^{2}\mu(E)}$ $\displaystyle\geq\sum\limits_{Q_{r,i}\in\mathcal{C}_{R}}2^{\epsilon r/4}\frac{\mu(E\cap Q_{r,i})}{r^{2}\mu(E)}$ $\displaystyle\geq 2^{\epsilon R/4}\sum\limits_{Q_{r,i}\in\mathcal{C}_{R}}\frac{\mu(E\cap Q_{r,i})}{\mu(E)}$ $\displaystyle\geq 2^{\epsilon R/4}\frac{\mu(N)}{\mu(E)}.$ Since $R$ can be arbitrarily large, and $0<\mu(E)<\infty$, the conclusion follows. ∎ We now have the machinery in place to prove the main theorem of this section. ###### Theorem 13. Let $E\subseteq\mathbb{R}^{n}$ with $\dim_{H}(E)=s$. Suppose there is a metric outer measure $\mu$ such that $0<\mu(E)<\infty$, and either 1. 1. $\mu\ll\mathcal{H}^{s-\delta}$, for every $\delta>0$, or 2. 2. $\mathcal{H}^{s}\ll\mu$ and $\mathcal{H}^{s}(E)>0$. Then $E$ has an optimal Hausdorff oracle $A$. ###### Proof. Let $A\subseteq\mathbb{N}$ be a Hausdorff oracle for $E$ such that $p_{\mu,E}$ is computable relative to $A$. Note that such an oracle exists by the point- to-set principle and routine encoding. We will show that $A$ is optimal for $E$. For the sake of contradiction, suppose that there is an oracle $B$ and $\epsilon>0$ such that, for every $x\in E$ either 1. 1. $\dim^{A,B}(x)<s-\epsilon$, or 2. 2. there are infinitely many $r$ such that $K^{A,B}_{r}(x)<K^{A}_{r}(x)-\epsilon r$. Let $N$ be the set of all $x$ for which the second item holds. By Lemma 12, $\mu(N)=0$. We also note that, by the point-to-set principle, $\dim_{H}(E-N)\leq s-\epsilon$, and so $\mathcal{H}^{s}(E-N)=0$. To achieve the desired contradiction, we first assume that $\mu\ll\mathcal{H}^{s-\delta}$, for every $\delta>0$. Since $\mu\ll\mathcal{H}^{s-\delta}$, and $\dim_{H}(E-N)<s-\epsilon$, $\mu(E-N)=0$. Since $\mu$ is a metric outer measure, $\displaystyle 0$ $\displaystyle<\mu(E)$ $\displaystyle\leq\mu(N)+\mu(E-N)$ $\displaystyle=0,$ a contradiction. Now suppose that $\mathcal{H}^{s}\ll\mu$ and $\mathcal{H}^{s}(E)>0$. Then, since $\mathcal{H}^{s}$ is an outer measure, $\mathcal{H}^{s}(E)>0$ and $\mathcal{H}^{s}(E-N)=0$ we must have $\mathcal{H}^{s}(N)>0$. However this implies that $\mu(N)>0$, and we again have the desired contradiction. Thus $A$ is an optimal Hausdorff oracle for $E$ and the proof is complete. ∎ Recall that $E\subseteq[0,1)^{n}$ is called an $s$-set if $0<\mathcal{H}^{s}(E)<\infty$. Since $\mathcal{H}^{s}$ is a metric outer measure, and trivially absolutely continuous with respect to itself, we have the following corollary. ###### Corollary 14. Let $E\subseteq[0,1)^{n}$ be an $s$-set. Then there is an optimal Hausdorff oracle for $E$. ### 3.2 Sets with optimal Hausdorff oracles We now show that every analytic set has optimal Hausdorff oracles. ###### Lemma 15. Every analytic set $E$ has optimal Hausdorff oracles. ###### Proof. We begin by assuming that $E$ is compact, and let $s=\dim_{H}(E)$. Then for every $t<s$, $\mathcal{H}^{t}(E)>0$. Thus, by Theorem 1(1), there is a sequence of compact subsets $F_{1},F_{2},\ldots$ of $E$ such that $\dim_{H}(\bigcup_{n}F_{n})=\dim_{H}(E)$, and, for each $n$, $0<\mathcal{H}^{s_{n}}(F_{n})<\infty$, where $s_{n}=s-1/n$. Therefore, by Theorem 13, each set $F_{n}$ has optimal Hausdorff oracles. Hence, by Proposition 7, $E$ has optimal Hausdorff oracles and the conclusion follows. We now show that every $\Sigma^{0}_{2}$ set has optimal Hausdorff oracles. Suppose $E=\cup_{n}F_{n}$ is $\Sigma^{0}_{1}$, where each $F_{n}$ is compact. As we have just seen, each $F_{n}$ has optimal Hausdorff oracles. Therefore, by Proposition 7, $E$ has optimal Hausdorff oracles and the conclusion follows. Finally, let $E$ be analytic. By Theorem 1(2), there is a $\Sigma^{0}_{2}$ subset $F$ of the same Hausdorff dimension as $E$. We have just seen that $F$ must have an optimal Hausdorff oracle. Since $\dim_{H}(F)=\dim_{H}(E)$, by Observation 6 $E$ has optimal Hausdorff oracles, and the proof is complete ∎ Crone, Fishman and Jackson [4] have recently shown that, assuming the Axiom of Determinacy (AD)555Note that AD is inconsistent with the axiom of choice., every subset $E$ has a Borel subset $F$ such that $\dim_{H}(F)=\dim_{H}(E)$. This, combined with Lemma 15, yields the following corollary. ###### Corollary 16. Assuming AD, every set $E\subseteq\mathbb{R}^{n}$ has optimal Hausdorff oracles. ###### Lemma 17. Suppose that $E\subseteq\mathbb{R}^{n}$ satisfies $\dim_{H}(E)=\dim_{P}(E)$. Then $E$ has an optimal Hausdorff oracle. Moreover, the join $(A,B)$ is an optimal Hausdorff oracle, where $A$ and $B$ are the Hausdorff and packing oracles, respectively, of $E$. ###### Proof. Let $A$ be a Hausdorff oracle for $E$ and let $B$ be a packing oracle for $E$. We claim that that the join $(A,B)$ is an optimal Hausdorff oracle for $E$. By the point-to-set principle, and the fact that extra information cannot increase effective dimension, $\displaystyle\dim_{H}(E)$ $\displaystyle=\sup\limits_{x\in E}\dim^{A}(x)$ $\displaystyle\geq\sup\limits_{x\in E}\dim^{A,B}(x)$ $\displaystyle\geq\dim_{H}(E).$ Therefore $\dim_{H}(E)=\sup\limits_{x\in E}\dim^{A,B}(x)$, and the first condition of optimal Hausdorff oracles is satisfied. Let $C\subseteq\mathbb{N}$ be an oracle and $\epsilon>0$. By the point-to-set principle, $\dim_{H}(E)\leq\sup\limits_{x\in E}\dim^{A,B,C}(x)$, so there is an $x\in E$ such that $\dim_{H}(E)-\epsilon/4<\dim^{A,B,C}(x)$. Let $r$ be sufficiently large. Then, by our choice of $B$ and the fact that additional information cannot increase the complexity of a point, $\displaystyle K^{A,B}_{r}(x)$ $\displaystyle\leq K^{B}_{r}(x)$ $\displaystyle\leq\dim_{P}(E)r+\epsilon r/4$ $\displaystyle=\dim_{H}(E)r+\epsilon r/4$ $\displaystyle<\dim^{A,B,C}(x)r+\epsilon r/2$ $\displaystyle\leq K_{r}^{A,B,C}(x)+\epsilon r.$ Since the oracle $C$ and $\epsilon$ were arbitrarily, the proof is complete. ∎ ### 3.3 Sets without optimal Hausdorff oracles In the previous section, we gave general conditions for a set $E$ to have optimal Hausdorff oracles. Indeed, we saw that under the axiom of determinacy, every set has optimal Hausdorff oracles. However, assuming the axiom of choice (AC) and the continuum hypothesis (CH), we are able to construct sets without optimal Hausdorff oracles. ###### Lemma 18. Assume AC and CH. Then, for every $s\in(0,1)$, there is a subset $E\subseteq\mathbb{R}$ with $\dim_{H}(E)=s$ such that $E$ does not have optimal Hausdorff oracles. Let $s\in(0,1)$. We begin by defining two sequences of natural numbers, $\\{a_{n}\\}$ and $\\{b_{n}\\}$. Let $a_{1}=2$, and $b_{1}=\lfloor 2/s\rfloor$. Inductively define $a_{n+1}=b_{n}^{2}$ and $b_{n+1}=\lfloor a_{n+1}/s\rfloor$. Note that $\lim_{n}a_{n}/b_{n}=s$. Using AC and CH, we order the subsets of the natural numbers such that every subset has countably many predecessors. For every countable ordinal $\alpha$, let $f_{\alpha}:\mathbb{N}\to\\{\beta\,|\,\beta<\alpha\\}$ be a function such that each ordinal $\beta$ strictly less than $\alpha$ is mapped to by infinitely many $n$. Note that such a function exists, since the range is countable assuming CH. We will define real numbers $x_{\alpha}$, $y_{\alpha}$ via transfinite induction. Let $x_{1}$ be a real which is random relative to $A_{1}$. Let $y_{1}$ be the real whose binary expansion is given by $\displaystyle y_{1}[r]=\begin{cases}0&\text{ if }a_{n}<r\leq b_{n}\text{ for some }n\in\mathbb{N}\\\ x_{1}[r]&\text{ otherwise}\end{cases}$ For the induction step, suppose we have defined our points up to $\alpha$. Let $x_{\alpha}$ be a real number which is random relative to the join of $\bigcup_{\beta<\alpha}(A_{\beta},x_{\beta})$ and $A_{\alpha}$. This is possible, as we are assuming that this union is countable. Let $y_{\alpha}$ be the point whose binary expansion is given by $\displaystyle y_{\alpha}[r]=\begin{cases}x_{\beta}[r]&\text{ if }a_{n}<r\leq b_{n},\text{ where }f_{\alpha}(n)=\beta\\\ x_{\alpha}[r]&\text{ otherwise}\end{cases}$ Finally, we define our set $E=\\{y_{\alpha}\\}$. We now claim that $\dim_{H}(E)=s$, and that $E$ does not have an optimal Hausdorff oracle. ###### Lemma 19. The Hausdorff dimension of $E$ is $s$. ###### Proof. We first upper bound the dimension. Let $A$ be an oracle encoding $x_{1}$. From our construction, for every element $y\in E$, there are infinitely many intervals $[a_{n},b_{n}]$ such that $y[a_{n},b_{n}]=x_{1}[a_{n},b_{n}]$. Hence, for every $y\in E$, there are infinitely many $n$ such that $\displaystyle K^{A}_{b_{n}}(y)$ $\displaystyle=K^{A}_{a_{n}}(y)+K^{A}_{b_{n},a_{n}}(y)+o(b_{n})$ $\displaystyle\leq K^{A}_{a_{n}}(y)+o(b_{n})$ $\displaystyle\leq a_{n}+o(b_{n}).$ Therefore, by the point to set principle, $\displaystyle\dim_{H}(E)$ $\displaystyle\leq\sup_{y}\dim^{A}(y)$ $\displaystyle=\sup_{y}\liminf_{r}\frac{K^{A}_{r}(y)}{r}$ $\displaystyle\leq\sup_{y}\liminf_{n}\frac{K^{A}_{b_{n}}(y)}{b_{n}}$ $\displaystyle\leq\sup_{y}\liminf_{n}\frac{a_{n}+o(b_{n})}{b_{n}}$ $\displaystyle\leq\sup_{y}\liminf_{n}s$ $\displaystyle=s,$ and the proof that $\dim_{H}(E)\leq s$ is complete. For the lower bound, let $A$ be a Hausdorff oracle for $E$, and let $\alpha$ be the ordinal corresponding to $A$. By our construction of $y_{\alpha}$, for every $n$, $\displaystyle K^{A_{\alpha}}_{a_{n}}(y_{\alpha})$ $\displaystyle\geq K^{A_{\alpha}}_{a_{n}}(x_{\alpha})-b_{n-1}$ $\displaystyle\geq a_{n}-b_{n-1}-o(a_{n})$ $\displaystyle\geq a_{n}-a_{n}^{\frac{1}{2}}-o(a_{n}).$ Hence, for every $n$, and every $a_{n}<r\leq b_{n}$, $\displaystyle K^{A_{\alpha}}_{r}(y_{\alpha})$ $\displaystyle\geq K^{A_{\alpha}}_{a_{n}}(y_{\alpha})$ $\displaystyle\geq a_{n}-a_{n}^{\frac{1}{2}}-o(a_{n}).$ This implies that $\frac{K^{A_{\alpha}}_{r}(y_{\alpha})}{r}\geq s-o(1)$, for every $n$, and every $a_{n}<r\leq b_{n}$. We can also conclude that, for every $n$ and every $b_{n}<r\leq a_{n+1}$, $\displaystyle K^{A_{\alpha}}_{r}(y_{\alpha})$ $\displaystyle=K^{A_{\alpha}}_{b_{n}}(y_{\alpha})+K^{A_{\alpha}}_{r,b_{n}}(y_{\alpha})-o(r)$ $\displaystyle\geq a_{n}-a_{n}^{\frac{1}{2}}+r-b_{n}-o(r).$ This implies that $\displaystyle\frac{K^{A_{\alpha}}_{r}(y_{\alpha})}{r}$ $\displaystyle=1+\frac{a_{n}}{r}-\frac{b_{n}}{r}-o(1)$ $\displaystyle=1-\frac{a_{n}(1/s-1)}{r}-o(1)$ $\displaystyle\geq 1-s(1/s-1)-o(1)$ $\displaystyle=s-o(1).$ for every $n$, and every $a_{n}<r\leq b_{n}$. Together, these inequalities and the point-to-set principle show that $\displaystyle\dim_{H}(E)$ $\displaystyle=\sup_{x}\dim^{A}(x)$ $\displaystyle\geq\dim^{A}(y_{\alpha})$ $\displaystyle=\liminf_{r}\frac{K^{A}(y_{\alpha})}{r}$ $\displaystyle\geq\liminf_{r}s-o(1)$ $\displaystyle=s,$ and the proof is complete. ∎ ###### Lemma 20. $E$ does not have optimal Hausdorff oracles. ###### Proof. Let $A_{\alpha}\subseteq\mathbb{N}$ be an oracle. It suffices to show that $A_{\alpha}$ is not optimal. With this goal in mind, let $B$ be an oracle encoding $x_{\alpha}$ and the set $\\{y_{\beta}\,|\,\beta<\alpha\\}$. Note that we can encode this information since this set is countable. Let $y_{\beta}\in E$. First, suppose that $\beta\leq\alpha$. Then by our choice of $B$, $\dim^{A_{\alpha},B}(y_{\beta})=0$. So then suppose that $\beta>\alpha$. We first note that, since $x_{\beta}$ is random relative to $A_{\alpha}$ $\displaystyle K^{A_{\alpha}}_{a_{n}}(y_{\beta})$ $\displaystyle\geq K^{A_{\alpha}}(y_{\beta}[b_{n-1}\ldots a_{n}])-O(\log a_{n})$ $\displaystyle=K^{A_{\alpha}}(x_{\beta}[b_{n-1}\ldots a_{n}])-O(\log a_{n})$ $\displaystyle\geq a_{n}-b_{n-1}-O(\log a_{n})$ $\displaystyle\geq a_{n}-o(a_{n}),$ for every $n\in\mathbb{N}$. By our construction, there are infinitely many $n$ such that $y_{\beta}[a_{n}\ldots b_{n}]=x_{\alpha}[a_{n}\ldots b_{n}]$ (6) Since $x_{\alpha}$ is random relative to $A_{\alpha}$, for any $n$ such that (6) holds, $\displaystyle K^{A_{\alpha}}_{b_{n}}(y_{\beta})$ $\displaystyle=K^{A_{\alpha}}_{a_{n}}(y_{\beta})+K^{A_{\alpha}}_{b_{n},a_{n}}(y_{\beta})$ $\displaystyle\geq a_{n}-o(a_{n})+K^{A_{\alpha}}(y_{\beta}[a_{n}\ldots b_{n}])-O(\log b_{n})$ $\displaystyle=a_{n}-o(a_{n})+K^{A_{\alpha}}(x_{\alpha}[a_{n}\ldots b_{n}])$ $\displaystyle\geq a_{n}-o(a_{n})+b_{n}-a_{n}-o(b_{n})$ $\displaystyle=b_{n}-o(b_{n}).$ However, since we can compute $x_{\alpha}$ given $B$, $\displaystyle K^{A_{\alpha},B}_{b_{n}}(y_{\beta})$ $\displaystyle=K^{A_{\alpha},B}_{a_{n}}(y_{\beta})+K^{A_{\alpha},B}_{b_{n},a_{n}}(y_{\beta})$ $\displaystyle=K^{A_{\alpha},B}_{a_{n}}(y_{\beta})$ $\displaystyle\leq a_{n}-o(a_{n})$ $\displaystyle=sb_{n}-o(a_{n})$ $\displaystyle=sK^{A_{\alpha}}_{b_{n}}(y_{\beta})-o(a_{n}).$ Therefore $A_{\alpha}$ is not optimal, and the claim follows. ∎ #### 3.3.1 Generalization to $\mathbb{R}^{n}$ In this section, we use Lemma 18 to show that there are sets without optimal Hausdorff oracles in $\mathbb{R}^{n}$ of every possible dimension. We will need the following lemma on giving sufficient conditions for a product set to have optimal Hausdorff oracles. Interestingly, we need the product formula to hold for arbitrary sets, first proven by Lutz [17]. Under the assumption that $F$ is regular, the product formula gives $\dim_{H}(F\times G)=\dim_{H}(F)+\dim_{H}(G)=\dim_{P}(F)+\dim_{H}(G)$, for every set $G$. ###### Lemma 21. Let $F\subseteq\mathbb{R}^{n}$ be a set such that $\dim_{H}(F)=\dim_{P}(F)$, let $G\subseteq\mathbb{R}^{m}$ and let $E=F\times G$. Then $E$ has optimal Hausdorff oracles if and only if $G$ has optimal Hausdorff oracles. ###### Proof. Assume that $G$ has an optimal Hausdorff oracle $A_{1}$. Let $A_{2},A_{3}$ be Hausdorff oracles for $E$ and $F$, respectively. Let $A\subseteq\mathbb{N}$ be the join of all three oracles. We claim that $A$ is optimal for $E$. Let $B$ be any oracle and let $\epsilon>0$. Since $A$ is optimal for $G$, by Lemma 5, there is a point $z\in G$ such that $\dim^{A,B}(z)\geq\dim_{H}(G)-\epsilon/2$ and $K^{A,B}_{r}(z)\geq K^{A}_{r}(z)-\epsilon r/2$, for almost every $r$. By the point-to-set principle, we may choose a $y\in F$ such that $\dim^{A,B,z}(y)\geq\dim_{H}(F)-\epsilon/2$. Let $x=(y,z)\in E$. Then $\displaystyle K^{A,B}_{r}(x)$ $\displaystyle=K^{A,B}_{r}(y,z)$ $\displaystyle=K^{A,B}_{r}(z)+K^{A,B}_{r}(y\,|\,z)$ $\displaystyle\geq K^{A}_{r}(z)-\epsilon r/2+K^{A,B,z}_{r}(y)$ $\displaystyle\geq K^{A}_{r}(z)-\epsilon r/2+(\dim_{H}(F)-\epsilon/2)r$ $\displaystyle=K^{A}_{r}(z)-\epsilon r/2+(\dim_{P}(F)-\epsilon/2)r$ $\displaystyle\geq K^{A}_{r}(z)-\epsilon r/2+K^{A}_{r}(y)-\epsilon r/2$ $\displaystyle\geq K^{A}_{r}(z)-\epsilon r/2+K^{A}_{r}(y\,|\,z)-\epsilon r/2$ $\displaystyle\geq K^{A}_{r}(y,z)-\epsilon r$ $\displaystyle=K^{A}_{r}(x)-\epsilon r.$ Since $B$ and $\epsilon$ were arbitrary, $A$ is optimal for $E$. Suppose that $G$ does not have optimal Hausdorff oracles. Let $A$ be a Hausdorff oracle for $E$. It suffices to show that $A$ is not optimal for $E$. Since optimal Hausdorff oracles are closed under the join operation, we may assume that $A$ is a Hausdorff oracle for $F$ and $G$ as well. Since $G$ does not have optimal Hausdorff oracles, there is an oracle $B$ and $\epsilon>0$ such that, for every $z\in G$, either $\dim^{A,B}(z)<\dim_{H}(G)-\epsilon$ or $K^{A,B}_{r}(z)<K^{A}_{r}(z)-\epsilon r/2$, for infinitely many $r$. Let $x\in E$, such that $\dim^{A,B}(x)\geq\dim_{H}(E)-\epsilon/2$. Let $x=(y,z)$ for some $y\in F$ and $z\in G$. Then we have $\displaystyle\dim_{H}(F)+\dim_{H}(G)$ $\displaystyle=\dim_{H}(E)$ $\displaystyle\leq\dim^{A,B}(x)+\epsilon/2$ $\displaystyle=\dim^{A,B}(y)+\dim^{A,B}(z\,|\,y)+\epsilon/2$ $\displaystyle\leq\dim_{H}(F)+\dim^{A,B}(z)+\epsilon/2.$ Hence, $\dim^{A,B}(z)\geq\dim_{H}(G)-\epsilon/2$. We conclude that there are infinitely many $r$ such that $\displaystyle K^{A,B}_{r}(x)$ $\displaystyle=K^{A,B}_{r}(z)+K^{A,B}_{r}(y\,|\,z)$ $\displaystyle<K^{A}_{r}(z)-\epsilon r/2+K^{A,B}_{r}(y\,|\,z)$ $\displaystyle\leq K^{A}_{r}(z)-\epsilon r/2+K^{A}_{r}(y\,|\,z)$ $\displaystyle=K^{A,B}_{r}(x)-\epsilon r/2.$ Thus $E$ does not have optimal Hausdorff oracles. ∎ ###### Theorem 22. Assume AC and CH. Then for every $n\in\mathbb{N}$ and $s\in(0,n)$, there is a subset $E\subseteq\mathbb{R}^{n}$ with $\dim_{H}(E)=s$ such that $E$ does not have optimal Hausdorff oracles. ###### Proof. We will show this via induction on $n$. For $n=1$, the conclusion follows from Lemma 18. Suppose the claim holds for all $m<n$. Let $s\in(0,n)$. First assume that $s<n-1$. Then by our induction hypothesis, there is a set $G\subseteq\mathbb{R}^{n-1}$ without optimal Hausdorff oracles such that $\dim_{H}(G)=s$. Let $E=\\{0\\}\times G$. Note that, since $\\{0\\}$ is a singleton, $\dim_{H}(\\{0\\})=\dim_{P}(\\{0\\})=0$. Therefore, by Lemma 21, $E$ does not have optimal Hausdorff oracles. By the well-known product formula for Hausdorff dimension, $\displaystyle\dim_{H}(G)$ $\displaystyle\leq\dim_{H}(\\{0\\})+\dim_{H}(G)$ $\displaystyle\leq\dim_{H}(E)$ $\displaystyle\leq\dim_{P}(\\{0\\})+\dim_{H}(G)$ $\displaystyle=\dim_{H}(G),$ and the claim follows. We now assume that $s\geq n-1$. Let $d=s-1$. By our induction hypothesis, there is a set $G\subseteq\mathbb{R}^{n-1}$ without optimal Hausdorff oracles such that $\dim_{H}(G)=d$. Let $E=[0,1]\times G$. Note that, since $[0,1]$ has (Lebesgue) measure one, $\dim_{H}([0,1])=\dim_{P}([0,1])=1$. Thus, by Lemma 21, $E$ is a set without optimal Hausdorff oracles. By the product formula, $\displaystyle 1+\dim_{H}(G)$ $\displaystyle\leq\dim_{H}([0,1])+\dim_{H}(G)$ $\displaystyle\leq\dim_{H}(E)$ $\displaystyle\leq\dim_{P}([0,1])+\dim_{H}(G)$ $\displaystyle=1+\dim_{H}(G),$ and the claim follows. ∎ ## 4 Marstrand’s Projection Theorem The following theorem, due to Lutz and Stull [19], gives sufficient conditions for strong lower bounds on the complexity of projected points. ###### Theorem 23. Let $z\in\mathbb{R}^{2}$, $\theta\in[0,\pi]$, $C\subseteq\mathbb{N}$, $\eta\in\mathbb{Q}\cap(0,1)\cap(0,\dim(z))$, $\varepsilon>0$, and $r\in\mathbb{N}$. Assume the following are satisfied. 1. 1. For every $s\leq r$, $K_{s}(\theta)\geq s-\log(s)$. 2. 2. $K^{C,\theta}_{r}(z)\geq K_{r}(z)-\varepsilon r$. Then, $K^{C,\theta}_{r}(p_{\theta}z)\geq\eta r-\varepsilon r-\frac{4\varepsilon}{1-\eta}r-O(\log r)\,.$ The second condition of this theorem requires the oracle $(C,\theta)$ to give essentially no information about $z$. The existence of optimal Hausdorff oracles gives a sufficient condition for this to be true, for all sufficiently large precisions. Thus we are able to show that Marstrands projection theorem holds for any set with optimal Hausdorff oracles. ###### Theorem 24. Suppose $E\subseteq\mathbb{R}^{2}$ has an optimal Hausdorff oracle. Then for almost every $\theta\in[0,\pi]$, $\dim_{H}(p_{\theta}E)=\min\\{\dim_{H}(E),1\\}$. ###### Proof. Let $A$ be an optimal Hausdorff oracle for $E$. Let $\theta$ be random relative to $A$. Let $B$ be oracle testifying to the point-to-set principle for $p_{\theta}E$. It suffices to show that $\sup\limits_{z\in E}\dim^{A,B}(p_{\theta}z)=\min\\{1,\dim_{H}(E)\\}$. Since $E$ has optimal Hausdorff oracles, for each $n\in\mathbb{N}$, we may choose a point $z_{n}\in E$ such that * • $\dim^{A,B,\theta}(z_{n})\geq\dim_{H}(E)-\frac{1}{2n}$, and * • $K^{A,B,\theta}_{r}(z_{n})\geq K^{A}_{r}(z_{n})-\frac{r}{2n}$ for almost every $r$. Fix a sufficiently large $n$, and let $\varepsilon=1/2n$. Let $\eta\in\mathbb{Q}$ be a rational such that $\min\\{1,\dim_{H}(E)\\}-5\varepsilon^{1/2}<\eta<1-4\varepsilon^{1/2}$. We now show that the conditions of Theorem 23 are satisfied for $\eta,\varepsilon$, relative to $A$. By our choice of $\theta$, $K^{A}_{r}(\theta)\geq r-O(\log r)$, for every $r\in\mathbb{N}$. By our choice of $z_{n}$ and the Hausdorff optimality of $A$, $K^{A,B,\theta}_{r}(z_{n})\geq K_{r}(z_{n})-\varepsilon r$, for all sufficiently large $r$. We may therefore apply Theorem 23, to see that, for all sufficiently large $r$, $K^{A,B,\theta}_{r}(p_{\theta}z_{n})\geq\eta r-\varepsilon r-\frac{4\varepsilon}{1-\eta}r-O(\log r)\,.$ Thus, $\displaystyle\dim^{A,B}(p_{\theta}z_{n})$ $\displaystyle\geq\dim^{A,B,\theta}(p_{\theta}z_{n})$ $\displaystyle=\limsup_{r}\frac{K^{A,B,\theta}_{r}(p_{\theta}z_{n})}{r}$ $\displaystyle\geq\limsup_{r}\frac{\eta r-\varepsilon r-\frac{4\varepsilon}{1-\eta}r-O(\log r)}{r}$ $\displaystyle=\limsup_{r}\eta-\varepsilon-\frac{4\varepsilon}{1-\eta}-o(1)$ $\displaystyle>\eta-\varepsilon-\varepsilon^{1/2}-o(1)$ $\displaystyle>\min\\{1,\dim_{H}(E)\\}-\varepsilon-6\varepsilon^{1/2}-o(1).$ Hence, $\lim_{n}\dim^{A,B}(p_{\theta}z_{n})=\min\\{1,\dim_{H}(E)\\}$, and the proof is complete. ∎ This shows that Marstrand’s theorem holds for every set $E$ with $\dim_{H}(E)=s$ satisfying any of the following: 1. 1. $E$ is analytic. 2. 2. $\dim_{H}(E)=\dim_{P}(E)$. 3. 3. $\mu\ll\mathcal{H}^{s-\delta}$, for every $\delta>0$ for some metric outer measure $\mu$ such that $0<\mu(E)<\infty$. 4. 4. $\mathcal{H}^{s}\ll\mu$ and $\mathcal{H}^{s}(E)>0$, for some metric outer measure $\mu$ such that $0<\mu(E)<\infty$. For example, the existence of exact gauged Hausdorff measures on $E$ guarantee the existence of optimal Hausdorff oracles. ###### Example. Let $E$ be a set with $\dim_{H}(E)=s$ and $\mathcal{H}^{s}(E)=0$. Suppose that $0<\mathcal{H}^{\phi}(E)<\infty$, where $\phi(t)=\frac{t^{s}}{\log\frac{1}{t}}$. Since $\mathcal{H}^{\phi}\ll\mathcal{H}^{s-\delta}$ for every $\delta>0$, Theorem 13 implies that $E$ has optimal Hausdorff oracles, and thus Marstrand’s theorem holds for $E$. ###### Example. Let $E$ be a set with $\dim_{H}(E)=s$ and $\mathcal{H}^{s}(E)=\infty$. Suppose that $0<\mathcal{H}^{\phi}(E)<\infty$, where $\phi(t)=t^{s}\log\frac{1}{t}$. Since $\mathcal{H}^{s}\ll\mathcal{H}^{\phi}$, Theorem 13 implies that $E$ has optimal Hausdorff oracles, and thus Marstrand’s theorem holds for $E$. ### 4.1 Counterexample to Marstrand’s theorem In this section we show that there are sets for which Marstrand’s theorem does not hold. While not explicitly mentioning optimal Hausdorff oracles, the construction is very similar to the construction in Section 3.3. ###### Theorem 25. Assuming AC and CH, for every $s\in(0,1)$ there is a set $E$ such that $\dim_{H}(E)=1+s$ but $\dim_{H}(p_{\theta}E)=s$ for every $\theta\in(\pi/4,3\pi/4)$. This is a modest generalization of Davies’ theorem to sets with Hausdorff dimension strictly greater than one. In the next section we give a new proof of Davies’ theorem by generalizing this construction to the endpoint $s=0$. We will need the following simple observation. ###### Observation 26. Let $r\in\mathbb{N}$, $s\in(0,1)$, and $\theta\in(\pi/8,3\pi/8)$. Then for every dyadic rectangle $R=[d_{x}-2^{-r},d_{x}+2^{-r}]\times[d_{y}-2^{-sr},d_{y}+2^{-sr}]$, there is a point $z\in R$ such that $K^{\theta}_{r}(p_{\theta}z)\leq sr+o(r)$. ###### Proof. Note that $p_{\theta}$ is a Lipschitz function. Thus, for any rectangle $R=[d_{x}-2^{-r},d_{x}+2^{-r}]\times[d_{y}-2^{-sr},d_{y}+2^{-sr}]$, The length of its projection (which is an interval) is $|p_{\theta}R|\geq c2^{-sr}$ for some constant $c$. It is well known that any interval of length $2^{-\ell}$ contains points $x$ such that $K_{r}(x)\leq\ell r+o(r)$. ∎ For every $r\in\mathbb{N}$, $\theta\in(\pi/4,3\pi/4)$, binary string $x$ of length $r$ and string $y$ of length $sr$, let $g_{\theta}(x,y)\mapsto z$ be a function such that $K^{\theta}_{r}(p_{\theta}\,(x,z))\leq sr+o(r)$. That is, $g_{\theta}$, given a rectangle $R=[d_{x}-2^{-r},d_{x}+2^{-r}]\times[d_{y}-2^{-sr},d_{y}+2^{-sr}]$, outputs a value $z$ such that $K_{r}(p_{\theta}(x,z))$ is small. Let $s\in(0,1)$. We begin by defining two sequences of natural numbers, $\\{a_{n}\\}$ and $\\{b_{n}\\}$. Let $a_{1}=2$, and $b_{1}=\lfloor 2/s\rfloor$. Inductively define $a_{n+1}=b_{n}^{2}$ and $b_{n+1}=\lfloor a_{n+1}/s\rfloor$. We will also need, for every ordinal $\alpha$, a function $f_{\alpha}:\mathbb{N}\to\\{\beta\,|\,\beta<\alpha\\}$ such that each ordinal $\beta<\alpha$ is mapped to by infinitely many $n$. Note that such a function exists, since the range is countable assuming CH. Using AC and CH, we first order the subsets of the natural numbers and we order the angles $\theta\in(\pi/4,3\pi/4)$ so that each has at most countably many predecessors. We will define real numbers $x_{\alpha}$, $y_{\alpha}$ and $z_{\alpha}$ inductively. Let $x_{1}$ be a real which is random relative to $A_{1}$. Let $y_{1}$ be a real which is random relative to $(A_{1},x_{1})$. Define $z_{1}$ to be the real whose binary expansion is given by $\displaystyle z_{1}[r]=\begin{cases}g_{\theta_{1}}(x_{1},y_{1})[r]&\text{ if }a_{n}<r\leq b_{n}\text{ for some }n\in\mathbb{N}\\\ y_{1}[r]&\text{ otherwise}\end{cases}$ For the induction step, suppose we have defined our points up to ordinal $\alpha$. Let $x_{\alpha}$ be a real number which is random relative to the join of $\bigcup_{\beta<\alpha}(A_{\beta},x_{\beta})$ and $A_{\alpha}$. Let $y_{\alpha}$ be random relative to the join of $\bigcup_{\beta<\alpha}(A_{\beta},x_{\beta})$, $A_{\alpha}$ and $x_{\alpha}$. This is possible, as we are assuming CH, and so this union is countable. Let $z_{\alpha}$ be the point whose binary expansion is given by $\displaystyle z_{\alpha}[r]=\begin{cases}g_{\theta_{\beta}}(x_{\alpha},y_{\alpha})[r]&\text{ if }a_{n}<r\leq b_{n},\text{ for }f_{\alpha}(n)=\beta\\\ y_{\alpha}[r]&\text{ otherwise}\end{cases}$ Finally, we define our set $E=\\{(x_{\alpha},z_{\alpha})\\}$. ###### Lemma 27. For every $\theta\in(\pi/4,3\pi/4)$, $\dim_{H}(p_{\theta}E)\leq s$ ###### Proof. Let $\theta\in(\pi/4,3\pi/4)$ and $\alpha$ be its corresponding ordinal. Let $A$ be an oracle encoding $\theta$ and $\bigcup\limits_{\beta\leq\alpha}(x_{\beta},y_{\beta},z_{\beta})$. Note that, since we assumed CH, this is a countable union, and so the oracle is well defined. Let $z=(x_{\beta},z_{\beta})\in E$. First assume that $\beta\leq\alpha$. Then, by our construction of $A$, all the information of $p_{\theta}z$ is already encoded in our oracle, and so $K^{A}_{r}(p_{\theta}z)=o(r)$. Now assume that $\beta>\alpha$. Then by our construction of $E$, there are infinitely many $n$ such that $f_{\beta}(n)=\alpha$. Therefore there are infinitely many $n$ such that $z_{\beta}[r]=g_{\theta_{\alpha}}(x_{\beta},y_{\beta})[r]$, for $a_{n}<r\leq b_{n}$. Recalling the definition of $g_{\theta_{\alpha}}$, this means that, for each such $n$, $K^{\theta}_{b_{n}}(p_{\theta}z)=sb_{n}+o(r)$. Therefore, by the point-to-set principle, $\displaystyle\dim_{H}(p_{\theta}E)$ $\displaystyle\leq\sup_{z\in E}\dim^{A}(p_{\theta}z)$ $\displaystyle\leq\sup_{\beta>\alpha}\liminf_{n}\frac{K^{A}_{b_{n}}(p_{\theta}z)}{b_{n}}$ $\displaystyle\leq\sup_{\beta>\alpha}\liminf_{n}\frac{sb_{n}}{b_{n}}$ $\displaystyle=s,$ and the proof is complete. ∎ ###### Lemma 28. The Hausdorff dimension of $E$ is $1+s$. ###### Proof. We first give an upper bound on the dimension. Let $A$ be an oracle encoding $\theta_{1}$. Let $z=(x_{\alpha},z_{\alpha})$. By our construction of $E$, there are infinitely many $n$ such that $f_{\alpha}(n)=1$. Therefore there are infinitely many $n$ such that $z_{\beta}[r]=g_{\theta_{1}}(x_{\beta},y_{\beta})[r]$, for $a_{n}<r\leq b_{n}$. Recalling the definition of $g_{\theta_{1}}$, this means that, for each such $n$, $K^{\theta_{1}}_{b_{n}}(p_{\theta_{1}}z)=sb_{n}+o(r)$. Moreover, $\displaystyle K^{\theta_{1}}_{b_{n}}(x_{\alpha},z_{\alpha})$ $\displaystyle\leq K^{\theta_{1}}_{b_{n}}(x_{\alpha})+K^{\theta_{1}}_{b_{n}}(z_{\alpha}\mid x_{\alpha})+o(r)$ $\displaystyle\leq b_{n}+K^{\theta_{1}}_{b_{n}}(p_{\theta_{1}}z)+o(r)$ $\displaystyle\leq b_{n}+sb_{n}+o(b_{n})).$ Therefore, by the point-to-set principle, $\displaystyle\dim_{H}(E)$ $\displaystyle\leq\sup_{z\in E}\dim^{A}(z)$ $\displaystyle\leq\sup_{z\in E}\liminf_{n}\frac{K^{A}_{b_{n}}(z)}{b_{n}}$ $\displaystyle\leq\sup_{z\in E}\liminf_{n}\frac{(1+s)b_{n}+o(b_{n})}{b_{n}}$ $\displaystyle=1+s.$ For the upper bound, let $A$ be a Hausdorff oracle for $E$, and let $\alpha$ be the ordinal corresponding to $A$. By construction of $z=(x_{\alpha},z_{\alpha})$, $K^{A}_{r}(x_{\alpha})\geq r-o(r)$, for all $r\in\mathbb{N}$. We also have, for every $n$, $\displaystyle K^{A}_{a_{n}}(z_{\alpha}\,|\,x_{\alpha})$ $\displaystyle\geq K^{A}_{a_{n}}(y_{\alpha}\,|\,x_{\alpha})-b_{n-1}-o(a_{n})$ $\displaystyle\geq a_{n}-b_{n-1}-o(a_{n})$ $\displaystyle=a_{n}-a^{\frac{1}{2}}_{n}-o(a_{n}).$ Hence, for every $n$ and every $a_{n}<r\leq b_{n}$, $\displaystyle K^{A}_{r}(z_{\alpha}\,|\,x_{\alpha})$ $\displaystyle\geq K^{A}_{a_{n}}(z_{\alpha}\,|\,x_{\alpha})$ $\displaystyle\geq a_{n}-a^{\frac{1}{2}}_{n}-o(a_{n}).$ This implies that $\displaystyle\frac{K^{A}_{r}(x_{\alpha},z_{\alpha})}{r}$ $\displaystyle=\frac{K^{A}_{r}(x_{\alpha})+K^{A}_{r}(z_{\alpha}\,|\,x_{\alpha})}{r}$ $\displaystyle\geq\frac{r+a_{n}-a^{\frac{1}{2}}_{n}-o(a_{n})}{r}$ $\displaystyle=1+s-o(1).$ We can also conclude that, for every $n$ and every $b_{n}<r\leq a_{n+1}$, $\displaystyle K^{A}_{r}(z_{\alpha}\,|\,x_{\alpha})$ $\displaystyle\geq K^{A}_{b_{n}}(z_{\alpha}\,|\,x_{\alpha})K^{A}_{a_{n},b_{n}}(z_{\alpha}\,|\,x_{\alpha})-o(r)$ $\displaystyle\geq a_{n}-a^{\frac{1}{2}}_{n}+r-b_{n}-o(r).$ This implies that $\displaystyle\frac{K^{A}_{r}(x_{\alpha},z_{\alpha})}{r}$ $\displaystyle=\frac{K^{A}_{r}(x_{\alpha})+K^{A}_{r}(z_{\alpha}\,|\,x_{\alpha})}{r}$ $\displaystyle\geq\frac{r+a_{n}-a^{\frac{1}{2}}_{n}+r-b_{n}-o(r)}{r}$ $\displaystyle\geq 1+s-o(1).$ These inequalities, combined with the point-to-set principle show that $\displaystyle\dim_{H}(E)$ $\displaystyle=\sup_{z\in E}\dim^{A}(z)$ $\displaystyle\geq\sup_{z\in E}\liminf_{r}\frac{K^{A}_{r}(z)}{r}$ $\displaystyle\geq\sup_{z\in E}\liminf_{r}1+s$ $\displaystyle=1+s,$ and the proof is complete. ∎ ### 4.2 Generalization to the endpoint $s=0$ ###### Theorem 29. Assuming AC and CH, there is a set $E$ such that $\dim_{H}(E)=1$ but $\dim_{H}(p_{\theta}E)=0$ for every $\theta\in(\pi/4,3\pi/4)$. For every $r\in\mathbb{N}$, $\theta\in(\pi/4,3\pi/4)$, binary string $x$ of length $r$ and string $y$ of length $sr$, let $g^{s}_{\theta}(x,y)\mapsto z$ be a function such that $K^{\theta}_{r}(p_{\theta}\,(x,z))\leq sr+o(r)$. That is, $g^{s}_{\theta}$, given a rectangle $R=[d_{x}-2^{-r},d_{x}+2^{-r}]\times[d_{y}-2^{-sr},d_{y}+2^{-sr}]$, outputs a value $z$ such that $K_{r}(p_{\theta}(x,z))$ is small. We begin by defining two sequences of natural numbers, $\\{a_{n}\\}$ and $\\{b_{n}\\}$. Let $a_{1}=2$, and $b_{1}=4$. Inductively define $a_{n+1}=b_{n}^{2}$ and $b_{n+1}=(n+1)\lfloor a_{n+1}\rfloor$. We will also need, for every ordinal $\alpha$, a function $f_{\alpha}:\mathbb{N}\to\\{\beta\,|\,\beta<\alpha\\}$ such that each ordinal $\beta<\alpha$ is mapped to by infinitely many $n$. Note that such a function exists, since the range is countable assuming CH. Using AC and CH, we first order the subsets of the natural numbers and we order the angles $\theta\in(\pi/4,3\pi/4)$ so that each has at most countably many predecessors. We will define real numbers $x_{\alpha}$, $y_{\alpha}$ and $z_{\alpha}$ inductively. Let $x_{1}$ be a real which is random relative to $A_{1}$. Let $y_{1}$ be a real which is random relative to $(A_{1},x_{1})$. Define $z_{1}$ to be the real whose binary expansion is given by $\displaystyle z_{1}[r]=\begin{cases}g^{1}_{\theta_{1}}(x_{1},y_{1})[r]&\text{ if }a_{n}<r\leq b_{n}\text{ for some }n\in\mathbb{N}\\\ y_{1}[r]&\text{ otherwise}\end{cases}$ For the induction step, suppose we have defined our points up to ordinal $\alpha$. Let $x_{\alpha}$ be a real number which is random relative to the join of $\bigcup_{\beta<\alpha}(A_{\beta},x_{\beta})$ and $A_{\alpha}$. Let $y_{\alpha}$ be random relative to the join of $\bigcup_{\beta<\alpha}(A_{\beta},x_{\beta})$, $A_{\alpha}$ and $x_{\alpha}$. This is possible, as we are assuming CH, and so this union is countable. Let $z_{\alpha}$ be the point whose binary expansion is given by $\displaystyle z_{\alpha}[r]=\begin{cases}g^{1/n}_{\theta_{\beta}}(x_{\alpha},y_{\alpha})[r]&\text{ if }a_{n}<r\leq b_{n},\text{ for }f_{\alpha}(n)=\beta\\\ y_{\alpha}[r]&\text{ otherwise}\end{cases}$ Finally, we define our set $E=\\{(x_{\alpha},z_{\alpha})\\}$. ###### Lemma 30. For every $\theta\in(\pi/4,3\pi/4)$, $\dim_{H}(p_{\theta}E)=0$. ###### Proof. Let $\theta\in(\pi/4,3\pi/4)$ and $\alpha$ be its corresponding ordinal. Let $A$ be an oracle encoding $\theta$ and $\bigcup\limits_{\beta\leq\alpha}(x_{\beta},y_{\beta},z_{\beta})$. Note that, since we assumed CH, this is a countable union, and so the oracle is well defined. Let $z=(x_{\beta},z_{\beta})\in E$. First assume that $\beta\leq\alpha$. Then, by our construction of $A$, all the information of $p_{\theta}z$ is already encoded in our oracle, and so $K^{A}_{r}(p_{\theta}z)=o(r)$. Now assume that $\beta>\alpha$. Then by our construction of $E$, there are infinitely many $n$ such that $f_{\beta}(n)=\alpha$. Therefore there are infinitely many $n$ such that $z_{\beta}[r]=g^{1/n}_{\theta_{\alpha}}(x_{\beta},y_{\beta})[r]$, for $a_{n}<r\leq b_{n}$. Recalling the definition of $g^{1/n}_{\theta_{\alpha}}$, this means that, for each such $n$, $K^{\theta}_{b_{n}}(p_{\theta}z)=\frac{b_{n}}{n}+o(r)$. Therefore, by the point-to-set principle, $\displaystyle\dim_{H}(p_{\theta}E)$ $\displaystyle\leq\sup_{z\in E}\dim^{A}(p_{\theta}z)$ $\displaystyle\leq\sup_{\beta>\alpha}\liminf_{n}\frac{K^{A}_{b_{n}}(p_{\theta}z)}{b_{n}}$ $\displaystyle\leq\sup_{\beta>\alpha}\liminf_{n}\frac{\frac{b_{n}}{n}}{b_{n}}$ $\displaystyle=\frac{1}{n},$ and the proof is complete. ∎ ###### Lemma 31. The Hausdorff dimension of $E$ is $1$. ###### Proof. We first give an upper bound on the dimension. Let $A$ be an oracle encoding $\theta_{1}$. Let $z=(x_{\alpha},z_{\alpha})$. By our construction of $E$, there are infinitely many $n$ such that $f_{\alpha}(n)=1$. Therefore there are infinitely many $n$ such that $z_{\beta}[r]=g^{1/n}_{\theta_{1}}(x_{\beta},y_{\beta})[r]$, for $a_{n}<r\leq b_{n}$. Recalling the definition of $g^{1/n}_{\theta_{1}}$, this means that, for each such $n$, $K^{\theta_{1}}_{b_{n}}(p_{\theta_{1}}z)=\frac{b_{n}}{n}+o(r)$. Moreover, $\displaystyle K^{\theta_{1}}_{b_{n}}(x_{\alpha},z_{\alpha})$ $\displaystyle\leq K^{\theta_{1}}_{b_{n}}(x_{\alpha})+K^{\theta_{1}}_{b_{n}}(z_{\alpha}\mid x_{\alpha})+o(r)$ $\displaystyle\leq b_{n}+K^{\theta_{1}}_{b_{n}}(p_{\theta_{1}}z)+o(r)$ $\displaystyle\leq b_{n}+\frac{b_{n}}{n}+o(b_{n})).$ Therefore, by the point-to-set principle, $\displaystyle\dim_{H}(E)$ $\displaystyle\leq\sup_{z\in E}\dim^{A}(z)$ $\displaystyle\leq\sup_{z\in E}\liminf_{n}\frac{K^{A}_{b_{n}}(z)}{b_{n}}$ $\displaystyle\leq\sup_{z\in E}\liminf_{n}\frac{(b_{n}+b_{n}/n+o(b_{n})}{b_{n}}$ $\displaystyle=1.$ For the upper bound, let $A$ be a Hausdorff oracle for $E$, and let $\alpha$ be the ordinal corresponding to $A$. By construction of $z=(x_{\alpha},z_{\alpha})$, $K^{A}_{r}(x_{\alpha})\geq r-o(r)$, for all $r\in\mathbb{N}$. We also have, for every $n$, $\displaystyle K^{A}_{a_{n}}(z_{\alpha}\,|\,x_{\alpha})$ $\displaystyle\geq K^{A}_{a_{n}}(y_{\alpha}\,|\,x_{\alpha})-b_{n-1}-o(a_{n})$ $\displaystyle\geq a_{n}-b_{n-1}-o(a_{n})$ $\displaystyle=a_{n}-a^{\frac{1}{2}}_{n}-o(a_{n}).$ Hence, for every $n$ and every $a_{n}<r\leq b_{n}$, $\displaystyle K^{A}_{r}(z_{\alpha}\,|\,x_{\alpha})$ $\displaystyle\geq K^{A}_{a_{n}}(z_{\alpha}\,|\,x_{\alpha})$ $\displaystyle\geq a_{n}-a^{\frac{1}{2}}_{n}-o(a_{n}).$ This implies that $\displaystyle\frac{K^{A}_{r}(x_{\alpha},z_{\alpha})}{r}$ $\displaystyle=\frac{K^{A}_{r}(x_{\alpha})+K^{A}_{r}(z_{\alpha}\,|\,x_{\alpha})}{r}$ $\displaystyle\geq\frac{r+a_{n}-a^{\frac{1}{2}}_{n}-o(a_{n})}{r}$ $\displaystyle=1-o(1).$ We can also conclude that, for every $n$ and every $b_{n}<r\leq a_{n+1}$, $\displaystyle K^{A}_{r}(z_{\alpha}\,|\,x_{\alpha})$ $\displaystyle\geq K^{A}_{b_{n}}(z_{\alpha}\,|\,x_{\alpha})+K^{A}_{a_{n},b_{n}}(z_{\alpha}\,|\,x_{\alpha})-o(r)$ $\displaystyle\geq a_{n}-a^{\frac{1}{2}}_{n}+r-b_{n}-o(r).$ This implies that $\displaystyle\frac{K^{A}_{r}(x_{\alpha},z_{\alpha})}{r}$ $\displaystyle=\frac{K^{A}_{r}(x_{\alpha})+K^{A}_{r}(z_{\alpha}\,|\,x_{\alpha})}{r}$ $\displaystyle\geq\frac{r+a_{n}-a^{\frac{1}{2}}_{n}+r-b_{n}-o(r)}{r}$ $\displaystyle\geq 1-o(1).$ These inequalities, combined with the point-to-set principle show that $\displaystyle\dim_{H}(E)$ $\displaystyle=\sup_{z\in E}\dim^{A}(z)$ $\displaystyle\geq\sup_{z\in E}\liminf_{r}\frac{K^{A}_{r}(z)}{r}$ $\displaystyle\geq\sup_{z\in E}1$ $\displaystyle=1,$ and the proof is complete. ∎ ## 5 Optimal Packing Oracles Similarly, we can define optimal packing oracles for a set. ###### Definition 32. Let $E\subseteq\mathbb{R}^{n}$ and $A\subseteq\mathbb{N}$. We say that $A$ is an optimal packing oracle (or packing optimal) for $E$ if the following conditions are satisfied. 1. 1. $A$ is a packing oracle for $E$. 2. 2. For every $B\subseteq\mathbb{N}$ and every $\epsilon>0$ there is a point $x\in E$ such that $\operatorname{Dim}^{A,B}(x)\geq\dim_{P}(E)-\epsilon$ and for almost every $r\in\mathbb{N}$ $K^{A,B}_{r}(x)\geq K^{A}_{r}(x)-\epsilon r$. Let $E\subseteq\mathbb{R}^{n}$ and $A\subseteq\mathbb{N}$. For $B\subseteq\mathbb{N}$, $\epsilon>0$ define the set $N(A,B,\epsilon)=\\{x\in E\,|\,(\forall^{\infty}r)\,K^{A,B}_{r}(x)\geq K^{A}_{r}(x)-\epsilon r\\}$ ###### Proposition 33. Let $E\subseteq\mathbb{R}^{n}$ be a set such that $\dim_{P}(E)>0$ and let $A$ be an oracle. Then $A$ is packing optimal for $E$ if and only if $A$ is a packing oracle and for every $B\subseteq\mathbb{N}$ and $\epsilon>0$, $\dim_{P}(N(A,B,\epsilon))=\dim_{P}(E)$. ###### Proof. For the forward direction, let $A$ be an optimal packing oracle for $E$. Then by the first condition of the definition, $A$ is a packing oracle. Let $B\subseteq\mathbb{N}$ and $\epsilon>0$. Let $C$ be a packing oracle for $N(A,B,\epsilon)$. For the sake of contradiction, suppose that $\dim_{P}(N(A,B,\epsilon))<\dim_{P}(E)-\gamma$, for some $\gamma>0$. We will, without loss of generality, assume that $\gamma<\epsilon$. Then, by the point to set principle, for every $x\in N(A,B,\epsilon)$, $\displaystyle\operatorname{Dim}^{A,(B,C)}(x)$ $\displaystyle\leq\operatorname{Dim}^{C}(x)$ $\displaystyle\leq\dim_{P}(N(A,B,\epsilon))$ $\displaystyle<\dim_{P}(E)-\gamma.$ Since, $A$ is an optimal packing oracle for $E$, there is a point $x\in E$ such that $\operatorname{Dim}^{A,(B,C)}(x)\geq\dim_{P}(E)-\gamma$ and for almost every $r\in\mathbb{N}$ $K^{A,(B,C)}_{r}(x)\geq K^{A}_{r}(x)-\gamma r$. By our previous discussion, any such point $x$ cannot be in $N(A,B,\epsilon)$. However, if $x\notin N(A,B,\epsilon)$, then for infinitely many $r$, $K^{A,(B,C)}_{r}(x)<K^{A}_{r}(x)-\epsilon r$. Thus, no such $x$ exists, contradicting the fact that $A$ is packing optimal. For the backward direction, let $A$ be an oracle satisfying the hypothesis. Then $A$ is a Hausdorff oracle for $E$ and the first condition of optimal Hausdorff oracles is satisfied. Let $B\subseteq\mathbb{N}$ and $\epsilon>0$. By our hypothesis and the point-to-set principle, $\displaystyle\dim_{H}(E)$ $\displaystyle=\dim_{H}(N(A,B,\epsilon))$ $\displaystyle\leq\sup\limits_{x\in N(A,B,\epsilon)}\dim^{A,B}(x).$ Therefore, there is certainly a point $x\in E$ such that $\dim^{A,B}(x)\geq\dim_{H}(E)-\epsilon$ and $K^{A,B}_{r}(x)\geq K^{A}_{r}(x)-\epsilon r$, for almost every $r\in\mathbb{N}$. ∎ ###### Lemma 34. Let $E\subseteq\mathbb{R}^{n}$. If $A$ is packing optimal for $E$, then the join $C=(A,B)$ is packing optimal for $E$ for every oracle $B$. ###### Proof. Let $A$ be an optimal packing oracle for $E$, let $B$ be an oracle and let $C=(A,B)$. By the point-to-set principle (Theorem 2), $\displaystyle\dim_{P}(E)$ $\displaystyle=\sup\limits_{x\in E}\operatorname{Dim}^{A}(x)$ $\displaystyle\geq\sup\limits_{x\in E}\operatorname{Dim}^{C}(x)$ $\displaystyle\geq\dim_{P}(E).$ Hence, the oracle $C=(A,B)$ is a packing oracle for $E$. Let $B^{\prime}\subseteq\mathbb{N}$ be an oracle, and let $\epsilon>0$. Let $x\in E$ be a point such that $\operatorname{Dim}^{A,(B,B^{\prime})}(x)\geq\dim_{P}(E)-\epsilon/2,$ (7) and $K_{r}^{A,(B,B^{\prime})}(x)\geq K^{A}_{r}(x)-\epsilon r/2,$ (8) for almost every precision $r$. Note that such a point exists since $A$ is packing optimal for $E$. For all sufficiently large $r$, $\displaystyle K^{(A,B),B^{\prime}}_{r}(x)$ $\displaystyle=K^{A,(B,B^{\prime})}_{r}(x)$ $\displaystyle\geq K^{A}_{r}(x)-\epsilon r/2$ $\displaystyle\geq K^{A,B}_{r}(x)-\epsilon r/2$ $\displaystyle=K^{C}_{r}(x)-\epsilon r/2.$ Therefore, $C=(A,B)$ is packing optimal for $E$. ∎ We now give some basic closure properties of the class of sets with optimal packing oracles. ###### Observation 35. Let $F\subseteq E$. If $\dim_{P}(F)=\dim_{P}(E)$ and $F$ has an optimal packing oracle, then $E$ has an optimal packing oracle. We can also show that having optimal packing oracles is closed under countable unions. ###### Lemma 36. Let $E_{1},E_{2},\ldots$ be a countable sequence of sets and let $E=\cup_{n}E_{n}$. If every set $E_{n}$ has an optimal packing oracle, then $E$ has an optimal Hausdorff oracle. ###### Proof. We first note that $\dim_{P}(E)=\sup_{n}\dim_{P}(E_{n})$. For each $n$, let $A_{n}$ be an optimal packing oracle for $E_{n}$. Let $A$ be the join of $A_{1},A_{2},\ldots$. Let $B$ be an oracle guaranteed by Theorem 2 such that $\sup_{x}\operatorname{Dim}^{B}(x)=\sup_{n}\dim_{P}(E_{n})$. Note that, by Lemma 5, for every $n$, $(A,B)$ is packing optimal for $E_{n}$. We now claim that $(A,B)$ is an optimal packing oracle for $E$. Theorem 2 shows that item (1) of the definition of optimal packing oracles is satisfied. For item (2), let $C\subseteq\mathbb{N}$ be an oracle, and let $\epsilon>0$. Let $n$ be a number such that $\dim_{P}(E_{n})>\dim_{P}(E)-\epsilon$. Since $(A,B)$ is packing optimal for $E_{N}$, there is a point $x\in E_{n}$ such that 1. (i) $\dim^{(A,B),C}(x)\geq\dim_{P}(E_{n})-\epsilon\geq\dim_{P}(E)-\epsilon$, and 2. (ii) for almost every $r$, $K^{(A,B),C}_{r}(x)\geq K^{(A,B)}_{r}(x)-\epsilon r$. Therefore, item (2) of the definition of optimal packing oracles is satisfied, and so $(A,B)$ is Hausdorff optimal for $E$. ∎ For every $0\leq\alpha<\beta\leq 1$ define the set $D_{\alpha,\beta}=\\{x\in(0,1)\,|\,\dim(x)=\alpha\text{ and }\operatorname{Dim}(x)=\beta\\}$. ###### Lemma 37. For every $0\leq\alpha<\beta\leq 1$, $D_{\alpha,\beta}$ has optimal Hausdorff and optimal packing oracles and $\displaystyle\dim_{H}(D_{\alpha,\beta})=\alpha$ $\displaystyle\dim_{P}(D_{\alpha,\beta})=\beta.$ ###### Proof. We begin by noting that $D_{\alpha,\beta}$ is Borel. Therefore, by Theorems 13 and 39, $D_{\alpha,\beta}$ has optimal Hausdorff and optimal packing oracles. Thus, it suffices to show prove the dimension equalities. Define the increasing sequence of natural numbers $\\{h_{j}\\}$ inductively as follows. Let $h_{1}=2$, and let $h_{j+1}=2^{h_{j}}$. For every oracle $A$ let $z_{A}$ be a point such that, for every $\delta>0$ and all sufficiently large $r$, $K^{A}_{(1+\delta)r,r}(z_{A}\,|\,z_{A})=\alpha\delta r=K_{(1+\delta)r,r}(z_{A}\,|\,z_{A})$. Let $y_{A}$ be random relative to $A$ and $z_{A}$. Let $x_{A}$ be the point whose binary expansion is given by $\displaystyle x_{A}[r]=\begin{cases}z_{A}[r]&\text{ if }h_{j}<r\leq\frac{1-\beta}{1-\alpha}h_{j+1}\text{ for some }j\in\mathbb{N}\\\ y_{A}[r]&\text{ otherwise}\end{cases}$ Let $A$ be an oracle, and consider the point $x_{A}$. Let $r\in\mathbb{N}$ be sufficiently large and let $j\in\mathbb{N}$ such that $h_{j}<r\leq h_{j+1}$. We first suppose that $r\leq\frac{1-\beta}{1-\alpha}h_{j+1}$. Then $\displaystyle K_{r}(x_{A})$ $\displaystyle\geq K^{A}_{r}(x_{A})$ $\displaystyle=K^{A}_{h_{j}}(x_{A})+K^{A}_{r,h_{j}}(x_{A}\,|\,x_{A})$ $\displaystyle=O(\log r)+K^{A}_{r,h_{j-1}}(z_{A}\,|\,z_{A})$ $\displaystyle=\alpha r+O(\log r)$ $\displaystyle\geq K_{r}(x_{A}).$ Now suppose that $r>\frac{1-\beta}{1-\alpha}h_{j+1}$. Let $t=\frac{1-\beta}{1-\alpha}h_{j+1}$. Then $\displaystyle K_{r}(x_{A})$ $\displaystyle\geq K^{A}_{r}(x_{A})$ $\displaystyle=K^{A}_{t}(x_{A})+K^{A}_{r,t}(x_{A}\,|\,x_{A})+O(\log r)$ $\displaystyle=\alpha t+K^{A}_{r,t}(x_{A}\,|\,x_{A})+O(\log r)$ $\displaystyle=\alpha t+r-t+O(\log r)$ $\displaystyle=r-t(1-\alpha)+O(\log r)$ $\displaystyle=r-(1-\beta)h_{j+1}+O(\log r)$ $\displaystyle\geq K_{r}(x_{A}).$ In particular, $K^{A}_{r}(x_{A})\geq\alpha r$ for every $h_{j}<r\leq h_{j+1}$. Hence for every oracle $A$, $\dim^{A}(x_{A})=\alpha=\dim(x_{A})$. For all sufficiently large $j$, $\displaystyle K_{h_{j}}(x_{A})$ $\displaystyle=K^{A}_{r}(x_{A})$ $\displaystyle=h_{j}-(1-\beta)h_{j}+O(\log r)$ $\displaystyle=\beta h_{j}+O(\log r),$ and so $\operatorname{Dim}^{A}(x_{A})=\beta=\operatorname{Dim}(x_{A})$. Therefore, for every $A$, $x_{A}\in D_{\alpha,\beta}$. Finally, by the above bounds, $\displaystyle\dim_{H}(D_{\alpha,\beta})=\alpha$ $\displaystyle\dim_{P}(D_{\alpha,\beta})=\beta.$ ∎ ### 5.1 Sufficient conditions for optimal packing oracles ###### Lemma 38. Let $E\subseteq\mathbb{R}^{n}$ be a set such that $\dim_{H}(E)=\dim_{P}(E)=s$. Then $E$ has optimal Hausdorff and optimal packing oracles. ###### Proof. Lemma 17 shows that $E$ has optimal Hausdorff oracles. Let $A_{1}$ be an optimal Hausdorff oracle for $E$. Let $A_{2}$ be a packing oracle for $E$. Let $A=(A_{1},A_{2})$. By Lemma 5, $A$ is an optimal Hausdorff oracle for $E$. We now show that $A$ is an optimal packing oracle for $E$. It is clear that $A$ is a packing oracle for $E$. Let $B\subseteq\mathbb{N}$ and $\epsilon>0$. Since $A$ is Hausdorff optimal for $E$, there is a point $x\in E$ such that $\dim^{A,B}(x)\geq s-\epsilon$ and $K^{A,B}_{r}(x)\geq K^{A}_{r}(x)-\epsilon r$ for almost every $r$. Therefore $\displaystyle\operatorname{Dim}^{A,B}(x)$ $\displaystyle\geq\dim^{A,B}(x)$ $\displaystyle\geq s-\epsilon$ $\displaystyle=\dim_{P}(E)-\epsilon.$ Therefore $x$ satisfies the second condition of optimal packing oracles, and the conclusion follows. ∎ ###### Theorem 39. Let $E\subseteq\mathbb{R}^{n}$ with $\dim_{P}(E)=s$. Suppose there is a metric outer measure $\mu$ such that $0<\mu(E)<\infty$, and either 1. 1. $\mu\ll\mathcal{P}^{s}$, or 2. 2. $\mathcal{P}^{s}\ll\mu$ and $\mathcal{P}^{s}(E)>0$. Then $E$ has an optimal packing oracle $A$. ###### Proof. Let $A\subseteq\mathbb{N}$ be a packing oracle for $E$ such that $p_{\mu,E}$ is computable relative to $A$. Note that such an oracle exists by the point- to-set principle and routine encoding. We will show that $A$ is packing optimal for $E$. For the sake of contradiction, suppose that there is an oracle $B$ and $\epsilon>0$ such that, for every $x\in E$ either 1. 1. $\operatorname{Dim}^{A,B}(x)<s-\epsilon$, or 2. 2. there are infinitely many $r$ such that $K^{A,B}_{r}(x)<K^{A}_{r}(x)-\epsilon r$. Let $N$ be the set of all $x$ for which the second item holds. By Lemma 12, $\mu(N)=0$. We also note that, by the point-to-set principle, $\operatorname{Dim}_{H}(E-N)\leq s-\epsilon$, and so $\mathcal{P}^{s}(E-N)=0$. To achieve the desired contradiction, we first assume that $\mu\ll\mathcal{P}^{s}$. In this case, it suffices to show that $\mu(E-N)>0$. Since $\mu\ll\mathcal{P}^{s}$, $\mu(E-N)=0$. Since $\mu$ is a metric outer measure, $\displaystyle 0$ $\displaystyle<\mu(E)$ $\displaystyle\leq\mu(N)+\mu(E-N)$ $\displaystyle=0,$ a contradiction. Now suppose that $\mathcal{P}^{s}\ll\mu$ and $\mathcal{P}^{s}(E)>0$. Then, since $\mathcal{P}^{s}$ is an outer measure, $\mathcal{P}^{s}(E)>0$ and $\mathcal{P}^{s}(E-N)=0$ we must have $\mathcal{P}^{s}(N)>0$. However this implies that $\mu(N)>0$, and we again have the desired contradiction. Thus $A$ is an optimal packing oracle for $E$ and the proof is complete. ∎ We now show that every analytic set has optimal packing oracles. ###### Lemma 40. Every analytic set $E$ has optimal packing oracles. ###### Proof. A set $E\subseteq\mathbb{R}^{n}$ is called an packing $s$-set if $0<\mathcal{P}^{s}(E)<\infty$. Since $\mathcal{P}^{s}$ is a metric outer measure, and trivially absolutely continuous with respect to itself, Theorem 39 shows that if $E$ is a packing $s$-set then there is an optimal packing oracle for $E$. Now assume that $E$ is compact, and let $s=\dim_{H}(E)$. Then for every $t<s$, $\mathcal{H}^{t}(E)>0$. Thus, by Theorem 1, there is a sequence of compact subsets $F_{1},F_{2},\ldots$ of $E$ such that $\dim_{P}(\bigcup_{n}F_{n})=\dim_{P}(E)$, and, for each $n$, $0<\mathcal{P}^{s_{n}}(F_{n})<\infty$, where $s_{n}=s-1/n$. Therefore, each set $F_{n}$ has optimal packing oracles. Hence, by Lemma 36, $E$ has optimal packing oracles and the conclusion follows. We now show that every $\Sigma^{0}_{2}$ set has optimal packing oracles. Suppose $E=\cup_{n}F_{n}$ is $\Sigma^{0}_{1}$, where each $F_{n}$ is compact. As we have just seen, each $F_{n}$ has optimal packing oracles. Therefore, by Lemma 36, $E$ has optimal packing oracles and the conclusion follows. Finally, let $E$ be analytic. By Theorem 1, there is a $\Sigma^{0}_{2}$ subset $F$ of the same packing dimension as $E$. We have just seen that $F$ must have an optimal packing oracle. Since $\dim_{P}(F)=\dim_{P}(E)$, by Observation 35 $E$ has optimal packing oracles, and the proof is complete ∎ ### 5.2 Sets without optimal oracles ###### Theorem 41. Assuming CH and AC, for every $0<s_{1}<s_{2}\leq 1$ there is a set $E\subseteq\mathbb{R}$ which does not have Hausdorff optimal nor packing optimal oracles such that $\dim_{H}(E)=s_{1}$ and $\dim_{P}(E)=s_{2}$. ###### Proof. Let $\delta=s_{2}-s_{1}$. We begin by defining two sequences of natural numbers, $\\{a_{n}\\}$ and $\\{b_{n}\\}$. Let $a_{1}=2$, and $b_{1}=4$. Inductively define $a_{n+1}=2^{b_{n}}$ and $b_{n+1}=2^{a_{n+1}}$. Using AC and CH, we order the subsets of the natural numbers such that every subset has countably many predecessors. For every countable ordinal $\alpha$, let $f_{\alpha}:\mathbb{N}\to\\{\beta\,|\,\beta<\alpha\\}$ be a function such that each ordinal $\beta$ strictly less than $\alpha$ is mapped to by infinitely many $n$. Note that such a function exists, since the range is countable assuming CH. We will define real numbers $w_{\alpha}$, $x_{\alpha}$, $y_{\alpha}$ and $z_{\alpha}$ via transfinite induction. Let $x_{1}$ be a real such that, for every $\gamma>0$ and all sufficiently large $r$, $K^{A_{1}}_{(1+\gamma)r,r}(w_{1}\,|\,w_{1})=s_{1}\gamma r=K_{(1+\gamma)r,r}(w_{1}\,|\,w_{1})$. Let $x_{1}$ be random relative to $A_{1}$ and $w_{1}$. Let $y_{1}$ be a real such that, for every $\gamma>0$ and all sufficiently large $r$, $K^{A_{1}}_{(1+\gamma)r,r}(y_{1}\,|\,y_{1})=(s_{1}+\frac{\delta}{2})\gamma r=K_{(1+\gamma)r,r}(y_{1}\,|\,y_{1})$. Let $z_{1}$ be the real whose binary expansion is given by $\displaystyle z_{1}[r]=\begin{cases}w_{1}[r]&\text{ if }a_{n}<r\leq\frac{1-s_{2}}{1-s_{1}}b_{n}\text{ for some }n\in\mathbb{N}\\\ x_{1}[r]&\text{ if }\frac{1-s_{2}}{1-s_{1}}b_{n}<r\leq b_{n}\text{ for some }n\in\mathbb{N}\\\ y_{1}[r]&\text{ if }b_{n}<r\leq(1-\delta)a_{n+1}<\text{ for some }n\in\mathbb{N}\\\ 0&\text{ if }(1-\delta)a_{n+1}<r\leq(1+\delta)a_{n+1}<\text{ for some }n\in\mathbb{N}\\\ \end{cases}$ For the induction step, suppose we have defined our points up to $\alpha$. Let $A$ be the join of $\bigcup_{\beta<\alpha}(A_{\beta},w_{\beta}.x_{\beta},y_{\beta},z_{\beta})$ and $A_{\alpha}$. Let $x_{\alpha}$ be a real such that, for every $\gamma>0$ and all sufficiently large $r$, $K^{A}_{(1+\gamma)r,r}(w_{\alpha}\,|\,w_{\alpha})=s_{1}\gamma r=K_{(1+\gamma)r,r}(w_{\alpha}\,|\,w_{\alpha})$. Let $x_{\alpha}$ be random relative to $A$ and $w_{\alpha}$. Let $y_{\alpha}$ be a real such that, for every $\gamma>0$ and all sufficiently large $r$, $K^{A,w_{\alpha},x_{\alpha}}_{(1+\gamma)r,r}(y_{\alpha}\,|\,y_{\alpha})=(s_{1}+\frac{\delta}{2})\gamma r=K_{(1+\gamma)r,r}(y_{\alpha}\,|\,y_{\alpha})$. Let $z_{\alpha}$ be the real whose binary expansion is given by $\displaystyle z_{\alpha}[r]=\begin{cases}w_{\alpha}[r]&\text{ if }a_{n}<r\leq\frac{1-s_{2}}{1-s_{1}}b_{n}\text{ for some }n\in\mathbb{N}\\\ x_{\alpha}[r]&\text{ if }\frac{1-s_{2}}{1-s_{1}}b_{n}<r\leq b_{n}\text{ for some }n\in\mathbb{N}\\\ y_{\alpha}[r]&\text{ if }b_{n}<r\leq(1-\delta)a_{n+1}<\text{ for some }n\in\mathbb{N}\\\ x_{\beta}&\text{ if }(1-s_{1}\delta/2)a_{n+1}<r\leq a_{n+1}<\text{ where }f(\beta)=n\\\ \end{cases}$ Finally, we define our set $E=\\{z_{\alpha}\\}$. We begin by collecting relevant aspects of our construction. Let $\alpha$ be an ordinal, let $A=A_{\alpha}$ be the corresponding oracle in the order, and let $z=z_{\alpha}$ be the point constructed at ordinal $\alpha$. Let$n$ be sufficiently large. Let $a_{n}<r\leq\frac{1-s_{2}}{1-s_{1}}b_{n}$, $\displaystyle K^{A}_{r}(z)$ $\displaystyle=K^{A}_{a_{n}}(z)+K^{A}_{r,a_{n}}(w_{\alpha}\,|\,z)$ $\displaystyle=K^{A}_{a_{n}}(z)+(r-a_{n})s_{1}+O(\log r).$ (9) Let $t=\frac{1-s_{2}}{1-s_{1}}b_{n}<r\leq b_{n}$, $\displaystyle K^{A}_{r}(z)$ $\displaystyle=K^{A}_{t}(z)+K^{A}_{r,t}(x_{\alpha}\,|\,z)$ $\displaystyle=K^{A}_{t}(z)+(r-t)+O(\log r)$ $\displaystyle=(t-a_{n})s_{1}+(r-t)+O(\log r)$ $\displaystyle=ts_{1}+r-t+O(\log r)$ $\displaystyle=r-(1-s_{1})t+O(\log r)$ $\displaystyle=r-(1-s_{2})b_{n}+O(\log r).$ (10) Let $b_{n}<r\leq(1-s_{1}\delta/2)a_{n+1}$. Then, $\displaystyle K^{A}_{r}(z)$ $\displaystyle=K^{A}_{b_{n}}(z)+K^{A}_{r,b_{n}}(y_{\alpha}\,|\,z)$ $\displaystyle=b_{n}-(1-s_{2})b_{n}+K^{A}_{r,b_{n}}(y_{\alpha}\,|\,z)$ $\displaystyle=s_{2}b_{n}+K^{A}_{r,b_{n}}(y_{\alpha}\,|\,z)$ $\displaystyle=s_{2}b_{n}+(s_{1}+\frac{\delta}{2})(r-b_{n}).$ (11) Finally, let $t=(1-s_{1}\delta/2)a_{n+1}<r\leq a_{n+1}$ and let $\beta<\alpha$ be the ordinal such that $f(\beta)=n$. Then, $\displaystyle K^{A}_{r}(z)$ $\displaystyle=K^{A}_{t}(z)+K^{A}_{r,t}(x_{\beta}\,|\,z)$ $\displaystyle=s_{2}b_{n}+(s_{1}+\frac{\delta}{2})(t-b_{n})+K^{A}_{r,t}(x_{\beta}\,|\,z)$ $\displaystyle=(s_{1}+\frac{\delta}{2})t+K^{A}_{r,t}(x_{\beta}\,|\,z)$ (12) In particular, $\displaystyle s_{1}r$ $\displaystyle\leq K^{A}_{r}(z)$ $\displaystyle=(s_{1}+\frac{\delta}{2})t+K^{A}_{r,t}(x_{\beta}\,|\,z)$ $\displaystyle\leq(s_{1}+\frac{\delta}{2})r+r-t$ $\displaystyle\leq(s_{1}+\frac{\delta}{2})r+\frac{\delta r}{2}$ $\displaystyle=s_{2}r$ (13) Let $a_{n}<r\leq a_{n+1}$. The above inequalities show that, if $r>\frac{1-s_{2}}{1-s_{1}}b_{n}$, then $K^{A}_{r}(z)\geq s_{1}r$. When $r\leq\frac{1-s_{2}}{1-s_{1}}b_{n}$, by combining equality (9) and inequality (13), $\displaystyle K^{A}_{r}(z)$ $\displaystyle=K^{A}_{a_{n}}(z)+(r-a_{n})s_{1}+O(\log r)$ $\displaystyle\geq s_{1}a_{n}+(r-a_{n})s_{1}+O(\log r)$ $\displaystyle=s_{1}r+O(\log r).$ Therefore, $K^{A}_{r}(z)\geq s_{1}r$. for all $r$. For the lower bound, let $r=\frac{1-s_{2}}{1-s_{1}}b_{n}$. Then, $\displaystyle K^{A}_{r}(z)$ $\displaystyle=K^{A}_{a_{n}}(z)+(r-a_{n})s_{1}+O(\log r)$ $\displaystyle\leq a_{n}++(r-a_{n})s_{1}+O(\log r)$ $\displaystyle\leq s_{1}r+O(\log r),$ and so $\dim^{A}(z)=s_{1}$. Similarly, the above inequalities show that $K^{A}_{r}(z)\leq s_{2}r$. To prove the lower bound, let $r=b_{n}$. Then $\displaystyle K^{A}_{r}(z)$ $\displaystyle=r-(1-s_{2})b_{n}+O(\log r)$ $\displaystyle=s_{2}r+O(\log r),$ and so $\operatorname{Dim}^{A}(z)=s_{2}$. To complete the proof, we must show that $E$ does not have an optimal Hausdorff oracle, nor an optimal packing oracle. Let $A=A_{\alpha}$ be any Hausdorff oracle for $E$. Let $B$ be an oracle encoding the set $\\{w_{\beta},x_{\beta},y_{\beta}\,|\,\beta\leq\alpha\\}$. Note that we can encode this information since this set is countable. Let $z_{\beta}\in E$. First, suppose that $\beta\leq\alpha$. Then by our choice of $B$, $\dim^{A_{\alpha},B}(z_{\beta})=0$. So then suppose that $\beta>\alpha$. Let $n$ be a sufficiently large natural such that $f(\alpha)=n$. Then, since $x_{\alpha}$ is random relative to $A_{\alpha}$ $\displaystyle K^{A_{\alpha}}_{a_{n+1}}(z_{\beta})$ $\displaystyle=(s_{1}+\frac{\delta}{2})t+K^{A}_{a_{n+1},t}(x_{\alpha}\,|\,z)$ $\displaystyle\geq(s_{1}+\frac{\delta}{2})t+a_{n+1}-t,$ where $t=(1-s_{1}\delta/2)a_{n+1}$. However, by our construction on $B$, $\displaystyle K^{A_{\alpha},B}_{a_{n+1}}(z_{\beta})$ $\displaystyle=(s_{1}+\frac{\delta}{2})t+K^{A,B}_{r,t}(x_{\alpha}\,|\,z)$ $\displaystyle\geq(s_{1}+\frac{\delta}{2})t+O(1).$ Therefore, $\displaystyle K^{A_{\alpha}}_{a_{n+1}}(z_{\beta})-K^{A_{\alpha},B}_{a_{n+1}}(z_{\beta})$ $\displaystyle=a_{n+1}-t$ $\displaystyle=\frac{s_{1}\delta a_{n+1}}{2}.$ Since $z_{\beta}$ was arbitrary, it follows that $B$ reduces the complexity of every point $z\in E$ infinitely often. Since $A_{\alpha}$ was arbitrary, we conclude that $E$ does not have optimal Hausdorff nor optimal packing oracles. ∎ ###### Corollary 42. Assuming CH and AC, for every $0<s_{1}<s_{2}\leq 1$ there is a set $E\subseteq\mathbb{R}$ which has optimal Hausdorff oracles but does not have optimal packing oracles such that $\dim_{H}(E)=s_{1}$ and $\dim_{P}(E)=s_{2}$. ###### Proof. Let $F$ be a set such that $\dim_{H}(F)=\dim_{P}(F)=s_{1}$. Then, by Lemma 17, $F$ has optimal Hausdorff oracles. Let $G$ be a set, guaranteed by Theorem 41, with $\dim_{H}(G)<s_{1}$, $\dim_{P}(G)=s_{2}$ such that $G$ does not have optimal Hausdorff nor optimal packing oracles. Let $E=F\cup G$. Then $\dim_{H}(E)=s_{1}$ and $\dim_{P}(E)=s_{2}$ by the union formula for Hausdorff and packing dimension. By Observation 6, $E$ has optimal Hausdorff oracles. We now prove that $E$ does not have optimal packing oracles. Let $A$ be a packing oracle for $E$. By possibly joining $A$ with a packing oracle for $G$, we may assume that $A$ is a packing oracle for $G$ as well. Since $G$ does not have optimal packing oracles, there is an oracle $B\subseteq\mathbb{N}$ and $\epsilon>s_{2}-s_{1}$ such that, for every $x\in G$ where $\operatorname{Dim}^{A,B}(x)\geq s_{2}-\epsilon$, $K^{A,B}_{r}(x)<K^{A}_{r}(x)-\epsilon r$ for infinitely many $r$. Let $x\in E$ such that $\operatorname{Dim}^{A,B}(x)\geq s_{2}-\epsilon$. Then, by our choice of $F$, $x$ must be in $G$. Therefore $K^{A,B}_{r}(x)<K^{A}_{r}(x)-\epsilon r$ for infinitely many $r$, and so $A$ is not an optimal packing oracle for $E$. Since $A$ was arbitrary, the conclusion follows. ∎ ###### Theorem 43. Assuming CH and AC, for every $0<s_{1}<s_{2}\leq 1$ there is a set $E\subseteq\mathbb{R}$ which has optimal packing oracles but does not have optimal Hausdorff oracles such that $\dim_{H}(E)=s_{1}$ and $\dim_{P}(E)=s_{2}$. ###### Proof. Let $F=\\{x\in(0,1)\,|\,\dim(x)=0\text{ and }\operatorname{Dim}(x)=s_{2}\\}$. By Lemma 37, $\dim_{H}(F)=0$, $\dim_{P}(F)=s_{2}$ and $F$ has optimal packing oracles. Let $G$ be a set, guaranteed by Theorem 41, with $\dim_{H}(G)=s_{1}$, $\dim_{P}(G)=s_{2}$ such that $G$ does not have optimal Hausdorff nor optimal packing oracles. Let $E=F\cup G$. Then $\dim_{H}(E)=s_{1}$ and $\dim_{P}(E)=s_{2}$ by the union formula for Hausdorff and packing dimension. By Observation 35, $E$ has optimal packing oracles. We now prove that $E$ does not have optimal Hausdorff oracles. Let $A$ be a Hausdorff oracle for $E$. By possibly joining $A$ with a Hausdorff oracle for $G$, we may assume that $A$ is a Hausdorff oracle for $G$ as well. Since $G$ does not have optimal Hausdorff oracles, there is an oracle $B\subseteq\mathbb{N}$ and $\epsilon>s_{1}$ such that, for every $x\in G$ where $\dim^{A,B}(x)\geq s_{1}-\epsilon$, $K^{A,B}_{r}(x)<K^{A}_{r}(x)-\epsilon r$ for infinitely many $r$. Let $x\in E$ such that $\dim^{A,B}(x)\geq s_{1}-\epsilon$. Then, since $\dim_{H}(F)=0$, $x$ must be in $G$. Therefore $K^{A,B}_{r}(x)<K^{A}_{r}(x)-\epsilon r$ for infinitely many $r$, and so $A$ is not an optimal Hausdorff oracle for $E$. Since $A$ was arbitrary, the conclusion follows. ∎ ## 6 Acknowledgments I would like to thank Denis Hirschfeldt, Jack Lutz and Chris Porter for very valuable discussions and suggestions. I would also like to thank the participants of the recent AIM workshop on Algorithmic Randomness. ## References * [1] Krishna B. Athreya, John M. Hitchcock, Jack H. Lutz, and Elvira Mayordomo. Effective strong dimension in algorithmic information and computational complexity. SIAM J. Comput., 37(3):671–705, 2007. * [2] Christopher J. Bishop and Yuval Peres. Fractals in probability and analysis, volume 162 of Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge, 2017. * [3] Adam Case and Jack H. Lutz. Mutual dimension. ACM Transactions on Computation Theory, 7(3):12, 2015. * [4] Logan Crone, Lior Fishman, and Stephen Jackson. Hausdorff dimension regularity properties and games. arXiv preprint arXiv:2003.11578, 2020. * [5] Roy O. Davies. Two counterexamples concerning Hausdorff dimensions of projections. Colloq. Math., 42:53–58, 1979. * [6] Rod Downey and Denis Hirschfeldt. Algorithmic Randomness and Complexity. Springer-Verlag, 2010. * [7] Kenneth Falconer. Fractal Geometry: Mathematical Foundations and Applications. Wiley, third edition, 2014. * [8] Kenneth Falconer, Jonathan Fraser, and Xiong Jin. Sixty years of fractal projections. In Fractal geometry and stochastics V, pages 3–25. Springer, 2015\. * [9] Leonid A. Levin. On the notion of a random sequence. Soviet Math Dokl., 14(5):1413–1416, 1973. * [10] Leonid Anatolevich Levin. Laws of information conservation (nongrowth) and aspects of the foundation of probability theory. Problemy Peredachi Informatsii, 10(3):30–35, 1974. * [11] Ming Li and Paul M.B. Vitányi. An Introduction to Kolmogorov Complexity and Its Applications. Springer, third edition, 2008. * [12] Jack H. Lutz. Dimension in complexity classes. SIAM J. Comput., 32(5):1236–1259, 2003. * [13] Jack H. Lutz. The dimensions of individual strings and sequences. Inf. Comput., 187(1):49–79, 2003. * [14] Jack H. Lutz and Neil Lutz. Algorithmic information, plane Kakeya sets, and conditional dimension. ACM Trans. Comput. Theory, 10(2):Art. 7, 22, 2018. * [15] Jack H Lutz and Neil Lutz. Who asked us? how the theory of computing answers questions about analysis. In Complexity and Approximation, pages 48–56. Springer, 2020. * [16] Jack H. Lutz and Elvira Mayordomo. Dimensions of points in self-similar fractals. SIAM J. Comput., 38(3):1080–1112, 2008. * [17] Neil Lutz. Fractal intersections and products via algorithmic dimension. In 42nd International Symposium on Mathematical Foundations of Computer Science (MFCS 2017), 2017. * [18] Neil Lutz and D. M. Stull. Bounding the dimension of points on a line. In Theory and applications of models of computation, volume 10185 of Lecture Notes in Comput. Sci., pages 425–439. Springer, Cham, 2017\. * [19] Neil Lutz and D. M. Stull. Projection theorems using effective dimension. In 43rd International Symposium on Mathematical Foundations of Computer Science (MFCS 2018), 2018. * [20] J. M. Marstrand. Some fundamental geometrical properties of plane sets of fractional dimensions. Proc. London Math. Soc. (3), 4:257–302, 1954. * [21] Pertti Mattila. Hausdorff dimension, orthogonal projections and intersections with planes. Ann. Acad. Sci. Fenn. Ser. AI Math, 1(2):227–244, 1975. * [22] Pertti Mattila. Geometry of sets and measures in Euclidean spaces: fractals and rectifiability. Cambridge University Press, 1999. * [23] Pertti Mattila. Hausdorff dimension, projections, and the fourier transform. Publicacions matematiques, pages 3–48, 2004. * [24] Pertti Mattila. Hausdorff dimension, projections, intersections, and besicovitch sets. In New Trends in Applied Harmonic Analysis, Volume 2, pages 129–157. Springer, 2019. * [25] Elvira Mayordomo. A Kolmogorov complexity characterization of constructive Hausdorff dimension. Inf. Process. Lett., 84(1):1–3, 2002. * [26] Elvira Mayordomo. Effective fractal dimension in algorithmic information theory. In S. Barry Cooper, Benedikt Löwe, and Andrea Sorbi, editors, New Computational Paradigms: Changing Conceptions of What is Computable, pages 259–285. Springer New York, 2008. * [27] Andre Nies. Computability and Randomness. Oxford University Press, Inc., New York, NY, USA, 2009. * [28] Tuomas Orponen. Combinatorial proofs of two theorems of Lutz and Stull. arXiv preprint arXiv:2002.01743, 2020. * [29] D. M. Stull. Results on the dimension spectra of planar lines. In 43rd International Symposium on Mathematical Foundations of Computer Science, volume 117 of LIPIcs. Leibniz Int. Proc. Inform., pages Art. No. 79, 15. Schloss Dagstuhl. Leibniz-Zent. Inform., Wadern, 2018.
††thanks: NHFP Einstein fellow # Constraining gravitational wave amplitude birefringence and Chern-Simons gravity with GWTC-2 Maria Okounkova 0000-0001-7869-5496<EMAIL_ADDRESS>Center for Computational Astrophysics, Flatiron Institute, 162 5th Ave, New York, NY 10010, United States Will M. Farr 0000-0003-1540-8562 <EMAIL_ADDRESS>Department of Physics and Astronomy, Stony Brook University, Stony Brook NY 11794, United States Center for Computational Astrophysics, Flatiron Institute, 162 5th Ave, New York, NY 10010, United States Maximiliano Isi 0000-0001-8830-8672<EMAIL_ADDRESS>Center for Computational Astrophysics, Flatiron Institute, 162 5th Ave, New York, NY 10010, United States LIGO Laboratory and Kavli Institute for Astrophysics and Space Research, Massachusetts Institute of Technology, Cambridge, Massachusetts 02139, USA Leo C. Stein 0000-0001-7559-9597 <EMAIL_ADDRESS>Department of Physics and Astronomy, The University of Mississippi, University, MS 38677, United States ###### Abstract We perform a new test of general relativity (GR) with signals from GWTC-2, the LIGO and Virgo catalog of gravitational wave detections. We search for the presence of amplitude birefringence, in which left versus right circularly polarized modes of gravitational waves are exponentially enhanced and suppressed during propagation. Such an effect is present in various beyond-GR theories but is absent in GR. We constrain the amount of amplitude birefringence consistent with the data through an opacity parameter $\kappa$, which we bound to be $\kappa\lesssim 0.74\textrm{ Gpc}^{-1}$. We then use these theory-agnostic results to constrain Chern-Simons gravity, a beyond-GR theory with motivations in quantum gravity. We bound the canonical Chern- Simons lengthscale to be $\ell_{0}\lesssim 1.0\times 10^{3}$ km, improving on previous long-distance measurement results by a factor of two. ## I Introduction At some length scale, Einstein’s theory of general relativity (GR) must break down and be reconciled with quantum mechanics in a beyond-GR theory of gravity. Gravitational waves (GWs) from binary black hole (BBH) mergers, such as those recently detected by LIGO Aasi _et al._ (2015) and Virgo Acernese _et al._ (2015) could contain signatures of beyond-GR effects, which has motivated significant efforts to test GR with LIGO and Virgo detections Abbott _et al._ (2019a, b); Isi _et al._ (2019a); Nair _et al._ (2019); Isi _et al._ (2019b); Abbott _et al._ (2020a, b). In this study, we perform a new test of GR with the second LIGO-Virgo catalog, GWTC-2 Abbott _et al._ (2020a, 2019c); LIGO Scientific Collaboration and Virgo Collaboration (2019).111Note that GWTC-2 contains GWTC-1 Abbott _et al._ (2019a), the first LIGO and Virgo catalog, as a subset. In several beyond-GR theories, GWs, exhibit amplitude birefringence: when propagating from the source to the detector, the amplitudes of left versus right polarized modes are exponentially enhanced or suppressed. We use the confident BBH detections in GWTC-2 to constrain this effect, which is absent in GR. We characterize the strength of the amplitude birefringence in terms of an opacity parameter, that can then be mapped on to various beyond-GR theories. A thorough review of theories that exhibit amplitude birefringence is provided in Zhao _et al._ (2020a), including Chern-Simons gravity Alexander and Yunes (2009), ghost-free scalar-tensor theories Crisostomi _et al._ (2018), symmetric teleparallel equivalents of GR Conroy and Koivisto (2019), and Hor̆ava-Lifshitz gravity Horava (2009). As a specific application, we use our limit on amplitude birefringence to constrain non-dynamical Chern-Simons gravity, a parity-violating beyond-GR effective field theory with origins in string theory, loop quantum gravity, and inflation Alexander and Yunes (2009); Green and Schwarz (1984); Taveras and Yunes (2008); Mercuri and Taveras (2009); Weinberg (2008). Previous works have addressed the possibility of detecting Chern-Simons amplitude birefringence with GW detectors Nojiri _et al._ (2019); Zhao _et al._ (2020b); Alexander _et al._ (2008); Yunes _et al._ (2010); Yunes and Finn (2009); Yagi and Yang (2018), and in this study we perform such a measurement on real GW data. In Sec. II, we give an overview of the observational effects of amplitude birefringence on GW detections. We then use GWTC-2 to bound the amount of amplitude birefrigence in BBH signals. In Sec. III, we consider these results in the context of Chern-Simons gravity, and bound the canonical Chern-Simons lengthscale. We conclude in Sec. IV. We set $G=c=1$ throughout. $H_{0}$ refers to the present day value of the Hubble parameter, with dimensions of $[H_{0}]=L^{-1}$, and $z$ refers to cosmological redshift. ## II Constraints on amplitude birefringence ### II.1 Observational effects In GR, for the dominant $(2,\pm 2)$ angular mode of non-precessing compact binary inspirals, the ratio of the gravitational wave strain $h$, in right $h_{\mathrm{R}}$, versus left $h_{\mathrm{L}}$, circularly polarized modes is purely a function of the inclination angle of the binary, of the form $\displaystyle\left(\frac{h_{\mathrm{R}}}{h_{\mathrm{L}}}\right)_{\mathrm{GR}}=\left(\frac{1+\cos\iota}{1-\cos\iota}\right)^{2}\,.$ (1) Here, the inclination $\iota$ is the angle from the total angular momentum of the binary to the line of sight of the observer. In terms of the plus, $h_{+}$, and cross, $h_{\times}$, polarizations, the circular polarizations are given by $h_{\mathrm{R},\mathrm{L}}=h_{+}\pm ih_{\times}$. A system with $\cos\iota=1$ has power purely in $h_{\mathrm{R}}$, and is face-on, while one with $\cos\iota=-1$ has power purely in $h_{\mathrm{L}}$ and is face-off. Thus $\displaystyle\textrm{pure }h_{\mathrm{R}}$ $\displaystyle\Longleftrightarrow\cos\iota=+1\Longleftrightarrow\textrm{face- on}\,,$ (2) $\displaystyle\textrm{pure }h_{\mathrm{L}}$ $\displaystyle\Longleftrightarrow\cos\iota=-1\Longleftrightarrow\textrm{face- off}\,.$ (3) We assume that the universe is homogeneous and isotropic at cosmological scales, and that gravitational physics does not have any preferred direction. This implies that the underlying distribution for $\cos\iota$ is flat, meaning no preference for face-on versus face-off events. The picture in Eq. (1) changes in beyond-GR theories that exhibit amplitude birefringence. In this case, the amplitudes of left- versus right-polarized modes are exponentially enhanced and suppressed during propagation, leading to an expression of the form $\displaystyle\left(\frac{h_{\mathrm{R_{obs}}}}{h_{\mathrm{L_{obs}}}}\right)_{\mathrm{Biref}}=\frac{e^{-d_{C}\kappa}(1+\cos\iota)^{2}}{e^{d_{C}\kappa}(1-\cos\iota)^{2}}\,.$ (4) Here, $d_{C}$ is the comoving distance to the source (with units of length $L^{1}$), and $\kappa$ is an opacity parameter with units of $L^{-1}$ that governs the strength of the birefringence.222Note that Eq. (4) is only correct for every theory at linear order in $\kappa d_{C}$; it is exactly correct at all orders for some theories and field profiles. We thus write $\kappa$ as function of $d_{C}$, so $\kappa\left(d_{C}\right)=\kappa_{0}+\mathcal{O}\left(d_{C}\right)$, and specialize to constant $\kappa$ for the remainder of the paper. Note that $\kappa=0$ is consistent with GR. Throughout this study, we will assume that $\kappa d_{C}\ll 1$, that is, beyond-GR effects are small enough that the effective field theory is valid. In traditional GW parameter estimation, however, we do not have access to the true value, $\cos\iota$, of the inclination angle, but rather observe some effective value, $\cos\iota_{\mathrm{obs}}$. Thus, from Eq. (4), in the presence of amplitude birefringence, we would measure a ratio $\displaystyle\frac{1+\cos\iota_{\mathrm{obs}}}{1-\cos\iota_{\mathrm{obs}}}=\frac{e^{-d_{C}\kappa/2}(1+\cos\iota)}{e^{d_{C}\kappa/2}(1-\cos\iota)}\,.$ (5) Let us think about how amplitude birefringence would affect the values $\cos\iota_{\mathrm{obs}}$ for multiple events. Statistical isotropy of BBH orientation requires that $p(\cos\iota)$, the distribution on the true inclination angle over the population of BBH mergers, be flat. The _observed_ distribution of inclinations is influenced by selection effects, but to a very good approximation these are independent of the _sign_ of $\cos\iota$ (Abbott _et al._ , 2020a, c). Thus, if there are no beyond-GR effects and $\kappa=0$, we expect to see an equal number of face-on and face-off events. Meanwhile, if $\kappa>0$, we will preferentially measure $\cos\iota_{\mathrm{obs}}\sim-1$ for isotropically distributed events. In other words, we will preferentially see more face-off, rather than face-on mergers. Similarly, if $\kappa<0$, we will preferentially see more face-on mergers. Thus, we expect $p(\cos\iota_{\mathrm{obs}})$ to not be symmetric about zero. We can quantify the number of observed face-on versus face-off events by defining the on/off (or right/left) asymmetry statistic $\Delta$, in the range $-1\leq\Delta\leq+1$, computed as $\displaystyle\Delta\equiv\frac{N(\cos\iota_{\mathrm{obs}}>0)-N(\cos\iota_{\mathrm{obs}}<0)}{N}\,,$ (6) where $N(\cos\iota_{\mathrm{obs}}>0)$ is the number of face-on observations, etc. From an underlying distribution on $\cos\iota$ and the birefringence result in Eq. (5), we induce a distribution on $\cos\iota_{\mathrm{obs}}$ and thus $\Delta$ (birefringent theories that lead to a different relationship between $\iota$ and $\iota_{\mathrm{obs}}$ will give a different induced distribution). Solving for $\cos\iota_{\mathrm{obs}}$ we get $\displaystyle\cos\iota_{\mathrm{obs}}=\frac{-(1-\cos\iota)e^{\kappa d_{C}/2}+(1+\cos\iota)e^{-\kappa d_{C}/2}}{(1-\cos\iota)e^{\kappa d_{C}/2}+(1+\cos\iota)e^{-\kappa d_{C}/2}}\,.$ (7) Working solely with a quantity such as $\Delta$ provides a robust framework for many beyond-GR theories, and does not require making assumptions about the underlying theory, as is done when producing template waveforms. Note that we do not consider beyond-GR modifications to the gravitational waveform itself as generated from the source, assuming such modifications to be small since they are not amplified with distance (unlike birefringence, which is a propagation effect). The effect of birefringence on the observed inclination depends on the product of the opacity parameter and the comoving distance to each event; a full analysis would take account of the varying distances to the events in GWTC-2, which we consider in Appendix A. To obtain an approximate constraint averaging over BBH detections, however, it is sufficient to approximate a common comoving distance, $d_{C}$, for all events. If we take the flat distribution $\cos\iota\sim\mathcal{U}(-1,1)$,333It is appropriate to match the value of $\Delta$ inferred from the data to the effect of $\kappa$ on the astrophysical population rather than the _selected_ population (events that pass some detection threshold) for the following reason. Selection effects are, to a very good approximation, independent of the _sign_ of $\cos\iota_{\mathrm{obs}}$ (Abbott _et al._ , 2019a, 2020a); due to this symmetry, the same fraction of the population of mergers will be detectable for _any_ value of $\Delta$ in our simplified model where the distribution of $\cos\iota_{\mathrm{obs}}$ is piecewise-flat. The usual factor correcting for selection effects, conventionally written $\alpha\left(\Delta\right)$ (Mandel _et al._ , 2019), appearing in the denominator of the likelihood is therefore constant. Our analysis, ignoring the constant $\alpha$ factor, infers the true _population_ value of $\Delta$; and it is therefore appropriate to match inferred $\Delta$ values to the actual effect on the population from $\kappa$ rather than the _selected_ population. and assuming that $\kappa d_{C}$ is the same for all observations, we get the simple expected value $\displaystyle\hat{\Delta}\equiv\langle\Delta\rangle=\tanh\frac{\kappa d_{C}}{2}\,,$ (8) which can be inverted to estimate $\kappa$ from $\hat{\Delta}$, $\displaystyle\hat{\kappa}=\frac{1}{d_{C}}\log\left[\frac{1+\hat{\Delta}}{1-\hat{\Delta}}\right]\,.$ (9) For our constraints on $\kappa$ and our projections, we use a common comoving distance to our BBH mergers of $d_{C}=d_{C}\left(z=0.3\right)\simeq 1.23\,\mathrm{Gpc}$, corresponding to the median detected redshift in GWTC-2. We will additionally consider an analysis with $d_{C}=d_{C}(z=0.3\pm 0.1)$ in order to provide some error region for our results. Birefringence also changes the signal _amplitude_ measured at the detector, and therefore the inferred luminosity distance to the source, via $\frac{d_{L,\mathrm{obs}}}{d_{L}}=\\\ \frac{\sqrt{1+\cos^{2}\iota}}{\sqrt{\left(1+\cos^{2}\iota_{\mathrm{obs}}\right)\cosh 2\kappa d_{C}+2\cos\iota_{\mathrm{obs}}\sinh 2\kappa d_{C}}}\\\ =1+\frac{\cos\iota_{\mathrm{obs}}\left(\cos^{2}\iota_{\mathrm{obs}}-5\right)}{2\left(1+\cos^{2}\iota_{\mathrm{obs}}\right)}\kappa d_{C}+\mathcal{O}\left(\kappa d_{C}\right)^{2}\,,$ (10) where we have used $d_{L}^{-1}\propto\sqrt{h_{+}^{2}+h_{\times}^{2}}\sim\sqrt{(1+\cos^{2}\iota)^{2}-4\cos^{2}\iota}$. The effect here is to modify the observed distance or redshift distribution of sources from the true distribution. Since the effect enters at linear order in $\kappa d_{C}$, it is degenerate with a variation in the BBH merger rate with redshift; this is in contrast to effects which modify the leading-order relation between the merger rate and distance or redshift, such as extra spacetime dimensions (Fishbach _et al._ , 2018; Pardo _et al._ , 2018). The latter are, in principle, observable even in a nearby sample of BBH mergers, with $z\to 0$. In this study, we use the values for $d_{C}$ reported in GWTC-2, without considering these higher-order corrections. Nevertheless, a full analysis could fit an evolving merger rate and birefringence effects on inclination and amplitude, incorporating selection effects. Given the existing uncertainty about the evolution of the merger rate with redshift (Fishbach _et al._ , 2018; Abbott _et al._ , 2019d) and the difficulty in measuring $\cos\iota_{\mathrm{obs}}$ with existing data (typical uncertanties are $\sim 0.3$ (Abbott _et al._ , 2019a)), our approximate analysis captures the majority of the information about birefringence in the data at this time. Note that in this study we assume that amplitude birefringence is the only phenomenon that modifies the observed inclination angle from its true value. In particular, we do expect strong gravitational lensing to affect fewer than $10^{-3}$ of the detected events Dai _et al._ (2020); Smith _et al._ (2018), and hence do not consider strong lensing effects in this study. ### II.2 GWTC-2 constraint Figure 1: Likelihood distributions on $\cos\iota_{\mathrm{obs}}$, the observed inclination angle from GWTC-2 Abbott _et al._ (2019a); LIGO Scientific Collaboration and Virgo Collaboration (2019); Abbott _et al._ (2020a). Each solid curve (including the gray curves) corresponds to a BBH detection, and the dashed black curve corresponds to the mean of $\cos\iota_{\mathrm{obs}}$ across these events. While most events do not provide a confident measurement of $\cos\iota_{\mathrm{obs}}$, we have highlighted (in thick, colored lines) the events that do show a strong preference for being face-off or face-on. Note that a population consistent with GR will have a mean distribution for $\cos\iota$ symmetric about zero. In Fig. 1, we show the posterior distributions on the observed inclination angle, $\cos\iota_{\mathrm{obs}}$, from GWTC-2 Abbott _et al._ (2019a); LIGO Scientific Collaboration and Virgo Collaboration (2019); Abbott _et al._ (2020a).444When available, we use the NRSur7dq4 parameter estimation results. Otherwise, if available, we use the SEOBNRv4PHM results, and finally we otherwise use the SEOBNRv4P results. We estimate that any systematic difference between which waveform model we use is well below the uncertainty in $\cos\iota$. The first two Advanced LIGO and Virgo observing runs, O1 and O2, contain 10 significant BBH detections, three of which have an inclination constraint sufficient to confidently identify the handedness of the wave, with each preferring a left-handed polarization (i.e. come from a binary orbiting in a left-handed sense with respect to the line-of-sight). The O3a observing run, meanwhile, contains approximately 37 candidate BBH detections, four of which provide a sufficient inclination constraint, with one left-handed polarization event, and three right-handed polarization events. While this results in a total of seven _confident_ inclination angle measurements, we will consider all of the $\cos\iota_{\mathrm{obs}}$ distributions in our analysis, incorporating even weak preferences for left or right handed orbits from each one into our analysis. Note that in the presence of strong amplitude birefringence, we would expect to observe such events with only one inclination angle preference. Thus, GWTC-2 rules out the possibility of purely right or left-handed gravitational events. Due to their relative proximities and the correspondingly weak expected opacity constraints, we simplify our analysis by excluding the binary neutron star events. Thus, we exclude GW170817 and GW190425, as well ad the neutron star - black hole candidate GW190426_152155. Note that we do include GW190814, which provides a strong inclination constraint, but does have an (uncategorized) component mass of $2.59M_{\odot}$ Abbott _et al._ (2020a). Using these measures of $\cos\iota_{\mathrm{obs}}$, we then compute a distribution on $\Delta$ from these observations using Eq. (6) and applying a flat prior on $-1<\Delta<1$, which we show in Fig. 2. We see that the distribution on $\Delta$ from the O1-O2 observing runs disfavors face-on events, while preferring face-off events, and that the distribution on $\Delta$ from O3a disfavors face-off events, while preferring face-on events. Together, all of the detections are consistent with $\Delta=0\pm 0.4$ consistent with no amplitude birefringence. Figure 2: Posterior distribution for $\Delta$, which measures preference for face-on versus face-off observed events, as defined in Eq. (6). Without amplitude birefringence, the distribution should be symmetric around $\Delta=0$. We show the distribution for $\Delta$ from O1-O2 events (light blue curve), and for O3a events (pink curve). We see that O1-O2 have a preference for face-off events, while O3a has a preference for face-on events. The resulting distribution is consistent with $\Delta=0$, with a standard deviation of $0.4$, supporting no amplitude birefringence. The red dashed line, meanwhile corresponds to the values of $\Delta$ obtained by drawing from a distribution uniform in $\cos\iota_{\mathrm{obs}}$ (thus corresponding to no information). Given $\Delta$, we can now use Eq. (9) to obtain a distribution on the absolute values of the opacity parameter $\kappa$, defined in Eq. (4). This will provide a physical measure of the amount of amplitude birefringence, the magnitude of which can then be used to constrain various beyond-GR theories. We show the resulting distribution on $\kappa$ in Fig. 3. We observe that for a common comoving distance of $d_{C}=d_{C}(z=0.3)$ (median detected redshift in GWTC-2), we can bound, at $1\sigma$: O1-O2: $\displaystyle\kappa\lesssim 2.0\textrm{ Gpc}^{-1}\,,$ O3a: $\displaystyle\kappa\lesssim 1.3\textrm{ Gpc}^{-1}\,,$ All: $\displaystyle\kappa\lesssim 0.74\textrm{ Gpc}^{-1}\,.$ In Fig. 3, we also show results for $\kappa$ for common comoving distances of $d_{C}=d_{C}(z=0.3\pm 0.1)$ for all of the detections, in order to qualitatively show the effect of a spread in the distance measurements on the inferred value of $\kappa$. These differences of $z\pm 0.1$ shift the inferred value for all of the detections by $\pm 0.25\textrm{ Gpc}^{-1}$. Recall that for the effective field theory to be valid, we require that $\kappa d_{C}\ll 1$. The analysis presented in this paper in terms of the observed inclination angle works for any value of $\kappa$, but we must be careful about the distances $d_{C}$. Thus, in Fig. 3 we shade the region for which $\kappa d_{C}>1$, where this condition is violated given our choice of $d_{C}=d_{C}(z=0.3)$. In order to see how much information we have gained from these detections, let us consider a distribution flat in $\cos\iota_{\mathrm{obs}}$ (meaning that all measured inclination angles are equally likely and $\cos\iota_{\mathrm{obs}}$ carries no information about the system). The posterior on $\Delta$ for 47 events from this distribution using Eq. (6) should be uniform on $\Delta$ (the events carry no information about which handedness is preferred). For such uninformative measurements if we wish to recover the correct flat distribution for $\Delta$ from our computations, we must satisfy the criterion that the number of samples used for each event is much larger than the number of detections as detailed in Appendix B. If we then compute $\kappa$ from these values of $\Delta$ in Fig. 3, we obtain a distribution that looks like that of O1-O2. We can thus conclude that the measurements of $\cos\iota_{\mathrm{obs}}$ in O1-O2 are not sufficient to provide an informative constraint on $\kappa$; almost all of our constraint on $\kappa$ comes from the assumed prior on $\Delta$ transformed through Eq. (9) into a prior on $\kappa$. However, adding in O3a does make the result deviate from the prior, thus showing that we can indeed constrain the level of amplitude birefringence with all of the BBH detections. In order to quantify this information, we can compute the Kullback–Leibler (KL) divergence $D_{\mathrm{KL}}(p(\lambda)\|q(\lambda))$ of a distribution $p$ with respect to $q$, $\displaystyle D_{\mathrm{KL}}(p(\lambda)\,\|\,q(\lambda))=\int p(\lambda)\log_{2}\left[\frac{p(\lambda)}{q(\lambda)}\right]d\lambda\,.$ (11) While technical details can be found in Kullback and Leibler (1951), the KL divergence is a distance measure of how a probability distribution is different from a reference probability distribution, thus allowing us to compare the curves in Fig. 3. Using the flat distribution as our reference distribution, we find $\displaystyle D_{\mathrm{KL}}(P_{\textrm{O1-O2}}(\kappa)\,\|\,P_{\textrm{Flat}}(\kappa))$ $\displaystyle=2.7\times 10^{-3}\,,$ $\displaystyle D_{\mathrm{KL}}(P_{\textrm{O3a}}(\kappa)\,\|\,P_{\textrm{Flat}}(\kappa))$ $\displaystyle=6.5\times 10^{-2}\,,$ $\displaystyle D_{\mathrm{KL}}(P_{\textrm{All}}(\kappa)\,\|\,P_{\textrm{Flat}}(\kappa))$ $\displaystyle=3.8\times 10^{-1}\,,$ in units of bits. Figure 3: Distribution for the norm of the opacity parameter $\kappa$, as given in Eq. (4), which measures the strength of the amplitude birefringence effect. In the absence of amplitude birefringence, we expect $\kappa=0$. Here, we compute the posterior on $|\kappa|$ with O1-O2 (light blue curve), and O3a (pink curve), combining all of the detections in the black curve. In order to show the effect of our assumption of a common comoving distance of $d_{C}(z=0.3)$ for all events, we also plot lines (light gray), for $d_{C}(z=0.3\pm 0.1)$. The shaded region corresponds to $\kappa=1/d_{C}(z=0.3)$, in which the effective field theory assumption that $\kappa d_{C}\ll 1$ does not hold. The O1-O2 result on its own is uninformative, as it qualitatively agrees with a constraint generated from a flat, uninformative distribution in $\cos\iota_{\mathrm{obs}}$ (dashed thick line). Adding in the O3a results, however, does result in an informative constraint. ## III Constraints on Chern-Simons gravity We now use the inferred opacity $\kappa$ from Sec. II to place constraints on Chern-Simons gravity (CS). CS modifies the Einstein-Hilbert action of GR through the inclusion of a scalar field coupled to a term quadratic in spacetime curvature. In CS, amplitudes of left versus right circularly- polarized modes are exponentially enhanced and suppressed during propagation, with the strength of this amplitude birefringence being governed by properties of the CS scalar field Alexander and Yunes (2009). Thus, by placing constraints on the opacity parameter with GWTC-2, we can place observational constraints on CS. Following the conventions of Alexander and Yunes (2009), the action of Chern- Simons gravity takes the form $\displaystyle S=\int d^{4}x\sqrt{-g}\Big{(}\frac{R}{16\pi G}+\frac{1}{4}\alpha\vartheta{}^{*}\\!RR-\beta\frac{1}{2}\nabla_{a}\vartheta\nabla^{a}\vartheta\Big{)}\,,$ (12) where $g_{ab}$ is the spacetime metric with covariant derivative $\nabla_{a}$. The first term corresponds to the Einstein-Hilbert action of GR, where $R$ is the spacetime Ricci scalar. The second term couples the CS scalar field $\vartheta$ to spacetime curvature via the Pontryagin density ${}^{*}\\!RR\equiv\,{}^{*}\\!R_{abcd}R^{abcd}$, which is the spacetime Riemann tensor contracted with its dual Alexander and Yunes (2009). The last term is a kinetic term for the scalar field, with constant $\beta$. We follow the choice of Jackiw and Pi (2003); Alexander _et al._ (2008), and set $\alpha=\kappa_{\mathrm{E}}$, which gives $\vartheta$ units of length squared, $[\vartheta]=L^{2}$. In non-dynamical CS gravity, we set $\beta=0$, and $\vartheta$ is ‘frozen-in’ with some pre-defined profile Jackiw and Pi (2003), which we will leave unspecified for now. Note that $\vartheta$ cannot be constant, otherwise the ${}^{*}\\!RR$ term, a topological invariant, would integrate out of the action in Eq. (12). As calculated by Alexander et al. Alexander _et al._ (2008), in CS, GWs propagating through a Friedmann-Robertson-Walker universe are exponentially suppressed and enhanced depending on helicity. For compact-binary sources, this birefringence effect manifests in a change in the observed inclination of the binary, $\cos\iota_{\mathrm{obs}}$, from the true inclination angle of the source, $\cos\iota$, as $\displaystyle\left(\frac{h_{\mathrm{R_{obs}}}}{h_{\mathrm{L_{obs}}}}\right)_{\mathrm{CS}}$ $\displaystyle=\left(\frac{1+\cos\iota}{1-\cos\iota}\right)^{2}\exp\left[\frac{2k(t)}{H_{0}}\zeta(\vartheta)\right]$ $\displaystyle=\left(\frac{1+\cos\iota_{\mathrm{obs}}}{1-\cos\iota_{\mathrm{obs}}}\right)^{2}\,.$ (13) Here, $k(t)$ is the wavenumber for the given Fourier propagating mode, with units of $L^{-1}$, and $\zeta(\vartheta)$ is a dimensionless function of the integrated history of the CS scalar field. While Eq. (III) is a function of the wavenumber, we will estimate that $k(t)$ covers a narrow frequency range, and thus write $k(t)\sim k$, where $k$ is a typical value in this range, without treating each mode separately. ### III.1 General constraint Comparing Eq. (III) with Eq. (4), we can directly relate $\zeta(\vartheta)$, which captures all of the dependence on the CS field, to the measured value of $\kappa$ as $\displaystyle\zeta(\vartheta)=\frac{\kappa d_{C}H_{0}}{k}\,.$ (14) Thus, setting $d_{C}\left(z=0.3\right)\simeq\sim 1.23$ Gpc for a typical Advanced LIGO BBH source distance (corresponding to the median detected redshift in GWTC-2) Abbott _et al._ (2019a), and setting $k\sim 2\pi\times 100\,\mathrm{Hz}/c\sim 2\times 10^{-6}\,\mathrm{m}$ for the approximate value of the region of greatest sensitivity of LIGO (cf. Abbott _et al._ (2016, 2019a)), we obtain the dimensionless result $\displaystyle\zeta(\vartheta)=\left(\frac{\kappa}{1\textrm{ Gpc}^{-1}}\right)\times 6.6\times 10^{-21}\,.$ (15) From the results for GWTC-2 in Sec. II, we compute O1-O2: $\displaystyle\zeta(\vartheta)\lesssim 1.3\times 10^{-20}\,,$ O3a: $\displaystyle\zeta(\vartheta)\lesssim 8.6\times 10^{-21}\,,$ All: $\displaystyle\zeta(\vartheta)\lesssim 4.9\times 10^{-21}\,.$ As stated before, $\zeta(\vartheta)$ is dependent on the integrated history of the CS scalar field. In Alexander _et al._ (2008), the authors calculate $\zeta(\vartheta)$ for a matter-dominated universe (with scale factor $a(\eta)=a_{0}\eta^{2}$, where $a_{0}$ is the present-day value and $\eta$ is conformal time). Since the LIGO sources are found at redshifts $z<1$ (300–3000 Mpc) Abbott _et al._ (2019a), we focus on a dark-energy dominated universe, with $a(t)=a_{0}e^{H_{0}t}$. We compute the corresponding $\zeta$, in terms of dimensionless conformal time $\eta$, to be $\displaystyle\zeta(\vartheta)=\frac{H_{0}^{2}}{2}\int_{\eta}^{1}\left(\eta^{2}\vartheta^{\prime\prime}(\eta)-2\eta\vartheta^{\prime}(\eta)\right)d\eta\,.$ (16) We give the full calculation in Appendix C. In the above expressions, we have left the ‘frozen-in’ profile of $\vartheta$ unspecified. Let us suppose that $\vartheta$ is dependent on some CS parameter $P$. For some specified profile $\vartheta[P]$, the reader can thus use Eqs. (15) and (16) to compute a value of $P$ given a value of $\kappa$. ### III.2 Constraint on canonical $\vartheta$ profile Let us now consider the ‘canonical’ profile for $\vartheta$ given in Jackiw and Pi (2003); Alexander and Yunes (2009); Yunes and Spergel (2009), where $\vartheta$ has an isotropic, time-dependent profile of the form $\displaystyle\vartheta=\frac{t}{\mu}\,,$ (17) where $\mu$ is a mass scale with units $[\mu]=L^{-1}$. Note that when $\mu$ is large, we recover GR. Let us define $\displaystyle\ell_{0}\equiv\frac{1}{\mu}$ (18) to be the CS lengthscale for this field profile. With this profile, $\zeta(\vartheta)$ in Eq. (16) becomes $\displaystyle\zeta(\vartheta)=\frac{3H_{0}\ell_{0}}{2c}(1-\eta)=\frac{3H_{0}\ell_{0}d_{C}}{2d_{H}}\,.$ (19) where we have re-introduced a factor of c and have set $(1-\eta)\sim d_{C}/d_{H}$, where $d_{H}\equiv c/H_{0}$ is the Hubble distance. Combining Eqs. (14) and (19), we obtain $\displaystyle\ell_{0}=\frac{2cd_{H}\kappa}{3k}\,.$ (20) which becomes $\displaystyle\ell_{0}=\left(\frac{\kappa}{1\textrm{ Gpc}^{-1}}\right)\times 1400\textrm{ km}\,.$ (21) Given the posterior on $\kappa$ computed in Sec. II, we show the posterior on $\ell_{0}$, computed using Eq. (21) in Fig. 4. We can thus bound O1-O2: $\displaystyle\ell_{0}\lesssim 2.8\times 10^{3}\textrm{ km}\,,$ O3a: $\displaystyle\ell_{0}\lesssim 1.8\times 10^{3}\textrm{ km}\,,$ All: $\displaystyle\ell_{0}\lesssim 1.0\times 10^{3}\textrm{ km}\,.$ Figure 4: Posterior on $\ell_{0}$, the CS field length scale for the canonical CS field profile given in Eqs. (17) and (18). We compute the likelihood from the observations in O1-O2 (light blue curve), O3a (pink curve), and both catalogs (black curve). Each vertical line corresponds to $1-\sigma$. ### III.3 Projected value of $\ell_{0}$ with more detections We can project future constrains on $\ell_{0}$ using the difference in constraints we have obtained with GWTC-2 results to see how this constraint would improve with future detections. At fixed detector sensitivity, we expect that the constraint will go as $\ell_{0}(N)\sim 1/\sqrt{N}$, where $N$ is the number of detections. But as the detector sensitivity changes so does the typical distance to a detected merger. Since advanced LIGO at design sensitivity is expected to have a larger reach in distance Abbott _et al._ (2018), we set the typical value of the redshift to $z=0.75$. Repeating the previous analysis with the O1-O2 detections and the O3a detections, and with $z=0.75$ instead of $z=0.3$, we find that $\displaystyle\ell_{0}(N=10,z=0.75)$ $\displaystyle=1300\textrm{ km}$ (22) $\displaystyle\ell_{0}(N=47,z=0.75)$ $\displaystyle=490\textrm{ km}$ (23) With 1000 BBH detections at design sensitivity, for example, we would expect to bound $\ell_{0}\lesssim 100\,\mathrm{km}$. This projection is the result of two anticipated improvements—first in the greater reach in redshift of LIGO at design sensitivity, and second in the number of detections. ### III.4 Implications of Chern-Simons constraint Let us compare the physical constraint on the canonical Chern-Simons lengthscale from Sec. III.2 to additional observed bounds on the non-dynamical theory, using the $1000$ km bound we obtain from GWTC-2. Smith et al. Smith _et al._ (2008) used Solar-System measurements of frame-dragging from LAGEOS and Gravity Probe B to bound $|\dot{\vartheta}|\leq 3000(\kappa/\alpha)$ km. We have chosen $\alpha=\kappa$ in this study, and for the canonical profile, we have $\dot{\vartheta}=\ell_{0}$. Hence, the Smith et al. constraint becomes $\ell_{0}\leq 3000$ km. The bound from GWTC-2 is smaller than this number, indicating that LIGO events can constrain the non-dynamical theory more tightly than this Solar-System test. Alexander et al. proposed an amplitude birefringence analysis with LISA Alexander _et al._ (2008), estimating that for a $10^{6}\,M_{\odot}$ BBH at redshift $z\sim 15$, one could bound $\ell_{0}\leq 10^{-2}$ km Alexander and Yunes (2009).555Note that the analysis in this paper was performed for a dark- energy dominated universe, which is applicable to LIGO sources with $z\sim 1$, while the LISA analysis required a matter-dominated universe. This is a stronger bound that the one obtained in this paper, and attempting to achieve such a bound with LIGO-Virgo events would require $N\sim 10^{11}$ detections (cf. Sec. III.3). The authors of Alexander _et al._ (2008) perform a Fisher- matrix analysis for a source sweeping through $10^{-4}-10^{-2}$ Hz, keeping track of the frequency dependence in $k(t)$ and hence $\iota_{\mathrm{obs}}(t)$. In this study, we have approximated $k(t)$ as a constant $2\pi\times 100$ Hz, which in turn corresponds to setting $\iota_{\mathrm{obs}}(t)$ to a constant function of time-varying apparent inclination angle described in Alexander _et al._ (2008). While LISA is sensitive to this effect due to probing long BBH inspirals, LIGO is not sensitive to this effect, as there are not enough cycles in the LIGO band to probe precession for most events Abbott _et al._ (2019a). Additionally, Hu et al. Hu _et al._ (2020) performed a study analyzing the capability of a network of future space-based detectors (LISA, Taiji, and TianQin) to constrain parity violations in gravitational wave propagation, finding that for a $10^{6}\,M_{\odot}$ event at 20 Gpc, the parity violating scale from amplitude birefringence could be bounded to $M_{\mathrm{PV}}>\mathcal{O}(10^{-15})$ eV, corresponding to $2\times 10^{5}$ km. This, as the authors note, is a weaker bound than the constraint from ground-based detectors. Yunes and Spergel Yunes and Spergel (2009) performed a binary pulsar test with PSR J0737–3039, finding $\ell_{0}\lesssim 6\times 10^{-9}$ km, a bound much stronger than the one reported in this paper. The periastron precession of a system is corrected in CS, with the gradient of $\vartheta$ selecting a preferred direction in spacetime for the correction. The strength of this correction relative to GR is governed by $a^{2}/R^{2}$, where $a$ is the semimajor axis of the system, and $R$ is the radius of the object. With a large separation ($\sim 10^{6}$ stellar radii in this case), and small radii, a binary pulsar system produces a very strong constraint. However, as shown in Ali-Haimoud (2011), this analysis failed to account for several effects that lead to a suppression of the rate of periastron precession. In particular, Yunes and Spergel (2009) modeled PSR J0737–3039B as a point particle, rather than an extended body with radius $R_{B}$. If $R_{B}$ is larger than $2\pi\ell_{0}$ (the CS wavelength), the average force per unit mass is suppressed by a factor of $\sim 15(\ell_{0}/R_{B})^{3}$. Thus, in order to match the observed constraint on periastron precession, $\ell_{0}$ must be $\gtrsim R_{B}$. Indeed, Ali-Haimoud (2011) computed a corrected constraint of $\ell_{0}\lesssim 0.4$ km. In addition, Yunes and Spergel (2009) probes a different physical regime than we probe in this paper. Yunes and Spergel assume the canonical, global $\vartheta=\ell_{0}t$ profile, but use a local measurement to probe $\ell_{0}$. This involves assuming that the canonical profile, which has no spatial dependence, truly holds within our galaxy, and that there are no spatial density variations in the field near PSR J0737-3039. In this paper, however, we use an integrated history of $\vartheta$, sampling its temporal evolution, all the way from redshift $z\sim 1$ to present day. Over such cosmological distances, choosing the smooth, isotropic profile $\vartheta=\ell_{0}t$ may be justified, as any spatial effects can be presumed to integrate out. Thus, our analysis differs from binary pulsar tests in that we have used a global measurement to constraint a global quantity, without making any local assumptions. Recently, Wang et al. Wang _et al._ (2020) analyzed the presence of amplitude and velocity birefringence in GWTC-1, the first catalog of LIGO and Virgo detections, finding no evidence of parity violation. Their methods are different from the ones presented in this paper, as they match GWTC-1 data against GW templates that include birefringence effects, rather than looking at an ensemble of inclination angles. The constraint on the parity violating energy scale found in Wang _et al._ (2020) is $M_{\mathrm{PV}}>0.07$ GeV, which corresponds to a lengthscale of $\hbar c/M_{\mathrm{PV}}\sim 10^{-18}$ km. However, this comes from velocity birefringence effects, as LIGO is more sensitive to phase, rather than amplitude, modifications. Indeed, the constraint from amplitude birefringence effects only is $M_{\mathrm{PV}}>10^{-22}$ GeV which corresponds to $\sim 2000$ km. Similarly, Yamada et al. Yamada and Tanaka (2020) performed a parametrized tests of parity violation in gravitational wave propagation for GWTC-1, finding a minimum bound of $\ell_{0}\leq 1422$ km for GW151226 for CS gravity. Our GWTC-2 result of $\ell_{0}\leq 1000$ km improves on both of these results. ## IV Conclusion In this study, we have used GWTC-2 Abbott _et al._ (2020a); LIGO Scientific Collaboration and Virgo Collaboration (2019), including events from the first three observation runs, to perform a new test of general relativity (GR). We have placed an observational bound on gravitational wave amplitude birefringence, which is absent in GR, but present in various beyond-GR theories. Namely, we have bounded the opacity parameter governing the strength of the amplitude birefringence to $\kappa\lesssim 0.74\textrm{ Gpc}^{-1}$ (Sec. II.2). This general opacity constraint can then be mapped onto any beyond-GR theory exhibiting amplitude birefringence (see Zhao _et al._ (2020a) for a review). We have focused on (non-dynamical) Chern-Simons gravity, a beyond-GR theory with motivations in string theory and loop quantum gravity (Sec. III). We have used our results for $\kappa$ to bound $\zeta(\theta)$, a general CS parameter governing the CS scalar field, to $\zeta(\vartheta)\lesssim 4.9\times 10^{-21}$. We then computed the constraint on the CS lengthscale of the canonical scalar field profile, to give $\ell_{0}\lesssim 1.0\times 10^{3}$ km (Sec. III.2). One of the main benefits of our analysis is that it is simple and fast (of order minutes), and only requires looking at inclination angle posterior distributions for gravitational wave events, which are readily available from LIGO and Virgo catalogs, without performing an independent parameter estimation analysis. We plan to repeat this analysis with future LIGO and Virgo observations, obtaining an even tighter bound on this beyond-GR effect. ## Acknowledgements MO and WF are funded by the Center for Computational Astrophysics at the Flatiron Institute, which is supported by the Simons Foundation. MI is supported by NASA through the NASA Hubble Fellowship grant #HST–HF2–51410.001–A awarded by the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., for NASA, under contract NAS5–26555. ## Appendix A Joint distance-$\kappa$ analysis Let us now consider dropping the assumption used in the main analysis of Sec. II that all of the observed GW events are at the same distance. This requires performing a joint analysis for $d_{L}$ and $\kappa$. For ease of notation, let us write $\displaystyle c$ $\displaystyle\equiv\cos\iota\,,$ (24) $\displaystyle c_{o}$ $\displaystyle\equiv\cos\iota_{\mathrm{obs}}\,,$ $\displaystyle d_{Lo}$ $\displaystyle\equiv d_{L,\mathrm{obs}}\,.$ We can model this entire system as a probabilistic graphical model (PGM), as illustrated in Fig. 5. The directions in the PGM denote the influences between various variables. In our case, $\kappa$, $d_{L}$, and $c$, which are true astrophysical parameters, influence the observed variables $d_{Lo}$ and $c_{o}$. In turn, $d_{Lo}$ and $c_{o}$ influence the observed gravitational wave data $D_{\mathrm{GW}}$. In this model, $\kappa$ plays a special role, because it is shared by the entire population. Figure 5: Probabilistic graphical model illustrating the relationship between the variables (see Eq. (24) for abbreviations). The gray region represents data that applies to each gravitational wave event, while $\kappa$ is a universal constant, independent of each event. The true luminosity distance $d_{L}$, the true inclination angle $c$, and $\kappa$ affect the observed luminosity distance $d_{Lo}$ and the observed inclination angle $c_{o}$. These in turn affect the observed gravitational wave data $D_{\mathrm{GW}}$. For each event, we can marginalize over the distributions for the observed variables, $\\{d_{Lo},c_{o}\\}$ to obtain $p(D_{\mathrm{GW}}|\kappa)$ as $\displaystyle p(D_{\mathrm{GW}}|\kappa)=\int\mathrm{d}c_{o}\,\mathrm{d}d_{Lo}\,p(D_{\mathrm{GW}}|c_{o},d_{Lo})p(c_{o},d_{Lo}|\kappa)\,.$ (25) The above expression is a standard marginalization using the PGM, without making any astrophysical arguments. We can compute the likelihood for $M_{\mathrm{obs}}$ gravitational wave observations as the product $\displaystyle p\left(\left\\{D_{\mathrm{GW},j}\mid j=1,\ldots,M_{\mathrm{obs}}\right\\}\mid\kappa\right)=\prod_{j=1}^{M_{\mathrm{obs}}}p\left(D_{\mathrm{GW},j}\mid\kappa\right)\,,$ (26) where we use Eq. (25) to compute each of the individual likelihoods in the product. Let us now work with Eq. (25), further marginalizing over $d_{Lo}$ as $\displaystyle p(D_{\mathrm{GW}}|\kappa)$ $\displaystyle=\int\mathrm{d}c_{o}\,\mathrm{d}d_{Lo}\,p(D_{\mathrm{GW}}|c_{o},d_{Lo})$ (27) $\displaystyle\quad\times p(c_{o}|\kappa,d_{Lo})p(d_{Lo})\,.$ In order to compute $p(d_{Lo})$, we assert that the distribution of observed luminosity distances tracks the star formation rate, with $\displaystyle p\left(d_{Lo}\right)\propto\frac{\left(1+z\right)^{\alpha}}{1+\left(\frac{1+z}{1+z_{p}}\right)^{\beta}}\frac{\mathrm{d}V}{\mathrm{d}z}\frac{\mathrm{d}z}{\mathrm{d}d_{L}}\frac{1}{1+z}$ (28) where $\alpha=2.7$, $z_{p}=1.9$, and $\beta=5.6$ from Madau and Dickinson (2014). Effectively, we adjust the true merger rate evolution with redshift to match the observed distribution to the star formation rate (this is consistent with the population analysis in Abbott _et al._ (2019d)). We do this to avoid learning anything about $\kappa$ from any imposed prior on the _true_ merger rate evolution, since we are a priori very uncertain about it. Now, we need to compute $p(c_{o}|\kappa,d_{Lo})$, We assume from isotropy that the true inclination angle at the source, $c$, is independent of $d_{L}$ and $\kappa$, giving $\displaystyle p(c|d_{L},\kappa)=\frac{1}{2}\,.$ (29) Then, we can compute $p(c_{o}|\kappa,d_{Lo})$ through a substitution of variables as $\displaystyle p(c_{o}|\kappa,d_{Lo})=p(c|d_{L},\kappa)\left|\frac{\partial c}{\partial c_{o}}\right|\,.$ (30) From Eq. (7), we can compute $\displaystyle\frac{\partial c}{\partial c_{o}}=1+c_{o}\kappa d_{C}+\mathcal{O}(\kappa d_{C})^{2}\,.$ (31) Thus, we obtain $\displaystyle p(c_{o}|\kappa,d_{Lo})=\frac{1}{2}\left(1+c_{o}\kappa d_{C}+\mathcal{O}(\kappa d_{C})^{2}\right)\,.$ (32) Now we have all of the pieces of Eq. (27). To make the above expressions valid, we impose that $\kappa d_{C}\ll 1$. We enforce this by choosing a flat prior on $\kappa$ symmetric about zero, with support up to maximum allowed value $\kappa_{\mathrm{max}}$ determined by the largest value of the distance, $d_{C\mathrm{,max}}$. We choose the 99th percentile value of $d_{C}$ in each dataset to give $d_{C\mathrm{,max}}$ (cf. Fig. 6 for an illustration). In practice, we have access not to continuous probability distributions, but rather to $N$ samples from each gravitational wave events. Thus, we express the integral in Eq. (27) as a sum over $N$ samples, giving $\displaystyle p\left(D_{\mathrm{GW}}\mid\kappa\right)$ $\displaystyle\simeq\frac{1}{N}\times$ (33) $\displaystyle\sum_{n=1}^{N}\frac{p\left(c_{o,n}\mid d_{Lo,n},\kappa\right)p\left(d_{Lo,n}\right)p\left(\vec{\theta}_{n,\mathrm{other}}\right)}{p\left(\vec{\theta}_{n}\right)}\,.$ The quantity, $\vec{\theta}_{\mathrm{n}}$ refers to all of the parameters of the model. The quantity $\vec{\theta}_{\mathrm{n,other}}$, meanwhile, refers to all of the parameters besides the distances, inclination angles, and $\kappa$ in the model, such as the masses and spins of the black holes. We can use priors on $p(\vec{\theta}_{\mathrm{n,other}})$ to re-sample the distributions on parameters given in GWTC-2, with weights $\displaystyle w_{n}=\frac{p\left(d_{Lo,n}\right)p\left(\vec{\theta}_{\mathrm{other},n}\right)}{p\left(\vec{\theta}_{n}\right)}\,,$ (34) to give $\displaystyle p\left(D_{\mathrm{GW}}\mid\kappa\right)\simeq\frac{1}{N^{\prime}}\sum_{n=1}^{N^{\prime}}w_{n}p\left(c_{o,n}\mid d_{Lo,n},\kappa\right)\,,$ (35) for the sum in Eq. (33). In particular, in keeping with Eq. (28), we want to choose the prior on masses and distances to track the star formation rate. The prior $p(m_{1},m_{2},d_{Lo})$ used in GWTC-2 is flat in detector frame masses and flat in $c_{o}$, of the form $\displaystyle p\left(m_{1},m_{2},d_{Lo}\right)\propto\frac{\partial m_{1}^{\mathrm{det}}}{\partial m_{1}}\frac{\partial m_{2}^{\mathrm{det}}}{\partial m_{2}}d_{L}^{2}=\left(1+z\right)^{2}d_{L}^{2}\,.$ (36) We will re-weight using a prior on the masses and is proportional to $m_{1}^{-1.6}$ and flat in mass ratio, $q$, the approximate best-fit distribution from Abbott _et al._ (2019d), of the form $\displaystyle p\left(m_{1},m_{2},d_{Lo}\right)\propto m_{1}^{-1.6}\frac{\partial q}{\partial m_{2}}p\left(d_{Lo}\right)=m_{1}^{-2.6}p\left(d_{Lo}\right),$ (37) where $p\left(d_{Lo}\right)$ tracks the star formation rate as given in Eq. (28). Note that we do not consider the parameter space of other physical binary black hole populations in this study, in part because population models are not presently well-constrained with GWTC-2 Abbott _et al._ (2020c). We then combine all of the events using Eq. (26) to give the likelihood across all events. From this likelihood, we can then compute the posterior $p\left(\kappa\mid\left\\{D_{\mathrm{GW},j}\mid j=1,\ldots,M_{\mathrm{obs}}\right\\}\right)$ using a flat prior on $\kappa$, normalizing to integrate to 1. We show the resulting posterior on $\kappa$ for GWTC-2 in Fig. 6. We see that we do not get an informative constraint on $\kappa$ from O1-O2, as in the analysis presented in Sec. II. However, adding in O3a, we can get a constraint consistent with $\kappa=0$. Since $\kappa=0$ corresponds to GR, by comparing the value of the posterior to the prior at $\kappa=0$, we can obtain an evidence for GR. While we see that for O1-O2 we effectively recover the prior value at $\kappa=0$, giving us no information, in the case of the simulated detections, we can recover informative evidence for GR. However, for O3a and all of the detections, the result does give a constraint around $\kappa=0$. Figure 6: Posterior distribution on $\kappa$ using a joint distance-$\kappa$ analysis. We show the posteriors for the O1-O2 detections (light blue curve), O3a detections (pink curve), and all detections (purple curve). The combined result prefers $\kappa=0$, thus showing consistency with GR. Compare to Fig. 6, which assumes a fixed distance for all events. For each dataset, we show the corresponding prior on $\kappa$ with a dot-dashed line, given through the condition that $\kappa d_{C}\ll 1$. The priors are different for the two datasets as they have different maximum values of $d_{C}$. We also fit a Gaussian to all of the detections to estimate a variance for the distribution (black dashed curve, visually overlapping with the data). Note that this analysis requires that $\kappa d_{C}\ll 1$, and thus we must limit the values of $\kappa$ considered consistent with our events with large comoving distances for our analysis to be valid; going beyond linear order in the above relations is possible, but the solutions for $c_{o}\left(c,d_{L,o}\right)$ become multi-valued, significantly complicated the analysis. The events with confident constraints on inclination angle shown in Fig. 1 are at redshifts of $0.05\lesssim z\lesssim 0.38$. GWTC-2 does contain events at redshifts up to $z=1$ Abbott _et al._ (2020c), but the inclination measurements from these events are uninformative. For future observations, however, we have to be cautious of the $\kappa d_{C}\ll 1$ requirement when bounding $\kappa$ with events at large redshifts in order for the linear analysis to remain valid. We fit a Gaussian to the computed distribution on $\kappa$ (for all of the gravitational wave events) in Fig. 6, finding a mean of $-0.035\textrm{ Gpc}^{-1}$, and a standard deviation of $\sigma=0.4\textrm{ Gpc}^{-1}$. This value of $\sigma$ is larger than the width of the prior support we impose to satisfy the $\kappa d_{C}\ll 1$ constraint. We can estimate, however, how many future detections it will take for $\sigma$ to lie inside of the prior. For the same distance distribution of observed sources, $\sigma$ will decrease by a factor of $\sqrt{N}$ for $N$ more detections. For $\sigma$ to decrease by a factor of two from $0.4\textrm{ Gpc}^{-1}$ to $0.2\textrm{ Gpc}^{-1}$, we thus require $N\sim 30$ more informative events. However, for future gravitational wave detections, we know that we will be able to observe further distances, which will affect the number of detections and hence the behavior of $\sigma$. Specifically, the rate at which we observe new events increases with distance $d_{C}$ as $d_{C}^{3}$ (since the overall observable volume increases). Thus, $\sigma$ will decrease with distance as $d_{C}^{-3/2}$.666Here we make the assumption that $\sigma$ is otherwise independent of distance, conservatively ignoring the fact that events that are further can give larger constraints on amplitude birefringence, and assuming that the inclination angle can be measured with similar accuracy at various distances. This increased distance, however, will decrease the allowed value of $\kappa$ (from the constraint $\kappa d_{C}\ll 1$) by a factor of $d_{C}^{-1}$. Thus, as the observable distance increases, $\sigma$, the variance on the measured $\kappa$, will decrease faster than the prior on the allowed values of $\kappa$. Hence, in time, we will be able to make a more precise and valid measurement of $\kappa$. ## Appendix B Uninformative inclination distributions In order the quantify the amount of information about $\kappa$ contained in the GWTC-2 detections, we must compare the results (whether qualitatively or quantitatively through a Kullback-Leibler divergence) to the distribution on $\kappa$ that we would get from detections that are completely uninformative about $\cos\iota$. Of course, such uninformative measurements must generate a posterior for $\kappa$ that is equal to the prior (that is, they must generate a flat likelihood function); but it is an interesting test for any practical inference method that it satisfies this condition. To generate such a test for our methods here, we produce an uninformative distribution on $\cos\iota$ for all detections. We generate $N_{\mathrm{samp}}$ mock samples from a distribution that is $\mathcal{U}[-1,1]$. We can then take the ensemble of $N_{\mathrm{det}}$ such detections and compute a likelihood distribution on $\Delta$ using the procedure in Sec. II.1, following with a computation of $\kappa$. However, when generating these samples, we must be careful about the fact that we are considering an uninformative distribution. For each detection, we obtain a certain amount of Poisson noise given that we only have $N_{\mathrm{samp}}$ discrete samples. Naively, one would expect these Poisson fluctuations to cancel one another out as we accumulate more detections, converging to some ‘true value’. However, because each successive uninformative ‘measurement’ of $\cos\iota\in\mathcal{U}[-1,1]$ offers no new information, there is no such sense of convergence. Instead, the detections essentially result in a random walk in the slope of the likelihood with the $\Delta$ parameter. We compute a log-likelihood distribution on $\Delta$ over all of the detections using $\displaystyle\log\mathcal{L}(\Delta)=\sum_{\mathrm{Detections}}\log\left[N_{-}\frac{(1-\Delta)}{2N_{\mathrm{samp}}}+N_{+}\frac{(1+\Delta)}{2N_{\mathrm{samp}}}\right]$ (38) where for each detection, $N_{-}$ is the number of samples with $\cos\iota<0$ and $N_{+}=N_{\mathrm{samp}}-N_{-}$ is the number of samples with $\cos\iota>0$. For a uniform distribution, we would expect to have $N_{-}=N_{+}=\frac{1}{2}$, so let us write, to linear order, $N_{-}/N_{\mathrm{samp}}=\frac{1}{2}+\epsilon$ and $N_{+}/N_{\mathrm{samp}}=\frac{1}{2}-\epsilon$. For any particular detection, assuming $N_{\mathrm{samp}}\gg 1$, $\epsilon$ is approximately normally distributed with mean zero and standard deviation $1/\sqrt{N_{\mathrm{samp}}}$. Eq. (38) then results in $\displaystyle\log\mathcal{L}(\Delta)=\sum_{\mathrm{Detections}}\log\left[\frac{1}{2}-\Delta\epsilon\right]\,,$ (39) which for each detection results in a line with slope linearly dependent on $\epsilon$. Summing the independent, normally-distributed random variables $\epsilon$ gives $\log\mathcal{L}(\Delta)=\mathrm{const}-\Delta\sum_{\mathrm{Detections}}\epsilon.$ (40) The sum of normally-distributed $\epsilon$ results in a random-walk for the slope of the likelihood with $\Delta$; the sum is, itself, normally- distributed with mean zero and standard deviation $\sqrt{N_{\mathrm{det}}/N_{\mathrm{samp}}}$. In order to ensure that uninformative detections do not accumulate a significant slope in $\mathcal{L}(\Delta)$, we must ensure that $\displaystyle N_{\mathrm{samp}}\gg N_{\mathrm{det}}$ (41) and thus have a number of samples that is dependent on the number of detections in the uninformative case. Note that this is different from what we do in practice, where we assume that the gravitational wave events are informative about $\cos\iota$ and hence $\Delta$, and we use a fixed number of samples (1024 in this study) from each posterior distribution in our calculations. We can see the outcome of this in Fig. 7, where we plot the resulting distribution on $\Delta$ from uninformative samples with and without imposing the criterion in Eq. (41), where we obtain convergence to a flat distribution when we satisfy the criterion. Figure 7: Posterior distribution on $\Delta$ computed from uninformative distributions of $\cos\iota\in\mathcal{U}[-1,1]$ for each detection. Each solid curve corresponds to the combined posterior distribution on $\Delta$ for the given number of detections, and the number of samples for each detection. The top panel corresponds to using a constant number of samples for each $N_{\mathrm{det}}$, which does not converge to the expected flat distribution on $\Delta$ (dashed grey line) with increasing detections. The bottom panel, however, shows the case where $N_{\mathrm{samp}}$ changes with $N_{\mathrm{det}}$ to satisfy the criterion in Eq. (41), indeed showing convergence to the expected flat distribution. The slope of the posterior is, in each case, comparable to $\sqrt{N_{\mathrm{det}}/N_{\mathrm{samp}}}$. ## Appendix C Derivation of $\zeta(\vartheta)$ for dark-energy dominated universe We now work through the derivation of $\zeta(\vartheta)$ (cf. Eq. (III)) for a dark-energy dominated universe. We follow the steps of Alexander _et al._ (2008), which computed $\zeta(\vartheta)$ for a matter-dominated universe. We work in units of conformal time $\eta$, with $[\eta]=L^{0}$, and where $\eta=1$ corresponds to present-day. The scale factor $a$ has units of $[a]=L$. The conformal time and proper time $t$ are related as $dt=ad\eta$. We use notation for derivatives $\dot{f}=\partial_{t}f$ and $f^{\prime}=\partial_{\eta}f$. $H\equiv\dot{a}/a$ is the Hubble parameter, with $[H]=L^{-1}$, and $\mathcal{H}\equiv a^{\prime}/a$ is the conformal Hubble parameter with dimensions $[\mathcal{H}]=L^{0}$. Quantities with subscript $0$, such as $\\{a_{0},H_{0},\mathcal{H}_{0}\\}$, refer to present- day values of the parameters. As stated before, the CS scalar field $\vartheta$ has dimensions of $[\vartheta]=L^{2}$, for the choice of $\alpha=\kappa$ for the CS coupling constant (cf. Eq. (12)). We set $G=c=1$ for this calculation. Let us assume that right and left polarized gravitational waves have the following profile (cf. Eq. 189 in Alexander and Yunes (2009)), $\displaystyle h_{\mathrm{R},\mathrm{L}}=A(1+\lambda_{\mathrm{R},\mathrm{L}}\cos\iota)^{2}\exp[-i(\phi_{0}+\Delta\phi_{\mathrm{R},\mathrm{L}})]\,,$ (42) where $\iota$ is the inclination angle between the angular momentum of the source and the observer’s line of sight, and $A$ is an amplitude dependent on parameters of the source that is the same for both polarizations. The quantity $\lambda_{\mathrm{R}}=+1$ for right-handed polarizations, and $\lambda_{\mathrm{L}}=-1$ for left-handed polarizations. The quantity $\phi_{0}$ is the gravitational wave phase as given by GR, and $\Delta\phi_{\mathrm{R},\mathrm{L}}$ is the CS modification to the gravitational wave phase. Let us write the total phase as $\displaystyle\phi_{\mathrm{R},\mathrm{L}}(\eta)=\phi_{0}(\eta)+\Delta\phi_{\mathrm{R},\mathrm{L}}(\eta)\,,$ (43) With the profile in Eq. (42), the ratio between the right and left polarized strain becomes $\displaystyle\frac{h_{\mathrm{R}}}{h_{\mathrm{L}}}=\frac{(1+\cos\iota)^{2}}{(1-\cos\iota)^{2}}\exp[-i(\Delta\phi_{\mathrm{R}}-\Delta\phi_{\mathrm{L}})]\,.$ (44) It is the quantity $\displaystyle\Delta\phi_{\mathrm{R}}-\Delta\phi_{\mathrm{L}}$ (45) that we are thus interested in computing, and which is related to $\zeta$ (cf. Eq. (III)) as $\displaystyle\frac{2k}{H_{0}}\zeta=-i(\Delta\phi_{\mathrm{R}}-\Delta\phi_{\mathrm{L}})\,.$ (46) The standard linearized Einstein equations for metric perturbations in a Friedmann-Robertson-Walker (FRW) universe are modified through the inclusion of CS coupling to a scalar field. The equation for the phase of circularly polarized modes thus takes the form (cf. Sec. 2.B in Alexander and Yunes (2009) for a full derivation) $\displaystyle\left[i\phi_{\mathrm{R},\mathrm{L}}^{\prime\prime}+(\phi_{\mathrm{R},\mathrm{L}}^{\prime})^{2}+\mathcal{H}^{\prime}+\mathcal{H}^{2}-\kappa^{2}\right]\left(1-\frac{\lambda_{\mathrm{R},\mathrm{L}}\kappa\vartheta^{\prime}}{a^{2}}\right)$ (47) $\displaystyle\quad=\frac{i\lambda_{\mathrm{R},\mathrm{L}}\kappa}{a^{2}}(\vartheta^{\prime\prime}-2\mathcal{H}\vartheta^{\prime})(\phi_{\mathrm{R},\mathrm{L}}^{\prime}-i\mathcal{H})\,,$ where $\kappa$ is the co-moving wave-number with units $[\kappa]=L^{0}$. For ease of notation, let us drop the ${\mathrm{R},\mathrm{L}}$ subscript and focus on a polarization with a generic $\lambda\in\\{-1,1\\}$. Following Alexander _et al._ (2008), we put Eq. (47) in terms of a host of other variables, namely $\displaystyle\;\;y\equiv\frac{\phi^{\prime}}{k}\;\;\;\;\gamma\equiv\frac{\mathcal{H}_{0}}{\kappa}\;\;\;\;\Gamma\equiv\frac{\mathcal{H}}{\mathcal{H}_{0}}$ (48) $\displaystyle\;\;\delta\equiv\frac{\mathcal{H}_{0}^{\prime}}{\kappa^{2}}\;\;\;\;\Delta\equiv\frac{\mathcal{H}^{\prime}}{\mathcal{H}_{0}^{\prime}}\;\;\;\;\epsilon=\frac{\vartheta_{0}^{\prime\prime}}{a_{0}^{2}}$ $\displaystyle\;\;\zeta\equiv\frac{\kappa\vartheta_{0}^{\prime}}{a_{0}^{2}}\;\;\;\;E\equiv\frac{\vartheta^{\prime\prime}}{a^{2}\epsilon}\;\;\;\;Z\equiv\frac{\kappa\vartheta^{\prime}}{a^{2}\zeta}\,.$ Eq. (47) thus becomes $\displaystyle\frac{y^{\prime}}{\kappa}+i(1-\gamma^{2}\Gamma^{2}-\delta\Delta-y^{2})=\frac{\lambda(\epsilon E-2\gamma\zeta\Gamma Z)}{1-\lambda\zeta Z}(y-i\gamma\Gamma)\,.$ (49) Thus far, nothing has been assumed about the scale factor or matter-energy content of the FRW universe. Let us assume, however, following Alexander _et al._ (2008) that $\vartheta$ and $\mathcal{H}$ evolve on cosmological timescales (with $f^{\prime}\sim\mathcal{H}f$), and so $\displaystyle\epsilon^{2}\sim(\gamma\zeta)^{2}\ll\gamma^{2}\sim\delta\,.$ (50) Then, we can say that all of the terms with factors of $\epsilon$ and $\gamma\zeta$ are perturbations, and hence we can write the solution to Eq. (49) as $\displaystyle y=y_{0}+\epsilon y_{0,1}+\gamma\zeta y_{1,0}+\ldots\,,$ (51) where $y_{0}$ is the value of $y$ obtained from pure GR (setting $\vartheta=0$ in Eq. (49)), and $\\{\epsilon,\gamma,\zeta\\}$ are given in Eq. (48). Next, we require that the perturbations vanish at some initial conformal time $\eta_{i}$, we obtain that (cf. Eq. 2.23 in Alexander and Yunes (2009)) $\displaystyle y_{0,1}(\eta)$ $\displaystyle=\lambda\mathcal{Y}[E](\eta)\,,$ (52) $\displaystyle y_{1,0}(\eta)$ $\displaystyle=-2\lambda\mathcal{Y}[\Gamma Z](\eta)\,,$ (53) where $\\{E,\Gamma,Z\\}$ are functions of $\vartheta$ given in Eq. (48), and $\displaystyle\mathcal{Y}[g](\eta)\equiv\kappa e^{2i\phi_{0}(\eta)}\int_{\eta_{i}}^{\eta}dxe^{-2i\phi_{0}(x)}y_{0}(x)g(x)\,,$ (54) for some function $g(\eta)$, where $\phi_{0}(\eta)$ is the gravitational wave phase from pure GR (obtained from solving Eq. (47) with $\vartheta=0$). The CS correction to the accumulated phase as the wave propagates from $\eta_{i}$ to $\eta$ (cf. Eq. 2.24 in Alexander and Yunes (2009)) is thus $\displaystyle\Delta\phi(\eta_{i},\eta)=\kappa\lambda\int_{\eta_{i}}^{\eta}d\eta\\{\epsilon\mathcal{Y}[E](\eta)-2\gamma\zeta\mathcal{Y}[\Gamma Z](\eta)\\}\,.$ (55) To summarize, our goal is to integrate Eq. (55) to obtain the CS modification to the gravitational wave phase, which will allow us to compute the ratio between right and left polarized stain modes for a given $\vartheta$, as expressed in Eq. (44). If we assume that $\gamma\ll 1$ (which is justified for the LIGO frequency range) then the function $\mathcal{Y}[g]$ (for some function $g[\eta]$) has the asymptotic expansion (cf. Eq. 2.25 in Alexander _et al._ (2008)) $\displaystyle\mathcal{Y}[g](\eta)\sim\frac{ie^{2i\phi_{0}(\eta)}}{2}\left[e^{-2i\phi_{0}(\eta)}\sum_{\ell=0}^{n}\left(\frac{1}{2ik}\right)^{\ell}+\left(\frac{1}{y_{0}}\frac{d}{d\eta}\right)^{\ell}g\right]^{\eta}_{\eta_{i}}\,.$ (56) We will follow Alexander _et al._ (2008) in going to order $\ell=0$ in this calculation, giving $\displaystyle\mathcal{Y}[g](\eta)$ $\displaystyle\sim\frac{ie^{2i\phi_{0}(\eta)}}{2}\left(e^{-2i\phi_{0}(\eta)}g(\eta)-e^{-2i\phi_{0}(\eta_{i})}g(\eta_{i})\right)$ $\displaystyle=\frac{i}{2}g(\eta)-\frac{i}{2}e^{2i(\phi_{0}(\eta)-\phi_{0}(\eta_{i}))}g(\eta_{i})\,.$ (57) Now our calculation diverges from that in Alexander and Yunes (2009), as we work in a dark-energy dominated (rather than matter-dominated) universe, with scale factor $\displaystyle a(t)=a_{0}e^{H_{0}t}\,.$ (58) Working in units of conformal time, we obtain $\displaystyle\eta(t)$ $\displaystyle=-\frac{1}{a_{0}H_{0}}e^{-H_{0}t}\,,$ (59) $\displaystyle t(\eta)$ $\displaystyle=\frac{\log\left(-\frac{1}{a_{0}H_{0}\eta}\right)}{H_{0}}\,,$ (60) which gives $\displaystyle a(\eta)$ $\displaystyle=-\frac{1}{H_{0}\eta}\,.$ (61) With the convention of $\eta=1$ corresponding to present day, we obtain $\displaystyle a_{0}$ $\displaystyle=-\frac{1}{H_{0}}\,,\;\;\;\;a^{\prime}(\eta)=\frac{1}{H_{0}\eta^{2}}\,,$ (62) $\displaystyle\mathcal{H}$ $\displaystyle\equiv\frac{a^{\prime}(\eta)}{a(\eta)}=\frac{1}{\eta}\,,\;\;\;\;\mathcal{H}_{0}=1\,,$ (63) $\displaystyle\mathcal{H}^{\prime}$ $\displaystyle=-\frac{1}{\eta^{2}}\,,\;\;\;\;\mathcal{H}^{\prime}_{0}=-1\,.$ (64) Now, we can compute all of the quantities in Eq. (48) for a dark-energy dominated universe as $\displaystyle\;\;y\equiv\frac{\phi^{\prime}}{\kappa}\;\;\;\;\gamma\equiv\frac{1}{\kappa}\;\;\;\;\Gamma\equiv\frac{1}{\eta}$ (65) $\displaystyle\;\;\delta\equiv\frac{-1}{\kappa^{2}}\;\;\;\;\Delta\equiv\frac{1}{\eta^{2}}\;\;\;\;\epsilon=H_{0}^{2}\vartheta_{0}^{\prime\prime}$ $\displaystyle\;\;\zeta\equiv\kappa\vartheta_{0}^{\prime}H_{0}^{2}\;\;\;\;E\equiv\eta^{2}\frac{\vartheta^{\prime\prime}}{\vartheta_{0}^{\prime\prime}}\;\;\;\;Z\equiv\eta^{2}\frac{\vartheta^{\prime}}{\vartheta_{0}^{\prime}}\,.$ Our aim is thus to evaluate Eq. (55) to obtain the CS correction to the phase, $\Delta\phi$. Now, we must first obtain $y_{0}$, the value of $\phi_{0}^{\prime}/k$ without a perturbation. Thus, we solve Eq. (49) with zero RHS to give $\displaystyle\frac{y_{0}^{\prime}}{k}+i(1-\gamma^{2}\Gamma^{2}-\delta\Delta-y^{2})=0$ $\displaystyle\frac{y_{0}^{\prime}}{k}+i(1-\frac{1}{\kappa^{2}\eta^{2}}-\frac{-1}{\kappa^{2}\eta^{2}}-y^{2})=0$ $\displaystyle\frac{y_{0}^{\prime}}{k}+i(1-y^{2})=0\,,$ (66) which gives solutions of the form $\displaystyle y_{0}=-i\tan(\kappa\eta-iC_{0})\,,$ (67) where $C_{0}$ is a constant of integration that we will leave unspecified for now. Integrating $\displaystyle\phi_{0}^{\prime}=\kappa y_{0}\,,$ (68) we obtain $\displaystyle\phi_{0}(\eta)=C_{1}+i\log(\cosh(C_{0}-i\kappa\eta))$ (69) we can freely set $C_{1}=0$ since we are interested in the difference between two values of $\phi_{0}$. Now, let us find $\Delta\phi$, the CS phase accumulated by the perturbations using Eq. (55). Using the form of $\mathcal{Y}[g]$ from Eq. (57), and the solution in Eq. (69), we compute $\displaystyle\phi_{0}(\eta)-\phi_{0}(\eta_{i})=$ (70) $\displaystyle\quad-i\log(\cosh(C_{0}-i\kappa\eta))+i\log(\cosh(C_{0}-i\kappa\eta_{i}))$ which gives $\displaystyle e^{2i(\phi_{0}(\eta)-\phi_{0}(\eta_{i}))}=\frac{\cosh(C_{0}-i\kappa\eta)^{2}}{\cosh(C_{0}-i\kappa\eta_{0})^{2}}$ (71) Thus, we have $\displaystyle\epsilon\mathcal{Y}[E](\eta)$ $\displaystyle=\frac{i}{2}H_{0}^{2}\vartheta_{0}^{\prime\prime}\left(\eta^{2}\frac{\vartheta^{\prime\prime}}{\vartheta_{0}^{\prime\prime}}-\eta_{i}^{2}\frac{\vartheta_{i}^{\prime\prime}}{\vartheta_{0}^{\prime\prime}}\frac{\cosh(C_{0}-i\kappa\eta)^{2}}{\cosh(C_{0}-i\kappa\eta_{0})^{2}}\right)$ $\displaystyle=\frac{i}{2}H_{0}^{2}\left(\eta^{2}\vartheta^{\prime\prime}-\eta_{i}^{2}\vartheta_{i}^{\prime\prime}\frac{\cosh(C_{0}-i\kappa\eta)^{2}}{\cosh(C_{0}-i\kappa\eta_{0})^{2}}\right)$ (72) and similarly $\displaystyle\gamma\zeta\mathcal{Y}[\Gamma Z](\eta)$ $\displaystyle=\frac{i}{2}\vartheta_{0}^{\prime}H_{0}^{2}\left(\eta\frac{\vartheta^{\prime}}{\vartheta_{0}^{\prime}}-\eta_{i}\frac{\vartheta_{i}^{\prime}}{\vartheta_{0}^{\prime}}\frac{\cosh(C_{0}-i\kappa\eta)^{2}}{\cosh(C_{0}-i\kappa\eta_{0})^{2}}\right)$ $\displaystyle=\frac{i}{2}H_{0}^{2}\left(\eta\vartheta^{\prime}-\eta_{i}\vartheta_{i}^{\prime}\frac{\cosh(C_{0}-i\kappa\eta)^{2}}{\cosh(C_{0}-i\kappa\eta_{0})^{2}}\right)\,.$ (73) Let us follow the logic below Eq. 3.4 of Alexander and Yunes (2009) to drop the oscillatory pieces, thus obtaining the overall integral from Eq. (55) of $\displaystyle\Delta\phi_{\mathrm{R},\mathrm{L}}\sim i\frac{\kappa}{2}\lambda_{\mathrm{R},\mathrm{L}}H_{0}^{2}\int_{\eta}^{1}\left(\eta^{2}\vartheta^{\prime\prime}(\eta)-2\eta\vartheta^{\prime}(\eta)\right)d\eta\,,$ (74) where we have reintroduced the ${\mathrm{R},\mathrm{L}}$ notation. Following Eq. (46), we have $\displaystyle\frac{2k}{H_{0}}\zeta=-i(\Delta\phi_{\mathrm{R}}-\Delta\phi_{\mathrm{L}})$ (75) and thus, using $\lambda_{\mathrm{R}}-\lambda_{\mathrm{L}}=2$, we obtain $\displaystyle\frac{2k}{H_{0}}\zeta=\kappa H_{0}^{2}\int_{\eta}^{1}\left(\eta^{2}\vartheta^{\prime\prime}(\eta)-2\eta\vartheta^{\prime}(\eta)\right)d\eta\,.$ (76) Writing $\kappa=k_{0}/H_{0}$ (cf. Eq. 3.8 in Alexander _et al._ (2008)), we obtain, $\displaystyle\zeta=\frac{H_{0}^{2}}{2}\int_{\eta}^{1}\left(\eta^{2}\vartheta^{\prime\prime}(\eta)-2\eta\vartheta^{\prime}(\eta)\right)d\eta\,.$ (77) Eq. (77) precisely gives us $\zeta$ for a dark-energy dominated universe. Let us double-check the units. In Eq. (46), $\zeta$ must be dimensionless. In this study, $[\vartheta]=L^{2}$ and $[H_{0}]=L^{-1}$, so indeed $[\zeta]=L^{0}$. ## References * Aasi _et al._ (2015) J. Aasi _et al._ (LIGO Scientific), Class. Quant. Grav. 32, 074001 (2015), arXiv:1411.4547 [gr-qc] . * Acernese _et al._ (2015) F. Acernese _et al._ (VIRGO), Class. Quant. Grav. 32, 024001 (2015), arXiv:1408.3978 [gr-qc] . * Abbott _et al._ (2019a) B. P. Abbott _et al._ (LIGO Scientific, Virgo), Phys. Rev. X9, 031040 (2019a), arXiv:1811.12907 [astro-ph.HE] . * Abbott _et al._ (2019b) B. P. Abbott _et al._ (LIGO Scientific, Virgo), (2019b), arXiv:1903.04467 [gr-qc] . * Isi _et al._ (2019a) M. Isi, M. Giesler, W. M. Farr, M. A. Scheel, and S. A. Teukolsky, Phys. Rev. Lett. 123, 111102 (2019a), arXiv:1905.00869 [gr-qc] . * Nair _et al._ (2019) R. Nair, S. Perkins, H. O. Silva, and N. Yunes, Phys. Rev. Lett. 123, 191101 (2019), arXiv:1905.00870 [gr-qc] . * Isi _et al._ (2019b) M. Isi, K. Chatziioannou, and W. M. Farr, Phys. Rev. Lett. 123, 121101 (2019b), arXiv:1904.08011 [gr-qc] . * Abbott _et al._ (2020a) R. Abbott _et al._ (LIGO Scientific, Virgo), (2020a), arXiv:2010.14527 [gr-qc] . * Abbott _et al._ (2020b) R. Abbott _et al._ (LIGO Scientific, Virgo), (2020b), arXiv:2010.14529 [gr-qc] . * Abbott _et al._ (2019c) R. Abbott _et al._ (LIGO Scientific, Virgo), (2019c), arXiv:1912.11716 [gr-qc] . * LIGO Scientific Collaboration and Virgo Collaboration (2019) LIGO Scientific Collaboration and Virgo Collaboration, “Parameter estimation sample release for GWTC-1,” https://dcc.ligo.org/LIGO-P1800370/public (2019). * Zhao _et al._ (2020a) W. Zhao, T. Zhu, J. Qiao, and A. Wang, Phys. Rev. D 101, 024002 (2020a), arXiv:1909.10887 [gr-qc] . * Alexander and Yunes (2009) S. Alexander and N. Yunes, Phys. Rept. 480, 1 (2009), arXiv:0907.2562 [hep-th] . * Crisostomi _et al._ (2018) M. Crisostomi, K. Noui, C. Charmousis, and D. Langlois, Phys. Rev. D97, 044034 (2018), arXiv:1710.04531 [hep-th] . * Conroy and Koivisto (2019) A. Conroy and T. Koivisto, JCAP 12, 016 (2019), arXiv:1908.04313 [gr-qc] . * Horava (2009) P. Horava, Phys. Rev. D79, 084008 (2009), arXiv:0901.3775 [hep-th] . * Green and Schwarz (1984) M. B. Green and J. H. Schwarz, Phys. Lett. 149B, 117 (1984). * Taveras and Yunes (2008) V. Taveras and N. Yunes, Phys. Rev. D78, 064070 (2008), arXiv:0807.2652 [gr-qc] . * Mercuri and Taveras (2009) S. Mercuri and V. Taveras, Phys. Rev. D80, 104007 (2009), arXiv:0903.4407 [gr-qc] . * Weinberg (2008) S. Weinberg, Phys. Rev. D77, 123541 (2008), arXiv:0804.4291 [hep-th] . * Nojiri _et al._ (2019) S. Nojiri, S. D. Odintsov, V. K. Oikonomou, and A. A. Popov, Phys. Rev. D100, 084009 (2019), arXiv:1909.01324 [gr-qc] . * Zhao _et al._ (2020b) W. Zhao, T. Liu, L. Wen, T. Zhu, A. Wang, Q. Hu, and C. Zhou, Eur. Phys. J. C 80, 630 (2020b), arXiv:1909.13007 [gr-qc] . * Alexander _et al._ (2008) S. Alexander, L. S. Finn, and N. Yunes, Phys. Rev. D78, 066005 (2008), arXiv:0712.2542 [gr-qc] . * Yunes _et al._ (2010) N. Yunes, R. O’Shaughnessy, B. J. Owen, and S. Alexander, Phys. Rev. D82, 064017 (2010), arXiv:1005.3310 [gr-qc] . * Yunes and Finn (2009) N. Yunes and L. S. Finn, _Laser Interferometer Space Antenna. Proceedings, 7th international LISA Symposium, Barcelona, Spain, June 16-20, 2008_ , J. Phys. Conf. Ser. 154, 012041 (2009), arXiv:0811.0181 [gr-qc] . * Yagi and Yang (2018) K. Yagi and H. Yang, Phys. Rev. D97, 104018 (2018), arXiv:1712.00682 [gr-qc] . * Abbott _et al._ (2020c) R. Abbott _et al._ (LIGO Scientific, Virgo), (2020c), arXiv:2010.14533 [astro-ph.HE] . * Mandel _et al._ (2019) I. Mandel, W. M. Farr, and J. R. Gair, Mon. Not. Roy. Astron. Soc. 486, 1086 (2019), arXiv:1809.02063 [physics.data-an] . * Fishbach _et al._ (2018) M. Fishbach, D. E. Holz, and W. M. Farr, Astrophys. J. 863, L41 (2018), [Astrophys. J. Lett.863,L41(2018)], arXiv:1805.10270 [astro-ph.HE] . * Pardo _et al._ (2018) K. Pardo, M. Fishbach, D. E. Holz, and D. N. Spergel, JCAP 1807, 048 (2018), arXiv:1801.08160 [gr-qc] . * Abbott _et al._ (2019d) B. P. Abbott _et al._ (LIGO Scientific, Virgo), Astrophys. J. 882, L24 (2019d), arXiv:1811.12940 [astro-ph.HE] . * Dai _et al._ (2020) L. Dai, B. Zackay, T. Venumadhav, J. Roulet, and M. Zaldarriaga, (2020), arXiv:2007.12709 [astro-ph.HE] . * Smith _et al._ (2018) G. P. Smith, M. Jauzac, J. Veitch, W. M. Farr, R. Massey, and J. Richard, Mon. Not. Roy. Astron. Soc. 475, 3823 (2018), arXiv:1707.03412 [astro-ph.HE] . * Kullback and Leibler (1951) S. Kullback and R. A. Leibler, Ann. Math. Statist. 22, 79 (1951). * Jackiw and Pi (2003) R. Jackiw and S. Y. Pi, Phys. Rev. D68, 104012 (2003), arXiv:gr-qc/0308071 [gr-qc] . * Abbott _et al._ (2016) B. P. Abbott _et al._ (LIGO Scientific, Virgo), Phys. Rev. Lett. 116, 131103 (2016), arXiv:1602.03838 [gr-qc] . * Yunes and Spergel (2009) N. Yunes and D. N. Spergel, Phys. Rev. D80, 042004 (2009), arXiv:0810.5541 [gr-qc] . * Abbott _et al._ (2018) B. Abbott _et al._ (KAGRA, LIGO Scientific, VIRGO), Living Rev. Rel. 21, 3 (2018), arXiv:1304.0670 [gr-qc] . * Smith _et al._ (2008) T. L. Smith, A. L. Erickcek, R. R. Caldwell, and M. Kamionkowski, Phys. Rev. D77, 024015 (2008), arXiv:0708.0001 [astro-ph] . * Hu _et al._ (2020) Q. Hu, M. Li, R. Niu, and W. Zhao, (2020), arXiv:2006.05670 [gr-qc] . * Ali-Haimoud (2011) Y. Ali-Haimoud, Phys. Rev. D83, 124050 (2011), arXiv:1105.0009 [astro-ph.HE] . * Wang _et al._ (2020) Y.-F. Wang, R. Niu, T. Zhu, and W. Zhao, (2020), arXiv:2002.05668 [gr-qc] . * Yamada and Tanaka (2020) K. Yamada and T. Tanaka, PTEP 2020, 093E01 (2020), arXiv:2006.11086 [gr-qc] . * Madau and Dickinson (2014) P. Madau and M. Dickinson, Annual Review of Astronomy and Astrophysics 52, 415 (2014), https://doi.org/10.1146/annurev-astro-081811-125615 .
∎ 11institutetext: S. Mishra 22institutetext: Indian Institute of Technology Kanpur, India 22email<EMAIL_ADDRESS>33institutetext: S. Prasad44institutetext: University of Illinois at Urbana-Champaign, USA 44email<EMAIL_ADDRESS>55institutetext: S. Mishra∗ [Corresponding Author]66institutetext: University of Illinois at Urbana-Champaign, USA 66email<EMAIL_ADDRESS> # Exploring multi-task multi-lingual learning of transformer models for hate speech and offensive speech identification in social media Sudhanshu Mishra Shivangi Prasad Shubhanshu Mishra∗ (Received: date / Accepted: date) ###### Abstract Hate Speech has become a major content moderation issue for online social media platforms. Given the volume and velocity of online content production, it is impossible to manually moderate hate speech related content on any platform. In this paper we utilize a multi-task and multi-lingual approach based on recently proposed Transformer Neural Networks to solve three sub- tasks for hate speech. These sub-tasks were part of the 2019 shared task on hate speech and offensive content (HASOC) identification in Indo-European languages. We expand on our submission to that competition by utilizing multi- task models which are trained using three approaches, a) multi-task learning with separate task heads, b) back-translation, and c) multi-lingual training. Finally, we investigate the performance of various models and identify instances where the Transformer based models perform differently and better. We show that it is possible to to utilize different combined approaches to obtain models that can generalize easily on different languages and tasks, while trading off slight accuracy (in some cases) for a much reduced inference time compute cost. We open source an updated version of our HASOC 2019 code with the new improvements at https://github.com/socialmediaie/MTML_HateSpeech. ###### Keywords: Hate Speech Offensive content Transformer Models BERT Language Models Neural Networks Multi-lingual Multi-Task Learning Social Media Natural Language Processing Machine Learning Deep Learning Open Source ## 1 Introduction With increased access to the internet, the number of people that are connected through social media is higher than ever (Perrin,, 2015). Thus, social media platforms are often held responsible for framing the views and opinions of a large number of people (Duggan et al.,, 2017). However, this freedom to voice our opinion has been challenged by the increase in the use of hate speech (Mondal et al.,, 2017). The anonymity of the internet grants people the power to completely change the context of a discussion and suppress a person’s personal opinion (Sticca and Perren,, 2013). These hateful posts and comments not only affect the society at a micro scale but also at a global level by influencing people’s views regarding important global events like elections, and protests (Duggan et al.,, 2017). Given the volume of online communication happening on various social media platforms and the need for more fruitful communication, there is a growing need to automate the detection of hate speech. For the scope of this paper we adopt the definition of hate speech and offensive speech as defined in the Mandl et al., (2019) as “insulting, hurtful, derogatory, or obscene content directed from one person to another person” (quoted from (Mandl et al.,, 2019)). In order to automate hate speech detection the Natural Language Processing (NLP) community has made significant progress which has been accelerated by organization of numerous shared tasks aimed at identifying hate speech (Mandl et al.,, 2019; Kumar et al.,, 2020, 2018). Furthermore, there has been a proliferation of new methods for automated hate speech detection in social media text (Salminen et al.,, 2018; Mishra et al., 2020b, ; Mishra and Mishra,, 2019; Mishra, 2020a, ; Waseem et al.,, 2017; Struß et al.,, 2019; Mandl et al.,, 2019; Mondal et al.,, 2017). However, working with social media text is difficult (Eisenstein,, 2013; Mishra and Diesner,, 2016; Mishra et al.,, 2014; Mishra and Diesner,, 2019; Mishra,, 2019; Mishra, 2020b, ; Mishra, 2020a, ), as people use combinations of different languages, spellings and words that one may never find in any dictionary. A common pattern across many hate speech identification tasks Mandl et al., (2019); Kumar et al., (2020); Waseem et al., (2017); Zampieri et al., (2019); Basile et al., (2019); Struß et al., (2019) is the identification of various aspects of hate speech, e.g., in HASOC 2019 (Mandl et al.,, 2019), the organizers divided the task into three sub-tasks, which focused on identifying the presence of hate speech; classification of hate speech into offensive, profane, and hateful; and identifying if the hate speech is targeted towards an entity. Many researchers have tried to address these types of tiered hate speech classification tasks using separate models, one for each sub-task (see review of recent shared tasks Zampieri et al., (2019); Kumar et al., (2018, 2020); Mandl et al., (2019); Struß et al., (2019)). However, we consider this approach limited for application to systems which consume large amounts of data, and are computationally constrained for efficiently flagging hate speech. The limitation of existing approach is because of the requirement to run several models, one for each language and sub-task. In this work, we propose a unified modeling framework which identifies the relationship between all tasks across multiple languages. Our aim is to be able to perform as good if not better than the best model for each task language combination. Our approach is inspired from the promising results of multi-task learning on some of our recent works (Mishra,, 2019; Mishra, 2020b, ; Mishra, 2020a, ). Additionally, while building a unified model which can perform well on all tasks is challenging, an important benefit of these models is their computational efficiency, achieved by reduced compute and maintenance costs, which can allow the system to trade-off slight accuracy for efficiency. In this paper, we propose the development of such universal modelling framework, which can leverage recent advancements in machine learning to achieve competitive and in few cases state-of-the-art performance of a variety of hate speech identification sub-tasks across multiple languages. Our framework encompasses a variety of modelling architectures which can either train on all tasks, all languages, or a combination of both. We extend the our prior work in Mishra and Mishra, (2019); Mishra et al., 2020b ; Mishra, 2020b ; Mishra, 2020a ; Mishra, (2019) to develop efficient models for hate speech identification and benchmark them against the HASOC 2019 corpus, which consists of social media posts in three languages, namely, English, Hindi, and German. We open source our implementation to allow its usage by the wider research community. Our main contributions are as follows: 1. 1. Investigate more efficient modeling architectures which use a) multi-task learning with separate task heads, b) back-translation, and c) multi-lingual training. These architectures can generalize easily on different languages and tasks, while trading off slight accuracy (in some cases) for a much reduced inference time compute cost. 2. 2. Investigate the performance of various models and identification of instances where our new models differ in their performance. 3. 3. Open source pre-trained models and model outputs at Mishra et al., 2020a along with the updated code for using these models at: https://github.com/socialmediaie/MTML_HateSpeech ## 2 Related Work Prior work (see Schmidt and Wiegand, (2017) for a detailed review on prior methods) in the area of hate speech identification, focuses on different aspects of hate speech identification, namely analyzing what constitutes hate speech, high modality and other issues encountered when dealing with social media data and finally, model architectures and developments in NLP, that are being used in identifying hate speech these days. There is also prior literature focusing on the different aspects of hateful speech and tackling the subjectivity that it imposes. There are many shared tasks Mandl et al., (2019); Kumar et al., (2018, 2020); Struß et al., (2019); Basile et al., (2019) that tackle hate speech detection by classifying it into different categories. Each shared task focuses on a different aspect of hate speech. Waseem et al., (2017) proposed a typology on the abusive nature of hate speech, classifying it into generalized, explicit and implicit abuse. Basile et al., (2019) focused on hateful and aggressive posts targeted towards women and immigrants. Mandl et al., (2019) focused on identifying targeted and un- targeted insults and classifying hate speech into hateful, offensive and profane content. Kumar et al., (2018, 2020) tackled aggression and misogynistic content identification for trolling and cyberbullying posts. Vidgen et al., (2019) identifies that most of these shared tasks broadly fall into these three classes, individual directed abuse, identity directed abuse and concept directed abuse. It also puts into context the various challenges encountered in abusive content detection. Unlike other domains of information retrieval, there is a lack of large data- sets in this field. Moreover, the data-sets available are highly skewed and focus on a particular type of hate speech. For example, Davidson et al., (2017) models the problem as a generic abusive content identification challenge, however, these posts are mostly related towards racism and sexism. Furthermore, in the real world, hateful posts do not fall into to a single type of hate speech. There is a huge overlapping between different hateful classes, making hate speech identification a multi label problem. A wide variety of system architectures, ranging from classical machine learning to recent deep learning models, have been tried for various aspects of hate speech identification. Facebook, YouTube, and Twitter are the major sources of data for most data-sets. Burnap and Williams, (2015) used SVM and ensemble techniques on identifying hate speech in Twitter data. Razavi et al., (2010) approach for abuse detection using an insulting and abusive language dictionary of words and phrases. Van Hee et al., (2015) used bag of words n-gram features and trained an SVM model on a cyberbullying dataset. Salminen et al., (2018) achieved an F1-score of 0.79 on classification of hateful YouTube and Facebook posts using a linear SVM model employing TF-IDF weighted n-grams. Recently, models based on deep learning techniques have also been applied to the task of hate speech identification. These models often rely on distributed representations or embeddings, e.g., FastText embeddings (Joulin et al., (2017), and paragraph2vec distributed representations (Le and Mikolov, (2014). Badjatiya et al., (2017) employed an LSTM architecture to tune Glove word embeddings on the DATA-TWITTER-TWH data-set. Risch and Krestel, (2018) used a neural network architecture using a GRU layer and ensemble methods for the TRAC 2018 (Kumar et al.,, 2018) shared task on aggression identification. They also tried back-translation as a data augmentation technique to increase the data-set size. Wang, (2018) illustrated the use of sequentially combining CNNs with RNNs for abuse detection. They show that this approach was better than using only the CNN architecture giving a 1% improvement in the F1-score. One of the most recent developments in NLP are the transformer architecture introduced by Vaswani et al., (2017). Utilizing the transformer architecture, Devlin et al., (2019) provide methods to pre-train models for language understanding (BERT) that have achieved state of the art results in many NLP tasks and are promising for hate speech detection as well. BERT based models achieved competitive performance in HASOC 2019 shared tasks. We Mishra and Mishra, (2019) fine tuned BERT base models for the various HASOC shared tasks being the top performing model in some of the sub-tasks. A similar approach was also used for the TRAC 2020 shared tasks on Aggression identification by us Mishra et al., 2020b still achieving competitive performance with the other models without using an ensemble techniques. An interesting approach was the use of multi-lingual models by joint training on different languages. This approach presents us with a unified model for different languages in abusive content detection. Ensemble techniques using BERT models (Risch and Krestel,, 2020) was the top performing model in many of the shared tasks in TRAC 2020. Recently, multi-task learning has been used for improving performance on NLP tasks (Liu et al.,, 2016; Søgaard and Goldberg,, 2016), especially social media information extraction tasks (Mishra,, 2019), and more simpler variants have been tried for hate speech identification in our recent works (Mishra and Mishra,, 2019; Mishra et al., 2020b, ). Florio et al., (2020) investigated the usage of AlBERTo on monitoring hate speech against Italian on Twitter. Their results show that even though AlBERTo is sensitive to the fine tuning set, it’s performance increases given enough training time. Mozafari et al., (2020) employ a transfer learning approach using BERT for hate speech detection. Ranasinghe and Zampieri, (2020) use cross-lingual embeddings to identify offensive content in multilingual setting. Our multi-lingual approach is similar in spirit to the method proposed in Plank, (2017) which use the same model architecture and aligned word embedding to solve the tasks. There has also been some work on developing solutions for multilingual toxic comments which can be related to hate speech.111https://www.kaggle.com/c/jigsaw- multilingual-toxic-comment-classification Recently, Mishra, 2020c also used a single model across various tasks which performed very well for event detection tasks for five Indian languages. There have been numerous competitions dealing with hate speech evaluation. OffensEval Zampieri et al., (2019) is one of the popular shared tasks dealing with offensive language in social media, featuring three sub-tasks for discriminating between offensive and non-offensive posts. Another popular shared task in SemEval is the HateEval Basile et al., (2019) task on the detection of hate against women and immigrants. The 2019 version of HateEval consists of two sub-task for determination of hateful and aggressive posts. GermEval Struß et al., (2019) is another shared task quite similar to HASOC. It focused on the Identification of Offensive Language in German Tweets. It features two sub-tasks following a binary and multi-class classification of the German tweets. An important aspect of hate speech is that it is primarily multi-modal in nature. A large portion of the hateful content that is shared on social media is in the form of memes, which feature multiple modalities like audio, text, images and videos in some cases as well. Yang et al., (2019) present different fusion approaches to tackle multi-modal information for hate speech detection. Gomez et al., (2020) explore multi-modal hate speech consisting of text and image modalities. They propose various multi-modal architectures to jointly analyze both the textual and visual information. Facebook recently released the hateful memes data-set for the Hateful Memes challenge Kiela et al., (2020) to provide a complex data-set where it is difficult for uni-modal models to achieve good performance. ## 3 Methods For this paper, we extend some of the techniques that we have used in TRAC 2020 in Mishra et al., 2020b as well as Mishra, (2019); Mishra, 2020a ; Mishra, 2020b , and apply them to the HASOC data-set Mandl et al., (2019). Furthermore, we extend the work that we did as part of the HASOC 2019 shared task Mishra and Mishra, (2019) by experimenting with multi-lingual training, back-translation based data-augmentation, and multi-task learning to tackle the data sparsity issue of the HASOC 2019 data-set. Figure 1: Task Description ### 3.1 Task Definition and Data All of the experiments reported hereafter have been done on the HASOC 2019 data-set (Mandl et al.,, 2019) consisting of posts in English (EN), Hindi (HI) and German (DE). The shared tasks of HASOC 2019 had three sub-tasks (A,B,C) for both English and Hindi languages and two sub-tasks (A,B) for the German language. The description of each sub-task is as follows (see Figure 1 for details): * • Sub-Task A : Posts have to be classified into hate speech HOF and non- offensive content NOT. * • Sub-Task B : A fine grained classification of the hateful posts in sub-task A. Hate Speech posts have to be identified into the type of hate they represent, i.e containing hate speech content (HATE), containing offensive content (OFFN) and those containing profane words (PRFN). * • Sub-Task C : Another fine grained classification of the hateful posts in sub- tasks A. This sub-task required us to identify whether the hate speech was targeted towards an individual or group TIN or whether it was un-targeted UNT. Table 1: Distribution of number of tweets in different data-sets and splits. task | DE | EN | HI ---|---|---|--- | train | dev | test | train | dev | test | train | dev | test A | 3,819 | 794 | 850 | 5,852 | 505 | 1,153 | 4,665 | 136 | 1,318 B | 407 | 794 | 850 | 2,261 | 302 | 1,153 | 2,469 | 136 | 1,318 C | | | | 2,261 | 299 | 1,153 | 2,469 | 72 | 1,318 The HASOC 2019 data-set consists of posts taken from Twitter and Facebook. The data-set only consists of text and labels and does not include any contextual information or meta-data of the original post e.g. time information. The data distribution for each language and sub-task is mentioned in Table 1. We can observe, that the sample size for each language is of the order of a few thousand post, which is an order smaller to other datasets like OffenseEval (13,200 posts), HateEval (19,000 posts), and Kaggle Toxic Comments datasets (240,000 posts). This can pose a challenge for training deep learning models, which often consists of large number of parameters, from scratch. Class wise data distribution for each language is available in the appendix .1 figures 4, 5, and 6. These figures show that the label distribution is highly skewed for task C, such as the label UNT, which is quite underrepresented. Similarly, for German the task A data is quite unbalanced. For more details on the dataset along with the details on its creation and motivation we refer the reader to Mandl et al., (2019). Mandl et al., (2019) reports that the inter-annotator agreement is in the range of 60% to 70% for English and Hindi. Furthermore, the inter-annotator agreement is more than 86% for German. Figure 2: An overview of various model architectures we used. Shaded task boxes represent that we first compute a marginal representation of labels only belonging to that task before computing the loss. ### 3.2 Fine-tuning transformer based models The transformer based models especially BERT (Devlin et al.,, 2019), have proven to be successful in achieving very good results on a range of NLP tasks. Upon its release, BERT based models became state of the art for 11 NLP tasks (Devlin et al.,, 2019). This motivated us to try out BERT for hate speech detection. We had used multiple variants of the BERT model during HASOC 2019 shared tasks Mishra and Mishra, (2019). We also experimented with other transformer models and BERT during TRAC2020 Mishra et al., 2020b . However, based on our experiments, we find the original BERT models to be best performing for most tasks. Hence, for this paper we only implement our models on those. For our experiments we use the open source implementations of BERT provided by Wolf et al., (2019)222https://github.com/huggingface/transformers. A common practice for using BERT based models, is to fine-tune an existing pre-trained model on data from a new task. For fine tuning the pre-trained BERT models we used the BERT for Sequence Classification paradigm present in the HuggingFace library. We fine tune BERT using various architectures. A visual description of these architectures is shown in Figure 2. These models are explained in detail in later sections. To process the text, we first use a pre-trained BERT tokenizer to convert the input sentences into tokens. These tokens are then passed to the model which generate a BERT specific embeddings for each token. The special part about BERT is that its decoder is supplied all the hidden states of the encoder unlike other transformer models before BERT. This helps it to capture better contextual information even for larger sequences. Each sequence of tokens is padded with a [CLS] and [SEP] token. The pre-trained BERT model generates an output vector for each of the tokens. For sequence classification tasks, the vector corresponding to the [CLS] token is used as it holds the contextual information about the complete sentence. Additional fine-tuning is done on this vector to generate the classification for specific data-sets. To keep our experiments consistent, the following hyper-parameters were kept constant for all our experiments. For training our models we used the standard hyper-parameters as mentioned in the huggingface transformers documentation. We used the Adam optimizer (with $\epsilon=1e-8$) for 5 epochs, with a training/eval batch size of 32. Maximum allowable length for each sequence was kept as $128$. We use a linearly decreasing learning rate with a starting value as $5e-5$ with a weight decay of 0.0 and a max gradient norm of $1.0$. All models were trained using Google Colab’s333https://colab.research.google.com/ GPU runtimes. This limited us to a model run-time of 12 hours with a GPU which constrained our batch size as well as number of training epochs based on the GPU allocated by Colab. We refer to models which fine tune BERT on using data set from a single language for a single task, as _Single models_ with an indicator _(S)_ , this is depicted in Figure 2 (1st row left). All other models types which we discuss later are identified by their model types and naems in Figure 2. ### 3.3 Training a model for all tasks One of the techniques that we had used for our work in HASOC 2019 Mishra and Mishra, (2019) was creating an additional sub-task D by combining the labels for all of the sub-tasks. We refer to models which use this technique as _Joint task models_ which an indicator _(D)_ (see Figure 2 models marked with D). This allowed us to train a single model for all of the sub-tasks. This also helps in overcoming the data sparsity issue for sub-tasks B and sub-tasks C for which the no. of data points is very small. The same technique was also employed in our submission to TRAC 2020 Mishra et al., 2020b aggression and misogyny identification tasks. Furthermore, when combining labels, we only consider valid combination of labels, which allows us to reduce the possible output space. For HASOC, the predicted output labels for the joint-training are as follows : NOT-NONE-NONE, HOF-HATE-TIN, HOF-HATE-UNT, HOF-OFFN-TIN, HOF- OFFN-UNT, HOF-PRFN-TIN, HOF-PRFN-UNT. The task specific labels can be easily extracted from the output labels, using post-processing of predicted labels. ### 3.4 Multi-lingual training Inspired from joint training of all tasks, as described above, we also implement the training of a single model for all languages for a given sub- task. Similar approach was utilized in our prior submission to TRAC 2020 Mishra et al., 2020b . We refer to models which use this technique as _All models_ with an indicator _(ALL)_ (see Figure 2 models marked with ALL). In this method, we combine the data-sets from all the languages and train a single multi-lingual model on this combined data-set. The multi-lingual model is able to learn data from multiple languages thus providing us with a single unified model for different languages. A major motivation for taking this approach was that social media data often does not belong to one particular language. It is quite common to find code-mixed posts on Twitter and Facebook. Thus, a multi-lingual model is the best choice in this scenario. During our TRAC 2020 work, we had found out that this approach works really well and was one of our top models in almost all of the shared tasks. From a deep learning point of view, this technique seems promising as doing this also increases the size of the data-set available for training without adding new data points from other data-sets or from data augmentation techniques. As a natural extension of the above two approaches, we combine the multi- lingual training with the joint training approach to train a single model on all tasks for all languages. We refer to models which use this technique as _All joint task models_ with an indicator _(ALL) (D)_ (see Figure 2). ### 3.5 Multi-task learning While the joint task setting, can be considered as a multi-task setting, it is not, in the common sense, and hence our reservation in calling it multi-task. The joint task training can be considered an instance of multi-class prediction, where the number of classes is based on the combination of tasks. This approach does not impose any sort of task specific structure on the model, or computes and combines task specific losses. The core idea of multi- task learning is to use similar tasks as regularizers for the model. This is done by simply adding the loss functions specific to each task in the final loss function of the model. This way the model is forced to optimize for all of the different tasks simultaneously, thus producing a model that is able to generalize on multiple tasks on the data-set. However, this may not always prove to be beneficial as it has been reported that when the tasks differ significantly the model fails to optimize on any of the tasks. This leads to significantly worse performance compared to single task approaches. However, sub-tasks in hate speech detection are often similar or overlapping in nature. Thus, this approach seems promising for hate speech detection. Our multi-task setup is inspired from the marginalized inference technique which was used in Mishra et al., 2020b . In the marginalized inference, we post-process the probabilities of each label in the joint model, and compute the task specific label probability by marginalizing the probability of across all the other tasks. This ensures that the probabilities of labels for each sub-task make a valid probability distribution and sum to one. For example, $p(\textbf{HOF-HATE-TIN})>p(\textbf{HOF-PRFN-TIN})$ does not guarantee that $p(\textbf{HOF-HATE-UNT})>p(\textbf{HOF-PRFN-UNT})$. As described above, we can calculate the task specific probabilities by marginalizing the output probabilities of that task. For example, $p(\textbf{HATE})=p(\textbf{HOF-HATE- TIN})+p(\textbf{HOF-HATE-UNT})$. However, using this technique did not lead to a significant improvement in the predictions and the evaluation performance. In some cases, it was even lower than the original method. A reason we suspect for this low performance is that the model was not trained to directly optimize its loss for this marginal inference. Next, we describe our multi- task setup inspired from this approach. For our multi-task experiments, we first use our joint training approach (sub- task D) to generate the logits for the different class labels. These logits are then marginalized to generate task specific logits (marginalizing logits is simpler than marginalizing the probability for each label, as we do not need to compute the partition function). For each task, we take a cross- entropy loss using the new task specific logits. Finally we add the respective losses for each sub-task along with the sub-task D loss. This added up loss is the final multi-task loss function of our model. We then train our model to minimize this loss function. In this loss, each sub-task loss acts as a regularizer for the other task losses. Since, we are computing the multi-task loss for each instance, we include a special label _NONE_ for sub-tasks B and C, for the cases where the label of sub-task A is _NOT_. We refer to models which use this technique as _Multi-task models_ with an indicator _(MTL) (D)_ (see Figure 2). One important point to note is that we restrict the output space of the multi- task model by using the task 4 labels. This is an essential constraint that we put on the model because of which there is no chance of any inconsistency in the prediction. By inconsistency we say that it is not possible for our multi- task model to predict a data point that belongs to _NOT_ for task A and to any label other than _NONE_ for the tasks B and C. If we follow the general procedure for training a multi-task model, we would have $2*(3+1)*(2+1)=24$ , with $+1$ for the additional _NONE_ label, combinations of outputs from our model, which would produce the inconsistencies mentioned above. Like the methods mentioned before, we extend multi-task learning to all languages, which results in _Multi-task all model_ , which are indicated with an indicator _(MTL) (ALL) (D)_. ### 3.6 Training with Back-Translated data One approach for increasing the size of the training data-set, is to generate new instances based on existing instances using data augmentation techniques. These new instances are assigned the same label as the original instance. Training model with instances generated using data augmentation techniques assumes that the label remains same if the data augmentation does not change the instance significantly. We utilized a specific data augmentation technique used in NLP models, called Back-Translation (Koehn,, 2005; Sennrich et al.,, 2016). Back-translation uses two machine translation models, one, to translate a text from its original language to a target language; and another, to translate the new text in target language back to the original language. This technique was successfully used in the submission of Risch and Krestel, (2018, 2020) during TRAC 2018 and 2020. Data augmentation via back-translation assumes that current machine translation systems when used in back-translation settings give a different text which expresses a similar meaning as the original. This assumption allows us to reuse the label of the original text for the back-translated text. We used the Google translate API444https://cloud.google.com/translate/docs to back-translate all the text in our data-sets. For each language in our data-set we use the following source $\rightarrow$ target $\rightarrow$ source pairs: * • EN: _English_ $\rightarrow$ _French_ $\rightarrow$ _English_ * • HI: _Hindi_ $\rightarrow$ _English_ $\rightarrow$ _Hindi_ * • DE: _German_ $\rightarrow$ _English_ $\rightarrow$ _German_ To keep track of the back-translated texts we added a flag to the text id. In many cases, there were minimal changes to the text. In some cases there were no changes to the back-translated texts. However the no. of such texts where there was no change after back-translation was very low. For example, among 4000 instances in the English training set around 100 instances did not have any changes. So while using the back-translated texts for our experiments, we simply used all the back-translated texts whether they under-went a change or not. The data-set size doubled after using the back-translation data augmentation technique. An example of back-translated English text is as follows (changed text is emphasized): 1. 1. Original: @politico No. We should remember very clearly that #Individual1 just admitted to treason . #TrumpIsATraitor #McCainsAHero #JohnMcCainDay 2. 2. Back-translated: @politico No, we must not forget that very clear #Individual1 just admitted to treason. #TrumpIsATraitor #McCainsAHero #JohnMcCainDay ## 4 Results We present our results for sub-tasks A, B and C in Table 2, 3, and 4 respectively. To keep the table concise we use the following convention. 1. 1. (ALL): A _bert-base-multi-lingual-uncased_ model was used with multi-lingual joint training. 2. 2. (BT): The data-set used for this experiment is augmented using back- translation. 3. 3. (D): A joint training approach has been used. 4. 4. (MTL): The experiment is performed using a multi-task learning approach. 5. 5. (S): This is the best model which was submitted to HASOC 2019 in Mishra and Mishra, (2019). The pre-trained BERT models which were fine-tuned for each language in a single language setting, are as follows: 1. 1. EN \- _bert-base-uncased_ 2. 2. HI \- _bert-base-multi-lingual-uncased_ 3. 3. DE \- _bert-base-multi-lingual-uncased_ ### 4.1 Model performance We evaluate our models against each other and also against the top performing models of HASOC 2019 for each task. We use the same benchmark scores, namely, weighted F1-score and macro F1-score, as were used in Mandl et al., (2019), with macro F1-score being the scores which were used for overall ranking in HASOC 2019. #### 4.1.1 Sub-Task A Table 2: Sub-task A results. Models in HASOC 2019 (Mandl et al.,, 2019) were ranked based on Macro F1. | | Weighted F1 | Macro F1 ---|---|---|--- lang | model | dev | train | test | dev | train | test EN | (ALL) | 0.562 | 0.949 | 0.804 | 0.568 | 0.946 | 0.753 (ALL) (D) | 0.481 | 0.894 | 0.797 | 0.497 | 0.886 | 0.740 (BT) | 0.535 | 0.973 | 0.756 | 0.545 | 0.971 | 0.690 (BT) (ALL) | 0.493 | 0.986 | 0.803 | 0.509 | 0.985 | 0.747 (BT) (ALL) (D) | 0.474 | 0.981 | 0.806 | 0.492 | 0.980 | 0.750 (MTL) (ALL) (D) | 0.552 | 0.823 | 0.801 | 0.559 | 0.812 | 0.755 (MTL) (D) | 0.543 | 0.745 | 0.819 | 0.557 | 0.725 | 0.765 | (S) | 0.606 | 0.966 | 0.790 | 0.610 | 0.964 | 0.740 | (S) (D) | 0.596 | 0.908 | 0.801 | 0.603 | 0.902 | 0.747 | HASOC Best | - | - | 0.840 | - | - | 0.788 HI | (ALL) | 0.786 | 0.976 | 0.793 | 0.785 | 0.976 | 0.793 (ALL) (D) | 0.815 | 0.959 | 0.811 | 0.815 | 0.959 | 0.810 (BT) | 0.654 | 0.967 | 0.746 | 0.654 | 0.967 | 0.744 (BT) (ALL) | 0.815 | 0.982 | 0.795 | 0.814 | 0.982 | 0.795 (BT) (ALL) (D) | 0.772 | 0.975 | 0.793 | 0.772 | 0.975 | 0.792 (MTL) (ALL) (D) | 0.860 | 0.921 | 0.808 | 0.860 | 0.921 | 0.807 (MTL) (D) | 0.748 | 0.893 | 0.814 | 0.749 | 0.893 | 0.814 | (S) | 0.742 | 0.961 | 0.802 | 0.742 | 0.961 | 0.802 | (S) (D) | 0.822 | 0.941 | 0.814 | 0.823 | 0.941 | 0.811 | HASOC Best | - | - | 0.820 | - | - | 0.815 DE | (ALL) | 0.899 | 0.993 | 0.794 | 0.706 | 0.981 | 0.584 (ALL) (D) | 0.906 | 0.988 | 0.779 | 0.730 | 0.968 | 0.566 (BT) | 0.878 | 0.988 | 0.777 | 0.628 | 0.969 | 0.533 (BT) (ALL) | 0.908 | 0.999 | 0.800 | 0.742 | 0.998 | 0.612 (BT) (ALL) (D) | 0.902 | 0.998 | 0.783 | 0.712 | 0.994 | 0.584 (MTL) (ALL) (D) | 0.917 | 0.969 | 0.786 | 0.764 | 0.915 | 0.582 (MTL) (D) | 0.878 | 0.898 | 0.789 | 0.593 | 0.683 | 0.526 | (S) | 0.606 | 0.966 | 0.789 | 0.610 | 0.964 | 0.577 | HASOC Best | - | - | 0.792 | - | - | 0.616 The best scores for sub-task A are mentioned in Table 2. The best scores for this task belong to Wang et al., (2019), Bashar and Nayak, (2020) and Saha et al., (2019) for English, Hindi and German respectively. All the models that we experimented with in sub-task A are very closely separated by the macro-F1 score. Hence, all of them give a similar performance for this task. The difference between the macro F1-scores of these models is $<3\%$ . For both English and Hindi, the multi-task learning model performed the best while for the German language, the model that was trained on the back-translated data using the multi-lingual joint training approach and task D ( (ALL) (BT) (D) ) worked best. However, it is interesting to see that the multi-task model gives competitive performance on all of the languages within the same computation budget. One thing to notice is that, the train macro-F1 scores of the multi- task model are much lower than the other models. This suggests that the (MTL) model, given additional training time might improve the results even further. We are unable to provide longer training time due to lack of computational resources available to us. The (ALL) (MTL) model also gives a similar performance compared to the (MTL) model. This suggests that the additional multi-lingual training comes with a trade off with a slightly lower macro-F1 score. However, the difference between the scores of the two models is $\sim 1\%$. In order to address the additional training time the (MTL) models required, we trained the (ALL) (MTL) model for 15 epochs. However, this training time was too large as the models over-fitted the data. This finally resulted in a degradation in the performance of these models. A sweet spot for the training time may be found for the (MTL) models which may result in an increase in the performance of the model whilst avoiding over-fitting. We were not able to conduct more experiments to do the same due to time constraints. This may be evaluated in the additional future work on these models. We, however, cannot compare the German (MTL) models with the (MTL) models of the other languages as the German data did not have not have sub-task C, so the (MTL) approach did not have sub-task C for German. As we will see in the next section, the (MTL) models performed equally well in sub-task B. This might be because both tasks A and B involve identifying hate and hence are in a sense co-related. This co-relation is something that the (MTL) models can utilize for their advantage. It has been found in other multi-task approaches that the models learn more effectively when the different tasks are co-related. However, their performance can degrade if the tasks are un-related. The lower performance on the German data may be because of the unavailability of the sub-task C. However, the results are still competitive with the other models. For German, the (ALL) (MTL) model performed better than our submission for HASOC 2019. The (MTL) model for Hindi was able to match the best model for this task at HASOC 2019. The (ALL) and (ALL) (D) training methods show an improvement from our single models submitted at HASOC. These models present us with an interesting option for abuse detection tasks as they are able to work on all of the shared tasks at the same time, leveraging the multi-lingual abilities of the model whilst still having a computation budget equivalent to that of a single model. These results show that these models give a competitive performance with the single models. They even outperform the single model, e.g., they outperform the bert- base-uncased single models that were used in English sub-task A, which have been specially tuned for English tasks. While for German and Hindi, the single models themselves utilized a bert-base-uncased model, so they are better suited for analyzing the improvements by the multi-lingual joint training approach. On these languages we see, that the (ALL) and (ALL) (D) techniques do improve the macro-F1 scores on for this task. The back-translation technique does not seem to improve the models much. The approach had a mixed performance. For all the languages, back-translation alone does not improve the model and hints at over-fitting, resulting in a decrease in test results. However, when it is combined with the (ALL) and (D) training methods we see an increase in the performance. The (ALL) and (D) training methods are able to leverage the data-augmentation applied in the back-translated data. Back-translation when used with (ALL) or (ALL) (D) are better than the single models that we submitted at HASOC 2019. The (BT) (ALL) model comes really close to the best model at HASOC, coming second according to the results in Mandl et al., (2019). #### 4.1.2 Sub-Task B Table 3: sub-task B results. Models in HASOC 2019 (Mandl et al.,, 2019) were ranked based on Macro F1. | | Weighted F1 | Macro F1 ---|---|---|--- lang | model | dev | train | test | dev | train | test EN | (ALL) | 0.361 | 0.826 | 0.501 | 0.290 | 0.805 | 0.467 (ALL) (D) | 0.201 | 0.776 | 0.556 | 0.190 | 0.580 | 0.392 (BT) | 0.422 | 0.965 | 0.532 | 0.352 | 0.960 | 0.510 (BT) (ALL) | 0.396 | 0.962 | 0.626 | 0.375 | 0.957 | 0.591 (BT) (ALL) (D) | 0.201 | 0.950 | 0.576 | 0.153 | 0.708 | 0.408 (MTL) (ALL) (D) | 0.397 | 0.927 | 0.635 | 0.277 | 0.915 | 0.590 (MTL) (D) | 0.344 | 0.899 | 0.638 | 0.341 | 0.881 | 0.600 | (S) | 0.349 | 0.867 | 0.728 | 0.314 | 0.846 | 0.545 | (S) (D) | 0.401 | 0.875 | 0.698 | 0.332 | 0.839 | 0.537 | HASOC Best | - | - | 0.728 | - | - | 0.545 HI | (ALL) | 0.494 | 0.832 | 0.500 | 0.340 | 0.802 | 0.494 (ALL) (D) | 0.678 | 0.792 | 0.564 | 0.293 | 0.566 | 0.415 (BT) | 0.231 | 0.807 | 0.507 | 0.160 | 0.767 | 0.501 (BT) (ALL) | 0.589 | 0.890 | 0.667 | 0.283 | 0.875 | 0.662 (BT) (ALL) (D) | 0.630 | 0.849 | 0.519 | 0.180 | 0.617 | 0.381 (MTL) (ALL) (D) | 0.819 | 0.883 | 0.647 | 0.499 | 0.864 | 0.641 (MTL) (D) | 0.553 | 0.802 | 0.602 | 0.348 | 0.764 | 0.593 | (S) | 0.466 | 0.749 | 0.688 | 0.322 | 0.701 | 0.553 | (S) (D) | 0.757 | 0.826 | 0.715 | 0.459 | 0.736 | 0.581 | HASOC Best | - | - | 0.715 | - | - | 0.581 DE | (ALL) | 0.326 | 0.876 | 0.459 | 0.315 | 0.861 | 0.343 (ALL) (D) | 0.285 | 0.813 | 0.154 | 0.255 | 0.603 | 0.128 (BT) | 0.285 | 0.620 | 0.413 | 0.328 | 0.581 | 0.285 (BT) (ALL) | 0.478 | 0.985 | 0.495 | 0.484 | 0.984 | 0.397 (BT) (ALL) (D) | 0.179 | 0.945 | 0.242 | 0.153 | 0.707 | 0.177 (MTL) (ALL) (D) | 0.463 | 0.946 | 0.527 | 0.346 | 0.707 | 0.344 (MTL) (D) | 0.468 | 0.923 | 0.541 | 0.482 | 0.918 | 0.416 | (S) | 0.112 | 0.367 | 0.756 | 0.140 | 0.247 | 0.249 | (S) (D) | 0.865 | 0.918 | 0.778 | 0.282 | 0.409 | 0.276 | HASOC Best | - | - | 0.775 | - | - | 0.347 The best scores for sub-task B are mentioned in Table 3. The best scores for this task belong to Ruiter et al., (2019) for German. For English and Hindi sub-task B our submissions had performed the best at HASOC 2019. For sub-task B, many of our models were able to significantly outperform the best HASOC models. For English, the multi-task approach results in a new best macro-F1 score of $0.600$, a $6\%$ increase from the previous best. For Hindi, our (BT) (ALL) results in a macro-F1 score of $0.662$ which is $8\%$ more than the previous best. For Germans, our (MTL) model has a macro-F1 score on the test set of 0.416 which is almost $7\%$ more than the previous best. For the English task, even our (MTL) (ALL) and (BT) (ALL) models were able to beat the previous best. However, our results show that unlike sub-task A where our models had similar performances, in sub-task B there is huge variation in their performance. Many outperform the best, however some of them also show poor results. The (ALL) and (ALL) (D) perform poorly for the three languages, except (ALL) in German, and show very small macro-F1 scores even on the training set. Thus, training these models for longer may change the results. The (MTL) models are able to give competitive performance in task-A and is able to outperform the previous best, thus showing it’s capability to leverage different co-related tasks and generalize well on all of them. Here again we see that back-translation alone does not improve the macro-F1 scores. However, an interesting thing to notice is that the (ALL) and (BT) models which perform poorly individually, tend to give good results when used together. Outperforming the previous best HASOC models, in all the three languages. This hints that data sparsity alone is not the major issue of this task. This is also evident from the performance of the (MTL) model which only utilizes the data-set of a single language, which is significantly smaller than the back-translated (twice the original data-set) and the multi-lingual joint model (sum of the sizes of the original model). But the (BT) (ALL) (D) model performed poorly in all of the three languages. Thus, the use of sub- task (D) with these models only degrades performance. The results from this task confirm that the information required to predict task - A is important for task - B as well. This information is shared better by the modified loss function of the (MTL) models rather than the loss function for sub-task (D). The (MTL) models build up on the sub-task (D) approach and do not utilize it explicitly. The sub-task (D) approach seems like a multi-task learning method, however, it is not complete and is not able to learn from the other tasks, thus does not offer huge improvement. These (MTL) models do show a variation in their performance but it is always on the higher side of the macro-F1 scores. #### 4.1.3 Sub-Task C The best scores for sub-task C are mentioned in Table 4. The best scores for this task belong to Mujadia et al., (2019) for Hindi while our submissions performed the best for English. The results for sub-task C also show appreciable variation. Except the (ALL) (D) and (BT) (ALL) (D) models which also performed poorly in sub-task B, the variation in their performance, especially for English, is not as significant as that present in sub-task B. This may be due to the fact that the two way fine-grained classification is a much easier task than the three way classification in sub-task B. One important point to note is that sub-task C focused on identifying the context of the hate speech, specifically it focused on finding out whether it is targeted or un-targeted, while sub-task A and sub-task B both focused on identifying the type of hate speech. The (MTL) models do not perform as well as they did in the previous two tasks. They were still able to outperform the best models for English but perform poorly for Hindi. An important point to notice here is that the train macro-F1 scores for the (MTL) models is significantly low. This suggests that the (MTL) model was not able learn well even for the training instances of this task. This can be attributed to the point mentioned above that this task is inherently not as co-related to sub-task A and sub-task B as previously assumed. The task structure itself is not beneficial for a (MTL) approach. The main reason for this is that this task focuses on identifying targeted and un- targeted hate-speech. However, a non hate-speech text can also be an un- targeted or targeted text. As the (MTL) model receives texts belonging to both hate (HOF) and non-hate speech NOT, the information contained in the texts belonging to this task are counter acted by those targeted and un-targeted texts belong to the (NOT) class. Thus, a better description of this task is not a fine-grain classification of hate speech text, but that which involves targeted and un-targeted labels for both (HOF) and (NOT) classes. In that setting, we can fully utilize the advantage of the multi-task learning model and can expect better performance on this task as well. The (ALL) and (BT) models performed really well in sub-task C. The (ALL), (BT) and (ALL) (BT) models outperform the previous best for English. The combination of these models with (D) still does not improve them and they continue to give poor performance. This provides more evidence to our previous inference that sub-task (D) alone does not improve the our performance. Table 4: sub-task C results. Models in HASOC 2019 (Mandl et al.,, 2019) were ranked based on Macro F1. | | Weighted F1 | Macro F1 ---|---|---|--- lang | model | dev | train | test | dev | train | test EN | (ALL) | 0.842 | 0.986 | 0.737 | 0.524 | 0.958 | 0.547 (ALL) (D) | 0.380 | 0.792 | 0.658 | 0.141 | 0.328 | 0.292 (BT) | 0.836 | 0.991 | 0.771 | 0.465 | 0.974 | 0.543 (BT) (ALL) | 0.839 | 0.987 | 0.718 | 0.534 | 0.962 | 0.514 (BT) (ALL) (D) | 0.374 | 0.967 | 0.600 | 0.173 | 0.618 | 0.297 (MTL) (ALL) (D) | 0.839 | 0.704 | 0.658 | 0.311 | 0.465 | 0.518 (MTL) (D) | 0.844 | 0.747 | 0.692 | 0.470 | 0.506 | 0.538 | (S) | 0.880 | 0.980 | 0.756 | 0.627 | 0.942 | 0.511 | (S) (D) | 0.548 | 0.874 | 0.764 | 0.393 | 0.651 | 0.476 | HASOC Best | - | - | 0.756 | - | - | 0.511 HI | (ALL) | 0.844 | 0.765 | 0.827 | 0.525 | 0.740 | 0.557 (ALL) (D) | 0.594 | 0.666 | 0.740 | 0.216 | 0.417 | 0.336 (BT) | 0.861 | 0.766 | 0.817 | 0.652 | 0.744 | 0.568 (BT) (ALL) | 0.797 | 0.941 | 0.775 | 0.517 | 0.937 | 0.527 (BT) (ALL) (D) | 0.682 | 0.779 | 0.673 | 0.288 | 0.507 | 0.317 (MTL) (ALL) (D) | 0.374 | 0.530 | 0.626 | 0.292 | 0.524 | 0.456 (MTL) (D) | 0.557 | 0.577 | 0.628 | 0.397 | 0.573 | 0.451 | (S) | 0.800 | 0.877 | 0.727 | 0.550 | 0.866 | 0.565 | (S) (D) | 0.769 | 0.724 | 0.758 | 0.537 | 0.622 | 0.550 | HASOC Best | - | - | 0.736 | - | - | 0.575 Overall most of our models, show an improvement from the single models submitted at HASOC.The sub-par performance of the back-translated models across the sub-tasks suggest that data sparsity is not the central issue of this challenge. To take advantage of the augmented-data additional methods have to be used. The sub-task (D) does not significantly adds as an improvement to the models. It can be seen that it actually worsens the situation for sub-tasks B and sub-tasks C. This can be attributed to it changing the task to a much harder 7-class distribution task.The combined model approaches that we have mentioned above offer a resource efficient way for hate speech detection. The (ALL), (ALL) (MTL) and (MTL) models are able to generalize well for the different tasks and different languages. They present themselves as good candidates for a unified model for hate speech detection. ### 4.2 Error analysis (a) sub-task A (b) sub-task B (c) sub-task C Figure 3: Variation in label F1-scores for all sub-tasks across all models After identifying the best model and the variation in evaluation scores for each model, we investigate the overall performance of these models for each label belonging to each task. In Figure 3, we can observe how the various labels have a high variance in their predictive performance. #### 4.2.1 Sub-Task A For English, the models show decent variation for both the labels on the training set. However, this variation is not as significant for the dev and test sets for the (NOT) label. There is an appreciable variation for (HOF) label on the dev set, but it is not transferred to the test set. The scores for the train sets is very high compared to the dev and test sets. For Hindi, the predictions for both the labels show minimum variation in the F1-score across the three data-sets, with similar scores for each label. For German, the F1-scores for the (NOT) class is quite high compared to that of (HOF) class. The models, have a very low F1-score for the (HOF) label on the test set with appreciable variation across the different models. #### 4.2.2 Sub-Task B For English, the F1-scores for all the labels are quite high on the train set with decent variation for the (OFFN) label. All of the labels show appreciable variance on the dev and test sets. The (OFFN) has the lowest F1-score amongst the labels on both the dev and test sets with the other two labels having similar scores for the test set. For Hindi, the train F1-scores are similar for all of the labels. The F1-scores are on the lower end for the (HATE) and (OFFN) labels on the dev set with appreciable variance across the models. This may be due to the fact that the Hindi dev set contains very few samples from these two labels. For German, the variation among the F1-scores is high across all the three sets. The (HATE) label and the (OFFN) label have a large variation in their F1-scores across the models on the dev and test set respectively. The F1-score for the (OFFN) label is much higher than the other labels on the test set. #### 4.2.3 Sub-Task C For English, the (UNT) label has exceptionally high variance across the models for the train set. This is due to the exceptionally low scores by the (BT) (ALL) (D) model. This label has extremely low F1-score on the dev set. Furthermore, there is also large variation in the (TIN) scores across the models in all the sets. For Hindi, the (TIN) label has similar F1-scores with large variations across the models on all of the three sets. However, the (UNT) label has small variance across the models on the dev and test sets. #### 4.2.4 Back Translation We also looked at the issue with back-translated results. In order to assess the back-translated data we looked at the new words added and removed from a sentence after back-translation. Aggregating these words over all sentences we find that the top words which are often removed and introduced are stop words, e.g. the, of, etc. In order to remove these stop words from our analysis and assess the salient words introduced and removed per label we remove the overall top 50 words from the introduced and removed list aggregated over each label. This highlights words which are often removed from offensive and hateful labels are indeed offensive words. A detailed list of words for English and German can be found in appendix .2 (we excluded results for Hindi because of Latex encoding issues). ## 5 Discussion ### 5.1 Computational benefits From the results mentioned above, we can easily conclude that multi-task models present us with robust models for hate speech detection that can generalize well across different languages and different tasks for a given domain. Even the combined models, present us with models that can be deployed easily and give competitive performance on different languages with an efficient computation budget. Many of our models, perform better than the best scoring models of HASOC 2019 while maintaing a low inference cost. ### 5.2 Additional evaluation Our current evaluation was limited to the HASOC dataset, additional evaluation needs to be done to assess out of domain and out of language capabilities of these techniques. Furthermore, the back-translation approach needs to be assessed even further using qualitative analysis of generated back- translations. ### 5.3 Architectures and Training improvements There are additional combination of architectures which we plan to try in future iterations of this work. Some of the combinations which we have not considered in this work are the (BT) (MTL) models and the (BT) (MTL) (ALL) models. We have seen that the (ALL) and (BT) models work well in unison and the (MTL) (ALL) models also give competitive performance with the (MTL) model. Therefore, a (BT) (MTL) (ALL) model is expected to bring out the best of both worlds. The (MTL) models we have used can still be tuned further, which may increase their results on the test sets. We trained (ALL) (MTL) model for 15 epochs instead of the usual 5, but it over-fitted the training set. Further experiments have to be conducted to identify the ideal training time for these models. ### 5.4 Real world usage Even though our models have performed really well on the HASOC dataset, yet the results are far from ideal. Given that HASOC dataset is quite small, our models may not generalize well outside of the domain of HASOC, however, our focus was on assessing the improvements we get using our multi-task and multi- lingual techniques on this datasets. We also conducted similar experiments in our work for the TRAC 2020 dataset Mishra et al., 2020b . In order to make these model more robust for general purpose hate-speech detection we need to train it on more-diverse and larger dataset. Furthermore, we also would like to highlight that our models need to be further evaluated for demographic bias as it has been found in Davidson et al., (2019) that hate speech and abusive language datasets exhibit racial bias towards African American English usage. ## 6 Conclusion We would like to conclude this paper by highlighting the promise shown by multi-lingual and multi-task models on solving the hate and abusive speech detection in a computationally efficient way while maintaining comparable accuracy of single task models. We do highlight that our pre-trained models need to be further evaluated before being used on large scale, however the architecture and the training framework is something which can easily scale to large dataset without sacrificing performance as was shown in Mishra, 2020b ; Mishra, 2020a ; Mishra, (2019). ## Compliance with Ethical Standards Conflict of Interest: The authors declare that they have no conflict of interest. ## References * Badjatiya et al., (2017) Badjatiya, P., Gupta, S., Gupta, M., and Varma, V. (2017). Deep learning for hate speech detection in tweets. In Proceedings of the 26th International Conference on World Wide Web Companion, WWW ’17 Companion, page 759–760, Republic and Canton of Geneva, CHE. International World Wide Web Conferences Steering Committee. * Bashar and Nayak, (2020) Bashar, M. A. and Nayak, R. (2020). Qutnocturnal@ hasoc’19: Cnn for hate speech and offensive content identification in hindi language. arXiv preprint arXiv:2008.12448. * Basile et al., (2019) Basile, V., Bosco, C., Fersini, E., Nozza, D., Patti, V., Rangel Pardo, F. M., Rosso, P., and Sanguinetti, M. (2019). SemEval-2019 task 5: Multilingual detection of hate speech against immigrants and women in twitter. In Proceedings of the 13th International Workshop on Semantic Evaluation, pages 54–63, Minneapolis, Minnesota, USA. Association for Computational Linguistics. * Burnap and Williams, (2015) Burnap, P. and Williams, M. L. (2015). Cyber hate speech on twitter: An application of machine classification and statistical modeling for policy and decision making. Policy & Internet, 7(2):223–242. * Davidson et al., (2019) Davidson, T., Bhattacharya, D., and Weber, I. (2019). Racial Bias in Hate Speech and Abusive Language Detection Datasets. In Proceedings of the Third Workshop on Abusive Language Online, pages 25–35, Stroudsburg, PA, USA. Association for Computational Linguistics. * Davidson et al., (2017) Davidson, T., Warmsley, D., Macy, M. W., and Weber, I. (2017). Automated hate speech detection and the problem of offensive language. In Proceedings of the International AAAI Conference on Web and Social Media 2017. * Devlin et al., (2019) Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. (2019). BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. * Duggan et al., (2017) Duggan, M., Smith, A., and Caiazza, T. (2017). Online Harassment 2017. Technical report, Pew Research Center. * Eisenstein, (2013) Eisenstein, J. (2013). What to do about bad language on the internet. In Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 359–369, Atlanta, Georgia. Association for Computational Linguistics. * Florio et al., (2020) Florio, K., Basile, V., Polignano, M., Basile, P., and Patti, V. (2020). Time of your hate: The challenge of time in hate speech detection on social media. Applied Sciences, 10(12):4180. * Gomez et al., (2020) Gomez, R., Gibert, J., Gomez, L., and Karatzas, D. (2020). Exploring hate speech detection in multimodal publications. In 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1459–1467. * Joulin et al., (2017) Joulin, A., Grave, E., Bojanowski, P., and Mikolov, T. (2017). Bag of tricks for efficient text classification. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 427–431, Valencia, Spain. Association for Computational Linguistics. * Kiela et al., (2020) Kiela, D., Firooz, H., Mohan, A., Goswami, V., Singh, A., Ringshia, P., and Testuggine, D. (2020). The hateful memes challenge: Detecting hate speech in multimodal memes. * Koehn, (2005) Koehn, P. (2005). Europarl : A Parallel Corpus for Statistical Machine Translation. MT Summit. * Kumar et al., (2018) Kumar, R., Ojha, A. K., Malmasi, S., and Zampieri, M. (2018). Benchmarking aggression identification in social media. In Proceedings of the First Workshop on Trolling, Aggression and Cyberbullying (TRAC-2018), pages 1–11, Santa Fe, New Mexico, USA. Association for Computational Linguistics. * Kumar et al., (2020) Kumar, R., Ojha, A. K., Malmasi, S., and Zampieri, M. (2020). Evaluating aggression identification in social media. In Kumar, R., Ojha, A. K., Lahiri, B., Zampieri, M., Malmasi, S., Murdock, V., and Kadar, D., editors, Proceedings of the Second Workshop on Trolling, Aggression and Cyberbullying (TRAC-2020), Paris, France. European Language Resources Association (ELRA). * Le and Mikolov, (2014) Le, Q. V. and Mikolov, T. (2014). Distributed representations of sentences and documents. CoRR, abs/1405.4053. * Liu et al., (2016) Liu, P., Qiu, X., and Huang, X. (2016). Deep Multi-Task Learning with Shared Memory for Text Classification. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 118–127, Stroudsburg, PA, USA. Association for Computational Linguistics. * Mandl et al., (2019) Mandl, T., Modha, S., Majumder, P., Patel, D., Dave, M., Mandlia, C., and Patel, A. (2019). Overview of the hasoc track at fire 2019: Hate speech and offensive content identification in indo-european languages. In Proceedings of the 11th Forum for Information Retrieval Evaluation, FIRE ’19, page 14–17, New York, NY, USA. Association for Computing Machinery. * Mishra, (2019) Mishra, S. (2019). Multi-dataset-multi-task Neural Sequence Tagging for Information Extraction from Tweets. In Proceedings of the 30th ACM Conference on Hypertext and Social Media - HT ’19, pages 283–284, New York, New York, USA. ACM Press. * (21) Mishra, S. (2020a). Information Extraction from Digital Social Trace Data with Applications to Social Media and Scholarly Communication Data. ACM SIGIR Forum, 54(1). * (22) Mishra, S. (2020b). Information Extraction from Digital Social Trace Data with Applications to Social Media and Scholarly Communication Data. PhD thesis, University of Illinois at Urbana-Champaign. * (23) Mishra, S. (2020c). Non-neural Structured Prediction for Event Detection from News in Indian Languages. In Mehta, P., Mandl, T., Majumder, P., and Mitra, M., editors, Working Notes of FIRE 2020 - Forum for Information Retrieval Evaluation, Hyderabad, India. CEUR Workshop Proceedings, CEUR-WS.org. * Mishra et al., (2014) Mishra, S., Agarwal, S., Guo, J., Phelps, K., Picco, J., and Diesner, J. (2014). Enthusiasm and support: alternative sentiment classification for social movements on social media. In Proceedings of the 2014 ACM conference on Web science - WebSci ’14, pages 261–262, Bloomington, Indiana, USA. ACM Press. * Mishra and Diesner, (2016) Mishra, S. and Diesner, J. (2016). Semi-supervised Named Entity Recognition in noisy-text. In Proceedings of the 2nd Workshop on Noisy User-generated Text (WNUT), pages 203–212, Osaka, Japan. The COLING 2016 Organizing Committee. * Mishra and Diesner, (2019) Mishra, S. and Diesner, J. (2019). Capturing Signals of Enthusiasm and Support Towards Social Issues from Twitter. In Proceedings of the 5th International Workshop on Social Media World Sensors - SIdEWayS’19, pages 19–24, New York, New York, USA. ACM Press. * Mishra and Mishra, (2019) Mishra, S. and Mishra, S. (2019). 3Idiots at HASOC 2019: Fine-tuning Transformer Neural Networks for Hate Speech Identification in Indo-European Languages. In Proceedings of the 11th annual meeting of the Forum for Information Retrieval Evaluation, pages 208–213, Kolkata, India. * (28) Mishra, S., Prasad, S., and Mishra, S. (2020a). Model and predictions for multi-task multi-lingual learning of transformer models for hate speech and offensive speech identification in social media. Accessible at: https://doi.org/10.13012/B2IDB-3565123_V1. * (29) Mishra, S., Prasad, S., and Mishra, S. (2020b). Multilingual Joint Fine-tuning of Transformer models for identifying Trolling,Aggression and Cyberbullying at TRAC 2020. In Proceedings of the Second Workshop on Trolling, Aggression and Cyberbullying (TRAC-2020). * Mondal et al., (2017) Mondal, M., Silva, L. A., and Benevenuto, F. (2017). A Measurement Study of Hate Speech in Social Media. In Proceedings of the 28th ACM Conference on Hypertext and Social Media - HT ’17, pages 85–94, New York, New York, USA. ACM Press. * Mozafari et al., (2020) Mozafari, M., Farahbakhsh, R., and Crespi, N. (2020). A bert-based transfer learning approach for hate speech detection in online social media. In Cherifi, H., Gaito, S., Mendes, J. F., Moro, E., and Rocha, L. M., editors, Complex Networks and Their Applications VIII, pages 928–940, Cham. Springer International Publishing. * Mujadia et al., (2019) Mujadia, V., Mishra, P., and Sharma, D. M. (2019). Iiit-hyderabad at hasoc 2019: Hate speech detection. * Perrin, (2015) Perrin, A. (2015). Social Media Usage:2005-2015. Technical report, Pew Research Center. * Plank, (2017) Plank, B. (2017). All-in-1 at IJCNLP-2017 task 4: Short text classification with one model for all languages. In Proceedings of the IJCNLP 2017, Shared Tasks, pages 143–148, Taipei, Taiwan. Asian Federation of Natural Language Processing. * Ranasinghe and Zampieri, (2020) Ranasinghe, T. and Zampieri, M. (2020). Multilingual offensive language identification with cross-lingual embeddings. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5838–5844, Online. Association for Computational Linguistics. * Razavi et al., (2010) Razavi, A. H., Inkpen, D., Uritsky, S., and Matwin, S. (2010). Offensive language detection using multi-level classification. In Proceedings of the 23rd Canadian Conference on Advances in Artificial Intelligence, AI’10, page 16–27, Berlin, Heidelberg. Springer-Verlag. * Risch and Krestel, (2018) Risch, J. and Krestel, R. (2018). Aggression identification using deep learning and data augmentation. In Proceedings of the First Workshop on Trolling, Aggression and Cyberbullying (co-located with COLING), pages 150–158. * Risch and Krestel, (2020) Risch, J. and Krestel, R. (2020). Bagging bert models for robust aggression identification. In Proceedings of the Workshop on Trolling, Aggression and Cyberbullying (TRAC@LREC). * Ruiter et al., (2019) Ruiter, D., Rahman, M. A., and Klakow, D. (2019). Lsv-uds at HASOC 2019: The problem of defining hate. In Mehta, P., Rosso, P., Majumder, P., and Mitra, M., editors, Working Notes of FIRE 2019 - Forum for Information Retrieval Evaluation, Kolkata, India, December 12-15, 2019, volume 2517 of CEUR Workshop Proceedings, pages 263–270. CEUR-WS.org. * Saha et al., (2019) Saha, P., Mathew, B., Goyal, P., and Mukherjee, A. (2019). Hatemonitors: Language agnostic abuse detection in social media. * Salminen et al., (2018) Salminen, J., Almerekhi, H., Milenković, M., gyo Jung, S., An, J., Kwak, H., and Jansen, B. (2018). Anatomy of online hate: Developing a taxonomy and machine learning models for identifying and classifying hate in online news media. * Schmidt and Wiegand, (2017) Schmidt, A. and Wiegand, M. (2017). A Survey on Hate Speech Detection using Natural Language Processing. In Proceedings of the Fifth International Workshop on Natural Language Processing for Social Media, pages 1–10, Stroudsburg, PA, USA. Association for Computational Linguistics. * Sennrich et al., (2016) Sennrich, R., Haddow, B., and Birch, A. (2016). Improving Neural Machine Translation Models with Monolingual Data. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 86–96, Stroudsburg, PA, USA. Association for Computational Linguistics. * Søgaard and Goldberg, (2016) Søgaard, A. and Goldberg, Y. (2016). Deep multi-task learning with low level tasks supervised at lower layers. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 231–235. Association for Computational Linguistics. * Sticca and Perren, (2013) Sticca, F. and Perren, S. (2013). Is Cyberbullying Worse than Traditional Bullying? Examining the Differential Roles of Medium, Publicity, and Anonymity for the Perceived Severity of Bullying. Journal of Youth and Adolescence, 42(5):739–750. * Struß et al., (2019) Struß, J., Siegel, M., Ruppenhofer, J., Wiegand, M., and Klenner, M. (2019). Overview of germeval task 2, 2019 shared task on the identification of offensive language. In KONVENS. * Van Hee et al., (2015) Van Hee, C., Lefever, E., Verhoeven, B., Mennes, J., Desmet, B., De Pauw, G., Daelemans, W., and Hoste, V. (2015). Detection and fine-grained classification of cyberbullying events. In Angelova, G., Bontcheva, K., and Mitkov, R., editors, Proceedings of Recent Advances in Natural Language Processing, Proceedings, pages 672–680. * Vaswani et al., (2017) Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., and Polosukhin, I. (2017). Attention is all you need. In Advances in neural information processing systems, pages 5998–6008. * Vidgen et al., (2019) Vidgen, B., Harris, A., Nguyen, D., Tromble, R., Hale, S., and Margetts, H. (2019). Challenges and frontiers in abusive content detection. In Proceedings of the Third Workshop on Abusive Language Online, pages 80–93, Florence, Italy. Association for Computational Linguistics. * Wang et al., (2019) Wang, B., Ding, Y., Liu, S., and Zhou, X. (2019). Ynu_wb at hasoc 2019: Ordered neurons lstm with attention for identifying hate speech and offensive language. * Wang, (2018) Wang, C. (2018). Interpreting neural network hate speech classifiers. In Proceedings of the 2nd Workshop on Abusive Language Online (ALW2), pages 86–92, Brussels, Belgium. Association for Computational Linguistics. * Waseem et al., (2017) Waseem, Z., Davidson, T., Warmsley, D., and Weber, I. (2017). Understanding abuse: A typology of abusive language detection subtasks. In Proceedings of the First Workshop on Abusive Language Online, pages 78–84, Vancouver, BC, Canada. Association for Computational Linguistics. * Wolf et al., (2019) Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., Cistac, P., Rault, T., Louf, R., Funtowicz, M., and Brew, J. (2019). Huggingface’s transformers: State-of-the-art natural language processing. * Yang et al., (2019) Yang, F., Peng, X., Ghosh, G., Shilon, R., Ma, H., Moore, E., and Predovic, G. (2019). Exploring deep multimodal fusion of text and photo for hate speech classification. In Proceedings of the Third Workshop on Abusive Language Online, pages 11–18, Florence, Italy. Association for Computational Linguistics. * Zampieri et al., (2019) Zampieri, M., Malmasi, S., Nakov, P., Rosenthal, S., Farra, N., and Kumar, R. (2019). SemEval-2019 task 6: Identifying and categorizing offensive language in social media (OffensEval). In Proceedings of the 13th International Workshop on Semantic Evaluation, pages 75–86, Minneapolis, Minnesota, USA. Association for Computational Linguistics. ## Appendix ### .1 Label Distribution Figure 4: English Data class wise distribution Figure 5: German Data class wise distribution Figure 6: Hindi Data class wise distribution ### .2 Back translation top changed words Here we list the top 5 words per label for each task obtained after removing the top 50 words which were either introduced or removed via back translation. We do not list the top words for Hindi because of the encoding issue in Latex. Listing 1: Changed words in English Training Data ⬇ 1task_1 introduced_words 2HOF [(’asset’, 30), (”you’re”, 29), (’so’, 28), (”it’s”, 26), (’there’, 25)] 3NOT [(’worldcup2019’, 47), (’at’, 41), (”i’m”, 40), (”it’s”, 38), (’us’, 38)] 4task_1 removed_words 5HOF [(’fuck’, 52), (’he’s’, 43), (’what’, 38), (’don’t’, 37), (’them’, 36)] 6NOT [(’happy’, 49), (’than’, 47), (’being’, 45), (’every’, 45), (’been’, 43)] 7task_2 introduced_words 8HATE [(’there’, 17), (’worldcup2019’, 16), (’so’, 14), (’because’, 14), (’dhoni’, 13)] 9NONE [(’worldcup2019’, 47), (’at’, 41), (”i’m”, 40), (”it’s”, 38), (’us’, 38)] 10OFFN [(’impeach45’, 11), (’asset’, 11), (’now’, 8), (’lie’, 6), (’trump2020’, 6)] 11PRFN [(”it’s”, 16), (’fucking’, 16), (’damn’, 14), (’f’, 14), (”you’re”, 12)] 12task_2 removed_words 13HATE [(’which’, 20), (’ground’, 20), (’such’, 18), (’its’, 18), (”doesn’t”, 18)] 14NONE [(’happy’, 49), (’than’, 47), (’being’, 45), (’every’, 45), (’been’, 43)] 15OFFN [(’he’s’, 12), (’them’, 11), (”he’s”, 9), (’what’, 9), (’been’, 9)] 16PRFN [(’fuck’, 47), (’fucking’, 21), (’he’s’, 16), (’off’, 16), (’don’t’, 15)] 17task_3 introduced_words 18NONE [(’worldcup2019’, 47), (’at’, 41), (”i’m”, 40), (”it’s”, 38), (’us’, 38)] 19TIN [(’asset’, 30), (”you’re”, 27), (’so’, 25), (’which’, 25), (’because’, 25)] 20UNT [(’f’, 8), (’nmy’, 5), (’***’, 5), (’there’, 4), (’these’, 4)] 21task_3 removed_words 22NONE [(’happy’, 49), (’than’, 47), (’being’, 45), (’every’, 45), (’been’, 43)] 23TIN [(’fuck’, 48), (’he’s’, 39), (’what’, 35), (’don’t’, 34), (”he’s”, 34)] 24UNT [(’them’, 7), (’f***’, 6), (’being’, 5), (’such’, 5), (’does’, 5)] Listing 2: Changed words in German Training Data ⬇ 1task_1 introduced_words 2HOF [(’!!’, 15), (’etwas’, 15), (’diese’, 12), (’sein’, 11), (’werden’, 11)] 3NOT [(’einen’, 59), (’jetzt’, 58), (’war’, 57), (’menschen’, 57), (’was’, 56)] 4task_1 removed_words 5HOF [(’du’, 18), (’wohl’, 14), (’haben’, 12), (’eure’, 11), (’mir’, 11)] 6NOT [(’wieder’, 56), (’uber’, 55), (’vom’, 52), (’haben’, 51), (’einem’, 49)] 7task_2 introduced_words 8HATE [(’diese’, 6), (’werden’, 5), (’grun’, 4), (’sollte’, 4), (’konnen’, 3)] 9NONE [(’einen’, 59), (’jetzt’, 58), (’war’, 57), (’menschen’, 57), (’was’, 56)] 10OFFN [(’!!’, 11), (’sein’, 8), (’dumm’, 8), (’etwas’, 8), (’sein,’, 7)] 11PRFN [(’ich’, 6), (’scheibe’, 5), (’etwas’, 5), (’keine’, 5), (’alle’, 4)] 12task_2 removed_words 13HATE [(’diesen’, 5), (’dass’, 5), (’kann’, 4), (’wohl’, 4), (’also’, 4)] 14NONE [(’wieder’, 56), (’uber’, 55), (’vom’, 52), (’haben’, 51), (’einem’, 49)] 15OFFN [(’du’, 12), (’nur’, 8), (’muss’, 8), (’eure’, 8), (’haben’, 7)] 16PRFN [(’bin’, 5), (’was’, 5), (’wohl’, 4), (’keine’, 4), (’fressen’, 4)]
[a]Wen-Chen Chang # QCD effects in lepton angular distributions of Drell-Yan/$Z$ production and jet discrimination Randall Evan McClellan Jen-Chieh Peng Oleg Teryaev ###### Abstract We present a comparison of data of lepton angular distributions of Drell- Yan/$Z$ production with the fixed-order pQCD calculations by which the baseline of pQCD effects is illustrated. As for the $Z$ production, we predict that $A_{0}$ and $A_{2}$ for $Z$ plus single gluon-jet events are very different from that of $Z$ plus single quark-jet events, allowing a new experimental tool for checking various algorithms which attempt to discriminate quark jets from gluon jets. Using an intuitive geometric approach, we show that the violation of the Lam-Tung relation, appearing at large transverse-momentum region, is attributed to the presence of a non- coplanarity effect. This interpretation is consistent with the appearance of violation beyond LO-QCD effect. ## 1 Introduction Measuring lepton angular distributions of Drell-Yan (D-Y) process [1] provides a powerful tool to explore the reaction mechanisms and the parton distributions of colliding hadrons. For example, the Lam-Tung (L-T) relation [2] has been proposed as a benchmark of the perturbative QCD (pQCD) effect in D-Y process. Violations of L-T relation were observed in the measurements of D-Y production by the fixed-target experiments as well as $\gamma^{*}$/$Z$ production by the CMS [3] and ATLAS [4] experiments at LHC. It is important to understand the origin of these violations. It is found that the violation of the L-T relation seen in CMS and ATLAS data in the region of transverse momentum ($q_{T}$) greater than 5 GeV could be well described taking into account NNLO pQCD effect [5]. On the other hand, the agreement is not as good in a similar comparison [6, 7] for the fixed- target data of NA10 [8], E615 [9] and E866 [10] at $q_{T}<3$ GeV. Transverse- momentum dependent Boer-Mulders functions [11], correlating the nucleon spin with the intrinsic transverse momentum of partons, have been suggested to account for a violation of the L-T relation observed at small $q_{T}$ in the fixed-target experiments. In this proceedings, we show that the $q_{T}$ dependence of the angular distribution coefficients, as well as the violation of the Lam-Tung violation, could be obtained if the angular distribution coefficients were analyzed as a function of the number of accompanying jets in $Z$-boson production measured by the CMS and ATLAS Collaborations [12]. Furthermore we compare the data of dilepton angular parameters $\lambda$, $\mu$, $\nu$ and the L-T violation quantity $1-\lambda-2\nu$ measured by E615 [9] with the fixed-order pQCD calculations. Finally we interpret some notable features of pQCD results using the geometric model [13, 14, 15]. More results and greater details can be found in Ref. [7, 12]. ## 2 Lepton angular distributions of $Z$ production and jet discrimination The lepton angular distribution in the $Z$ rest frame can be expressed as [3, 4] $\displaystyle\frac{d\sigma}{d\Omega}$ $\displaystyle\propto$ $\displaystyle(1+\cos^{2}\theta)+\frac{A_{0}}{2}(1-3\cos^{2}\theta)+A_{1}\sin 2\theta\cos\phi+\frac{A_{2}}{2}\sin^{2}\theta\cos 2\phi$ (1) $\displaystyle+$ $\displaystyle A_{3}\sin\theta\cos\phi+A_{4}\cos\theta+A_{5}\sin^{2}\theta\sin 2\phi+A_{6}\sin 2\theta\sin\phi+A_{7}\sin\theta\sin\phi,$ where $\theta$ and $\phi$ are the polar and azimuthal angles of leptons in the rest frame of $Z$. Since the intrinsic transverse momenta of the annihilating quark and antiquark is neglected in the original Drell-Yan model, the angular distribution is simply $1+\cos^{2}\theta$ and all angular distribution coefficients, $A_{i}$, vanish. For a non-zero dilepton transverse momentum, $q_{T}$, these coefficients can deviate from zero due to QCD effects. However, it was predicted that the coefficients $A_{0}$ and $A_{2}$ should remain identical, $A_{0}=A_{2}$, i.e. the Lam-Tung relation [2]. Figure 1: Comparison between the CMS data [3] of $A_{0}$ and $A_{0}-A_{2}$ for $Z$ production from $p-p$ collisions with fixed-order pQCD calculations. Curves correspond to calculations described in the text. Figure 1 shows the CMS data for $A_{0}$ and $A_{0}-A_{2}$. Pronounced $q_{T}$ dependence of $A_{0}$ is observed and the Lam-Tung relation, $A_{0}-A_{2}=0$, is clearly violated. There are two NLO QCD subprocesses for $Z$ production: $q\bar{q}\rightarrow Zg$ annihilation process, and $qg\rightarrow Zq$ quark Compton scattering process. In the Collins-Soper frame [16], the NLO pQCD predictions of $A_{0}$ and $A_{2}$ as a function of $q_{T}$ of $Z$ for these two processes are $A_{0}=A_{2}=q^{2}_{T}/(M_{Z}^{2}+q^{2}_{T})$ [17] and $A_{0}=A_{2}=5q^{2}_{T}/(M_{Z}^{2}+5q^{2}_{T})$ [18, 19], respectively. The dotted and dashed curves in Fig. 1(a) correspond to these NLO expressions. As the $q\bar{q}$ and $qg$ processes contribute to the $pp\to ZX$ reaction incoherently, the observed $q_{T}$ dependence of $A_{0}$ reflects the combined effect of these two contributions. A best-fit to the CMS $A_{0}$ data, shown as the solid line in Fig. 1(a), gives a mixture of 58.5$\pm$1.6% $qg$ and 41.5$\pm$1.6% $q\bar{q}$ processes. For $pp$ collisions at the LHC, the $qg$ process is expected to be more important than the $q\bar{q}$ process, in agreement with the best-fit result. For $Z$ plus single-jet events, Fig. 1(a) shows that there is remarkable difference in the $q_{T}$ dependence for $A_{0}$ between the $q\bar{q}$ annihilation process and the $qg$ Compton process. Since it is a high-$p_{T}$ gluon (quark) jet associated with the $q\bar{q}(qg)$ process at the $\alpha_{s}$ level, one could first utilize the existing algorithms for quark (gluon) jet identification to separate the $q\bar{q}$ annihilation events from the $qg$ Compton events and investigate the angular distribution of individual event samples. Their angular distribution coefficients for $Z$ plus single jet data would also provide a powerful tool for testing various algorithms designed to distinguish quark jets from gluon jets. For the $Z$ plus multi-jet data, the L-T relation is expected to be violated at a higher level than that of the inclusive production data. Exclusion of the $Z$ plus single-jet events satisfying the L-T relation, would enhance the violation of the L-T relation. We have carried out pQCD calculations using DYNNLO [20, 21] to demonstrate this effect. Figure 1(b) compares the DYNNLO calculations with the CMS $A_{0}-A_{2}$ data. The black band corresponds to the NNLO calculation including contributions from the events of $Z$ with single jet and two jets. The blue band singles out the contributions to $A_{0}-A_{2}$ from $Z$ plus two jets only, showing that the violation of the Lam-Tung relation is indeed amplified for the multi-jet events. This can be readily tested with the data collected at the LHC. ## 3 Lepton angular distributions of Drell-Yan production in fixed-target experiments In the rest frame of the virtual photon in the D-Y process, another expression for the lepton angular distributions commonly used by the fixed-target experiments is given as [2] $\frac{d\sigma}{d\Omega}\propto 1+\lambda\cos^{2}\theta+\mu\sin 2\theta\cos\phi+\frac{\nu}{2}\sin^{2}\theta\cos 2\phi,$ (2) where $\theta$ and $\phi$ refer to the polar and azimuthal angles of leptons. The $\lambda,\mu,\nu$ are related to $A_{0},A_{1},A_{2}$ in Eq. (1) via $\displaystyle\lambda=\frac{2-3A_{0}}{2+A_{0}};~{}~{}~{}\mu=\frac{2A_{1}}{2+A_{0}};~{}~{}~{}\nu=\frac{2A_{2}}{2+A_{0}}.$ (3) Equation (3) shows that the L-T relation, $1-\lambda-2\nu=0$, is equivalent to $A_{0}=A_{2}$. Figure 2: Comparison of NLO (red points) and NNLO (blue points) fixed-order pQCD calculations with the E615 $\pi^{-}+W$ D-Y data at 252 GeV [9] (black points) for $\lambda$, $\mu$, $\nu$ and $1-\lambda-2\nu$. In Fig. 2, we compare the results of $\lambda$, $\mu$, $\nu$, and the L-T violation, $1-\lambda-2\nu$, from the fixed-order pQCD calculations with 252-GeV $\pi^{-}+W$ data from E615 experiment [9]. The angular parameters are evaluated as a function of $q_{T}$ in the Collins-Soper frame. Overall, the calculated $\lambda$, $\mu$ and $\nu$ exhibit distinct $q_{T}$ dependencies. At $q_{T}\rightarrow 0$, $\lambda$, $\mu$ and $\nu$ approach the values predicted by the collinear parton model: $\lambda=1$ and $\mu=\nu=0$. As $q_{T}$ increases, Fig. 2 shows that $\lambda$ decreases toward its large-$q_{T}$ limit of $-1/3$ while $\nu$ increases toward $2/3$, for both $q\bar{q}$ and $qG$ processes [18, 19]. The $q_{T}$ dependence of $\mu$ is relatively mild compared to $\lambda$ and $\nu$. This is understood as a result of some cancellation effect, to be discussed in Sec. 4. Comparing the results of the NLO with the NNLO calculations, $\lambda{\rm(NNLO)}$ is smaller than $\lambda\rm{(NLO)}$ while $\mu$ and $\nu$ are very similar for NLO and NNLO. The amount of L-T violation, $1-\lambda-2\nu$, is zero in the NLO calculation, and nonzero and positive in the NNLO calculation. As seen in Fig. 2, the pQCD predicts a sizable magnitude for $\nu$, comparable to the data. Such pQCD effect should be included in the determination of nonperturbative Boer-Mulders effect from the data of $\nu$. ## 4 Geometric model As introduced above, both CMS and E615 data of lepton angular distributions for $Z$ and D-Y production can be reasonably well described by the NLO and NNLO pQCD calculations. It is interesting to see that various salient features of pQCD calculations could be understood using a geometric approach developed in Refs. [13, 14]. In the Collins-Soper $\gamma^{*}/Z$ rest frame, the hadron plane, the quark plane, and the lepton plane of collision geometry are defined [13, 14]. A pair of collinear $q$ and $\bar{q}$ with equal momenta annihilate into a $\gamma^{*}/Z$. The momentum unit vector of $q$ is defined as $\hat{z}^{\prime}$, and the quark plane is formed by the $\hat{z}^{\prime}$ and the $\hat{z}$ axes of Collins-Soper frame. The angular coefficients $A_{i}$ in Eq. (3) can be expressed in term of $\theta_{1}$ and $\phi_{1}$ as follows: $\displaystyle A_{0}=\langle\sin^{2}\theta_{1}\rangle,~{}~{}~{}A_{1}=\frac{1}{2}\langle\sin 2\theta_{1}\cos\phi_{1}\rangle,~{}~{}~{}A_{2}=\langle\sin^{2}\theta_{1}\cos 2\phi_{1}\rangle,$ (4) where the $\theta_{1}$ and $\phi_{1}$ are the polar and azimuthal angles of the natural quark axis $\hat{z}^{\prime}$ of the quark plane in the Collins- Soper frame. Up to NLO ($\mathcal{O}(\alpha_{S})$) in pQCD, the quark plane coincides with the hadron plane and $\phi_{1}=0$. Therefore $A_{0}=A_{2}$ or $1-\lambda-2\nu=0$, i.e., the L-T relation is satisfied. Higher order pQCD processes allow the quark plane to deviate from the hadron plane, i.e., $\phi_{1}\neq 0$. This acoplanarity effect leads to a violation of the L-T relation. For a nonzero $\phi_{1}$, Eq. (4) shows that $A_{2}<A_{0}$. Therefore, when the L-T relation is violated, $A_{0}$ must be greater than $A_{2}$ or, equivalently, $1-\lambda-2\nu>0$. This expectation of $1-\lambda-2\nu>0$ in this geometric approach agrees with the results of NNLO pQCD calculations shown in Fig. 2. The geometric approach offers a simple and intuitive interpretation for this result. Furthermore the sign of $\mu$ could be either positive or negative, depending on which parton and hadron the gluon is emitted from [14, 7]. Hence, one expects some cancellation effects for $\mu$ among contributions from various processes. Each process is weighted by the corresponding parton density distributions. At mid-rapidity, the momentum fraction carried by the beam parton is comparable to that of the target parton. Therefore, the weighting factors for various processes are of similar magnitude and the cancellation effect could be very significant, resulting in a small value of $\mu$. ## 5 Summary We have presented a comparison of the measurements of the angular parameters $A_{0}$ and $A_{0}-A_{2}$ of the $Z$ production from the CMS experiment as well as $\lambda$, $\mu$, $\nu$ and $1-\lambda-2\nu$ of the D-Y process from the fixed-target E615 experiment with the corresponding results from the NLO and NNLO pQCD calculations. Qualitatively the transverse momentum dependence of measured angular parameters could be described by pQCD. The L-T violation part $A_{0}-A_{2}$ or $1-\lambda-2\nu$ remains zero in the NLO pQCD calculation and turns positive in NNLO pQCD. The measurement of $A_{0}$ and $A_{2}$ coefficients in $Z$ plus single-jet or multi-jet events would provide valuable insights on the origin of the violation of the L-T relation and could be used an an index in discriminating the intrinsic property of high-$q_{T}$ jets. Within the geometric picture, the occurrence of acoplanarity between the quark plane and the hadron plane ($\phi_{1}\neq 0$), for the pQCD processes beyond NLO offers an interpretation of a violation of the L-T relation. The predicted positive value of $1-\lambda-2\nu$, or $A_{0}>A_{2}$ when $\phi_{1}$ is nonzero, is consistent with the NNLO pQCD results. ## References * [1] J. C. Peng and J. W. Qiu, Prog. Part. Nucl. Phys. 76, 43 (2014). * [2] C. S. Lam and W. K. Tung, Phys. Rev. D 21, 2712 (1980). * [3] CMS Collaboration, V. Khachatryan et al., Phys. Lett. B 750, 154 (2015). * [4] ATLAS Collaboration, G. Aad et al., JHEP 08, 159 (2016). * [5] R. Gauld, A. Gehrmann-De Ridder, T. Gehrmann, E. W. N. Glover and A. Huss, JHEP 1711, 003 (2017). * [6] M. Lambertsen and W. Vogelsang, Phys. Rev. D 93, 114013 (2016). * [7] W. C. Chang, R. E. McClellan, J. C. Peng and O. Teryaev, Phys. Rev. D 99, 014032 (2019). * [8] NA10 Collaboration, S. Falciano et al., Z. Phys. C 31, 513 (1986); M. Guanziroli et al., Z. Phys. C 37, 545 (1988). * [9] E615 Collaboration, J. S. Conway et al., Phys. Rev. D 39, 92 (1989); J. G. Heinrich et al., Phys. Rev. D 44, 1909 (1991). * [10] E866/NuSea Collaboration, R. S. Towell et al., Phys. Rev. D 64, 052002 (2001). * [11] D. Boer, Phys. Rev. D 60, 014012 (1999). * [12] J. C. Peng, W. C. Chang, R. E. McClellan and O. Teryaev, Phys. Lett. B 797, 134895 (2019). * [13] J. C. Peng, W. C. Chang, R. E. McClellan, and O. Teryaev, Phys. Lett. B 758, 384 (2016). * [14] W. C. Chang, R. E. McClellan, J. C. Peng and O. Teryaev, Phys. Rev. D 96, 054020 (2017). * [15] J. C. Peng, D. Boer, W. C. Chang, R. E. McClellan and O. Teryaev, Phys. Lett. B 789, 356 (2019). * [16] J. C. Collins and D. E. Soper, Phys. Rev. D16, 2219 (1977). * [17] J. C. Collins, Phys. Rev. Lett. 42, 291 (1979). * [18] R. L. Thews, Phys. Rev. Lett. 43, 987 (1979). * [19] J. Lindfors, Phys. Scr. 20, 19 (1979). * [20] S. Catani and M. Grazzini, Phys. Rev. Lett. 98, 222002 (2007). * [21] S. Catani et al., Phys. Rev. Lett. 103, 082001 (2009).
# Inadequacy of Linear Methods for Minimal Sensor Placement and Feature Selection in Nonlinear Systems; a New Approach Using Secants Samuel E. Otto Department of Mechanical and Aerospace Engineering, Princeton University, Princeton, NJ 08544<EMAIL_ADDRESS>and Clarence W. Rowley ###### Abstract. Sensor placement and feature selection are critical steps in engineering, modeling, and data science that share a common mathematical theme: the selected measurements should enable solution of an inverse problem. Most real- world systems of interest are nonlinear, yet the majority of available techniques for feature selection and sensor placement rely on assumptions of linearity or simple statistical models. We show that when these assumptions are violated, standard techniques can lead to costly over-sensing without guaranteeing that the desired information can be recovered from the measurements. In order to remedy these problems, we introduce a novel data- driven approach for sensor placement and feature selection for a general type of nonlinear inverse problem based on the information contained in secant vectors between data points. Using the secant-based approach, we develop three efficient greedy algorithms that each provide different types of robust, near- minimal reconstruction guarantees. We demonstrate them on two problems where linear techniques consistently fail: sensor placement to reconstruct a fluid flow formed by a complicated shock-mixing layer interaction and selecting fundamental manifold learning coordinates on a torus. ###### Key words and phrases: nonlinear inverse problems, state estimation, feature selection, manifold learning, greedy algorithms, submodular optimization, shock-turbulence interaction, reduced-order modeling This research was supported by the Army Research Office under grant number W911NF-17-1-0512. S.E.O. was supported by a National Science Foundation Graduate Research Fellowship Program under Grant No. DGE-2039656. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation. ## 1\. Introduction Reconstructing the state of complex systems like fluid flows, chemical processes, and biological networks from measurements taken by a few carefully chosen sensors is a crucial task for controlling, forecasting, and building simplified models of these systems. In this setting it is important to be able to reconstruct the relevant information about the system using the smallest total number of measurements which includes minimizing the number of sensors to reduce cost, and using the shortest possible measurement histories to shorten response time. Feature selection in statistics and machine learning is a related task where one tries to find a small subset of measured variables (features) in the available data that allow one to reliably predict a quantity of interest. Nonlinear reconstruction can yield large improvements over linear reconstruction when the sensors or features are carefully selected [20]. Successful nonlinear reconstruction techniques include neural networks [37],[38], deep nonlinear state estimators [23], [29], and convex $\ell^{1}$ minimization to reveal sparse coefficients in learned libraries [64], [9]. The need for nonlinear representation and reconstruction is also recognized in the reduced-order-modeling community where it is called “nonlinear Galerkin” approximation [31], [47], [34]. These methods are necessary because in many systems of interest, the state is found to lie near a low-dimensional underlying manifold that is curved in such a way that it is not contained in any low-dimensional subspace [40]. We will show that the best possible linear reconstruction accuracy is fundamentally limited by the number of measurements (features) and the fraction of the variance that is captured in the principal subspace [24] of that dimension. In essence, any linear representation in a subspace is “too loose” and demands an excessive number of measurements to even have a hope of accurately reconstructing the state using linear functions. Nonlinear reconstruction is much more powerful, as Whitney’s celebrated embedding theorem (Theorem. 5, [62]) shows that states on any $r$-dimensional smooth manifold can be reconstructed using $2r$ carefully chosen measurements. If the measurements must be linear functions of the state on a compact sub-manifold of $\mathbb{R}^{n}$ then $2r+1$ can be found [61]. With many measurements available from our sensors (though not necessarily ones that achieve Whitney’s results), the problem that remains is to properly choose them so that nonlinear reconstruction is possible and robust to noise. While nonlinear reconstruction has proved to be extremely advantageous, the overwhelming majority of sensor placement and feature selection methods rely on measures of linear or Gaussian reconstruction accuracy as an optimization criteria. Such methods include techniques based on sampling modal bases [66], [33], [14], [18], [8], linear dynamical system models [36], [17], [53], [54], [59], Bayesian and maximum likelihood optimality in linear inverse problems [13], [26], [51], information-theoretic criteria under Gaussian or other simple statistical models [28], [12], [11], [52], [50], and sparse linear approximation in dictionaries using LASSO [57], [67] or orthogonal matching pursuit [41], [58]. We provide an overview of a representative collection of these methods that we shall use as a basis for comparison in Section 2. We show that relying on these linear, Gaussian techniques to identify sensors that will be used for nonlinear reconstruction can lead to costly over-sensing when the underlying manifold is low-dimensional, but the data do not lie in an equally low-dimensional subspace. This effect is most pronounced when the most energetic (highest variance) components of the data are actually functions of less-energetic components, but not vice versa. In such cases, the linear techniques are consistently “tricked” into sensing the most energetic components while failing to capture the important less energetic ones that can actually be used for minimal reconstruction. These situations are not merely academic, and they actually abound in physics and in data science. As we shall discuss in Section 3, the problem appears in mixing layer fluid flows and in the presence of shock waves, which are both ubiquitous in aerodynamics. The presence of important low-energy sub-harmonic frequencies is also generic behavior after a period-doubling bifurcation, which is a common route to chaos, for instance in ecosystem collapse [60] and cardiac arrhythmia [45]. In data science, the problem is most pronounced when we try to select fundamental nonlinear embedding coordinates for a data set using manifold learning techniques like kernel PCA [49], Laplacian eigenmaps [1], diffusion maps [16], and Isomap [56] as we shall discuss in Section 3.3. In order to address the limitations of linear, Gaussian methods for sensor placement and feature selection demonstrated in the first half of the paper, we develop a novel data-driven approach based on consideration of secant vectors between states in Section 4. Related secant-based approaches have been pioneered by [5], [25], [21], [55] for the purpose of finding optimal embedding subspaces. While their considerations of secants lead to continuous optimization problems over subspaces, our considerations of secants lead to combinatorial optimization problems over sets of sensors. We develop three different secant-based objectives together with greedy algorithms that each provide different types of robust, near-minimal reconstruction guarantees for very general types of nonlinear inverse problems. The guarantees stem from the underlying geometric information that is captured by secants and encoded in our optimization objectives. Moreover, the objectives we consider each have the celebrated diminishing returns property called _submodularity_ , allowing us to leverage the classical results of G. L. Nemhauser and L. A. Wolsey et al. [39], [63] to guarantee the performance of efficient greedy algorithms for sensor placement. We also leverage concentration of measure results in order to prove performance guarantees when the secants are randomly down-sampled, enabling computational scalability to very large data sets. Each of these techniques demonstrates greatly improved performance compared to a large collection of linear techniques on a canonical shock-mixing layer flow problem [65] as well as for selecting fundamental manifold learning coordinates. ## 2\. Background on Linear, Gaussian, Techniques The predominant sensor placement, feature selection, and experimental design techniques available today rely on linear and/or Gaussian assumptions about the underlying data: that is, that the data live in a low-dimensional subspace and/or have a Gaussian distribution. Under these assumptions, it becomes easy to quantify the performance of sensors, features, or experiments, using a variety of information theoretic, Bayesian, maximum likelihood, or other optimization criteria. A comprehensive review is beyond the scope of this paper, and of course we do not claim that linear methods always fail. Rather, we argue that because the underlying linear, Gaussian assumptions are violated in many real-world problems, we cannot expect them to find small collections of sensors that enable nonlinear reconstruction of the desired quantities. We shall briefly review the collection of linear techniques that we shall compare to throughout this work and that we think are representative of the current literature. ### 2.1. (Group) LASSO The Least Absolute Shrinkage and Selection Operator (LASSO) method introduced by R. Tibshirani [57] is a highly successful technique for feature selection in machine learning that has found additional applications in compressive sampling recovery [10] and system identification [7]. A generalization by M. Yuan and Y. Lin [67] called group LASSO is especially relevant for sensor placement since it allows measurements to be selected in groups that might come from the same sensor at different instants of time. Suppose we are given a collection of data consisting of available measurements ${\mathbf{m}}_{j}({\mathbf{x}}_{i})$, $j=1,\ldots,M$ along with relevant quantities ${\mathbf{g}}({\mathbf{x}}_{i})$ that we wish to reconstruct over a collection of states ${\mathbf{x}}_{i}$, $i=1,\ldots,N$. The group LASSO convex optimization problem takes the form (1) $\operatorname*{\min\\!imize\enskip}_{{\mathbf{A}}_{1},\ldots,{\mathbf{A}}_{M}}\sum_{i=1}^{N}\Big{\|}{\mathbf{g}}({\mathbf{x}}_{i})-\sum_{j=1}^{M}{\mathbf{A}}_{j}{\mathbf{m}}_{j}({\mathbf{x}}_{i})\Big{\|}_{2}^{2}+\gamma\sum_{j=1}^{M}\left\|{\mathbf{A}}_{j}\right\|_{F}$ and tries to reconstruct the targets as accurately as possible using a linear combination of the measurements subject to a sparsity-promoting penalty. The strength of the penalty depends on the user-specified parameter $\gamma\geq 0$ and forces the coefficient matrices ${\mathbf{A}}_{j}$ on many of the measurement groups to be identically zero. Those coefficient matrices with nonzero entries indicate the sensors that should be used to _linearly_ reconstruct the target variables with high accuracy. ### 2.2. Determinantal “D”-Optimal Selection Suppose the state ${\mathbf{x}}$ has a prior probability distribution with covariance ${\mathbf{C}}_{{\mathbf{x}}}$ and the target variables ${\mathbf{g}}({\mathbf{x}})$ and measurements ${\mathbf{m}}_{j}({\mathbf{x}})$, $j=1,\ldots,M$ are linear functions of the state (2) ${\mathbf{g}}({\mathbf{x}})={\mathbf{T}}{\mathbf{x}},\qquad{\mathbf{m}}_{j}({\mathbf{x}})={\mathbf{M}}_{j}{\mathbf{x}}+{\mathbf{n}}_{j}$ where ${\mathbf{n}}_{j}$ is the mean-zero, state independent, noise from the $j$th sensor with covariance ${\mathbf{C}}_{{\mathbf{n}}_{j}}$. Then, if ${\mathbf{M}}_{\mathscr{S}}$ is a matrix with rows given by ${\mathbf{M}}_{j}$ and ${\mathbf{C}}_{{\mathbf{n}}_{\mathscr{S}}}$ is a block diagonal matrix formed from ${\mathbf{C}}_{{\mathbf{n}}_{j}}$, for $j$ in a given set of sensors $\mathscr{S}$, then the optimum (least-squares) linear estimate of ${\mathbf{g}}({\mathbf{x}})$ given ${\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})$ has error covariance (3) ${\mathbf{C}}_{{\mathbf{e}}}(\mathscr{S})={\mathbf{T}}\left({\mathbf{C}}_{{\mathbf{x}}}^{-1}+{\mathbf{M}}_{\mathscr{S}}^{T}{\mathbf{C}}_{{\mathbf{n}}_{\mathscr{S}}}^{-1}{\mathbf{M}}_{\mathscr{S}}\right)^{-1}{\mathbf{T}}^{T}.$ If ${\mathbf{x}}$ and the noise are independent Gaussian random variables then Eq. 3 is the covariance of the posterior distribution for ${\mathbf{g}}({\mathbf{x}})$ given ${\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})$. A low-dimensional representation of the state and its covariance are usually found from data via principal component analysis (PCA) [24] or proper orthogonal decomposition (POD) [4] when an analytical model is not available. A common technique, referred to as the Bayesian approach in the optimal design of experiments [44] is to quantify performance using functions of ${\mathbf{C}}_{{\mathbf{e}}}(\mathscr{S})$ [13]. In particular, Bayesian determinantal or “D”-optimality entails minimizing $\log{\det{{\mathbf{C}}_{{\mathbf{e}}}(\mathscr{S})}}$, which, under Gaussian assumptions, is equivalent to minimizing the conditional entropy [52], [50] or the volumes of confidenece ellipsoids about the maximum a posteriori (MAP) estimate of ${\mathbf{g}}({\mathbf{x}})$ given ${\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})$ [26]. This approach is widely used for sensor placement since it readily admits efficient approximations based on convex relaxation [26] and greedy algorithms [51], [59] with guaranteed performance. Similar objectives have been used to quantify observability and controllability for sensor and actuator placement in linear dynamical systems [53], [54]. When there is no prior probability distribution for ${\mathbf{x}}$ and we want to estimate the full state ${\mathbf{g}}({\mathbf{x}})={\mathbf{x}}$ from measurements corrupted by Gaussian noise, we can construct the maximum likelihood estimate whose error covariance is (4) ${\mathbf{C}}_{{\mathbf{e}}}(\mathscr{S})=\left({\mathbf{M}}_{\mathscr{S}}^{T}{\mathbf{C}}_{{\mathbf{n}}_{\mathscr{S}}}^{-1}{\mathbf{M}}_{\mathscr{S}}\right)^{-1}.$ Minimizing the volumes of confidence ellipsoids in this setting as is done in [26] is referred to as maximum likelihood “D”-optimality since it entails maximizing $\log\det{\left({\mathbf{M}}_{\mathscr{S}}^{T}{\mathbf{C}}_{{\mathbf{n}}_{\mathscr{S}}}^{-1}{\mathbf{M}}_{\mathscr{S}}\right)}$. In the absence of the regularizing effect the prior distribution has on the estimate, we must have at least as many sensor measurements as state variables in the maximum likelihood setting. ### 2.3. Pivoted QR Factorization Pivoted matrix factorization techniques, and QR pivoting in particular, have become a popular choice for sensor placement [33], [6] and feature selection in reduced-order modeling [14], [18], where the method is often referred to as the Discrete Empirical Interpolation Method (DEIM). This approach dates back to P. Businger and G. H. Golub’s seminal work [8], which introduced Householder-pivoted QR factorization for the purpose of feature selection in least squares fitting problems. The approach is also closely related to orthogonal matching pursuit [41] and simultaneous orthogonal matching pursuit [58], which are widely used sparse approximation algorithms. In its simplest form, one supposes that the underlying state to be estimated ${\mathbf{g}}({\mathbf{x}})={\mathbf{x}}$ is low dimensional (e.g., using its PCA or POD coordinate representation) and selects the linear measurements from among the rows of a matrix ${\mathbf{M}}$ by forming a pivoted QR decomposition of the form (5) ${\mathbf{M}}^{T}\begin{bmatrix}[c|c]{\mathbf{P}}_{1}&{\mathbf{P}}_{2}\end{bmatrix}={\mathbf{Q}}\begin{bmatrix}[c|c]{\mathbf{R}}_{1}&{\mathbf{R}}_{2}\end{bmatrix},$ where $\begin{bmatrix}[c|c]{\mathbf{P}}_{1}&{\mathbf{P}}_{2}\end{bmatrix}$ is a permutation. The first $K=\dim{\mathbf{x}}$ pivot columns forming ${\mathbf{P}}_{1}$ determine a set of sensor measurements ${\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})={\mathbf{M}}_{\mathscr{S}}{\mathbf{x}}={\mathbf{P}}_{1}^{T}{\mathbf{M}}{\mathbf{x}}$ from which ${\mathbf{x}}$ can be robustly recovered as (6) ${\mathbf{x}}=\left({\mathbf{P}}_{1}^{T}{\mathbf{M}}\right)^{-1}{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})={\mathbf{Q}}\left({\mathbf{R}}_{1}^{T}\right)^{-1}{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}).$ This approach is successful because at each step of the QR pivoting process, the measurement that maximizes the corresponding diagonal entry of the upper triangular matrix ${\mathbf{R}}_{1}$ is selected. The resulting large diagonal entries of ${\mathbf{R}}_{1}$ mean that measurement errors are not strongly amplified by the linear reconstruction map ${\mathbf{Q}}\left({\mathbf{R}}_{1}^{T}\right)^{-1}$. ## 3\. Problems with Linear Techniques In this section, we illustrate the problems with employing linear state reconstruction and sensor placement techniques for nonlinear systems and data sets by means of an example. We consider the shock-mixing layer interaction proposed by Yee et al. [65], which has become a canonical problem for studying jet noise production as well as high-order numerical methods. This problem captures many key elements of shock wave-turbulent boundary layer interactions that, according to S. Priebe and M. P. Martín [42] “occur in many external and internal compressible flow applications such as transonic aerofoils, high- speed engine inlets, internal flowpaths of scramjets, over-expanded rocket engine nozzles and deflected control surfaces or any other discontinuities in the surface geometry of high-speed vehicles.” The resulting pressure and heat transfer fluctuations can be large, so it is important to monitor the state of these flows to ensure the safety of a vehicle. Our goal will be to choose a small number of sensor locations in this flow at which to measure either the horizontal, $u$, or vertical, $v$, velocity component in order to reconstruct the entire velocity field. A snapshot of these velocity fields from the fully-developed flow computed using the high- fidelity local WENO-type characteristic filtering method of S.-C. Lo et al. [32] is shown in Fig. 1. While the flow is very nearly periodic, and hence lives near a one-dimensional loop in state space, the complicated physics arising from the interaction of the oblique shock with vortices in the spatially-evolving mixing layer results in data that do not lie near any low- dimensional subspace. In addition to being high dimensional, this flow exhibits the low-frequency unsteadiness characteristic of shock wave–turbulent boundary layer interactions [42], [15], [43] and of spatial mixing layer flows in general [22]. (a) stream-wise $u$ velocity component (b) transverse $v$ velocity component (c) available sensor locations Figure 1. A snapshot of the $u$ and $v$ velocity components in the shock mixing-layer flow is shown in (a) and (b) along with the sensors selected using various methods from among the two components at $1105$ available locations shown in (c). These methods include LASSO with PCA (black o), LASSO with Isomap (red x) greedy Bayes D-optimality (magenta x), convex Bayes D-optimality (black $>$), convex D-optimality for modes $3$ and $4$ (black v), QR pivoting (green +), and secant-based techniques using detectable differences ($\\#1,\\#2$: green star, $\\#3$: black star) and the amplification threshold method (black square). ### 3.1. The Need for Nonlinear Reconstruction Linear reconstruction is fundamentally confined to a subspace whose dimension is at most equal to the total number of sensor measurements. Hence the fraction of the variance that linear reconstruction can capture using $d$ measurements is at most the fraction of the variance along the first $d$ principal components: in particular, the coefficient of determination is bounded by (7) $R^{2}\leq\frac{\sigma_{1}^{2}+\cdots+\sigma_{d}^{2}}{\sigma_{1}^{2}+\cdots+\sigma_{n}^{2}}.$ Examining the fraction of the variance captured by the leading principal subspaces in Figure 2a leads us to the rather disappointing conclusion that in order to capture $90\%$ of the variance in the shock-mixing layer flow via linear reconstruction, we need at least $11$ independent measurements, and to capture $98\%$ we need at least $33$. variance fraction remainingprincipal subspace dimension (a) variance orthogonal to principal subspaces $\theta$$\phi_{1}$$\phi_{2}$ (b) Isomap coordinates phase angle, $\theta$modal coefficient, $z_{k}$ (c) PCA coefficients Figure 2. The linear and nonlinear dimension reduction techniques PCA (a.k.a POD) and Isomap are applied to the shock-mixing layer data. (a) shows the remaining fraction of the total variance orthogonal to each leading principal subspace. (b) plots the data in the leading two Isomap embedding coordinates, revealing that it lies very near a loop in state space. (c) shows how the leading principal components (modal coefficients) vary with the phase angle around the loop. The black vertical lines reveal distinct points where the leading three principal components are identical. The best possible linear reconstruction performance can be arbitrarily poor even though the underlying manifold is low-dimensional. We illustrate this fact with the following toy model that resembles the phase dependence of principal components in the shock-mixing layer problem shown in Figures 2b and 2c. Let $\theta$ be uniformly distributed over the interval $[0,2\pi]$ and let the components of the state vector have sinusoidal dependence on the phase given by (8) $x_{2k-1}=\sqrt{2}\cos(k\theta),\ \ x_{2k}=\sqrt{2}\sin(k\theta),\ \ k=1,\ldots,n/2.$ Since these components are orthonormal functions of $\theta$ with respect to the uniform probability measure on $[0,2\pi]$, the state vector has isotropic covariance $\mathbb{E}{\mathbf{x}}{\mathbf{x}}^{T}={\mathbf{I}}_{n}$ and the fraction of the variance captured by the leading $d$ principal components is $d/n$. As the dimension increases, the highest possible coefficient of determination for linear reconstruction approaches zero since $R^{2}\leq d/n\to 0$ as $n\to\infty$. Meanwhile, it’s obvious that the state vector can be perfectly reconstructed as a nonlinear function of $x_{1}$ and $x_{2}$ alone. Indeed, it is possible to reconstruct the entire shock-mixing layer flow-field as a nonlinear function of the velocity measurements at two carefully chosen locations. In particular, the measurements made at the locations marked by the two green stars in Figure 1 are one-to-one with the phase and hence the state of the flow. This is seen in Figure 3g, where the phase angle (color) — hence the full state — can be determined uniquely from the values of the measurements. Meanwhile, the best possible linear reconstruction performance using two measurements is $R^{2}<0.5$. In practice, many nonlinear reconstruction techniques are available including neural networks [37], Gaussian process regression [46], and recurrent neural networks for time-delayed measurements [29]. Using Gaussian process regression and the two sensor locations marked by green stars in Figure 1, we obtain near-perfect, robust reconstruction of the leading $100$ principal components. The resulting reconstruction accuracy for the flow-fields on a held-out set of $250$ snapshots is $R^{2}=0.986$. ### 3.2. The Need for Nonlinear Sensor Placement With such poor reconstruction afforded by linear techniques, we cannot expect sensor placement methods based on them to perform any better. This is not to say that a practitioner won’t ever find lucky sensor locations for nonlinear reconstruction by employing a sensor placement technique that maximizes linear reconstruction accuracy. However, this kind of luck is not guaranteed as illustrated when we apply state of the art linear sensor placement techniques to the shock mixing-layer problem. Indeed Figures 3a, 3b, 3c, 3d, and 3e provide visual proof that three sensors chosen using LASSO to reconstruct the leading $100$ principal components, LASSO to reconstruct the leading two Isomap coordinates, the greedy Bayes D-optimality approach, the convex Bayes D-optimality approach, and pivoted QR factorization do not produce measurements that are one-to-one with the state. Implementation details can be found in Appendix A. In each case, there are at least two distinct states with different phases on the orbit (color) for which the sensors measure the same values and hence cannot be used to tell them apart. Even measuring the leading three principal components directly, which are optimal for linear reconstruction, cannot always reveal the state of the shock-mixing layer flow. The black vertical lines in Figure 2c indicate the phases of two distinct states for which the leading three principal components agree, yet the fourth differs. One may wonder whether the fact that the third and fourth principal components are one-to-one with the state can be leveraged for sensor placement. Even our attempt to place three maximum likelihood D-optimal sensors using the convex optimization approach of [26] to reconstruct the third and fourth principal components fails to produce measurement that can recover the phase of the flow as seen in Figure 3f. On the other hand, it is possible to find two sensor locations whose measurements are one-to-one with the state as shown in Figure 3g. The resulting curve near which the data lie has a kink in the lower-right region indicating that while the measurements are one-to-one, the time-derivative of the state cannot be determined at this point. Capturing time-derivatives is necessary for building reduced-order models, and this can be accomplished using the three sensors marked by black squares in Figure 1 and whose measurements are plotted in Figure 3i. We note, however, that these locations are far apart in space, and so will be more sensitive to perturbations of the shear-layer thickness which affects the horizontal spacing of vortices. (a) LASSO+PCA (b) LASSO+Isomap (c) greedy Bayes D-opt. (d) convex Bayes D-opt. (e) pivoted QR (f) convex M.L. D-opt., modes $3,4$ (g) secant detect. diffs., $\\#(\mathscr{S})=2$ (h) secant detect. diffs., $\\#(\mathscr{S})=3$ (i) secant amplification tol. Figure 3. these plots show the measurements made by sensors selected using various methods on the shock-mixing layer flow problem. Each dot indicates the values measured by the sensors and its color indicates the phase of the corresponding flowfield. The sensors selected using linear methods shown in (a)-(f) each make identical or nearly identical measurements on distinct flowfields, indicated by overlapping points with different colors. These sensors cannot tell those flowfields apart since the measurements are the same. The sensors selected using secant-based methods shown in (g)-(i) make distinct measurements for distinct states and have no such overlaps. The linear techniques, LASSO, greedy and convex Bayesian D-optimal selection, pivoted QR, and even direct measurement of principal components fail to reveal the minimum number of sensors needed to reconstruct the state because there is important information about the flow contained in less-energetic principal components. In particular, Figure 2c shows that the most energetic two principal components oscillate with twice the frequency of the third and fourth most energetic components as one moves around the orbit. In trying to maximize the variance captured by a linear estimator, the linear sensor placement techniques are doomed to choose sensors whose measurements return to the same values twice in one period as in Figures 3a, 3c, and 3e. In addition, the convex Bayesian D-optimal approach finds sensors that achieve a superior value of the objective $\log\det{{\mathbf{C}}_{{\mathbf{e}}}(\mathscr{S})}$ than the greedy Bayesian D-optimal approach, yet the resulting measurements in Figure 3d have many more self-intersections than the greedy method in Figure 3c. We are forced to conclude that sensor placement based on linear reconstruction is totally unconnected with nonlinear reconstructability when the underlying manifold and principal dimensions do not agree. This can be seen most clearly from the fact that by simply re-scaling each coordinate in the toy model Eq. 8 by positive constants $\alpha_{1},\ldots,\alpha_{n}$, we can trick these techniques into selecting any given collection of coordinates. Under this scaling, the covariance matrix becomes $\text{diag}(\alpha_{1}^{2},\ldots,\alpha_{n}^{2})$ and if we sort the constants in decreasing order $\alpha_{k_{1}}\geq\alpha_{k_{2}}\geq\cdots$ then the variance captured by a linear reconstruction from $d$ measurements cannot exceed (9) $R^{2}\leq\frac{\alpha_{k_{1}}^{2}+\cdots+\alpha_{k_{d}}^{2}}{\alpha_{1}^{2}+\cdots+\alpha_{n}^{2}},$ according to the bound in Eq. 7. Equality is achieved by the optimal linear estimator based on measured coordinates $x_{k_{1}},\ldots,x_{k_{d}}$. Meanwhile, the only pair of coordinates needed for nonlinear reconstruction are $x_{1}$ and $x_{2}$. The key point is that sensor placement approaches based on linear reconstruction tend to pick sensor locations that have high variance over other choices that can be more informative. The linear approach works well when a small number of principal components contain essentially all of the variance or when all higher modal components are very nearly determined by the lower ones. But as we have shown, linear approaches to sensor placement can fail catastrophically when genuinely informative fluctuations, e.g. sub- harmonics, produce significant variance orthogonal to the leading principal subspace. In order to reveal minimal sensor locations that can be used for nonlinear reconstruction in such situations, we cannot rely on linear reconstruction performance as an optimization criteria, and an entirely new approach is needed. In Section 4 we discuss an approach that can recover the correct coordinates from which all others can be nonlinearly reconstructed. ### 3.3. Selecting Manifold Learning Coordinates The examples presented in the previous Section 3.2 involved data lying near a one-dimensional underlying manifold. Essentially the same problems can occur for data lying near higher-dimensional manifolds, and an especially illustrative and practically useful application where this situation is routinely encountered is manifold learning. In general, manifold learning seeks to find a small collection of nonlinear coordinates that fully describe the structure of a dataset, i.e., that embed it in a lower-dimensional space. Many techniques including kernel PCA [49], Laplacian eigenmaps [1], diffusion maps [16], and Isomap [56] accomplish this via eigen-decomposition of various symmetric matrices (10) $\mathbf{G}=\boldsymbol{\Phi}\boldsymbol{\Lambda}^{2}\boldsymbol{\Phi}^{T},\qquad\boldsymbol{\Phi}=\begin{bmatrix}\boldsymbol{\phi}_{1}&\cdots&\boldsymbol{\phi}_{r}\end{bmatrix}$ derived from pair-wise similarity among data points. The $k$th eigen- coordinate of each point in the data set is given by the elements of $\boldsymbol{\phi}_{k}$, which can be viewed as a discrete approximation of an eigenfunction of some kernel integral operator on the underlying manifold. These methods suffer from a well-known issue when the dataset has multiple length scales: namely, there may be several redundant harmonically related eigen-coordinates with higher salience (determined by the eigenvalues) before one encounters a new fundamental eigen-coordinate describing a new set of features. This makes the search for a fundamental set of eigen-coordinates that embed the underlying manifold a potentially large combinatorial search problem. As a concrete example, consider the Isomap eigen-coordinates shown in Figure 4 computed from 2000 points lying on the torus in $\mathbb{R}^{3}$, (11) $\mathbf{x}=\left(\left(5+\cos{\theta_{2}}\right)\cos{\theta_{1}},\ \left(5+\cos{\theta_{2}}\right)\sin{\theta_{1}},\ \sin{\theta_{2}}\right),$ with $(\theta_{1},\theta_{2})$ drawn uniformly at random from the square $[0,2\pi]\times[0,2\pi]$. Toroidal dynamics are known to occur in combustion instabilities where multiple incommensurate frequencies are observed [19], [30], producing data that winds around a torus in high-dimensional state space. One may want to build simplified reduced-order models of these dynamics by finding a small set of nonlinear coordinates that described the state on the torus using manifold learning. Considering the torus in Eq. 11, the underlying kernel integral operators associated with each manifold learning technique mentioned above are equivariant with respect to rotations about $\theta_{1}$, meaning that among their eigenfunctions are always those of the symmetry’s generator, namely $\phi_{k}({\mathbf{x}})=e^{ik\theta_{1}({\mathbf{x}})}$. Unsurprisingly, the leading six Isomap eigen-coordinates, ranked by their associated eigenvalues, are all harmonically related modes resembling the real and imaginary parts of $e^{ik\theta_{1}}$, which provide redundant information about $\theta_{1}$ and no information about $\theta_{2}$. The coordinate $\theta_{1}$ corresponds to larger spatial variations among points and it is not until we encounter the seventh eigen-coordinate that we learn about the smaller variations associated with $\theta_{2}$. A naïve user of Isomap might plot the data in the leading three coordinates and falsely conclude that the data lies on a two-dimensional gasket. We’d like to provide an efficient method for selecting the fundamental eigen-coordinates $\phi_{1}$, $\phi_{2}$ and $\phi_{7}$, from which all others can be (nonlinearly) reconstructed; yet again, linear methods fundamentally cannot be used to select them. (a) Isomap $\phi_{1}$ (b) Isomap $\phi_{2}$ (c) Isomap $\phi_{3}$ (d) Isomap $\phi_{7}$ Figure 4. Isomap coordiantes computed from $2000$ randomly sampled points on the torus defined by Eq. 11. The leading six coordinates resemble the real and imaginary components of $e^{ik\theta_{1}}$, $k=1,2,3$, due to the rotational symmetry, providing redundant information about $\theta_{1}$ and no information about $\theta_{2}$. The fundamental coordinates $\phi_{1}$, $\phi_{2}$, and $\phi_{7}$ provide an embedding of the data that captures its toroidal structure. Linear methods cannot be used to select manifold learning eigen-coordinates for essentially the same reason why they failed on the toy models in Section 3.2: the coordinates are all mutually orthogonal as functions supported on the data! In particular, the covariance among the eigen-coordinates over the data is isotropic $\mathbb{E}[\phi_{i}(\mathbf{x})\phi_{j}(\mathbf{x})]=\frac{1}{m}\boldsymbol{\phi}_{i}^{T}\boldsymbol{\phi}_{j}=\frac{1}{m}\delta_{i,j}$ and so all sub-collections of a given size capture the same fraction of the total eigen-coordinate variance. The methods presented in the following Section 4 remedy this issue and are capable of selecting the correct set of fundamental eigen-coordinates on the torus example in Eq. 11. ## 4\. Greedy Algorithms using Secants With the failure of techniques based on linear reconstruction to select minimal collections of sensors for nonlinear reconstruction, we propose an alternative approach that relies on a collection of “secant” vectors between distinct data points. In this section, we develop this approach, yielding three related greedy selection techniques with classical theoretical guarantees on their performance. We also discuss some theoretical results that provide deterministic performance guarantees for the sensors selected by our algorithms on unseen data drawn from an underlying set. We consider a very general type of sensor placement problem that can be stated as follows. Let the set $\mathcal{X}\subset\mathbb{R}^{n}$ represent the possible states of the system and suppose that we are interested in some relevant information about the state described by a function ${\mathbf{g}}:\mathcal{X}\to\mathbb{R}^{q}$. The sensors are also described as functions of the state ${\mathbf{m}}_{j}:\mathcal{X}\to\mathbb{R}^{d_{j}}$, $j=1,\ldots,M$ where, with a slight abuse of notation, we will denote the set of all sensors and the set of all sensor indices by $\mathscr{M}$ interchangeably. Our goal is to choose a small subset of sensors $\mathscr{S}=\\{j_{1},\ldots,j_{K}\\}\subseteq\mathscr{M}$ so that the relevant information ${\mathbf{g}}({\mathbf{x}})$ about any state ${\mathbf{x}}\in\mathcal{X}$ can be recovered from the combined measurements we have selected (12) ${\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})=\left({\mathbf{m}}_{j_{1}}({\mathbf{x}}),\ldots,{\mathbf{m}}_{j_{K}}({\mathbf{x}})\right)\in\mathbb{R}^{d_{\mathscr{S}}},$ where the measurement dimension is $d_{\mathscr{S}}=\sum_{j\in\mathscr{S}}d_{j}$. That is, we want to choose $\mathscr{S}$ in such a way that there exists a reconstruction function $\boldsymbol{\Phi}_{\mathscr{S}}:\mathbb{R}^{d_{\mathscr{S}}}\to\mathbb{R}^{q}$ so that (13) ${\mathbf{g}}({\mathbf{x}})=\boldsymbol{\Phi}_{\mathscr{S}}\left({\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})\right)$ for every ${\mathbf{x}}\in\mathcal{X}$. For such a reconstruction function $\boldsymbol{\Phi}_{\mathscr{S}}$ to exist, we must meet the modest condition that any two states ${\mathbf{x}},{\mathbf{x}}^{\prime}\in\mathcal{X}$ with different target values ${\mathbf{g}}({\mathbf{x}})\neq{\mathbf{g}}({\mathbf{x}}^{\prime})$ produce different measured values ${\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})\neq{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})$. This is nothing but the vertical line test for $\boldsymbol{\Phi}_{\mathscr{S}}$, ensuring that it is a true function that does not take multiple values. However, this condition may be met for a variety of different choices of measurements $\mathscr{S}$ and we shall introduce three different ways to quantify their performance and choose among them. In these methods, the notion that $\boldsymbol{\Phi}_{\mathscr{S}}$ should not be sensitive to perturbations of the measurements is key in quantifying the performance of the sensors. The techniques we propose each rely on secants, defined below, to measure the sensitivity of $\boldsymbol{\Phi}_{\mathscr{S}}$. ###### Definition 4.1 (Secant). A _secant_ is a pair of states $({\mathbf{x}},{\mathbf{x}}^{\prime})$, where ${\mathbf{x}},{\mathbf{x}}^{\prime}\in\mathcal{X}$ and ${\mathbf{x}}\neq{\mathbf{x}}^{\prime}$. By carefully choosing the objective functions $f:2^{\mathscr{M}}\to\mathbb{R}$, we can rely on classical results by G. L. Nemhauser and L. A. Wolsey et al. [39], [63] to prove that greedy algorithms can be used to place the sensors with near-optimal performance. In particular, each objective that we propose is normalized so that $f(\emptyset)=0$, monotone non-decreasing so that $\mathscr{S}\subseteq\mathscr{S}^{\prime}$ implies $f(\mathscr{S})\leq f(\mathscr{S}^{\prime})$, and has a diminishing returns property called _submodularity_. ###### Definition 4.2 (Submodular Function). Let $\mathscr{M}$ be a finite set and denote the set of all subsets of $\mathscr{M}$ by $2^{\mathscr{M}}$. A real-valued function of the subsets $f:2^{\mathscr{M}}\to\mathbb{R}$ is called “submodular” when it has the following diminishing returns property: for any element $j\in\mathscr{M}$ and subsets $\mathscr{S},\mathscr{S}^{\prime}\subseteq\mathscr{M}$, (14) $\mathscr{S}\subseteq\mathscr{S}^{\prime}\subseteq{\mathscr{M}}\setminus\\{j\\}\quad\Rightarrow\quad f(\mathscr{S}\cup\\{j\\})-f(\mathscr{S})\geq f(\mathscr{S}^{\prime}\cup\\{j\\})-f(\mathscr{S}^{\prime}).$ That is, adding any new element $j$ to the smaller set $\mathscr{S}$ increases $f$ at least as much as adding the same element to the larger set $\mathscr{S}^{\prime}\supseteq\mathscr{S}$. Note that in applications we often do not have direct access to the full set $\mathcal{X}$, which may be continuous. Rather, we have a discrete collection of data $\mathcal{X}_{N}=\\{{\mathbf{x}}_{1},\ldots,{\mathbf{x}}_{N}\\}\subset\mathcal{X}$, which we assume is large enough to achieve suitable approximations of the underlying set. ### 4.1. Maximizing Detectable Differences As we have seen in the first half of this paper, a set of sensors can be considered good if nearby measurements come only from states whose target variables are also close together. Otherwise a small perturbation to the measurements results in a large change in the quantities of interest. One way to quantify this intuition is to select measurements that minimize the sum of squared differences in the target variables associated with states whose measurements are closer together than a fixed detection threshold $\gamma>0$, i.e., (15) $F_{\gamma}(\mathscr{S}):=\sum_{\begin{subarray}{c}{\mathbf{x}},{\mathbf{x}}^{\prime}\in\mathcal{X}_{N}\ :\\\ \|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}<\gamma\end{subarray}}\left\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\right\|_{2}^{2}.$ The length scale $\gamma$ determines how large of a difference between measurements the user deems to be significant enough to tell the two states ${\mathbf{x}}$ and ${\mathbf{x}}^{\prime}$ apart. For instance, $\gamma^{2}$ might be selected to be proportional to the noise variance using a desired number $\\#(\mathscr{S})=K$ of sensors. Let the sum of squared differences in the target variables along each secant be denoted by (16) $F_{\infty}:=\sum_{{\mathbf{x}},{\mathbf{x}}^{\prime}\in\mathcal{X}_{N}}\left\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\right\|_{2}^{2}.$ Then it is clear that minimizing the sum of squared “undetectable” differences given by Eq. 15 is equivalent to maximizing an objective function (17) $\tilde{f}_{\gamma}(\mathscr{S})=F_{\infty}-F_{\gamma}(\mathscr{S})=\sum_{{\mathbf{x}},{\mathbf{x}}^{\prime}\in\mathcal{X}_{N}}\tilde{w}_{\gamma,{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}^{2},$ where $\tilde{w}_{\gamma,{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})$ is one if $\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}\geq\gamma$ and is zero otherwise. This weight function indicates whether our measurements ${\mathbf{m}}_{\mathscr{S}}$ can distinguish the states ${\mathbf{x}}$ and ${\mathbf{x}}^{\prime}$ using the detection threshold $\gamma$, and may be written (18) $\tilde{w}_{\gamma,{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})=\mathbbm{1}\left\\{\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}\geq\gamma\right\\},$ where $\mathbbm{1}\\{A\\}=1$ if $A$ is true and $0$ if $A$ is false. Therefore, we can view the objective in Eq. 17 as the sum of squared differences that are “detectable.” Maximizing the objective in Eq. 17 over a fixed number of sensors $\\#(\mathscr{S})\leq K$ is a combinatorial optimization problem and to our knowledge does not admit an efficient direct approximation algorithm. However, if we reformulate the objective using a relaxed weight function (19) $w_{\gamma,{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})=\min\left\\{\frac{1}{\gamma^{2}}\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}^{2},\ 1\right\\},$ then (20) $\boxed{f_{\gamma}(\mathscr{S})=\sum_{{\mathbf{x}},{\mathbf{x}}^{\prime}\in\mathcal{X}_{N}}w_{\gamma,{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})\left\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\right\|_{2}^{2},}$ obtained by replacing $\tilde{w}$ with $w$ in Eq. 17, becomes a normalized, monotone, submodular function on subsets $\mathscr{S}\subseteq\mathscr{M}$ (Lemma B.3 in the Appendix) and a simple greedy approximation algorithm guarantees near-optimal performance on this problem! The greedy algorithm produces a sequence of sets $\mathscr{S}_{1},\mathscr{S}_{2},\ldots$, by starting with $\mathscr{S}_{0}=\emptyset$ and adding the sensor $j_{k}$ to $\mathscr{S}_{k-1}$ that maximizes the objective $f_{\gamma}(\mathscr{S}_{k-1}\cup\\{j\\})$ over all $j\in\mathscr{M}\setminus\mathscr{S}_{k-1}$. If $\mathscr{S}^{*}_{K}$ maximizes $f_{\gamma}(\mathscr{S})$ over all subsets of size $\\#(\mathscr{S})=K$ then the classical result of G. L. Nemhauser et al. [39] states that the objective values attained by the greedily chosen sets satisfy (21) $f_{\gamma}(\mathscr{S}_{k})\geq\left(1-e^{-k/K}\right)f_{\gamma}(\mathscr{S}^{*}_{K}),\qquad k=1,\ldots,\\#(\mathscr{M}).$ The objective function $f_{\gamma}$ given by Eq. 20 can be viewed as a “submodular relaxation” of the original sum of squared differences $\tilde{f}_{\gamma}$ given by Eq. 17. While $f_{\gamma}(\mathscr{S})\geq\tilde{f}_{\gamma}(\mathscr{S})$ for every $\mathscr{S}\subseteq\mathscr{M}$, Theorem 4.3, below, shows that $f_{\gamma}$ also provides a lower bound on $\tilde{f}_{\gamma^{\prime}}$ at reduced values of the detection threshold $\gamma^{\prime}<\gamma$. Hence, maximization of $f_{\gamma}$ is justified as a proxy for maximizing $\tilde{f}_{\gamma^{\prime}}$. Moreover, the relaxed objective bounds the total square differences among target variables that are _not detectable_ due to corresponding measurement differences smaller than reduced threshold via Eq. 23 of Theorem 4.3. ###### Theorem 4.3 (Relaxation Bound on Undetectable Differences). Consider the rigid and relaxed objectives given by Eq. 17 and Eq. 20. Then for every $\mathscr{S}\subseteq\mathscr{M}$ and constant $0<\alpha<1$, we have (22) $\tilde{f}_{\alpha\gamma}(\mathscr{S})\geq\frac{1}{1-\alpha^{2}}\left[f_{\gamma}(\mathscr{S})-\alpha^{2}F_{\infty}\right].$ Furthermore, the total fluctuation between target variables associated with states whose measurements are closer together than the reduced detection threshold $\alpha\gamma$, given by Eq. 15, is bounded above by (23) $F_{\alpha\gamma}(\mathscr{S})\leq\frac{1}{1-\alpha^{2}}\left[F_{\infty}-f_{\gamma}(\mathscr{S})\right].$ ###### Proof. We observe that (24) $\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}\geq\alpha\gamma\quad\Leftrightarrow\quad w_{\gamma,{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})\geq\alpha^{2}$ and so we have (25) $\displaystyle\tilde{w}_{\alpha\gamma,{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})$ $\displaystyle=\mathbbm{1}\left\\{\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}\geq\alpha\gamma\right\\}$ (26) $\displaystyle=\mathbbm{1}\left\\{w_{\gamma,{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})\geq\alpha^{2}\right\\}.$ Since $0\leq w_{\gamma,{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})\leq 1$, we obtain the following linear lower bound (27) $\tilde{w}_{\alpha\gamma,{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})\geq\frac{1}{1-\alpha^{2}}\left[w_{\gamma,{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})-\alpha^{2}\right].$ Summing this lower bound over all secants gives (28) $\tilde{f}_{\alpha\gamma}(\mathscr{S})\geq\frac{1}{1-\alpha^{2}}\left[f_{\gamma}(\mathscr{S})-\alpha^{2}F_{\infty}\right]$ and subtracting each side from $F_{\infty}$ yields the final result. ∎ When applied to the shock-mixing layer problem with the leading Isomap coordinates taken as the target variables ${\mathbf{g}}({\mathbf{x}})=(\phi_{1}({\mathbf{x}}),\phi_{2}({\mathbf{x}}))$, the greedy algorithm maximizing $f_{\gamma}$ first reveals the two sensor locations marked by green stars and then the black star in Figure 1 over the range of $0.02\leq\gamma\leq 0.06$. These choices produce the measurements shown in Figs. 3g and 3h, which can be used to reveal the exact phase of the system. Choosing smaller values of $\gamma$ yields different sensors that can also be used to reveal the phase, but with reduced robustness to measurement perturbations. This method of maximizing detectable differences also reveals the correct $K=3$ fundamental Isomap eigen-coordinates from among the leading $100$ on the torus example in Eq. 11 over a wide range $0.05\leq\gamma\leq 3.0$. For implementation details, see the Appendix. ### 4.2. Minimal Sensing to Meet an Error Tolerance The approach presented above relies on an average and so does not guarantee that the target value ${\mathbf{g}}({\mathbf{x}})$ can be recovered from the selected measurements ${\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})$ for every ${\mathbf{x}}\in\mathcal{X}$. In this section, we modify the technique developed above in order to provide such a guarantee by trying to find the minimum number of sensors so that every pair of states in the sampled set $\mathcal{X}_{N}$ with target values separated by at least $\varepsilon$ correspond to measurements separated by at least $\gamma$. If our sampled points $\mathcal{X}_{N}$ come sufficiently close to every point of $\mathcal{X}$ in the sense of Definition 4.4, then Proposition 4.5, given below, allows us to draw a similar conclusion about the measurements from all points in the underlying set $\mathcal{X}$. ###### Definition 4.4 ($\varepsilon_{0}$-net). An $\varepsilon_{0}$-net of $\mathcal{X}$ is a finite subset $\mathcal{X}_{N}\subset\mathcal{X}$ satisfying (29) $\forall{\mathbf{x}}\in\mathcal{X},\quad\exists{\mathbf{x}}_{i}\in\mathcal{X}_{N}\quad\mbox{such that}\quad\|{\mathbf{x}}-{\mathbf{x}}_{i}\|_{2}<\varepsilon_{0}.$ We use the subscript $N$ to denote the number of points in $\mathcal{X}_{N}$. In particular, if $\mathcal{X}_{N}$ forms a fine enough $\varepsilon_{0}$-net of $\mathcal{X}$, then Proposition 4.5 guarantees that small measurement differences never correspond to large target value differences. ###### Proposition 4.5 (Separation Guarantee on Underlying Set). Let $\mathcal{X}_{N}$ be an $\varepsilon_{0}$-net of $\mathcal{X}$ (see Definition 4.4) and let $\mathscr{S}$ be a subset of $\mathscr{M}$ satisfying (30) $\forall{\mathbf{x}}_{i},{\mathbf{x}}_{j}\in\mathcal{X}_{N}\qquad\|{\mathbf{g}}({\mathbf{x}}_{i})-{\mathbf{g}}({\mathbf{x}}_{j})\|_{2}\geq\varepsilon\quad\Rightarrow\quad\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{i})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{j})\|_{2}\geq\gamma.$ If ${\mathbf{m}}_{\mathscr{S}}$ and ${\mathbf{g}}$ are Lipschitz functions with Lipschitz constants $\|{\mathbf{m}}_{\mathscr{S}}\|_{\text{lip}}$ and $\|{\mathbf{g}}\|_{\text{lip}}$ respectively, then (31) $\forall{\mathbf{x}},{\mathbf{x}}^{\prime}\in\mathcal{X}\qquad\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}\geq\varepsilon+2\varepsilon_{0}\|{\mathbf{g}}\|_{\text{lip}}\\\ \quad\Rightarrow\quad\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}>\gamma-2\varepsilon_{0}\|{\mathbf{m}}_{\mathscr{S}}\|_{\text{lip}}.$ ###### Proof. The proof follows immediately from successive applications of the triangle inequality and so we relegate it to Appendix C ∎ Consequently, the approach described in this section allows one to reconstruct ${\mathbf{g}}({\mathbf{x}})$ from a perturbed measurement ${\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})+{\mathbf{n}}$ by taking the value ${\mathbf{g}}({\mathbf{x}}^{\prime})$ from its nearest neighbor ${\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})$ with ${\mathbf{x}}^{\prime}\in\mathcal{X}$ and achieve small error $\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}$ as long as the perturbation $\|{\mathbf{n}}\|_{2}$ is below a threshold. Supposing that the desired separation can be obtained using all of the sensors, i.e., $\mathscr{S}=\mathscr{M}$, then we can take the sum in the objective $f_{\gamma}$ given by Eq. 20 only over those pairs ${\mathbf{x}},{\mathbf{x}}^{\prime}\in\mathcal{X}_{N}$ with targets separated by at least $\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}\geq\varepsilon$, i.e., (32) $\boxed{f_{\gamma,\varepsilon}(\mathscr{S})=\sum_{\begin{subarray}{c}{\mathbf{x}},{\mathbf{x}}^{\prime}\in\mathcal{X}_{N}\ :\\\ \|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}\geq\varepsilon\end{subarray}}w_{\gamma,{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}^{2},}$ and state the problem formally as (33) $\operatorname*{\min\\!imize\enskip}_{\mathscr{S}\subseteq\mathscr{M}}\\#(\mathscr{S})\quad\mbox{subject to}\quad f_{\gamma,\varepsilon}(\mathscr{S})=f_{\gamma,\varepsilon}(\mathscr{M}).$ We observe that if all points ${\mathbf{x}},{\mathbf{x}}^{\prime}\in\mathcal{X}_{N}$ with $\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}\geq\varepsilon$ can be separated by at least $\gamma$ using $\mathscr{S}=\mathscr{M}$ then $w_{\gamma,{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{M})=1$ for each term in Eq. 32. On the other hand if there is such a pair ${\mathbf{x}},{\mathbf{x}}^{\prime}$ with $\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}<\gamma$ then that term has $w_{\gamma,{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})<1$ and $f_{\gamma,\varepsilon}(\mathscr{S})<f_{\gamma,\varepsilon}(\mathscr{M})$ as a consequence. One can show, by using the same argument as in Lemma B.3 of the Appendix, that the objective Eq. 32 is submodular in addition to being normalized and monotone non-decreasing. It follows that Eq. 33 is a classical submodular set cover problem for which a greedy algorithm maximizing $f_{\gamma,\varepsilon}$ and stopping when $f_{\gamma,\varepsilon}(\mathscr{S}_{K})=f_{\gamma,\varepsilon}(\mathscr{M})$ will always find, up to a logarithmic factor, the minimum possible number of sensors [63]. In particular, suppose that $\mathscr{S}^{*}$ is a subset of minimum size with $f_{\gamma,\varepsilon}(\mathscr{S}^{*})=f(\mathscr{M})$ and that the greedy algorithm chooses a sequence of subsets $\mathscr{S}_{1},\ldots,\mathscr{S}_{K}$ with $f_{\gamma,\varepsilon}(\mathscr{S}_{K})=f_{\gamma,\varepsilon}(\mathscr{M})$. If we define the “increment condition number” to be the ratio of the largest and smallest increments in the objective during greedy optimization (34) $\kappa=\frac{f_{\gamma,\varepsilon}(\mathscr{S}_{1})}{f_{\gamma,\varepsilon}(\mathscr{S}_{K})-f_{\gamma,\varepsilon}(\mathscr{S}_{K-1})},$ then the classical result of L. A. Wolsey [63] proves that the greedily chosen set is no larger than (35) $\\#(\mathscr{S}_{K})\leq(1+\ln{\kappa})\\#(\mathscr{S}^{*}).$ ### 4.3. Minimal Sensing to Meet an Amplification Tolerance The approaches discussed above are capable of choosing measurements that separate states with distant target values by at least a fixed distance $\gamma$. However, the nearby measurements separated by less than $\gamma$ may not adequately capture the local behavior of the target variables as illustrated by the kink in the measurements made by these sensors in the shock-mixing layer flow shown in Figure 3g. This means that while the state can be reconstructed from the measurements, its time derivative cannot. This would be a major problem if we wish to build a reduced-order model of this system based only on the fluid velocities measured at the chosen points. In addition, we may want the separation between the measurements to grow with the corresponding separation in target values, rather than potentially saturating at the $\gamma$ threshold. Attempting to select sensors $\mathscr{S}$ whose measurements capture both the local and global structure of the target variables leads us to consider disturbance amplification as a performance metric. In this section, we try to find the minimum number of sensors so that the Lipschitz constant of the reconstruction function does not exceed a user-specified threshold $L$. In practice, we do not have access to the true Lipschitz constant, so instead we bound a proxy defined below: (36) $\|\boldsymbol{\Phi}_{\mathscr{S}}\|_{\text{lip}}\approx\|\boldsymbol{\Phi}_{\mathscr{S}}\|_{\mathcal{X}_{N},\text{lip}}=\max_{{\mathbf{x}},{\mathbf{x}}^{\prime}\in\mathcal{X}_{N}}\frac{\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}}{\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}}\leq L.$ Proposition 4.6, below, shows that it suffices to enforce this condition over an $\varepsilon_{0}$-net, $\mathcal{X}_{N}$, of $\mathcal{X}$ (see Definition 4.4) in order to bound the amplification over all of $\mathcal{X}$ up to a slight relaxation for measurement differences on the same scale $\varepsilon_{0}$ as the sampling. ###### Proposition 4.6 (Amplification Guarantee on Underlying Set). Let $\mathcal{X}_{N}$ be an $\varepsilon_{0}$-net of $\mathcal{X}$ and let $\mathscr{S}$ be a subset of $\mathscr{M}$ satisfying (37) ${\mathbf{x}}_{i},{\mathbf{x}}_{j}\in\mathcal{X}_{N}\qquad\|{\mathbf{g}}({\mathbf{x}}_{i})-{\mathbf{g}}({\mathbf{x}}_{j})\|_{2}\leq L\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{i})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{j})\|_{2}.$ If ${\mathbf{m}}_{\mathscr{S}}$ and ${\mathbf{g}}$ are Lipschitz functions, with Lipschitz constants $\|{\mathbf{m}}_{\mathscr{S}}\|_{\text{lip}}$ and $\|{\mathbf{g}}\|_{\text{lip}}$ respectively, then (38) $\forall{\mathbf{x}},{\mathbf{x}}^{\prime}\in\mathcal{X}\qquad\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}<L\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})+{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}\\\ +2\left(\|{\mathbf{g}}\|_{\text{lip}}+L\|{\mathbf{m}}_{\mathscr{S}}\|_{\text{lip}}\right)\varepsilon_{0}.$ ###### Proof. The proof is a direct application of the triangle inequality and so it is relegated Appendix C. ∎ If the Lipschitz condition in Eq. 36 over $\mathcal{X}_{N}$ can be met using all of the sensors $\mathscr{S}=\mathscr{M}$ then the problem we hope to solve can be stated formally as in Eq. 33, where the condition Eq. 36 is imposed using a different normalized, monotone, submodular function (39) $\boxed{f_{L}(\mathscr{S})=\sum_{\begin{subarray}{c}{\mathbf{x}},{\mathbf{x}}^{\prime}\in\mathcal{X}_{N}\\\ {\mathbf{g}}({\mathbf{x}})\neq{\mathbf{g}}({\mathbf{x}}^{\prime})\end{subarray}}\min\left\\{\frac{\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}^{2}}{\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}^{2}},\ \frac{1}{L^{2}}\right\\}.}$ See Lemma B.4 in the Appendix for proof of these properties. We observe that if there is any secant $({\mathbf{x}},{\mathbf{x}}^{\prime})\in\mathcal{X}_{N}\times\mathcal{X}_{N}$ for which Eq. 36 is not satisfied for a given $\mathscr{S}\subset\mathscr{M}$, then the corresponding term of Eq. 39 is less than $1/L^{2}$ and $f_{L}(\mathscr{S})<f_{L}(\mathscr{M})$. Otherwise, each term of Eq. 39 is $1/L^{2}$ and we have $f_{L}(\mathscr{S})=f_{L}(\mathscr{M})$. Again, the classical result in [63] shows that a greedy approximation algorithm maximizing Eq. 39 and stopping when $f_{L}(\mathscr{S}_{K})=f_{L}(\mathscr{M})$ finds the minimum possible number of sensors up to a logarithmic factor so that the Lipschitz condition Eq. 36 is satisfied. In particular, the same guarantee stated in Eq. 35 holds for the Lipschitz objective too. In some applications, we may instead want to find the measurements that minimize the reconstruction Lipschitz constant $\|\boldsymbol{\Phi}_{\mathscr{S}}\|_{\mathcal{X}_{N},\text{lip}}$ using a fixed sensor budget $\\#(\mathscr{S})\leq C$. By running the greedy algorithm repeatedly using different thresholds $L$ it is possible to obtain upper and sometimes lower bounds on this budget-constrained minimum Lipschitz constant $L^{*}$. This idea is closely related to the approach of [27]. If the greedy algorithm using Lipschitz constant $L$ chooses sensors $\mathscr{S}$ that meet the budget $\\#(\mathscr{S})\leq C$ then $L$ is obviously an upper bound on $L^{*}$. In practice, we can use a bisection search over $L$ to find nearly the smallest $L$ to any given tolerance for which $\\#(\mathscr{S})\leq C$. To get the lower bound, the greedy algorithm is run with a small enough $L$ so that the bound on the minimum possible cost from Eq. 35 exceeds the budget (40) $C<\\#(\mathscr{S})/(1+\ln\kappa).$ If this is the case, there is no collection of measurements with amplification at most $L$ that meets the cost constraint. Thus, such an $L$ is a lower bound on the minimum possible amplification using measurement budget $C$. Again, bisection search can be used to find nearly the largest $L$ so that $C<\\#(\mathscr{S})/(1+\ln\kappa)$. With the leading Isomap coordinates taken as the target variables ${\mathbf{g}}({\mathbf{x}})=(\phi_{1}({\mathbf{x}}),\phi_{2}({\mathbf{x}}))$, a bisection search over $L$ identifies the three sensor locations marked by black squares in Figure 1 on the shock-mixing layer problem and the correct fundamental Isomap eigenfunctions $\phi_{1},\phi_{2},\phi_{7}$ on the torus example in Eq. 11. The measurements made by these sensors on the shock-mixing layer problem are shown in Figure 3i and indicate, by the lack of self- intersections, that they can be used to recover the phase. The minimum number of sensors selected by the greedy algorithm that allow one to reconstruct both the relevant information ${\mathbf{g}}({\mathbf{x}})$ and its time derivative is usually persistent over a wide range of Lipschitz constants with fewer sensors not being chosen until $L$ is made extremely large. In the shock-mixing layer problem, three sensors that successfully reveal the underlying phase are found for values of $L$ ranging from $1868$ to $47624$, above which only two sensors that cannot reveal the underlying phase are selected. The fact that a smaller set of inadequate sensors are selected for extremely large $L$ reflects our use of a discrete approximation $\mathcal{X}_{N}$ of the continuous set $\mathcal{X}$. Measurements from $\mathcal{X}_{N}$ will almost never truly overlap to give $\|\boldsymbol{\Phi}_{\mathscr{S}}\|_{\text{lip}}=\infty$ as they would for measurements from $\mathcal{X}$. We also find that with $L=129$, the minimum possible number of sensors exceeds $\\#(\mathscr{S}_{K})/(1+\ln{\kappa})=3.18>3$ on the shock-mixing layer problem. Therefore, the minimum possible reconstruction Lipschitz constant using three sensors that one might find by an exhaustive search over the $\binom{2210}{3}\approx 1.8\times 10^{9}$ possible combinations must be greater than $129$. For implementation details, see the Appendix. ## 5\. Computational Considerations and Down-Sampling So far, the three secant-based methods we presented involve objectives that sum over $\mathcal{O}(N^{2})$ pairs of points from the sampled set $\mathcal{X}_{N}$. In this section, we discuss how this large collection of secants can be sub-sampled to produce high-probability performance guarantees using a number of secants that scales more favorably with the size of the data set. By sub-sampling we do pay a price in the sense that some “bad” secants may escape our sampling scheme and so we cannot draw the same conclusions about every point in the underlying set as we did in Propositions 4.5 and 4.6 for the sensors chosen using the methods in Sections 4.2 and 4.3. Instead, we can bound the size of the set of these “bad” secants with high probability by using a sampled collection of secants that scales linearly with $N$. In the case of the total detectable difference-based objective discussed in Section 4.1, we can prove high-probability bounds for the sum of squared undetectable differences in the target variables using a constant number of secants that doesn’t depend on $N$ at all. Before getting started with our discussion of down-sampling, let us first mention that the calculation of each of the objectives formulated in Section 4 is easily parallelizable, whether or not they are down-sampled. Even though the computation of each objective function given by Eq. 20, 32, or 39 requires $\mathcal{O}(N^{2})$ operations, the terms being summed can be distributed among many processors without the need for any communication except at the end when each processor reports the sum over the secants allocated to it. Furthermore, because each secant-based objective we consider in this paper is submodular, it is not actually necessary to evaluate the objectives over all of the remaining sensors during each step of the greedy algorithm. By employing the “accelerated greedy” algorithm of M. Minoux [35], the same set of sensors can be found using a minimal number of evaluations of the objective. We provide a summary of the accelerated greedy algorithm in Section D of the Appendix. The computational cost of evaluating the objectives in Sections 4.2 and 4.3 during each step of the greedy algorithm may also be reduced by exploiting the fact that each term in the sum is truncated once the measurements achieve a certain level of separation. This means that only the nearest neighbors within a known distance of each ${\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})$, ${\mathbf{x}}\in\mathcal{X}_{N}$ need to be computed and rest of the terms all achieve the threshold and need not be computed explicitly. To compute the sum efficiently, fixed-radius near neighbors algorithms [2], [3] could be employed. ### 5.1. Maximizing Detectable Differences The main results of this section are Theorems 5.2 and 5.3, which show that with high probability we can obtain guaranteed performance in terms of mean undetectable differences by sampling a constant number of secants (i.e., independent of $N$) selected at random. In particular, Theorem 5.2 bounds the worst-case performance of the greedy algorithm with high probability using the sampled objective. Theorem 5.3, on the other hand, shows that if one only considers randomly sampled secants with target variables separated by at least $\varepsilon$ (see Section 4.2), then the mean square undetectable difference between target values is less than $2\varepsilon^{2}$ with high probability. While the original mean square fluctuation objective in Eq. 20 was formulated over the discrete set $\mathcal{X}_{N}$, we can actually prove more versatile approximation results about an objective defined as an average over the entire, possibly continuous, set $\mathcal{X}$ with respect to a probability measure $\mu$. In particular, we assume the target variables ${\mathbf{g}}$ and measurements ${\mathbf{m}}_{j}$, $j\in\mathscr{M}$ are measurable functions on $\mathcal{X}$ and consider an average detectable difference objective (41) $f_{\gamma}(\mathscr{S})=\int_{\mathcal{X}\times\mathcal{X}}w_{\gamma,{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})\left\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\right\|_{2}^{2}\ d\mu({\mathbf{x}})d\mu({\mathbf{x}}^{\prime})$ with $w_{\gamma,{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})$ defined by Eq. 19. We also denote the average fluctuations between target variables associated with states whose measurements are closer together than the detection threshold $\gamma$ by (42) $F_{\gamma}(\mathscr{S}):=\int_{\begin{subarray}{c}({\mathbf{x}},{\mathbf{x}}^{\prime})\in\mathcal{X}\times\mathcal{X}\ :\\\ \|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}<\gamma\end{subarray}}\left\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\right\|_{2}^{2}\ d\mu({\mathbf{x}})d\mu({\mathbf{x}}^{\prime})$ and the total fluctuation among target variables by (43) $F_{\infty}:=\int_{\mathcal{X}\times\mathcal{X}}\left\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\right\|_{2}^{2}\ d\mu({\mathbf{x}})d\mu({\mathbf{x}}^{\prime}).$ Note that the original objective formulated in Section 4.1 as well as Eq. 15 are special cases of Eq. 41 and Eq. 42, up to an irrelevant constant factor, when $\mu=\frac{1}{N}\sum_{{\mathbf{x}}\in\mathcal{X}_{N}}\delta_{{\mathbf{x}}}$ and $\delta_{{\mathbf{x}}}(A)=\mathbbm{1}\\{{\mathbf{x}}\in A\\}$ is the Dirac measure on Borel sets $A\subseteq\mathcal{X}$. By Lemma B.3, Eq. 41 is submodular in addition to being normalized and monotone non-decreasing. Furthermore, by an identical argument to Theorem 4.3, we know that the mean square fluctuation between target variables associated with states whose measurements are closer together than a reduced detection thereshold $\alpha\gamma$ with $0<\alpha<1$ is bounded above by (44) $F_{\alpha\gamma}(\mathscr{S})\leq\frac{1}{1-\alpha^{2}}\left[F_{\infty}-f_{\gamma}(\mathscr{S})\right].$ We begin with Lemma 5.1, which shows that by sampling a large enough collection of points ${\mathbf{x}}_{1},{\mathbf{x}}^{\prime}_{1},\ldots,{\mathbf{x}}_{m},{\mathbf{x}}^{\prime}_{m}\in\mathcal{X}$ independently according to $\mu$, the objective $f_{\gamma}$ can be uniformly approximated by a sample-based average (45) $f_{\gamma,m}(\mathscr{S})=\frac{1}{m}\sum_{i=1}^{m}w_{\gamma,{\mathbf{x}}_{i},{\mathbf{x}}_{i}^{\prime}}(\mathscr{S})\left\|{\mathbf{g}}({\mathbf{x}}_{i})-{\mathbf{g}}({\mathbf{x}}_{i}^{\prime})\right\|_{2}^{2}$ over all $\mathscr{S}\subseteq\mathscr{M}$ of size $\\#(\mathscr{S})\leq L$ with high probability over the sample points. Most importantly, the number of sample points needed for this approximation guarantee is independent of the distribution $\mu$. Consequently if we have access to $N$ points making up $\mathcal{X}_{N}$ that have been sampled independently according to $\mu$, we need only keep the first $2m$ of them to accurately approximate the objective. The number $m$ of such sub-sampled points depends only on the quality of the probabilistic guarantee and not on the size of the data set $N$. ###### Lemma 5.1 (Accuracy of the Down-Sampled Objective). Consider the objectives $f_{\gamma}$ and $f_{\gamma,m}$ defined according to Eq. 41 and Eq. 45. Assume that the target function is bounded over $\mathcal{X}$ so that (46) $D=\operatorname{diam}{\mathbf{g}}(\mathcal{X})=\sup_{{\mathbf{x}},{\mathbf{x}}^{\prime}\in\mathcal{X}}\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}<\infty.$ and that ${\mathbf{x}}_{1},{\mathbf{x}}^{\prime}_{1},\ldots,{\mathbf{x}}_{m},{\mathbf{x}}^{\prime}_{m}\in\mathcal{X}$ are sampled independently according to a probability measure $\mu$ on $\mathcal{X}$. If the number of sampled pairs is at least (47) $m\geq\frac{D^{4}}{2\varepsilon^{2}}\left[L\ln{\\#(\mathscr{M})}-\ln{\left((L-1)!\right)}-\ln{\left(\frac{p}{2}\right)}\right],$ then $|f_{\gamma,m}(\mathscr{S})-f_{\gamma}(\mathscr{S})|<\varepsilon$ for every $\mathscr{S}\subseteq\mathscr{M}$ of size $\\#(\mathscr{S})\leq L$ with probability at least $1-p$. ###### Proof. For simplicity, we will drop $\gamma$ from the subscripts on our objectives since $\gamma$ remains fixed throughout the proof. Let us begin by fixing a set $\mathscr{S}\subseteq\mathscr{M}$ of size $\\#(\mathscr{S})\leq L$ and denoting $M=\\#(\mathscr{M})$ for short. Under the assumption that the points ${\mathbf{x}}_{i},{\mathbf{x}}^{\prime}_{i}$ are sampled independently and identically under $\mu$, the random variables (48) $Z_{i}(\mathscr{S})=w_{{\mathbf{x}}_{i},{\mathbf{x}}_{i}^{\prime}}(\mathscr{S})\|{\mathbf{g}}({\mathbf{x}}_{i})-{\mathbf{g}}({\mathbf{x}}_{i}^{\prime})\|_{2}^{2},\quad i=1,\ldots,m,$ are independent and bounded by $0\leq Z_{i}(\mathscr{S})\leq D^{2}$. The value of the optimization objective is the expectation $f(\mathscr{S})=\mathbb{E}[Z_{i}(\mathscr{S})]$ and the value of our sub- sampled objective is the empirical average (49) $f_{m}(\mathscr{S})=\frac{1}{m}\sum_{i=1}^{m}Z_{i}(\mathscr{S}).$ Hoeffding’s inequality allows us to bound the probability that $f_{m}(\mathscr{S})$ differs from $f(\mathscr{S})$ by more than $\varepsilon$ according to (50) $\mathbb{P}\left\\{\left|f_{m}(\mathscr{S})-f(\mathscr{S})\right|\geq\varepsilon\right\\}\leq 2\exp\left(-\frac{2m\varepsilon^{2}}{D^{4}}\right).$ We want the objective to be accurately approximated with tolerance $\varepsilon$ uniformly over all collections of sensors of size $\\#(\mathscr{S})\leq L$. We unfix $\mathscr{S}$ by taking the union bound (51) $\mathbb{P}\bigcup_{\begin{subarray}{c}\mathscr{S}\subseteq\mathscr{M}:\\\ \\#(\mathscr{S})\leq L\end{subarray}}\left\\{\left|f_{m}(\mathscr{S})-f(\mathscr{S})\right|\geq\varepsilon\right\\}\leq\sum_{\begin{subarray}{c}\mathscr{S}\subseteq\mathscr{M}:\\\ \\#(\mathscr{S})\leq L\end{subarray}}2\exp\left(-\frac{2m\varepsilon^{2}}{D^{4}}\right).$ The combinatorial inequality (52) $\\#\left(\\{\mathscr{S}\subseteq\mathscr{M}\ :\ \\#(\mathscr{S})\leq L\\}\right)=\sum_{k=1}^{L}\binom{M}{k}\leq\sum_{k=1}^{L}\frac{M^{k}}{k!}\leq L\frac{M^{L}}{L!}=\frac{M^{L}}{(L-1)!}$ yields the bound (53) $\mathbb{P}\bigcup_{\begin{subarray}{c}\mathscr{S}\subseteq\mathscr{M}:\\\ \\#(\mathscr{S})\leq L\end{subarray}}\left\\{\left|f_{m}(\mathscr{S})-f(\mathscr{S})\right|\geq\varepsilon\right\\}\leq 2\exp\left(L\ln{M}-\ln{\left((L-1)!\right)}-\frac{2m\varepsilon^{2}}{D^{4}}\right)\leq p$ when the number of sampled pairs ${\mathbf{x}}_{i},{\mathbf{x}}_{i}^{\prime}$ satisfies Eq. 47. ∎ The uniform accuracy of the sampled objective $f_{\gamma,m}$ over the feasible subsets $\mathscr{S}$ in our optimization problem (54) $\operatorname*{\max\\!imize\enskip}_{\begin{subarray}{c}\mathscr{S}\subseteq\mathscr{M}\ :\ \\#(\mathscr{S})\leq K\end{subarray}}f_{\gamma}(\mathscr{S})$ established in Lemma 5.1 leads to performance guarantees for the greedy approximation algorithm when the sampled objective $f_{\gamma.m}$ is used in place of $f_{\gamma}$. In particular, Theorem 5.2 shows that the greedy algorithm can be applied to the sampled objective Eq. 45 and still achieve near-optimal performance with respect to the original objective Eq. 41 on the underlying set $\mathcal{X}$ with high probability. This sampling-based approach therefore completely eliminates the $\mathcal{O}(N^{2})$ dependence of the computational complexity involved in evaluating the objective at a penalty on the worst case performance that can be made arbitrarily small by sampling more points. ###### Theorem 5.2 (Greedy Performance using Sampled Objective). Assume the same hypotheses as Lemma 5.1 and let $\mathscr{S}^{*}$ denote an optimal solution of (55) $\operatorname*{\max\\!imize\enskip}_{\begin{subarray}{c}\mathscr{S}\subseteq\mathscr{M}\ :\ \\#(\mathscr{S})\leq K\end{subarray}}f_{\gamma}(\mathscr{S}),$ with $f_{\gamma}$ given by Eq. 41 and $K\leq L$. If $\mathscr{S}_{1},\ldots,\mathscr{S}_{L}$ are the sequence of subsets selected by the greedy algorithm using the sampled objective $f_{\gamma,m}$ given by Eq. 45, then (56) $f_{\gamma}(\mathscr{S}_{k})\geq\left(1-e^{-k/K}\right)f_{\gamma}(\mathscr{S}^{*})-\left(2-e^{-k/K}\right)\varepsilon,\qquad k=1,\ldots,L,$ with probability at least $1-p$ over the sample points. ###### Proof. For simplicity, we will drop $\gamma$ from the subscripts on our objectives since $\gamma$ remains fixed throughout the proof. Let $\mathscr{S}_{m}^{*}$ denote the optimal solution of (57) $\operatorname*{\max\\!imize\enskip}_{\begin{subarray}{c}\mathscr{S}\subseteq\mathscr{M}\ :\ \\#(\mathscr{S})\leq K\end{subarray}}f_{m}(\mathscr{S}),$ using the sampled objective and assume that $|f(\mathscr{S})-f_{m}(\mathscr{S})|<\varepsilon$ for every subset $\mathscr{S}$ of $\mathscr{M}$ with $\\#(\mathscr{S})\leq L$. According to Lemma 5.1, this happens with probability at least $1-p$ over the sample points. Using this uniform approximation and the guarantee on the performance of the greedy algorithm for $f_{m}$, we have (58) $f(\mathscr{S}_{k})\geq f_{m}(\mathscr{S}_{k})-\varepsilon\geq\left(1-e^{-k/K}\right)f_{m}(\mathscr{S}_{m}^{*})-\varepsilon.$ Since $\mathscr{S}_{m}^{*}$ is the optimal solution using the sampled objective, we must have $f_{m}(\mathscr{S}_{m}^{*})\geq f_{m}(\mathscr{S}^{*})$. Using this fact and the uniform approximation gives (59) $\displaystyle f(\mathscr{S}_{k})$ $\displaystyle\geq\left(1-e^{-k/K}\right)f_{m}(\mathscr{S}^{*})-\varepsilon$ (60) $\displaystyle\geq\left(1-e^{-k/K}\right)\left(f(\mathscr{S}^{*})-\varepsilon\right)-\varepsilon.$ Combining the terms on $\varepsilon$ completes the proof. ∎ ###### Remark 5.1. While Theorem 5.2 tells us that down-sampling has a small effect on the worst- case performance of the greedy algorithm, unfortunately, we cannot say much beyond that. It may be the case that the greedy solution using the sampled objective $f_{\gamma,m}$ produces a very different value of $f_{\gamma}$ than the greedy solution using $f_{\gamma}$ directly, even though these functions are both submodular and differ by no more than an arbitrarily small $\varepsilon>0$. Consider the following example in Table 1 where we have two submodular objectives, $f$ and $\tilde{f}$, that differ by no more than $\varepsilon\ll 1$, yet the greedy algorithm applied to $f$ and $\tilde{f}$ yield results that differ by $\mathcal{O}(1)$. $\mathscr{S}$ | $f(\mathscr{S})$ | $\tilde{f}(\mathscr{S})$ ---|---|--- $\emptyset$ | $0$ | $0$ $\\{a\\}$ | $2+\varepsilon$ | $2$ $\\{b\\}$ | $2$ | $2+\varepsilon$ $\\{c\\}$ | $1$ | $1$ $\\{a,b\\}$ | $2+\varepsilon$ | $2+2\varepsilon$ $\\{a,c\\}$ | $3+\varepsilon$ | $3$ $\\{b,c\\}$ | $2$ | $2+\varepsilon$ $\\{a,b,c\\}$ | $3$ | $3$ Table 1. Two submodular functions are given that differ by no more than $\varepsilon\ll 1$, yet produce very different greedy solutions and objective values. One can easily verify that both functions in Table 1 are normalized, monotone, and submodular. When selecting subsets of size $2$, the greedy algorithm for $f$ picks $\emptyset\to\\{a\\}\to\\{a,c\\}$ and the greedy algorithm for for $\tilde{f}$ picks $\emptyset\to\\{b\\}\to\\{a,b\\}$. The values of $f$ on the chosen sets, $f(\\{a,c\\})=3+\varepsilon$ and $f(\\{a,b\\})=2+2\varepsilon$, differ by $1-\varepsilon\gg\varepsilon$, and similarly for $\tilde{f}(\\{a,c\\})=3$ and $\tilde{f}(\\{a,b\\})=2+\varepsilon$, which also differ by $1-\varepsilon\gg\varepsilon$. Thus the performance of the greedy algorithm can be sensitive to small perturbations of the objective even though the lower bound on performance is not sensitive. It turns out that by solving the error tolerance problem in Section 4.2 greedily using a down-sampled objective, we can provide high probability bounds directly on the mean square undetectable differences in Eq. 42. We will use the down-sampled objective (61) $f_{\gamma,\varepsilon,m}(\mathscr{S})=\frac{1}{m}\sum_{\begin{subarray}{c}i\in\\{1,\ldots,m\\}:\\\ \|{\mathbf{g}}({\mathbf{x}}_{i})-{\mathbf{g}}({\mathbf{x}}^{\prime}_{i})\|_{2}\geq\varepsilon\end{subarray}}w_{\gamma,{\mathbf{x}}_{i},{\mathbf{x}}^{\prime}_{i}}(\mathscr{S})\|{\mathbf{g}}({\mathbf{x}}_{i})-{\mathbf{g}}({\mathbf{x}}^{\prime}_{i})\|_{2}^{2},$ with the relaxed weight function in Eq.19 in a greedy approximation algorithm for the submodular set-cover problem (62) $\operatorname*{\min\\!imize\enskip}_{\mathscr{S}\subseteq\mathscr{M}}\\#(\mathscr{S})\quad\mbox{subject to}\quad f_{\gamma,\varepsilon,m}(\mathscr{S})=f_{\gamma,\varepsilon,m}(\mathscr{M}).$ Using the resulting greedy solution $\mathscr{S}_{K}$ that satisfies $f_{\gamma,\varepsilon,m}(\mathscr{S}_{K})=f_{\gamma,\varepsilon,m}(\mathscr{M})=\tilde{f}_{\gamma,\varepsilon,m}(\mathscr{M})$, Theorem 5.3 provides a high-probability bound on the mean square undetectable difference in the target variables, Eq. 42, over the entire set $\mathcal{X}\times\mathcal{X}$ rather than merely $\mathcal{X}_{N}\times\mathcal{X}_{N}$. ###### Theorem 5.3 (Sample Separation Bound on Undetectable Differences). Consider the functions $f_{\gamma,\varepsilon,m}$ and $F_{\gamma}$ defined by Eqs. 62 and 42 and assume that the condition $\|{\mathbf{m}}_{\mathscr{M}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{M}}({\mathbf{x}}^{\prime})\|_{2}\geq\gamma$ holds for $\mu$-almost every ${\mathbf{x}},{\mathbf{x}}^{\prime}\in\mathcal{X}$ such that $\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}\geq\varepsilon$. Suppose that the target function is bounded over $\mathcal{X}$ so that (63) $D=\operatorname{diam}{\mathbf{g}}(\mathcal{X})=\sup_{{\mathbf{x}},{\mathbf{x}}^{\prime}\in\mathcal{X}}\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}<\infty.$ and that ${\mathbf{x}}_{1},{\mathbf{x}}^{\prime}_{1},\ldots,{\mathbf{x}}_{m},{\mathbf{x}}^{\prime}_{m}\in\mathcal{X}$ are sampled independently according to the probability measure $\mu$ on $\mathcal{X}$. If the number of sampled pairs is at least (64) $m\geq\frac{D^{4}}{2\varepsilon^{4}}\left(\\#(\mathscr{M})\ln 2-\ln{p}\right),$ and the greedy approximation of Eq. 62 produces a set $\mathscr{S}_{K}$, then (65) $F_{\gamma}(\mathscr{S}_{K})<2\varepsilon^{2}$ with probability at least $1-p$. ###### Proof. For simplicity, we will drop $\gamma,\varepsilon$ from the subscripts on our objectives since $\gamma$ and $\varepsilon$ remain fixed throughout the proof. Let (66) $\mathcal{D}=\left\\{({\mathbf{x}},{\mathbf{x}}^{\prime})\in\mathcal{X}\times\mathcal{X}\ :\ \|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}\geq\varepsilon\right\\}$ and (67) $\tilde{f}(\mathscr{S})=\mathbb{E}\left[\tilde{f}_{m}(\mathscr{S})\right]\\\ =\int_{\mathcal{X}\times\mathcal{X}}\chi_{\mathcal{D}}({\mathbf{x}},{\mathbf{x}}^{\prime})\tilde{w}_{\gamma,{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}^{2}\ d\mu({\mathbf{x}})d\mu({\mathbf{x}}^{\prime}),$ where $\chi_{\mathcal{D}}$ is the characteristic function of the set $\mathcal{D}$. From our assumption that $\|{\mathbf{m}}_{\mathscr{M}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{M}}({\mathbf{x}})\|_{2}\geq\gamma$ for $\mu$-almost every ${\mathbf{x}},{\mathbf{x}}^{\prime}\in\mathcal{X}$ with $\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}\geq\varepsilon$, it follows (68) $\tilde{f}(\mathscr{M})=\int_{\mathcal{X}\times\mathcal{X}}\chi_{\mathcal{D}}({\mathbf{x}},{\mathbf{x}}^{\prime})\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}^{2}\ d\mu({\mathbf{x}})d\mu({\mathbf{x}}^{\prime}).$ Expanding our definition of $F_{\gamma}$ in Eq.42, we find (69) $F_{\gamma}(\mathscr{S})=\tilde{f}(\mathscr{M})-\tilde{f}(\mathscr{S})\\\ +\int_{\mathcal{X}\times\mathcal{X}}\chi_{\mathcal{D}^{c}}({\mathbf{x}},{\mathbf{x}}^{\prime})\left[1-\tilde{w}_{\gamma,{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})\right]\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}^{2}\ d\mu({\mathbf{x}})d\mu({\mathbf{x}}^{\prime})$ and therefore (70) $F_{\gamma}(\mathscr{S})\leq\tilde{f}(\mathscr{M})-\tilde{f}(\mathscr{S})+\varepsilon^{2}.$ We shall now use a similar Hoeffding and union bound argument as in Thm. 5.1 to relate $\tilde{f}(\mathscr{M})-\tilde{f}(\mathscr{S})$ to $\tilde{f}_{m}(\mathscr{M})-\tilde{f}_{m}(\mathscr{S})$ uniformly over every subset $\mathscr{S}\subseteq\mathscr{M}$. Fixing such $\mathscr{S}\subset\mathscr{M}$, the one-sided Hoeffding inequality tells us that (71) $\mathbb{P}\left\\{\left[\tilde{f}(\mathscr{M})-\tilde{f}(\mathscr{S})\right]-\left[\tilde{f}_{m}(\mathscr{M})-\tilde{f}_{m}(\mathscr{S})\right]\geq\varepsilon^{2}\right\\}\leq\exp{\left(-\frac{2m\varepsilon^{4}}{D^{4}}\right)}.$ Unfixing $\mathscr{S}$ using the union bound tells us that (72) $\tilde{f}(\mathscr{M})-\tilde{f}(\mathscr{S})<\tilde{f}_{m}(\mathscr{M})-\tilde{f}_{m}(\mathscr{S})+\varepsilon^{2}$ uniformly over all $\mathscr{S}\subset\mathscr{M}$ with probability at least $1-p$. Since the greedy algorithm terminates when $\tilde{f}_{m}(\mathscr{S}_{K})=\tilde{f}_{m}(\mathscr{M})$, it follows by substitution into Eq. 70 that (73) $F_{\gamma}(\mathscr{S})<2\varepsilon^{2}$ with probability at least $1-p$ over the sample points. ∎ ### 5.2. Minimal Sensing to Meet Separation or Amplification Tolerances If we want to draw stronger conclusions about the underlying set $\mathcal{X}$ than are captured by the mean square (un)detectable differences, then we must increase the number of sample points. The following Theorems 5.4 and 5.5 show that similar conclusions about the separation of points as in Propositions 4.5 and 4.6 can be achieved over large subsets of $\mathcal{X}$ with high probability by considering secants between a randomly chosen set of “base points” and an the full data set. More precisely, we will consider secants between an $\varepsilon_{0}$-net $\mathcal{X}_{N}$ of $\mathcal{X}$ and a collection of base point $\mathcal{B}_{m}\subset\mathcal{X}$ with size $m$ independent of $N$. This leads to linear $\mathcal{O}(N)$ scaling of the cost to evaluate the down-sampled versions of the objectives given by Eqs. 32 and 39 in Sections 4.2 and 4.3 to achieve these relaxed guarantees. The strong guarantee of Proposition 4.5 requires that we use an objective like Eq. 32 in the submodular set-cover problem Eq. 33 where the sum in Eq. 32 is taken over $\mathcal{X}_{N}\times\mathcal{X}_{N}$ and $\mathcal{X}_{N}$ is an $\varepsilon_{0}$-net of the underlying set $\mathcal{X}$. The problem is that the $\varepsilon_{0}$-net $\mathcal{X}_{N}$ may be quite large and the number of operations needed to evaluate the sum in the objective scales with the square of the size of $\mathcal{X}_{N}$. Here we will prove that a similar guarantee as in Proposition 4.5 holds with high probability over a large subset of $\mathcal{X}$ when the sum in Eq. 32 is taken over secants between a randomly chosen collection of base points $\mathcal{B}_{m}=\\{{\mathbf{b}}_{1},\ldots,{\mathbf{b}}_{m}\\}\subseteq\mathcal{X}$ and the $\varepsilon_{0}$-net $\mathcal{X}_{N}$. Most importantly, the number of base points depends on the quality of the guarantee and not on size of the $\varepsilon_{0}$-net, so that the computational cost can be reduced to linear dependence on the size of $\mathcal{X}_{N}$. Specifically, in place of Eq. 32, we can consider the sampled objective (74) $f_{\gamma,\varepsilon,m}(\mathscr{S})=\frac{1}{mN}\sum_{\begin{subarray}{c}1\leq i\leq m,\ 1\leq j\leq N:\\\ \|{\mathbf{g}}({\mathbf{b}}_{i})-{\mathbf{g}}({\mathbf{x}}_{j})\|_{2}\geq\varepsilon\end{subarray}}w_{\gamma,{\mathbf{b}}_{i},{\mathbf{x}}_{j}}(\mathscr{S})\|{\mathbf{g}}({\mathbf{b}}_{i})-{\mathbf{g}}({\mathbf{x}}_{j})\|_{2}^{2}$ with $w_{\gamma,{\mathbf{b}}_{i},{\mathbf{x}}_{j}}(\mathscr{S})$ defined by Eq. 19 in the optimization problem Eq. 33. The greedy approximation algorithm produces a set of sensors $\mathscr{S}_{K}$ such that (75) $\|{\mathbf{g}}({\mathbf{b}}_{i})-{\mathbf{g}}({\mathbf{x}}_{j})\|_{2}\geq\varepsilon\quad\Rightarrow\quad\|{\mathbf{m}}_{\mathscr{S}_{K}}({\mathbf{b}}_{i})-{\mathbf{m}}_{\mathscr{S}_{K}}({\mathbf{x}}_{j})\|_{2}\geq\gamma$ for every ${\mathbf{b}}_{i}\in\mathcal{B}_{m}$ and ${\mathbf{x}}_{j}\in\mathcal{X}_{N}$. Theorem 5.4 guarantees that with high probability, only a small subset of points in $\mathcal{X}$ have target values that cannot be distinguished from the rest by measurements separated by a relaxed detection threshold. This size of this “bad set” is determined by its $\mu$-measure, which can be made arbitrarily small with high probability by taking more sample base points $m$. ###### Theorem 5.4 (Sampled Separation Guarantee). Let $\mathcal{X}_{N}$ be an $\varepsilon_{0}$-net of $\mathcal{X}$ and let the base points $\mathcal{B}_{m}$ be sampled independently according to a probability measure $\mu$ on $\mathcal{X}$ with (76) $m\geq\frac{1}{2\delta^{2}}\left(\\#(\mathscr{M})\ln{2}-\ln{p}\right),$ where $p,\delta\in(0,1)$. Consider the objective $f_{\gamma,\varepsilon,m}$ given by Eq. 74 for a certain choice of $\gamma>0$ and $\varepsilon>0$ for which every ${\mathbf{b}}_{i}\in\mathcal{B}_{m}$ and ${\mathbf{x}}_{j}\in\mathcal{X}_{N}$ satisfies (77) $\|{\mathbf{g}}({\mathbf{b}}_{i})-{\mathbf{g}}({\mathbf{x}}_{j})\|_{2}\geq\varepsilon\quad\Rightarrow\quad\|{\mathbf{m}}_{\mathscr{M}}({\mathbf{b}}_{i})-{\mathbf{m}}_{\mathscr{M}}({\mathbf{x}}_{j})\|_{2}\geq\gamma.$ Suppose also that ${\mathbf{g}}$ and the measurement functions ${\mathbf{m}}_{k}$, $k\in\mathscr{M}$ are all Lipschitz over $\mathcal{X}$. If $f_{\gamma,\varepsilon,m}(\mathscr{S})=f_{\gamma,\varepsilon,m}(\mathscr{M})$, then the $\mu$ measure of points ${\mathbf{x}}\in\mathcal{X}$ such that (78) $\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}\geq\varepsilon+\varepsilon_{0}\|{\mathbf{g}}\|_{\text{lip}}\\\ \Rightarrow\quad\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}>\gamma-\varepsilon_{0}\|{\mathbf{m}}_{\mathscr{S}}\|_{\text{lip}}$ for every ${\mathbf{x}}^{\prime}\in\mathcal{X}$ is at least $1-\delta$ with probability at least $1-p$. ###### Proof. For simplicity, we will drop $\gamma,\varepsilon$ from the subscript on our objective since $\gamma$ and $\varepsilon$ remain fixed throughout the proof. Let us begin by fixing a set $\mathscr{S}\subseteq\mathscr{M}$ and define the random variables (79) $Z_{\mathscr{S}}({\mathbf{b}}_{i})=\max_{{\mathbf{x}}\in\mathcal{X}_{N}}\mathbbm{1}\big{\\{}\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{b}}_{i})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})\|_{2}<\gamma\quad\mbox{and}\quad\|{\mathbf{g}}({\mathbf{b}}_{i})-{\mathbf{g}}({\mathbf{x}})\|_{2}\geq\varepsilon\big{\\}}.$ If $Z_{\mathscr{S}}({\mathbf{b}}_{i})=0$ then every ${\mathbf{x}}\in\mathcal{X}_{N}$ with $\|{\mathbf{g}}({\mathbf{b}}_{i})-{\mathbf{g}}({\mathbf{x}})\|_{2}\geq\varepsilon$ also satisfies $\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{b}}_{i})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})\|_{2}\geq\gamma$, otherwise $Z_{\mathscr{S}}({\mathbf{b}}_{i})=1$. We observe that $Z_{\mathscr{S}}({\mathbf{b}}_{i})$, $i=1,\ldots,m$ are independent, identically distributed Bernoulli random variables whose expectation (80) $\mathbb{E}\left[Z_{\mathscr{S}}({\mathbf{b}}_{i})\right]=\mu\big{(}\big{\\{}{\mathbf{x}}\in\mathcal{X}\ :\ \exists{\mathbf{x}}^{\prime}\in\mathcal{X}_{N}\quad\mbox{s.t.}\\\ \|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}<\gamma,\quad\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}\geq\varepsilon\big{\\}}\big{)}$ is the $\mu$-measure of points in $\mathcal{X}$ for which target values differing by at least $\varepsilon$ with points of $\mathcal{X}_{N}$ are separated by measurements differing by less than $\gamma$. Suppose that for a fixed ${\mathbf{x}}\in\mathcal{X}$ we have (81) $\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}_{j})\|_{2}\geq\varepsilon\quad\Rightarrow\quad\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{j})\|_{2}\geq\gamma$ for every ${\mathbf{x}}_{j}\in\mathcal{X}_{N}$. For any ${\mathbf{x}}^{\prime}\in\mathcal{X}$, there is an ${\mathbf{x}}_{j}\in\mathcal{X}_{N}$ with $\|{\mathbf{x}}^{\prime}-{\mathbf{x}}_{j}\|_{2}<\varepsilon_{0}$ and so we have (82) $\begin{split}\varepsilon+\varepsilon_{0}\|{\mathbf{g}}\|_{\text{lip}}&\leq\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}\\\ &\leq\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}_{j})\|_{2}+\|{\mathbf{g}}({\mathbf{x}}_{j})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}\\\ &<\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}_{j})\|_{2}+\varepsilon_{0}\|{\mathbf{g}}\|_{\text{lip}}.\end{split}$ Hence, $\varepsilon\leq\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}_{j})\|_{2}$, which implies that $\gamma\leq\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{j})\|_{2}$ by assumption. From this we obtain (83) $\begin{split}\gamma&\leq\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}+\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{j})\|_{2}\\\ &<\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}+\varepsilon_{0}\|{\mathbf{m}}_{\mathscr{S}}\|_{\text{lip}}.\end{split}$ Therefore, for such an ${\mathbf{x}}\in\mathcal{X}$ we have (84) $\forall{\mathbf{x}}^{\prime}\in\mathcal{X}\qquad\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}\geq\varepsilon+\varepsilon_{0}\|{\mathbf{g}}\|_{\text{lip}}\\\ \Rightarrow\quad\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}>\gamma-\varepsilon_{0}\|{\mathbf{m}}_{\mathscr{S}}\|_{\text{lip}}.$ It follows that $\mathbb{E}\left[Z_{\mathscr{S}}({\mathbf{b}}_{i})\right]$ is an upper bound on the $\mu$-measure of points in $\mathcal{X}$ for which there is another point in $\mathcal{X}$ with a close measurement and distant target value, that is (85) $\mathbb{E}\left[Z_{\mathscr{S}}({\mathbf{b}}_{i})\right]\geq\mu\big{(}\big{\\{}{\mathbf{x}}\in\mathcal{X}\ :\ \exists{\mathbf{x}}^{\prime}\in\mathcal{X}\quad\mbox{s.t.}\\\ \|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}\leq\gamma-\varepsilon_{0}\|{\mathbf{m}}_{\mathscr{S}}\|_{\text{lip}},\\\ \|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}\geq\varepsilon+\varepsilon_{0}\|{\mathbf{g}}\|_{\text{lip}}\big{\\}}\big{)}.$ By assumption, we have a set $\mathscr{S}\subset\mathscr{M}$ so that $Z_{\mathscr{S}}({\mathbf{b}}_{i})=0$ for each $i=1,\ldots,m$. And so it remains to bound the difference between the empirical and true expectation of $Z_{\mathscr{S}}({\mathbf{b}}_{i})$ uniformly over every subset $\mathscr{S}\subset\mathscr{M}$. For fixed $\mathscr{S}$, the one-sided Hoeffding inequality gives (86) $\mathbb{P}\Big{\\{}\frac{1}{m}\sum_{i=1}^{m}\left(\mathbb{E}[Z_{\mathscr{S}}({\mathbf{b}}_{i})]-Z_{\mathscr{S}}({\mathbf{b}}_{i})\right)\geq\delta\Big{\\}}\leq e^{-2m\delta^{2}}.$ Unfixing $\mathscr{S}$ via the union bound over all $\mathscr{S}\subset\mathscr{M}$ and applying our assumption about the number of base points $m$ yields (87) $\mathbb{P}\bigcup_{\mathscr{S}\subseteq\mathscr{M}}\Big{\\{}\frac{1}{m}\sum_{i=1}^{m}\left(\mathbb{E}[Z_{\mathscr{S}}({\mathbf{b}}_{i})]-Z_{\mathscr{S}}({\mathbf{b}}_{i})\right)\geq\delta\Big{\\}}\leq\exp{\left[\\#(\mathscr{M})\ln{2}-2m\delta^{2}\right]}\\\ \leq p.$ Since our assumed choice of $\mathscr{S}$ has $f_{m}(\mathscr{S})=f_{m}(\mathscr{M})$ it follows that all $Z_{\mathscr{S}}({\mathbf{b}}_{i})=0$, $i=1,\ldots,m$, hence we have (88) $\mathbb{E}[Z_{\mathscr{S}}({\mathbf{b}}_{i})]<\delta$ with probability at least $1-p$. Combining this with Eq. 85 completes the proof. ∎ It is also possible to use a down-sampled objective to greedily choose sensors that satisfy a similarly relaxed version of the amplification guarantee given by Proposition 4.6 with high probability over a large subset of $\mathcal{X}$. In order to do this, we take the sum in Eq. 39 over secants between a randomly chosen collection of base points $\mathcal{B}_{m}=\\{{\mathbf{b}}_{1},\ldots,{\mathbf{b}}_{m}\\}\subseteq\mathcal{X}$ and the $\varepsilon_{0}$-net $\mathcal{X}_{N}$. Again, the number of base points depends on the quality of the guarantee and not on size of the $\varepsilon_{0}$-net, so that the computational cost can be reduced to linear dependence on the size of $\mathcal{X}_{N}$. Specifically, in place of Eq. 39, we consider (89) $f_{L,m}(\mathscr{S})=\sum_{\begin{subarray}{c}1\leq i\leq m,\ 1\leq j\leq N,\\\ {\mathbf{g}}({\mathbf{b}}_{i})\neq{\mathbf{g}}({\mathbf{x}}_{j})\end{subarray}}\min\left\\{\frac{\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{b}}_{i})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{j})\|_{2}^{2}}{\|{\mathbf{g}}({\mathbf{b}}_{i})-{\mathbf{g}}({\mathbf{x}}_{j})\|_{2}^{2}},\ \frac{1}{L^{2}}\right\\}.$ In Theorem 5.5 we show that when a sufficiently small set of sensors $\mathscr{S}$ is found, e.g., using the greedy algorithm with the sampled objective $f_{L,m}$, that satisfies the amplification tolerance over $\mathcal{B}_{m}\times\mathcal{X}_{N}$, we can conclude that that a slightly relaxed amplification bound holds with high probability over a large subset of $\mathcal{X}$. In particular, the subset of “bad points” in ${\mathbf{x}}\in\mathcal{X}$ for which there is another point ${\mathbf{x}}^{\prime}\in\mathcal{X}$ with a different target value, but not a sufficiently different measured value, has small $\mu$-measure with high probability. ###### Theorem 5.5 (Sampled Amplification Guarantee). Let $\mathcal{X}_{N}$ be an $\varepsilon_{0}$-net of $\mathcal{X}$ and let the base points $\mathcal{B}_{m}$ be sampled independently according to a probability measure $\mu$ on $\mathcal{X}$ with (90) $m\geq\frac{1}{2\delta^{2}}\left(\\#(\mathscr{M})\ln{2}-\ln{p}\right).$ Consider the objective $f_{m}$ given by Eq. 89 for a certain choice of $L>0$ for which (91) $\|{\mathbf{g}}({\mathbf{b}}_{i})-{\mathbf{g}}({\mathbf{x}}_{j})\|_{2}\leq L\|{\mathbf{m}}_{\mathscr{M}}({\mathbf{b}}_{i})-{\mathbf{m}}_{\mathscr{M}}({\mathbf{x}}_{j})\|_{2}$ is achieved for all ${\mathbf{b}}_{i}\in\mathcal{B}_{m}$, ${\mathbf{x}}_{j}\in\mathcal{X}_{N}$. Suppose also that ${\mathbf{g}}$ and the measurement functions ${\mathbf{m}}_{k}$, $k\in\mathscr{M}$ are all Lipschitz functions over $\mathcal{X}$. If $f_{L,m}(\mathscr{S})=f_{L,m}(\mathscr{M})$, then the $\mu$-measure of points ${\mathbf{x}}\in\mathcal{X}$ such that (92) $\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}<L\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})+{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}+\left(\|{\mathbf{g}}\|_{\text{lip}}+L\|{\mathbf{m}}_{\mathscr{S}}\|_{\text{lip}}\right)\varepsilon_{0}$ for every ${\mathbf{x}}^{\prime}\in\mathcal{X}$ is at least $1-\delta$ with probability at least $1-p$. ###### Proof. The proof is analogous to Theorem 5.4 and so we relegate it to Appendix C. ∎ ## 6\. Working with Noisy Data So far, we have considered maximizing different measures of robust reconstructability given a collection of noiseless data. That is, the resulting sensors are selected in order to be noise robust, but we have assumed that the measurements ${\mathbf{m}}_{j}({\mathbf{x}}_{i})$, $j\in\mathscr{M}$ and target variables ${\mathbf{g}}({\mathbf{x}}_{i})$ used during the sensor selection process are noiseless over the sampled states ${\mathbf{x}}_{i}\in\mathcal{X}_{N}$. In many applications, however, our data may contain noisy measurements, target variables, or both. In this section, we study the effect of noisy data on the performance of our proposed secant-based greedy algorithms. By “noise” we mean specifically that we are given a collection of available measurements $\left\\{{\mathbf{\tilde{m}}}_{i,\mathscr{M}}={\mathbf{m}}_{\mathscr{M}}({\mathbf{x}}_{i})+{\mathbf{u}}_{i,\mathscr{M}}\right\\}_{i=1}^{N}$ that are corrupted by unknown noise ${\mathbf{u}}_{i,\mathscr{M}}$ together with the corresponding target values $\left\\{{\mathbf{\tilde{g}}}_{i}={\mathbf{g}}({\mathbf{x}}_{i})+{\mathbf{v}}_{i}\right\\}_{i=1}^{N}$ that are also corrupted by unknown noise ${\mathbf{v}}_{i}$. That is, we do not have access to the measurement functions ${\mathbf{m}}_{\mathscr{M}}$ or the target function ${\mathbf{g}}$ and must rely solely on noisy data generated by them. First, we mention that the minimal sensing method to meet an error tolerance discussed in Section 4.2 is robust to bounded noise in the measurements and target variables. In particular, since the selected sensors $\mathscr{S}$ using the approach described in Section 4.2 automatically satisfy Eq. 94, Proposition 6.1, below, shows that the true measurements coming from states with sufficiently distant true target values must also be separated by the measurements. ###### Proposition 6.1 (Noisy Separation Guarantee). Let $\mathcal{X}_{N}$ be an $\varepsilon_{0}$-net of $\mathcal{X}$ (see Definition 4.4) and let ${\mathbf{v}}_{i}\in\mathbb{R}^{\dim{\mathbf{g}}}$, ${\mathbf{u}}_{i,\mathscr{S}}\in\mathbb{R}^{d_{\mathscr{S}}}$, $i=1,\ldots,N$ be bounded vectors with (93) $\forall i=1,\ldots,N\qquad\left\|{\mathbf{u}}_{i,\mathscr{S}}\right\|_{2}\leq\delta_{u},\qquad\left\|{\mathbf{v}}_{i}\right\|_{2}\leq\delta_{v}.$ Suppose that there exists $\epsilon>0$ and $\gamma>0$ such that (94) $\forall{\mathbf{x}}_{i},{\mathbf{x}}_{j}\in\mathcal{X}_{N}\qquad\|\left({\mathbf{g}}({\mathbf{x}}_{i})+{\mathbf{v}}_{i}\right)-\left({\mathbf{g}}({\mathbf{x}}_{j})+{\mathbf{v}}_{j}\right)\|_{2}\geq\varepsilon\\\ \Rightarrow\quad\|\left({\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{i})+{\mathbf{u}}_{i,\mathscr{S}}\right)-\left({\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{j})+{\mathbf{u}}_{j,\mathscr{S}}\right)\|_{2}\geq\gamma.$ If ${\mathbf{m}}_{\mathscr{S}}$ and ${\mathbf{g}}$ are Lipschitz functions with Lipschitz constants $\|{\mathbf{m}}_{\mathscr{S}}\|_{\text{lip}}$ and $\|{\mathbf{g}}\|_{\text{lip}}$ respectively, then (95) $\forall{\mathbf{x}},{\mathbf{x}}^{\prime}\in\mathcal{X}\qquad\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}\geq\varepsilon+2\delta_{v}+2\varepsilon_{0}\|{\mathbf{g}}\|_{\text{lip}}\\\ \quad\Rightarrow\quad\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}>\gamma-2\delta_{u}-2\varepsilon_{0}\|{\mathbf{m}}_{\mathscr{S}}\|_{\text{lip}}.$ ###### Proof. The proof is analogous to Proposition 4.5 and has been relegated to Appendix C. ∎ As a consequence of Proposition 6.1, the reconstruction error for the desired quantities using these sensors can still be bounded if the thresholds $\varepsilon$ and $\gamma$ exceed twice the noise level of the target variable and measurements respectively (with a little extra padding based on the sampling fineness). On the other hand, the minimal sensing method to meet an amplification tolerance discussed in Section 4.3 is very sensitive to noisy data. This is because measurement noise can bring two nearby measurements ${\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})$ and ${\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})$ arbitrarily close together while the corresponding target variables ${\mathbf{g}}({\mathbf{x}})$ and ${\mathbf{g}}({\mathbf{x}}^{\prime})$ remain separated. Such terms can result in arbitrarily large data-driven estimates of the reconstruction Lipschitz constant. Consequently it may not be possible to find a small set of sensors $\mathscr{S}$ such that (96) $\max_{1\leq i<j\leq N}\frac{\left\|{\mathbf{\tilde{g}}}_{i}-{\mathbf{\tilde{g}}}_{j}\right\|_{2}}{\left\|{\mathbf{\tilde{m}}}_{i,\mathscr{S}}-{\mathbf{\tilde{m}}}_{j,\mathscr{S}}\right\|_{2}}\leq L$ for acceptable values of $L$. One way to deal with this problem is to smooth out the target variables. For instance, given the available noisy measurement and target pairs $\left\\{\left({\mathbf{\tilde{m}}}_{i,\mathscr{M}},\ {\mathbf{\tilde{g}}}_{i}\right)\right\\}_{i=1}^{N}$, one can find an approximation of the reconstruction function $\boldsymbol{\Phi}_{\mathscr{M}}$ via regression. Using the predicted target variables (97) ${\mathbf{\hat{g}}}_{i}:=\boldsymbol{\Phi}_{\mathscr{M}}\left({\mathbf{\tilde{m}}}_{i,\mathscr{M}}\right)$ in place of the noisy data ${\mathbf{\tilde{g}}}_{i}$ fixes the problem of infinite Lipschitz constants. This is because the amplification-based approach using these data seeks to find the minimal set of sensors $\mathscr{S}$ such that (98) $\max_{1\leq i<j\leq N}\frac{\left\|{\mathbf{\hat{g}}}_{i}-{\mathbf{\hat{g}}}_{j}\right\|_{2}}{\left\|{\mathbf{\tilde{m}}}_{i,\mathscr{S}}-{\mathbf{\tilde{m}}}_{j,\mathscr{S}}\right\|_{2}}\leq L$ rather than satisfying Eq. 96. We use a similar type of smoothing approach for the shock-mixing layer problem by choosing the leading two Isomap coordinates ${\mathbf{g}}({\mathbf{x}})=\left(\phi_{1}({\mathbf{x}}),\ \phi_{2}({\mathbf{x}})\right)$ rather than simply taking ${\mathbf{g}}({\mathbf{x}})={\mathbf{x}}$. This is because the full state ${\mathbf{x}}$ contains some small noise, meaning that it does not lie exactly on the one-dimensional loop in state space, but rather on a very thin manifold with full dimensionality. If we were to use the Lipschitz-based approach to reconstruct ${\mathbf{x}}$ directly, we would need enough sensors to reconstruct this noise. By seeking to reconstruct the leading Isomap coordinates instead, we have regularized our selection algorithm to choose only those sensors that are needed to reconstruct the dominant periodic behavior. Reconstructing smoothed target variables turns out to be a robust method for sensor placement, as we show by introducing increasing levels of noise in the shock-mixing layer problem. We added independent Gaussian noise with standard deviations $\sigma_{\text{noise}}=0.01$, $0.02$, $0.03$, $0.04$, and $0.05$ to each velocity component at every location on the computational grid, yielding noisy snapshots like the one shown in Figure 5. This reflects the typical situation when the underlying data given to us are noisy. At each noise level we selected three sensors using the detectable difference-based method of Section 4.1 as well as the amplification tolerance-based method of Section 4.3, with a bisection search over the threshold Lipschitz constant $L$, to reconstruct the leading two Isomap coordinates of the noisy data. Despite the noise, the leading two Isomap coordinates continued to accurately capture the dominant periodic behavior of the underlying system, making them good reconstruction target variables. The thresholds for the detectable difference method were fixed at $\gamma=0.04$ except in the $\sigma_{\text{noise}}=0.02$ case, where better performance was achieved using $\gamma=0.02$. (a) (b) (c) (d) (e) (f) (g) (h) (i) (j) Figure 5. We show the stream-wise (first column) and transverse (second column) components of velocity for a single snapshot of the shock-mixing layer flow with increasing levels of noise added in each successive row. Independent Gaussian noise with standard deviations $\sigma_{\text{noise}}=0.01$, $0.02$, $0.03$, $0.04$, and $0.05$ are added to each velocity component at each location on the computational grid. The first two sensors chosen by detectable difference method of Section 4.1 are indicated by green stars and the third is indicated by a black star. The three sensors selected using the amplification tolerance method of Section 4.3 with bisection search over $L$ are indicated by black squares. We found that the amplification tolerance-based method identified the same sensors across each of the first four noise levels $\sigma_{\text{noise}}=0.01$, $0.02$, $0.03$, and $0.04$. While these sensor locations differed slightly from the ones selected without noise (shown in Figure 1), they too were capable of robustly recovering the underlying phase of the system as illustrated by their corresponding measurements in the third column of Figure 6. At the largest noise level $\sigma_{\text{noise}}=0.5$, the sensors selected using this method changed, but were still capable of revealing the phase as shown in the bottom right plot of Figure 6. The detectable difference-based method selected the same three sensors as in the zero noise case when $\sigma_{\text{noise}}=0.01$ with the first two remaining the same up to $\sigma_{\text{noise}}=0.02$. At these noise levels the first two sensors are sufficient to reveal the underlying phase of the system as shown in the first two plots in the first column of Figure 6. Beyond this level of noise, the first two sensors were no longer able to reveal the phase as illustrated by the self-intersections in the last three plots in the first column of Figure 6. While it is admittedly difficult to see from the last three plots in the middle column of Figure 6, the third sensor eliminated these self-intersections by raising one of the two intersecting branches and allowing the phase to be determined. (a) (b) (c) (d) (e) (f) (g) (h) (i) (j) (k) (l) (m) (n) (o) Figure 6. These plots show the measurements made by sensors selected using the detectable difference method of Section 4.1 with two (first column) and three (second column) sensors along with the amplification tolerance method of Section 4.3 with three sensors (third column) on the shock-mixing layer flow problem with various levels of added noise. Each row shows the result of adding independent Gaussian noise with standard deviations $\sigma_{\text{noise}}=0.01$, $0.02$, $0.03$, $0.04$, and $0.05$ to each velocity component at each location on the computational grid. ## 7\. Conclusion In this paper we have identified a common type of nonlinear structure that causes techniques for sensor placement relying on linear reconstruction accuracy as an optimization criterion to consistently fail to identify minimal sets of sensors. Specifically, these techniques break down and lead to costly over-sensing when the data is intrinsically low dimensional, but is curved in such a way that energetic components are functions of less energetic ones, but not vice versa. This problem occurs commonly in fluid flows, period-doubling bifurcations in ecology and cardiology, as well as in spectral methods for manifold learning. We demonstrated that a representative collection of linear techniques fail to identify sensors from which the state of a shock-mixing layer flow can be reconstructed, and we provide a simple example that illustrates that the performance of the linear techniques can be arbitrarily bad. In addition, we demonstrated that it is impossible to use linear feature selection methods to choose fundamental nonlinear eigen-coordinates in manifold learning problems. To remedy these issues, we proposed a new approach for sensor placement that relies on the information contained in secant vectors between data points to quantify nonlinear reconstructability of desired quantities from measurements. The resulting secant-based optimization problems turn out to have useful diminishing returns properties that enable efficient greedy approximation algorithms to achieve guaranteed high levels of performance. We also describe how down-sampling can be used to improve the computational scaling of these algorithms while still providing guarantees regarding the reconstructability of states in the underlying set from which the available data is sampled. Finally, these methods prove to be capable of selecting minimal collections of sensors in the shock-mixing layer problem as well as selecting the minimal set of fundamental manifold learning coordinates on a torus — both of which are problems where the linear techniques fail. ### Acknowledgements The authors would like to thank Gregory Blaisdell, Shih-Chieh Lo, Tasos Lyrintzis, and Kurt Aikens for providing the code used to compute the shock- mixing layer interaction. We also want to thank Alberto Padovan and Anastasia Bizyaeva for providing key references that motivate our main example, provide connections with period doubling, and reveal how linear methods can fail to find adequate sensor and actuator locations in real-world problems. ## References * [1] M. Belkin and P. Niyogi. Laplacian eigenmaps for dimensionality reduction and data representation. Neural computation, 15(6):1373–1396, 2003. * [2] J. L. Bentley. A survey of techniques for fixed radius near neighbor searching. Technical report, Stanford University, Stanford, CA, USA, 1975. * [3] J. L. Bentley, D. F. Stanat, and E. H. Williams Jr. The complexity of finding fixed-radius near neighbors. Information Processing Letters, 6(6):209–212, 1977. * [4] G. Berkooz, P. Holmes, and J. L. Lumley. The proper orthogonal decomposition in the analysis of turbulent flows. Annual Review of Fluid Mechanics, 25(1):539–575, 1993. * [5] D. Broomhead and M. Kirby. Dimensionality reduction using secant-based projection methods: The induced dynamics in projected systems. Nonlinear Dynamics, 41(1-3):47–67, 2005. * [6] S. L. Brunton and J. N. Kutz. Data-Driven Science and Engineering: Machine Learning, Dynamical Systems, and Control. Cambridge University Press, 2019. * [7] S. L. Brunton, J. L. Proctor, and J. N. Kutz. Discovering governing equations from data by sparse identification of nonlinear dynamical systems. Proceedings of the National Academy of Sciences, 113(15):3932–3937, 2016. * [8] P. Businger and G. H. Golub. Linear least squares solutions by householder transformations. Numerische Mathematik, 7(3):269–276, 1965. * [9] J. L. Callaham, K. Maeda, and S. L. Brunton. Robust flow reconstruction from limited measurements via sparse representation. Physical Review Fluids, 4(10):103907, 2019. * [10] E. J. Candès, Y. Plan, et al. Near-ideal model selection by $\ell_{1}$ minimization. The Annals of Statistics, 37(5A):2145–2177, 2009. * [11] W. F. Caselton, L. Kan, and J. V. Zidek. Quality data networks that minimize entropy. In Statistics in the Environmental and Earth Sciences, pages 10–38. Halsted Press, 1992. * [12] W. F. Caselton and J. V. Zidek. Optimal monitoring network designs. Statistics & Probability Letters, 2(4):223–227, 1984. * [13] K. Chaloner and I. Verdinelli. Bayesian experimental design: A review. Statistical Science, pages 273–304, 1995. * [14] S. Chaturantabut and D. C. Sorensen. Nonlinear model reduction via discrete empirical interpolation. SIAM Journal on Scientific Computing, 32(5):2737–2764, 2010. * [15] N. T. Clemens and V. Narayanaswamy. Low-frequency unsteadiness of shock wave/turbulent boundary layer interactions. Annual Review of Fluid Mechanics, 46:469–492, 2014. * [16] R. R. Coifman and S. Lafon. Diffusion maps. Applied and Computational Harmonic Analysis, 21(1):5–30, 2006. * [17] N. K. Dhingra, M. R. Jovanović, and Z.-Q. Luo. An ADMM algorithm for optimal sensor and actuator selection. In 53rd IEEE Conference on Decision and Control, pages 4039–4044. IEEE, 2014. * [18] Z. Drmac and S. Gugercin. A new selection operator for the discrete empirical interpolation method—improved a priori error bound and extensions. SIAM Journal on Scientific Computing, 38(2):A631–A648, 2016. * [19] W. J. Dunstan, R. R. Bitmead, and S. M. Savaresi. Fitting nonlinear low-order models for combustion instability control. Control Engineering Practice, 9(12):1301–1317, 2001. * [20] I. Guyon and A. Elisseeff. An introduction to variable and feature selection. Journal of Machine Learning Research, 3(Mar):1157–1182, 2003. * [21] C. Hegde, A. C. Sankaranarayanan, W. Yin, and R. G. Baraniuk. Numax: A convex approach for learning near-isometric linear embeddings. IEEE Transactions on Signal Processing, 63(22):6109–6121, 2015\. * [22] C.-M. Ho and L.-S. Huang. Subharmonics and vortex merging in mixing layers. Journal of Fluid Mechanics, 119:443–473, 1982. * [23] S. Hosseinyalamdary. Deep Kalman filter: Simultaneous multi-sensor integration and modelling; a GNSS/IMU case study. Sensors, 18(5):1316, 2018. * [24] H. Hotelling. Analysis of a complex of statistical variables into principal components. Journal of Educational Psychology, 24(6):417, 1933. * [25] A. A. Jamshidi and M. J. Kirby. Towards a black box algorithm for nonlinear function approximation over high-dimensional domains. SIAM Journal on Scientific Computing, 29(3):941–963, 2007. * [26] S. Joshi and S. Boyd. Sensor selection via convex optimization. IEEE Transactions on Signal Processing, 57(2):451–462, 2008. * [27] A. Krause, H. B. McMahan, C. Guestrin, and A. Gupta. Robust submodular observation selection. Journal of Machine Learning Research, 9(Dec):2761–2801, 2008. * [28] A. Krause, A. Singh, and C. Guestrin. Near-optimal sensor placements in Gaussian processes: Theory, efficient algorithms and empirical studies. Journal of Machine Learning Research, 9(Feb):235–284, 2008. * [29] R. G. Krishnan, U. Shalit, and D. Sontag. Structured inference networks for nonlinear state space models. In Thirty-First AAAI Conference on Artificial Intelligence, 2017\. * [30] A. Lamraoui, F. Richecoeur, S. Ducruix, and T. Schuller. Experimental analysis of simultaneous non-harmonically related unstable modes in a swirled combustor. In Proceedings of the ASME 2011 Turbo Expo: Turbine Technical Conference and Exposition, volume 2, pages 1289–1299, 2011. * [31] K. Lee and K. T. Carlberg. Model reduction of dynamical systems on nonlinear manifolds using deep convolutional autoencoders. Journal of Computational Physics, 404:108973, 2020. * [32] S.-C. Lo, G. A. Blaisdell, and A. S. Lyrintzis. High-order shock capturing schemes for turbulence calculations. International Journal for Numerical Methods in Fluids, 62(5):473–498, 2010. * [33] K. Manohar, B. W. Brunton, J. N. Kutz, and S. L. Brunton. Data-driven sparse sensor placement for reconstruction: Demonstrating the benefits of exploiting known patterns. IEEE Control Systems Magazine, 38(3):63–86, 2018. * [34] M. Marion and R. Temam. Nonlinear Galerkin methods. SIAM Journal on Numerical Analysis, 26(5):1139–1157, 1989. * [35] M. Minoux. Accelerated greedy algorithms for maximizing submodular set functions. In Optimization Techniques, pages 234–243. Springer, 1978. * [36] V. Mons, J.-C. Chassaing, and P. Sagaut. Optimal sensor placement for variational data assimilation of unsteady flows past a rotationally oscillating cylinder. Journal of Fluid Mechanics, 823:230–277, 2017. * [37] N. J. Nair and A. Goza. Integrating sensor data into reduced-order models with deep learning. Bulletin of the American Physical Society, 64, 2019. * [38] N. J. Nair and A. Goza. Leveraging reduced-order models for state estimation using deep learning. arXiv preprint arXiv:1912.10553, 2019. * [39] G. L. Nemhauser, L. A. Wolsey, and M. L. Fisher. An analysis of approximations for maximizing submodular set functions—I. Mathematical Programming, 14(1):265–294, 1978. * [40] M. Ohlberger and S. Rave. Reduced basis methods: Success, limitations and future challenges. In Proceedings of ALGORITMY, pages 1–12, 2016. * [41] Y. C. Pati, R. Rezaiifar, and P. S. Krishnaprasad. Orthogonal matching pursuit: Recursive function approximation with applications to wavelet decomposition. In Proceedings of 27th Asilomar conference on signals, systems and computers, pages 40–44. IEEE, 1993. * [42] S. Priebe and M. P. Martín. Low-frequency unsteadiness in shock wave–turbulent boundary layer interaction. Journal of Fluid Mechanics, 699:1–49, 2012. * [43] S. Priebe, J. H. Tu, C. W. Rowley, and M. P. Martín. Low-frequency dynamics in a shock-induced separated flow. Journal of Fluid Mechanics, 807:441–477, 2016. * [44] F. Pukelsheim. Optimal Design of Experiments. SIAM, 2006. * [45] T. Quail, A. Shrier, and L. Glass. Predicting the onset of period-doubling bifurcations in noisy cardiac systems. Proceedings of the National Academy of Sciences, 112(30):9358–9363, 2015. * [46] C. E. Rasmussen. Gaussian processes in machine learning. In Summer School on Machine Learning, pages 63–71. Springer, 2003\. * [47] G. Rega and H. Troger. Dimension reduction of dynamical systems: methods, models, applications. Nonlinear Dynamics, 41(1-3):1–15, 2005. * [48] C. W. Rowley, T. Colonius, and R. M. Murray. Model reduction for compressible flows using POD and Galerkin projection. Physica D: Nonlinear Phenomena, 189(1-2):115–129, 2004. * [49] B. Schölkopf, A. Smola, and K.-R. Müller. Nonlinear component analysis as a kernel eigenvalue problem. Neural computation, 10(5):1299–1319, 1998. * [50] P. Sebastiani and H. P. Wynn. Maximum entropy sampling and optimal Bayesian experimental design. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 62(1):145–157, 2000. * [51] M. Shamaiah, S. Banerjee, and H. Vikalo. Greedy sensor selection: Leveraging submodularity. In 49th IEEE Conference on Decision and Control, pages 2572–2577. IEEE, 2010. * [52] M. C. Shewry and H. P. Wynn. Maximum entropy sampling. Journal of applied statistics, 14(2):165–170, 1987. * [53] T. H. Summers, F. L. Cortesi, and J. Lygeros. On submodularity and controllability in complex dynamical networks. IEEE Transactions on Control of Network Systems, 3(1):91–101, 2015\. * [54] T. H. Summers, F. L. Cortesi, and J. Lygeros. On submodularity and controllability in complex dynamical networks. IEEE Transactions on Control of Network Systems, 3(1):91–101, March 2016. * [55] W. Sun, G. Yang, B. Du, L. Zhang, and L. Zhang. A sparse and low-rank near-isometric linear embedding method for feature extraction in hyperspectral imagery classification. IEEE Transactions on Geoscience and Remote Sensing, 55(7):4032–4046, 2017. * [56] J. B. Tenenbaum, V. De Silva, and J. C. Langford. A global geometric framework for nonlinear dimensionality reduction. science, 290(5500):2319–2323, 2000. * [57] R. Tibshirani. Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society: Series B (Methodological), 58(1):267–288, 1996. * [58] J. A. Tropp, A. C. Gilbert, and M. J. Strauss. Simultaneous sparse approximation via greedy pursuit. In Proceedings.(ICASSP’05). IEEE International Conference on Acoustics, Speech, and Signal Processing, 2005., volume 5, pages v–721. IEEE, 2005. * [59] V. Tzoumas, A. Jadbabaie, and G. J. Pappas. Sensor placement for optimal Kalman filtering: Fundamental limits, submodularity, and algorithms. In 2016 American Control Conference, pages 191–196. IEEE, 2016\. * [60] O. Tzuk, S. R. Ujjwal, C. Fernandez-Oto, M. Seifan, and E. Meron. Period doubling as an indicator for ecosystem sensitivity to climate extremes. Scientific reports, 9(1):1–10, 2019. * [61] H. Whitney. Differentiable manifolds. Annals of Mathematics, pages 645–680, 1936. * [62] H. Whitney. The self-intersections of a smooth $n$-manifold in $2n$-space. Annals of Mathematics, pages 220–246, 1944. * [63] L. A. Wolsey. An analysis of the greedy algorithm for the submodular set covering problem. Combinatorica, 2(4):385–393, 1982. * [64] J. Wright, A. Y. Yang, A. Ganesh, S. S. Sastry, and Y. Ma. Robust face recognition via sparse representation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 31(2):210–227, 2008. * [65] H. C. Yee, N. D. Sandham, and M. J. Djomehri. Low-dissipative high-order shock-capturing methods using characteristic-based filters. Journal of Computational Physics, 150(1):199–238, 1999. * [66] B. Yildirim, C. Chryssostomidis, and G. Karniadakis. Efficient sensor placement for ocean measurements using low-dimensional concepts. Ocean Modelling, 27(3-4):160–173, 2009. * [67] M. Yuan and Y. Lin. Model selection and estimation in regression with grouped variables. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 68(1):49–67, 2006. ## Appendix A Implementation Details ### A.1. Principal Component Analysis (PCA) and Isomap In this paper, we used principal component analysis (PCA) [24] in order to find a modal basis for pivoted QR factorization and to identify a low- dimensional representation of the state and its covariance for determinantal D-optimal selection techniques on the shock-mixing layer flow. In order to perform PCA, one needs an appropriate inner product on the space in which the data lives. In the case of the shock-mixing layer problem, we use the energy- based inner product for compressible flows developed in [48] together with trapezoidal quadrature weights to approximate the integrals of the spatial fields over a stretched computational grid. In this problem, the data consists of vectors ${\mathbf{z}}$ whose elements are the streamwise velocity $u$, transverse velocity $v$, and the local speed of sound $a$ over a $321\times 81$ computational grid. The inner product between two snapshots ${\mathbf{z}}$ and ${\mathbf{z}}^{\prime}$ is defined by (99) $\left\langle{\mathbf{z}},\ {\mathbf{z}}^{\prime}\right\rangle={\mathbf{z}}^{T}{\mathbf{W}}{\mathbf{z}}^{\prime}=\sum_{i=1}^{321}\sum_{j=1}^{81}w_{i,j}\left(u_{i,j}^{2}+v_{i,j}^{2}+a_{i,j}^{2}\right)\\\ \approx\int_{\Omega}\left[u(\xi_{1},\xi_{2})^{2}+v(\xi_{1},\xi_{2})^{2}+a(\xi_{1},\xi_{2})^{2}\right]d\xi_{1}d\xi_{2},$ where the weights $\\{w_{i,j}\\}$ are selected to perform trapezoidal quadrature. Principal component analysis is performed by computing an economy- sized singular value decomposition of the mean-subtracted data matrix (100) $\tilde{{\mathbf{U}}}\boldsymbol{\Sigma}{\mathbf{V}}^{T}={\mathbf{W}}^{1/2}\begin{bmatrix}({\mathbf{z}}_{1}-\bar{{\mathbf{z}}})&\cdots&({\mathbf{z}}_{N}-\bar{{\mathbf{z}}})\end{bmatrix},\quad\bar{{\mathbf{z}}}=\frac{1}{N}\sum_{i=1}^{N}{\mathbf{z}}_{i}$ and forming the matrix of principal vectors ${\mathbf{U}}={\mathbf{W}}^{-1/2}\tilde{{\mathbf{U}}}$. These vectors, making up the columns of ${\mathbf{U}}$, are orthonormal with respect to the ${\mathbf{W}}$-weighted inner product. If we represent the states in this basis so that ${\mathbf{z}}_{i}=\bar{{\mathbf{z}}}+{\mathbf{U}}{\mathbf{x}}_{i}$ then ${\mathbf{x}}$ has empirical covariance ${\mathbf{C}}_{{\mathbf{x}}}=\frac{1}{N}\boldsymbol{\Sigma}^{2}$. The same weighted inner product was used to compute the distances between each data point ${\mathbf{z}}_{i}$ and its $10$ nearest neighbors in order to compute the leading $50$ Isomap coordinates using scikit learn’s implementation found at https://scikit- learn.org/stable/modules/generated/sklearn.manifold.Isomap.html. ### A.2. (Group) LASSO We use the Python implementation of group LASSO [67] by Yngve Mardal Moe at the University of Oslo that can be found at https://group- LASSO.readthedocs.io/en/latest/index.html. We select among $2210$ sensor measurements of $u$ and $v$ velocity components over a grid of $1105$ spatial locations taken directly from the shock-mixing layer snapshot data. We tried two different kinds of target variables to be reconstructed via group LASSO. For the method we call “LASSO+PCA”, the target variables were the data’s leading $100$ principal components which capture over $99\%$ of the data’s variance. For the method we call “LASSO+Isomap”, the target variables were the leading two Isomap coordinates ${\mathbf{g}}({\mathbf{x}})=\left(\phi_{1}({\mathbf{x}}),\phi_{2}({\mathbf{x}})\right)$, which reveal the phase angle $\theta$. The sparsity-promoting regularization parameter was found using a bisection search in each case and was the smallest value, to within a tolerance of $10^{-5}$, for which group LASSO selected $3$ sensors. ### A.3. Bayesian D-Optimal Selection We use two different approaches for Bayesian D-optimal sensor placement: the greedy technique of [51] and the convex relaxation approach by [26]. In the greedy approach, we leverage the submodularity of the objective in the case when ${\mathbf{T}}={\mathbf{I}}$ in order to use the accelerated greedy algorithm of M. Minoux [35]. For the convex approach, we wrote a direct Python translation of a MATLAB code written by S. Joshi and S. Boyd that implements a Newton method with line search, and may be found at https://web.stanford.edu/~boyd/papers/matlab/sensor_selection/. We use the gradient and Hessian matrices for the Bayesian D-optimal objective from their paper [26]. In both the greedy and convex approach for the shock-mixing layer problem, we take the state to be its representation using $100$ principal components with covariance given by ${\mathbf{C}}_{{\mathbf{x}}}=\frac{1}{N}\boldsymbol{\Sigma}^{2}$ as computed by PCA. These principal components were also used as the relevant information to be reconstructed, i.e., ${\mathbf{T}}={\mathbf{I}}$. The sensor noise was assumed to be isotropic with covariance ${\mathbf{C}}_{{\mathbf{n}}_{\mathscr{S}}}=\sigma^{2}{\mathbf{I}}_{d_{\mathscr{S}}}$ with $\sigma=0.02$. We tried many other values of $\sigma$, yielding different sensor locations, none of which could be used for nonlinear reconstruction. The ones we show at $\sigma=0.02$ are representative. ### A.4. Maximum Likelihood D-Optimal Selection We used the maximum likelihood D-optimal selection technique based on convex relaxation found in [26] in order to choose sensors to try to reconstruct only the $3$rd and $4$th principal components of the shock-mixing layer snapshots. That is, if ${\mathbf{U}}=\begin{bmatrix}{\mathbf{u}}_{1}&{\mathbf{u}}_{2}&\cdots\end{bmatrix}$ is the matrix of principal components, we model the state as a linear combination of ${\mathbf{u}}_{3}$ and ${\mathbf{u}}_{4}$ together with isotropic Gaussian noise. We try to find the sensors so that the correct coefficients on ${\mathbf{u}}_{3}$ and ${\mathbf{u}}_{4}$ can be recovered with high confidence from the measurements. The rationale for doing so is the fact that these two components are sufficient to nonlinearly reconstruct the state of the system if they can be measured. As in Section A.3 above, we use a direct Python translation of a MATLAB code written by S. Joshi and S. Boyd, which may be found at https://web.stanford.edu/~boyd/papers/matlab/sensor_selection/. ### A.5. Pivoted QR Factorization For the pivoted QR factorization method [18, 8] applied to the shock-mixing layer flow, we represent the state approximately as a linear combination of the leading three principal components. Scipy’s implementation of pivoted QR factorization found at https://docs.scipy.org/doc/scipy/reference/generated/scipy.linalg.qr.html was used to select among the $2210$ allowable sensors those that allow robust reconstruction of these first three principal components. We also tried representing the state using more principal components and taking the first three sensor locations chosen via pivoted QR factorization. As with the case when only three principal components are used, these sensors do not enable nonlinear reconstruction of the state. ### A.6. Secant-Based Detectable Differences The secant-based detectable difference method was implemented using the accelerated greedy algorithm of M. Minoux [35] to optimize the objective computed over all secants between points in the training data set consisting of $N=750$ snapshots of the shock-mixing layer velocity field. We select among the $2210$ sensor measurements of $u$ and $v$ velocity components on a grid of $1105$ spatial locations taken directly from the shock-mixing layer snapshot data. The target variables were chosen to be the leading two Isomap coordinates ${\mathbf{g}}({\mathbf{x}})=\left(\phi_{1}({\mathbf{x}}),\phi_{2}({\mathbf{x}})\right)$, which reveal the phase angle $\theta$. The greedy algorithm first reveals the two sensor locations marked by green stars and then the black star in Figure 1 over the range of $0.02\leq\gamma\leq 0.06$, which can be used to reveal the exact phase of the system. Choosing smaller values of $\gamma$ produce different sensors that can also be used to reveal the phase, but with reduced robustness to measurement perturbations. Gaussian process regression [46] was used to reconstruct the leading $100$ principal components of the flowfields from the sensor measurements. We used scikit learn’s implementation which can be found at https://scikit- learn.org/stable/modules/generated/sklearn.gaussian_process.GaussianProcessRegressor.html together with a Matérn and white noise kernel whose parameters were optimized during the fit. For the torus example, the relevant information we wish to reconstruct are the leading $100$ Isomap eigen-coordinates ${\mathbf{g}}({\mathbf{x}})=\left(\phi_{1}({\mathbf{x}}),\ldots,\phi_{100}({\mathbf{x}})\right)$ computed from $2000$ points sampled from the torus according to Eq. 11. The objective function was evaluated using secants between $\\#(\mathcal{B})=100$ randomly sampled base points and the original set of $N=2000$ points. The correct three coordinates $\phi_{1},\phi_{2},\phi_{7}$ are selected from among the first $100$ consistently across a wide range of measurement separation values $0.05\leq\gamma\leq 3.0$. We note that these values vary slightly with the selected base points and these particular values hold only for one instance. ### A.7. Secant-Based Amplification Tolerance Like the secant-based detectable difference method described above, the secant-based amplification tolerance method was implemented using the same data, secant vectors, and target variables with the accelerated greedy algorithm. A bisection search was used to find the smallest Lipschitz constant $L=1868$ to within a tolerance of $1$ for which the algorithm selects three sensors on the shock-mixing layer flow. Three (different) sensors that correctly reveal the state of the flow are selected by this algorithm over a range $1868\leq L\leq 47624$, above which only two sensors that cannot reveal the state are selected. We also find that with $L=129$, the minimum possible number of sensors exceeds $\\#(\mathscr{S}_{K})/(1+\ln{\kappa})=3.18>3$. Therefore, the minimum possible reconstruction Lipschitz constant using three sensors that one might find by an exhaustive combinatorial search must be greater than $129$. We admit that this is likely a rather pessimistic bound, but we cannot check it as there are $\binom{2210}{3}\approx 1.8\times 10^{9}$ possible choices for three sensors in this problem. When applied to select from among the leading $100$ Isomap eigen-coordinates on the torus example with the same setup as the secant-based detectable differences method, the amplification tolerance method selects the appropriate collection $\phi_{1},\phi_{2},\phi_{7}$ over the range $7.1\leq L\leq 25$. We note that these value vary slightly with the selected base points and these particular values hold only for one instance. ## Appendix B Submodularity of Objectives We will need the definition of a modular function given below. ###### Definition B.1 (Modular Function). Denote the set of all subsets of $\mathscr{M}$ by $2^{\mathscr{M}}$. A real- valued function of the subsets $f:2^{\mathscr{M}}\to\mathbb{R}$ is called “modular” when it can be written as a sum (101) $f(\mathscr{S})=\sum_{j\in\mathscr{S}}a_{j}$ of constants $a_{j}$, $j\in\mathscr{M}$. The key ingredient needed to prove submodularity for the objectives described in Section 4 is the following lemma. ###### Lemma B.2 (Concave Composed with Modular is Submodular). Let $h:\mathbb{R}\to\mathbb{R}$ be a concave function and let $a:2^{\mathscr{M}}\to\mathbb{R}$ defined by (102) $a(\mathscr{S})=\sum_{j\in\mathscr{S}}a_{j}$ be a modular function (Def. B.1) of subsets $\mathscr{S}\subseteq\mathscr{M}$ with $a_{j}\geq 0$ for all $j\in\mathscr{M}$. Then the function $f:2^{\mathscr{M}}\to\mathbb{R}$ defined by (103) $f(\mathscr{S})=h(a(\mathscr{S}))$ is submodular. ###### Proof. Suppose that $\mathscr{S}\subseteq\mathscr{S}^{\prime}\subseteq{\mathscr{M}}\setminus\\{j\\}$. By concavity of $h$ we have (104) $h_{\alpha}=h((1-\alpha)a(\mathscr{S})+\alpha(a(\mathscr{S}^{\prime})+a_{j}))\geq(1-\alpha)h_{0}+\alpha h_{1}$ for every $\alpha\in[0,1]$, where we note that $h_{0}=f(\mathscr{S})$ and $h_{1}=f(\mathscr{S}^{\prime}\cup\\{j\\})$. Since $\\{a_{l}\\}$ are non-negative we have $a(\mathscr{S})\leq a(\mathscr{S})+a_{j}\leq a(\mathscr{S}^{\prime})+a_{j}$ and $a(\mathscr{S})\leq a(\mathscr{S}^{\prime})\leq a(\mathscr{S}^{\prime})+a_{j}$. We can therefore find (105) $\alpha_{1}=\frac{a_{j}}{a(\mathscr{S}^{\prime})+a_{j}-a(\mathscr{S})},\quad\alpha_{2}=\frac{a(\mathscr{S}^{\prime})-a(\mathscr{S})}{a(\mathscr{S}^{\prime})+a_{j}-a(\mathscr{S})}$ so that $h_{\alpha_{1}}=f(\mathscr{S}\cup\\{j\\})$ and $h_{\alpha_{2}}=f(\mathscr{S}^{\prime})$. Note that $\alpha_{1}+\alpha_{2}=1$. We now use Eq. 104 at $\alpha_{1}$ and $\alpha_{2}$ to bound the increments of $f$: (106) $f(\mathscr{S}\cup\\{j\\})-f(\mathscr{S})=h_{\alpha_{1}}-h_{0}\geq\alpha_{1}(h_{1}-h_{0}),$ (107) $f(\mathscr{S}^{\prime}\cup\\{j\\})-f(\mathscr{S}^{\prime})=h_{1}-h_{\alpha_{2}}\leq(1-\alpha_{2})(h_{1}-h_{0})$ Combining the bounds Eq. 106 and Eq. 107 on the increments using $1-\alpha_{2}=\alpha_{1}$ we conclude that $f$ is submodular (108) $f(\mathscr{S}\cup\\{j\\})-f(\mathscr{S})\geq f(\mathscr{S}^{\prime}\cup\\{j\\})-f(\mathscr{S}^{\prime}).$ ∎ Using Lemma B.2 it suffices to observe that each of the objectives described in Section 4 can be written as the composition of a concave function and a modular function. We carry this out below in addition to proving normalization and monotonicity for these objectives. ###### Lemma B.3 (Detectable Difference Objective is Submodular). Suppose that the target variables ${\mathbf{g}}$ and measurements ${\mathbf{m}}_{j}$, $j\in\mathscr{M}$ are measurable functions. If $\mu$ and $\nu$ are measures on $\mathcal{X}$, then the function defined by (109) $f(\mathscr{S})=\int_{\begin{subarray}{c}({\mathbf{x}},{\mathbf{x}}^{\prime})\in\mathcal{X}\times\mathcal{X}:\\\ \|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}\geq\varepsilon\end{subarray}}w_{\gamma,{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}^{2}\ d\mu({\mathbf{x}})\nu(d{\mathbf{x}}^{\prime}),$ for any $\varepsilon\geq 0$ with (110) $w_{\gamma,{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})=\min\left\\{\frac{1}{\gamma^{2}}\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}^{2},\ 1\right\\},$ is normalized so that $f(\emptyset)=0$, monotone non-decreasing so that $\mathscr{S}\subseteq\mathscr{S}^{\prime}\ \Rightarrow\ f(\mathscr{S})\leq f(\mathscr{S}^{\prime})$, and submodular (Def. 4.2). ###### Proof. Normalization is obvious. It suffices to prove that the function $w_{{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})$ is monotone and submodular for any fixed ${\mathbf{x}},{\mathbf{x}}^{\prime}\in\mathcal{X}$. For if we suppose that (111) $\mathscr{S}\subseteq\mathscr{S}^{\prime}\subseteq{\mathscr{M}}\setminus\\{j\\}\quad\Rightarrow\quad w_{\gamma,{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S}\cup\\{j\\})-w_{\gamma,{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})\\\ \geq w_{\gamma,{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S}^{\prime}\cup\\{j\\})-w_{\gamma,{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S}^{\prime}),$ then multiplying both sides of the inequality by $\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}^{2}$ and integrating proves that $f$ is submodular. The same argument also proves monotonicity. Let ${\mathbf{x}},{\mathbf{x}}^{\prime}\in\mathcal{X}$ be fixed. The squared separation between the measurements is given by a modular (Def. B.1) sum (112) $\mathscr{S}\ \mapsto\ \|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}^{2}=\sum_{j\in\mathscr{S}}\|{\mathbf{m}}_{j}({\mathbf{x}})-{\mathbf{m}}_{j}({\mathbf{x}}^{\prime})\|_{2}^{2}$ of non-negative constants $\|{\mathbf{m}}_{j}({\mathbf{x}})-{\mathbf{m}}_{j}({\mathbf{x}}^{\prime})\|_{2}^{2}$ over each $j\in\mathscr{S}$. Since $x\mapsto\min\\{x/\gamma^{2},\ 1\\}$ is a non-decreasing function, it follows that $\mathscr{S}\subseteq\mathscr{S}^{\prime}\ \Rightarrow\ w_{{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})\leq w_{{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S}^{\prime})$, proving monotonicity. Submodularity of $w_{{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})$ follows from Lemma B.2 since $w_{{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})$ is the composition of a concave function $x\mapsto\min\\{x/\gamma^{2},\ 1\\}$ with the modular function in Eq. 112. ∎ ###### Lemma B.4 (Lipschitz Objective is Submodular). Suppose that the target variables ${\mathbf{g}}$ and measurements ${\mathbf{m}}_{j}$, $j\in\mathscr{M}$ are measurable functions. If $\mu$ and $\nu$ are measures on $\mathcal{X}$, then the function defined by (113) $f(\mathscr{S})=\int_{\begin{subarray}{c}({\mathbf{x}},{\mathbf{x}}^{\prime})\in\mathcal{X}\times\mathcal{X}:\\\ {\mathbf{g}}({\mathbf{x}})\neq{\mathbf{g}}({\mathbf{x}}^{\prime})\end{subarray}}g_{{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})\ d\mu({\mathbf{x}})\nu(d{\mathbf{x}}^{\prime}),$ with (114) $g_{{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})=\min\left\\{\frac{\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}^{2}}{\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}^{2}},\ \frac{1}{L^{2}}\right\\},$ is normalized so that $f(\emptyset)=0$, monotone non-decreasing so that $\mathscr{S}\subseteq\mathscr{S}^{\prime}\ \Rightarrow\ f(\mathscr{S})\leq f(\mathscr{S}^{\prime})$, and submodular (Def. 4.2). ###### Proof. Normalization is obvious. It suffices to prove that the function $g_{{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})$ is monotone and submodular for any fixed ${\mathbf{x}},{\mathbf{x}}^{\prime}\in\mathcal{X}$. For if we suppose that (115) $\mathscr{S}\subseteq\mathscr{S}^{\prime}\subseteq{\mathscr{M}}\setminus\\{j\\}\quad\Rightarrow\quad g_{{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S}\cup\\{j\\})-g_{{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})\geq g_{{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S}^{\prime}\cup\\{j\\})-g_{{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S}^{\prime}),$ then integrating both sides of the inequality proves that $f$ is submodular. The same argument also proves monotonicity. Let ${\mathbf{x}},{\mathbf{x}}^{\prime}\in\mathcal{X}$ be fixed. The squared separation between the measurements is given by a modular (Def. B.1) sum (116) $\mathscr{S}\ \mapsto\ \|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}^{2}=\sum_{j\in\mathscr{S}}\|{\mathbf{m}}_{j}({\mathbf{x}})-{\mathbf{m}}_{j}({\mathbf{x}}^{\prime})\|_{2}^{2}$ of non-negative constants $\|{\mathbf{m}}_{j}({\mathbf{x}})-{\mathbf{m}}_{j}({\mathbf{x}}^{\prime})\|_{2}^{2}$ over each $j\in\mathscr{S}$. Since (117) $x\ \mapsto\ \min\left\\{\frac{x}{\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}^{2}},\ \frac{1}{L^{2}}\right\\}$ is a non-decreasing function, it follows that $\mathscr{S}\subseteq\mathscr{S}^{\prime}\ \Rightarrow\ g_{{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})\leq g_{{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S}^{\prime})$, proving monotonicity. Submodularity of $g_{{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})$ follows from Lemma B.2 since $g_{{\mathbf{x}},{\mathbf{x}}^{\prime}}(\mathscr{S})$ is the composition of the concave function in Eq. 117 with the modular function in Eq. 116. ∎ ## Appendix C Proofs ###### Proposition 4.5: Separation Guarantee on Underlying Set. The result follows immediately from the triangle inequality. Let ${\mathbf{x}},{\mathbf{x}}^{\prime}\in\mathcal{X}$ and ${\mathbf{x}}_{i},{\mathbf{x}}_{j}\in\mathcal{X}_{N}$ so that $\|{\mathbf{x}}-{\mathbf{x}}_{i}\|_{2}<\varepsilon_{0}$ and $\|{\mathbf{x}}^{\prime}-{\mathbf{x}}_{j}\|_{2}<\varepsilon_{0}$. Then $\varepsilon+2\varepsilon_{0}\|{\mathbf{g}}\|_{\text{lip}}\leq\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}$ implies that (118) $\begin{split}\varepsilon+2\varepsilon_{0}\|{\mathbf{g}}\|_{\text{lip}}&\leq\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}\\\ &\leq\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}_{i})\|_{2}+\|{\mathbf{g}}({\mathbf{x}}^{\prime})-{\mathbf{g}}({\mathbf{x}}_{j})\|_{2}+\|{\mathbf{g}}({\mathbf{x}}_{i})-{\mathbf{g}}({\mathbf{x}}_{j})\|_{2}\\\ &<\|{\mathbf{g}}({\mathbf{x}}_{i})-{\mathbf{g}}({\mathbf{x}}_{j})\|_{2}+2\varepsilon_{0}\|{\mathbf{g}}\|_{\text{lip}},\end{split}$ hence, $\|{\mathbf{g}}({\mathbf{x}}_{i})-{\mathbf{g}}({\mathbf{x}}_{j})\|_{2}\geq\varepsilon$. By assumption, this implies that $\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{i})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{j})\|_{2}\geq\gamma$ and (119) $\begin{split}\gamma&\leq\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{i})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{j})\|_{2}\\\ &\leq\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{i})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})\|_{2}+\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{j})\|_{2}+\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}\\\ &<2\varepsilon_{0}\|{\mathbf{m}}_{\mathscr{S}}\|_{\text{lip}}+\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2},\end{split}$ hence, $\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}>\gamma-2\varepsilon_{0}\|{\mathbf{m}}_{\mathscr{S}}\|_{\text{lip}}$ as claimed. ∎ ###### Proposition 4.6: Amplification Guarantee on Underlying Set. The result follows immediately from the triangle inequality. Let ${\mathbf{x}},{\mathbf{x}}^{\prime}\in\mathcal{X}$ and ${\mathbf{x}}_{i},{\mathbf{x}}_{j}\in\mathcal{X}_{N}$ so that $\|{\mathbf{x}}-{\mathbf{x}}_{i}\|_{2}<\varepsilon_{0}$ and $\|{\mathbf{x}}^{\prime}-{\mathbf{x}}_{j}\|_{2}<\varepsilon_{0}$, then (120) $\begin{split}\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}&\leq\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}_{i})\|_{2}+\|{\mathbf{g}}({\mathbf{x}}^{\prime})-{\mathbf{g}}({\mathbf{x}}_{j})\|_{2}+\|{\mathbf{g}}({\mathbf{x}}_{i})-{\mathbf{g}}({\mathbf{x}}_{j})\|_{2}\\\ &<2\varepsilon_{0}\|{\mathbf{g}}\|_{\text{lip}}+L\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{i})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{j})\|_{2}\\\ &\leq 2\varepsilon_{0}\|{\mathbf{g}}\|_{\text{lip}}+L\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{i})\|_{2}+L\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{j})\|_{2}\\\ &\hskip 56.9055pt+L\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}\\\ &<2\varepsilon_{0}\|{\mathbf{g}}\|_{\text{lip}}+2L\varepsilon_{0}\|{\mathbf{m}}_{\mathscr{S}}\|_{\text{lip}}+L\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}.\end{split}$ Gathering terms on $\varepsilon_{0}$ completes the proof. ∎ ###### Proposition 6.1: Noisy Separation Guarantee. Choose ${\mathbf{x}}_{i},{\mathbf{x}}_{j}\in\mathcal{X}_{N}$ and suppose that (121) $\|{\mathbf{g}}({\mathbf{x}}_{i})-{\mathbf{g}}({\mathbf{x}}_{j})\|_{2}\geq\varepsilon+2\delta_{v}.$ Then we have (122) $\begin{split}\|\left({\mathbf{g}}({\mathbf{x}}_{i})+{\mathbf{v}}_{i}\right)-\left({\mathbf{g}}({\mathbf{x}}_{j})+{\mathbf{v}}_{j}\right)\|_{2}&\geq\|{\mathbf{g}}({\mathbf{x}}_{i})-{\mathbf{g}}({\mathbf{x}}_{j})\|_{2}-\|{\mathbf{v}}_{i}\|-\|{\mathbf{v}}_{j}\|\\\ &\geq\|{\mathbf{g}}({\mathbf{x}}_{i})-{\mathbf{g}}({\mathbf{x}}_{j})\|_{2}-2\delta_{v}\\\ &\geq\varepsilon\end{split}$ By our assumption, this implies (123) $\|\left({\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{i})+{\mathbf{u}}_{i,\mathscr{S}}\right)-\left({\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{j})+{\mathbf{u}}_{j,\mathscr{S}}\right)\|_{2}\geq\gamma,$ and so we have (124) $\begin{split}\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{i})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{j})+{\mathbf{u}}_{j,\mathscr{S}}\|_{2}&\geq\|\left({\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{i})+{\mathbf{u}}_{i,\mathscr{S}}\right)-\left({\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{j})+{\mathbf{u}}_{j,\mathscr{S}}\right)\|_{2}-\|{\mathbf{u}}_{i,\mathscr{S}}\|-\|{\mathbf{u}}_{j,\mathscr{S}}\|\\\ &\geq\gamma-2\delta_{u}.\end{split}$ Therefore, we have established that (125) $\forall{\mathbf{x}}_{i},{\mathbf{x}}_{j}\in\mathcal{X}_{N}\qquad\|{\mathbf{g}}({\mathbf{x}}_{i})-{\mathbf{g}}({\mathbf{x}}_{j})\|_{2}\geq\varepsilon+2\delta_{v}\\\ \Rightarrow\quad\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{i})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{j})+{\mathbf{u}}_{j,\mathscr{S}}\|_{2}\geq\gamma-2\delta_{u}.$ The conclusion follows immediately by Proposition 4.5. ∎ ###### Theorem 5.5: Down-Sampled Amplification Guarantee. For simplicity, we will drop $L$ from the subscript on our objective since the threshold $L$ for the Lipschitz constant remains fixed throughout the proof. Let us begin by fixing a set $\mathscr{S}\subseteq\mathscr{M}$ and define the random variables (126) $Z_{\mathscr{S}}({\mathbf{b}}_{i})=\max_{{\mathbf{x}}\in\mathcal{X}_{N}}\mathbbm{1}\big{\\{}\|{\mathbf{g}}({\mathbf{b}}_{i})-{\mathbf{g}}({\mathbf{x}})\|_{2}>L\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{b}}_{i})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})\|_{2}\big{\\}},$ for $i=1,\ldots,m$. If $Z_{\mathscr{S}}({\mathbf{b}}_{i})=0$ then every secant between ${\mathbf{b}}_{i}$ and points of $\mathcal{X}_{N}$ satisfies the desired bound on the amplification. Otherwise, there is some point ${\mathbf{x}}\in\mathcal{X}$ for which (127) $\|{\mathbf{g}}({\mathbf{b}}_{i})-{\mathbf{g}}({\mathbf{x}})\|_{2}>L\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{b}}_{i})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})\|_{2}$ and so $Z_{\mathscr{S}}({\mathbf{b}}_{i})=1$. We observe that $Z_{\mathscr{S}}({\mathbf{b}}_{1}),\ldots,Z_{\mathscr{S}}({\mathbf{b}}_{m})$ are independent, identically distributed Bernoulli random variables whose expectation (128) $\mathbb{E}[Z_{\mathscr{S}}({\mathbf{b}}_{i})]=\mu\big{(}\big{\\{}{\mathbf{x}}\in\mathcal{X}\ :\ \exists{\mathbf{x}}_{j}\in\mathcal{X}_{N}\quad\mbox{s.t.}\\\ \|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}_{j})\|_{2}>L\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{j})\|_{2}\big{\\}}\big{)}$ is the $\mu$-measure of points in $\mathcal{X}$ that are not adequately separated from points in the $\varepsilon_{0}$-net $\mathcal{X}_{N}$ by the measurements ${\mathbf{m}}_{\mathscr{S}}$. Suppose that for a fixed ${\mathbf{x}}\in\mathcal{X}$ we have (129) $\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}_{j})\|_{2}\leq L\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{j})\|_{2}$ for every ${\mathbf{x}}_{j}\in\mathcal{X}_{N}$. By definition of $\mathcal{X}_{N}$, for any ${\mathbf{x}}^{\prime}\in\mathcal{X}$, there is an ${\mathbf{x}}_{j}\in\mathcal{X}_{N}$ with $\|{\mathbf{x}}^{\prime}-{\mathbf{x}}_{j}\|_{2}<\varepsilon_{0}$ and so we have (130) $\begin{split}\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}&\leq\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}_{j})\|_{2}+\|{\mathbf{g}}({\mathbf{x}}_{j})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}\\\ &<L\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{j})\|_{2}+\varepsilon_{0}\|{\mathbf{g}}\|_{\text{lip}}\\\ &\leq L\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}+L\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}_{j})\|_{2}+\varepsilon_{0}\|{\mathbf{g}}\|_{\text{lip}}\\\ &<L\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}+\left(\|{\mathbf{g}}\|_{\text{lip}}+L\|{\mathbf{m}}_{\mathscr{S}}\|_{\text{lip}}\right)\varepsilon_{0}.\end{split}$ It follows that $\mathbb{E}[Z_{\mathscr{S}}({\mathbf{b}}_{i})]$ is an upper bound on the $\mu$-measure of points in $\mathcal{X}$ for which the relaxed amplification threshold is exceeded, that is, (131) $\mathbb{E}[Z_{\mathscr{S}}({\mathbf{b}}_{i})]\geq\mu\big{(}\big{\\{}{\mathbf{x}}\in\mathcal{X}\ :\ \exists{\mathbf{x}}^{\prime}\in\mathcal{X}\quad\mbox{s.t.}\quad\|{\mathbf{g}}({\mathbf{x}})-{\mathbf{g}}({\mathbf{x}}^{\prime})\|_{2}\\\ \geq L\|{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}})-{\mathbf{m}}_{\mathscr{S}}({\mathbf{x}}^{\prime})\|_{2}+\left(\|{\mathbf{g}}\|_{\text{lip}}+L\|{\mathbf{m}}_{\mathscr{S}}\|_{\text{lip}}\right)\varepsilon_{0}\big{\\}}\big{)}.$ By assumption, we have a set $\mathscr{S}\subseteq\mathscr{M}$ so that $Z_{\mathscr{S}}({\mathbf{b}}_{i})=0$ for each $i=1,\ldots,m$. And so it remains to bound the difference between the empirical and true expectation of $Z_{\mathscr{S}}({\mathbf{b}}_{i})$ uniformly over every subset $\mathscr{S}\subseteq\mathscr{M}$. For fixed $\mathscr{S}$, the one-sided Hoeffding inequality gives (132) $\mathbb{P}\Big{\\{}\frac{1}{m}\sum_{i=1}^{m}\left(\mathbb{E}[Z_{\mathscr{S}}({\mathbf{b}}_{i})]-Z_{\mathscr{S}}({\mathbf{b}}_{i})\right)\geq\delta\Big{\\}}\leq e^{-2m\delta^{2}}.$ Unfixing $\mathscr{S}$ via the union bound over all $\mathscr{S}\subseteq\mathscr{M}$ and applying our assumption about the number of base points $m$ yields (133) $\mathbb{P}\bigcup_{\mathscr{S}\subseteq\mathscr{M}}\Big{\\{}\frac{1}{m}\sum_{i=1}^{m}\left(\mathbb{E}[Z_{\mathscr{S}}({\mathbf{b}}_{i})]-Z_{\mathscr{S}}({\mathbf{b}}_{i})\right)\geq\delta\Big{\\}}\leq e^{\\#(\mathscr{M})\ln{2}-2m\delta^{2}}\leq p.$ Since our assumed choice of $\mathscr{S}$ has $f_{m}(\mathscr{S})=f_{m}(\mathscr{M})$ it follows that all $Z_{\mathscr{S}}({\mathbf{b}}_{i})=0$, $i=1,\ldots,m$, hence we have (134) $\mathbb{E}[Z_{\mathscr{S}}({\mathbf{b}}_{i})]<\delta$ with probability at least $1-p$. Combining this with Eq. 131 completes the proof. ∎ ## Appendix D Description of the Accelerated Greedy Algorithm Since each objective function $f$ presented in Section 4 is submodular, it is possible to use an “accelerated greedy” (AG) algorithm to obtain the same solution as the naive greedy algorithm with a provably minimal number of objective function evaluations compared to a broad class of algorithms [35]. Let the increase in the objective function obtained by adding the sensor $j$ to the set $\mathscr{S}$ be called $\Delta_{j}(\mathscr{S})=f(\mathscr{S}\cup\\{j\\})-f(\mathscr{S})$. Instead of evaluating $\Delta_{j}(\mathscr{S}_{k-1})$ for every measurement in $\mathscr{M}\setminus\mathscr{S}_{k-1}$, AG keeps track of an upper bound $\hat{\Delta}_{j}\geq\Delta_{j}(\mathscr{S}_{k-1})$ on the increments for each sensor. Since submodularity of $f$ means that the increments $\Delta_{j}(\mathscr{S})$ can only decrease as the size of $\mathscr{S}$ increases, it is sufficient to have the maximum upper bound $\hat{\Delta}_{j^{*}}\geq\hat{\Delta}_{j}$, $\forall j\in\mathscr{M}\setminus\mathscr{S}_{k-1}$ be tight $\hat{\Delta}_{j^{*}}=\Delta_{j^{*}}(\mathscr{S}_{k-1})$ in order to conclude that $\Delta_{j^{*}}(\mathscr{S}_{k-1})$ is the largest increment. The rest of the upper bounds on the increments can remain loose since they are smaller than the tight maximum upper bound. The AG algorithm finds largest upper bound $\hat{\Delta}_{j^{*}}$ and updates it so that it is tight. If $\hat{\Delta}_{j^{*}}$ is still the greatest upper bound, then $j^{*}=j_{k}$ achieves the largest increment and is added to $\mathscr{S}_{k-1}$. Otherwise if $\hat{\Delta}_{j^{*}}$ is no longer the largest upper bound, the new largest upper bound is selected and process repeated until a tight maximum upper bound is obtained.
# $\ell$-covering $k$-hypergraphs are quasi-eulerian Mateja Šajna111Corresponding author. Email<EMAIL_ADDRESS>Mailing address: Department of Mathematics and Statistics, University of Ottawa, 150 Louis-Pasteur Private, Ottawa, ON, K1N 6N5, Canada. and Andrew Wagner University of Ottawa ###### Abstract An Euler tour in a hypergraph $H$ is a closed walk that traverses each edge of $H$ exactly once, and an Euler family is a family of closed walks that jointly traverse each edge of $H$ exactly once. An $\ell$-covering $k$-hypergraph, for $2\leq\ell<k$, is a $k$-uniform hypergraph in which every $\ell$-subset of vertices lie together in at least one edge. In this paper we prove that every $\ell$-covering $k$-hypergraph, for $k\geq 3$, admits an Euler family. Keywords: $\ell$-covering hypergraph; Euler family; Euler tour; Lovasz’s $(g,f)$-factor theorem. ## 1 Introduction The complete characterization of graphs that admit an Euler tour is a classic result covered by any introductory graph theory course. The concept naturally extends to hypergraphs; that is, an Euler tour of a hypergraph is a closed walk that traverses every edge exactly once. However, the study of eulerian hypegraphs is a much newer and largely unexplored territory. The first results on Euler tours in hypergraphs were obtained by Lonc and Naroski [4]. Most notably, they showed that the problem of existence of an Euler tour is NP-complete on the set of $k$-uniform hypergraphs, for any $k\geq 3$, as well as when restricted to a particular subclass of 3-uniform hypergraphs. Bahmanian and Šajna [2] attempted a systematic study of eulerian properties of general hypergraphs; some of their techniques and results will be used in this paper. In particular, they introduced the notion of an Euler family — a collection of closed walks that jointly traverse each edge exactly once — and showed that the problem of existence of an Euler family is polynomial on the class of all hypergraphs. In this paper, we define an $\ell$-covering $k$-hypergraph, for $2\leq\ell<k$, to be a non-empty $k$-uniform hypergraph in which every $\ell$-subset of vertices appear together in at least one edge. In [2], the authors proved that every 2-covering 3-hypergraph with at least two edges admits an Euler family, and the present authors gave a short proof [6] to show that every triple system — that is, a 3-uniform hypergraph in which every pair of vertices lie together in the same number of edges — admits an Euler tour as long as it has at least two edges. Most recently, the present authors proved the following result. ###### Theorem 1.1. [7] Let $k\geq 3$, and let $H$ be a $(k-1)$-covering $k$-hypergraph. Then $H$ admits an Euler tour if and only if it has at least two edges. In this paper, we aim to extend Theorem 1.1 to all $\ell$-covering $k$-hypergraphs. Our main result is as follows. ###### Theorem 1.2. Let $\ell$ and $k$ be integers, $2\leq\ell<k$, and let $H$ be an $\ell$-covering $k$-hypergraph. Then $H$ admits an Euler family if and only if it has at least two edges. As the concept of an Euler family is a relaxation of the concept of an Euler tour, the conclusion of Theorem 1.2 is weaker than that of Theorem 1.1; however, it holds for a much larger class of hypergraphs. We prove Theorem 1.2 by induction on $\ell$. The base case $\ell=2$ is stated as Theorem 5.1; its proof is essentially a counting argument and requires most of the work. The main part of the proof is presented in Section 5, while some special cases and technical details are handled in Sections 3 and 4. In particular, in Section 4, using the Lovasz $(g,f)$-factor theorem, we develop a sufficient condition for a $k$-uniform hypergraph without cut edges to admit an Euler family. ## 2 Preliminaries We use hypergraph terminology established in [1, 2], which applies to loopless graphs as well. Any graph theory terms not explained here can be found in [3]. A hypergraph $H$ is a pair $(V,E)$, where $V$ is a non-empty set, and $E$ is a multiset of elements from $2^{V}$. The elements of $V=V(H)$ and $E=E(H)$ are called the vertices and edges of $H$, respectively. The order of $H$ is $|V|$, and the size is $|E|$. A hypergraph of order 1 is called trivial, and a hypergraph with no edges is called empty. Distinct vertices $u$ and $v$ in a hypergraph $H=(V,E)$ are called adjacent (or neighbours) if they lie in the same edge, while a vertex $v$ and an edge $e$ are said to be incident if $v\in e$. The degree of $v$ in $H$, denoted $\deg_{H}(v)$, is the number of edges of $H$ incident with $v$. An edge $e$ is said to cover the vertex pair $\\{u,v\\}$ if $\\{u,v\\}\subseteq e$. A hypergraph $H$ is called $k$-uniform if every edge of $H$ has cardinality $k$. ###### Definition 2.1. Let $\ell$ and $k$ be integers, $2\leq\ell<k$. An $\ell$-covering $k$-hypergraph is a $k$-uniform hypergraph in which every $\ell$-subset of vertices lie together in at least one edge. The incidence graph of a hypergraph $H=(V,E)$ is a bipartite simple graph $G$ with vertex set $V\cup E$ and bipartition $\\{V,E\\}$ such that vertices $v\in V$ and $e\in E$ of $G$ are adjacent if and only if $v$ is incident with $e$ in $H$. The elements of $V$ and $E$ are called v-vertices and e-vertices of $G$, respectively. A hypergraph $H^{\prime}=(V^{\prime},E^{\prime})$ is called a subhypergraph of the hypergraph $H=(V,E)$ if $V^{\prime}\subseteq V$ and $E^{\prime}=\\{e\cap V^{\prime}:e\in E^{\prime\prime}\\}$ for some submultiset $E^{\prime\prime}$ of $E$. For $e\in E$, the symbol $H{\backslash}e$ denotes the subhypergraph $(V,E-\\{e\\})$ of $H$, and for $v\in V$, the symbol $H-v$ denotes the subhypergraph $(V-\\{v\\},E^{\prime})$ where $E^{\prime}=\\{e-\\{v\\}:e\in E,e-\\{v\\}\neq\emptyset\\}$. A $(v_{0},v_{k})$-walk in $H$ is a sequence $W=v_{0}e_{1}v_{1}e_{2}\ldots e_{k}v_{k}$ such that $v_{0},\ldots,v_{k}\in V$; $e_{1},\ldots,e_{k}\in E$; and $v_{i-1},v_{i}\in e_{i}$ with $v_{i-1}\neq v_{i}$ for all $i=1,\ldots,k$. A walk is said to traverse each of the vertices and edges in the sequence. The vertices $v_{0},v_{1},\ldots,v_{k}$ are called the anchors of $W$. If $e_{1},e_{2},\ldots,e_{k}$ are pairwise distinct, then $W$ is called a trail (strict trail in [1, 2]); if $v_{0}=v_{k}$ and $k\geq 2$, then $W$ is closed. A hypergraph $H$ is connected if every pair of vertices are connected in $H$; that is, if for any pair $u,v\in V(H)$, there exists a $(u,v)$-walk in $H$. A connected component of $H$ is a maximal connected subhypergraph of $H$ without empty edges. The number of connected components of $H$ is denoted by ${\rm c}(H)$. We call $v\in V(H)$ a cut vertex of $H$, and $e\in E(H)$ a cut edge of $H$, if ${\rm c}(H-v)>{\rm c}(H)$ and ${\rm c}(H{\backslash}e)>{\rm c}(H)$, respectively. An Euler family of a hypergraph $H$ is a collection of pairwise anchor- disjoint and edge-disjoint closed trails that jointly traverse every edge of $H$, and an Euler tour is a closed trail that traverses every edge of $H$. A hypergraph that is either empty or admits an Euler tour (family) is called eulerian (quasi-eulerian). Note that an Euler tour corresponds to an Euler family of cardinality 1, so every eulerian hypergraph is also quasi-eulerian. The following theorem allows us to determine whether a hypergraph is eulerian or quasi-eulerian from its incidence graph. ###### Theorem 2.2. [2, Theorem 2.18] Let $H$ be a hypergraph and $G$ its incidence graph. Then the following hold. (1) $H$ is quasi-eulerian if and only if $G$ has a spanning subgraph $G^{\prime}$ such that $\deg_{G^{\prime}}(e)=2$ for all $e\in E(H)$, and $\deg_{G^{\prime}}(v)$ is even for all $v\in V(H)$. (2) $H$ is eulerian if and only if $G$ has a spanning subgraph $G^{\prime}$ with at most one non-trivial connected component such that $\deg_{G^{\prime}}(e)=2$ for all $e\in E(H)$, and $\deg_{G^{\prime}}(v)$ is even for all $v\in V(H)$. ## 3 Technical Lemmas In this section, we take care of some special cases and prove some technical results that will aid in the proof of our base case, Theorem 5.1. ###### Lemma 3.1. Let $k\geq 4$, and let $H$ be a 2-covering $k$-hypergraph with at least 2 edges. Then $H$ has no cut edges. ###### Proof. Suppose $e$ is a cut edge of $H$. Then there exist vertices $u,v\in e$ that are disconnected in $H{\backslash}e$. Since $H$ has at least 2 edges, it must be that $k\neq|V(H)|$ and $e\neq V(H)$. Hence there exists $w\in V(H)-e$. Let $e_{1},e_{2}$ be edges of $H$ containing $u$ and $w$, and $v$ and $w$, respectively. As $e\not\in\\{e_{1},e_{2}\\}$, we can see that $ue_{1}we_{2}v$ is a $(u,v)$-walk in $H{\backslash}e$, a contradiction. ∎ ###### Lemma 3.2. Let $k\geq 4$, and let $H$ be a 2-covering $k$-hypergraph of order $n>\frac{3k}{2}$ and size $m\geq 2$. Then $m\geq 2\lfloor\frac{n+3}{k}\rfloor$. ###### Proof. If $n\leq 2k-4$, then $2\lfloor\frac{n+3}{k}\rfloor\leq 2\leq m$. Hence assume $n\geq 2k-3$. Suppose first that $n\geq 3k-3$. Since there are $\binom{n}{2}$ pairs of vertices to cover, and each edge covers $\binom{k}{2}$ pairs, we know that $m\geq\frac{n(n-1)}{k(k-1)}$. As $k\geq 4$, we have $\displaystyle m\geq\frac{n(n-1)}{k(k-1)}$ $\displaystyle\geq\frac{(3k-3)(n-1)}{k(k-1)}=\frac{3(n-1)}{k}=\frac{2n+n-3}{k}$ $\displaystyle\geq\frac{2n+3k-6}{k}\geq\frac{2n+6}{k}\geq 2\Big{\lfloor}\frac{n+3}{k}\Big{\rfloor}.$ Finally, assume $2k-3\leq n\leq 3k-4$. As $2\lfloor\frac{n+3}{k}\rfloor\leq 4$, it suffices to show that $m\geq 4$. Suppose $m\leq 3$. Since $H$ is a 2-covering $k$-hypergraph with $n>k$ and $m\geq 2$, every vertex has degree at least 2. Thus $2n\leq\sum_{v\in V(H)}\deg(v)=km\leq 3k$ and $n\leq\frac{3k}{2}$, contradicting the assumption that $n>\frac{3k}{2}$. Therefore, in all cases we have $m\geq 2\lfloor\frac{n+3}{k}\rfloor.$ ∎ ###### Lemma 3.3. Let $H$ be a hypergraph with $|E(H)|\geq 2$ satifying the following. * • For all $e,f\in E(H)$, we have $|e\cap f|\geq 2$; and * • there exist distinct $e,f\in E(H)$ such that $|e\cap f|\geq 3$. Then $H$ is eulerian. ###### Proof. Let $E(H)=\\{e_{1},\ldots,e_{m}\\}$ and assume $e_{1}$ and $e_{m}$ are distinct edges such that $|e_{1}\cap e_{m}|\geq 3$. Take any $v_{1}\in e_{1}\cap e_{2}$. For $i=2,\ldots,m-1$, let $v_{i}$ be a vertex in $(e_{i}\cap e_{i+1})-\\{v_{i-1}\\}$, and let $v_{0}\in(e_{1}\cap e_{m})-\\{v_{1},v_{m-1}\\}$. It is easy to verify that $v_{0}e_{1}v_{1}\ldots v_{m-1}e_{m}v_{0}$ is an Euler tour of $H$. ∎ ###### Corollary 3.4. Let $H$ be a 2-covering $k$-hypergraph of order $n$. If $n\leq 2k-3$ or $(k,n)=(4,6)$, then $H$ is eulerian. ###### Proof. If $n\leq 2k-3$, then every pair of edges $e,f\in E(H)$ satisfies $|e\cap f|\geq 3$, so $H$ is eulerian by Lemma 3.3. Assume now that $(k,n)=(4,6)$. For all $e,f\in E(H)$, we have $|e\cap f|\geq 2$. If there exist distinct edges $e,f\in E(H)$ such that $|e\cap f|\geq 3$, then $H$ is eulerian by Lemma 3.3. Hence assume $|e\cap f|=2$ for all $e,f\in E(H)$, and let $V(H)=\\{v_{1},\ldots,v_{6}\\}$. It is not difficult to see that we must have $E(H)=\\{e_{1},e_{2},e_{3}\\}$ where, without loss of generality, the edges are $e_{1}=v_{1}v_{2}v_{3}v_{4}$, $e_{2}=v_{1}v_{2}v_{5}v_{6}$, and $e_{3}=v_{3}v_{4}v_{5}v_{6}$. It follows that $W=v_{3}e_{1}v_{2}e_{2}v_{5}e_{3}v_{3}$ is an Euler tour of $H$. ∎ ###### Lemma 3.5. Let $n,k,q\in\mathbb{Z}^{+}$ be such that $n\geq qk$. Let $S=\big{\\{}(x_{1},\ldots,x_{q})\in(\mathbb{Z}^{+})^{q}:x_{1}+\cdots+x_{q}=n,x_{i}\geq k\mbox{ for all }i\big{\\}},$ and define $f:S\to\mathbb{Z}^{+}$ by $f(x_{1},\ldots,x_{q})=\binom{x_{1}}{2}+\cdots+\binom{x_{q}}{2}$. Then $f$ attains its maximum on $S$ at the point $\big{(}k,\ldots,k,n-k(q-1)\big{)}$. ###### Proof. Since the domain $S$ is finite, function $f$ indeed attains a maximum on $S$. Let ${\bf x}=(x_{1},\ldots,x_{q})\in S$ be such that $f({\bf x})$ is maximum. By symmetry of $f$, we may assume that $x_{1}\leq x_{2}\leq\ldots\leq x_{q}$. As $x_{1}\geq k$ and $x_{q}=n-(x_{1}+\ldots+x_{q-1})$, we observe that $x_{q}\leq n-k(q-1)$. Suppose that $x_{q}<n-k(q-1)$. Then there exists $i\in\\{1,\ldots,q-1\\}$ such that $x_{i}>k$. Let $i$ be the smallest index with this property, and let ${\bf y}=(x_{1},\ldots,x_{i-1},x_{i}-1,x_{i+1},\ldots,x_{q-1},x_{q}+1).$ Then ${\bf y}\in S$ and $\displaystyle f({\bf y})=$ $\displaystyle\sum_{\begin{subarray}{c}j=1\\\ j\neq i\end{subarray}}^{q-1}\binom{x_{j}}{2}+\binom{x_{i}-1}{2}+\binom{x_{q}+1}{2}$ $\displaystyle=$ $\displaystyle\sum_{\begin{subarray}{c}j=1\\\ j\neq i\end{subarray}}^{q-1}\binom{x_{j}}{2}+\frac{x_{i}(x_{i}-1)}{2}-\frac{2(x_{i}-1)}{2}+\frac{x_{q}(x_{q}-1)}{2}+\frac{2x_{q}}{2}$ $\displaystyle=$ $\displaystyle\sum_{j=1}^{q}\binom{x_{j}}{2}+(x_{q}-x_{i}+1)>f({\bf x}),$ contradicting the choice of ${\bf x}$. Hence $x_{q}=n-k(q-1)$, and consequently $x_{1}=\ldots=x_{q-1}=k$. Thus $f$ attains its maximum on $S$ at the point ${\bf x}=\big{(}k,\ldots,k,n-k(q-1)\big{)}$ as claimed. ∎ ## 4 A sufficient condition In this section, we state and prove Proposition 4.2, which gives a sufficient condition for a $k$-uniform hypergraph to admit an Euler family. This sufficient condition will be our main tool in the proof of Theorem 5.1. It is based on the $(g,f)$-factor theorem by Lovász [5], stated below as Theorem 4.1. For a graph $G$ and functions $f,g:V(G)\to\mathbb{N}$, a $(g,f)$-factor of $G$ is a spanning subgraph $F$ of $G$ such that $g(x)\leq\deg_{F}(x)\leq f(x)$ for all $x\in V(G)$. An $f$-factor is simply an $(f,f)$-factor. For any sets $U,W\subseteq V(G)$, let $\varepsilon_{G}(U,W)$ denote the number of edges of $G$ with one endpoint in $U$ and the other in $W$. ###### Theorem 4.1. [5] Let $G=(V,E)$ be a graph and $f,g:V\rightarrow\mathbb{N}$ be functions such that $g(x)\leq f(x)$ and $g(x)\equiv f(x)$ (mod 2) for all $x\in V$. Then $G$ has a $(g,f)$-factor $F$ such that $\deg_{F}(x)\equiv f(x)$ (mod 2) for all $x\in V$ if and only if, for all disjoint $S,T\subseteq V$, we have $\sum_{x\in S}f(x)+\sum_{x\in T}(\deg_{G}(x)-g(x))-\varepsilon_{G}(S,T)-q(S,T)\geq 0,$ (1) where $q(S,T)$ is the number of connected components $C$ of $G-(S\cup T)$ such that $\sum_{x\in V(C)}f(x)+\varepsilon_{G}(V(C),T)\equiv 1\text{ (mod 2).}$ ###### Proposition 4.2. Let $k\geq 3$, and let $H=(V,E)$ be a $k$-uniform hypergraph of order $n$ and size $m$. Let $G$ be the incidence graph of $H$, and $G^{*}$ the graph obtained from $G$ by appending $2(m+n)^{2}$ loops to every v-vertex. Assume that $H$ has no cut edges and that for all $X\subseteq E$ with $|X|\geq 2$, we have that $|X|\geq 2\lfloor\frac{{\rm c}(G^{*}-X)+3}{k}\rfloor$. Then $H$ is quasi-eulerian. ###### Proof. Let $r=2(m+n)^{2}$, and define $f:V(G^{*})\rightarrow\mathbb{Z}$ by $f(x)=\left\\{\begin{array}[]{l l}r&\mbox{ if }x\in V,\\\ 2&\mbox{ if }x\in E.\end{array}\right.$ We shall use Theorem 4.1 to show that $G^{*}$ has an $(f,f)$-factor, so let $S,T\subseteq V(G^{*})$ be disjoint sets, and denote $\gamma(S,T)=\sum_{x\in S}f(x)+\sum_{x\in T}(\deg_{G^{*}}(x)-f(x))-\varepsilon_{G^{*}}(S,T)-q(S,T),$ where $q(S,T)$ is the number of connected components $C$ of $G^{*}-(S\cup T)$ such that $\varepsilon_{G^{*}}(V(C),T)$ is odd. Observe that Condition (1) for $G^{*}$ with $g=f$ is equivalent to $\gamma(S,T)\geq 0$. Since $G$ is a subgraph of $K_{n,m}$, we have $\varepsilon_{G^{*}}(S,T)\leq mn$ and $q(S,T)\leq m+n$, and therefore $\varepsilon_{G^{*}}(S,T)+q(S,T)\leq(m+n)^{2}=\frac{r}{2}$. In addition, we have $\deg_{G^{*}}(x)-f(x)\geq r$ for all $x\in V$, and $\deg_{G^{*}}(x)-f(x)\geq k-2$ for all $x\in E$. Case 1: $(S\cup T)\cap V\neq\emptyset$. If $S\cap V\neq\emptyset$, then $\displaystyle\sum_{x\in S}f(x)\geq r$, and if $T\cap V\neq\emptyset$, then $\displaystyle\sum_{x\in T}(\deg_{G^{*}}(x)-f(x))\geq r$. Thus, in both cases $\displaystyle\gamma(S,T)$ $\displaystyle=\Big{(}\sum_{x\in S}f(x)+\sum_{x\in T}(\deg_{G^{*}}(x)-f(x))\Big{)}-\Big{(}\varepsilon_{G^{*}}(S,T)+q(S,T)\Big{)}$ $\displaystyle\geq r-\frac{r}{2}\geq 0.$ Case 2: $(S\cup T)\cap V=\emptyset$. Then $\varepsilon_{G^{*}}(S,T)=0$ since $S\cup T\subseteq E$. First, suppose $T=\emptyset$. Then $\varepsilon_{G^{*}}(V(C),T)=0$ for all connected components $C$ of $G^{*}-(S\cup T)$, so $q(S,T)=0$. Hence $\displaystyle\gamma(S,T)=\sum_{x\in S}f(x)\geq 0$. Next, suppose $S=\emptyset$ and $|T|=1$. Then $S\cup T=\\{e\\}$ for some $e\in E$. By assumption, edge $e$ is not a cut edge of $H$ and hence by [1, Theorem 3.23], e-vertex $e$ is not a cut vertex of $G^{*}$, and $G^{*}-(S\cup T)$ is connected. It follows that $q(S,T)\leq 1$ and $\gamma(S,T)=(\deg_{G^{*}}(e)-f(e))-q(S,T)\geq(k-2)-1\geq 0.$ We may now assume that $T\neq\emptyset$ and $|S\cup T|\geq 2$. Since each connected component $C$ of $G^{*}-(S\cup T)$ with $\varepsilon_{G^{*}}(V(C),T)$ odd corresponds to at least one edge incident with a vertex in $T$, the number of such components is at most $k|T|$. Hence $q(S,T)\leq\min\\{{\rm c}(G^{*}-(S\cup T)),k|T|\\}$, and $\displaystyle\gamma(S,T)$ $\displaystyle=2|S|+(k-2)|T|-q(S,T)$ $\displaystyle\geq 2|S\cup T|+(k-4)|T|-\min\\{{\rm c}(G^{*}-(S\cup T)),k|T|\\}.$ (2) Define $t=\lfloor\frac{{\rm c}(G^{*}-(S\cup T))+3}{k}\rfloor$, so that $kt-3\leq{\rm c}(G^{*}-(S\cup T))\leq kt+k-4.$ If $|T|\geq t+1$, then $\min\\{{\rm c}(G^{*}-(S\cup T)),k|T|\\}={\rm c}(G^{*}-(S\cup T))\leq kt+k-4$, so Inequality (2) yields $\gamma(S,T)\geq 2|S\cup T|+(k-4)(t+1)-(kt+k-4)=2|S\cup T|-4t.$ The same bound is obtained if $|T|\leq t$: in this case, we have $\min\\{{\rm c}(G^{*}-(S\cup T)),k|T|\\}\leq k|T|$, so that (2) yields $\gamma(S,T)\geq 2|S\cup T|+(k-4)|T|-k|T|=2|S\cup T|-4|T|\geq 2|S\cup T|-4t.$ In both cases, as $S\cup T\subseteq E$ and $|S\cup T|\geq 2$, the assumption of the proposition implies $|S\cup T|\geq 2\lfloor\frac{{\rm c}(G^{*}-(S\cup T))+3}{k}\rfloor=2t$, so that $\gamma(S,T)\geq 0$. Therefore, $\gamma(S,T)\geq 0$ for all disjoint $S,T\subseteq V(G^{*})$, and by Theorem 4.1, we conclude that $G^{*}$ has an $(f,f)$-factor $F$. Deleting the loops of $F$, we obtain a spanning subgraph $F^{\prime}$ of $G$ in which all v-vertices have even degree and all e-vertices have degree 2. Thus $H$ admits an Euler family by Theorem 2.2. ∎ ## 5 Proof of the main result We shall now prove our main result, Theorem 1.2. We use induction on $\ell$, and most of the work is required to prove the basis of induction, which we state below as Theorem 5.1. ###### Theorem 5.1. Let $k\geq 4$, and let $H$ be a 2-covering $k$-hypergraph with at least two edges. Then $H$ is quasi-eulerian. ###### Proof. Let $H=(V,E)$ with $n=|V|$ and $m=|E|$. If $n\leq 2k-3$, then $H$ is eulerian by Corollary 3.4, so we may assume that $n\geq 2k-2.$ If $n\leq\frac{3k}{2}$, it then follows that $(k,n)=(4,6)$. Again, $H$ is eulerian by Corollary 3.4. Hence $n>\frac{3k}{2}$, and Lemma 3.2 implies that $m\geq 2\big{\lfloor}\frac{n+3}{k}\big{\rfloor}$. In the rest of the proof we show that $H$ satisfies the conditions of Proposition 4.2. Let $G^{*}$ be the graph obtained from the incidence graph of $H$ by adjoining $2(m+n)^{2}$ loops to every v-vertex. Fix any $X\subseteq E$ with $|X|\geq 2$, and denote $q={\rm c}(G^{*}-X)$. Suppose that $|X|<2\Big{\lfloor}\frac{q+3}{k}\Big{\rfloor}$. If $q\leq 2k-4$, then this supposition implies that $|X|<2$, a contradiction. Hence we may assume that $q\geq 2k-3$, and hence $q\geq 5$. Moreover, our supposition implies $|X|\leq 2\frac{q+3}{k}-1.$ (3) Let $\ell$ denote the number of v-vertices that are isolated in $G^{*}-X$. Case 1: $\ell\geq 1$. If $\ell=n$, then $X=E$, $q=n$, and $|X|=|E|\geq 2\lfloor\frac{n+3}{k}\rfloor=2\lfloor\frac{q+3}{2}\rfloor$, contradicting our assumption on $X$. Thus we may assume $\ell<n$, and hence $\ell<q$. Since $G^{*}-X$ has $q-\ell$ non-trivial connected components, each with at least $k$ v-vertices, we have $n\geq\ell+k(q-\ell).$ (4) Since $q>\ell$, this inequality also implies $n\geq\ell+k.$ (5) Let $S$ be the set of pairs $\\{u,v\\}$ of v-vertices such that $u$ is isolated in $G^{*}-X$, and $v$ is not. Then $|S|=\ell(n-\ell)$. Observe that every edge of $H$ covers at most $\frac{k^{2}}{4}$ pairs from $S$, which implies that $|X|\geq\frac{\ell(n-\ell)}{\frac{k^{2}}{4}}$. Combining this inequality with (3), we obtain $\frac{4\ell(n-\ell)}{k^{2}}\leq\frac{2q+6-k}{k}.$ (6) Substituting $q\leq\ell+\frac{n-\ell}{k}$ from Inequality (4) and rearranging yields $n(4\ell-2)\leq 4\ell^{2}-k^{2}+2\ell k-2\ell+6k.$ Further substituting $n\geq\ell+k$ from (5) and isolating $\ell$, we obtain $\ell\leq 4-\frac{k}{2}$, which implies $\ell\in\\{1,2\\}$ as $k\geq 4$. However, if on the left-hand side of Inequality (6) we apply $\frac{n-\ell}{k}\geq q-\ell$ from (4) and simplify, then we obtain $(4\ell-2)q-4\ell^{2}\leq 6-k\leq 2.$ Now substituting either $\ell=1$ or $\ell=2$ yields $q\leq 3$, a contradiction. Case 2: $\ell=0$. Let $C_{1},C_{2},\dotso,C_{q}$ be the connected components of $G^{*}-X$, and let $n_{i}$ denote the number of v-vertices of $C_{i}$. Note that $n_{i}\geq k$ for all $i$. The number of pairs of v-vertices that lie in distinct connected components of $G^{*}-X$ is $\binom{n}{2}-\sum_{i=1}^{q}\binom{n_{i}}{2}$, and these pairs must all be covered by the edges of $X$. As $n\geq qk$, $n_{1}+\ldots+n_{q}=n$, and $n_{i}\geq k$, for all $i$,we know that $\sum_{i=1}^{q}\binom{n_{i}}{2}\leq(q-1)\binom{k}{2}+\binom{n-k(q-1)}{2}$ by Lemma 3.5. Therefore, $\binom{n}{2}-\sum_{i=1}^{q}\binom{n_{i}}{2}\geq\binom{n}{2}-(q-1)\binom{k}{2}-\binom{n-k(q-1)}{2}.$ Since each edge of $X$ covers up to $\binom{k}{2}$ pairs of v-vertices in distinct connected components, we deduce that $|X|\geq\frac{\binom{n}{2}-(q-1)\binom{k}{2}-\binom{n-k(q-1)}{2}}{\binom{k}{2}}.$ On the other hand, by (3), we have $|X|\leq\frac{2q+6-k}{k}$, so $\frac{\binom{n}{2}-(q-1)\binom{k}{2}-\binom{n-k(q-1)}{2}}{\binom{k}{2}}\leq\frac{2q+6-k}{k}.$ (7) We now substitute $x=q-1$, noting that $x\geq 4$ as $q\geq 5$. Rearranging Inequality (7), we then obtain $2kxn\leq k^{2}x^{2}+(k^{2}+2k-2)x-(k-8)(k-1).$ Applying $n\geq qk=(x+1)k$ further yields $k^{2}x^{2}+(k^{2}-2k+2)x+(k-8)(k-1)\leq 0.$ Denote the left-hand side by $f(x)=ax^{2}+bx+c$, where $a=k^{2}$, $b=k^{2}-2k+2$, and $c=(k-8)(k-1)$, and observe that $a,b>0$ as $k\geq 4$. If $b^{2}-4ac<0$, then $f(x)>0$ for all $x$, a contradiction. Hence assume $b^{2}-4ac\geq 0$. Let $x_{2}$ be the larger of the two roots of $f(x)=0$. If $x_{2}<4$, then $f(x)>0$ for all $x\geq 4$, a contradiction. Hence we must have $4\leq\frac{-b+\sqrt{b^{2}-4ac}}{2a}.$ Since $a,b>0$, it is straightforward to show that $16a+4b+c\leq 0$ follows. However, $16a+4b+c=k(21k-17)+16>0,$ a contradiction. Since each case leads to a contradiction, we conclude that $|X|\geq\lfloor\frac{{\rm c}(G^{*}-X)+3}{k}\rfloor$. By Lemma 3.1, hypergraph $H$ has no cut edges, so we may apply Proposition 4.2 to conclude that $H$ is quasi-eulerian. ∎ We are now ready to prove our main result, restated below. ###### Theorem 1.2. Let $\ell$ and $k$ be integers, $2\leq\ell<k$, and let $H$ be an $\ell$-covering $k$-hypergraph. Then $H$ is quasi-eulerian if and only if it has at least two edges. ###### Proof. Since $H$ is non-empty, and since a hypergraph with a single edge does not admit a closed trail, necessity is easy to see. To prove sufficiency, for $s\geq 1$ and $\ell\geq 2$, define the proposition $P_{s}(\ell):\mbox{ ``Every }\ell\mbox{-covering }(\ell+s)\mbox{-hypergraph with at least two edges is quasi-eulerian.''}$ Theorem 1.1 implies that $P_{1}(\ell)$ holds for all $\ell\geq 2$. Hence fix any $s\geq 2$. We prove $P_{s}(\ell)$ by induction on $\ell$. As $\ell+s\geq 4$, the basis of induction, $P_{s}(2)$, follows from Theorem 5.1. Suppose that, for some $\ell\geq 2$, the proposition $P_{s}(\ell)$ holds; that is, every $\ell$-covering $(\ell+s)$-hypergraph with at least two edges is quasi- eulerian. Let $H=(V,E)$ be an $(\ell+1)$-covering $\big{(}(\ell+1\big{)}+s)$-hypergraph with $|E|\geq 2.$ Fix any $v\in V$ and let $V^{*}=V-\\{v\\}$. Define a mapping $\varphi:E\to 2^{V^{*}}$ by $\varphi(e)=e-\\{v\\}\quad\mbox{ if }v\in e,$ and otherwise, $\varphi(e)=e-\\{u\\}\quad\mbox{ for any }u\in e.$ Then let $E^{*}=\\{\varphi(e):e\in E\\}$ and $H^{*}=(V^{*},E^{*})$, so that $\varphi$ is a bijection from $E$ to $E^{*}$. It is straightforward to verify that $H^{*}$ is an $\ell$-covering $(\ell+s)$-hypergraph. As $|E^{*}|=|E|\geq 2$, by induction hypothesis, hypergraph $H^{*}$ admits an Euler family $\mathcal{F}^{*}$. In each closed trail in $\mathcal{F}^{*}$, replace each $e\in E^{*}$ with $\varphi^{-1}(e)$ to obtain a set $\mathcal{F}$ of closed trails of $H$. It is not difficult to verify that $\mathcal{F}$ is an Euler family of $H$, so $P_{s}(\ell+1)$ follows. By induction, we conclude that $P_{s}(\ell)$ holds for all $\ell\geq 2$, and any $s\geq 1$. Therefore, every $\ell$-covering $k$-hypergraph with at least two edges is quasi-eulerian. ∎ ## Acknowledgements The first author gratefully acknowledges support by the Natural Sciences and Engineering Research Council of Canada (NSERC), Discovery Grant RGPIN-2016-04798. ## References * [1] M. A. Bahmanian and M. Šajna, Connection and separation in hypergraphs, Theory Appl. Graphs 2 (2015), Art. 5, 24 pp. * [2] M. A. Bahmanian and M. Šajna, Quasi-Eulerian hypergraphs, Elec. J. Combin. 24 (2017), #P3.30, 12 pp. * [3] J. A. Bondy and U. S. R. Murty, Graph Theory, Springer, 2008. * [4] Z. Lonc and P. Naroski, On tours that contain all edges of a hypergraph, Elec. J. Combin. 17 (2010), #R144, 31 pp. * [5] L. Lovász, The factorization of graphs II, Acta Math. Acad. Sci. Hungar. 23 (1972), 223–246. * [6] M. Šajna and A. Wagner, Triple systems are eulerian, J. Combin. Des. 25 (2017), 185–191. * [7] M. Šajna and A. Wagner, Covering hypergraphs are eulerian, submitted, 24 pp. arXiv:2101.04561.
# Using edge cuts to find Euler tours and Euler families in hypergraphs Mateja Šajna111Corresponding author. Email<EMAIL_ADDRESS>Mailing address: Department of Mathematics and Statistics, University of Ottawa, 150 Louis-Pasteur Private, Ottawa, ON, K1N 6N5, Canada. and Andrew Wagner University of Ottawa ###### Abstract An Euler tour in a hypergraph is a closed walk that traverses each edge of the hypergraph exactly once, while an Euler family is a family of closed walks that jointly traverse each edge exactly once and cannot be concatenated. In this paper, we show how the problem of existence of an Euler tour (family) in a hypergraph $H$ can be reduced to the analogous problem in some smaller hypergraphs that are derived from $H$ using an edge cut of $H$. In the process, new techniques of edge cut assignments and collapsed hypergraphs are introduced. Moreover, we describe algorithms based on these characterizations that determine whether or not a hypergraph admits an Euler tour (family), and can also construct an Euler tour (family) if it exists. Keywords: Hypergraph; Euler tour; Euler family; edge cut; edge cut assignment; collapsed hypergraph; algorithm. ## 1 Introduction It is common knowledge that a connected graph admits an Euler tour — that is, a closed walk traversing each edge of the graph exactly once — if and only if the graph has no vertices of odd degree. The notion of an Euler tour can be generalized to hypergraphs in the obvious way, and has been studied as such in [5, 1]. In addition, Bahmanian and Šajna [1] also introduced the notion of an Euler family, which is a family of closed walks that jointly traverse each edge of the hypergraph exactly once and cannot be concatenated. For a connected graph, an Euler family precisely corresponds to an Euler tour; for general connected hypergraphs, however, the two notions give rise to two rather distinct problems: Euler family, which is of polynomial complexity [1], and Euler tour, which is NP-complete [5]. The question of how to reduce the search for an Euler family or tour in a hypergraph to smaller hypergraphs is thus particularly pertinent in the case of Euler tours. An Euler tour (family) is called spanning if it traverses every vertex of the hypergraph. In [6], Steimle and Šajna showed how to use certain vertex cuts to reduce the problem of finding a spanning Euler tour (family) in a hypergraph $H$ to some smaller hypergraphs derived from subhypergraphs of $H$. In the present paper, with the same goal in mind, we shall use edge cuts instead. This new approach represents a major improvement over the results from [6]. First, it can be used for general Euler tours and families, not just for spanning Euler tours and families. Second, while the main results from [6] apply only to hypergraphs with vertex cuts of cardinality at most two, the present approach applies to edge cuts of any size, and hence to all hypergraphs, as every non-trivial hypergraph has an edge cut. In addition, we introduce new elegant techniques of edge cut assignments and collapsed hypergraphs, which will be useful in problems beyond the scope of this paper. After reviewing the terminology and providing some basic tools in Section 2, we shall introduce edge cuts in Section 3, and the first main tool — edge cut assignments — in Section 4. The key result of this section is Theorem 4.4, where we show that a hypergraph admits an Euler family (tour) if and only if there exists an edge cut assignment such that the associated hypergraph admits an Euler family (tour). This theorem forms the basis of all reduction results and algorithms. In Section 5, we present our first main reduction result. Namely, in Theorems 5.2 and 5.3 we show that a hypergraph $H$ with a minimal edge cut $F$ admits an Euler family (tour) if and only if certain subhypergraphs of H (obtained from the connected components of $H{\backslash}F$) admit an Euler family (tour). We begin Section 6 by introducing a collapsed hypergraph of a hypegraph $H$; that is, a hypergraph obtained from $H$ by “collapsing” (identifying) the vertices in a given subset of the vertex set of $H$. We then present our second main reduction result. In Theorem 6.2, we show that a hypergraph $H$ admits an Euler family if and only if certain collapsed hypergraphs obtained via an edge cut assignment admit Euler families. The analogous result for Euler tours is presented in Corollaries 6.3 and 6.4; the former contains the proof of necessity, while the latter contains the proof of sufficiency, but requires an additional assumption. Each of Sections 5 and 6 concludes with a description of pertinent algorithms that determine whether or not a hypergraph admits an Euler family (tour). Since all of our proofs are constructive, the algorithms can easily be modified to construct an Euler family (tour) if one exists. ## 2 Preliminaries We begin with some basic concepts related to hypergraphs, which will be used in later discussions. For any graph- and hypergraph-theoretic terms not defined here, we refer the reader to [3] and [2], respectively. A hypergraph $H$ is an ordered pair $(V,E)$, where $V$ is a non-empty finite set and $E$ is a finite multiset of $2^{V}$. To denote multisets, we shall use double braces, $\left\\{\\!\\!\left\\{.\right\\}\\!\\!\right\\}$. The elements of $V=V(H)$ and $E=E(H)$ are called vertices and edges, respectively. A hypergraph is said to be trivial if it has only one vertex, and empty if it has no edges. Let $H=(V,E)$ be a hypergraph, and $u,v\in V$. If $u\neq v$ and there exists an edge $e\in E$ such that $u,v\in e$, then we say that $u$ and $v$ are adjacent (via the edge $e$); this is denoted $u\sim_{H}v$. If $v\in V$ and $e\in E$ are such that $v\in e$, then $v$ is said to be incident with $e$, and the ordered pair $(v,e)$ is called a flag of $H$. The set of flags of $H$ is denoted by $F(H)$. The degree of a vertex $v\in V$ is the number of edges in $E$ incident with $v$, and is denoted by $\deg_{H}(v)$, or simply $\deg(v)$ when there is no ambiguity. A hypergraph $H^{\prime}=(V^{\prime},E^{\prime})$ is called a subhypergraph of the hypergraph $H=(V,E)$ if $V^{\prime}\subseteq V$ and $E^{\prime}=\left\\{\\!\\!\left\\{e\cap V^{\prime}:e\in E^{\prime\prime}\right\\}\\!\\!\right\\}$ for some submultiset $E^{\prime\prime}$ of $E$. For any subset $V^{\prime}\subseteq V$, we define the subhypergraph of $H$ induced by $V^{\prime}$ to be the hypergraph $(V^{\prime},E^{\prime})$ with $E^{\prime}=\left\\{\\!\\!\left\\{e\cap V^{\prime}:e\in E,e\cap V^{\prime}\neq\emptyset\right\\}\\!\\!\right\\}$. Thus, we obtain the subhypergraph induced by $V^{\prime}$ by deleting all vertices in $V-V^{\prime}$ from $V$ and from each edge of $H$, and subsequently deleting all empty edges. For any subset $E^{\prime}\subseteq E$, we denote the subhypergraph $(V,E-E^{\prime})$ of $H$ by $H{\backslash}E^{\prime}$, and for $e\in E$, we write $H{\backslash}e$ instead of $H{\backslash}\\{e\\}$. For any multiset $E^{\prime}$ of $2^{V}$, the symbol $H+E^{\prime}$ will denote the hypergraph obtained from $H$ by adjoining all edges in $E^{\prime}$. A $(v_{0},v_{k})$-walk of length $k$ in a hypergraph $H$ is an alternating sequence $W=v_{0}e_{1}v_{1}\ldots$ $v_{k-1}e_{k}v_{k}$ of (possibly repeated) vertices and edges such that $v_{0},\ldots,v_{k}\in V$, $e_{1},\ldots,e_{k}\in E$, and for each $i\in\\{1,\ldots,k\\}$, the vertices $v_{i-1}$ and $v_{i}$ are adjacent in $H$ via the edge $e_{i}$. Note that since adjacent vertices are by definition distinct, no two consecutive vertices in a walk can be the same. It follows that no walk in a hypergraph contains an edge of cardinality less than 2. The vertices in $V_{a}(W)=\\{v_{0},\ldots,v_{k}\\}$ are called the anchors of $W$, vertces $v_{0}$ and $v_{k}$ are the endpoints of $W$, and $v_{1},\ldots,v_{k-1}$ are the internal vertices of $W$. We also define the edge set of $W$ as $E(W)=\\{e_{1},\ldots,e_{k}\\}$, and the set of anchor flags of $W$ as $F(W)=\\{(v_{0},e_{1}),(v_{1},e_{1}),(v_{2},e_{2}),\ldots,(v_{k-1},e_{k}),(v_{k},e_{k})\\}$. Walks $W$ and $W^{\prime}$ in a hypergraph $H$ are said to be edge-disjoint if $E(W)\cap E(W^{\prime})=\emptyset$, and anchor-disjoint if $V_{a}(W)\cap V_{a}(W^{\prime})=\emptyset$. A walk $W=v_{0}e_{1}v_{1}\ldots v_{k-1}e_{k}v_{k}$ is called closed if $v_{0}=v_{k}$ and $k\geq 2$; a trail if the edges $e_{1},\ldots,e_{k}$ are pairwise distinct; a path if it is a trail and the vertices $v_{0},\ldots,v_{k}$ are pairwise distinct; and a cycle if it is a closed trail and the vertices $v_{0},\ldots,v_{k-1}$ are pairwise distinct. (Note that in [2], a “trail” was defined as a walk with no repeated anchor flags, and a walk with no repeated edges was called a “strict trail”. In this paper, we shall consider only strict trails, and hence use the shorter term “trail” to mean a “strict trail”.) A walk $W=v_{0}e_{1}v_{1}\ldots v_{k-1}e_{k}v_{k}$ is said to traverse a vertex $v$ and edge $e$ if $v\in V_{a}(W)$ and $e\in E(W)$, respectively. More specifically, the walk $W$ is said to traverse the edge $e$ via vertex $v$, as well as traverse vertex $v$ via edge $e$, if either $ev$ or $ve$ is a subsequence of $W$. Vertices $u$ and $v$ are connected in a hypergraph $H$ if there exists a $(u,v)$-walk — or equivalently, a $(u,v)$-path [2, Lemma 3.9] — in $H$, and $H$ itself is connected if every pair of vertices in $V$ are connected in $H$. The connected components of $H$ are the maximal connected subhypergraphs of $H$ without empty edges. The number of connected components of $H$ is denoted by $c(H)$. An Euler tour of a hypergraph $H$ is a closed trail of $H$ traversing every edge of $H$, and a hypergraph is said to be eulerian if it either admits an Euler tour or is trivial and empty. An Euler family of $H$ is a set of pairwise edge-disjoint and anchor-disjoint closed trails of $H$ jointly traversing every edge of $H$. Clearly, an Euler family of cardinality 1 corresponds to an Euler tour, and vice-versa. An Euler family $\cal F$ of $H$ is said to be spanning if every vertex of $H$ is an anchor of a closed trail in $\cal F$, and an Euler tour of $H$ is spanning if it traverses every vertex of $H$. The following two observations will be frequently used tools in the rest of the paper. Their easy proofs are omitted. ###### Lemma 2.1 Let $H$ be a hypergraph with connected components $H_{i},$ for $i\in I$. Then the following hold. 1. (i) If, for each $i\in I$, we have that $H_{i}$ has an Euler family ${\mathcal{F}}_{i}$, then $\bigcup_{i\in I}{\mathcal{F}}_{i}$ is an Euler family of $H$. If each ${\mathcal{F}}_{i}$ is spanning in $H_{i}$, then ${\mathcal{F}}$ is spanning in $H$. 2. (ii) If $H$ has an Euler family ${\mathcal{F}}$, then ${\mathcal{F}}$ has a partition $\\{{\mathcal{F}}_{i}:i\in I\\}$ such that, for each $i\in I$, we have that ${\mathcal{F}}_{i}$ is an Euler family of $H_{i}$. If ${\mathcal{F}}$ is spanning in $H$, then ${\mathcal{F}}_{i}$ is spanning in $H_{i}$, for each $i\in I$. ###### Lemma 2.2 Let $H_{1}$ and $H_{2}$ be hypergraphs such that $V(H_{1})\subseteq V(H_{2})$ and there exists a bijection $\varphi:E(H_{1})\to E(H_{2})$ satisfying $e\subseteq\varphi(e)$ for all $e\in E(H_{1})$. Then the following hold. 1. (i) If $H_{1}$ has an Euler family ${\mathcal{F}}_{1}$, then $H_{2}$ has an Euler family ${\mathcal{F}}_{2}$ obtained from ${\mathcal{F}}_{1}$ by replacing each edge $e$ with $\varphi(e)$. 2. (ii) If $H_{2}$ has an Euler family ${\mathcal{F}}_{2}$ such that for all $e\in E(H_{2})$ we have that $e$ is traversed in ${\mathcal{F}}_{2}$ via vertices in $\varphi^{-1}(e)$, then $H_{1}$ has an Euler family ${\mathcal{F}}_{1}$ obtained from ${\mathcal{F}}_{2}$ by replacing each edge $e$ with $\varphi^{-1}(e)$. ## 3 Edge cuts in hypergraphs As in graphs [3], an edge cut in a hypergraph $H=(V,E)$ is a set of edges of the form $[S,V-S]_{H}$ for some non-empty proper subset $S$ of $V$. Here, for any $S,T\subseteq V$, we denote $[S,T]_{H}=\\{e\in E:e\cap S\neq\emptyset\mbox{ and }e\cap T\neq\emptyset\\}.$ An edge $e$ in a hypergraph $H$ is said to be a cut edge of $H$ if $c(H{\backslash}e)>c(H)$. Thus, an edge $e$ is a cut edge if and only if $\\{e\\}$ is an edge cut. ###### Lemma 3.1 Let $H=(V,E)$ be a hypergraph and $F\subseteq E.$ Then $H{\backslash}F$ is disconnected if and only if $H$ has an edge cut $F^{\prime}\subseteq F$. Proof. Assume $H{\backslash}F$ is disconnected, let $H_{1}$ be one connected component of $H{\backslash}F$, and $S=V(H_{1})$. Then $F^{\prime}=[S,V-S]_{H}$ is an edge cut of $H$ contained in $F$. Conversely, assume $H$ has an edge cut $F^{\prime}\subseteq F$, where $F^{\prime}=[S,V-S]_{H}$ for $\emptyset\subsetneq S\subsetneq V$. Then $H{\backslash}F$ has no edge intersecting both $S$ and $V-S$, so it is disconnected. a As we shall see in the next lemma, minimal edge cuts — that is, edge cuts that are not properly contained in other edge cuts — have a very nice property that will be much exploited in subsequent results. Note that, for a hypergraph $H=(V,E)$ and its subhypergraph $H^{\prime}$, we say that an edge $e\in E$ intersects $H^{\prime}$ if $e\cap V(H^{\prime})\neq\emptyset$. ###### Lemma 3.2 Let $H=(V,E)$ be a hypergraph. An edge cut $F$ of $H$ is minimal if and only if every edge of $F$ intersects each connected component of $H{\backslash}F$. Proof. Assume $F$ is a minimal edge cut of $H$. Suppose $H_{1}$ is a connected component of $H{\backslash}F$, and $f\in F$ is such that $f$ does not intersect $H_{1}$. Let $S=V(H_{1})$ and $F^{\prime}=F-\\{f\\}$. Then $[S,V-S]_{H}\subseteq F^{\prime}\subsetneq F$, contradicting the minimality of $F$. Thus every edge of $F$ intersects each connected component of $H{\backslash}F$. Conversely, assume each edge of $F$ intersects every connected component of $H{\backslash}F$. Suppose $F^{\prime}$ is an edge cut of $H$ and $F^{\prime}\subsetneq F$. Take any $f\in F-F^{\prime}$. Since $f$ intersects every connected component of $H{\backslash}F$, the hypergraph $H+f=H{\backslash}F^{\prime}$ is connected. Hence by Lemma 3.1, the set $F^{\prime}$ contains no edge cut of $H$, a contradiction. Hence the edge cut $F$ is minimal. a ## 4 Edge Cut Assignments We are now ready to present the first of the two main tools that we shall use to study eulerian properties of hypergraphs via their edge cuts; namely, an edge cut assignment $\alpha$ associated with an edge cut $F$ of a hypergraph $H$, which maps each edge of $F$ to an unordered pair of connected components (or, more generally, unions of connected components) of $H{\backslash}F$. Note that, for a finite set $I$, we denote the set of all unordered pairs of elements in $I$, with repetitions in a pair permitted, by $I^{[2]}$; that is, $I^{[2]}=\\{ij:i,j\in I\\}$. Thus, $|I^{[2]}|=\frac{1}{2}n(n+1)$. ###### Definition 4.1 Let $H$ be a connected hypergraph with a minimal edge cut $F$, and let $\\{V_{i}:i\in I\\}$, be a partition of $V(H)$ such that each $V_{i}$ is a union of the vertex sets of the connected components of $H{\backslash}F$. * • A mapping $\alpha:F\to I^{[2]}$ is called an edge cut assignment (associated with $F$ and $H$) if, for all $f\in F$, we have that $\alpha(f)=ij$ implies that $f\cap V_{i}\neq\emptyset$ and $f\cap V_{j}\neq\emptyset$, and $\alpha(f)=i$ implies that $|f\cap V_{i}|\geq 2$. * • An edge cut assignment $\alpha:F\to I^{[2]}$ is called standard if $V_{i}$, for each $i\in I$, is the vertex set of a connected component of $H{\backslash}F$. * • Given an edge cut assignment $\alpha$, we define, for each $e\in E(H)$, $e^{\alpha}=\left\\{\begin{array}[]{ll}e\cap\left(V_{i}\cup V_{j}\right)&\mbox{ if }e\in F\mbox{ and }\alpha(e)=ij\\\ e&\mbox{ if }e\not\in F\end{array}\right..$ * • A hypergraph $H^{\alpha}$, defined by $V(H^{\alpha})=V(H)\quad\mbox{and}\quad E(H^{\alpha})=\left\\{\\!\\!\left\\{e^{\alpha}:e\in E(H)\right\\}\\!\\!\right\\},$ is called the hypergraph associated with $H$ and the edge cut assignment $\alpha$. * • A multigraph $G^{\alpha}$ with vertex set $V(G^{\alpha})=I$ and edge multiset $E(G^{\alpha})=\left\\{\\!\\!\left\\{\alpha(f):f\in F\right\\}\\!\\!\right\\}$ is called the multigraph associated with the edge cut assignment $\alpha$. Thus $\alpha$ can be viewed as a bijection $F\to E(G^{\alpha})$. The usefulness of these concepts will be conveyed in the following three observations, the last of which yields necessary and sufficient conditions for a hypergraph to admit an Euler family (tour) via an edge cut assignment $\alpha$ and the associated hypergraph $H^{\alpha}$. ###### Lemma 4.2 Let $H$ be a connected hypergraph with a minimal edge cut $F$, and let $\\{V_{i}:i\in I\\}$ be a partition of $V(H)$ into unions of the vertex sets of the connected components of $H{\backslash}F$. Furthermore, let $\alpha:F\to I^{[2]}$ be an edge cut assignment. If $H^{\alpha}$ has an Euler family (tour), then so does $G^{\alpha}$. Proof. Let ${\mathcal{F}}$ be an Euler family of $H^{\alpha}$, and ${\mathcal{F}}^{\prime}$ the subset of ${\mathcal{F}}$ consisting of all closed trails that traverse at least on edge of $F$. Take any closed trail $T\in{\mathcal{F}}^{\prime}$. Then, without loss of generality, $T$ is of the form $T=v_{0}T_{0}v_{0}^{\prime}f_{0}^{\alpha}v_{1}T_{1}v_{1}^{\prime}f_{1}^{\alpha}v_{2}\ldots v_{k-1}T_{k-1}v_{k-1}^{\prime}f_{k-1}^{\alpha}v_{0}$ for some $f_{0},\ldots,f_{k-1}\in F$ and $(v_{i},v_{i}^{\prime})$-trails $T_{i}$ in $H_{i}$, for $i\in\\{0,\ldots,k-1\\}$. Obtain the sequence $T^{\alpha}$ from $T$ by replacing each subsequence $v_{i}T_{i}v_{i}^{\prime}$ with $i$, and each edge $f_{i}^{\alpha}$ with $\alpha(f_{i})$. Then $T^{\alpha}$ is a closed trail in $G^{\alpha}$, and $\\{T^{\alpha}:T\in{\mathcal{F}}^{\prime}\\}$ is an Euler family of $G^{\alpha}$. Moreover, if $T$ is an Euler tour of $H^{\alpha}$, then $T^{\alpha}$ is Euler tour of $G^{\alpha}$. a ###### Lemma 4.3 Let $H$ be a connected hypergraph with a minimal edge cut $F$, and let $\\{V_{i}:i\in I\\}$ be a partition of $V(H)$ into unions of the vertex sets of the connected components of $H{\backslash}F$. 1. (i) Suppose $H$ has an Euler family ${\mathcal{F}}$. Let $\alpha:F\to I^{[2]}$ be an edge cut assignment defined by $\alpha(f)=ij$ if the edge $f\in F$ is traversed by a trail in ${\mathcal{F}}$ via a vertex in $V_{i}$ and a vertex in $V_{j}$ (where $i=j$ is possible). Then the hypergraph $H^{\alpha}$ has an Euler family obtained from ${\mathcal{F}}$ by replacing each $f\in F$ with $f^{\alpha}$. 2. (ii) Conversely, if for some edge cut assignment $\alpha:F\to I^{[2]}$, the associated hypergraph $H^{\alpha}$ has an Euler family ${\mathcal{F}}^{\alpha}$, then $H$ has an Euler family obtained from ${\mathcal{F}}^{\alpha}$ by replacing each $f^{\alpha}$, for $f\in F$, with $f$. Proof. 1. (i) Define a bijection $\varphi:E(H^{\alpha})\to E(H)$ by $\varphi(e^{\alpha})=e$, for all $e\in E(H)$. Then $e^{\alpha}\subseteq\varphi(e^{\alpha})$ for all $e^{\alpha}\in E(H^{\alpha})$. By Lemma 2.2(ii), since each edge $e$ of $H$ is traversed in ${\mathcal{F}}$ via vertices in $\varphi^{-1}(e)$, an Euler family of $H^{\alpha}$ is obtained from ${\mathcal{F}}$ by replacing each edge $e$ with $\varphi^{-1}(e)$, which effectively means replacing each $f\in F$ with $f^{\alpha}$. 2. (ii) Define $\varphi$ as in (i) and use Lemma 2.2(i). a ###### Theorem 4.4 Let $H$ be a connected hypergraph with a minimal edge cut $F$, and let $\\{V_{i}:i\in I\\}$ be a partition of $V(H)$ into unions of the vertex sets of the connected components of $H{\backslash}F$. Then 1. (i) $H$ has an Euler family if and only if for some edge cut assignment $\alpha:F\to I^{[2]}$, each connected component of $H^{\alpha}$ has an Euler family; and 2. (ii) $H$ has an Euler tour if and only if for some edge cut assignment $\alpha:F\to I^{[2]}$, the hypergraph $H^{\alpha}$ has a unique non-empty connected component, which has an Euler tour. Proof. Observe that, by Lemma 4.3, a hypergraph $H$ has an Euler family of cardinality $k$ if and only if for some edge cut assignment $\alpha$, the associated hypergraph $H^{\alpha}$ has an Euler family of cardinality $k$. 1. (i) Since by Lemma 2.1, $H^{\alpha}$ has an Euler family if and only if each of its connected components has an Euler family, the statement follows. 2. (ii) From the above observation, $H$ has an Euler tour if and only if $H^{\alpha}$, for some $\alpha$, has an Euler tour. Since it is clear that $H^{\alpha}$ has an Euler tour if and only if it has a unique non-empty connected component, which itself has an Euler tour, the statement follows. a We point out that Theorem 4.4 forms the basis for all other, more specific reductions, as well as for the algorithms presented in the rest of this paper. ## 5 Reduction using standard edge cut assignments In this section, we shall be using only standard edge cut assignments; that is, edge cut assignments arising from partitions of the form $\\{V(H_{i}):i\in I\\}$, where the $H_{i}$, for $i\in I$, are the connected components of $H{\backslash}F$, and $F$ is a minimal edge cut in a hypergraph $H$. ###### Lemma 5.1 Let $H$ be a connected hypergraph with a minimal edge cut $F$, and let $H_{i}$, for $i\in I$, be the connected components of $H{\backslash}F$. Furthermore, let $\alpha:F\to I^{[2]}$ be a standard edge cut assignment. 1. (i) If $G^{\prime}$ is a connected component of $G^{\alpha}$, then $H^{\alpha}[\bigcup_{i\in V(G^{\prime})}V(H_{i})]$ is a connected component of $H^{\alpha}$. 2. (ii) If $H^{\prime}$ is a connected component of $H^{\alpha}$, then $V(H^{\prime})=\bigcup_{i\in J}V(H_{i})$ for some $J\subseteq I$, and $G^{\alpha}[J]$ is a connected component of $G^{\alpha}$. Proof. 1. (i) Let $G^{\prime}$ be a connected component of $G^{\alpha}$, let $J=V(G^{\prime})$ and $H^{\prime}=H^{\alpha}[\bigcup_{i\in J}V(H_{i})]$. Take any $i,j\in J$ such that $i\sim_{G^{\alpha}}j$. Then there exists $f\in F$ such that $\alpha(f)=ij$ and $f^{\alpha}=f\cap(V(H_{i})\cup V(H_{j}))$. Since $f$ intersects both $H_{i}$ and $H_{j}$, it follows that $H^{\alpha}[V(H_{i})\cup V(H_{j})]$ is connected. Therefore $H^{\prime}$ is connected. Moreover, if there exist $i\in J$, $j\in I-J$, $u\in V(H_{i})$, and $v\in V(H_{j})$ such that $u\sim_{H^{\alpha}}v$, then for some $f\in F$ we have that $\alpha(f)=ij$, and hence $i\sim_{G^{\alpha}}j$, a contradiction. We conclude that $H^{\prime}$ is a connected component of $H^{\alpha}$. 2. (ii) Let $H^{\prime}$ be a connected component of $H^{\alpha}$. Since $H^{\alpha}=\left(\bigcup_{i\in I}H_{i}\right)+\\{f^{\alpha}:f\in F\\}$, there exists $J\subseteq I$ such that $V(H^{\prime})=\bigcup_{i\in J}V(H_{i})$. Take any $i,j\in J$. Then for all $u\in V(H_{i})$ and $v\in V(H_{j})$ we have that $u$ and $v$ are connected in $H^{\alpha}$. It is then easy to see that $i$ and $j$ are connected in $G^{\alpha}$. Hence $G^{\alpha}[J]$ is connected. It then follows from (i) that $G^{\alpha}[J]$ is a connected component of $G^{\alpha}$. a We are now ready to prove our first reduction theorem, Theorem 5.2, which states that a hypergraph $H$ with a minimal edge cut $F$ admits an Euler family if and only if certain subhypergraphs of $H$ obtained from the connected components of $H{\backslash}F$ admit Euler families. The analogous result for Euler tours follows in Theorem 5.3. ###### Theorem 5.2 Let $H$ be a connected hypergraph with a minimal edge cut $F$, and let $H_{i}$, for $i\in I$, be the connected components of $H{\backslash}F$. Then $H$ has a (spanning) Euler family if and only if there exists $J\subseteq I$ with $1\leq|J|\leq|F|$ such that 1. (i) $H\left[\bigcup_{j\in J}V(H_{j})\right]$ has a non-empty (spanning) Euler family, and 2. (ii) $H_{i}$ has a (spanning) Euler family for all $i\in I-J$. Proof. Assume $H$ has an Euler family ${\mathcal{F}}$, and let $\alpha:F\to I^{[2]}$ be the edge cut assignment defined by $\alpha(f)=ij$ if the edge $f\in F$ is traversed by ${\mathcal{F}}$ via a vertex in $H_{i}$ and a vertex in $H_{j}$. Then by Lemma 4.3(i), the associated hypergraph $H^{\alpha}$ has an Euler family as well. Let $G^{\alpha}$ be the associated multigraph, and $J$ the union of the vertex sets of all non-empty connected components of $G^{\alpha}$. Since $H$ is connected, we have $|F|\geq 1$, and hence $|J|\geq 1$. Since $H^{\alpha}$ has an Euler family, by Lemma 4.2, so does $G^{\alpha}$. Hence $G^{\alpha}[J]$ is an even graph with $|J|$ vertices, $|F|$ edges, and minimum degree at least 2; it follows that $|J|\leq|F|$. To show (i) and (ii), let $V^{\prime}=\bigcup_{j\in J}V(H_{j})$. By Lemma 5.1(i), we have that $H^{\alpha}[V^{\prime}]$ is a union of connected components of $H^{\alpha}$, and $H_{i}$, for each $i\in I-J$, is a connected component of $H^{\alpha}$. Since $H^{\alpha}$ has an Euler family, it follows from Lemma 2.1 that $H^{\alpha}[V^{\prime}]$ has an Euler family, as does $H_{i}$, for each $i\in I-J$. It remains to show that $H[V^{\prime}]$ has a Euler family. Define a bijection $\varphi:E\left(H^{\alpha}[V^{\prime}]\right)\to E\left(H[V^{\prime}]\right)$ by $\varphi(e^{\alpha})=e\cap V^{\prime}$. Then $e^{\alpha}\subseteq\varphi(e^{\alpha})$ for all $e^{\alpha}\in E\left(H^{\alpha}[V^{\prime}]\right)$, so by Lemma 2.2(i), since $H^{\alpha}[V^{\prime}]$ has an Euler family, so does $H[V^{\prime}]$. Moreover, since $H[V^{\prime}]$ is non-empty, this Euler family is non-empty. Conversely, assume that there exists $J\subseteq I$ with $1\leq|J|\leq|F|$ such that (i) and (ii) hold. Let $V^{\prime}=\bigcup_{j\in J}V(H_{j})$ and $H^{\prime}=H[V^{\prime}]\cup\bigcup_{i\in I-J}H_{i}$. Then by Lemma 2.1(i), the hypergraph $H^{\prime}$ admits an Euler family. Observe that $E(H^{\prime})=\bigcup_{i\in I}E(H_{i})\cup\\{f\cap V^{\prime}:f\in F\\}$. Define a bijection $\varphi:E(H^{\prime})\to E(H)$ by $\varphi(f\cap V^{\prime})=f$ for all $f\in F$, and $\varphi(e)=e$ for all $e\in\bigcup_{i\in I}E(H_{i})$. Hence $e\subseteq\varphi(e)$ for all $e\in E(H^{\prime})$, and since $H^{\prime}$ has an Euler family, by Lemma 2.2(i), the hyperhgraph $H$ has an Euler family as well. It is easy to see that if the Euler family ${\mathcal{F}}$ of $H$ is spanning, then so are the resulting Euler families of $H[V^{\prime}]$ and $H_{i}$, for all $i\in I-J$, and vice-versa. a ###### Theorem 5.3 Let $H$ be a connected hypergraph with a minimal edge cut $F$, and let $H_{i}$, for $i\in I$, be the connected components of $H{\backslash}F$. Then $H$ has an Euler tour if and only if there exists $J\subseteq I$ with $1\leq|J|\leq|F|$ such that 1. (i) $H\left[\bigcup_{j\in J}V(H_{j})\right]$ has an Euler tour, and 2. (ii) $H_{i}$ is empty for all $i\not\in J$. Proof. Assume $H$ has an Euler tour $T$, and let $\alpha:F\to I^{[2]}$ be the edge cut assignment defined by $\alpha(f)=ij$ if the edge $f\in F$ is traversed by $T$ via a vertex in $H_{i}$ and a vertex in $H_{j}$. Let $G^{\alpha}$ be the associated multigraph. By Lemmas 4.3(i) and 4.2, respectively, the hypergraph $H^{\alpha}$ and multigraph $G^{\alpha}$ both have Euler tours. Hence $G^{\alpha}$ has a unique non-empty connected component; let $J$ be its vertex set. As in the proof of Theorem 5.2, it can be shown that $1\leq|J|\leq|F|$. Let $V^{\prime}=\bigcup_{j\in J}V(H_{j})$. By Lemma 5.1(i), we have that $H^{\alpha}[V^{\prime}]$ is a connected component of $H^{\alpha}$, as is $H_{i}$, for each $i\in I-J$. Since $H^{\alpha}$ has an Euler tour and $H^{\alpha}[V^{\prime}]$ is not empty, it follows that $H^{\alpha}[V^{\prime}]$ has an Euler tour, and $H_{i}$, for each $i\in I-J$, is empty. Conversely, assume that there exists $J\subseteq I$ with $1\leq|J|\leq|F|$ such that (i) and (ii) hold. Let $V^{\prime}=\bigcup_{j\in J}V(H_{j})$, and $H^{\prime}=H[V^{\prime}]\cup\bigcup_{i\in I-J}H_{i}$. Similarly to the proof of Theorem 5.2, we observe that since $H[V^{\prime}]$ has an Euler tour and $H_{i}$, for each $i\in I-J$, is empty, Lemma 2.1(i) shows that $H^{\prime}$ has an Euler tour as well. By Lemma 2.2(i), it follows that $H$ has an Euler tour. a We observe that Theorem 5.3 cannot be extended to spanning Euler tours in any meaningful way. The following immediate consequence of Theorems 5.2 and 5.3 gives more transparent necessary and sufficient conditions for the existence of an Euler family and Euler tour for a hypergraph with a cut edge. ###### Corollary 5.4 Let $H$ be a connected hypergraph with a cut edge $f$. Let $H_{i}$, for $i\in I$, be the connected components of $H{\backslash}f$. Then the following hold. 1. (i) $H$ has a (spanning) Euler family if and only if there exists $j\in I$ such that * • $H[V(H_{j})]$ has a non-empty (spanning) Euler family, and * • $H_{i}$ has a (spanning) Euler family for all $i\in I-\\{j\\}$. 2. (ii) $H$ has an Euler tour if and only if there exists $j\in I$ such that * • $H[V(H_{j})]$ has an Euler tour, and * • $H_{i}$ is empty for all $i\in I-\\{j\\}$. 3. (iii) $H$ has no spanning Euler tour. 4. (iv) If $H$ has no vertex of degree 1, then $H$ has no Euler tour. In Algorithm 5.5 below, we shall now put to use the results of Lemma 4.2 and Theorem 4.4 (applied to standard edge cut assignments) to determine whether a given hypergraph has an Euler family. ###### Algorithm 5.5 Does a connected hypergraph $H$ admit an Euler family? 1. (1) Sequentially delete vertices of degree at most 1 from $H$ until no such vertices remain or $H$ is trivial. If at any step $H$ has an edge of cardinality less than 2, then $H$ has no Euler family — exit. 2. (2) If $H$ is trivial (and empty), then it has an Euler family — exit. 3. (3) Find a minimal edge cut $F$. 4. (4) Let $H_{i}$, for $i\in I$, be the connected components of $H{\backslash}F$. 5. (5) For all edge cut assignments $\alpha:F\to I^{[2]}$: 1. (a) If $G^{\alpha}$ has no Euler family, then $H^{\alpha}$ has no Euler family — discard this $\alpha$. 2. (b) For each connected component $H^{\prime}$ of $H^{\alpha}$: if $H^{\prime}$ has no Euler family (recursive call), discard this $\alpha$. 3. (c) If every connected component of $H^{\alpha}$ has an Euler family, then $H$ has an Euler family — exit. 6. (6) $H$ has no Euler family — exit. Similarly, in Algorithm 5.6 below, we use Theorem 5.3 and Corollary 5.4, in addition to Lemma 4.2 and Theorem 4.4, to determine whether a given hypergraph has an Euler tour. ###### Algorithm 5.6 Does a connected hypergraph $H$ admit an Euler tour? 1. (1) Sequentially delete vertices of degree at most 1 from $H$ until no such vertices remain or $H$ is trivial. If at any step $H$ has an edge of cardinality less than 2, then $H$ is not eulerian — exit. 2. (2) If $H$ is trivial (and empty), then it is eulerian — exit. 3. (3) Find a minimal edge cut $F$. If $|F|=1$, then $H$ is not eulerian — exit. 4. (4) Let $H_{i}$, for $i\in I$, be the connected components of $H{\backslash}F$. 5. (5) Let $J=\\{i\in I:H_{i}\mbox{ is non-empty}\\}$. If $|F|<|J|$, then $H$ is not eulerian — exit. 6. (6) For all edge cut assignments $\alpha:F\to I^{[2]}$: 1. (a) If $G^{\alpha}$ is not eulerian, then $H^{\alpha}$ is not eulerian — discard this $\alpha$. 2. (b) If $H^{\alpha}$ has at least 2 non-empty connected components, then $H^{\alpha}$ is not eulerian — discard this $\alpha$. 3. (c) Let $H^{\prime}$ be the unique non-empty connected component of $H^{\alpha}$. If $H^{\prime}$ is eulerian (recursive call), then $H$ is eulerian — exit. Otherwise, discard this $\alpha$. 7. (7) $H$ is not eulerian — exit. Figure 1: Isomorphism classes of even graphs of small size without isolated vertices. ###### Remarks 5.7 1. (i) In both algorithms, the input hypergraph is being modified during the execution of the algorithm. However, at any step, the current hypergraph $H$ admits an Euler family (tour) if and only if the input hypergraph does. 2. (ii) As each non-empty proper subset $S$ of $V(H)$ corresponds to an edge cut, minimal edge cuts are easy to construct. As we comment below, it is likely that using a minimum edge cut would be more efficient. An ${\cal O}(p+n^{2}\lambda)$ algorithm for finding a minimum edge cut in a hypergraph $H$ with size $p=\sum_{e\in E(H)}|e|$, order $n=|V(H)|$, and cardinality of a minimum edge cut $\lambda$ was described in [4]. 3. (iii) Recall that a non-empty graph has an Euler family if and only if it is even, and is eulerian if and only if it is even and has at most one non-empty connected component. Hence (5a) in Algorithm 5.5 and (6a) in Algorithm 5.6 are easy to verify. 4. (iv) The most time-consuming part of both algorithms is likely going to be the sequence of recursive calls to determine whether hypergraphs $H^{\prime}$ have an Euler family or tour. The smaller the size of the largest of the hypergraphs $H^{\prime}$, the better; however, this size is impossible to predict. Instead, we suggest minimizing the number of possible edge cut assignments $\alpha$ to be verified. The number of all possible mappings $F\to I^{[2]}$ is ${|I|+1\choose 2}^{|F|}$, which indeed suggests choosing $F$ to be a minimum edge cut. Among minimum edge cuts $F$, should we seek one that also minimizes $|I|$? We do not have a clear answer: on one hand, this approach would minimize the number of mappings $\alpha$ to verify; one the other hand, the expected size of the hypergraphs $H^{\prime}$ for our recursive calls is inversely proportional to $|I|$. — We shall comment more on the complexity of Algorithms 5.5 and 5.6 in Remark 6.8(iii). 5. (v) The number of edge cut assignments $\alpha:F\to I^{[2]}$ resulting in an even graph $G^{\alpha}$ is likely much smaller than ${|I|+1\choose 2}^{|F|}$. To give some idea of this number, we show in Figure 1, for $|F|\leq 4$, the isomorphism classes of the suitable graphs $G^{\alpha}$ with the isolated vertices removed. 6. (vi) Observe that Algorithms 5.5 and 5.6 are formulated to solve decision problems — does $H$ contain an Euler family (tour)? — however, since they are based on results that use constructive proofs, they can easily be adapted to construct an Euler family (tour), if it exists. ## 6 Reduction Using Collapsed Hypergraphs We shall now introduce our second main tool — the collapsed hypergraph — which allows for a binary-type of a reduction. ###### Definition 6.1 Let $H$ be a hypergraph, and $S$ a subset of its vertex set with $\emptyset\subsetneq S\subsetneq V(H)$. The collapsed hypergraph of $H$ with respect to $S$ is the hypergraph $H\circ S=(V^{\circ},E^{\circ})$ defined by $V^{\circ}=(V-S)\cup\\{u^{\circ}\\}\quad\mbox{ and }\quad E^{\circ}=\left\\{\\!\\!\left\\{e^{\circ}:e\in E(H),|e\cap(V-S)|\geq 1\right\\}\\!\\!\right\\},$ where $e^{\circ}=\left\\{\begin{array}[]{ll}e&\mbox{ if }e\cap S=\emptyset\\\ (e-S)\cup\\{u^{\circ}\\}&\mbox{otherwise}\end{array}\right..$ Here, $u^{\circ}$ is a new vertex, which we call the collapsed vertex of $H\circ S$. We are now ready for our second main reduction theorem, Theorem 6.2, which states that a hypergraph $H$ admits an Euler family if and only if some collapsed hypergraphs of $H$ (obtained using an edge cut assignment) admit Euler families. ###### Theorem 6.2 Let $H$ be a connected hypergraph with a minimal edge cut $F$, and $\\{V_{0},V_{1}\\}$ a partition of $V(H)$ into unions of the vertex sets of the connected components of $H{\backslash}F$. Then $H$ admits an Euler family if and only if there exists an edge cut assignment $\alpha:F\to\mathbb{Z}_{2}^{[2]}$ such that $|\alpha^{-1}(01)|$ is even, and either 1. (i) $\alpha^{-1}(01)=\emptyset$ and $H^{\alpha}[V_{i}]$ has an Euler family for each $i\in\mathbb{Z}_{2}$, or 2. (ii) $\alpha^{-1}(01)\neq\emptyset$, and for each $i\in\mathbb{Z}_{2}$, the collapsed hypergraph $H^{\alpha}\circ V_{i}$ has an Euler family that traverses the collapsed vertex $u_{i}^{\circ}$ via each of the edges in $\\{f^{\circ}:f\in\alpha^{-1}(01)\\}$. Proof. Let ${\mathcal{F}}$ be an Euler family of $H$, and $\alpha:F\to\mathbb{Z}_{2}^{[2]}$ the edge cut assignment with $\alpha(f)=ij$ if and only if the edge $f$ is traversed by ${\mathcal{F}}$ via a vertex in $V_{i}$ and a vertex in $V_{j}$ (where $i=j$ is possible). By Lemma 4.3(i), the hypergraph $H^{\alpha}$ has an Euler family as well, say ${\mathcal{F}}^{\alpha}$. Since the only edges of $H^{\alpha}$ that intersect both $V_{0}$ and $V_{1}$ are edges of the form $f^{\alpha}$ with $f\in F$ and $\alpha(f)=01$, it is easy to see that each closed trail in ${\mathcal{F}}^{\alpha}$ traverses an even number of such edges. Hence $|\alpha^{-1}(01)|$ is indeed even. Suppose first that $\alpha^{-1}(01)=\emptyset$. Then $H^{\alpha}[V_{i}]$, for each $i\in\mathbb{Z}_{2}$, is a union of connected components of $H^{\alpha}$, and hence has an Euler family by Lemma 2.1(ii). Suppose now that $\alpha^{-1}(01)\neq\emptyset$. By symmetry, it suffices to show that $H^{\alpha}\circ V_{1}$ has an Euler family that traverses the collapsed vertex $u_{1}^{\circ}$ via each of the edges in $\\{f^{\circ}:f\in\alpha^{-1}(01)\\}$. Let $T$ be a closed trail in the Euler family ${\mathcal{F}}^{\alpha}$ of $H^{\alpha}$ that traverses an edge $f^{\alpha}$, for some $f\in\alpha^{-1}(01)$. Then $T$ has the form $T=v_{0}^{(0)}f_{0}^{\alpha}u_{0}^{(1)}T_{0}^{(1)}v_{0}^{(1)}f_{0}^{\prime\alpha}u_{0}^{(0)}T_{0}^{(0)}v_{1}^{(0)}\ldots u_{k-1}^{(0)}T_{k-1}^{(0)}v_{0}^{(0)}$ for some vertices $v_{0}^{(0)},u_{0}^{(0)},v_{1}^{(0)},u_{1}^{(0)},\ldots,u_{k-1}^{(0)}\in V_{0}$ and $u_{0}^{(1)},v_{0}^{(1)},u_{1}^{(1)},v_{1}^{(1)},\ldots,v_{k-1}^{(1)}\in V_{1}$; for some edges $f_{0},f_{0}^{\prime},f_{1},f_{1}^{\prime},\ldots,f_{k-1}^{\prime}\in\alpha^{-1}(01)$; and for $i\in\mathbb{Z}_{k}$, for some $(u_{i}^{(0)},v_{i+1}^{(0)})$-trails $T_{i}^{(0)}$ in $H^{\alpha}[V_{0}]$ and $(u_{i}^{(1)},v_{i}^{(1)})$-trails $T_{i}^{(1)}$ in $H^{\alpha}[V_{1}]$. Let $T^{\circ}$ be the sequence obtained from $T$ by replacing each subsequence of the form $v_{i}^{(0)}f_{i}^{\alpha}u_{i}^{(1)}T_{i}^{(1)}v_{i}^{(1)}f_{i}^{\prime\alpha}u_{i}^{(0)}$ with $v_{i}^{(0)}f_{i}^{\circ}u_{1}^{\circ}f_{i}^{\prime\circ}u_{i}^{(0)}$. Then $T^{\circ}$ is a closed trail in $H^{\alpha}\circ V_{1}$, and it traverses the collapsed vertex $u_{1}^{\circ}$ via each of the edges of the form $f^{\circ}$ for $f\in\alpha^{-1}(01)$ such that $T$ traverses $f^{\alpha}$. If we additionally define $T^{\circ}=T$ for all closed trails $T\in{\mathcal{F}}^{\alpha}$ that do not traverse any edges of the form $f^{\alpha}$, for some $f\in\alpha^{-1}(01)$, then it is clear that $\\{T^{\circ}:T\in{\mathcal{F}}^{\alpha}\\}$ is a family of closed trails in $H^{\alpha}\circ V_{1}$ that jointly traverse each edge exactly once, and also traverse the collapsed vertex $u_{1}^{\circ}$ via each of the edges in $\\{f^{\circ}:f\in\alpha^{-1}(01)\\}$. To obtain an Euler family, we just need to concatenate all those closed trails in this family that traverse the collapsed vertex $u_{1}^{\circ}$. To prove the converse, let $\alpha:F\to\mathbb{Z}_{2}^{[2]}$ be an edge cut assignment such that $|\alpha^{-1}(01)|$ is even, and either (i) or (ii) holds. Suppose first that (i) holds. Since $\alpha^{-1}(01)=\emptyset$, for each $i\in\mathbb{Z}_{2}$, the hypergraph $H^{\alpha}[V_{i}]$ is a union of connected components of $H^{\alpha}$, and by Lemma 2.1, since each of $H^{\alpha}[V_{0}]$ and $H^{\alpha}[V_{1}]$ admits an Euler family, so does $H^{\alpha}$. By Lemma 4.3(ii), it follows that $H$ admits an Euler family. Suppose now that (ii) holds and $|\alpha^{-1}(01)|=2k$. For each $i\in\mathbb{Z}_{2}$, let $\alpha^{-1}(01)=\\{f_{0}^{(i)},f_{1}^{(i)},\ldots,f_{2k-1}^{(i)}\\}$. By assumption, the collapsed hypergraph $H^{\alpha}\circ V_{1-i}$ admits an Euler family ${\mathcal{F}}^{(i)}$ that traverses the collapsed vertex $u_{1-i}^{\circ}$ via each of the edges $f^{\circ}$, for $f\in\alpha^{-1}(01)$. Let $T^{(i)}$ be the unique closed trail in ${\mathcal{F}}^{(i)}$ that traverses $u_{1-i}^{\circ}$. Then, without loss of generality, $T^{(i)}$ must be of the form $T^{(i)}=v_{0}^{(i)}\,\,(f_{0}^{(i)})^{\circ}\,\,u_{1-i}^{\circ}\,\,(f_{1}^{(i)})^{\circ}\,\,v_{1}^{(i)}\,\,T_{1}^{(i)}\,\,v_{2}^{(i)}\,\,(f_{2}^{(i)})^{\circ}\ldots\,\,v_{2k-1}^{(i)}\,\,T_{k}^{(i)}\,\,v_{0}^{(i)}$ for some vertices $v_{0}^{(i)},v_{1}^{(i)},\ldots,v_{2k-1}^{(i)}\in V_{i}$ and, for $j=1,2,\ldots,k$, some $(v_{2j-1}^{(i)},v_{2j}^{(i)})$-trails $T_{j}^{(i)}$ in $H^{\alpha}[V_{i}]$. (Here, subscripts are evaluated modulo $2k$.) Let $\pi:\mathbb{Z}_{2k}\to\mathbb{Z}_{2k}$ be a bijection such that $f_{\pi(j)}^{(1)}=f_{j}^{(0)}$, for all $j\in\mathbb{Z}_{2k}$. We thus have that $v_{\pi(j)}^{(1)}\in f_{j}^{(0)}$, for all $j\in\mathbb{Z}_{2k}$. We now link the $(v_{2j-1}^{(0)},v_{2j}^{(0)})$-trails $T_{j}^{(0)}$ and $(v_{2j-1}^{(1)},v_{2j}^{(1)})$-trails $T_{j}^{(1)}$, for $j=1,2,\ldots,k$, into a family ${\cal T}$ of closed trails in $H^{\alpha}$ using the edges $(f_{j}^{(0)})^{\alpha}$, for $j\in\mathbb{Z}_{2k}$. In particular, the new closed trails will traverse each edge $(f_{j}^{(0)})^{\alpha}$ via anchors $v_{j}^{(0)}$ and $v_{\pi(j)}^{(1)}$. Finally, let ${\mathcal{F}}=\left({\mathcal{F}}^{(0)}-\\{T^{(0)}\\}\right)\cup\left({\mathcal{F}}^{(1)}-\\{T^{(1)}\\}\right)\cup{\cal T}.$ Since for each $i\in\mathbb{Z}_{2}$, the closed trails in ${\mathcal{F}}^{(i)}-\\{T^{(i)}\\}$ traverse all edges of $H^{\alpha}[V_{i}]$ that are not traversed by $T^{(i)}$, and the closed trails in ${\cal T}$ traverse all edges of $H^{\alpha}[V_{0}]\cup H^{\alpha}[V_{1}]$ that are traversed by $T^{(0)}$ or $T^{(1)}$, as well as all edges in $\\{f^{\alpha}:f\in\alpha^{-1}(01)\\}$, we conclude that ${\mathcal{F}}$ is an Euler family of $H^{\alpha}$. It then follows by Lemma 4.3(ii) that $H$ admits an Euler family. a Observe that in the proof of sufficiency in Theorem 6.2 we have no control over the number of closed trails in the family ${\cal T}$; hence the analogous result for Euler tours does not hold in general. A weaker result is proved below: necessity is guaranteed by Corollary 6.3, while sufficiency is proved in Corollary 6.4 with an additional assumption on the edge cut assignment. This additional assumption, however, always holds for edge cuts of cardinality at most 3. ###### Corollary 6.3 Let $H$ be a connected hypergraph with a minimal edge cut $F$, and $\\{V_{0},V_{1}\\}$ a partition of $V(H)$ into unions of the vertex sets of the connected components of $H{\backslash}F$. If $H$ admits an Euler tour, then there exists an edge cut assignment $\alpha:F\to\mathbb{Z}_{2}^{[2]}$ such that $|\alpha^{-1}(01)|$ is even, and either 1. (i) $\alpha^{-1}(01)=\emptyset$ and, without loss of generality, $H^{\alpha}[V_{0}]$ has an Euler tour and $H^{\alpha}[V_{1}]$ is empty, or 2. (ii) $\alpha^{-1}(01)\neq\emptyset$, and for each $i\in\mathbb{Z}_{2}$, the collapsed hypergraph $H^{\alpha}\circ V_{i}$ has an Euler tour that traverses the collapsed vertex $u_{i}^{\circ}$ via each of the edges in $\\{f^{\circ}:f\in\alpha^{-1}(01)\\}$. Proof. Let $T$ be an Euler tour of $H$, and $\alpha:F\to\mathbb{Z}_{2}^{[2]}$ the edge cut assignment with $\alpha(f)=ij$ if and only if the edge $f$ is traversed by $T$ via a vertex in $V_{i}$ and a vertex in $V_{j}$ (where $i=j$ is possible). We establish that $|\alpha^{-1}(01)|$ is even just as in the proof of necessity in Theorem 6.2. By Lemma 4.3(i), the hypergraph $H^{\alpha}$ has an Euler tour as well. If $\alpha^{-1}(01)=\emptyset$, then $H^{\alpha}$ is disconnected. Hence without loss of generality, $H^{\alpha}[V_{0}]$ has an Euler tour and $H^{\alpha}[V_{1}]$ is empty. Suppose now that $\alpha^{-1}(01)\neq\emptyset$. Following the proof of necessity in Theorem 6.2, we can show that for each $i\in\mathbb{Z}_{2}$, the Euler tour $T$ of $H$ gives rise to an Euler tour $T_{i}^{\circ}$ of $H^{\alpha}\circ V_{i}$ that traverses the collapsed vertex $u_{i}^{\circ}$ via each of the edges in $\\{f^{\circ}:f\in\alpha^{-1}(01)\\}$. a ###### Corollary 6.4 Let $H$ be a connected hypergraph with a minimal edge cut $F$, and $\\{V_{0},V_{1}\\}$ a partition of $V(H)$ into unions of the vertex sets of the connected components of $H{\backslash}F$. Assume that there exists an edge cut assignment $\alpha:F\to\mathbb{Z}_{2}^{[2]}$ such that $|\alpha^{-1}(01)|\in\\{0,2\\}$, and either 1. (i) $\alpha^{-1}(01)=\emptyset$ and, without loss of generality, $H^{\alpha}[V_{0}]$ has an Euler tour and $H^{\alpha}[V_{1}]$ is empty, or 2. (ii) $\alpha^{-1}(01)\neq\emptyset$, and for each $i\in\mathbb{Z}_{2}$, the collapsed hypergraph $H^{\alpha}\circ V_{i}$ has an Euler tour that traverses the collapsed vertex $u_{i}^{\circ}$. Then $H$ admits an Euler tour. Proof. If $\alpha^{-1}(01)=\emptyset$, then $H^{\alpha}=H^{\alpha}[V_{0}]\cup H^{\alpha}[V_{1}]$, and since $H^{\alpha}[V_{1}]$ is empty and $H^{\alpha}[V_{0}]$ admits an Euler tour, so does $H^{\alpha}$. By Lemma 2.2(ii), it follows that $H$ admits an Euler tour. The case $\alpha^{-1}(01)\neq\emptyset$ is similar to the proof of sufficiency in Theorem 6.2, assuming Condition (ii). We have $\alpha^{-1}(01)=\\{f_{0}^{(0)},f_{1}^{(0)}\\}=\\{f_{0}^{(1)},f_{1}^{(1)}\\}$. By assumption, for each $i\in\mathbb{Z}_{2}$, the collapsed hypergraph $H^{\alpha}\circ V_{1-i}$ admits an Euler tour $T^{(i)}$ that traverses the collapsed vertex $u_{1-i}^{\circ}$, necessarily via the edges $f^{\circ}$, for $f\in\alpha^{-1}(01)$. Then $T^{(i)}$ must be of the form $T^{(i)}=v_{0}^{(i)}\,\,(f_{0}^{(i)})^{\circ}\,\,u_{1-i}^{\circ}\,\,(f_{1}^{(i)})^{\circ}\,\,v_{1}^{(i)}\,\,T_{1}^{(i)}\,\,v_{0}^{(i)}$ for some vertices $v_{0}^{(i)},v_{1}^{(i)}\in V_{i}$ and a $(v_{1}^{(i)},v_{0}^{(i)})$-trail $T_{1}^{(i)}$ in $H^{\alpha}[V_{i}]$. Since either $f_{j}^{(0)}=f_{j}^{(1)}$ or $f_{j}^{(0)}=f_{1-j}^{(1)}$ for $j\in\mathbb{Z}_{2}$, linking the $(v_{1}^{(0)},v_{0}^{(0)})$-trail $T_{1}^{(0)}$ and $(v_{1}^{(1)},v_{0}^{(1)})$-trail $T_{1}^{(1)}$ using the edges $(f_{0}^{(0)})^{\alpha}$ and $(f_{1}^{(0)})^{\alpha}$ clearly results in a closed trail $T$ in $H^{\alpha}$ that traverses all edges of $H^{\alpha}$ traversed by $T^{(0)}$ and $T^{(1)}$, as well as the edges $(f_{0}^{(0)})^{\alpha}$ and $(f_{1}^{(0)})^{\alpha}$. We conclude that $T$ is an Euler tour of $H^{\alpha}$. By Lemma 4.3(ii), we have that $H$ admits an Euler tour. a Observe that in the case $|F|\leq 3$, Corollary 6.4 gives a full converse to Corollary 6.3. In Theorem 6.2 and Corollaries 6.3 and 6.4, we seem to be translating the problem of finding an Euler family (tour) to a different problem, namely, of finding an Euler family (tour) that traverses a specified vertex via each of the specified edges. In the next lemma, we show that an algorithm to find the former can be used to find the latter as well. ###### Lemma 6.5 Let $H$ be a hypergraph, $u\in V(H)$, and $F\subseteq E(H)$ such that each edge in $F$ is incident with the vertex $u$. Construct a hypergraph $H^{\prime}$ from $H$ as follows: * • for each edge $f\in F$, adjoin a new vertex $u_{f}$, and * • replace each edge $f\in F$ with edges $f^{\prime}=(f-\\{u\\})\cup\\{u_{f}\\}$ and $e_{f}=\\{u_{f},u\\}$. Then $H$ has an Euler family (tour) traversing vertex $u$ via each edge in $F$ if and only if $H^{\prime}$ has an Euler family (tour). Proof. Assume ${\mathcal{F}}$ is an Euler family of $H$ traversing vertex $u$ via each edge in $F$. For each closed trail $T$ in ${\mathcal{F}}$, obtain a sequence $T^{\prime}$ by replacing each subsequence of the form $fu$ in $T$ with the subsequence $f^{\prime}u_{f}e_{f}u$, and each subsequence of the form $uf$ with the subsequence $ue_{f}u_{f}f^{\prime}$. It is clear that ${\mathcal{F}}^{\prime}=\\{T^{\prime}:T\in{\mathcal{F}}\\}$ is an Euler family of $H^{\prime}$. Conversely, if ${\mathcal{F}}^{\prime}$ is an Euler family of $H^{\prime}$, then it must traverse each edge of the form $e_{f}$, for some $f\in F$, via either the subtrail $f^{\prime}u_{f}e_{f}u$ or the subtrail $ue_{f}u_{f}f^{\prime}$. For each closed trail $T^{\prime}\in{\mathcal{F}}^{\prime}$, obtain a sequence $T$ by replacing each subsequence of the form $f^{\prime}u_{f}e_{f}u$ with $fu$, and each subsequence of the form $ue_{f}u_{f}f^{\prime}$ with $uf$. It is then easy to see that ${\mathcal{F}}=\\{T:T^{\prime}\in{\mathcal{F}}^{\prime}\\}$ is an Euler family of $H$ that traverses vertex $u$ via each edge in $F$. Since in both parts of the proof we have $|{\mathcal{F}}|=|{\mathcal{F}}^{\prime}|$, the statement for Euler tours holds as well. a We shall now describe an algorithm based on Theorem 6.2 and Lemma 6.5 that determines whether a hypergraph admits an Euler family. ###### Algorithm 6.6 Does a connected hypergraph $H$ admit an Euler family? 1. (1) Sequentially delete vertices of degree at most 1 from $H$ until no such vertices remain or $H$ is trivial. If at any step $H$ has an edge of cardinality less than 2, then $H$ has no Euler family — exit. 2. (2) If $H$ is trivial (and empty), then it has an Euler family — exit. 3. (3) Find a minimal edge cut $F$. 4. (4) Find a bipartition $\\{V_{0},V_{1}\\}$ of $V(H)$ into unions of the vertex sets of the connected components of $H{\backslash}F$. 5. (5) For all edge cut assignments $\alpha:F\to\mathbb{Z}_{2}^{[2]}$: 1. (a) If $|\alpha^{-1}(01)|$ is odd — discard this $\alpha$. 2. (b) If $\alpha^{-1}(01)=\emptyset$: if $H^{\alpha}[V_{0}]$ and $H^{\alpha}[V_{1}]$ both have Euler families (recursive call), then $H$ has an Euler family — exit. 3. (c) If $\alpha^{-1}(01)\neq\emptyset$: if $H^{\alpha}\circ V_{1}$ and $H^{\alpha}\circ V_{0}$ both have Euler families (recursive call) that traverse the collapsed vertex via each of the edges of the form $f^{\circ}$, for $f\in\alpha^{-1}(01)$, then $H$ has an Euler family — exit. 6. (6) $H$ has no Euler family — exit. The analogous algorithm for Euler tours, Algorithm 6.7 below, relies on Corollary 6.4 and Lemma 6.5. If the hypergraph has no edge cut satisfying the assumptions of Corollary 6.4, then Algorithm 5.6 is invoked. ###### Algorithm 6.7 Does a connected hypergraph $H$ admit an Euler tour? 1. (1) Sequentially delete vertices of degree at most 1 from $H$ until no such vertices remain or $H$ is trivial. If at any step $H$ has an edge of cardinality less than 2, then $H$ is not eulerian — exit. 2. (2) If $H$ is trivial (and empty), then it is eulerian — exit. 3. (3) Find a minimum edge cut $F$. If $|F|=1$, then $H$ is not eulerian — exit. 4. (4) For all bipartitions $\\{V_{0},V_{1}\\}$ of $V(H)$ into unions of the vertex sets of the connected components of $H{\backslash}F$ and for all edge cut assignments $\alpha:F\to\mathbb{Z}_{2}^{[2]}$: 1. (a) If $|\alpha^{-1}(01)|$ is odd — discard this $\alpha$. 2. (b) If $|\alpha^{-1}(01)|=0$: if $H^{\alpha}[V_{0}]$ is empty and $H^{\alpha}[V_{1}]$ has an Euler tour, or vice-versa (recursive call), then $H$ has an Euler tour — exit. 3. (c) If $|\alpha^{-1}(01)|=2$: if $H^{\alpha}\circ V_{1}$ and $H^{\alpha}\circ V_{0}$ both have an Euler tour that traverses the collapsed vertex (recursive call), then $H$ has an Euler tour — exit. 5. (5) Use Algorithm 5.6. ###### Remarks 6.8 1. (i) Remarks 5.7(i) and 5.7(vi) apply to Algorithms 6.6 and 6.7 as well. 2. (ii) In Step (3) of Algorithm 6.7, we chose to find a minimum edge cut — see Remark 5.7(ii) — in the hope of finding an edge cut of cardinality at most 3; in this case, a reduction using Corollary 6.4 is guaranteed, and Step (5) is not reached in this call. 3. (iii) Observe that for a given edge cut $F$ such that $H{\backslash}F$ has exactly $|I|$ connected components, the number of bipartitions $\\{V_{0},V_{1}\\}$ is $2^{|I|-1}-1$, and the number of edge cut assignments $\alpha:F\to\mathbb{Z}_{2}^{[2]}$ is at most $3^{|F|}$. Hence the loop at Step (4) of Algorithm 6.7 will be executed at most $(2^{|I|-1}-1)\cdot 3^{|F|}$ times. 4. (iv) For suitable bipartitions $\\{V_{0}V_{1}\\}$, we expect that the reduction at Step (4) in both algorithms is the most efficient (that is, the number of recursive calls is minimized) if $|V_{0}|$ and $|V_{1}|$ are as equal as possible. 5. (v) We shall now compare the time complexity functions $\tau(p)$ and $\sigma(p)$ of Algorithms 5.6 and 6.7, respectively, where $p=\sum_{e\in E(H)}|E|$ is the size of the input hypergraph $H$. To facilitate the comparison, we shall assume that at any step of the algorithm (including recursive calls) we have $1\leq|F|\leq k$ and $2\leq|I|\leq m$ for some constants $k$ and $m$, and that at any recursive call, the size of the hypergraph $H^{\prime}$ is at most $\frac{p}{c}$ for some constant $c>1$. In addition, we assume that Step (5) of Algorithm 6.7 is never reached. With these assumptions, we have $\tau(p)\approx{m+1\choose 2}^{k}\cdot\tau\left(\frac{p}{c}\right)\approx m^{2k}\cdot\tau\left(\frac{p}{c}\right)\approx m^{2k\log_{c}p}=p^{\log_{c}(m^{2k})}$ and $\sigma(p)\approx 2^{m}3^{k}\cdot\sigma\left(\frac{p}{c}\right)\approx(2^{m}3^{k})^{\log_{c}p}=p^{\log_{c}(2^{m}3^{k})}.$ Thus, $\tau(p)$ and $\sigma(p)$ are both polynomial in $p$; however, which of the polynomial orders is smaller depends on the relationship between $k$ and $m$. Roughly speaking, $\log_{c}(2^{m}3^{k})=m\log_{c}2+k\log_{c}3$ is smaller when $k$ is larger than $m$, while $\log_{c}(m^{2k})=2k\log_{c}m$ is smaller when $m$ is much larger than $k$. 6. (vi) The time complexities of Algorithms 5.5 and 6.6 can be compared in a similar way, with very similar results. Acknowledgement The first author gratefully acknowledges support by the Natural Sciences and Engineering Research Council of Canada (NSERC), Discovery Grant RGPIN-2016-04798. ## References * [1] M. A. Bahmanian and M. Šajna, Quasi-eulerian hypergraphs, Electron. J. Combin. 24 (2017), #P3.30, 12 pp. * [2] M. A. Bahmanian and M. Šajna, Connection and separation in hypergraphs, Theory Appl. Graphs 2 (2015), no. 2, Art. 5, 24 pp. * [3] J. A. Bondy, U. S. R. Murty, Graph theory. Graduate Texts in Mathematics 244, Springer, New York, 2008. * [4] C. Chekuri, C. Xu, Computing minimum cuts in hypergraphs, Proceedings of the Twenty-Eighth Annual ACM-SIAM Symposium on Discrete Algorithms (2017), 1085–1100. * [5] Z. Lonc, P. Naroski, On tours that contain all edges of a hypergraph, Electron. J. Combin. 17 (2010), # R144, 31 pp. * [6] Y. D. Steimle, M. Šajna, Spanning Euler tours and spanning Euler families in hypergraphs with particular vertex cuts, Discrete Math. 341 (2018), 2808–2819.
# A proposal to improve Ni-based superconductors Zi-Jian Lang (gbsn郎子健) Tsung-Dao Lee Institute & School of Physics and Astronomy, Shanghai Jiao Tong University, Shanghai 200240, China Ruoshi Jiang (gbsn姜若诗) Tsung-Dao Lee Institute & School of Physics and Astronomy, Shanghai Jiao Tong University, Shanghai 200240, China Wei Ku (bsmi顧威) corresponding email<EMAIL_ADDRESS>Tsung-Dao Lee Institute & School of Physics and Astronomy, Shanghai Jiao Tong University, Shanghai 200240, China Key Laboratory of Artificial Structures and Quantum Control (Ministry of Education), Shanghai 200240, China ###### Abstract Recently discovered superconductivity in hole-doped nickelate Nd0.8Sr0.2NiO2 caught intensive attention in the field. An immediate question is how to improve its superconducting properties. Guided by the key characteristics of electronic structures of the cuprates and the nickelates, we propose that nickel chalcogenides with a similar lattice structure should be a promising family of materials. Using NdNiS2 as an example, we find this particular crystal structure a stable one, through first-principle structural optimization and phonon calculation. We justify our proposal by comparing with CaCuO2 and NdNiO2 the strength of the charge-transfer characteristics and the trend in their low-energy many-body effective Hamiltonians of doped hole carriers. These analysis indicates that nickel chalcogenides host low-energy physics closer to that of the cuprates, with stronger magnetic interaction than the nickelates, and thus deserve further experimental exploration. Our proposal also opens up the possibility of a wide range of parameter tuning through ligand substitution among chalcogenides, to further improve superconducting properties. Hole doped nickelate, Nd1-xSrxNiO2 Li _et al._ (2019, 2020a), as a first Ni- based high-temperature superconductor, has recently attracted great attention in condensed matter physics recently. It displays type-II superconductivity, dome-shaped superconducting phase Li _et al._ (2020a) and strange metal (linear resistivity) behavior in its normal state Li _et al._ (2019). All of these characteristics suggest that this material represents a new family of unconventional superconductors. Meanwhile, their strong temperature and doping dependent hall coefficient Li _et al._ (2020a), negative magneto-resistance Li _et al._ (2020b), absence of long-range magnetic order Hayward and Rosseinsky (2003); Sawatzky (2019) in the parent compound, and increasing normal-state resistivity in the overdoped regime Li _et al._ (2020a); Osada _et al._ (2020) also indicate rich underlying physics in this new superconductor that might be absent in the cuprates Ando _et al._ (2004a, b); Bozovic _et al._ (2016); Dagotto (1994). Obviously, these nickelate superconductors are of a promising family to help unravel the long-standing puzzles of high-temperature superconductivity and even to find higher transition temperature $T_{c}$ beyond the cuprates. So far within limited attempts, the highest $T_{c}$ of nickelates is only about 12K Li _et al._ (2020a); Osada _et al._ (2020), one order of magnitude lower than the best cuprates Schilling _et al._ (1993); Dagotto (1994). Furthermore, at present, superconductivity is only found in thin films but not in bulk samples Li _et al._ (2020b), for reasons yet to be understood. Significant experimental progress is thus expected upon improvement of sample quality. On the other hand, it is of equal importance to seek other approaches to improve the superconducting properties besides the sample quality. Here we address this timely issue by first comparing the high-energy electronic structure of the cuprates and nickelates to identify their key characteristics being the strength of charge transfer. Based on this, we propose a new family of material, nickel chalcogenides, as a promising candidate to improve the superconducting properties. Taking NdNiS2 as an example, through density functional structure optimization and phonon calculation, we first demonstrate that this compound is stable under the same crystal structure as NdNiO2. The corresponding high-energy electronic structure confirms our expectation of an enhanced charge-transfer characteristic. Furthermore, our local many-body diagonalization gives a ground state similar to those of the cuprates and nickelates, namely a spin- singlet state with doped holes mostly residing in ligand-$p$-orbitals Lang _et al._ (2020). As anticipated, the corresponding effective eV-scale Hamiltonian of hole carriers contains stronger spin interactions than NdNiO2, suggesting a higher temperature scale in the phase diagrams. Our study indicates that nickel chalcogenides are promising candidates for improved superconducting properties, and ligand substitution, e.g. NdNiS2-xOx and NdNiS2-xSex, would introduce a great deal of tunability for future experimental exploration. Figure 1: Comparison of LDA+$U$ band structures of CaCuO2, NdNiS2 and NdNiO2 under AFM order, unfolded in the one-Cu/Ni Brillouin zone. The red, blue and green colors represent the weights of Cu/Ni, Ca/Nd and O/S orbitals, and Nd $f$-orbitals are set transparent. The lower panel shows the magnified band structure of the purple dashed boxes in the upper panel. Notice the trend in the relative energies of O/S and Cu/Ni orbitals. To identify the key difference between the cuprates and nickelates, we compare their high-energy electronic structure using density functional theory(DFT). Since both the cuprates and nickelates host strong antiferromagnetic (AFM) correlation Lee _et al._ (2006); Cui _et al._ (2020) inherited from the unfrustrated square lattice of spin 1/2 local moment Anderson (1950), we calculate the band structures under AFM order within the LDA+$U$ approximation Anisimov _et al._ (1993); Liechtenstein _et al._ (1995); sup and unfold them to the one-Ni unit cell for a simpler visualization Ku _et al._ (2010). Fig. 1(a)(c) show that compared with NdNiO2, the main difference of CaCuO2 at the large energy scale is the much lower energy of its $d$-orbitals (in red) relative to the O $p$-orbitals (in green), reflecting a much stronger charge- transfer nature well known to the community Zaanen _et al._ (1985). Given that both families are doped spin 1/2 systems, it is reasonable to expect that promoting such a charge transfer characteristic should improve significantly the superconducting properties, due to various considerations of low-energy physics such as enhanced super-exchange interaction Anderson (1950) and renormalized kinetic energy. Since there is no chemical way to further lower the orbital energy of Ni (other than replacing it by Cu), we are left with no choice but to raise the energy of the ligand $p$-orbitals, for example by substituting O with S or Se. Taking NdNiS2 as an example, we first examine the stability of this compound under the same crystal structure [c.f. Fig. 2(a)] as the nickelates. Our structure optimization calculation sup gives lattice constants $a=b=4.505$Å and $c=3.703$Å. With these structural parameters, further phonon calculation finds that phonon frequencies are all positive, as shown in Fig. 2(b). This confirms a stable structure realizable in the lab. Figure 2: (a) Crystal structure of NdNiS2, where grey, orange and yellow balls represent the Ni, Nd and S atoms respectively. (b) Phonon dispersion and corresponding density of states of NdNiS2. The positivity of all phonon frequencies confirms the stability of the crystal structure. Next, we verify the enhanced charge-transfer characteristic of this material. Fig. 1(b) shows the similar unfolded band structure of AFM NdNiS2. As expected from above chemical intuition, substituting O by S raises the energy of the $p$-orbitals (in green) quite significantly, thereby enhancing the charge- transfer nature. The density of state (DOS) plots in Fig. 3 illustrate a similar trend. Right below the Fermi energy, the relative weight of the most relevant ligand $p_{\mathbin{\\!/\mkern-5.0mu/\\!}}$-orbitals (in green) to the $d_{x^{2}-y^{2}}$-orbital (in red) grows systematically from NdNiO2 to NdNiS2 and CaCuO2. (Here, $p_{\mathbin{\\!/\mkern-5.0mu/\\!}}$ refers to O/S $p$-orbitals pointing toward nearest Cu/Ni atoms.) Indeed, substituting O by S enhances the charge-transfer nature and brings nickel chalcogenides closer to the cuprates. To reveal the physical benefits of a stronger charge-transfer characteristic, we proceed to investigate the low-energy effective Hamiltonian using well- established approaches for the cuprates sup ; Ghijsen _et al._ (1988); Zhang and Rice (1988); Lang _et al._ (2020). Using DFT-parameterized high-energy many-body Hamiltonian, we calculate the local many-body ground state via exact diagonalization. The ground state with a doped hole is a spin-singlet state similar to the well-known Zhang-Rice singlet Zhang and Rice (1988) with (self-)doped hole mostly residing in the ligand $p$-orbitals. Note that such a strong singlet formation introduces an important correction to Fig. 1 and 3: it pulls the energy of $x^{2}-y^{2}$ orbital closest to the chemical potential, even beyond the $3z^{2}-r^{2}$ orbital. This effect, however, will still respect the above mentioned trend concerning the relative energies of O/S orbitals and Cu/Ni orbitals. Figure 3: Comparison of orbital resolved Density of States in CaCuO2, NdNiS2 and NdNiO2. Notice the gradual reduction of the relative weights of O/S (green) $p_{\mathbin{\\!/\mkern-5.0mu/\\!}}$-orbitals against (red) Cu/Ni $d_{x^{2}-y^{2}}$-orbitals right below the Fermi energy from CaCuO2 to NdNiO2. Using this singlet state as basis, the low-energy Hamiltonian of hole carriers resembles the well-known $t$-$J$ model: (The subspace spanned by this singlet state form the basis for the low-energy effective Hamiltonian, upon integrating out the rest of the Hilbert space perturbatively.) $\begin{split}H=\sum_{ii^{\prime}\nu}t_{ii^{\prime}}\tilde{c}_{i\nu}^{\dagger}\tilde{c}_{i^{\prime}\nu}+\sum_{<i,j>}J\mathbf{S}_{i}\cdot\mathbf{S}_{j},\end{split}$ (1) where $\tilde{c}_{i\nu}^{\dagger}$ create a dressed hole at site $i$ of spin $\nu$. $\mathbf{S}_{i}=\sum_{\nu,\nu^{\prime}}\tilde{c}^{\dagger}_{i\nu}\bm{\sigma}_{\nu,\nu^{\prime}}\tilde{c}_{i\nu^{\prime}}$ denotes the spin operator and $\bm{\sigma}_{\nu,\nu^{\prime}}$ is the vector of Pauli matrices. Table 1 shows our resulting nearest neighbor hopping parameters $t_{ii^{\prime}}$ and super-exchange parameters $J$ for the three materials. Despite the larger lattice constant in NdNiS2, $t_{ii^{\prime}}$ turns out to be similar in all three materials owing to the larger radius of S $p$-orbitals. In contrast, $J$ is systematically enhanced from NdNiO2, to NdNiS2 and CaCuO2. This is because a stronger charge-transfer nature (higher $p$-orbital energy) gives a reduced charge-transfer gap $\Delta_{CT}$ (approximate energy to return an electron from the $p_{\mathbin{\\!/\mkern-5.0mu/\\!}}$-orbital back to Cu/Ni $d_{x^{2}-y^{2}}$-orbital.) With the intra-atomic repulsion roughly the same in Cu and Ni, this in turn enhances the super-exchange processes ( $\propto\Delta_{CT}^{-1}$) Ogata and Fukuyama (2008); Lang _et al._ (2020). We stress that despite the simplicity of such an estimation, the qualitative trend among these materials is robust. The enhanced $J$ is likely very important for the superconducting properties. It would not only lead to a stronger magnetic correlation that dominates the low-energy physical Hilbert space, but also give rise to a larger renormalized kinetic energy Yin and Ku (2009); Dagotto (1994). In other words, a larger $J$ can stretch the energy scale of all the low-energy physics, effectively producing a larger temperature scale in the phase diagram. One can therefore expect that NdNiS2 should have better superconducting properties than the nickelates. An interesting feature of NdNiS2 is that the possible electron-carrier density in the parent compound will increase as a result of higher-energy $p$-orbitals (c.f. Fig. 1 and 3). On the one hand, since the electron carriers are shown to be nearly decoupled from the hole carriers Lang _et al._ (2020) in the nickelates (and the same is found in NdNiS2 sup ), their existence should not interfere much with the hole superconductivity. On the other hand, these weakly correlated electron carriers might introduce additional physical effects $absent$ in the cuprates (for example, strengthening the essential superconducting phase stiffness.) Further experimental investigation of the nickel chalcogenides will prove highly illuminating. Finally, we note that it is not just S that has a good $p$-orbital energy, Se having a similar chemical orbital energy should also be suitable from our consideration. This opens up a wide range of tunability in material design, for example NdNiS2-xOx or NdNiS2-xSex, to optimize superconducting properties, or to explore systematic trends for better physical understanding. Table 1: Comparison of energy difference of $d_{x^{2}-y^{2}}$ and $p_{\mathbin{\\!/\mkern-5.0mu/\\!}}$, $\Delta_{pd}=\epsilon_{p_{\mathbin{\\!/\mkern-5.0mu/\\!}}}-\epsilon_{d_{x^{2}-y^{2}}}$; estimated charge transfer gap, $\Delta_{CT}$; hybridization between $d_{x^{2}-y^{2}}$ and $p_{\mathbin{\\!/\mkern-5.0mu/\\!}}$ orbitals, $t_{pd}$; nearest neighbor hopping $t$ and exchange parameter $J$ in one band $t-J$ model and $T_{c}$ Li _et al._ (2020a); Balestrino _et al._ (2001) for three different materials, CaCuO2, NdNiS2 and NdNiO2. | $\Delta_{pd}$ | $\Delta_{CT}$ | $t_{pd}$ | $t$ | $J$ | $T_{c}$ ---|---|---|---|---|---|--- CaCuO2 | 3.7 | $\sim$ 3.5 | 1.3 | 0.3 | $\sim$ 0.3 | $>$50K NdNiS2 | 5.7 | $\sim$ 4.0 | 1.2 | 0.3 | $\sim$ 0.13 | ? NdNiO2 | 8.9 | $\sim$ 6.0 | 1.3 | 0.3 | $\sim$ 0.07 | $\sim 12$K In conclusion, aiming to improve the superconducting properties of the newly discovered unconventional nickelate superconductors, we identify the degree of charge transfer as the key difference with the cuprates in their high-energy electronic structure. Guided by this, we propose a new family of material nickel chalcogenides as a promising candidate for improved superconducting properties. Taking NdNiS2 as an example, we find this compound stable under the desired crystal structure and thus realizable in laboratory. The resulting high-energy electronic structure displays the anticipated enhancement of the charge-transfer nature. We then reveal the physical benefits of a stronger charge-transfer characteristic via derivation of low-energy effective Hamiltonian. The resulting Hamiltonian encapsulates a stronger super-exchange spin-interaction, implying a higher temperature scale for all low-energy physics, including superconductivity. Our study paves the way to discover more nickel-based superconductors in nickel chalcogenides with improved superconducting properties, for examples NdNiS2-xOx and NdNiS2-xSex. Further experimental exploitation of the wide range of tunability through ligand substitution would likely make significant contribution to the resolution of the long-standing puzzles of high-temperature superconductivity. ###### Acknowledgements. This work is supported by National Natural Science Foundation of China (NSFC) #11674220 and 11745006 and Ministry of Science and Technology #2016YFA0300500 and 2016YFA0300501. ## References * Li _et al._ (2019) D. Li, K. Lee, B. Y. Wang, M. Osada, S. Crossley, H. R. Lee, Y. Cui, Y. Hikita, and H. Y. Hwang, Nature 572, 624 (2019). * Li _et al._ (2020a) D. Li, B. Y. Wang, K. Lee, S. P. Harvey, M. Osada, B. H. Goodge, L. F. Kourkoutis, and H. Y. Hwang, Phys. Rev. Lett. 125, 027001 (2020a). * Li _et al._ (2020b) Q. Li, C. He, J. Si, X. Zhu, Y. Zhang, and H.-H. Wen, Communications Materials 1, 16 (2020b). * Hayward and Rosseinsky (2003) M. Hayward and M. Rosseinsky, Solid State Sciences 5, 839 (2003), international Conference on Inorganic Materials 2002. * Sawatzky (2019) G. A. Sawatzky, Nature 572, 592 (2019). * Osada _et al._ (2020) M. Osada, B. Y. Wang, K. Lee, D. Li, and H. Y. Hwang, arXiv e-prints , arXiv:2010.16101 (2020), arXiv:2010.16101 [cond-mat.supr-con] . * Ando _et al._ (2004a) Y. Ando, Y. Kurita, S. Komiya, S. Ono, and K. Segawa, Phys. Rev. Lett. 92, 197001 (2004a). * Ando _et al._ (2004b) Y. Ando, S. Komiya, K. Segawa, S. Ono, and Y. Kurita, Phys. Rev. Lett. 93, 267001 (2004b). * Bozovic _et al._ (2016) I. Bozovic, X. He, J. Wu, and A. T. Bollinger, Nature 536, 309 EP (2016). * Dagotto (1994) E. Dagotto, Rev. Mod. Phys. 66, 763 (1994). * Schilling _et al._ (1993) A. Schilling, M. Cantoni, J. D. Guo, and H. R. Ott, Nature 363, 56 (1993). * Lang _et al._ (2020) Z.-J. Lang, R. Jiang, and W. Ku, arXiv e-prints , arXiv:2005.00022 (2020), arXiv:2005.00022 [cond-mat.supr-con] . * Lee _et al._ (2006) P. A. Lee, N. Nagaosa, and X.-G. Wen, Rev. Mod. Phys. 78, 17 (2006). * Cui _et al._ (2020) Y. Cui, C. Li, Q. Li, X. Zhu, Z. Hu, Y.-f. Yang, J. S. Zhang, R. Yu, H.-H. Wen, and W. Yu, arXiv e-prints , arXiv:2011.09610 (2020), arXiv:2011.09610 [cond-mat.supr-con] . * Anderson (1950) P. W. Anderson, Phys. Rev. 79, 350 (1950). * Anisimov _et al._ (1993) V. I. Anisimov, I. V. Solovyev, M. A. Korotin, M. T. Czyżyk, and G. A. Sawatzky, Phys. Rev. B 48, 16929 (1993). * Liechtenstein _et al._ (1995) A. I. Liechtenstein, V. I. Anisimov, and J. Zaanen, Phys. Rev. B 52, R5467 (1995). * (18) See Supplemental Material for details of phonon and band structure calculation. * Ku _et al._ (2010) W. Ku, T. Berlijn, and C.-C. Lee, Phys. Rev. Lett. 104, 216401 (2010). * Zaanen _et al._ (1985) J. Zaanen, G. A. Sawatzky, and J. W. Allen, Phys. Rev. Lett. 55, 418 (1985). * Ghijsen _et al._ (1988) J. Ghijsen, L. H. Tjeng, J. van Elp, H. Eskes, J. Westerink, G. A. Sawatzky, and M. T. Czyzyk, Phys. Rev. B 38, 11322 (1988). * Zhang and Rice (1988) F. C. Zhang and T. M. Rice, Phys. Rev. B 37, 3759 (1988). * Ogata and Fukuyama (2008) M. Ogata and H. Fukuyama, Reports on Progress in Physics 71, 036501 (2008). * Yin and Ku (2009) W.-G. Yin and W. Ku, Phys. Rev. B 79, 214512 (2009). * Balestrino _et al._ (2001) G. Balestrino, S. Lavanga, P. G. Medaglia, P. Orgiani, A. Paoletti, G. Pasquini, A. Tebano, and A. Tucciarone, Applied Physics Letters 79, 99 (2001).
APCTP-Pre2021-001 PNUTP-21-A11 Axion-driven hybrid inflation over a barrier Jinn-Ouk Gonga,b and Kwang Sik Jeongc aDepartment of Science Education, Ewha Womans University, Seoul 03760, Korea bAsia Pacific Center for Theoretical Physics, Pohang 37673, Korea cDepartment of Physics, Pusan National University, Busan 46241, Korea We present a novel cosmological scenario that describes both inflation and dark matter. A concrete realization of our scenario is given based on a well- established particle physics model, where an axionlike field drives inflation until a potential barrier, which keeps a waterfall field at the origin, disappears to trigger a waterfall transition. Such a barrier makes the inflaton potential much flatter, improving significantly the naturalness and viability of the otherwise problematic setup adopted previously. The observed spectrum of the cosmic microwave background indicates that the inflationary Hubble scale, which is allowed to span a wide range, uniquely fixes the inflaton mass and decay constant. This raises an intriguing possibility of probing inflation via experimental searches for axionlike particles. Further, our model involves dark matter candidates including the inflaton itself. Also, for a complex waterfall field, we can determine cosmologically the Peccei- Quinn scale associated with the strong CP problem. ## 1 INTRODUCTION Cosmic inflation [1, 2, 3] has become an essential part of the standard cosmological model. Before the onset of the hot big bang evolution, it provides the necessary initial conditions–otherwise extremely finely tuned–as confirmed by the observations on the cosmic microwave background (CMB) [4]. Furthermore, it explains the origin of temperature fluctuations of the CMB and the inhomogeneous distribution of galaxies on large scales due to quantum fluctuations during inflation [5]. The properties of these primordial perturbations have been constrained by decades of observations, and are consistent with the predictions of inflation [6]. To implement inflation, we need typically an inflaton field with a sufficiently flat potential. The inflaton drives an inflationary epoch until the “slow-roll” period does not hold any longer [7, 8]. It is, however, a formidable task to maintain an unusually flat potential against various corrections [9]. A powerful way of protecting the flatness is to impose certain symmetries. An axionlike field is an appealing candidate for an inflaton because its mass requires breaking of the associated shift symmetry, naturally making it very light. However, the predictions of the minimal axion- driven inflation – natural inflation [10, 11] – are only marginally consistent with the most recent Planck observations. Additionally, for successful inflation, the axion should have a trans-Planckian decay constant $f\gtrsim 5m_{\rm Pl}$ with $m_{\rm Pl}\equiv(8\pi G)^{-1/2}$ [6], which may be outside the range of validity of an effective field theory description. In this article, we present a novel cosmological scenario in which an axionlike field can drive inflation successfully and at the same time contribute to dark matter. The end of the inflationary phase is triggered by a waterfall transition like hybrid inflation [12]. The distinctive features of our model are twofold. First, the waterfall field $\chi$ is trapped at the origin during inflation by a potential barrier. This implies that, differently from the previous hybrid inflation models, the scale of inflation is not tied to that of the waterfall phase transition. As a result, unlike the original natural inflation, the inflaton is allowed to have a decay constant well below $m_{\rm Pl}$ so that the effective field theory is trustable, yet maintains a flat potential. Thus the naturalness and viability of the models, which were problematic either in the effective theory viewpoint or in unnaturally fine- tuned initial conditions, are improved significantly within our scenario. The Planck results are accommodated in a broad range of the inflaton mass and decay constant, but with a certain relationship between them. This opens an interesting possibility to probe inflation via experimental searches for axionlike particles. Another merit of our scenario is that the inflaton itself can constitute dark matter, which is generally difficult in other inflation models. Further, because the inflationary Hubble scale $H_{\rm inf}$ is allowed to span a very wide range, $\chi$ has a potential to resolve other puzzles of the Standard Model (SM). If complex, we may identify U$(1)_{\chi}$ with the Peccei-Quinn (PQ) symmetry solving the strong CP problem [13]. Remarkably, the PQ scale is then determined cosmologically, and the contribution of the QCD axion to dark matter constrains $H_{\rm inf}$ to be below about $10^{4}$ GeV. ## 2 MODEL Our model consists of two scalar fields, the inflaton $\phi$ and the waterfall field $\chi$. During inflation, $\phi$ rolls down the potential slowly while $\chi$ is trapped at the origin by a barrier. There, the effective mass squared of $\chi$, $\mu_{\rm eff}^{2}$, is thus positive. As $\phi$ evolves, $\mu_{\rm eff}^{2}$ decreases monotonically and vanishes at a critical value $\phi_{c}$, removing the barrier. Then inflation ends almost instantaneously. Here, the barrier does bring the separation between the scales for inflation and waterfall phase transition. Our scenario is successfully realized if $\phi$ is an axionlike field with decay constant $f$. This is because its interactions are well controlled by shift symmetry, $\phi\to\phi+{\rm constant}$, presumably broken only by nonperturbative effects, and the size of its potential terms is finite and insensitive to $f$. A dangerous waterfall tadpole can be avoided by imposing a symmetry, for instance, U$(1)_{\chi}$ if $\chi$ is a complex scalar. Explicitly, we consider the potential $V(\phi,\chi)=V_{0}+\mu^{2}_{\rm eff}(\phi)|\chi|^{2}-\lambda|\chi|^{4}+\frac{1}{\Lambda^{2}}|\chi|^{6}+U(\phi)\,,$ (1) with $\lambda>0$, and the $\phi$-dependent terms given respectively by $\displaystyle\mu^{2}_{\rm eff}(\phi)$ $\displaystyle=m^{2}-\mu^{2}\cos\left(\frac{\phi}{f}+\alpha\right)\,,$ (2) $\displaystyle U(\phi)$ $\displaystyle=M^{4}\cos\left(\frac{\phi}{f}\right)\,,$ (3) where $\alpha$ is constant. The positive constant $V_{0}$ is fixed by demanding $V=0$ at the true vacuum, and $\Lambda$ is the cutoff scale of the theory. The parameter space of our interest is $m^{4}<\mu^{4}\ll\lambda V_{0}\quad\text{and}\quad M^{4}\ll V_{0}\,.$ (4) Then the true vacuum appears at $\chi_{0}\sim\sqrt{\lambda}\Lambda$, well below the cutoff scale $\Lambda$ as long as $\lambda\ll 1$, and $V_{0}$ reads $V_{0}\sim\lambda^{3}\Lambda^{4}\,.$ (5) Note that there are two minima along the $\chi$-direction for $\mu^{2}_{\rm eff}(\phi)>0$, and a barrier separates them. The position and height of the waterfall barrier are determined by $\mu$, whereas the value of $V_{0}$ is insensitive to it. Before going further, let us discuss the case $\lambda<0$ so that there is no barrier, similar to hybrid natural inflation [14, 15, 16]. In such a case, $V_{0}$ is fixed by $\mu$ roughly to be $\mu^{4}/|\lambda|$, and the possible range of $M^{4}/V_{0}$ is severely constrained because a closed loop of $\chi$ generally makes $|\lambda|\gtrsim 1/16\pi^{2}$ and $M^{4}\gtrsim\mu^{2}\Lambda^{2}_{\ast}/16\pi^{2}$ with $\mu<\Lambda_{\ast}$. Here $\Lambda_{\ast}$ is the cutoff scale of the $\phi$-dependent waterfall mass operator. In our scenario, $M^{4}/V_{0}$ can be arbitrarily small, making the inflaton potential much flatter than the case without a barrier. The rate of tunneling over a barrier is proportional to $\exp(-S_{E})$, where $S_{E}$ is the Euclidean action of $\chi$ evaluated on a bounce solution. Tunneling proceeds dominantly via the Coleman-De Luccia bounce [17] with $S_{E}>S_{0}\equiv 8\pi^{2}/(3\lambda)$ in the region with $\mu^{2}_{\rm eff}>2H^{2}_{\rm inf}$, while through the Hawking-Moss instantons [18] with $S_{E}=\mu^{4}_{\rm eff}/H^{4}_{\rm inf}\times S_{0}$ in the opposite region [19]. Here we have used that the bounce is insensitive to the $|\chi|^{6}$ term for $\mu^{4}\ll\lambda V_{0}$. For viable inflation, we thus impose the condition $\mu^{2}\gg H_{\rm inf}^{2}\,.$ (6) Then $\chi$ is heavy enough to be initially fixed at the origin. In addition, the tunneling rate is exponentially suppressed so that $\chi$ stays at the origin until the barrier disappears at $\phi_{c}$. Bubbles of true vacuum can be nucleated around the end of inflation, but the U$(1)_{\chi}$ phase transition occurs rather smoothly because the barrier soon disappears. As a simple ultraviolet completion of the inflaton potential, we consider a hidden QCD with U$(1)_{\chi}$ charged quarks. (2) and (3) are then generated in a controllable way while naturally satisfying the hierarchies (4). Vectorlike quarks $u+u^{c}$ and $d+d^{c}$ couple to $\chi$ through the U$(1)_{\chi}$ and gauge invariant interactions $m_{u}uu^{c}+y\chi u^{c}d+y^{\prime}\chi^{*}ud^{c}+m_{d}dd^{c}+\frac{1}{16\pi^{2}}\frac{\phi}{f}\,G_{\mu\nu}\widetilde{G}^{\mu\nu}\,,$ (7) where the hidden confining scale lies in the range $m_{d}\ll\Lambda_{h}\ll m_{u}$. Here we have taken the field basis where the quark mass parameters are real. Note that the last term above is an anomalous inflaton coupling to hidden gluons, which is the only source of shift symmetry breaking. At energy scales below $m_{u}$, $u+u^{c}$ are integrated out to give a $\chi$-dependent effective quark mass $\left(\frac{yy^{\prime}}{m_{u}}|\chi|^{2}+m_{d}+\delta m_{d}\right)dd^{c}\,.$ (8) Here we have included the radiative contribution from a closed loop of $\chi$ $\delta m_{d}=\frac{yy^{\prime}}{16\pi^{2}}m_{u}\log\left(\frac{\Lambda^{2}}{m^{2}_{\chi}}\right)\,,$ (9) with $m_{\chi}$ being the mass of the radial component of $\chi$. For small values of $\chi$, $d+d^{c}$ are lighter than $\Lambda_{h}$ and condensate to form a meson with mass and decay constant around $\Lambda_{h}$. The inflaton mixes with the meson in the presence of anomalous coupling to hidden gluons, and finally the effective potential at scales below $\Lambda_{h}$ is obtained by integrating out the heavy meson $\Delta V_{\rm eff}=-\left|\frac{yy^{\prime}}{m_{u}}\right|\Lambda^{3}_{h}\cos\left(\frac{\phi}{f}+\beta_{1}\right)|\chi|^{2}+\left|m_{d}+\delta m_{d}\right|\Lambda^{3}_{h}\cos\left(\frac{\phi}{f}+\beta_{2}\right)\,,$ (10) where the constant phases are given by $\beta_{1}=\arg(yy^{\prime}/m_{u})$ and $\beta_{2}=\arg(m_{d}+\delta m_{d})$. It is clear that the above reduces to (2) and (3) with $\alpha=\beta_{1}-\beta_{2}$. Also, the hierarchies $\mu^{4}\ll\lambda V_{0}$ and $M^{4}\ll V_{0}$ are satisfied naturally if $H_{\rm inf}\lesssim\Lambda_{h}\ll\Lambda\,,$ (11) where we have used that $H_{\rm inf}$ should be lower than $\Lambda_{h}$ since otherwise instanton effects become very weak. On the other hand, $m$ should be smaller than $\mu$ because inflation ends when the barrier disappears. The smallness of $m$ may be accommodated in more speculative idea like supersymmetry or anthropic selection. ## 3 COSMOLOGICAL DYNAMICS ### 3.1 Inflation The Universe undergoes an inflationary phase while $\chi$ is trapped at the origin. During this stage, $\phi$ evolves down the potential $V=V_{0}+U(\phi)=V_{0}+M^{4}\cos\left(\frac{\phi}{f}\right)\,.$ (12) Thus, the evolution of $\phi$ during inflation is essentially identical to hybrid natural inflation. $\mu^{2}_{\rm eff}$ crosses zero when $\phi$ reaches the critical value $\frac{\phi_{c}}{f}=\cos^{-1}\left(\frac{m^{2}}{\mu^{2}}\right)-\alpha\,.$ (13) The sign flip triggers the waterfall phase transition, because there is no potential barrier along the $\chi$-direction, and inflation ends almost instantaneously. Among the model parameters, $m$, $\mu$, and $\alpha$ affect inflation only through the above combination. Figure 1 shows schematically the inflationary and waterfall phases. Figure 1: Schematic display of inflationary and waterfall phases. The evolution of $\phi$ changes the waterfall potential dramatically near the origin, as illustrated in the right panel, but rarely around the true vacuum at $\chi=\chi_{0}$ if (4) holds. At this point, it is very important to note that two crucial ingredients are required to make our scenario distinctive. One is the shift symmetry of $\phi$, which naturally allows the hierarchies (4). The other is a $\phi$-dependent barrier between two extrema at and off the origin in the waterfall potential. The barrier makes the inflaton potential flatter, and consequently the value of $f$ required for viable inflation can naturally be much lower than $m_{\rm Pl}$. For $M^{4}\ll V_{0}$, inflation is driven by $V_{0}$ so that $H_{\rm inf}^{2}\approx\frac{V_{0}}{3m_{\rm Pl}^{2}}\,.$ (14) From (12), the slow-roll parameters are given by, with $\theta\equiv\phi/f$, $\displaystyle\epsilon$ $\displaystyle\equiv\frac{m_{\rm Pl}^{2}}{2}\left(\frac{V^{\prime}}{V}\right)^{2}\approx\frac{1}{2}\left(\frac{m_{\rm Pl}}{f}\right)^{2}\left(\frac{M^{4}}{V_{0}}\right)^{2}\sin^{2}\theta\,,$ (15) $\displaystyle\eta$ $\displaystyle\equiv m_{\rm Pl}^{2}\frac{V^{\prime\prime}}{V}\approx-\left(\frac{m_{\rm Pl}}{f}\right)^{2}\left(\frac{M^{4}}{V_{0}}\right)\cos\theta\,.$ (16) Thus, $|\eta|$ is parametrically much bigger than $\epsilon$. The slow-roll conditions, $\epsilon\ll 1$ and $|\eta|\ll 1$, are satisfied if the following condition holds $f\gtrsim\left(\frac{M^{4}}{V_{0}}\right)^{1/2}m_{\rm Pl}\,,$ (17) but it need not be above $m_{\rm Pl}$. The amplitude of the power spectrum of the curvature perturbation and its spectral index, and the tensor-to-scalar ratio in terms of the slow-roll parameters are constrained as [6] $\displaystyle A_{\cal R}$ $\displaystyle=\frac{V_{0}}{24\pi^{2}m_{\rm Pl}^{4}\epsilon_{*}}\approx 2.0989^{+0.0296}_{-0.0292}\times 10^{-9}\,,$ (18) $\displaystyle n_{\cal R}$ $\displaystyle=1-6\epsilon_{*}+2\eta_{*}\approx 0.9656\pm 0.0042\,,$ (19) $\displaystyle r$ $\displaystyle=16\epsilon_{*}<0.056\,,$ (20) where the subscript $*$ denotes the evaluation at the horizon exit. Since $\epsilon\ll|\eta|$, $n_{\cal R}$ is determined entirely by $\eta$. Hence, from (16) and (19), $f$ is written as $f=\sqrt{\frac{2}{1-n_{\cal R}}\cos\theta_{*}}\bigg{(}\frac{M^{4}}{V_{0}}\bigg{)}^{1/2}m_{\rm Pl}\approx 7.625\sqrt{\cos\theta_{*}}\bigg{(}\frac{M^{4}}{V_{0}}\bigg{)}^{1/2}m_{\rm Pl}\,,$ (21) while (20) is translated to the following mild constraint $\frac{M^{4}}{V_{0}}<\frac{0.056}{8}\frac{2}{1-n_{\cal R}}\cos\theta_{*}\approx 0.4070\cos\theta_{*}\,.$ (22) The number of $e$-folds $N$, before the onset of the waterfall phase transition, is estimated by $N=\frac{1}{m_{\rm Pl}}\int_{\phi_{c}}^{\phi}\frac{d\phi^{\prime}}{\sqrt{2\epsilon}}\approx\frac{V_{0}}{M^{4}}\bigg{(}\frac{f}{m_{\rm Pl}}\bigg{)}^{2}\log\bigg{[}\frac{\tan(\theta_{c}/2)}{\tan(\theta/2)}\bigg{]}\approx 58.14\cos\theta_{*}\log\bigg{[}\frac{\tan(\theta_{c}/2)}{\tan(\theta/2)}\bigg{]}\,,$ (23) thus we can use $\theta$ and $N$ interchangeably. The required number of $e$-folds is around 60, which fixes $\theta_{\ast}$ roughly as $\theta_{\ast}\approx 0.7126\tan\bigg{(}\frac{\theta_{c}}{2}\bigg{)}-0.1566\tan^{3}\bigg{(}\frac{\theta_{c}}{2}\bigg{)}\,,$ (24) neglecting terms of higher order in $\tan(\theta_{c}/2)$. Thus, $\theta_{\ast}$ does not need to be very close to the hilltop of the potential. The inflaton mass during inflation sets the lower bound of the inflaton mass at the true vacuum, $m_{\phi}|_{\rm min}=M^{2}/f$. It is interesting to note that both $f$ and $m_{\phi}|_{\rm min}$ are proportional to $H_{\text{inf}}$; combining (18) with (14) and (21), they are written respectively as $\displaystyle f$ $\displaystyle=\frac{H_{\text{inf}}}{\pi(1-n_{\cal R})\sqrt{A_{\cal R}}\tan\theta_{*}}\approx\frac{2.020\times 10^{5}}{\tan\theta_{*}}H_{\text{inf}}\,,$ (25) $\displaystyle m_{\phi}|_{\rm min}$ $\displaystyle=\sqrt{\frac{3(1-n_{\cal R})}{2\cos\theta_{*}}}H_{\text{inf}}\approx\frac{0.2272}{\sqrt{\cos\theta_{*}}}H_{\text{inf}}\,.$ (26) Figure 2 shows the relationship between $H_{\rm inf}$, $m_{\phi}|_{\rm min}$ and $f$. $\phi$ can couple to the SM sector, for instance, to gauge bosons through anomalous couplings as naturally expected from its axionic nature. Our scenario thus provides theoretical support for experimental searches for axionlike particles in a wide mass range. The rough relation, $f\sim 10^{6}\times m_{\phi}|_{\rm min}$, indicates that $\phi$ should be heavier than about $0.1$ GeV to avoid too rapid cooling of stars [20], if coupled to photons. Another plausible possibility is that it instead couples to the Higgs sector as in the cosmological relaxation model of the weak scale [21]. Then, it may be detectable at collider and beam-dump experiments. For instance, the mass range below a few GeV can be probed by experiments at SHiP [22] and NA62 [23]. Figure 2: Decay constant and lower bound on mass of the inflaton compatible with the Planck results on $n_{\cal R}$ and $A_{\cal R}$. We have taken $\theta_{\ast}$ lying between 0.01 (solid lines) and 1.5 (dotted lines). The postinflationary evolution leads to very rich phenomenologies. The quantitative predictions depend very much on the detail of the model, so here we are satisfied with describing briefly the subsequent evolution. After the barrier disappears, $\chi$ soon acquires a tachyonic mass much larger than $H_{\text{inf}}$ in magnitude for $\mu^{2}\gg H^{2}_{\text{inf}}$. This happens within an $e$-fold after $\phi=\phi_{c}$, so $\chi$ rolls fast down to the true vacuum. Unlike usual axion-driven inflation where the Universe is reheated via an anomalous coupling to photons, the reheating process depends greatly on the details of the model. Generally speaking, however, depending on the couplings, tachyonic preheating [24] is extremely effective so that the vacuum energy is rapidly transferred to the energy of inhomogeneous oscillations of $\phi$ and/or $\chi$ [25, 26], subsequently heating up the Universe to a radiation-dominated regime. After inflation, spontaneous U$(1)_{\chi}$ breaking occurs and leads to the formation of cosmic strings [27], which can survive in the late Universe and contribute to the CMB temperature anisotropies, depending on how the associated Nambu-Goldstone boson becomes massive. For instance, for global U$(1)_{\chi}$, it can obtain a mass nonperturbatively from some confining gauge sector. Then, topological defects are unstable and collapse by the wall tension if the domain-wall number is equal to unity, or if a small explicit symmetry-breaking term is added to lift the vacuum degeneracy [28, 29]. Further, cosmic string loops and large time-dependent inhomogeneities generated during tachyonic preheating can act as a source of gravitational waves (GWs). The corresponding GW spectrum can span a huge range of frequencies; from ${\cal O}(10^{-12})$ Hz to ${\cal O}(1)$ Hz for stable and metastable cosmic strings [30], within the reach of pulsar-timing arrays, LIGO, and LISA, and from ${\cal O}(1)$ Hz to ${\cal O}(10^{10})$ Hz for inhomogeneities from tachyonic preheating [31, 32, 33]. Such high-frequency GWs are unfortunately beyond the sensitivity of interferometric experiments due to the shot-noise fluctuations of photons. GWs in relatively low-frequency regimes may well be within the reach of future detectors like advanced LIGO, the Einstein Telescope, and the Big Bang Observer, which however is possible only for extremely small values of couplings. ### 3.2 Dark matter Another distinctive feature of our scenario is the possibility that the inflaton can contribute to dark matter if its potential arises from hidden QCD as in (7). Having Yukawa interactions with $\chi$, the hidden quarks have masses increasing with the waterfall field value. This implies that $\Lambda_{h}$ also increases, and thus the hidden QCD gets stronger after inflation. In the region of large waterfall field values where all the hidden quarks are heavier than $\Lambda_{h}$, we have $\mu^{2}=0\quad\text{and}\quad M^{4}=\Lambda^{4}_{h}\,,$ (27) in (2) and (3), because there are no mesons formed. Let us consider the case that $\chi_{0}$ is sufficiently large so that (27) holds in the present Universe. $\phi$ is then stabilized at a CP-conserving minimum, and consequently accidental $Z_{2}$ symmetry arises: $\phi\to-\phi$. The $Z_{2}$ forbids $\phi$ to mix with $\chi$, making it stable if does not couple to SM. $\phi$ starts coherent oscillations around the minimum when $H$ becomes comparable to its mass, i.e. at the temperature fixed by $m_{\phi}(T_{1})=3H(T_{1})\,.$ (28) If $T$ is below $\Lambda_{h}$ during reheating, oscillations start before reheating ends. Then, the inflaton relic density from oscillation is roughly estimated by $\Omega_{\phi}h^{2}\sim 0.24\,\theta^{2}_{c}\left(\frac{T_{1}}{\Lambda_{h}}\right)^{n}\left(\frac{f}{10^{11}{\rm GeV}}\right)^{2}\left(\frac{T_{\rm reh}}{10^{5}{\rm GeV}}\right)\,,$ (29) with $T_{1}<\Lambda_{h}$, and $n=11N/6-2$ for confining SU$(N)$. Here $T_{\rm reh}$ is the reheating temperature at which the Universe becomes completely radiation-dominated, and we have used that the scale factor scales as $H^{-2/3}$ during a matter-dominated era. If oscillations start after reheating, the relic density can be read off from (29) by replacing $T_{\rm reh}$ with $T_{1}$. $\phi$ can thus account for the observed dark matter in a wide range of $f$ depending on $T_{\rm reh}$. It is worth noting that $\alpha=0$ leads to accidental $Z_{2}$ even when (27) is not the case [34]. The inflation sector includes another candidate for dark matter associated with spontaneously broken U$(1)_{\chi}$. An interesting possibility arising due to a wide allowed range of $H_{\rm inf}$ is to identify U$(1)_{\chi}$ with the PQ symmetry so that $\arg{\chi}$ becomes the QCD axion explaining the absence of CP violation in QCD. This implies that the waterfall and PQ phase transitions are identical. Then, as corresponds to $\chi_{0}$, the axion decay constant is cosmologically determined by $f_{a}\approx\frac{3.8\times 10^{11}\,{\rm GeV}}{\lambda^{1/4}}\left(\frac{H_{\rm inf}}{10^{4}{\rm GeV}}\right)^{1/2}\,,$ (30) which should be above about $10^{9}$ GeV to avoid astrophysical bounds. The axion anomalous coupling to gluons, which is required to solve the strong CP problem, is generated by adding U$(1)_{\chi}$-charged heavy quarks or extra Higgs doublets [35, 36]. We also note that the domain-wall number should be equal to one since otherwise domain walls formed during the QCD phase transition overclose the Universe. In such a case, axions are produced from coherent oscillations and, more efficiently, from unstable domain-walls bounded by an axion string. The relic density is estimated by [37] $\Omega_{a}h^{2}\approx 0.54\times\left(\frac{\Lambda_{\rm QCD}}{400{\rm MeV}}\right)\left(\frac{f_{a}}{10^{11}{\rm GeV}}\right)^{1.19}\,.$ (31) Therefore, the observed dark-matter density indicates $H_{\rm inf}\lesssim\sqrt{\lambda}\times 10^{4}\,{\rm GeV}\,.$ (32) It would be also interesting to consider other cases where U$(1)_{\chi}$ is identified, for instance, with U$(1)_{L}$ associated with the seesaw mechanism or local U$(1)_{B-L}$ to extend SM. ## 4 CONCLUSIONS We have proposed a cosmological scenario that improves significantly the naturalness and viability of axion-driven inflation. During inflation, the waterfall field remains at the origin by a potential barrier, which disappears when the inflaton reaches at a critical point–then inflation ends almost instantaneously. The inflaton interaction responsible for such a barrier can naturally arise if the shift symmetry is broken nonperturbatively by hidden QCD with quarks coupled to the waterfall field. Interestingly, the Planck results indicate the possibility of probing our scenario by experimental searches for axionlike particles. It is also remarkable that the inflaton can be stable enough to constitute dark matter if all the hidden quarks get heavier than the confining scale at the true vacuum. Further, for the case of a complex waterfall field, its phase component can play the role of the QCD axion, contributing to dark matter for $H_{\rm inf}$ below about $10^{4}$ GeV. ### Acknowledgements This work is supported in part by the National Research Foundation of Korea Grants No. 2018R1C1B6006061, No. 2021R1A4A5031460 (K.S.J.) and No. 2019R1A2C2085023 (J.G.). We also acknowledge the Korea-Japan Basic Scientific Cooperation Program supported by the National Research Foundation of Korea and the Japan Society for the Promotion of Science (2020K2A9A2A08000097). J.G. is further supported in part by the Ewha Womans University Research Grant of 2020 (1-2020-1630-001-1). J.G. is grateful to the Asia Pacific Center for Theoretical Physics for hospitality while this work was under progress. ## References * [1] A. H. Guth, Phys. Rev. D 23, 347-356 (1981) * [2] A. D. Linde, Phys. Lett. B 108, 389-393 (1982) * [3] A. Albrecht and P. J. Steinhardt, Phys. Rev. Lett. 48, 1220-1223 (1982) * [4] N. Aghanim et al. [Planck], Astron. Astrophys. 641, A6 (2020) [arXiv:1807.06209 [astro-ph.CO]]. * [5] V. F. Mukhanov and G. V. Chibisov, JETP Lett. 33, 532-535 (1981) * [6] Y. Akrami et al. [Planck], Astron. Astrophys. 641, A10 (2020) [arXiv:1807.06211 [astro-ph.CO]]. * [7] V. Mukhanov, “Physical Foundations of Cosmology,” Cambridge, UK: Univ. Pr. (2005) 421 p. * [8] S. Weinberg, “Cosmology,” Oxford, UK: Oxford Univ. Pr. (2008) 593 p. * [9] D. H. Lyth and A. Riotto, Phys. Rept. 314, 1-146 (1999) [arXiv:hep-ph/9807278 [hep-ph]]. * [10] K. Freese, J. A. Frieman and A. V. Olinto, Phys. Rev. Lett. 65, 3233-3236 (1990) * [11] F. C. Adams, J. R. Bond, K. Freese, J. A. Frieman and A. V. Olinto, Phys. Rev. D 47, 426-455 (1993) [arXiv:hep-ph/9207245 [hep-ph]]. * [12] A. D. Linde, Phys. Rev. D 49, 748-754 (1994) [arXiv:astro-ph/9307002 [astro-ph]]. * [13] R. D. Peccei and H. R. Quinn, Phys. Rev. Lett. 38, 1440-1443 (1977) * [14] G. G. Ross and G. German, Phys. Lett. B 684, 199-204 (2010) [arXiv:0902.4676 [hep-ph]]. * [15] G. G. Ross and G. German, Phys. Lett. B 691, 117-120 (2010) [arXiv:1002.0029 [hep-ph]]. * [16] G. G. Ross, G. German and J. A. Vazquez, JHEP 05, 010 (2016) [arXiv:1601.03221 [astro-ph.CO]]. * [17] S. R. Coleman and F. De Luccia, Phys. Rev. D 21, 3305 (1980) * [18] S. W. Hawking and I. G. Moss, Phys. Lett. B 110 (1982), 35-38. * [19] A. Shkerin and S. Sibiryakov, Phys. Lett. B 746, 257-260 (2015) [arXiv:1503.02586 [hep-ph]]. * [20] G. G. Raffelt, Ann. Rev. Nucl. Part. Sci. 49, 163 (1999) [hep-ph/9903472]. * [21] P. W. Graham, D. E. Kaplan and S. Rajendran, Phys. Rev. Lett. 115, no.22, 221801 (2015) [arXiv:1504.07551 [hep-ph]]. * [22] S. Alekhin et al., Rept. Prog. Phys. 79, no. 12, 124201 (2016) [arXiv:1504.04855 [hep-ph]]. * [23] S. Martellotti, arXiv:1510.00172 [physics.ins-det]. * [24] G. N. Felder, J. Garcia-Bellido, P. B. Greene, L. Kofman, A. D. Linde and I. Tkachev, Phys. Rev. Lett. 87, 011601 (2001) [arXiv:hep-ph/0012142 [hep-ph]]. * [25] J. Garcia-Bellido and A. D. Linde, Phys. Rev. D 57, 6075-6088 (1998) [arXiv:hep-ph/9711360 [hep-ph]]. * [26] E. J. Copeland, S. Pascoli and A. Rajantie, Phys. Rev. D 65, 103517 (2002) [arXiv:hep-ph/0202031 [hep-ph]]. * [27] R. Jeannerot, J. Rocher and M. Sakellariadou, Phys. Rev. D 68, 103514 (2003) [arXiv:hep-ph/0308134 [hep-ph]]. * [28] G. B. Gelmini, M. Gleiser and E. W. Kolb, Phys. Rev. D 39, 1558 (1989). * [29] S. E. Larsson, S. Sarkar and P. L. White, Phys. Rev. D 55, 5129 (1997) [hep-ph/9608319]. * [30] P. Auclair, J. J. Blanco-Pillado, D. G. Figueroa, A. C. Jenkins, M. Lewicki, M. Sakellariadou, S. Sanidas, L. Sousa, D. A. Steer and J. M. Wachter, et al. JCAP 04, 034 (2020) [arXiv:1909.00819 [astro-ph.CO]]. * [31] J. Garcia-Bellido and D. G. Figueroa, Phys. Rev. Lett. 98, 061302 (2007) [arXiv:astro-ph/0701014 [astro-ph]]. * [32] J. Garcia-Bellido, D. G. Figueroa and A. Sastre, Phys. Rev. D 77, 043517 (2008) [arXiv:0707.0839 [hep-ph]]. * [33] J. F. Dufaux, G. Felder, L. Kofman and O. Navros, JCAP 03, 001 (2009) [arXiv:0812.2917 [astro-ph]]. * [34] S. H. Im and K. S. Jeong, Phys. Lett. B 799, 135044 (2019) [arXiv:1907.07383 [hep-ph]]. * [35] A. Ringwald, Phys. Dark Univ. 1, 116 (2012) [arXiv:1210.5081 [hep-ph]]. * [36] M. Kawasaki and K. Nakayama, Ann. Rev. Nucl. Part. Sci. 63, 69 (2013) [arXiv:1301.1123 [hep-ph]]. * [37] T. Hiramatsu, M. Kawasaki, K. Saikawa and T. Sekiguchi, JCAP 01, 001 (2013) [arXiv:1207.3166 [hep-ph]].
# Graph Neural Network for Traffic Forecasting: A Survey Weiwei Jiang Department of Electronic Engineering, Tsinghua University, Beijing, 100084, China Jiayun Luo School of Computer Science and Engineering, Nanyang Technological University, 639798, Singapore ###### Abstract Traffic forecasting is important for the success of intelligent transportation systems. Deep learning models, including convolution neural networks and recurrent neural networks, have been extensively applied in traffic forecasting problems to model spatial and temporal dependencies. In recent years, to model the graph structures in transportation systems as well as contextual information, graph neural networks have been introduced and have achieved state-of-the-art performance in a series of traffic forecasting problems. In this survey, we review the rapidly growing body of research using different graph neural networks, e.g. graph convolutional and graph attention networks, in various traffic forecasting problems, e.g. road traffic flow and speed forecasting, passenger flow forecasting in urban rail transit systems, and demand forecasting in ride-hailing platforms. We also present a comprehensive list of open data and source resources for each problem and identify future research directions. To the best of our knowledge, this paper is the first comprehensive survey that explores the application of graph neural networks for traffic forecasting problems. We have also created a public GitHub repository where the latest papers, open data, and source resources will be updated. ###### keywords: Traffic Forecasting , Graph Neural Networks , Graph Convolution Network , Graph Attention Network , Deep Learning ††journal: Journal of LaTeX Templates ## 1 Introduction Transportation systems are among the most important infrastructure in modern cities, supporting the daily commuting and traveling of millions of people. With rapid urbanization and population growth, transportation systems have become more complex. Modern transportation systems encompass road vehicles, rail transport, and various shared travel modes that have emerged in recent years, including online ride-hailing, bike-sharing, and e-scooter sharing. Expanding cities face many transportation-related problems, including air pollution and traffic congestion. Early intervention based on traffic forecasting is seen as the key to improving the efficiency of a transportation system and to alleviate transportation-related problems. In the development and operation of smart cities and intelligent transportation systems (ITSs), traffic states are detected by sensors (e.g. loop detectors) installed on roads, subway and bus system transaction records, traffic surveillance videos, and even smartphone GPS data collected in a crowd-sourced fashion. Traffic forecasting is typically based on consideration of historical traffic state data, together with the external factors which affect traffic states, e.g. weather and holidays. Both short-term and long-term traffic forecasting problems for various transport modes are considered in the literature. This survey focuses on the data-driven approach, which involves forecasting based on historical data. The traffic forecasting problem is more challenging than other time series forecasting problems because it involves large data volumes with high dimensionality, as well as multiple dynamics including emergency situations, e.g. traffic accidents. The traffic state in a specific location has both spatial dependency, which may not be affected only by nearby areas, and temporal dependency, which may be seasonal. Traditional linear time series models, e.g. auto-regressive and integrated moving average (ARIMA) models, cannot handle such spatiotemporal forecasting problems. Machine learning (ML) and deep learning techniques have been introduced in this area to improve forecasting accuracy, for example, by modeling the whole city as a grid and applying a convolutional neural network (CNN) as demonstrated by Jiang & Zhang [2018]. However, the CNN-based approach is not optimal for traffic foresting problems that have a graph-based form, e.g. road networks. In recent years, graph neural networks (GNNs) have become the frontier of deep learning research, showing state-of-the-art performance in various applications [Wu et al., 2020b]. GNNs are ideally suited to traffic forecasting problems because of their ability to capture spatial dependency, which is represented using non-Euclidean graph structures. For example, a road network is naturally a graph, with road intersections as the nodes and road connections as the edges. With graphs as the input, several GNN-based models have demonstrated superior performance to previous approaches on tasks including road traffic flow and speed forecasting problems. These include, for example, the diffusion convolutional recurrent neural network (DCRNN) [Li et al., 2018b] and Graph WaveNet [Wu et al., 2019] models. The GNN-based approach has also been extended to other transportation modes, utilizing various graph formulations and models. To the best of the authors’ knowledge, this paper presents the first comprehensive literature survey of GNN-related approaches to traffic forecasting problems. While several relevant traffic forecasting surveys exist [Shi & Yeung, 2018, Pavlyuk, 2019, Yin et al., 2021, Luca et al., 2020, Fan et al., 2020, Boukerche & Wang, 2020a, Manibardo et al., 2020, Ye et al., 2020a, Lee et al., 2021, Xie et al., 2020a, George & Santra, 2020, Haghighat et al., 2020, Boukerche et al., 2020, Tedjopurnomo et al., 2020, Varghese et al., 2020], most of them are not GNN-focused with only one exception [Ye et al., 2020a]. For this survey, we reviewed 212 papers published in the years 2018 to 2020. Additionally, because this is a very rapidly developing research field, we also included preprints that have not yet gone through the traditional peer review process (e.g., arXiv papers) to present the latest progress. Based on these studies, we identify the most frequently considered problems, graph formulations, and models. We also investigate and summarize publicly available useful resources, including datasets, software, and open-sourced code, for GNN-based traffic forecasting research and application. Lastly, we identify the challenges and future directions of applying GNNs to the traffic forecasting problem. Instead of giving a whole picture of traffic forecasting, our aim is to provide a comprehensive summary of GNN-based solutions. This paper is useful for both the new researchers in this field who want to catch up with the progress of applying GNNs and the experienced researchers who are not familiar with these latest graph-based solutions. In addition to this paper, we have created an open GitHub repository on this topic 111https://github.com/jwwthu/GNN4Traffic, where relevant content will be updated continuously. Our contributions are summarized as follows: 1) Comprehensive Review: We present the most comprehensive review of graph- based solutions for traffic forecasting problems in the past three years (2018-2020). 2) Resource Collection: We provide the latest comprehensive list of open datasets and code resources for replication and comparison of GNNs in future work. 3) Future Directions: We discuss several challenges and potential future directions for researchers in this field, when using GNNs for traffic forecasting problems. The remainder of this paper is organized as follows. In Section 2, we compare our work with other relevant research surveys. In Section 3, we categorize the traffic forecasting problems that are involved with GNN-based models. In Section 4, we summarize the graphs and GNNs used in the reviewed studies. In Section 5, we outline the open resources. Finally, in Section 6, we point out challenges and future directions. ## 2 Related Research Surveys In this section, we introduce the most recent relevant research surveys (most of which were published in 2020). The differences between our study and these existing surveys are pointed out when appropriate. We start with the surveys addressing wider ITS topics, followed by those focusing on traffic prediction problems and GNN application in particular. Besides traffic forecasting, machine learning and deep learning methods have been widely used in ITSs as discussed in Haghighat et al. [2020], Fan et al. [2020], Luca et al. [2020]. In Haghighat et al. [2020], GNNs are only mentioned in the task of traffic characteristics prediction. Among the major milestones of deep-learning driven traffic prediction (summarized in Figure 2 of Fan et al. [2020]), the state-of-the-art models after 2019 are all based on GNNs, indicating that GNNs are indeed the frontier of deep learning-based traffic prediction research. Roughly speaking, five different types of traffic prediction methods are identified and categorized in previous surveys [Xie et al., 2020a, George & Santra, 2020], namely, statistics-based methods, traditional machine learning methods, deep learning-based methods, reinforcement learning-based methods, and transfer learning-based methods. Some comparisons between different categories have been considered, e.g., statistics-based models have better model interpretability, whereas ML-based models are more flexible as discussed in Boukerche et al. [2020]. Machine learning models for traffic prediction are further categorized in Boukerche & Wang [2020a], which include the regression model, example-based models (e.g., k-nearest neighbors), kernel-based models (e.g. support vector machine and radial basis function), neural network models, and hybrid models. Deep learning models are further categorized into five different generations in Lee et al. [2021], in which GCNs are classified as the fourth generation and other advanced techniques that have been considered but are not yet widely applied are merged into the fifth generation. These include transfer learning, meta learning, reinforcement learning, and the attention mechanism. Before these advanced techniques become mature in traffic prediction tasks, GNNs remain the state-of-the-art technique. Some of the relevant surveys only focus on the progress of deep learning-based methods [Tedjopurnomo et al., 2020], while the others prefer to compare them with the statistics-based and machine learning methods [Yin et al., 2021, Manibardo et al., 2020]. In Tedjopurnomo et al. [2020], 37 deep neural networks for traffic prediction are reviewed, categorized, and discussed. The authors conclude that encoder-decoder long short term-memory (LSTM) combined with graph-based methods is the state-of-the-art prediction technique. A detailed explanation of various data types and popular deep neural network architectures is also provided, along with challenges and future directions for traffic prediction. Conversely, it is found that deep learning is not always the best modeling technique in practical applications, where linear models and machine learning techniques with less computational complexity can sometimes be preferable [Manibardo et al., 2020]. Additional research surveys consider aspects other than model selection. In Pavlyuk [2019], spatiotemporal feature selection and extraction pre-processing methods, which may also be embedded as internal model processes, are reviewed. A meta-analysis of prediction accuracy when applying deep learning methods to transport studies is given in Varghese et al. [2020]. In this study, apart from the models themselves, additional factors including sample size and prediction time horizon are shown to have a significant influence on prediction accuracy. To the authors’ best knowledge, there are no existing surveys focusing on the application of GNNs for traffic forecasting. Graph-based deep learning architectures are reviewed in Ye et al. [2020a], for a series of traffic applications, namely, traffic congestion, travel demand, transportation safety, traffic surveillance, and autonomous driving. Specific and practical guidance for constructing graphs in these applications is provided. The advantages and disadvantages of both GNNs and other deep learning models (e.g. RNN, TCN, Seq2Seq, and GAN) are examined. While the focus is not limited to traffic prediction problems, the graph construction process is universal in the traffic domain when GNNs are involved. ## 3 Problems In this section, we discuss and categorize the different types of traffic forecasting problems considered in the literature. Problems are first categorized by the traffic state to be predicted. Traffic flow, speed, and demand problems are considered separately while the remaining types are grouped together under “other problems”. Then, the problem-types are further broken down into levels according to where the traffic states are defined. These include road-level, region-level, and station-level categories. Different problem types have different modelling requirements for representing spatial dependency. For the road-level problems, the traffic data are usually collected from sensors, which are associated with specific road segments, or GPS trajectory data, which are also mapped into the road network with map matching techniques. In this case, the road network topology can be seen as the graph to use, which may contain hundreds or thousands of road segments potentially. The spatial dependency may be described by the road network connectivity or spatial proximity. For the station-level problems, the metro or bus station topology can be taken as the graph to use, which may contain tens or hundreds of stations potentially. The spatial dependency may be described by the metro lines or bus routes. For the region-level problem, the regular or irregular regions are used as the nodes in a graph. The spatial dependency between different regions can be extracted from the land use purposes, e.g., from the points-of-interest data. A full list of the traffic forecasting problems considered in the surveyed studies is shown in Table LABEL:tab:problems. Instead of giving the whole picture of traffic forecasting research, only those problems with GNN-based solutions in the literature are listed in Table I. Table 1: Traffic forecasting problems in the surveyed studies. Problem | Relevant Studies ---|--- Road Traffic Flow | Zhang et al. [2018b], Wei et al. [2019], Xu et al. [2020a], Guo et al. [2020a], Zheng et al. [2020b], Pan et al. [2020, 2019], Lu et al. [2019a], Mallick et al. [2020], Zhang et al. [2020j, l], Bai et al. [2020], Fang et al. [2019], Huang et al. [2020a], Wang et al. [2018b], Zhang et al. [2020e], Song et al. [2020a], Xu et al. [2020b], Wang et al. [2020g], Chen et al. [2020e], Lv et al. [2020], Kong et al. [2020], Fukuda et al. [2020], Zhang & Guo [2020], Boukerche & Wang [2020b], Tang et al. [2020b], Kang et al. [2019], Guo et al. [2019c], Li et al. [2019], Xu et al. [2019], Zhang et al. [2019d], Wu et al. [2018a], Sun et al. [2020], Wei & Sheng [2020], Li et al. [2020f], Cao et al. [2020], Yu et al. [2018, 2019b], Li et al. [2020b], Yin et al. [2020], Chen et al. [2020g], Zhang et al. [2020a], Wang et al. [2020a], Xin et al. [2020], Qu et al. [2020], Wang et al. [2020b], Xie et al. [2020d], Huang et al. [2020b], Guo et al. [2020b], Zhang et al. [2020h], Fang et al. [2020a], Li & Zhu [2021], Tian et al. [2020], Xu et al. [2020c], Chen et al. [2020c] Road OD Flow | Xiong et al. [2020], Ramadan et al. [2020] Intersection Traffic Throughput | Sánchez et al. [2020] Regional Taxi Flow | Zhou et al. [2020d], Sun et al. [2020], Chen et al. [2020d], Wang et al. [2018a], Peng et al. [2020], Zhou et al. [2019], Wang et al. [2020e], Qiu et al. [2020] Regional Bike Flow | Zhou et al. [2020d], Sun et al. [2020], Wang et al. [2018a, 2020e] Regional Ride-hailing Flow | Zhou et al. [2019] Regional Dockless E-Scooter Flow | He & Shin [2020a] Regional OD Taxi Flow | Wang et al. [2020e], Yeghikyan et al. [2020] Regional OD Bike Flow | Wang et al. [2020e] Regional OD Ride-hailing Flow | Shi et al. [2020], Wang et al. [2020h, 2019] Station-level Subway Passenger Flow | Fang et al. [2019, 2020a], Peng et al. [2020], Ren & Xie [2019], Li et al. [2018a], Zhao et al. [2020a], Han et al. [2019], Zhang et al. [2020b, c], Li et al. [2020e], Liu et al. [2020b], Ye et al. [2020b], Ou et al. [2020] Station-level Bus Passenger Flow | Fang et al. [2019, 2020a], Peng et al. [2020] Station-level Shared Vehicle Flow | Zhu et al. [2019] Station-level Bike Flow | He & Shin [2020b], Chai et al. [2018] Station-level Railway Passenger Flow | He et al. [2020] Road Traffic Speed | Li et al. [2018b], Wu et al. [2019], Zhang et al. [2018b], Wei et al. [2019], Xu et al. [2020a], Guo et al. [2020a], Zheng et al. [2020b], Pan et al. [2020, 2019], Lu et al. [2019a], Mallick et al. [2020], Zhang et al. [2020j], Lv et al. [2020], Li et al. [2020f], Yin et al. [2020], Guo et al. [2020b], Li & Zhu [2021], Chen et al. [2020d], Zhao et al. [2020a], Bai et al. [2021], Tang et al. [2020a], James [2020], Shin & Yoon [2020], Liu et al. [2020a], Zhang et al. [2018a, 2019f], Yu & Gu [2019], Xie et al. [2019], Zhang et al. [2019a], Guo et al. [2019a], Diao et al. [2019], Cirstea et al. [2019], Lu et al. [2019b], Zhang et al. [2019c], James [2019], Ge et al. [2019a, b], Zhang et al. [2019b], Lee & Rhee [2019a], Shleifer et al. [2019], Yu et al. [2020a], Ge et al. [2020], Lu et al. [2020b], Yang et al. [2020], Zhao et al. [2019], Cui et al. [2019], Chen et al. [2019], Zhang et al. [2019e], Yu et al. [2019a], Lee & Rhee [2019b], Bogaerts et al. [2020], Wang et al. [2020f], Cui et al. [2020b, a], Guo et al. [2020c], Zhou et al. [2020a], Cai et al. [2020], Zhou et al. [2020b], Wu et al. [2020c], Chen et al. [2020f], Opolka et al. [2019], Mallick et al. [2021], Oreshkin et al. [2021], Jia et al. [2020], Sun et al. [2021], Guo & Yuan [2020], Xie et al. [2020b], Zhang et al. [2020i], Zhu et al. [2020c], Feng et al. [2020], Zhu et al. [2020a], Fu et al. [2020], Zhang et al. [2020d], Xie et al. [2020c], Park et al. [2020], Agafonov [2020], Chen et al. [2020a], Lu et al. [2020a], Jepsen et al. [2019, 2020], Bing et al. [2020], Lewenfus et al. [2020], Zhu et al. [2020b], Liao et al. [2018], Maas & Bloem [2020], Li et al. [2020d], Song et al. [2020b], Zhao et al. [2020b], Guopeng et al. [2020], Kim et al. [2020] Road Travel Time | Guo et al. [2020a], Hasanzadeh et al. [2019], Fang et al. [2020b], Shao et al. [2020], Shen et al. [2020] Traffic Congestion | Dai et al. [2020], Mohanty & Pozdnukhov [2018], Mohanty et al. [2020], Qin et al. [2020a], Han et al. [2020] Time of Arrival | Hong et al. [2020] Regional OD Taxi Speed | Hu et al. [2018] Ride-hailing Demand | Pian & Wu [2020], Jin et al. [2020b], Li & Axhausen [2020], Jin et al. [2020a], Geng et al. [2019b], Lee et al. [2019], Bai et al. [2019b], Geng et al. [2019a], Bai et al. [2019a], Ke et al. [2021], Li et al. [2020c] Taxi Demand | Lee et al. [2019], Bai et al. [2019b, a], Ke et al. [2019], Hu et al. [2020], Zheng et al. [2020a], Xu & Li [2019], Davis et al. [2020], Chen et al. [2020h], Du et al. [2020], Li & Moura [2020], Wu et al. [2020a], Ye et al. [2021] Shared Vehicle Demand | Luo et al. [2020] Bike Demand | Lee et al. [2019], Bai et al. [2019b, a], Du et al. [2020], Ye et al. [2021], Chen et al. [2020b], Wang et al. [2020d], Qin et al. [2020b], Xiao et al. [2020], Yoshida et al. [2019], Guo et al. [2019b], Kim et al. [2019], Lin et al. [2018] Traffic Accident | Zhou et al. [2020e], Yu et al. [2020b], Zhang et al. [2020k], Zhou et al. [2020f] Traffic Anomaly | Liu et al. [2020c] Parking Availability | Zhang et al. [2020g], Yang et al. [2019], Zhang et al. [2020f] Transportation Resilience | Wang et al. [2020c] Urban Vehicle Emission | Xu et al. [2020d] Railway Delay | Heglund et al. [2020] Lane Occupancy | Wright et al. [2019] Generally speaking, traffic forecasting problems are challenging, not only for the complex temporal dependency, but only for the complex spatial dependency. While many solutions have been proposed for dealing with the time dependency, e.g., recurrent neural networks and temporal convolutional networks, the problem to capture and model the spatial dependency has not been fully solved. The spatial dependency, which refers to the complex and nonlinear relationship between the traffic state in one particular location with other locations. This location could be a road intersection, a subway station, or a city region. The spatial dependency may not be local, e.g., the traffic state may not only be affected by nearby areas, but also those which are far away in the spatial range but connected by a fast transportation tool. The graphs are necessary to capture such kind of spatial information as we would discuss in the next section. Before the usage of graph theories and GNNs, the spatial information is usually extracted by multivariate time series models or CNNs. Within a multivariate time series model, e.g., vector autoregression, the traffic states collected in different locations or regions are combined together as multivariate time series. However, the multivariate time series models can only extract the linear relationship among different states, which is not enough for modeling the complex and nonlinear spatial dependency. CNNs take a step further by modeling the local spatial information, e.g., the whole spatial range is divided into regular grids as the two-dimensional image format and the convolution operation is performed in the neighbor grids. However, the CNN-based approach is bounded to the case of Euclidean structure data, which cannot model the topological structure of the subway network or the road network. Graph neural networks bring new opportunities for solving traffic forecasting problems, because of their strong learning ability to capture the spatial information hidden in the non-Euclidean structure data, which are frequently seen in the traffic domain. Based on graph theories, both nodes and edges have their own attributes, which can be used further in the convolution or aggregation operations. These attributes describe different traffic states, e.g., volume, speed, lane numbers, road level, etc. For the dynamic spatial dependency, dynamic graphs can be learned from the data automatically. For the case of hierarchical traffic problems, the concepts of super-graphs and sub- graphs can be defined and further used. ### 3.1 Traffic Flow We consider three levels of traffic flow problems in this survey, namely, road-level flow, region-level flow, and station-level flow. Road-level flow problems are concerned with traffic volumes on a road and include road traffic flow, road origin-destination (OD) Flow, and intersection traffic throughput. In road traffic flow problems, the prediction target is the traffic volume that passes a road sensor or a specific location along the road within a certain time period (e.g. five minutes). In the road OD flow problem, the target is the volume between one location (the origin) and another (the destination) at a single point in time. The intersection traffic throughput problem considers the volume of traffic moving through an intersection. Region-level flow problems consider traffic volume in a region. A city may be divided into regular regions (where the partitioning is grid-based) or irregular regions (e.g. road-based or zip-code-based partitions). These problems are classified by transport mode into regional taxi flow, regional bike flow, regional ride-hailing flow, regional dockless e-scooter flow, regional OD taxi flow, regional OD bike flow, and regional OD ride-hailing flow problems. Station-level flow problems relate to the traffic volume measured at a physical station, for example, a subway or bus station. These problems are divided by station type into station-level subway passenger flow, station- level bus passenger flow, station-level shared vehicle flow, station-level bike flow, and station-level railway passenger flow problems. Road-level traffic flow problems are further divided into cases of unidirectional and bidirectional traffic flow, whereas region-level and station-level traffic flow problems are further divided into the cases of inflow and outflow, based on different problem formulations. ### 3.2 Traffic Speed We consider two levels of traffic speed problems in this survey, namely, road- level and region-level problems. We also include travel time and congestion predictions in this category because they are closely correlated to traffic speed. For example, in several studies, traffic congestion is judged by a threshold-based speed inference. The specific road-level speed problem categories considered are road traffic speed, road travel time, traffic congestion, and time of arrival problems; while the region-level speed problem considered is regional OD taxi speed. ### 3.3 Traffic Demand Traffic demand refers to the potential demand for travel, which may or may not be fulfilled completely. For example, on an online ride-hailing platform, the ride requests sent by passengers represent the demand, whereas only a subset of these requests may be served depending on the supply of drivers and vehicles, especially during rush hours. Accurate prediction of travel demand is a key element of vehicle scheduling systems (e.g. online ride-hailing or taxi dispatch platforms). However, in some cases, it is difficult to collect the potential travel demand from passengers and a compromise method using transaction records as an indication of the traffic demand is used. In such cases the real demand may be underestimated. Based on transport mode, the traffic demand problems considered include ride-hailing demand, taxi demand, shared vehicle demand, and bike demand. ### 3.4 Other Problems In addition to the above three categories of traffic forecasting problems, GNNs are also being applied to the following problems. Traffic accident and Traffic anomaly: the target is to predict the traffic accident number reported to the police system. A traffic accident is usually an accident in road traffic involving different vehicles, which may cause significant loss of life and property. The traffic anomaly has a broader definition that deviates from the normal traffic state, e.g., the traffic jam caused by a traffic accident or a public procession. Parking availability: the target is to predict the availability of vacant parking space for cars in the streets or in a car parking lot. Urban vehicle emission: while not directly related to traffic states, the prediction of urban vehicle emission is considered in Xu et al. [2020d]. Urban vehicle emission refers to the emission produced by motor vehicles, e.g., those use internal combustion engines. Urban vehicle emission is a major source of air pollutants and its amount is affected by different traffic states, e.g., the excess emission would be created in traffic congestion situations. Railway delay: the delay time of specific routes in the railway system is considered in Heglund et al. [2020]. Lane occupancy: With simulated traffic data, lane occupancy has been measured and predicted [Wright et al., 2019]. ## 4 Graphs and Graph Neural Networks In this section, we summarize the types of graphs and GNNs used in the surveyed studies, focusing on GNNs that are frequently used for traffic forecasting problems. The contributions of this section include an organized approach for classifying the different traffic graphs based on the domain knowledge, and a summary of the common ways for constructing adjacency matrices, which may not be encountered in other neural networks before and would be very helpful for those who would like to use graph neural networks. The different GNN structures already used for traffic forecasting problems are briefly introduced in this section too. For a wider and deeper discussion of GNNs, refer to Wu et al. [2020b], Zhou et al. [2020c], Zhang et al. [2020m]. ### 4.1 Traffic Graphs #### 4.1.1 Graph Construction A graph is the basic structure used in GNNs. It is defined as $G=(V,E,A)$, where $V$ is the set of vertices or nodes, $E$ is the set of edges between the nodes, and $A$ is the adjacency matrix Element $a_{ij}$ of $A$ represents the “edge weight” between nodes $i$ and $j$. Both the nodes and the edges may represent different attributes in different GNN problems. For traffic forecasting, the traffic state prediction target is usually one of the node features. We divide the time axis into discrete time steps, e.g. five minutes or one hour, depending on the specific scenario. In single step forecasting, the traffic state in the next time step is predicted, whereas in multiple step forecasting the traffic state several time steps later is the prediction target. The traffic state at time step $i$ is denoted by $\chi_{i}$, and the forecasting problem is formulated as: find the function $f$ which generates $y=f(\mathbf{\chi};G)$, where $y$ is the traffic state to be predicted, $\mathbf{\chi}=\\{\chi_{1},\chi_{2},...,\chi_{N}\\}$ is the historical traffic state defined on graph $G$, and $N$ is the number of time steps in the historical window size. As mentioned in Section 1, traffic states can be highly affected by external factors, e.g. weather and holidays. The forecasting problem formulation, extended to incorporate these external factors, takes the form $y=f(\mathbf{\chi},\varepsilon;G)$, where $\varepsilon$ represents the external factors. Various graph structures are used to model traffic forecasting problems depending on both the forecasting problem-type and the traffic datasets available. These graphs can be pre-defined static graphs, or dynamic graphs continuously learned from the data. The static graphs can be divided into two types, namely, natural graphs and similarity graphs. Natural graphs are based on a real-world transportation system, e.g. the road network or subway system; whereas similarity graphs are based solely on the similarity between different node attributes where nodes may be virtual stations or regions. We categorize the existing traffic graphs into the same three levels used in Section 3, namely, road-level, region-level and station-level graphs, as shown in the examples in Figures 1, 1, and 1, respectively. Figure 1: Examples of the different levels of graphs. 1 a road-level graph: the road network in the Performance Measurement System (PeMS) where each sensor is a node; source: http://pems.dot.ca.gov/; 1 a region-level graph: the zip codes of Manhattan where each zip code zone is a node; source: https://maps-manhattan.com/manhattan-zip-code-map; and 1 a station-level graph, the Beijing subway system where each subway station is a node; source: https://www.travelchinaguide.com/cityguides/beijing/transportation/subway.htm. Road-level graphs. These include sensor graphs, road segment graphs, road intersection graphs, and road lane graphs. Sensor graphs are based on traffic sensor data (e.g. the PeMS dataset) where each sensor is a node, and the edges are road connections. The other three graphs are based on road networks with the nodes formed by road segments, road intersections, and road lanes, respectively. Region-level graphs. These include irregular region graphs, regular region graphs, and OD graphs. In both irregular and regular region graphs the nodes are regions of the city. Regular region graphs, which have grid-based partitioning, are listed separately because of their natural connection to previous widely used grid-based forecasting using CNNs, in which the grids may be seen as image pixels. Irregular region graphs include all other partitioning approaches, e.g. road based, or zip code based Ke et al. [2019]. In the OD graph, the nodes are origin region - destination region pairs. In these graphs, the edges are usually defined with a spatial neighborhood or other similarities. Station-level graphs. These include subway station graphs, bus station graphs, bike station graphs, railway station graphs, car-sharing station graphs, parking lot graphs, and parking block graphs. Usually, there are natural links between stations that are used to define the edges, e.g. subway or railway lines, or the road network. A full list of the traffic graphs used in the surveyed studies is shown in Table 2. Sensor graphs and road segment graphs are most frequently used because they are compatible with the available public datasets as discussed in Section 5. It is noted that in some studies multiple graphs are used as simultaneous inputs and then fused to improve the forecasting performance [Lv et al., 2020, Zhu et al., 2019]. Table 2: Traffic graphs in the surveyed studies. Graph | Relevant Studies ---|--- Sensor Graph | Li et al. [2018b], Wu et al. [2019], Xu et al. [2020a], Zheng et al. [2020b], Pan et al. [2020, 2019], Lu et al. [2019a], Mallick et al. [2020], Zhang et al. [2020j], Bai et al. [2020], Huang et al. [2020a], Zhang et al. [2020e], Song et al. [2020a], Xu et al. [2020b], Wang et al. [2020g], Chen et al. [2020e], Lv et al. [2020], Kong et al. [2020], Fukuda et al. [2020], Zhang & Guo [2020], Boukerche & Wang [2020b], Tang et al. [2020b], Kang et al. [2019], Guo et al. [2019c], Li et al. [2019], Sun et al. [2020], Wei & Sheng [2020], Li et al. [2020f], Cao et al. [2020], Yu et al. [2018, 2019b], Li et al. [2020b], Yin et al. [2020], Chen et al. [2020g], Zhang et al. [2020a], Wang et al. [2020a], Xin et al. [2020], Xie et al. [2020d], Huang et al. [2020b], Li & Zhu [2021], Tian et al. [2020], Xu et al. [2020c], Chen et al. [2020c], Xiong et al. [2020], Chen et al. [2020d], Tang et al. [2020a], Zhang et al. [2018a, 2019a], Cirstea et al. [2019], Ge et al. [2019a, b], Shleifer et al. [2019], Ge et al. [2020], Yang et al. [2020], Zhao et al. [2019], Cui et al. [2019], Chen et al. [2019], Yu et al. [2019a], Wang et al. [2020f], Cui et al. [2020b, a], Zhou et al. [2020a], Cai et al. [2020], Zhou et al. [2020b], Wu et al. [2020c], Chen et al. [2020f], Opolka et al. [2019], Mallick et al. [2021], Oreshkin et al. [2021], Jia et al. [2020], Sun et al. [2021], Guo & Yuan [2020], Zhang et al. [2020i], Feng et al. [2020], Xie et al. [2020c], Park et al. [2020], Chen et al. [2020a], Lewenfus et al. [2020], Maas & Bloem [2020], Li et al. [2020d], Song et al. [2020b], Zhao et al. [2020b], Wang et al. [2020c] Road Segment Graph | Zhang et al. [2018b], Guo et al. [2020a], Pan et al. [2019], Zhang et al. [2020j, l], Wang et al. [2018b], Zhang et al. [2020e], Lv et al. [2020], Zhang et al. [2019d, 2020a], Qu et al. [2020], Guo et al. [2020b], Ramadan et al. [2020], Zhao et al. [2020a], Bai et al. [2021], Shin & Yoon [2020], Liu et al. [2020a], Yu & Gu [2019], Xie et al. [2019], Guo et al. [2019a], Diao et al. [2019], Lu et al. [2019b], Zhang et al. [2019c], James [2019], Zhang et al. [2019b], Lee & Rhee [2019a], Yu et al. [2020a], Lu et al. [2020b], Zhao et al. [2019], Cui et al. [2019], Zhang et al. [2019e], Lee & Rhee [2019b], Cui et al. [2020b, a], Guo et al. [2020c], Xie et al. [2020b], Zhu et al. [2020c, a], Fu et al. [2020], Zhang et al. [2020d], Agafonov [2020], Lu et al. [2020a], Jepsen et al. [2019, 2020], Zhu et al. [2020b], Liao et al. [2018], Guopeng et al. [2020], Kim et al. [2020], Hasanzadeh et al. [2019], Fang et al. [2020b], Dai et al. [2020], Han et al. [2020], Hong et al. [2020], Chen et al. [2020h], Yu et al. [2020b] Road Intersection Graph | Zhang et al. [2018b], Wei et al. [2019], Fang et al. [2019], Zhang et al. [2020e], Xu et al. [2019], Wu et al. [2018a], Sánchez et al. [2020], James [2020], Zhang et al. [2019f], Lu et al. [2019b], Zhang et al. [2019c], Bogaerts et al. [2020], Shao et al. [2020], Qin et al. [2020a] Road Lane Graph | Wright et al. [2019] Irregular Region Graph | Zhou et al. [2020d], Sun et al. [2020], Chen et al. [2020d], Bing et al. [2020], Mohanty & Pozdnukhov [2018], Mohanty et al. [2020], Hu et al. [2018], Li & Axhausen [2020], Bai et al. [2019b, a], Ke et al. [2021], Hu et al. [2020], Zheng et al. [2020a], Davis et al. [2020], Du et al. [2020], Li & Moura [2020], Ye et al. [2021], Zhang et al. [2020k], Liu et al. [2020c] Regular Region Graph | Pan et al. [2020], Wang et al. [2020b], Zhang et al. [2020h], Wang et al. [2018a], Zhou et al. [2019], Wang et al. [2020e], Qiu et al. [2020], He & Shin [2020a], Yeghikyan et al. [2020], Shi et al. [2020], Wang et al. [2019], Shen et al. [2020], Pian & Wu [2020], Jin et al. [2020b, a], Geng et al. [2019b], Lee et al. [2019], Geng et al. [2019a], Li et al. [2020c], Xu & Li [2019], Davis et al. [2020], Wu et al. [2020a], Zhou et al. [2020e, f], Xu et al. [2020d] OD Graph | Wang et al. [2020h], Ke et al. [2019] Subway Station Graph | Fang et al. [2019, 2020a], Ren & Xie [2019], Li et al. [2018a], Zhao et al. [2020a], Han et al. [2019], Zhang et al. [2020b, c], Li et al. [2020e], Liu et al. [2020b], Ye et al. [2020b], Ou et al. [2020] Bus Station Graph | Fang et al. [2019, 2020a] Bike Station Graph | He & Shin [2020b], Chai et al. [2018], Du et al. [2020], Chen et al. [2020b], Wang et al. [2020d], Qin et al. [2020b], Xiao et al. [2020], Yoshida et al. [2019], Guo et al. [2019b], Kim et al. [2019], Lin et al. [2018] Railway Station Graph | He et al. [2020], Heglund et al. [2020] Car-sharing Station Graph | Zhu et al. [2019], Luo et al. [2020] Parking Lot Graph | Zhang et al. [2020g, f] Parking Block Graph | Yang et al. [2019] #### 4.1.2 Adjacency Matrix Construction Adjacency matrices are seen as the key to capturing spatial dependency in traffic forecasting [Ye et al., 2020a]. While nodes may be fixed by physical constraints, the user typically has control over the design of the adjacency matrix, which can even be dynamically trained from continuously evolving data. We extend the categories of adjacency matrices used in previous studies [Ye et al., 2020a] and divide them into four types, namely, road-based, distance- based, similarity-based, and dynamic matrices. Road-based Matrix. This type of adjacency matrix relates to the road network and includes connection matrices, transportation connectivity matrices, and direction matrices. A connection matrix is a common way of representing the connectivity between nodes. It has a binary format, with an element value of 1 if connected and 0 otherwise. The transportation connectivity matrix is used where two regions are geographically distant but conveniently reachable by motorway, highway, or subway [Ye et al., 2020a]. It also includes cases where the connection is measured by travel time between different nodes, e.g. if a vehicle can travel between two intersections in less than 5 minutes then there is an edge between the two intersections [Wu et al., 2018a]. The less commonly used direction matrix takes the angle between road links into consideration. Distance-based Matrix. This widely used matrix-type represents the spatial closeness between nodes. It contains two sub-types, namely, neighbor and distance matrices. In neighbor matrices, the element values are determined by whether the two regions share a common boundary (if connected the value is set to 1, generally, or 1/4 for grids, and 0 otherwise). In distance matrices, the element values are a function of geometrical distance between nodes. This distance may be calculated in various ways, e.g. the driving distance between two sensors, the shortest path length along the road [Kang et al., 2019, Lee & Rhee, 2019a], or the proximity between locations calculated by the random walk with restart (RWR) algorithm [Zhang et al., 2019e]. Similarity-based Matrix. This type of matrix is divided into two sub-types, namely, traffic pattern and functional similarity matrices. Traffic pattern similarity matrices represent the correlations between traffic states, e.g. similarities of flow patterns, mutual dependencies between different locations, and traffic demand correlation in different regions. Functional similarity matrices represent, for example, the distribution of different types of point-of-interests in different regions. Dynamic Matrix. This type of matrix is used when no pre-defined static matrices are used. Many studies have demonstrated the advantages of using dynamic matrices, instead of a pre-defined adjacency matrix, for various traffic forecasting problems. A full list of the adjacency matrices applied in the surveyed studies is shown in Table 3. Dynamic matrices are listed at the bottom of the table, with no further subdivisions. The connection and distance matrices are the most frequently used types, because of their simple definition and representation of spatial dependency. Table 3: Adjacency matrices in the surveyed studies. Adjacency Matrix | Relevant Studies ---|--- Connection Matrix | Zhang et al. [2018b], Wei et al. [2019], Xu et al. [2020a], Guo et al. [2020a], Zhang et al. [2020l], Wang et al. [2018b], Song et al. [2020a], Zhang & Guo [2020], Xu et al. [2019], Cao et al. [2020], Yu et al. [2019b], Chen et al. [2020g], Zhang et al. [2020a], Qu et al. [2020], Wang et al. [2020b], Huang et al. [2020b], Xiong et al. [2020], Sánchez et al. [2020], Wang et al. [2020h], Zhang et al. [2020c], Li et al. [2020e], Liu et al. [2020b], Ou et al. [2020], He et al. [2020], Bai et al. [2021], Liu et al. [2020a], Zhang et al. [2019f], Yu & Gu [2019], Xie et al. [2019], Guo et al. [2019a], Lu et al. [2019b], Zhang et al. [2019c], James [2019], Zhang et al. [2019b], Zhao et al. [2019], Cui et al. [2019, 2020b, 2020a], Wu et al. [2020c], Opolka et al. [2019], Sun et al. [2021], Guo & Yuan [2020], Xie et al. [2020b], Zhu et al. [2020c, a], Zhang et al. [2020d], Agafonov [2020], Chen et al. [2020a], Lu et al. [2020a], Bing et al. [2020], Zhu et al. [2020b], Fang et al. [2020b], Shao et al. [2020], Shen et al. [2020], Qin et al. [2020a], Hong et al. [2020], Xu & Li [2019], Davis et al. [2020], Chen et al. [2020h], Wang et al. [2020d], Zhou et al. [2020e], Yu et al. [2020b], Liu et al. [2020c], Zhang et al. [2020g, f], Heglund et al. [2020], Yin et al. [2020], Zhang et al. [2020b] Transportation Connectivity Matrix | Pan et al. [2020, 2019], Lv et al. [2020], Wu et al. [2018a], Ye et al. [2020b], Geng et al. [2019b, a], Luo et al. [2020], Wright et al. [2019] Direction Matrix | Shin & Yoon [2020], Lee & Rhee [2019a, b] Neighbor Matrix | Wang et al. [2018a], Yeghikyan et al. [2020], Shi et al. [2020], Wang et al. [2019], Hu et al. [2018], Geng et al. [2019b], Lee et al. [2019], Ke et al. [2021, 2019], Hu et al. [2020], Zheng et al. [2020a], Yoshida et al. [2019] Distance Matrix | Li et al. [2018b], Zheng et al. [2020b], Pan et al. [2020, 2019], Lu et al. [2019a], Mallick et al. [2020], Huang et al. [2020a], Xu et al. [2020b], Wang et al. [2020g], Boukerche & Wang [2020b], Kang et al. [2019], Sun et al. [2020], Wei & Sheng [2020], Yu et al. [2018], Li et al. [2020b], Chen et al. [2020g], Wang et al. [2020a], Xin et al. [2020], Xie et al. [2020d], Li & Zhu [2021], Tian et al. [2020], Xu et al. [2020c], Chen et al. [2020c], Zhou et al. [2020d], Chen et al. [2020d], He & Shin [2020a], Ren & Xie [2019], Zhu et al. [2019], He & Shin [2020b], Chai et al. [2018], Shin & Yoon [2020], Zhang et al. [2018a], Ge et al. [2019a, b], Lee & Rhee [2019a], Shleifer et al. [2019], Ge et al. [2020], Yang et al. [2020], Chen et al. [2019], Zhang et al. [2019e], Lee & Rhee [2019b], Bogaerts et al. [2020], Wang et al. [2020f], Guo et al. [2020c], Zhou et al. [2020a], Cai et al. [2020], Zhou et al. [2020b], Chen et al. [2020f], Mallick et al. [2021], Jia et al. [2020], Zhang et al. [2020i], Feng et al. [2020], Xie et al. [2020c], Li et al. [2020d], Song et al. [2020b], Zhao et al. [2020b], Kim et al. [2020], Mohanty & Pozdnukhov [2018], Mohanty et al. [2020], Jin et al. [2020b], Li & Axhausen [2020], Jin et al. [2020a], Geng et al. [2019a], Ke et al. [2021], Li et al. [2020c], Ke et al. [2019], Luo et al. [2020], Chen et al. [2020b], Xiao et al. [2020], Guo et al. [2019b], Kim et al. [2019], Lin et al. [2018], Yang et al. [2019], Wang et al. [2020c], Xu et al. [2020d] Traffic Pattern Similarity Matrix | Lv et al. [2020], Li & Zhu [2021], Xu et al. [2020c], Zhou et al. [2020d], Sun et al. [2020], Wang et al. [2020e], He & Shin [2020a], Ren & Xie [2019], Han et al. [2019], Liu et al. [2020b], He & Shin [2020b], Chai et al. [2018], Lu et al. [2020a], Lewenfus et al. [2020], Dai et al. [2020], Han et al. [2020], Jin et al. [2020b], Li & Axhausen [2020], Jin et al. [2020a], Bai et al. [2019b, a], Li et al. [2020c], Ke et al. [2019], Chen et al. [2020b], Wang et al. [2020d], Yoshida et al. [2019], Kim et al. [2019], Lin et al. [2018], Zhou et al. [2020f] Functional Similarity Matrix | Lv et al. [2020], He & Shin [2020a], Shi et al. [2020], Zhu et al. [2019], Ge et al. [2019a, b, 2020], Jin et al. [2020b], Geng et al. [2019b, a], Ke et al. [2019], Luo et al. [2020], Zhang et al. [2020k] Dynamic Matrix | Wu et al. [2019], Bai et al. [2020], Fang et al. [2019], Zhang et al. [2020e], Chen et al. [2020e], Kong et al. [2020], Tang et al. [2020b], Guo et al. [2019c], Li et al. [2019], Zhang et al. [2019d], Li et al. [2020f], Guo et al. [2020b], Zhang et al. [2020h], Peng et al. [2020], Zhou et al. [2019], Shi et al. [2020], Li et al. [2018a], Tang et al. [2020a], Zhang et al. [2019a], Diao et al. [2019], Yu et al. [2020a], Fu et al. [2020], Maas & Bloem [2020], Li & Axhausen [2020], Du et al. [2020], Li & Moura [2020], Wu et al. [2020a], Ye et al. [2021] ### 4.2 Graph Neural Networks Previous neural networks, e.g. fully-connected neural networks (FNNs), CNNs, and RNNs, could only be applied to Euclidean data (i.e. images, text, and videos). As a type of neural network which directly operates on a graph structure, GNNs have the ability to capture complex relationships between objects and make inferences based on data described by graphs. GNNs have been proven effective in various node-level, edge-level, and graph-level prediction tasks. As mentioned in Section 2, GNNs are currently considered the state-of- the-art techniques for traffic forecasting problems. GNNs can be divided into four types, namely, recurrent GNNs, convolutional GNNs, graph autoencoders, and spatiotemporal GNNs [Wu et al., 2020b]. Because traffic forecasting is a spatiotemporal problem, the GNNs used in this field can all be categorized as the latter. However, certain components of the other types of GNNs have also been applied in the surveyed traffic forecasting studies. Spatiotemporal GNNs can be further categorized based on the approach used to capture the temporal dependency in particular. Most of the relevant studies in the literature can be split into two types, namely, RNN-based and CNN-based spatiotemporal GNNs [Wu et al., 2020b]. The RNN-based approach is used in Li et al. [2018b], Guo et al. [2020a], Pan et al. [2020, 2019], Lu et al. [2019a], Mallick et al. [2020], Zhang et al. [2020j, l], Bai et al. [2020], Huang et al. [2020a], Wang et al. [2018b, 2020g], Lv et al. [2020], Fukuda et al. [2020], Zhang & Guo [2020], Boukerche & Wang [2020b], Kang et al. [2019], Li et al. [2019], Xu et al. [2019], Wu et al. [2018a], Wei & Sheng [2020], Li et al. [2020f], Yu et al. [2019b], Yin et al. [2020], Xin et al. [2020], Qu et al. [2020], Huang et al. [2020b], Guo et al. [2020b], Fang et al. [2020a], Li & Zhu [2021], Chen et al. [2020c], Ramadan et al. [2020], Zhou et al. [2020d], Wang et al. [2018a], Peng et al. [2020], Zhou et al. [2019], Wang et al. [2020e], Qiu et al. [2020], Shi et al. [2020], Wang et al. [2020h, 2019], Zhang et al. [2020b], Liu et al. [2020b], Ye et al. [2020b], Zhu et al. [2019], Chai et al. [2018], He et al. [2020], Bai et al. [2021], Zhang et al. [2018a, 2019f], Xie et al. [2019], Zhang et al. [2019a], Guo et al. [2019a], Cirstea et al. [2019], Lu et al. [2019b], Zhang et al. [2019b], Lu et al. [2020b], Zhao et al. [2019], Cui et al. [2019], Chen et al. [2019], Zhang et al. [2019e], Bogaerts et al. [2020], Cui et al. [2020a], Zhou et al. [2020a], Mallick et al. [2021], Sun et al. [2021], Xie et al. [2020b], Zhu et al. [2020c, a], Fu et al. [2020], Chen et al. [2020a], Lewenfus et al. [2020], Zhu et al. [2020b], Liao et al. [2018], Zhao et al. [2020b], Guopeng et al. [2020], Shao et al. [2020], Shen et al. [2020], Mohanty & Pozdnukhov [2018], Mohanty et al. [2020], Hu et al. [2018], Pian & Wu [2020], Jin et al. [2020a], Geng et al. [2019a], Bai et al. [2019a], Li et al. [2020c], Ke et al. [2019], Hu et al. [2020], Xu & Li [2019], Davis et al. [2020], Chen et al. [2020h], Du et al. [2020], Wu et al. [2020a], Ye et al. [2021], Luo et al. [2020], Chen et al. [2020b], Wang et al. [2020d], Xiao et al. [2020], Guo et al. [2019b], Lin et al. [2018], Zhou et al. [2020f], Liu et al. [2020c], Zhang et al. [2020g], Yang et al. [2019], Zhang et al. [2020f], Wang et al. [2020c], Wright et al. [2019]; while the CNN-based approach is used in Wu et al. [2019], Fang et al. [2019], Zhang et al. [2020e], Xu et al. [2020b], Chen et al. [2020e], Kong et al. [2020], Tang et al. [2020b], Guo et al. [2019c], Sun et al. [2020], Yu et al. [2018], Li et al. [2020b], Wang et al. [2020a], Tian et al. [2020], Chen et al. [2020d], Zhao et al. [2020a], Zhang et al. [2020c], Ou et al. [2020], Tang et al. [2020a], Diao et al. [2019], Lee & Rhee [2019a, b], Wang et al. [2020f], Wu et al. [2020c], Guo & Yuan [2020], Zhang et al. [2020i], Feng et al. [2020], Zhang et al. [2020d], Xie et al. [2020c], Lu et al. [2020a], Maas & Bloem [2020], Li et al. [2020d], Song et al. [2020b], Dai et al. [2020], Hong et al. [2020], Zheng et al. [2020a], Zhou et al. [2020e], Yu et al. [2020b], Xu et al. [2020d], Heglund et al. [2020]. With the recent expansion of relevant studies, we add two sub-types of spatiotemporal GNNs in this survey, namely, attention-based and FNN-based. Attention mechanism is firstly proposed to memorize long source sentences in neural machine translation [Vaswani et al., 2017]. Then it is used for temporal forecasting problems. As a special case, Transformer is built entirely upon attention mechanisms, which makes it possible to access any part of a sequence regardless of its distance to the target [Xie et al., 2020d, Cai et al., 2020, Jin et al., 2020b, Li & Moura, 2020]. The attention-based approaches are used in Zheng et al. [2020b], Zhang et al. [2020a], Wang et al. [2020b], Xie et al. [2020d], Cai et al. [2020], Zhou et al. [2020b], Chen et al. [2020f], Park et al. [2020], Fang et al. [2020b], Jin et al. [2020b], Bai et al. [2019b], Li & Moura [2020], Zhang et al. [2020k], while the simpler FNN-based approach is used in Zhang et al. [2018b], Wei et al. [2019], Song et al. [2020a], Cao et al. [2020], Chen et al. [2020g], Zhang et al. [2020h], Sun et al. [2020], He & Shin [2020a], Yeghikyan et al. [2020], Ren & Xie [2019], Li et al. [2018a], Han et al. [2019], He & Shin [2020b], Zhang et al. [2019c], Ge et al. [2019a, b], Yu et al. [2020a], Ge et al. [2020], Yu et al. [2019a], Guo et al. [2020c], Agafonov [2020], Geng et al. [2019b], Qin et al. [2020b], Kim et al. [2019]. Apart from using neural networks to capture temporal dependency, other techniques that have also been combined with GNNs include autoregression [Lee et al., 2019], Markov processes [Cui et al., 2020b], and Kalman filters [Xiong et al., 2020]. Of the additional GNN components adopted in the surveyed studies, convolutional GNNs are the most popular, while recurrent GNN [Scarselli et al., 2008] and Graph Auto-Encoder (GAE) [Kipf & Welling, 2016] are used less frequently. We further categorize convolutional GNNs into the following five types: (1) Graph Convolutional Network (GCN) [Kipf & Welling, 2017], (2) Diffusion Graph Convolution (DGC) [Atwood & Towsley, 2016], (3) Message Passing Neural Network (MPNN) [Gilmer et al., 2017], (4) GraphSAGE [Hamilton et al., 2017], and (5) Graph Attention Network (GAT) [Veličković et al., 2018]. These relevant graph neural networks are listed chronologically in Figure 2. While different GNNs can be used for traffic forecasting, a general design pipeline is proposed in [Zhou et al., 2020c] and suggested for future studies as follows: 1. 1. Find graph structure. As discussed in Section IV, different traffic graphs are available. 2. 2. Specify graph type and scale. The graphs can be further classified into different types if needed, e.g., directed/undirected graphs, homogeneous/heterogeneous graphs, static/dynamic graphs. For most cases in traffic forecasting, the graphs of the same type are used in a single study. As for the graph scale, the graphs in the traffic domain are not as large as those for the social networks or academic networks with millions of nodes and edges. 3. 3. Design loss function. The training setting usually follows the supervised approach, which means the GNN-based models are firstly trained on a training set with labels and then evaluated on a test set. The forecasting task is usually designed as the node-level regression problem. Based on these considerations, the proper loss function and evaluation metrics can be chosen, e.g., Root Mean Square Error (RMSE) and Mean Absolute Error (MAE). 4. 4. Build model using computational modules. The GNNs discussed in this section are exactly those which have already been used as computational modules to build forecasting models in the surveyed studies. Figure 2: The relevant graph neural networks in this survey. GCNs are spectral-based convolutional GNNs, in which the graph convolutions are defined by introducing filters from graph signal processing. Spectral convoluted neural networking [Bruna et al., 2014] assumes that the filter is a set of learnable parameters and considers graph signals with multiple channels. The GCN is a first-order approximation of Chebyshev’s spectral CNN (ChebNet) [Defferrard et al., 2016], which approximates the filter using the Chebyshev polynomials of the diagonal matrix of eigenvalues. The alternative approach is spatial-based convolutional GNNs, in which the graph convolutions are defined by information propagation. DGC, MPNN, GraphSAGE, and GAT all follow this approach. The graph convolution is modeled as a diffusion process with a transition probability from one node to a neighboring node in DGC. An equilibrium is expected to be obtained after several rounds of information transition. The general framework followed is a message passing network, which models the graph convolutions as an information-passing process from one node to another connected node directly. To alleviate the computation problems caused by a large number of neighbors, sampling is used to obtain a fixed number of neighbors in GraphSAGE. Lastly, without using a predetermined adjacency matrix, the attention mechanism is used to learn the relative weights between two connected nodes in GAT. A full list of the GNN components used in the surveyed studies is shown in Table 4. Currently, the most widely used GNN is the GCN. However, we also notice a growing trend in the use of GAT in traffic forecasting. Table 4: GNNs in the surveyed studies. GNN | Relevant Studies ---|--- Recurrent GNN | Wang et al. [2018b, a], Lu et al. [2019b, 2020b] GAE | Xu et al. [2020a, 2019], Opolka et al. [2019], Shen et al. [2020] GCN | Wu et al. [2019], Zhang et al. [2018b], Guo et al. [2020a], Lu et al. [2019a], Zhang et al. [2020j, l], Bai et al. [2020], Fang et al. [2019], Zhang et al. [2020e], Song et al. [2020a], Xu et al. [2020b], Wang et al. [2020g], Lv et al. [2020], Boukerche & Wang [2020b], Tang et al. [2020b], Guo et al. [2019c], Li et al. [2019], Zhang et al. [2019d], Sun et al. [2020], Li et al. [2020f], Cao et al. [2020], Yu et al. [2018, 2019b], Li et al. [2020b], Chen et al. [2020g], Zhang et al. [2020a], Wang et al. [2020a], Xin et al. [2020], Qu et al. [2020], Wang et al. [2020b], Huang et al. [2020b], Guo et al. [2020b], Fang et al. [2020a], Li & Zhu [2021], Xu et al. [2020c], Chen et al. [2020c], Xiong et al. [2020], Ramadan et al. [2020], Zhou et al. [2020d], Sun et al. [2020], Peng et al. [2020], Zhou et al. [2019], Wang et al. [2020e], Qiu et al. [2020], He & Shin [2020a], Yeghikyan et al. [2020], Shi et al. [2020], Wang et al. [2020h], Ren & Xie [2019], Li et al. [2018a], Zhao et al. [2020a], Han et al. [2019], Zhang et al. [2020b, c], Liu et al. [2020b], Ye et al. [2020b], Zhu et al. [2019], Chai et al. [2018], He et al. [2020], Bai et al. [2021], Tang et al. [2020a], James [2020], Zhang et al. [2018a, 2019f], Yu & Gu [2019], Guo et al. [2019a], Diao et al. [2019], Zhang et al. [2019c], James [2019], Ge et al. [2019a, b], Zhang et al. [2019b], Lee & Rhee [2019a], Yu et al. [2020a], Ge et al. [2020], Zhao et al. [2019], Cui et al. [2019], Zhang et al. [2019e], Yu et al. [2019a], Lee & Rhee [2019b], Bogaerts et al. [2020], Cui et al. [2020b, a], Guo et al. [2020c], Cai et al. [2020], Wu et al. [2020c], Chen et al. [2020f], Jia et al. [2020], Sun et al. [2021], Xie et al. [2020b], Zhu et al. [2020c], Feng et al. [2020], Zhu et al. [2020a], Fu et al. [2020], Agafonov [2020], Chen et al. [2020a], Lu et al. [2020a], Jepsen et al. [2019, 2020], Bing et al. [2020], Lewenfus et al. [2020], Zhu et al. [2020b], Liao et al. [2018], Maas & Bloem [2020], Li et al. [2020d], Song et al. [2020b], Zhao et al. [2020b], Guopeng et al. [2020], Shao et al. [2020], Dai et al. [2020], Mohanty & Pozdnukhov [2018], Mohanty et al. [2020], Qin et al. [2020a], Han et al. [2020], Hong et al. [2020], Hu et al. [2018], Li & Axhausen [2020], Jin et al. [2020a], Geng et al. [2019b], Bai et al. [2019b], Geng et al. [2019a], Bai et al. [2019a], Ke et al. [2021], Li et al. [2020c], Ke et al. [2019], Hu et al. [2020], Zheng et al. [2020a], Davis et al. [2020], Chen et al. [2020h], Du et al. [2020], Li & Moura [2020], Ye et al. [2021], Luo et al. [2020], Chen et al. [2020b], Wang et al. [2020d], Qin et al. [2020b], Xiao et al. [2020], Yoshida et al. [2019], Guo et al. [2019b], Kim et al. [2019], Lin et al. [2018], Zhou et al. [2020e], Yu et al. [2020b], Zhang et al. [2020k], Zhou et al. [2020f], Liu et al. [2020c], Zhang et al. [2020g], Yang et al. [2019], Zhang et al. [2020f], Xu et al. [2020d], Heglund et al. [2020] DGC | Li et al. [2018b], Mallick et al. [2020], Chen et al. [2020e], Fukuda et al. [2020], Ou et al. [2020], Chen et al. [2019], Wang et al. [2020f], Zhou et al. [2020a, b], Mallick et al. [2021], Xie et al. [2020c], Kim et al. [2020], Wang et al. [2020c] MPNN | Wei et al. [2019], Xu et al. [2020b], Wang et al. [2019] GraphSAGE | Liu et al. [2020a] GAT | Zheng et al. [2020b], Pan et al. [2020, 2019], Huang et al. [2020a], Kong et al. [2020], Zhang & Guo [2020], Tang et al. [2020b], Kang et al. [2019], Wu et al. [2018a], Wei & Sheng [2020], Yin et al. [2020], Xie et al. [2020d], Zhang et al. [2020h], Tian et al. [2020], He & Shin [2020b], Tang et al. [2020a], Zhang et al. [2019a], Cirstea et al. [2019], Yang et al. [2020], Guo & Yuan [2020], Zhang et al. [2020i, d], Park et al. [2020], Song et al. [2020b], Fang et al. [2020b], Pian & Wu [2020], Jin et al. [2020b], Xu & Li [2019], Wu et al. [2020a], Wright et al. [2019] ## 5 Open Data and Source Resources In this section, we summarize the open data and source code used in the surveyed papers. These open data are suitable for GNN-related studies with graph structures discussed in Section IV, which can be used to formulate different forecasting problems in Section III. We also list the GNN-related code resources for those who want to replicate the previous GNN-based solutions as baselines in the follow-up studies. ### 5.1 Open Data We categorize the data used in the surveyed studies into three major types, namely, graph-related data, historical traffic data, and external data. Graph- related data refer to those data which exhibit a graph structure in the traffic domain, i.e., transportation network data. Historical traffic data refer to those data which record the historical traffic states, usually in different locations and time points. We further categorize the historical traffic data into sub-types as follows. External data refer to the factors that would affect the traffic states, i.e., weather data and calendar data. Some of these data can be used in the graph-based modeling directly, while the others may require some pre-processing steps before being Incorporated into GNN-based models. Transportation Network Data. These data represent the underlying transportation infrastructure, e.g., road, subway, and bus networks. They can be obtained from government transportation departments or extracted from online map services, e.g., OpenStreetMap. Based on their topology structure, these data can be used to build the graphs directly, e.g., the road segments or the stations are nodes and the road intersections or subway links are the edges. While this modeling approach is straightforward, the disadvantage is that only static graphs can be built from transportation network data. Traffic Sensor Data. Traffic sensors, e.g. loop detectors, are installed on roads to collect traffic information, e.g., traffic volume or speed. This type of data is widely used for traffic prediction, especially road traffic flow and speed prediction problems. For graph-based modeling, each sensor can be used as a node, with road connections as the edges. One advantage of using traffic sensor data for graph-based modeling is that the captured traffic information can be used directly as the node attributes, with little pre- processing overhead. One exception is that the sensors are prone to hardware faults, which causes the missing data or data noise problems and requires corresponding pre-processing techniques, e.g., data imputation and denoising methods. Another disadvantage of using traffic sensor data for graph-based modeling is that the traffic sensors can only be installed in a limited number of locations for a series of reasons, e.g., installation cost. With this constraint, only the part of the road networks with traffic sensors can be incorporated into a graph, while the uncovered areas are neglected. GPS Trajectory Data. Different types of vehicles (e.g. taxis, buses, online ride-hailing vehicles, and shared bikes) can be equipped with GPS receivers, which record GPS coordinates in 2-60 second intervals. The trajectory data calculated from these GPS coordinate samples can be matched to road networks and further used to derive traffic flow or speed. The advantage of using GPS trajectory data for graph-based modeling is both the low expense to collect GPS data with smartphones and the wider coverage with the massive number of vehicles, compared with traffic sensor data. However, GPS trajectory data contain no direct traffic information, which can be derived with corresponding definitions though. The data quality problems also remain with GPS trajectory data and more pre-processing steps are required, e.g., map matching. Location-based Service Data. GPS function is also embedded in smartphones, which can be used to collect various types of location-related data, e.g., check-in data, point-of-interest data, and route navigation application data. The pros and cons of using location-based service data are similar with GPS trajectory data. And the difference is that location-based service data are often collected in a crowd-sourced approach, with more data providers but potentially a lower data quality. Trip Record Data. These include departure and arrival dates/times, departure and arrival locations, and other trip information. Traffic speed and demand can derived from trip record data from various sources, e.g., taxis, ride- hailing services, buses, bikes, or even dock-less e-scooters used in He & Shin [2020a]. These data can be collected in public transportation systems with mature methods, for example, by AFC (Automatic Fare Collection) in the subway and bus systems. Trip record data have the advantage of being capable of constructing multiple graph-based problems, e.g., station-level traffic flow and demand problems. They are also easier to collect in existing public transportation systems. Traffic Report Data. This type of data is often used for abnormal cases, e.g., anomaly report data used in Liu et al. [2020c] and traffic accident report data used in Zhou et al. [2020e], Zhang et al. [2020k], Zhou et al. [2020f]. Traffic report data are less used in graph-based modeling because of their sparsity in both spatial and temporal dimensions, compared with trip record data. Multimedia Data. This type of data can be used as an additional input to deep learning models or for verifying the traffic status indicated by other data sources. Multimedia data used in the surveyed studies include the Baidu street-view images used in Qin et al. [2020a] for traffic congestion, as well as satellite imagery data [Zhang et al., 2020k], and video surveillance data [Shao et al., 2020]. Multimedia data are also less seen in graph-based modeling because of their higher requirement for data collection, transmission and storage, compared with traffic sensor data with similar functionalities. It is also more difficult to extract precise traffic information, e.g., vehicle counts, from images or videos through image processing and object detection techniques. Simulated Traffic Data. In addition to observed real-world datasets, microscopic traffic simulators are also used to build virtual training and testing datasets for deep learning models. Examples in the surveyed studies include the MATES Simulator used in Fukuda et al. [2020] and INTEGRATION software used in Ramadan et al. [2020]. With many real-world datasets available, simulated traffic data are rarely used in GNN-based and more broader ML-based traffic forecasting studies. Traffic simulations have the potential of modeling unseen graphs though, e.g., evaluating a planned road topology. Weather Data. Traffic states are highly affected by the meteorological factors including temperature, humidity, precipitation, barometer pressure, and wind strength. Calendar Data. This includes the information on weekends and holidays. Because traffic patterns vary significantly between weekdays and weekends/holidays, some studies consider these two cases separately. Both weather and calendar data have been proven useful for traffic forecasting in the literature and should not be neglected in graph-based modeling as external factors. While present road network and weather data can be easily found on the Internet, it is much more difficult to source historical traffic data, both due to data privacy concerns and the transmission and storage requirements of large data volumes. In Table 5 we present a list of the open data resources used in the surveyed studies. Most of these open data are already cleaned or preprocessed and can be readily used for benchmarking and comparing the performance of different models in future work. Table 5: Open data for traffic prediction problems. Dataset Name | Relevant Studies ---|--- METR-LA | Li et al. [2018b], Wu et al. [2019], Xu et al. [2020a], Pan et al. [2020, 2019], Lu et al. [2019a], Zhang et al. [2020e], Wang et al. [2020g], Zhang & Guo [2020], Boukerche & Wang [2020b], Cao et al. [2020], Yu et al. [2019b], Li & Zhu [2021], Tian et al. [2020], Chen et al. [2020d], Bai et al. [2021], Zhang et al. [2018a], Cirstea et al. [2019], Shleifer et al. [2019], Yang et al. [2020], Chen et al. [2019], Wang et al. [2020f], Cui et al. [2020b], Zhou et al. [2020a], Cai et al. [2020], Zhou et al. [2020b], Wu et al. [2020c], Chen et al. [2020f], Opolka et al. [2019], Oreshkin et al. [2021], Jia et al. [2020], Zhang et al. [2020i], Feng et al. [2020], Xie et al. [2020c], Park et al. [2020], Song et al. [2020b] PeMS all | Mallick et al. [2020, 2021] PeMS-BAY | Li et al. [2018b], Wu et al. [2019], Zheng et al. [2020b], Pan et al. [2020, 2019], Xu et al. [2020b], Wang et al. [2020g], Zhang & Guo [2020], Boukerche & Wang [2020b], Li et al. [2020f], Cao et al. [2020], Xie et al. [2020d], Li & Zhu [2021], Tian et al. [2020], Shleifer et al. [2019], Chen et al. [2019], Yu et al. [2019a], Wang et al. [2020f], Cui et al. [2020b], Zhou et al. [2020a], Cai et al. [2020], Zhou et al. [2020b], Wu et al. [2020c], Chen et al. [2020f], Oreshkin et al. [2021], Guo & Yuan [2020], Zhang et al. [2020i], Feng et al. [2020], Xie et al. [2020c], Park et al. [2020], Song et al. [2020b] PeMSD3 | Song et al. [2020a], Cao et al. [2020], Chen et al. [2020g], Wang et al. [2020a], Li & Zhu [2021] PeMSD4 | Bai et al. [2020], Huang et al. [2020a], Zhang et al. [2020e], Song et al. [2020a], Chen et al. [2020e], Tang et al. [2020b], Guo et al. [2019c], Li et al. [2019], Wei & Sheng [2020], Cao et al. [2020], Li et al. [2020b], Yin et al. [2020], Zhang et al. [2020a], Wang et al. [2020a], Xin et al. [2020], Huang et al. [2020b], Guo et al. [2020b], Li & Zhu [2021], Xu et al. [2020c], Chen et al. [2020c], Ge et al. [2019a, b, 2020], Zhao et al. [2020b] PeMSD7 | Zhang et al. [2020j], Huang et al. [2020a], Song et al. [2020a], Xu et al. [2020b], Tang et al. [2020b], Sun et al. [2020], Cao et al. [2020], Yu et al. [2018, 2019b], Chen et al. [2020g], Wang et al. [2020a], Xin et al. [2020], Xie et al. [2020d], Li & Zhu [2021], Zhang et al. [2019a], Ge et al. [2019a, b, 2020], Yu et al. [2019a], Zhao et al. [2020b] PeMSD8 | Bai et al. [2020], Huang et al. [2020a], Song et al. [2020a], Chen et al. [2020e], Guo et al. [2019c], Wei & Sheng [2020], Cao et al. [2020], Li et al. [2020b], Yin et al. [2020], Zhang et al. [2020a], Wang et al. [2020a], Guo et al. [2020b], Li & Zhu [2021] Seattle Loop | Cui et al. [2019, 2020a], Sun et al. [2021], Lewenfus et al. [2020] T-Drive | Pan et al. [2020, 2019] SHSpeed | Zhang et al. [2020j], Wang et al. [2018b], Guo et al. [2019a] TaxiBJ | Zhang et al. [2020h], Wang et al. [2018a], Bai et al. [2019b] TaxiSZ | Bai et al. [2021], Zhao et al. [2019] TaxiCD | Hu et al. [2018, 2020] TaxiNYC | Zhang et al. [2020h], Sun et al. [2020], Zhou et al. [2019], Hu et al. [2018], Jin et al. [2020b], Li & Axhausen [2020], Zheng et al. [2020a], Xu & Li [2019], Davis et al. [2020], Du et al. [2020], Li & Moura [2020], Ye et al. [2021], Zhou et al. [2020f] UberNYC | Jin et al. [2020b], Ke et al. [2021] DiDiChengdu | Zhang et al. [2019d], Qu et al. [2020], Wang et al. [2020b], Zhou et al. [2019], Wang et al. [2020h], Bogaerts et al. [2020], Li et al. [2020c] DiDiTTIChengdu | Lu et al. [2020a] DiDiXi’an | Qu et al. [2020], Bogaerts et al. [2020] DiDiHaikou | Pian & Wu [2020], Jin et al. [2020a] BikeDC | Sun et al. [2020], Wang et al. [2020d] BikeNYC | Zhang et al. [2020h], Sun et al. [2020], Wang et al. [2018a], He & Shin [2020b], Chai et al. [2018], Lee et al. [2019], Bai et al. [2019b], Du et al. [2020], Ye et al. [2021], Wang et al. [2020d], Guo et al. [2019b], Lin et al. [2018] BikeChicago | Chai et al. [2018] SHMetro | Liu et al. [2020b] HZMetro | Liu et al. [2020b] #### 5.1.1 Traffic Sensor Data The relevant open traffic sensor data are listed as follows. METR-LA 222Download link: https://github.com/liyaguang/DCRNN: This dataset contains traffic speed and volume collected from the highway of the Los Angeles County road network, with 207 loop detectors. The samples are aggregated in 5-minute intervals. The most frequently referenced time period for this dataset is from March 1st to June 30th, 2012. Performance Measurement System (PeMS) Data 333http://pems.dot.ca.gov/: This dataset contains raw detector data from over 18,000 vehicle detector stations on the freeway system spanning all major metropolitan areas of California from 2001 to 2019, collected with various sensors including inductive loops, side- fire radar, and magnetometers. The samples are captured every 30 seconds and aggregated in 5-minute intervals. Each data sample contains a timestamp, station ID, district, freeway ID, direction of travel, total flow, and average speed. Different subsets of PeMS data have been used in previous studies, for example: * 1. PeMS-BAY 444Download link: https://github.com/liyaguang/DCRNN: This subset contains data from 325 sensors in the Bay Area from January 1st to June 30th, 2017. * 2. PeMSD3: This subset uses 358 sensors in the North Central Area. The frequently referenced time period for this dataset is September 1st to November 30th, 2018. * 3. PeMSD4: This subset uses 307 sensors in the San Francisco Bay Area. The frequently referenced time period for this dataset is January 1st to February 28th, 2018. * 4. PeMSD7: This subset uses 883 sensors in the Los Angeles Area. The frequently referenced time period for this dataset is May to June, 2012. * 5. PeMSD8: This subset uses 170 sensors in the San Bernardino Area. The frequently referenced time period for this dataset is July to August, 2016. Seattle Loop 555Download link: https://github.com/zhiyongc/Seattle-Loop-Data: This dataset was collected by inductive loop detectors deployed on four connected freeways (I-5, I-405, I-90, and SR-520) in the Seattle area, from January 1st to 31st, 2015. It contains the traffic speed data from 323 detectors. The samples are aggregated in 5-minute intervals. #### 5.1.2 Taxi Data The open taxi datasets used in the surveyed studies are listed as follows. T-drive [Yuan et al., 2010]: This dataset contains a large number of taxicab trajectories collected by 30,000 taxis in Beijing from February 1st to June 2nd, 2015. SHSpeed (Shanghai Traffic Speed) [Wang et al., 2018b] 666Download link: https://github.com/xxArbiter/grnn: This dataset contains 10-minute traffic speed data, derived from raw taxi trajectory data, collected from 1 to 30 April 2015, for 156 urban road segments in the central area of Shanghai, China. TaxiBJ [Zhang et al., 2017]: This dataset contains inflow and outflow data derived from GPS data in more than 34,000 taxicabs in Beijing from four time intervals: (1) July 1st to October 30th, 2013; (2) March 1st to June 30th, 2014; (3) March 1st to June 30th, 2015; and (4) November 1st, 2015 to April 10th, 2016. The Beijing city map is divided into $32\times 32$ grids and the time interval of the flow data is 30 minutes. TaxiSZ [Zhao et al., 2019] 777Download link: https://github.com/lehaifeng/T-GCN: This dataset is derived from taxi trajectories in Shenzhen from January 1st to 31st, 2015. It contains the traffic speed on 156 major roads of the Luohu District every 15 minutes. TaxiCD 888 https://js.dclab.run/v2/cmptDetail.html?id=175: This dataset contains 1.4 billion GPS records from 14,864 taxis collected from August 3rd to 30th, 2014 in Chengdu, China. Each GPS record consists of a taxi ID, latitude, longitude, an indicator of whether the taxi is occupied, and a timestamp. TaxiNYC999http://www.nyc.gov/html/tlc/html/about/trip_record_data.shtml: The taxi trip records in New York starting from 2009, in both yellow and green taxis. Each trip record contains pick-up and drop-off dates/times, pick-up and drop-off locations, trip distances, itemized fares, rate types, payment types, and driver-reported passenger counts. #### 5.1.3 Ride-hailing Data The open ride-hailing data used in the surveyed studies are listed as follows. UberNYC 101010https://github.com/fivethirtyeight/uber-tlc-foil-response: This dataset comes from Uber, which is one of the largest online ride-hailing companies in the USA, and is provided by the NYC Taxi & Limousine Commission (TLC). It contains data from over 4.5 million Uber pickups in New York City from April to September 2014, and 14.3 million more Uber pickups from January to June 2015. Didi GAIA Open Data 111111https://outreach.didichuxing.com/research/opendata/: This open data plan is supported by Didi Chuxing, which is one of the largest online ride-hailing companies in China. * 1. DiDiChengdu: This dataset contains the trajectories of DiDi Express and DiDi Premier drivers within Chengdu, China. The data contains trips from October to November 2016. * 2. DiDiTTIChengdu: This dataset represents the DiDi Travel Time Index Data in Chengdu, China in the year of 2018, which contains the average speed of major roads every 10 minutes. * 3. DiDiXi’an: This dataset contains the trajectories of DiDi Express and DiDi Premier drivers within Xi’an, China. The data contains trips from October to November 2016. * 4. DiDiHaikou: The dataset contains DiDi Express and DiDi Premier orders from May 1st to October 31st, 2017 in the city of Haikou, China, including the coordinates of origins and destinations, pickup and drop-off timestamps, as well as other information. #### 5.1.4 Bike Data The open bike data used in the surveyed studies are listed as follows. BikeNYC 121212 https://www.citibikenyc.com/system-data: This dataset is from the NYC Bike System, which contains 416 stations. The frequently referenced time period for this dataset is from 1st July, 2013 to 31th December, 2016. BikeDC 131313 https://www.capitalbikeshare.com/system-data: This dataset is from the Washington D.C. Bike System, which contains 472 stations. Each record contains trip duration, start and end station IDs, and start and end times. BikeChicago 141414https://www.divvybikes.com/system-data: This dataset is from the Divvy System Data in Chicago, from 2015 to 2020. #### 5.1.5 Subway Data The subway data referenced in the surveyed studies are listed as follows. SHMetro [Liu et al., 2020b] 151515Download link: https://github.com/ivechan/PVCGN: This dataset is derived from 811.8 million transaction records of the Shanghai metro system collected from July 1st to September 30th, 2016. It contains 288 metro stations and 958 physical edges. The inflow and outflow of each station are provided in 15 minute intervals. HZMetro [Liu et al., 2020b] 161616Download link: https://github.com/ivechan/PVCGN: This dataset is similar to SHMetro, from the metro system in Hangzhou, China, in January 2019. It contains 80 metro stations and 248 physical edges, and the aggregation time length is also 15 minutes. ### 5.2 Open Source Codes Several open source frameworks for implementing general deep learning models, most of which are built with the Python programming language, can be accessed online, e.g. TensorFlow 171717https://www.tensorflow.org/, Keras 181818https://keras.io/, PyTorch 191919https://pytorch.org/, and MXNet 202020https://mxnet.apache.org/. Additional Python libraries designed for implementing GNNs are available. These include DGL 212121https://www.dgl.ai/, pytorch_geometric 222222https://pytorch-geometric.readthedocs.io/, and Graph Nets 232323https://github.com/deepmind/graph_nets. Many authors have also released open-source implementations of their proposed models. The open source projects for traffic flow, traffic speed, traffic demand, and other problems are summarized in Tables 6, 7, 8, and 9, respectively. In these open source projects, TensorFlow and PyTorch are the two frameworks that are used most frequently. Table 6: Open source projects for traffic flow related problems. Article | Year | Framework | Problem | Link ---|---|---|---|--- Zheng et al. [2020b] | 2020 | TensorFlow | Road Traffic Flow, Road Traffic Speed | https://github.com/zhengchuanpan/GMAN Bai et al. [2020] | 2020 | PyTorch | Road Traffic Flow | https://github.com/LeiBAI/AGCRN Song et al. [2020a] | 2020 | MXNet | Road Traffic Flow | https://github.com/wanhuaiyu/STSGCN Tang et al. [2020b] | 2020 | TensorFlow | Road Traffic Flow | https://github.com/sam101340/GAGCN-BC-20200720 Wang et al. [2020a] | 2020 | MXNet, PyTorch | Road Traffic Flow | https://github.com/zkx741481546/Auto-STGCN Guo et al. [2020b] | 2020 | PyTorch | Road Traffic Flow, Road Traffic Speed | https://github.com/guokan987/DGCN Li & Zhu [2021] | 2020 | MXNet | Road Traffic Flow, Road Traffic Speed | https://github.com/MengzhangLI/STFGNN Tian et al. [2020] | 2020 | PyTorch, DGL | Road Traffic Flow | https://github.com/Kelang-Tian/ST-MGAT Xiong et al. [2020] | 2020 | TensorFlow | Road OD Flow | https://github.com/alzmxx/OD_Prediction Peng et al. [2020] | 2020 | Keras | Road Station-level Subway Passenger Flow, Station-level Bus Passenger Flow, Regional Taxi Flow | https://github.com/RingBDStack/GCNN-In-Traffic Qiu et al. [2020] | 2020 | Pytorch | Regional Taxi Flow | https://github.com/Stanislas0/ToGCN-V2X Yeghikyan et al. [2020] | 2020 | PyTorch | Regional OD Taxi Flow | https://github.com/FelixOpolka/Mobility-Flows-Neural-Networks Zhang et al. [2020b] | 2020 | Keras | Station-level Subway Passenger Flow | https://github.com/JinleiZhangBJTU/ResNet-LSTM-GCN Zhang et al. [2020c] | 2020 | Keras | Station-level Subway Passenger Flow | https://github.com/JinleiZhangBJTU/Conv-GCN Liu et al. [2020b] | 2020 | PyTorch | Station-level Subway Passenger Flow | https://github.com/ivechan/PVCGN Ye et al. [2020b] | 2020 | Keras | Station-level Subway Passenger Flow | https://github.com/start2020/Multi-STGCnet Pan et al. [2019] | 2019 | MXNet, DGL | Road Traffic Flow, Road Traffic Speed | https://github.com/panzheyi/ST-MetaNet Guo et al. [2019c] | 2019 | MXNet | Road Traffic Flow | https://github.com/wanhuaiyu/ASTGCN Guo et al. [2019c] | 2019 | PyTorch | Road Traffic Flow | https://github.com/wanhuaiyu/ASTGCN-r-pytorch Wang et al. [2018b] | 2018 | PyTorch | Road Traffic Flow | https://github.com/xxArbiter/grnn Yu et al. [2018] | 2018 | TensorFlow | Road Traffic Flow | https://github.com/VeritasYin/STGCN_IJCAI-18 Li et al. [2018a] | 2018 | Keras | Station-level Subway Passenger Flow | https://github.com/RingBDStack/GCNN-In-Traffic Chai et al. [2018] | 2018 | TensorFlow | Bike Flow | https://github.com/Di-Chai/GraphCNN-Bike Table 7: Open source projects for traffic speed related problems. Article | Year | Framework | Problem | Link ---|---|---|---|--- Zhang et al. [2020h] | 2020 | Keras | Road Traffic Speed | https://github.com/jillbetty001/ST-CGA Bai et al. [2021] | 2020 | TensorFlow | Road Traffic Speed | https://github.com/lehaifeng/T-GCN/tree/master/A3T Yang et al. [2020] | 2020 | TensorFlow | Road Traffic Speed | https://github.com/fanyang01/relational-ssm Wu et al. [2020c] | 2020 | PyTorch | Road Traffic Speed | https://github.com/nnzhan/MTGNN Mallick et al. [2021] | 2020 | TensorFlow | Road Traffic Speed | https://github.com/tanwimallick/TL-DCRNN Chen et al. [2020a] | 2020 | PyTorch | Road Traffic Speed | https://github.com/Fanglanc/DKFN Lu et al. [2020a] | 2020 | PyTorch | Road Traffic Speed | https://github.com/RobinLu1209/STAG-GCN Guopeng et al. [2020] | 2020 | TensorFlow, Keras | Road Traffic Speed | https://github.com/RomainLITUD/DGCN_traffic_forecasting Shen et al. [2020] | 2020 | PyTorch | Road Travel Time | https://github.com/YibinShen/TTPNet Hong et al. [2020] | 2020 | TensorFlow | Time of Arrival | https://github.com/didi/heteta Wu et al. [2019] | 2019 | PyTorch | Road Traffic Speed | https://github.com/nnzhan/Graph-WaveNet Shleifer et al. [2019] | 2019 | PyTorch | Road Traffic Speed | https://github.com/sshleifer/Graph-WaveNet Zhao et al. [2019] | 2019 | TensorFlow | Road Traffic Speed | https://github.com/lehaifeng/T-GCN Cui et al. [2019] | 2019 | TensorFlow | Road Traffic Speed | https://github.com/zhiyongc/Graph_Convolutional_LSTM Jepsen et al. [2019, 2020] | 2019 | MXNet | Road Traffic Speed | https://github.com/TobiasSkovgaardJepsen/relational-fusion-networks Li et al. [2018b] | 2018 | TensorFlow | Road Traffic Speed | https://github.com/liyaguang/DCRNN Li et al. [2018b] | 2018 | PyTorch | Road Traffic Speed | https://github.com/chnsh/DCRNN_PyTorch Zhang et al. [2018a] | 2018 | MXNet | Road Traffic Speed | https://github.com/jennyzhang0215/GaAN Liao et al. [2018] | 2018 | TensorFlow | Road Traffic Speed | https://github.com/JingqingZ/BaiduTraffic Mohanty & Pozdnukhov [2018], Mohanty et al. [2020] | 2018 | TensorFlow | Traffic Congestion | https://github.com/sudatta0993/Dynamic-Congestion-Prediction Table 8: Open source projects for traffic demand related problems. Article | Year | Framework | Problem | Link ---|---|---|---|--- Hu et al. [2020] | 2020 | TensorFlow | Taxi Demand | https://github.com/hujilin1229/od-pred Davis et al. [2020] | 2020 | TensorFlow, PyTorch | Taxi Demand | https://github.com/NDavisK/Grids-versus-Graphs Ye et al. [2021] | 2020 | PyTorch | Taxi Demand, Bike Demand | https://github.com/Essaim/CGCDemandPrediction Lee et al. [2019] | 2019 | TensorFlow, Keras | Ride-hailing Demand, Bike Demand, Taxi Demand | https://github.com/LeeDoYup/TGGNet-keras Ke et al. [2019] | 2019 | Keras | Taxi Demand | https://github.com/kejintao/ST-ED-RMGC Table 9: Open source projects for other problems. Article | Year | Framework | Problem | Link ---|---|---|---|--- Zhou et al. [2020e] | 2020 | TensorFlow | Traffic Accident | https://github.com/zzyy0929/AAAI2020-RiskOracle/ Yu et al. [2020b] | 2020 | PyTorch, DGL | Traffic Accident | https://github.com/yule-BUAA/DSTGCN Zhang et al. [2020f] | 2020 | PyTorch, DGL | Parking Availability | https://github.com/Vvrep/SHARE-parking_availability_prediction-Pytorch Wang et al. [2020c] | 2020 | TensorFlow | Transportation Resilience | https://github.com/Charles117/resilience_shenzhen Wright et al. [2019] | 2019 | TensorFlow, Keras | Lane Occupancy | https://github.com/mawright/trafficgraphnn ## 6 Challenges and Future Directions In this section, we discuss general challenges for traffic prediction problems as well as specific new challenges when GNNs are involved. While GNNs achieve a better forecasting performance, they are not the panacea. Some existing challenges from the border topic of traffic forecasting remain unsolved in current graph-based studies. Based on these challenges, we discuss possible future directions as well as early attempts in these directions. Some of these future directions are inspired from the border traffic forecasting research and remain insightful for the graph-based modeling approach. We would also highlight the special opportunities with GNNs. ### 6.1 Challenges #### 6.1.1 Heterogeneous Data Traffic prediction problems involve both spatiotemporal data and external factors, e.g., weather and calendar information. Heterogeneous data fusion is a challenge that is not limited to the traffic domain. GNNs have enabled significant progress by taking the underlying graph structures into consideration. However, some challenges remain; for example, geographically close nodes may not be the most influential, both for CNN-based and GNN-based approaches. Another special challenge for GNNs is that the underlying graph information may not be correct or up to date. For example, the road topology data of OpenStreetMap, an online map services, are collected in a crowd- sourced approach, which may be inaccurate or lagged behind the real road network. The spatial dependency relationship extracted by GNNs with these inaccurate data may decrease the forecasting accuracy. Data quality concerns present an additional challenge with problems such as missing data, sparse data and noise potentially compromising forecasting results. Most of the surveyed models are only evaluated with processed high- quality datasets. A few studies do, however, take data quality related problems into consideration, e.g., using the Kalman filter to deal with the sensor data bias and noise [Chen et al., 2020a], infilling missing data with moving average filters [Hasanzadeh et al., 2019] or linear interpolation [Agafonov, 2020, Chen et al., 2020a]. Missing data problem could be more common in GNNs, with the potential missing phenomena happening with historical traffic data or underlying graph information, e.g., GCNs are proposed to fill data gaps in missing OD flow problems [Yao et al., 2020]. Traffic anomalies (e.g., congestion) are an important external factor that may affect prediction accuracy and it has been proven that under congested traffic conditions a deep neural network may not perform as well as under normal traffic conditions [Mena-Oreja & Gozalvez, 2020]. However, it remains a challenge to collect enough anomaly data to train deep learning models (including GNNs) in both normal and anomalous situations. The same concern applies for social events, public holidays, etc. #### 6.1.2 Multi-task Performance For the public service operation of ITSs, a multi-task framework is necessary to incorporate all the traffic information and predict the demand of multiple transportation modes simultaneously. For example, knowledge adaption is proposed to adapt the relevant knowledge from an information-intensive source to information-sparse sources for demand prediction [Li et al., 2020a]. Related challenges lie in data format incompatibilities as well as the inherent differences in spatial or temporal patterns. While some of the surveyed models can be used for multiple tasks, e.g., traffic flow and traffic speed prediction on the same road segment, most can only be trained for a single task at one time. Multi-task forecasting is a bigger challenge in graph-based modeling because different tasks may use different graph structures, e.g., road-level and station-level problems use different graphs and thus are difficult to be solved with a single GNN model. Some efforts that have been made in GNN-based models for multi-task prediction include taxi departure flow and arrival flow [Chen et al., 2020h], region-flow and transition-flow [Wang et al., 2020b], crowd flows, and OD of the flows [Wang et al., 2020e]. However, most of the existing attempts are based on the same graph with multiple outputs generated by feed forward layers. Nonetheless, GNN-based multi-task prediction for different types of traffic forecasting problems is a research direction requiring significant further development, especially those requiring multiple graph structures. #### 6.1.3 Practical Implementation A number of challenges prevent the practical implementation of the models developed in the surveyed studies in city-scale ITSs. First, there is significant bias introduced by the small amount of data considered in the existing GNN-based studies which, in most cases, spans less than one year. The proposed solutions are therefore not necessarily applicable to different time periods or different places. If longer traffic data are to be used in GNNs, the corresponding change of the underlying traffic infrastructures should be recorded and updated, which increases both the expense and difficulty of the associated data collection process in practice. A second challenge is the computation scalability of GNNs. To avoid the huge computation requirements of the large-scale real-world traffic network graphs, only a subset of the nodes and edges are typically considered. For example, most studies only use a subset of the PeMS dataset when considering the road traffic flow or speed problems. Their results can therefore only be applied to the selected subsets. Graph partitioning and parallel computing infrastructures have been proposed for solving this problem. The traffic speed and flow of the entire PeMS dataset with 11,160 traffic sensor locations are predicted simultaneously in Mallick et al. [2020], using a graph-partitioning method that decomposes a large highway network into smaller networks and trains a single DCRNN model on a cluster with graphics processing units (GPUs). However, increased modeling power can only improve the state-of-the- art results with narrow performance margins, compared to statistical and machine learning models with less complex structures and computational requirements. A third challenge is presented by changes in the transportation networks and infrastructure, which are essential to build the graphs in GNNs. The real- world network graphs change when road segments or bus lines are added or removed. Points-of-interest in a city also change when new facilities are built. Static graph formulations are not enough for handling these situations. Some efforts have been made to solve this problem with promising results. For example, a dynamic Laplacian matrix estimator is proposed to find the change of Laplacian matrix, according to changes in spatial dependencies hidden in the traffic data [Diao et al., 2019], and a Data Adaptive Graph Generation (DAGG) module is proposed to infer the inter-dependencies between different traffic series automatically, without using pre-defined graphs based on spatial connections [Bai et al., 2020]. #### 6.1.4 Model Interpretation The challenge of model interpretation is a point of criticism for all “black- box” machine learning or deep learning models, and traffic forecasting tasks are no exception [Wu et al., 2018b, Barredo-Arrieta et al., 2019]. while there have been remarkable progresses for visualizing and explaining other deep neural network structures, e.g., CNNs, the development of post-processing techniques to explain the predictions made by GNNs is still in an early phase [Baldassarre & Azizpour, 2019, Pope et al., 2019, Ying et al., 2019] and the application of these techniques to the traffic forecasting domain has not yet been addressed. ### 6.2 Future Directions #### 6.2.1 Centralized Data Repository A centralized data repository for GNN-based traffic forecasting resources would facilitate objective comparison of the performance of different models and be an invaluable contribution to the field. This future direction is proposed for the challenge of heterogeneous data as well as the data quality problem. Another unique feature of this repository could be the inclusion of graph-related data, which have not be provided directly in previous traffic forecasting studies. Some criteria for building such data repositories, e.g. a unified data format, tracking of dataset versions, public code and ranked results, and sufficient record lengths (longer than a year ideally), have been discussed in previous surveys [Manibardo et al., 2020]. Compiling a centralized and standardized data repository is particularly challenging for GNN-based models where natural graphs are collected and stored in a variety of data formats (e.g. Esri Shapefile and OSM XML used by Openstreetmap are used for digital maps in the GIS community) and various different similarity graphs can be constructed from the same traffic data in different models. Some previous attempts in this direction have been made in the machine learning community, e.g. setting benchmarks for several traffic prediction tasks in Papers With Code 242424https://paperswithcode.com/task/traffic- prediction, and in data science competitions, e.g., the Traffic4cast competition series 252525https://www.iarai.ac.at/traffic4cast/. However, the realization of a centralized data repository remains an open challenge. #### 6.2.2 Combination with Other Techniques GNNs may be combined with other advanced techniques to overcome some of their inherent challenges and achieve better performance. Data Augmentation. Data augmentation has been proven effective for boosting the performance of deep learning models, e.g. in image classification tasks and time series prediction tasks. Data augmentation is proposed for the challenge of the possible forecasting bias introduced by the small amount of available data. However, due to the complex structure of graphs, it is more challenging to apply data augmentation techniques to GNNs. Recently, data augmentation for GNNs has proven helpful in semi-supervised node classification tasks [Zhao et al., 2021]. However, it remains a question whether data augmentation may be effective in traffic forecasting GNN applications. Transfer Learning. Transfer learning utilizes knowledge or models trained for one task to solve related tasks, especially those with limited data. In the image classification field, pre-trained deep learning models from the ImageNet or MS COCO datasets are widely used in other problems. In traffic prediction problems, where a lack of historical data is a frequent problem, transfer learning is a possible solution. For GNNs, transfer learning can be used from a graph with more historical traffic data for the model training process to another graph with less available data. Transfer learning can also be used for the challenge caused by the changes in the transportation networks and infrastructure, when new stations or regions have not accumulated enough historical traffic data to train a GNN model. A novel transfer learning approach for DCRNN is proposed in Mallick et al. [2021], so that a model trained on data-rich regions of highway network can be used to predict traffic on unseen regions of the highway network. The authors demonstrated the efficacy of model transferability between the San Francisco and Los Angeles regions using different parts of the California road network from the PeMS. Meta-learning. Meta-learning, or learning how to learn, has recently become a potential learning paradigm that can absorb information from a task and effectively generalize it to an unseen task. Meta-learning is proposed for the challenge of GNN-based multi-task prediction, especially those involving mutiple graphs. There are different types of meta learning methods and some of them are combined with graph structures for describing relationships between tasks or data samples [Garcia & Bruna, 2017, Liu et al., 2019]. Based on a deep meta learning method called network weight generation, ST-MetaNet+ is proposed in Pan et al. [2020], which leverages the meta knowledge extracted from geo-graph attributes and dynamic traffic context learned from traffic states to generate the parameter weights in graph attention networks and RNNs, so that the inherent relationships between diverse types of spatiotemporal correlations and geo-graph attributes can be captured. Generative Adversarial Network (GAN) [Goodfellow et al., 2014]. GAN is a machine learning framework that has two components, namely, a generator, which learns to generate plausible data, and a discriminator, which learns to distinguish the generator’s fake data from real data. After training to a state of Nash equilibrium, the generator may generate undistinguished data, which helps to expand the training data size for many problems, including GNN- based traffic forecasting. GAN is proposed for the challenges caused by the small data amount used in previous studies or the changes in the transportation networks and infrastructure when not enough historical traffic data are available. In Xu et al. [2020a], the road network is used directly as the graph, in which the nodes are road state detectors and the edges are built based on their adjacent links. DeepWalk is used to embed the graph and the road traffic state sensor information is transferred into a low-dimensional space. Then, the Wasserstein GAN (WGAN) [Arjovsky et al., 2017] is used to train the traffic state data distribution and generate predicted results. Both public traffic flow (i.e., Caltrans PeMSD7) and traffic speed (i.e., METR-LA) datasets are used for evaluation, and the results demonstrate the effectiveness of the GAN-based solution when used in graph-based modeling. Automated Machine Learning (AutoML). The application of machine learning requires considerable manual intervention in various aspects of the process, including feature extraction, model selection, and parameter adjustment. AutoML automatically learns the important steps related to features, models, optimization, and evaluation, so that machine learning models can be applied without manual intervention. AutoML would help to improve the implementation of machine learning models, including GNNs. AutoML is proposed for the challenge for computational requirements in graph-based modeling, in which case the hyper parameter tuning for GNNs can be more efficient with state-of- the-art AutoML techniques. An early attempt to combine AutoML with GNNs for traffic prediction problems is an Auto-STGCN algorithm, proposed in Wang et al. [2020a]. This algorithm searches the parameter space for STGCN models quickly based on reinforcement learning and generates optimal models automatically for specific scenarios. Bayesian Network. Most of the existing studies aim for deterministic models that make mean predictions. However, some traffic applications rely on uncertainty estimates for the future situations. To tackle this gap, the Bayesian network, which is a type of probabilistic graphical model using Bayesian inference for probability computations, is a promising solution. The combination of GNNs with Bayesian networks is proposed for the challenge of GNN model interpretation. With probabilistic predictions, uncertainty estimates are generated for the future situations, especially the chance of extreme traffic states. A similar alternative is Quantile Regression, which estimates the quantile function of a distribution at chosen points, combined with Graph WaveNet for uncertainty estimates [Maas & Bloem, 2020]. #### 6.2.3 Applications in Real-World ITS Systems Last but not the least, most of the surveyed GNN-based studies are only based on the simulations with historical traffic data, without being validated or deployed in real-world ITS systems. However, there are a number of potential applications, especially for GNN-based models with the better forecasting performance. To name a few potential cases, the GNN-based forecasting model can be used for traffic light control in signalized intersections, when each intersection is modeled as a node in the graph and the corresponding traffic flow forecasting result can be used to design the traffic light control strategy. Another example is the application in map service and navigation applications, in which each road segment is modeled as a node in the graph and the corresponding traffic speed and travel time forecasting result can be used to calculate the estimated time of arrival. A third example is the application in online ride-hailing service providers, e.g., Uber and Lyft, in which each region is modeled as a node and the corresponding ride-hailing demand forecasting can be used to design a more profitable vehicle dispatching and scheduling system. Inspired by these potential application scenarios, there are a lot of potential research opportunities for researchers from both the academia and the industry. ## 7 Conclusion In this paper, a comprehensive review of the application of GNNs for traffic forecasting is presented. Three levels of traffic problems and graphs are summarized, namely, road-level, region-level and station-level. The usages of recurrent GNNs, convolutional GNNs and graph autoencoders are discussed. We also give the latest collection of open dataset and code resource for this topic. Challenges and future directions are further pointed out for the follow-up research. ## References * Agafonov [2020] Agafonov, A. (2020). Traffic flow prediction using graph convolution neural networks. In 2020 10th International Conference on Information Science and Technology (ICIST) (pp. 91–95). IEEE. * Arjovsky et al. [2017] Arjovsky, M., Chintala, S., & Bottou, L. (2017). Wasserstein gan. arXiv preprint arXiv:1701.07875, . * Atwood & Towsley [2016] Atwood, J., & Towsley, D. (2016). Diffusion-convolutional neural networks. In NIPS. * Bai et al. [2021] Bai, J., Zhu, J., Song, Y., Zhao, L., Hou, Z., Du, R., & Li, H. (2021). A3t-gcn: attention temporal graph convolutional network for traffic forecasting. ISPRS International Journal of Geo-Information, 10, 485. * Bai et al. [2019a] Bai, L., Yao, L., Kanhere, S. S., Wang, X., Liu, W., & Yang, Z. (2019a). Spatio-temporal graph convolutional and recurrent networks for citywide passenger demand prediction. In Proceedings of the 28th ACM International Conference on Information and Knowledge Management (pp. 2293–2296). * Bai et al. [2019b] Bai, L., Yao, L., Kanhere, S. S., Wang, X., & Sheng, Q. Z. (2019b). Stg2seq: spatial-temporal graph to sequence model for multi-step passenger demand forecasting. In Proceedings of the 28th International Joint Conference on Artificial Intelligence (pp. 1981–1987). AAAI Press. * Bai et al. [2020] Bai, L., Yao, L., Li, C., Wang, X., & Wang, C. (2020). Adaptive graph convolutional recurrent network for traffic forecasting. In Advances in Neural Information Processing Systems. * Baldassarre & Azizpour [2019] Baldassarre, F., & Azizpour, H. (2019). Explainability techniques for graph convolutional networks. In International Conference on Machine Learning (ICML) Workshops, 2019 Workshop on Learning and Reasoning with Graph-Structured Representations. * Barredo-Arrieta et al. [2019] Barredo-Arrieta, A., Laña, I., & Del Ser, J. (2019). What lies beneath: A note on the explainability of black-box machine learning models for road traffic forecasting. In 2019 IEEE Intelligent Transportation Systems Conference (ITSC) (pp. 2232–2237). IEEE. * Bing et al. [2020] Bing, H., Zhifeng, X., Yangjie, X., Jinxing, H., & Zhanwu, M. (2020). Integrating semantic zoning information with the prediction of road link speed based on taxi gps data. Complexity, 2020. * Bogaerts et al. [2020] Bogaerts, T., Masegosa, A. D., Angarita-Zapata, J. S., Onieva, E., & Hellinckx, P. (2020). A graph cnn-lstm neural network for short and long-term traffic forecasting based on trajectory data. Transportation Research Part C: Emerging Technologies, 112, 62–77. * Boukerche et al. [2020] Boukerche, A., Tao, Y., & Sun, P. (2020). Artificial intelligence-based vehicular traffic flow prediction methods for supporting intelligent transportation systems. Computer Networks, 182, 107484. * Boukerche & Wang [2020a] Boukerche, A., & Wang, J. (2020a). Machine learning-based traffic prediction models for intelligent transportation systems. Computer Networks, 181, 107530. * Boukerche & Wang [2020b] Boukerche, A., & Wang, J. (2020b). A performance modeling and analysis of a novel vehicular traffic flow prediction system using a hybrid machine learning-based model. Ad Hoc Networks, . * Bruna et al. [2014] Bruna, J., Zaremba, W., Szlam, A., & LeCun, Y. (2014). Spectral networks and deep locally connected networks on graphs. In 2nd International Conference on Learning Representations, ICLR 2014. * Cai et al. [2020] Cai, L., Janowicz, K., Mai, G., Yan, B., & Zhu, R. (2020). Traffic transformer: Capturing the continuity and periodicity of time series for traffic forecasting. Transactions in GIS, . * Cao et al. [2020] Cao, D., Wang, Y., Duan, J., Zhang, C., Zhu, X., Huang, C., Tong, Y., Xu, B., Bai, J., Tong, J. et al. (2020). Spectral temporal graph neural network for multivariate time-series forecasting. Advances in Neural Information Processing Systems, 33. * Chai et al. [2018] Chai, D., Wang, L., & Yang, Q. (2018). Bike flow prediction with multi-graph convolutional networks. In Proceedings of the 26th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems (pp. 397–400). * Chen et al. [2019] Chen, C., Li, K., Teo, S. G., Zou, X., Wang, K., Wang, J., & Zeng, Z. (2019). Gated residual recurrent graph neural networks for traffic prediction. In Proceedings of the AAAI Conference on Artificial Intelligence (pp. 485–492). volume 33. * Chen et al. [2020a] Chen, F., Chen, Z., Biswas, S., Lei, S., Ramakrishnan, N., & Lu, C.-T. (2020a). Graph convolutional networks with kalman filtering for traffic prediction. In Proceedings of the 28th International Conference on Advances in Geographic Information Systems (pp. 135–138). * Chen et al. [2020b] Chen, H., Rossi, R. A., Mahadik, K., & Eldardiry, H. (2020b). A context integrated relational spatio-temporal model for demand and supply forecasting. arXiv preprint arXiv:2009.12469, . * Chen et al. [2020c] Chen, J., Liao, S., Hou, J., Wang, K., & Wen, J. (2020c). Gst-gcn: A geographic-semantic-temporal graph convolutional network for context-aware traffic flow prediction on graph sequences. In 2020 IEEE International Conference on Systems, Man, and Cybernetics (SMC) (pp. 1604–1609). IEEE. * Chen et al. [2020d] Chen, K., Chen, F., Lai, B., Jin, Z., Liu, Y., Li, K., Wei, L., Wang, P., Tang, Y., Huang, J. et al. (2020d). Dynamic spatio-temporal graph-based cnns for traffic flow prediction. IEEE Access, 8, 185136–185145. * Chen et al. [2020e] Chen, L., Han, K., Yin, Q., & Cao, Z. (2020e). Gdcrn: Global diffusion convolutional residual network for traffic flow prediction. In International Conference on Knowledge Science, Engineering and Management (pp. 438–449). Springer. * Chen et al. [2020f] Chen, W., Chen, L., Xie, Y., Cao, W., Gao, Y., & Feng, X. (2020f). Multi-range attentive bicomponent graph convolutional network for traffic forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence. volume 34. * Chen et al. [2020g] Chen, X., Zhang, Y., Du, L., Fang, Z., Ren, Y., Bian, K., & Xie, K. (2020g). Tssrgcn: Temporal spectral spatial retrieval graph convolutional network for traffic flow forecasting. In 2020 IEEE International Conference on Data Mining (ICDM). IEEE. * Chen et al. [2020h] Chen, Z., Zhao, B., Wang, Y., Duan, Z., & Zhao, X. (2020h). Multitask learning and gcn-based taxi demand prediction for a traffic road network. Sensors, 20, 3776\. * Cirstea et al. [2019] Cirstea, R.-G., Guo, C., & Yang, B. (2019). Graph attention recurrent neural networks for correlated time series forecasting. MileTS19@KDD, . * Cui et al. [2019] Cui, Z., Henrickson, K., Ke, R., & Wang, Y. (2019). Traffic graph convolutional recurrent neural network: A deep learning framework for network-scale traffic learning and forecasting. IEEE Transactions on Intelligent Transportation Systems, . * Cui et al. [2020a] Cui, Z., Ke, R., Pu, Z., Ma, X., & Wang, Y. (2020a). Learning traffic as a graph: A gated graph wavelet recurrent neural network for network-scale traffic prediction. Transportation Research Part C: Emerging Technologies, 115, 102620. * Cui et al. [2020b] Cui, Z., Lin, L., Pu, Z., & Wang, Y. (2020b). Graph markov network for traffic forecasting with missing data. Transportation Research Part C: Emerging Technologies, 117, 102671. URL: http://www.sciencedirect.com/science/article/pii/S0968090X20305866. doi:https://doi.org/10.1016/j.trc.2020.102671. * Dai et al. [2020] Dai, R., Xu, S., Gu, Q., Ji, C., & Liu, K. (2020). Hybrid spatio-temporal graph convolutional network: Improving traffic prediction with navigation data. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining KDD ’20 (p. 3074–3082). New York, NY, USA: Association for Computing Machinery. URL: https://doi.org/10.1145/3394486.3403358. doi:10.1145/3394486.3403358. * Davis et al. [2020] Davis, N., Raina, G., & Jagannathan, K. (2020). Grids versus graphs: Partitioning space for improved taxi demand-supply forecasts. IEEE Transactions on Intelligent Transportation Systems, . * Defferrard et al. [2016] Defferrard, M., Bresson, X., & Vandergheynst, P. (2016). Convolutional neural networks on graphs with fast localized spectral filtering. In Proceedings of the 30th International Conference on Neural Information Processing Systems (pp. 3844–3852). * Diao et al. [2019] Diao, Z., Wang, X., Zhang, D., Liu, Y., Xie, K., & He, S. (2019). Dynamic spatial-temporal graph convolutional neural networks for traffic forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence (pp. 890–897). volume 33. * Du et al. [2020] Du, B., Hu, X., Sun, L., Liu, J., Qiao, Y., & Lv, W. (2020). Traffic demand prediction based on dynamic transition convolutional neural network. IEEE Transactions on Intelligent Transportation Systems, . * Fan et al. [2020] Fan, X., Xiang, C., Gong, L., He, X., Qu, Y., Amirgholipour, S., Xi, Y., Nanda, P., & He, X. (2020). Deep learning for intelligent traffic sensing and prediction: recent advances and future challenges. CCF Transactions on Pervasive Computing and Interaction, (pp. 1–21). * Fang et al. [2020a] Fang, S., Pan, X., Xiang, S., & Pan, C. (2020a). Meta-msnet: Meta-learning based multi-source data fusion for traffic flow prediction. IEEE Signal Processing Letters, . * Fang et al. [2019] Fang, S., Zhang, Q., Meng, G., Xiang, S., & Pan, C. (2019). Gstnet: Global spatial-temporal network for traffic flow prediction. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19 (pp. 2286–2293). International Joint Conferences on Artificial Intelligence Organization. URL: https://doi.org/10.24963/ijcai.2019/317. doi:10.24963/ijcai.2019/317. * Fang et al. [2020b] Fang, X., Huang, J., Wang, F., Zeng, L., Liang, H., & Wang, H. (2020b). Constgat: Contextual spatial-temporal graph attention network for travel time estimation at baidu maps. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining KDD ’20 (p. 2697–2705). New York, NY, USA: Association for Computing Machinery. URL: https://doi.org/10.1145/3394486.3403320. doi:10.1145/3394486.3403320. * Feng et al. [2020] Feng, D., Wu, Z., Zhang, J., & Wu, Z. (2020). Dynamic global-local spatial-temporal network for traffic speed prediction. IEEE Access, 8, 209296–209307. * Fu et al. [2020] Fu, J., Zhou, W., & Chen, Z. (2020). Bayesian spatio-temporal graph convolutional network for traffic forecasting. arXiv preprint arXiv:2010.07498, . * Fukuda et al. [2020] Fukuda, S., Uchida, H., Fujii, H., & Yamada, T. (2020). Short-term prediction of traffic flow under incident conditions using graph convolutional recurrent neural network and traffic simulation. IET Intelligent Transport Systems, . * Garcia & Bruna [2017] Garcia, V., & Bruna, J. (2017). Few-shot learning with graph neural networks. arXiv preprint arXiv:1711.04043, . * Ge et al. [2019a] Ge, L., Li, H., Liu, J., & Zhou, A. (2019a). Temporal graph convolutional networks for traffic speed prediction considering external factors. In 2019 20th IEEE International Conference on Mobile Data Management (MDM) (pp. 234–242). IEEE. * Ge et al. [2019b] Ge, L., Li, H., Liu, J., & Zhou, A. (2019b). Traffic speed prediction with missing data based on tgcn. In 2019 IEEE SmartWorld, Ubiquitous Intelligence & Computing, Advanced & Trusted Computing, Scalable Computing & Communications, Cloud & Big Data Computing, Internet of People and Smart City Innovation (SmartWorld/SCALCOM/UIC/ATC/CBDCom/IOP/SCI) (pp. 522–529). IEEE. * Ge et al. [2020] Ge, L., Li, S., Wang, Y., Chang, F., & Wu, K. (2020). Global spatial-temporal graph convolutional network for urban traffic speed prediction. Applied Sciences, 10, 1509. * Geng et al. [2019a] Geng, X., Li, Y., Wang, L., Zhang, L., Yang, Q., Ye, J., & Liu, Y. (2019a). Spatiotemporal multi-graph convolution network for ride-hailing demand forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence (pp. 3656–3663). volume 33. * Geng et al. [2019b] Geng, X., Wu, X., Zhang, L., Yang, Q., Liu, Y., & Ye, J. (2019b). Multi-modal graph interaction for multi-graph convolution network in urban spatiotemporal forecasting. arXiv preprint arXiv:1905.11395, . * George & Santra [2020] George, S., & Santra, A. K. (2020). Traffic prediction using multifaceted techniques: A survey. Wireless Personal Communications, 115, 1047–1106. * Gilmer et al. [2017] Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., & Dahl, G. E. (2017). Neural message passing for quantum chemistry. In International Conference on Machine Learning (pp. 1263–1272). PMLR. * Goodfellow et al. [2014] Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., & Bengio, Y. (2014). Generative adversarial nets. Advances in neural information processing systems, 27, 2672–2680. * Guo & Yuan [2020] Guo, G., & Yuan, W. (2020). Short-term traffic speed forecasting based on graph attention temporal convolutional networks. Neurocomputing, . * Guo et al. [2019a] Guo, J., Song, C., & Wang, H. (2019a). A multi-step traffic speed forecasting model based on graph convolutional lstm. In 2019 Chinese Automation Congress (CAC) (pp. 2466–2471). IEEE. * Guo et al. [2020a] Guo, K., Hu, Y., Qian, Z., Liu, H., Zhang, K., Sun, Y., Gao, J., & Yin, B. (2020a). Optimized graph convolution recurrent neural network for traffic prediction. IEEE Transactions on Intelligent Transportation Systems, . * Guo et al. [2020b] Guo, K., Hu, Y., Qian, Z., Sun, Y., Gao, J., & Yin, B. (2020b). Dynamic graph convolution network for traffic forecasting based on latent network of laplace matrix estimation. IEEE Transactions on Intelligent Transportation Systems, . * Guo et al. [2020c] Guo, K., Hu, Y., Qian, Z. S., Sun, Y., Gao, J., & Yin, B. (2020c). An optimized temporal-spatial gated graph convolution network for traffic forecasting. IEEE Intelligent Transportation Systems Magazine, . * Guo et al. [2019b] Guo, R., Jiang, Z., Huang, J., Tao, J., Wang, C., Li, J., & Chen, L. (2019b). Bikenet: Accurate bike demand prediction using graph neural networks for station rebalancing. In 2019 IEEE SmartWorld, Ubiquitous Intelligence & Computing, Advanced & Trusted Computing, Scalable Computing & Communications, Cloud & Big Data Computing, Internet of People and Smart City Innovation (SmartWorld/SCALCOM/UIC/ATC/CBDCom/IOP/SCI) (pp. 686–693). IEEE. * Guo et al. [2019c] Guo, S., Lin, Y., Feng, N., Song, C., & Wan, H. (2019c). Attention based spatial-temporal graph convolutional networks for traffic flow forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence (pp. 922–929). volume 33. * Guopeng et al. [2020] Guopeng, L., Knoop, V. L., & van Lint, H. (2020). Dynamic graph filters networks: A gray-box model for multistep traffic forecasting. In 2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC) (pp. 1–6). IEEE. * Haghighat et al. [2020] Haghighat, A. K., Ravichandra-Mouli, V., Chakraborty, P., Esfandiari, Y., Arabi, S., & Sharma, A. (2020). Applications of deep learning in intelligent transportation systems. Journal of Big Data Analytics in Transportation, 2, 115–145. * Hamilton et al. [2017] Hamilton, W., Ying, Z., & Leskovec, J. (2017). Inductive representation learning on large graphs. In Advances in neural information processing systems (pp. 1024–1034). * Han et al. [2020] Han, X., Shen, G., Yang, X., & Kong, X. (2020). Congestion recognition for hybrid urban road systems via digraph convolutional network. Transportation Research Part C: Emerging Technologies, 121, 102877. * Han et al. [2019] Han, Y., Wang, S., Ren, Y., Wang, C., Gao, P., & Chen, G. (2019). Predicting station-level short-term passenger flow in a citywide metro network using spatiotemporal graph convolutional neural networks. ISPRS International Journal of Geo-Information, 8, 243. * Hasanzadeh et al. [2019] Hasanzadeh, A., Liu, X., Duffield, N., & Narayanan, K. R. (2019). Piecewise stationary modeling of random processes over graphs with an application to traffic prediction. In 2019 IEEE International Conference on Big Data (Big Data) (pp. 3779–3788). IEEE. * He & Shin [2020a] He, S., & Shin, K. G. (2020a). Dynamic flow distribution prediction for urban dockless e-scooter sharing reconfiguration. In Proceedings of The Web Conference 2020 (pp. 133–143). * He & Shin [2020b] He, S., & Shin, K. G. (2020b). Towards fine-grained flow forecasting: A graph attention approach for bike sharing systems. In Proceedings of The Web Conference 2020 WWW ’20 (p. 88–98). New York, NY, USA: Association for Computing Machinery. URL: https://doi.org/10.1145/3366423.3380097. doi:10.1145/3366423.3380097. * He et al. [2020] He, Y., Zhao, Y., Wang, H., & Tsui, K. L. (2020). Gc-lstm: A deep spatiotemporal model for passenger flow forecasting of high-speed rail network. In 2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC) (pp. 1–6). IEEE. * Heglund et al. [2020] Heglund, J. S., Taleongpong, P., Hu, S., & Tran, H. T. (2020). Railway delay prediction with spatial-temporal graph convolutional networks. In 2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC) (pp. 1–6). IEEE. * Hong et al. [2020] Hong, H., Lin, Y., Yang, X., Li, Z., Fu, K., Wang, Z., Qie, X., & Ye, J. (2020). Heteta: Heterogeneous information network embedding for estimating time of arrival. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining KDD ’20 (p. 2444–2454). New York, NY, USA: Association for Computing Machinery. URL: https://doi.org/10.1145/3394486.3403294. doi:10.1145/3394486.3403294. * Hu et al. [2018] Hu, J., Guo, C., Yang, B., Jensen, C. S., & Chen, L. (2018). Recurrent multi-graph neural networks for travel cost prediction. arXiv preprint arXiv:1811.05157, . * Hu et al. [2020] Hu, J., Yang, B., Guo, C., Jensen, C. S., & Xiong, H. (2020). Stochastic origin-destination matrix forecasting using dual-stage graph convolutional, recurrent neural networks. In 2020 IEEE 36th International Conference on Data Engineering (ICDE) (pp. 1417–1428). IEEE. * Huang et al. [2020a] Huang, R., Huang, C., Liu, Y., Dai, G., & Kong, W. (2020a). Lsgcn: Long short-term traffic prediction with graph convolutional networks. In C. Bessiere (Ed.), Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI-20 (pp. 2355–2361). International Joint Conferences on Artificial Intelligence Organization. URL: https://doi.org/10.24963/ijcai.2020/326. doi:10.24963/ijcai.2020/326 main track. * Huang et al. [2020b] Huang, Y., Zhang, S., Wen, J., & Chen, X. (2020b). Short-term traffic flow prediction based on graph convolutional network embedded lstm. In International Conference on Transportation and Development 2020 (pp. 159–168). American Society of Civil Engineers Reston, VA. * James [2019] James, J. (2019). Online traffic speed estimation for urban road networks with few data: A transfer learning approach. In 2019 IEEE Intelligent Transportation Systems Conference (ITSC) (pp. 4024–4029). IEEE. * James [2020] James, J. (2020). Citywide traffic speed prediction: A geometric deep learning approach. Knowledge-Based Systems, (p. 106592). * Jepsen et al. [2019] Jepsen, T. S., Jensen, C. S., & Nielsen, T. D. (2019). Graph convolutional networks for road networks. In Proceedings of the 27th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems (pp. 460–463). * Jepsen et al. [2020] Jepsen, T. S., Jensen, C. S., & Nielsen, T. D. (2020). Relational fusion networks: Graph convolutional networks for road networks. IEEE Transactions on Intelligent Transportation Systems, . * Jia et al. [2020] Jia, C., Wu, B., & Zhang, X.-P. (2020). Dynamic spatiotemporal graph neural network with tensor network. arXiv preprint arXiv:2003.08729, . * Jiang & Zhang [2018] Jiang, W., & Zhang, L. (2018). Geospatial data to images: A deep-learning framework for traffic forecasting. Tsinghua Science and Technology, 24, 52–64. * Jin et al. [2020a] Jin, G., Cui, Y., Zeng, L., Tang, H., Feng, Y., & Huang, J. (2020a). Urban ride-hailing demand prediction with multiple spatio-temporal information fusion network. Transportation Research Part C: Emerging Technologies, 117, 102665. * Jin et al. [2020b] Jin, G., Xi, Z., Sha, H., Feng, Y., & Huang, J. (2020b). Deep multi-view spatiotemporal virtual graph neural network for significant citywide ride-hailing demand prediction. arXiv preprint arXiv:2007.15189, . * Kang et al. [2019] Kang, Z., Xu, H., Hu, J., & Pei, X. (2019). Learning dynamic graph embedding for traffic flow forecasting: A graph self-attentive method. In 2019 IEEE Intelligent Transportation Systems Conference (ITSC) (pp. 2570–2576). IEEE. * Ke et al. [2021] Ke, J., Feng, S., Zhu, Z., Yang, H., & Ye, J. (2021). Joint predictions of multi-modal ride-hailing demands: A deep multi-task multi-graph learning-based approach. Transportation Research Part C: Emerging Technologies, 127, 103063. * Ke et al. [2019] Ke, J., Qin, X., Yang, H., Zheng, Z., Zhu, Z., & Ye, J. (2019). Predicting origin-destination ride-sourcing demand with a spatio-temporal encoder-decoder residual multi-graph convolutional network. arXiv preprint arXiv:1910.09103, . * Kim et al. [2020] Kim, S.-S., Chung, M., & Kim, Y.-K. (2020). Urban traffic prediction using congestion diffusion model. In 2020 IEEE International Conference on Consumer Electronics-Asia (ICCE-Asia) (pp. 1–4). IEEE. * Kim et al. [2019] Kim, T. S., Lee, W. K., & Sohn, S. Y. (2019). Graph convolutional network approach applied to predict hourly bike-sharing demands considering spatial, temporal, and global effects. PLOS ONE, 14, e0220782. * Kipf & Welling [2016] Kipf, T. N., & Welling, M. (2016). Variational graph auto-encoders. arXiv preprint arXiv:1611.07308, . * Kipf & Welling [2017] Kipf, T. N., & Welling, M. (2017). Semi-supervised classification with graph convolutional networks. In International Conference on Learning Representations (ICLR ’17). * Kong et al. [2020] Kong, X., Xing, W., Wei, X., Bao, P., Zhang, J., & Lu, W. (2020). Stgat: Spatial-temporal graph attention networks for traffic flow forecasting. IEEE Access, . * Lee et al. [2019] Lee, D., Jung, S., Cheon, Y., Kim, D., & You, S. (2019). Demand forecasting from spatiotemporal data with graph networks and temporal-guided embedding. arXiv preprint arXiv:1905.10709, . * Lee et al. [2021] Lee, K., Eo, M., Jung, E., Yoon, Y., & Rhee, W. (2021). Short-term traffic prediction with deep neural networks: A survey. IEEE Access, 9, 54739–54756. * Lee & Rhee [2019a] Lee, K., & Rhee, W. (2019a). Ddp-gcn: Multi-graph convolutional network for spatiotemporal traffic forecasting. arXiv preprint arXiv:1905.12256, . * Lee & Rhee [2019b] Lee, K., & Rhee, W. (2019b). Graph convolutional modules for traffic forecasting. CoRR, abs/1905.12256. URL: http://arxiv.org/abs/1905.12256. arXiv:1905.12256. * Lewenfus et al. [2020] Lewenfus, G., Martins, W. A., Chatzinotas, S., & Ottersten, B. (2020). Joint forecasting and interpolation of time-varying graph signals using deep learning. IEEE Transactions on Signal and Information Processing over Networks, . * Li & Axhausen [2020] Li, A., & Axhausen, K. W. (2020). Short-term traffic demand prediction using graph convolutional neural networks. AGILE: GIScience Series, 1, 1--14. * Li et al. [2020a] Li, C., Bai, L., Liu, W., Yao, L., & Waller, S. T. (2020a). Knowledge adaption for demand prediction based on multi-task memory neural network. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management (pp. 715--724). * Li et al. [2018a] Li, J., Peng, H., Liu, L., Xiong, G., Du, B., Ma, H., Wang, L., & Bhuiyan, M. Z. A. (2018a). Graph cnns for urban traffic passenger flows prediction. In 2018 IEEE SmartWorld, Ubiquitous Intelligence & Computing, Advanced & Trusted Computing, Scalable Computing & Communications, Cloud & Big Data Computing, Internet of People and Smart City Innovation (SmartWorld/SCALCOM/UIC/ATC/CBDCom/IOP/SCI) (pp. 29--36). IEEE. * Li & Zhu [2021] Li, M., & Zhu, Z. (2021). Spatial-temporal fusion graph neural networks for traffic flow forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence (pp. 4189--4196). volume 35. * Li et al. [2020b] Li, W., Wang, X., Zhang, Y., & Wu, Q. (2020b). Traffic flow prediction over muti-sensor data correlation with graph convolution network. Neurocomputing, . * Li et al. [2020c] Li, W., Yang, X., Tang, X., & Xia, S. (2020c). Sdcn: Sparsity and diversity driven correlation networks for traffic demand forecasting. In 2020 International Joint Conference on Neural Networks (IJCNN) (pp. 1--8). IEEE. * Li & Moura [2020] Li, Y., & Moura, J. M. (2020). Forecaster: A graph transformer for forecasting spatial and time-dependent data. In Proceedings of the Twenty-fourth European Conference on Artificial Intelligence. * Li et al. [2018b] Li, Y., Yu, R., Shahabi, C., & Liu, Y. (2018b). Diffusion convolutional recurrent neural network: Data-driven traffic forecasting. In International Conference on Learning Representations (ICLR ’18). * Li et al. [2020d] Li, Z., Li, L., Peng, Y., & Tao, X. (2020d). A two-stream graph convolutional neural network for dynamic traffic flow forecasting. In 2020 IEEE 32nd International Conference on Tools with Artificial Intelligence (ICTAI) (pp. 355--362). IEEE. * Li et al. [2020e] Li, Z., Sergin, N. D., Yan, H., Zhang, C., & Tsung, F. (2020e). Tensor completion for weakly-dependent data on graph for metro passenger flow prediction. In Proceedings of the AAAI Conference on Artificial Intelligence. volume 34. * Li et al. [2019] Li, Z., Xiong, G., Chen, Y., Lv, Y., Hu, B., Zhu, F., & Wang, F.-Y. (2019). A hybrid deep learning approach with gcn and lstm for traffic flow prediction. In 2019 IEEE Intelligent Transportation Systems Conference (ITSC) (pp. 1929--1933). IEEE. * Li et al. [2020f] Li, Z., Xiong, G., Tian, Y., Lv, Y., Chen, Y., Hui, P., & Su, X. (2020f). A multi-stream feature fusion approach for traffic prediction. IEEE Transactions on Intelligent Transportation Systems, . * Liao et al. [2018] Liao, B., Zhang, J., Wu, C., McIlwraith, D., Chen, T., Yang, S., Guo, Y., & Wu, F. (2018). Deep sequence learning with auxiliary information for traffic prediction. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (pp. 537--546). * Lin et al. [2018] Lin, L., He, Z., & Peeta, S. (2018). Predicting station-level hourly demand in a large-scale bike-sharing network: A graph convolutional neural network approach. Transportation Research Part C: Emerging Technologies, 97, 258--276. * Liu et al. [2020a] Liu, J., Ong, G. P., & Chen, X. (2020a). Graphsage-based traffic speed forecasting for segment network with sparse data. IEEE Transactions on Intelligent Transportation Systems, . * Liu et al. [2020b] Liu, L., Chen, J., Wu, H., Zhen, J., Li, G., & Lin, L. (2020b). Physical-virtual collaboration modeling for intra-and inter-station metro ridership prediction. IEEE Transactions on Intelligent Transportation Systems, . * Liu et al. [2019] Liu, L., Zhou, T., Long, G., Jiang, J., & Zhang, C. (2019). Learning to propagate for graph meta-learning. In Advances in Neural Information Processing Systems (pp. 1039--1050). * Liu et al. [2020c] Liu, R., Zhao, S., Cheng, B., Yang, H., Tang, H., & Yang, F. (2020c). St-mfm: A spatiotemporal multi-modal fusion model for urban anomalies prediction. In Proceedings of the Twenty-fourth European Conference on Artificial Intelligence. * Lu et al. [2020a] Lu, B., Gan, X., Jin, H., Fu, L., & Zhang, H. (2020a). Spatiotemporal adaptive gated graph convolution network for urban traffic flow forecasting. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management (pp. 1025--1034). * Lu et al. [2019a] Lu, M., Zhang, K., Liu, H., & Xiong, N. (2019a). Graph hierarchical convolutional recurrent neural network (ghcrnn) for vehicle condition prediction. arXiv preprint arXiv:1903.06261, . * Lu et al. [2020b] Lu, Z., Lv, W., Cao, Y., Xie, Z., Peng, H., & Du, B. (2020b). Lstm variants meet graph neural networks for road speed prediction. Neurocomputing, . * Lu et al. [2019b] Lu, Z., Lv, W., Xie, Z., Du, B., & Huang, R. (2019b). Leveraging graph neural network with lstm for traffic speed prediction. In 2019 IEEE SmartWorld, Ubiquitous Intelligence & Computing, Advanced & Trusted Computing, Scalable Computing & Communications, Cloud & Big Data Computing, Internet of People and Smart City Innovation (SmartWorld/SCALCOM/UIC/ATC/CBDCom/IOP/SCI) (pp. 74--81). IEEE. * Luca et al. [2020] Luca, M., Barlacchi, G., Lepri, B., & Pappalardo, L. (2020). Deep learning for human mobility: a survey on data and models. arXiv preprint arXiv:2012.02825, . * Luo et al. [2020] Luo, M., Du, B., Klemmer, K., Zhu, H., Ferhatosmanoglu, H., & Wen, H. (2020). D3p: Data-driven demand prediction for fast expanding electric vehicle sharing systems. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 4, 1--21. * Lv et al. [2020] Lv, M., Hong, Z., Chen, L., Chen, T., Zhu, T., & Ji, S. (2020). Temporal multi-graph convolutional network for traffic flow prediction. IEEE Transactions on Intelligent Transportation Systems, . * Maas & Bloem [2020] Maas, T., & Bloem, P. (2020). Uncertainty intervals for graph-based spatio-temporal traffic prediction. arXiv preprint arXiv:2012.05207, . * Mallick et al. [2020] Mallick, T., Balaprakash, P., Rask, E., & Macfarlane, J. (2020). Graph-partitioning-based diffusion convolution recurrent neural network for large-scale traffic forecasting. Transportation Research Record, (p. 0361198120930010). * Mallick et al. [2021] Mallick, T., Balaprakash, P., Rask, E., & Macfarlane, J. (2021). Transfer learning with graph neural networks for short-term highway traffic forecasting. In 2020 25th International Conference on Pattern Recognition (ICPR) (pp. 10367--10374). IEEE. * Manibardo et al. [2020] Manibardo, E. L., Laña, I., & Del Ser, J. (2020). Deep learning for road traffic forecasting: Does it make a difference? arXiv preprint arXiv:2012.02260, . * Mena-Oreja & Gozalvez [2020] Mena-Oreja, J., & Gozalvez, J. (2020). A comprehensive evaluation of deep learning-based techniques for traffic prediction. IEEE Access, 8, 91188--91212. * Mohanty & Pozdnukhov [2018] Mohanty, S., & Pozdnukhov, A. (2018). Graph cnn+ lstm framework for dynamic macroscopic traffic congestion prediction. In International Workshop on Mining and Learning with Graphs. * Mohanty et al. [2020] Mohanty, S., Pozdnukhov, A., & Cassidy, M. (2020). Region-wide congestion prediction and control using deep learning. Transportation Research Part C: Emerging Technologies, 116, 102624. * Opolka et al. [2019] Opolka, F. L., Solomon, A., Cangea, C., Veličković, P., Liò, P., & Hjelm, R. D. (2019). Spatio-temporal deep graph infomax. In Representation Learning on Graphs and Manifolds, ICLR 2019 Workshop. * Oreshkin et al. [2021] Oreshkin, B. N., Amini, A., Coyle, L., & Coates, M. (2021). Fc-gaga: Fully connected gated graph architecture for spatio-temporal traffic forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence (pp. 9233--9241). volume 35. * Ou et al. [2020] Ou, J., Sun, J., Zhu, Y., Jin, H., Liu, Y., Zhang, F., Huang, J., & Wang, X. (2020). Stp-trellisnets: Spatial-temporal parallel trellisnets for metro station passenger flow prediction. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management (pp. 1185--1194). * Pan et al. [2019] Pan, Z., Liang, Y., Wang, W., Yu, Y., Zheng, Y., & Zhang, J. (2019). Urban traffic prediction from spatio-temporal data using deep meta learning. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (pp. 1720--1730). * Pan et al. [2020] Pan, Z., Zhang, W., Liang, Y., Zhang, W., Yu, Y., Zhang, J., & Zheng, Y. (2020). Spatio-temporal meta learning for urban traffic prediction. IEEE Transactions on Knowledge and Data Engineering, . * Park et al. [2020] Park, C., Lee, C., Bahng, H., Tae, Y., Jin, S., Kim, K., Ko, S., & Choo, J. (2020). St-grat: A novel spatio-temporal graph attention networks for accurately forecasting dynamically changing road speed. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management (pp. 1215--1224). * Pavlyuk [2019] Pavlyuk, D. (2019). Feature selection and extraction in spatiotemporal traffic forecasting: a systematic literature review. European Transport Research Review, 11, 6. * Peng et al. [2020] Peng, H., Wang, H., Du, B., Bhuiyan, M. Z. A., Ma, H., Liu, J., Wang, L., Yang, Z., Du, L., Wang, S. et al. (2020). Spatial temporal incidence dynamic graph neural networks for traffic flow forecasting. Information Sciences, 521, 277--290. * Pian & Wu [2020] Pian, W., & Wu, Y. (2020). Spatial-temporal dynamic graph attention networks for ride-hailing demand prediction. arXiv preprint arXiv:2006.05905, . * Pope et al. [2019] Pope, P. E., Kolouri, S., Rostami, M., Martin, C. E., & Hoffmann, H. (2019). Explainability methods for graph convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 10772--10781). * Qin et al. [2020a] Qin, K., Xu, Y., Kang, C., & Kwan, M.-P. (2020a). A graph convolutional network model for evaluating potential congestion spots based on local urban built environments. Transactions in GIS, . * Qin et al. [2020b] Qin, T., Liu, T., Wu, H., Tong, W., & Zhao, S. (2020b). Resgcn: Residual graph convolutional network based free dock prediction in bike sharing system. In 2020 21st IEEE International Conference on Mobile Data Management (MDM) (pp. 210--217). IEEE. * Qiu et al. [2020] Qiu, H., Zheng, Q., Msahli, M., Memmi, G., Qiu, M., & Lu, J. (2020). Topological graph convolutional network-based urban traffic flow and density prediction. IEEE Transactions on Intelligent Transportation Systems, . * Qu et al. [2020] Qu, Y., Zhu, Y., Zang, T., Xu, Y., & Yu, J. (2020). Modeling local and global flow aggregation for traffic flow forecasting. In International Conference on Web Information Systems Engineering (pp. 414--429). Springer. * Ramadan et al. [2020] Ramadan, A., Elbery, A., Zorba, N., & Hassanein, H. S. (2020). Traffic forecasting using temporal line graph convolutional network: Case study. In ICC 2020-2020 IEEE International Conference on Communications (ICC) (pp. 1--6). IEEE. * Ren & Xie [2019] Ren, Y., & Xie, K. (2019). Transfer knowledge between sub-regions for traffic prediction using deep learning method. In International Conference on Intelligent Data Engineering and Automated Learning (pp. 208--219). Springer. * Sánchez et al. [2020] Sánchez, C. S., Wieder, A., Sottovia, P., Bortoli, S., Baumbach, J., & Axenie, C. (2020). Gannster: Graph-augmented neural network spatio-temporal reasoner for traffic forecasting. In International Workshop on Advanced Analysis and Learning on Temporal Data (AALTD). Springer. * Scarselli et al. [2008] Scarselli, F., Gori, M., Tsoi, A. C., Hagenbuchner, M., & Monfardini, G. (2008). The graph neural network model. IEEE transactions on neural networks, 20, 61--80. * Shao et al. [2020] Shao, K., Wang, K., Chen, L., & Zhou, Z. (2020). Estimation of urban travel time with sparse traffic surveillance data. In Proceedings of the 2020 4th High Performance Computing and Cluster Technologies Conference & 2020 3rd International Conference on Big Data and Artificial Intelligence (pp. 218--223). * Shen et al. [2020] Shen, Y., Jin, C., & Hua, J. (2020). Ttpnet: A neural network for travel time prediction based on tensor decomposition and graph embedding. IEEE Transactions on Knowledge and Data Engineering, . * Shi et al. [2020] Shi, H., Yao, Q., Guo, Q., Li, Y., Zhang, L., Ye, J., Li, Y., & Liu, Y. (2020). Predicting origin-destination flow via multi-perspective graph convolutional network. In 2020 IEEE 36th International Conference on Data Engineering (ICDE) (pp. 1818--1821). IEEE. * Shi & Yeung [2018] Shi, X., & Yeung, D.-Y. (2018). Machine learning for spatiotemporal sequence forecasting: A survey. arXiv preprint arXiv:1808.06865, . * Shin & Yoon [2020] Shin, Y., & Yoon, Y. (2020). Incorporating dynamicity of transportation network with multi-weight traffic graph convolutional network for traffic forecasting. IEEE Transactions on Intelligent Transportation Systems, . * Shleifer et al. [2019] Shleifer, S., McCreery, C., & Chitters, V. (2019). Incrementally improving graph wavenet performance on traffic prediction. arXiv preprint arXiv:1912.07390, . * Song et al. [2020a] Song, C., Lin, Y., Guo, S., & Wan, H. (2020a). Spatial-temporal synchronous graph convolutional networks: A new framework for spatial-temporal network data forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence. volume 34. * Song et al. [2020b] Song, Q., Ming, R., Hu, J., Niu, H., & Gao, M. (2020b). Graph attention convolutional network: Spatiotemporal modeling for urban traffic prediction. In 2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC) (pp. 1--6). IEEE. * Sun et al. [2020] Sun, J., Zhang, J., Li, Q., Yi, X., Liang, Y., & Zheng, Y. (2020). Predicting citywide crowd flows in irregular regions using multi-view graph convolutional networks. IEEE Transactions on Knowledge and Data Engineering, (pp. 1--1). * Sun et al. [2020] Sun, X., Li, J., Lv, Z., & Dong, C. (2020). Traffic flow prediction model based on spatio-temporal dilated graph convolution. KSII Transactions on Internet & Information Systems, 14. * Sun et al. [2021] Sun, Y., Wang, Y., Fu, K., Wang, Z., Zhang, C., & Ye, J. (2021). Constructing geographic and long-term temporal graph for traffic forecasting. In 2020 25th International Conference on Pattern Recognition (ICPR) (pp. 3483--3490). IEEE. * Tang et al. [2020a] Tang, C., Sun, J., & Sun, Y. (2020a). Dynamic spatial-temporal graph attention graph convolutional network for short-term traffic flow forecasting. In 2020 IEEE International Symposium on Circuits and Systems (ISCAS) (pp. 1--5). IEEE. * Tang et al. [2020b] Tang, C., Sun, J., Sun, Y., Peng, M., & Gan, N. (2020b). A general traffic flow prediction approach based on spatial-temporal graph attention. IEEE Access, 8, 153731--153741. * Tedjopurnomo et al. [2020] Tedjopurnomo, D. A., Bao, Z., Zheng, B., Choudhury, F., & Qin, A. (2020). A survey on modern deep neural network for traffic prediction: Trends, methods and challenges. IEEE Transactions on Knowledge and Data Engineering, . * Tian et al. [2020] Tian, K., Guo, J., Ye, K., & Xu, C.-Z. (2020). St-mgat: Spatial-temporal multi-head graph attention networks for traffic forecasting. In 2020 IEEE 32nd International Conference on Tools with Artificial Intelligence (ICTAI) (pp. 714--721). IEEE. * Varghese et al. [2020] Varghese, V., Chikaraishi, M., & Urata, J. (2020). Deep learning in transport studies: A meta-analysis on the prediction accuracy. Journal of Big Data Analytics in Transportation, (pp. 1--22). * Vaswani et al. [2017] Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all you need. Advances in neural information processing systems, 30, 5998--6008. * Veličković et al. [2018] Veličković, P., Cucurull, G., Casanova, A., Romero, A., Liò, P., & Bengio, Y. (2018). Graph attention networks. In International Conference on Learning Representations. * Wang et al. [2018a] Wang, B., Luo, X., Zhang, F., Yuan, B., Bertozzi, A. L., & Brantingham, P. J. (2018a). Graph-based deep modeling and real time forecasting of sparse spatio-temporal data. arXiv preprint arXiv:1804.00684, . * Wang et al. [2020a] Wang, C., Zhang, K., Wang, H., & Chen, B. (2020a). Auto-stgcn: Autonomous spatial-temporal graph convolutional network search based on reinforcement learning and existing research results. arXiv preprint arXiv:2010.07474, . * Wang et al. [2020b] Wang, F., Xu, J., Liu, C., Zhou, R., & Zhao, P. (2020b). Mtgcn: A multitask deep learning model for traffic flow prediction. In International Conference on Database Systems for Advanced Applications (pp. 435--451). Springer. * Wang et al. [2020c] Wang, H.-W., Peng, Z.-R., Wang, D., Meng, Y., Wu, T., Sun, W., & Lu, Q.-C. (2020c). Evaluation and prediction of transportation resilience under extreme weather events: A diffusion graph convolutional approach. Transportation Research Part C: Emerging Technologies, 115, 102619. * Wang et al. [2020d] Wang, Q., Guo, B., Ouyang, Y., Shu, K., Yu, Z., & Liu, H. (2020d). Spatial community-informed evolving graphs for demand prediction. In Proceedings of The European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML-PKDD 2020). * Wang et al. [2020e] Wang, S., Miao, H., Chen, H., & Huang, Z. (2020e). Multi-task adversarial spatial-temporal networks for crowd flow prediction. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management (pp. 1555--1564). * Wang et al. [2018b] Wang, X., Chen, C., Min, Y., He, J., Yang, B., & Zhang, Y. (2018b). Efficient metropolitan traffic prediction based on graph recurrent neural network. arXiv preprint arXiv:1811.00740, . * Wang et al. [2020f] Wang, X., Guan, X., Cao, J., Zhang, N., & Wu, H. (2020f). Forecast network-wide traffic states for multiple steps ahead: A deep learning approach considering dynamic non-local spatial correlation and non-stationary temporal dependency. Transportation Research Part C: Emerging Technologies, 119, 102763. URL: http://www.sciencedirect.com/science/article/pii/S0968090X20306756. doi:https://doi.org/10.1016/j.trc.2020.102763. * Wang et al. [2020g] Wang, X., Ma, Y., Wang, Y., Jin, W., Wang, X., Tang, J., Jia, C., & Yu, J. (2020g). Traffic flow prediction via spatial temporal graph neural network. In Proceedings of The Web Conference 2020 WWW ’20 (p. 1082–1092). New York, NY, USA: Association for Computing Machinery. URL: https://doi.org/10.1145/3366423.3380186. doi:10.1145/3366423.3380186. * Wang et al. [2020h] Wang, Y., Xu, D., Peng, P., Xuan, Q., & Zhang, G. (2020h). An urban commuters’ od hybrid prediction method based on big gps data. Chaos: An Interdisciplinary Journal of Nonlinear Science, 30, 093128\. * Wang et al. [2019] Wang, Y., Yin, H., Chen, H., Wo, T., Xu, J., & Zheng, K. (2019). Origin-destination matrix prediction via graph convolution: a new perspective of passenger demand modeling. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (pp. 1227--1235). * Wei & Sheng [2020] Wei, C., & Sheng, J. (2020). Spatial-temporal graph attention networks for traffic flow forecasting. In IOP Conference Series: Earth and Environmental Science (p. 012065). IOP Publishing volume 587. * Wei et al. [2019] Wei, L., Yu, Z., Jin, Z., Xie, L., Huang, J., Cai, D., He, X., & Hua, X.-S. (2019). Dual graph for traffic forecasting. IEEE Access, . * Wright et al. [2019] Wright, M. A., Ehlers, S. F., & Horowitz, R. (2019). Neural-attention-based deep learning architectures for modeling traffic dynamics on lane graphs. In 2019 IEEE Intelligent Transportation Systems Conference (ITSC) (pp. 3898--3905). IEEE. * Wu et al. [2020a] Wu, M., Zhu, C., & Chen, L. (2020a). Multi-task spatial-temporal graph attention network for taxi demand prediction. In Proceedings of the 2020 5th International Conference on Mathematics and Artificial Intelligence (pp. 224--228). * Wu et al. [2018a] Wu, T., Chen, F., & Wan, Y. (2018a). Graph attention lstm network: A new model for traffic flow forecasting. In 2018 5th International Conference on Information Science and Control Engineering (ICISCE) (pp. 241--245). IEEE. * Wu et al. [2018b] Wu, Y., Tan, H., Qin, L., Ran, B., & Jiang, Z. (2018b). A hybrid deep learning based traffic flow prediction method and its understanding. Transportation Research Part C: Emerging Technologies, 90, 166--180. * Wu et al. [2020b] Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., & Philip, S. Y. (2020b). A comprehensive survey on graph neural networks. IEEE Transactions on Neural Networks and Learning Systems, . * Wu et al. [2020c] Wu, Z., Pan, S., Long, G., Jiang, J., Chang, X., & Zhang, C. (2020c). Connecting the dots: Multivariate time series forecasting with graph neural networks. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining KDD ’20 (p. 753–763). New York, NY, USA: Association for Computing Machinery. URL: https://doi.org/10.1145/3394486.3403118. doi:10.1145/3394486.3403118. * Wu et al. [2019] Wu, Z., Pan, S., Long, G., Jiang, J., & Zhang, C. (2019). Graph wavenet for deep spatial-temporal graph modeling. In Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19 (pp. 1907--1913). International Joint Conferences on Artificial Intelligence Organization. URL: https://doi.org/10.24963/ijcai.2019/264. doi:10.24963/ijcai.2019/264. * Xiao et al. [2020] Xiao, G., Wang, R., Zhang, C., & Ni, A. (2020). Demand prediction for a public bike sharing program based on spatio-temporal graph convolutional networks. Multimedia Tools and Applications, (pp. 1--19). * Xie et al. [2020a] Xie, P., Li, T., Liu, J., Du, S., Yang, X., & Zhang, J. (2020a). Urban flow prediction from spatiotemporal data using machine learning: A survey. Information Fusion, 59, 1--12. * Xie et al. [2020b] Xie, Q., Guo, T., Chen, Y., Xiao, Y., Wang, X., & Zhao, B. Y. (2020b). Deep graph convolutional networks for incident-driven traffic speed prediction. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management (pp. 1665--1674). * Xie et al. [2020c] Xie, Y., Xiong, Y., & Zhu, Y. (2020c). Istd-gcn: Iterative spatial-temporal diffusion graph convolutional network for traffic speed forecasting. arXiv preprint arXiv:2008.03970, . * Xie et al. [2020d] Xie, Y., Xiong, Y., & Zhu, Y. (2020d). Sast-gnn: A self-attention based spatio-temporal graph neural network for traffic prediction. In International Conference on Database Systems for Advanced Applications (pp. 707--714). Springer. * Xie et al. [2019] Xie, Z., Lv, W., Huang, S., Lu, Z., Du, B., & Huang, R. (2019). Sequential graph neural network for urban road traffic speed prediction. IEEE Access, . * Xin et al. [2020] Xin, Y., Miao, D., Zhu, M., Jin, C., & Lu, X. (2020). Internet: Multistep traffic forecasting by interacting spatial and temporal features. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management (pp. 3477--3480). * Xiong et al. [2020] Xiong, X., Ozbay, K., Jin, L., & Feng, C. (2020). Dynamic origin--destination matrix prediction with line graph neural networks and kalman filter. Transportation Research Record, (p. 0361198120919399). * Xu et al. [2019] Xu, D., Dai, H., Wang, Y., Peng, P., Xuan, Q., & Guo, H. (2019). Road traffic state prediction based on a graph embedding recurrent neural network under the scats. Chaos: An Interdisciplinary Journal of Nonlinear Science, 29, 103125\. * Xu et al. [2020a] Xu, D., Wei, C., Peng, P., Xuan, Q., & Guo, H. (2020a). Ge-gan: A novel deep learning framework for road traffic state estimation. Transportation Research Part C: Emerging Technologies, 117, 102635. * Xu et al. [2020b] Xu, M., Dai, W., Liu, C., Gao, X., Lin, W., Qi, G.-J., & Xiong, H. (2020b). Spatial-temporal transformer networks for traffic flow forecasting. arXiv preprint arXiv:2001.02908, . * Xu et al. [2020c] Xu, X., Zheng, H., Feng, X., & Chen, Y. (2020c). Traffic flow forecasting with spatial-temporal graph convolutional networks in edge-computing systems. In 2020 International Conference on Wireless Communications and Signal Processing (WCSP) (pp. 251--256). IEEE. * Xu & Li [2019] Xu, Y., & Li, D. (2019). Incorporating graph attention and recurrent architectures for city-wide taxi demand prediction. ISPRS International Journal of Geo-Information, 8, 414. * Xu et al. [2020d] Xu, Z., Kang, Y., Cao, Y., & Li, Z. (2020d). Spatiotemporal graph convolution multifusion network for urban vehicle emission prediction. IEEE Transactions on Neural Networks and Learning Systems, . * Yang et al. [2020] Yang, F., Chen, L., Zhou, F., Gao, Y., & Cao, W. (2020). Relational state-space model for stochastic multi-object systems. In International Conference on Learning Representations. * Yang et al. [2019] Yang, S., Ma, W., Pi, X., & Qian, S. (2019). A deep learning approach to real-time parking occupancy prediction in transportation networks incorporating multiple spatio-temporal data sources. Transportation Research Part C: Emerging Technologies, 107, 248--265. * Yao et al. [2020] Yao, X., Gao, Y., Zhu, D., Manley, E., Wang, J., & Liu, Y. (2020). Spatial origin-destination flow imputation using graph convolutional networks. IEEE Transactions on Intelligent Transportation Systems, . * Ye et al. [2021] Ye, J., Sun, L., Du, B., Fu, Y., & Xiong, H. (2021). Coupled layer-wise graph convolution for transportation demand prediction. In Proceedings of the AAAI Conference on Artificial Intelligence (pp. 4617--4625). volume 35. * Ye et al. [2020a] Ye, J., Zhao, J., Ye, K., & Xu, C. (2020a). How to build a graph-based deep learning architecture in traffic domain: A survey. IEEE Transactions on Intelligent Transportation Systems, . * Ye et al. [2020b] Ye, J., Zhao, J., Ye, K., & Xu, C. (2020b). Multi-stgcnet: A graph convolution based spatial-temporal framework for subway passenger flow forecasting. In 2020 International Joint Conference on Neural Networks (IJCNN) (pp. 1--8). IEEE. * Yeghikyan et al. [2020] Yeghikyan, G., Opolka, F. L., Nanni, M., Lepri, B., & Liò, P. (2020). Learning mobility flows from urban features with spatial interaction models and neural networks. In 2020 IEEE International Conference on Smart Computing (SMARTCOMP) (pp. 57--64). IEEE. * Yin et al. [2020] Yin, X., Wu, G., Wei, J., Shen, Y., Qi, H., & Yin, B. (2020). Multi-stage attention spatial-temporal graph networks for traffic prediction. Neurocomputing, . * Yin et al. [2021] Yin, X., Wu, G., Wei, J., Shen, Y., Qi, H., & Yin, B. (2021). Deep learning on traffic prediction: Methods, analysis and future directions. IEEE Transactions on Intelligent Transportation Systems, . * Ying et al. [2019] Ying, Z., Bourgeois, D., You, J., Zitnik, M., & Leskovec, J. (2019). Gnnexplainer: Generating explanations for graph neural networks. In Advances in neural information processing systems (pp. 9244--9255). * Yoshida et al. [2019] Yoshida, A., Yatsushiro, Y., Hata, N., Higurashi, T., Tateiwa, N., Wakamatsu, T., Tanaka, A., Nagamatsu, K., & Fujisawa, K. (2019). Practical end-to-end repositioning algorithm for managing bike-sharing system. In 2019 IEEE International Conference on Big Data (Big Data) (pp. 1251--1258). IEEE. * Yu et al. [2020a] Yu, B., Lee, Y., & Sohn, K. (2020a). Forecasting road traffic speeds by considering area-wide spatio-temporal dependencies based on a graph convolutional neural network (gcn). Transportation Research Part C: Emerging Technologies, 114, 189--204. * Yu et al. [2019a] Yu, B., Li, M., Zhang, J., & Zhu, Z. (2019a). 3d graph convolutional networks with temporal graphs: A spatial information free framework for traffic forecasting. arXiv preprint arXiv:1903.00919, . * Yu et al. [2018] Yu, B., Yin, H., & Zhu, Z. (2018). Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI-18 (pp. 3634--3640). International Joint Conferences on Artificial Intelligence Organization. URL: https://doi.org/10.24963/ijcai.2018/505. doi:10.24963/ijcai.2018/505. * Yu et al. [2019b] Yu, B., Yin, H., & Zhu, Z. (2019b). St-unet: A spatio-temporal u-network for graph-structured time series modeling. arXiv preprint arXiv:1903.05631, . * Yu & Gu [2019] Yu, J. J. Q., & Gu, J. (2019). Real-time traffic speed estimation with graph convolutional generative autoencoder. IEEE Transactions on Intelligent Transportation Systems, 20, 3940--3951. * Yu et al. [2020b] Yu, L., Du, B., Hu, X., Sun, L., Han, L., & Lv, W. (2020b). Deep spatio-temporal graph convolutional network for traffic accident prediction. Neurocomputing, . * Yuan et al. [2010] Yuan, J., Zheng, Y., Zhang, C., Xie, W., Xie, X., Sun, G., & Huang, Y. (2010). T-drive: driving directions based on taxi trajectories. In Proceedings of the 18th SIGSPATIAL International conference on advances in geographic information systems (pp. 99--108). * Zhang et al. [2019a] Zhang, C., James, J., & Liu, Y. (2019a). Spatial-temporal graph attention networks: A deep learning approach for traffic forecasting. IEEE Access, 7, 166246--166256. * Zhang et al. [2020a] Zhang, H., Liu, J., Tang, Y., & Xiong, G. (2020a). Attention based graph covolution networks for intelligent traffic flow analysis. In 2020 IEEE 16th International Conference on Automation Science and Engineering (CASE) (pp. 558--563). IEEE. * Zhang et al. [2020b] Zhang, J., Chen, F., Cui, Z., Guo, Y., & Zhu, Y. (2020b). Deep learning architecture for short-term passenger flow forecasting in urban rail transit. IEEE Transactions on Intelligent Transportation Systems, . * Zhang et al. [2020c] Zhang, J., Chen, F., & Guo, Y. (2020c). Multi-graph convolutional network for short-term passenger flow forecasting in urban rail transit. IET Intelligent Transport Systems, . * Zhang et al. [2018a] Zhang, J., Shi, X., Xie, J., Ma, H., King, I., & Yeung, D.-Y. (2018a). Gaan: Gated attention networks for learning on large and spatiotemporal graphs. arXiv preprint arXiv:1803.07294, . * Zhang et al. [2017] Zhang, J., Zheng, Y., & Qi, D. (2017). Deep spatio-temporal residual networks for citywide crowd flows prediction. In Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence (pp. 1655--1661). * Zhang et al. [2020d] Zhang, K., He, F., Zhang, Z., Lin, X., & Li, M. (2020d). Graph attention temporal convolutional network for traffic speed forecasting on road networks. Transportmetrica B: Transport Dynamics, (pp. 1--19). * Zhang et al. [2019b] Zhang, N., Guan, X., Cao, J., Wang, X., & Wu, H. (2019b). A hybrid traffic speed forecasting approach integrating wavelet transform and motif-based graph convolutional recurrent neural network. arXiv preprint arXiv:1904.06656, . * Zhang et al. [2020e] Zhang, Q., Chang, J., Meng, G., Xiang, S., & Pan, C. (2020e). Spatio-temporal graph structure learning for traffic forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence. volume 34. * Zhang et al. [2018b] Zhang, Q., Jin, Q., Chang, J., Xiang, S., & Pan, C. (2018b). Kernel-weighted graph convolutional network: A deep learning approach for traffic forecasting. In 2018 24th International Conference on Pattern Recognition (ICPR) (pp. 1018--1023). IEEE. * Zhang & Guo [2020] Zhang, T., & Guo, G. (2020). Graph attention lstm: A spatio-temperal approach for traffic flow forecasting. IEEE Intelligent Transportation Systems Magazine, . * Zhang et al. [2019c] Zhang, T., Jin, J., Yang, H., Guo, H., & Ma, X. (2019c). Link speed prediction for signalized urban traffic network using a hybrid deep learning approach. In 2019 IEEE Intelligent Transportation Systems Conference (ITSC) (pp. 2195--2200). IEEE. * Zhang et al. [2020f] Zhang, W., Liu, H., Liu, Y., Zhou, J., & Xiong, H. (2020f). Semi-supervised hierarchical recurrent graph neural network for city-wide parking availability prediction. In Proceedings of the AAAI Conference on Artificial Intelligence. volume 34. * Zhang et al. [2020g] Zhang, W., Liu, H., Liu, Y., Zhou, J., Xu, T., & Xiong, H. (2020g). Semi-supervised city-wide parking availability prediction via hierarchical recurrent graph neural network. IEEE Transactions on Knowledge and Data Engineering, . * Zhang et al. [2020h] Zhang, X., Huang, C., Xu, Y., & Xia, L. (2020h). Spatial-temporal convolutional graph attention networks for citywide traffic flow forecasting. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management (pp. 1853--1862). * Zhang et al. [2020i] Zhang, X., Zhang, Z., & Jin, X. (2020i). Spatial-temporal graph attention model on traffic forecasting. In 2020 13th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI) (pp. 999--1003). IEEE. * Zhang et al. [2019d] Zhang, Y., Cheng, T., & Ren, Y. (2019d). A graph deep learning method for short-term traffic forecasting on large road networks. Computer-Aided Civil and Infrastructure Engineering, 34, 877--896. * Zhang et al. [2020j] Zhang, Y., Cheng, T., Ren, Y., & Xie, K. (2020j). A novel residual graph convolution deep learning model for short-term network-based traffic forecasting. International Journal of Geographical Information Science, 34, 969--995. * Zhang et al. [2020k] Zhang, Y., Dong, X., Shang, L., Zhang, D., & Wang, D. (2020k). A multi-modal graph neural network approach to traffic risk forecasting in smart urban sensing. In 2020 17th Annual IEEE International Conference on Sensing, Communication, and Networking (SECON) (pp. 1--9). IEEE. * Zhang et al. [2020l] Zhang, Y., Lu, M., & Li, H. (2020l). Urban traffic flow forecast based on fastgcrnn. Journal of Advanced Transportation, 2020. * Zhang et al. [2019e] Zhang, Y., Wang, S., Chen, B., & Cao, J. (2019e). Gcgan: Generative adversarial nets with graph cnn for network-scale traffic prediction. In 2019 International Joint Conference on Neural Networks (IJCNN) (pp. 1--8). IEEE. * Zhang et al. [2020m] Zhang, Z., Cui, P., & Zhu, W. (2020m). Deep learning on graphs: A survey. IEEE Transactions on Knowledge and Data Engineering, . * Zhang et al. [2019f] Zhang, Z., Li, M., Lin, X., Wang, Y., & He, F. (2019f). Multistep speed prediction on traffic networks: A deep learning approach considering spatio-temporal dependencies. Transportation research part C: emerging technologies, 105, 297--322. * Zhao et al. [2020a] Zhao, B., Gao, X., Liu, J., Zhao, J., & Xu, C. (2020a). Spatiotemporal data fusion in graph convolutional networks for traffic prediction. IEEE Access, . * Zhao et al. [2020b] Zhao, H., Yang, H., Wang, Y., Wang, D., & Su, R. (2020b). Attention based graph bi-lstm networks for traffic forecasting. In 2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC) (pp. 1--6). IEEE. * Zhao et al. [2019] Zhao, L., Song, Y., Zhang, C., Liu, Y., Wang, P., Lin, T., Deng, M., & Li, H. (2019). T-gcn: A temporal graph convolutional network for traffic prediction. IEEE Transactions on Intelligent Transportation Systems, . * Zhao et al. [2021] Zhao, T., Liu, Y., Neves, L., Woodford, O., Jiang, M., & Shah, N. (2021). Data augmentation for graph neural networks. In Proceedings of the 30th International Joint Conference on Artificial Intelligence. AAAI Press. * Zheng et al. [2020a] Zheng, B., Hu, Q., Ming, L., Hu, J., Chen, L., Zheng, K., & Jensen, C. S. (2020a). Spatial-temporal demand forecasting and competitive supply via graph convolutional networks. arXiv preprint arXiv:2009.12157, . * Zheng et al. [2020b] Zheng, C., Fan, X., Wang, C., & Qi, J. (2020b). Gman: A graph multi-attention network for traffic prediction. In Proceedings of the AAAI Conference on Artificial Intelligence. volume 34. * Zhou et al. [2020a] Zhou, F., Yang, Q., Zhang, K., Trajcevski, G., Zhong, T., & Khokhar, A. (2020a). Reinforced spatio-temporal attentive graph neural networks for traffic forecasting. IEEE Internet of Things Journal, . * Zhou et al. [2020b] Zhou, F., Yang, Q., Zhong, T., Chen, D., & Zhang, N. (2020b). Variational graph neural networks for road traffic prediction in intelligent transportation systems. IEEE Transactions on Industrial Informatics, . * Zhou et al. [2020c] Zhou, J., Cui, G., Hu, S., Zhang, Z., Yang, C., Liu, Z., Wang, L., Li, C., & Sun, M. (2020c). Graph neural networks: A review of methods and applications. AI Open, 1, 57--81. * Zhou et al. [2020d] Zhou, Q., Gu, J.-J., Ling, C., Li, W.-B., Zhuang, Y., & Wang, J. (2020d). Exploiting multiple correlations among urban regions for crowd flow prediction. Journal of Computer Science and Technology, 35, 338--352. * Zhou et al. [2019] Zhou, X., Shen, Y., & Huang, L. (2019). Revisiting flow information for traffic prediction. arXiv preprint arXiv:1906.00560, . * Zhou et al. [2020e] Zhou, Z., Wang, Y., Xie, X., Chen, L., & Liu, H. (2020e). Riskoracle: A minute-level citywide traffic accident forecasting framework. In Proceedings of the AAAI Conference on Artificial Intelligence. volume 34. * Zhou et al. [2020f] Zhou, Z., Wang, Y., Xie, X., Chen, L., & Zhu, C. (2020f). Foresee urban sparse traffic accidents: A spatiotemporal multi-granularity perspective. IEEE Transactions on Knowledge and Data Engineering, . * Zhu et al. [2019] Zhu, H., Luo, Y., Liu, Q., Fan, H., Song, T., Yu, C. W., & Du, B. (2019). Multistep flow prediction on car-sharing systems: A multi-graph convolutional neural network with attention mechanism. International Journal of Software Engineering and Knowledge Engineering, 29, 1727--1740. * Zhu et al. [2020a] Zhu, H., Xie, Y., He, W., Sun, C., Zhu, K., Zhou, G., & Ma, N. (2020a). A novel traffic flow forecasting method based on rnn-gcn and brb. Journal of Advanced Transportation, 2020. * Zhu et al. [2020b] Zhu, J., Han, X., Deng, H., Tao, C., Zhao, L., Tao, L., & Li, H. (2020b). Kst-gcn: A knowledge-driven spatial-temporal graph convolutional network for traffic forecasting. arXiv preprint arXiv:2011.14992, . * Zhu et al. [2020c] Zhu, J., Tao, C., Deng, H., Zhao, L., Wang, P., Lin, T., & Li, H. (2020c). Ast-gcn: Attribute-augmented spatiotemporal graph convolutional network for traffic forecasting. arXiv preprint arXiv:2011.11004, .
# Narrow-line absorption at 689 nm in an ultracold strontium gas Fachao Hu These authors contributed equally to this work. Canzhu Tan These authors contributed equally to this work. Hefei National Laboratory for Physical Sciences at the Microscale and Shanghai Branch, University of Science and Technology of China, Shanghai 201315, China CAS Center for Excellence and Synergetic Innovation Center in Quantum Information and Quantum Physics, University of Science and Technology of China, Shanghai 201315, China Yuhai Jiang<EMAIL_ADDRESS>Shanghai Advanced Research Institute, Chinese Academy of Sciences, Shanghai 201210, China CAS Center for Excellence and Synergetic Innovation Center in Quantum Information and Quantum Physics, University of Science and Technology of China, Shanghai 201315, China Matthias Weidemüller<EMAIL_ADDRESS>Hefei National Laboratory for Physical Sciences at the Microscale and Shanghai Branch, University of Science and Technology of China, Shanghai 201315, China CAS Center for Excellence and Synergetic Innovation Center in Quantum Information and Quantum Physics, University of Science and Technology of China, Shanghai 201315, China Physikalisches Institut, Universität Heidelberg, Im Neuenheimer Feld 226, 69120 Heidelberg, Germany Bing Zhu<EMAIL_ADDRESS>Physikalisches Institut, Universität Heidelberg, Im Neuenheimer Feld 226, 69120 Heidelberg, Germany Hefei National Laboratory for Physical Sciences at the Microscale and Shanghai Branch, University of Science and Technology of China, Shanghai 201315, China CAS Center for Excellence and Synergetic Innovation Center in Quantum Information and Quantum Physics, University of Science and Technology of China, Shanghai 201315, China ###### Abstract We analyse the spectrum on the narrow-line transition $5\mathrm{s^{2}}\,^{1}\textrm{S}_{0}-5\mathrm{s}5\mathrm{p}\,^{3}\textrm{P}_{1}$ at 689 nm in an ultracold gas of 88Sr via absorption imaging. In the low saturation regime, the Doppler effect dominates in the observed spectrum giving rise to a symmetric Voigt profile. The atomic temperature and atom number can accurately be deduced from these low-saturation imaging data. At high saturation, the absorption profile becomes asymmetric due to the photon- recoil shift, which is of the same order as the natural line width. The line shape can be described by an extension of the optical Bloch equations including the photon recoil. A lensing effect of the atomic cloud induced by the dispersion of the atoms is also observed at higher atomic densities in both the low and strong saturation regimes. ## I Introduction The existence of metastable states and narrow-line transitions among the alkaline-earth and alkaline-earth-like atoms brings new opportunities for studying cold and ultracold atoms, for example the optical-lattice clocks Jun Ye (2008); Ludlow _et al._ (2015), the time variation of fundamental constants Safronova _et al._ (2018a, b); Kennedy _et al._ (2020), atom interferometersHu _et al._ (2017, 2019); Rudolph _et al._ (2020), nonlinear quantum optics Ye _et al._ (1998); Christensen _et al._ (2015); Westergaard _et al._ (2015), and strongly correlated Rydberg gases Dunning _et al._ (2016); Madjarov _et al._ (2020). While many of these applications rely on the clock transition ${}^{1}\textrm{S}_{0}-^{3}\textrm{P}_{0}$ with a linewidth on the level of mHz, the other narrow one ${}^{1}\textrm{S}_{0}-^{3}\textrm{P}_{1}$ with a $(1\sim 100)$-kHz linewidth enable the cooling of the atoms down to the photon-recoil-limited regime Curtis _et al._ (2001); Loftus _et al._ (2004); Guttridge _et al._ (2016) and a direct laser cooling to quantum degeneracy was demonstrated Stellmer _et al._ (2013). Very recently, the kHz-transitions play an essential role in the realizations of optical tweezer arrays of alkali-earth Norcia _et al._ (2018); Cooper _et al._ (2018) and alkali-earth-like Saskin _et al._ (2019) atoms. Benefiting from the narrow line width of the ${}^{1}\textrm{S}_{0}$-${}^{3}\textrm{P}_{1}$ transitions in alkali-earth or alkali-earth-like systems, fluorescence signals from these transitions can be employed for studying collective atomic scattering and motional effects Bromley _et al._ (2016), measuring atomic transition properties Ferrari _et al._ (2003); Ido _et al._ (2005); Schmitt _et al._ (2013), and detecting single atoms with high fidelities Saskin _et al._ (2019). On the other hand, absorption imaging using broad dipole-allowed transitions ($\sim$ 10 MHz) may be by far the most widely-used method in diagnosing ultracold-atom systems, providing accurate information on the spatial distribution of atoms, the atom number, and the atomic temperature Ketterle _et al._ (1999); Ketterle and Zwierlein (2008). However, absorption with narrow-line transitions were rarely studied in the ultracold regime, where the photon recoil energy is comparable to the absorption linewidth including the Doppler effect. Oates _et al._ studied the atomic-recoil-induced asymmetries in a form of saturation spectroscopy with a Ca optical-clock apparatus Oates _et al._ (2005), and the photon-recoil effect on the dispersion was observed in a Yb vapor cell in Ref. Grimm and Mlynek (1988). Stellmer _et al._ have implemented the absorption imaging on the 7.5-kHz transition at 689 nm to resolve the hyperfine structure of the fermionic 87Sr at a magnetic field of about 0.5 G Stellmer _et al._ (2011). They observed a Lorentzian lineshape with a full width at half maximum (FWHM) of about 40 kHz, without discussing further details on the spectrum. In this work, we study in detail the absorption spectrum on the narrow transition $5\mathrm{s^{2}}\,^{1}\textrm{S}_{0}-5\mathrm{s}5\mathrm{p}\,^{3}\textrm{P}_{1}$ at 689 nm with an ultracold 88Sr atomic cloud. We measure the spectrum in both the weak and strong saturation regimes. At low saturations, the absorption lineshape is close to a Gaussian shape essentially determined by the Doppler effect in the temperature range studied here. Thus, this regime can be exploited for thermometry of the atomic sample, which is confirmed by a comparison to the temperature obtained by the standard time-of-flight (TOF) method Ketterle _et al._ (1999); Ketterle and Zwierlein (2008) using the broadband transition $5\mathrm{s^{2}}\,^{1}\textrm{S}_{0}-5\mathrm{s}5\mathrm{p}\,^{1}\textrm{P}_{1}$ . The narrow-line absorption imaging at low saturation also provides information on the atom numbers and atomic densities with a comparable accuracy to detection methods based on the broad (blue) line. In the strong saturation regime, an asymmetric lineshape is observed. We have performed a theoretical simulation based on the optical Bloch equations (OBEs) involving the momentum transfers during the imaging process and confirmed that the photon recoil has important influence on the line shape. We also observe a density-dependent lensing effect in the absorption images at large detunings of the imaging light. The article is organized as follows: We show our experimental setup in Sec. II. The low- and high-saturation absorption spectra are described in Secs. III.1 and III.2, respectively. The theoretical simulation and comparison to experiments in the high-saturation regime are discussed in Sec. III.3. The observation of lensing effect is presented in Sec. IV. Sec. V concludes the paper. ## II Experimental setup Figure 1: (a) Schematic of the top view of experimental setup. HWP: half wave- plate; PBS: polarizing beam-splitter. g and B represent the gravity and magnetic field, respectively. See text for more details. (b) Time sequence for absorption imaging. See text for explanations of $\tau_{\textrm{TOF}}$ and $\tau_{\textrm{exp}}$. (c) Absorption spectrum showing all three Zeeman sublevels of ${}^{3}\textrm{P}_{1}$ state when the imaging light polarization is tuned to about 45∘ angled to the residual magnetic field. Black points are the measured peak OD, and the red curve is the fit to a multi-peak Gaussian function. The obtained Zeeman splitting is 167.7(1.2) kHz, corresponding to a magnetic field of 79.9(6) mG. Fig. 1(a) shows the experimental setup. The 88Sr atoms are first loaded into a two-stage magneto-optical trap (MOT) for the laser cooling and trapping Nosske _et al._ (2017); Qiao _et al._ (2019), operated on the broad $5\mathrm{s^{2}}\,^{1}\textrm{S}_{0}-5\mathrm{s}5\mathrm{p}\,^{1}\textrm{P}_{1}$ and narrow $5\mathrm{s^{2}}\,^{1}\textrm{S}_{0}-5\mathrm{s}5\mathrm{p}\,^{3}\textrm{P}_{1}$ transitions, respectively. We could create an atomic cloud of $10^{6}$ atoms with a density of about $10^{10}$ cm-3 and a temperature around 1 $\mu$K. A cigar-shaped optical dipole trap (ODT) formed by two horizontally propagating beams at the wavelength of 532 nm, is simultaneously switched on at the second-stage MOT. The two ODT beams both have a waist of about 60 $\mu$m and cross at an angle of 18∘. Holding atoms in the ODT for 200 ms to reach equilibrium after switching off the MOT, we obtain about $(0.5\cdots 5)\times 10^{5}$ atoms at a temperature of $0.7\cdots 6$ $\mu$K depending on the ODT power. At a power of 0.6 W for each beam the trap depth of the ODT is about $6\mu$K and the trap frequencies are $2\pi\times$(217, 34, 217) Hz along the $x$, $y$, and $z$ directions [see Fig. 1(a)], respectively, resulting in cloud radii of (27, 69, 27) $\mu$m and a peak density of $7\times 10^{11}$cm-3. The temperatures along the $y$ and $z$ directions are mapped out by the standard TOF method. The above-mentioned atom numbers, cloud sizes, and temperatures are measured using absorption imaging with the broad $5\mathrm{s^{2}}\,^{1}\textrm{S}_{0}-5\mathrm{s}5\mathrm{p}\,^{1}\textrm{P}_{1}$ transition. The lifetime of the atomic clouds in the ODT is about 2 s, limited by the collisions with background gas. The imaging light at 689 nm is delivered from a commercial tapered amplifier seeded by an external-cavity diode laser (Toptica TApro), used also for the narrow-line MOT cooling, which is frequency-stabilized to a passive ultra-low expansion cavity with a short-term noise of 1 kHz level and a long-term drift of 8 kHz/day Qiao _et al._ (2019). As shown in Fig. 1(a), the imaging beam propagates along the $z$ direction with a tunable linear polarization and has a $1/e^{2}$ diameter of 4.2 mm. The imaging pulse length and intensity are controlled by an accousto-optic modulator (not shown in the figure). The imaging system consists of two achromatic lenses with focal lengths of +200 mm and +300 mm, and maps the absorption to an EM-CCD camera from Andor with a magnification factor of 1.5. We have an imaging resolution of about 12 $\mu$m. The imaging sequence is described in Fig. 1(b). The absorption imaging on the narrow-line transition is performed after rapidly switching off the ODT to avoid the differential AC Stark shifts on the energy levels. A quantization magnetic field along the vertical direction is applied (rising time 2 ms) before the imaging pulse to split the Zeeman sublevels of ${}^{3}\textrm{P}_{1}$ state, as seen in Fig. 1(c). After a given time-of- flight (TOF) time $\tau_{\textrm{TOF}}$, the atoms are shined by the imaging light with an exposure time $\tau_{\textrm{exp}}=200$ $\mu$s. By tuning the $\tau_{\textrm{TOF}}$ we can tune the atomic density during the absorption, which plays an important role in observing the dispersive lensing effect discussed in Sec. III. As done in a standard absorption imaging sequence, two additional images with and without the imaging light are taken after the first pulse. The three images are then processed (see, e.g., Lewandowski _et al._ (2003)) to obtain the two-dimensional optical density (OD) distribution (see the insets of Fig. 5). By changing the linear imaging polarization angle in the $x-y$ plane, all three Zeeman sublevels of the ${}^{3}\textrm{P}_{1}$ state are addressable. An example is shown in Fig. 1(d). The peak OD is measured as a function of the imaging detuning showing three peaks at a magnetic field of about 80 mG. The relative line strengths are determined by the polarization and the different coupling strengths of the three corresponding transitions (see Fig. 1(d)). We have used this measurement to optimize the compensation of the background magnetic field to be better than 5 mG in our setup and to calibrate the quantization fields. For the absorption studies, we apply a field of 4 G to split the sublevels and the imaging polarization is tuned parallel to the quantization axis, so that the system is subjected only to the closed $\pi$ transition ($m_{j}=0\rightarrow m_{j^{\prime}}=0$), which can be treated as a perfect two-level system. ## III Measurements and analysis Thanks to the high sensitivity and large dynamical range of our imaging camera (Andor iXon 897) at 689 nm, we can study the absorption spectrum on the narrow-line transition with a saturation parameter $s$ ranging from 0.01 to more than 100. Meanwhile, the cloud temperature and the atomic density can be controlled via the ODT depth and the TOF time $\tau_{\textrm{TOF}}$ independently. Figure 2: (a) Low-saturation absorption spectra at temperatures of 1.3 $\mu K$ (black) and 5.7 $\mu K$ (red). The integrated absorption signal over the atomic cloud region is plotted as a function of the imaging detuning. The solid curves are fits to the Voigt profile. See text for more details. (b) Spectroscopic thermometry. The fitted Doppler widths from (a) are used to estimate the temperatures $T_{\textrm{Fit}}$, which are plotted against the TOF measurement results $T_{\textrm{TOF}}$ in the lower panel of (b). A linear fit to the data (black dashed line) gives a slope of 1.05(3). The red dashed line represents $T_{\textrm{Fit}}=T_{\textrm{TOF}}$. In the upper panel we also show the ratio (black open circles) between the atom numbers obtained from the narrow- ($N_{\textrm{red}}$) and broad-linewidth ($N_{\textrm{blue}}$) imaging, which is a constant of 1.06(2) (gray solid line). ### III.1 Low-saturation absorption In Fig. 2(a), we show two measured absorption spectra at temperatures of 1.3 $\mu$K (black points) and 5.7 $\mu$K (red points) with a saturation parameter of $s=0.1$. The TOF time $\tau_{\textrm{TOF}}$ (see Fig. 1(b)) is chosen to be 3.1 ms to minimize the lensing effect (see Sec. IV) as well as to keep large enough signal-to-noise ratios (SNRs) in the OD images. The plotted signals in Fig. 1(b) are the OD integrals over the whole atomic cloud region divided by the peak cross section $\sigma_{0}=3\lambda^{2}/2\pi$, which is the standard way to calculate the atom number in the absorption imaging (see the following paragraph for a correction). Symmetric lineshapes are observed in both cases and the linewidth increases with the increasing temperature. The spectra fit well to Voigt profiles with a fixed Lorentzian width of $v_{L}=10.01$ kHz, resulted from the power broadening $\Gamma\sqrt{1+s}/2\pi$ and the detection bandwidth $0.9/\tau_{\textrm{exp}}=4.5$ kHz due to the finite length of the square-shape imaging pulse (see Fig. 1(b)), where $\Gamma/2\pi=7.5$ kHz is the natural linewidth. The FWHM Gaussian width $v_{G}$ obtained from the Voigt profile fitting is used to deduce the temperature $T_{\textrm{Fit}}$ along the imaging propagation direction, from the relation $v_{G}=\frac{2}{\lambda}\sqrt{2\ln 2k_{b}T_{\textrm{Fit}}/m}$. Here $k_{b}$ is the Boltzmann constant, $\lambda$ is the transition wavelength, and $m$ is the atomic mass. $T_{\textrm{Fit}}$ obtained in this way are compared to those measured by the TOF method in the lower panel of Fig. 2(b). The linear fit between $T_{\textrm{Fit}}$ and $T_{\textrm{TOF}}$ (black dashed line) results in a slope of 1.05(3), which agrees excellently with the ideal case of $T_{\textrm{Fit}}=T_{\textrm{TOF}}$ (red dashed line). We also notice the empirical density broadening in the saturation fluorescence spectroscopy reported in Ido _et al._ (2005). The linear slope is only modified slightly to 1.05(6) even if we take the empirical density relation following Ref. Ido _et al._ (2005). Figure 3: High-saturation absorption. (a) The measured absorption lineshapes at low (blue dots) and high (red dots) saturations. The data in the high- saturation case ($s=35.8$) showing asymmetric profile is fitted to the numerical solution of Eq. (5) (red curve) and magnified by 6 times to have a better visualization. As a comparison, the low-saturation ($s=0.09$) data is symmetric and fits well to the Voigt profile (blue curve). (b) - (d), the population difference $\Delta\rho(p)$ obtained from the OBE solutions at there different detunings [$0,\pm 5\Gamma$, as marked by the grey vertical lines in (a)] after an exposure time of 100 $\mu$s and 200 $\mu$s, respectively. As a reference, we also show the initial distribution at $t=0$, which is the Maxwell-Boltzmann one determined by the cloud temperature. The black solid vertical lines mark the resonant momentum positions, where the probe detuning is compensated by the Doppler effect. The inset images show measurements of the two-dimensional OD distributions at low (upper) and high (lower) saturations for their respective detunings. In addition to the temperature, the atom number and atomic density can also be extracted from the narrow-linewidth absorption imaging in the low saturation regime. The broad (blue) transition typically used in determining the atom number and atomic density has a natural linewidth on the order of 10 MHz, much broader than the Doppler width. The absorption cross-section in the broad- transition imaging can hence be regarded as temperature-independent. However, for the narrow transition with a natural linewidth smaller than the Doppler width ($\Gamma/2\pi v_{G}<1$), the Doppler effect has to be considered when calculating the atom number Foot (2004). This is done by convolving the velocity-dependent Lorentzian absorption profile with the Maxwell-Boltzmann velocity distribution in the atomic sample (see Appendix Appendix). The convolution results in a relationship between the measured OD and the atom number similar to the broadband absorption imaging case, modified by a coefficient depending on the ratio between the Doppler-broadened width and the natural linewidth, $\displaystyle OD_{0}(x,y)$ $\displaystyle=n(x,y)\sigma_{0}\times C(\Gamma,v_{G})\,,$ (1) where $C(\Gamma,v_{G})=\sqrt{\pi}\alpha e^{\alpha^{2}}\textrm{Erfc}(\alpha)$ is the coefficient with $\alpha=\sqrt{\ln 2}\Gamma/2\pi v_{G}$, $OD_{0}(x,y)$ are the on-resonance OD spatial distribution, and $n(x,y)$ is the atomic column density. Erfc(x) is the complementary error function. The derivation of the coefficient is presented in the Appendix Appendix. With the on-resonance OD and the temperature-dependent $v_{G}$ determined from the spectrum fitting, the atom number and atomic density can be obtained with Eq. (1). The upper panel of Fig. 2(b) shows the ratio of the atom number determined by absorption imaging with the narrow $5\mathrm{s^{2}}\,^{1}\textrm{S}_{0}-5\mathrm{s}5\mathrm{p}\,^{3}\textrm{P}_{1}$ ($N_{\textrm{red}}$) and broad $5\mathrm{s^{2}}\,^{1}\textrm{S}_{0}-5\mathrm{s}5\mathrm{p}\,^{1}\textrm{P}_{1}$ ($N_{\textrm{blue}}$) transitions, which lies close to 1 (gray solid line). ### III.2 Strong-saturation absorption In this section we study the narrow-line absorption spectrum at strong saturations with $s\gg 1$. The question arises, to which extent the photon recoils impacts the absorption profile, as the the recoil shift is comparable to the natural linewidth ($\sim 4.78$ kHz vs. $7.5$ kHz for the narrow transition studied here). During the absorption process, absorption and spontaneous emission events give rise to the change of momentum distribution, and hence affecting the subsequent absorption. For narrow-line transitions, few photon-recoil events are enough to drive the atom out of resonance with the imaging laser. This is in stark contrast to broadband transitions, since the natural linewidth is usually much larger than the Doppler shifts induced by photon recoils in a cold atom sample. Nevertheless, the photon-recoil effect was already considered when imaging light species like Li using broad transitions Horikoshi _et al._ (2017), where the recoil-induced detuning and blurring place strong constraints on the proper imaging conditions. In Fig. 3(a) we show two absorption spectra at saturation parameters of $s=0.09$ and $s=35.8$, respectively. We observe a decrease of the integrated OD signal at all detunings due to the saturation effect (note that the data at higher saturation is magnified by a factor of 6 for a better view). More importantly, the lineshape is asymmetric at the high saturation, namely the integrated OD approaches zero more slowly on the negative-detuning side than that on the positive one, and the absorption peak is shifted by a few kHz to the positive detuning. At high saturation, differences can already be seen in the OD images at the two detuning sides [see lower rows in the insets of Figs. 3(b-d)], namely a wider spatial extension for the position detuning than that for the negative one. In this series of experiments, the influence of the lensing effect on the OD measurement(see Sec. IV) is negligible due to the low atomic densities involved here. The observed asymmetry and peak shift can be interpreted qualitatively by considering the absorption process including the influence of the photon recoil. The photon recoil associated with each absorption-spontaneous emission cycle redistributes the momentum of atoms, which depends strongly on the light detuning Stenholm (1978). Consequently, an asymmetric lineshape and the shift of the maximum of absorption emerges when more and more photons are scattered due to the momentum redistribution in the atomic cloud. In order to resolve such effects, the Doppler width has to be comparable to the power-broadened line width. In the case of the strong saturation in Fig. 3 (a), the power- broadened Lorentzian width $\Gamma\sqrt{1+s}\sim 45$ kHz is close to the Doppler one of $\sim 40$ kHz. In the following subsection a quantitative description is presented incorporating the photon-recoil effect in an OBE formalism. ### III.3 Spectrum lineshape simulation Figure 4: Absorption peak position shift. The relative position of the absorption peak at different saturation parameters $s$ are compared for imaging times of 100 $\mu$s (red circles and curve) and 200 $\mu$s (black circles and curve). The black and red curves are the calculated results without any free parameters, while the black and red circles are the fitted results from measurements by using the peak position and height as the two free fitting parameters. See text for more discussions. We take an OBE formalism including the method of so-called ’momentum families’ from Ref. Castin _et al._ (1989), originally developed to understand laser cooling on a narrow-line transition. The model considers a two-level atom system with an initial Maxwell-Boltzmann thermal distribution, interacting with a single near-resonant monochromatic homogeneous probe beam. The state of an atom with momentum $\bm{p}$ is expressed in the form of $\\{\ket{g,\bm{p}},\ket{e,\bm{p}}\\}$, where $\ket{g(e)}$ corresponds to the atomic ground (excited) state. The system Hamiltonian driven under a laser beam propagating along the $z$ axis is, $\centering H_{0}=\frac{\hat{\bm{p}}^{2}}{2m}+\hbar\omega_{0}\ket{e}\bra{e}-\hat{\bm{D}}\cdot\hat{\bm{E}}\@add@centering$ (2) where $\omega_{0},\hat{\bm{D}},\hat{\bm{E}}$ are the transition frequency, dipole moment operator, and laser electric field, respectively. In our case, only the $\pi$-transition branch $m_{j}=0\rightarrow m_{j^{\prime}}=0$ is considered, and only momentum along the light propagation axis $p=p_{z}$ is preserved, with the other two components $p_{x},p_{y}$ traced over. The system Hamiltonian under the rotating-wave approximation becomes, $\centering H_{S}=\frac{\hat{p}^{2}}{2m}-\hbar\delta\ket{e}\bra{e}+\frac{\hbar\Omega}{2}(e^{ikz}\ket{e}\bra{g}+\ket{g}\bra{e}e^{-ikz})\@add@centering$ (3) where $\delta,\Omega$ are the bare detuning and Rabi frequency. The evolution of states $\ket{g,p},\ket{e,p+\hbar k}$ with any momentum p remains globally closed under $H_{S}$ when the spontaneous emission is not considered, for which reason the states $\ket{g,p}$, $\ket{e,p+\hbar k}$ are grouped as a family $\mathcal{F}(p)$. The system density matrix $\rho$ expanded in this basis is, $\centering\begin{aligned} \rho_{gg}(p)&=\braket{g,p}{\rho}{g,p}\\\ \rho_{ee}(p)&=\braket{e,p+\hbar k}{\rho}{e,p+\hbar k}\\\ \rho_{ge}(p)&=\rho_{eg}^{*}(p)=\braket{g,p}{\rho}{e,p+\hbar k}\\\ \end{aligned}\,.\@add@centering$ (4) The equations of evolution under $H_{S}$ together with the spontaneous emission processes are, $\displaystyle\dot{\rho}_{gg}(p)$ $\displaystyle=\Gamma\bar{\pi}_{e}(p-\hbar k)-\frac{i\Omega}{2}(\rho_{eg}(p)-\rho_{ge}(p))\,,$ (5) $\displaystyle\dot{\rho}_{ee}(p)$ $\displaystyle=-\Gamma\bar{\pi}_{e}(p)+\frac{i\Omega}{2}(\rho_{eg}(p)-\rho_{ge}(p))\,,$ $\displaystyle\dot{\rho}_{ge}(p)$ $\displaystyle=\dot{\rho}_{eg}^{*}(p)$ $\displaystyle=-(i(\bar{\delta}-\frac{kp}{m})+\frac{\Gamma}{2})\rho_{ge}(p)+\frac{i\Omega}{2}(\rho_{gg}(p)-\rho_{ee}(p))\,,$ where $\bar{\delta}=\delta-\hbar k^{2}/(2m)$ and the term $\bar{\pi}_{e}$ represents the impact of spontaneous decay on the system evolution, defined as $\displaystyle\bar{\pi}_{e}(p)=$ $\displaystyle\int\limits_{-\infty}^{+\infty}dp_{x}\int\limits_{-\infty}^{+\infty}dp_{y}\int\limits_{-\hbar k}^{+\hbar k}dp^{\prime}\mathcal{N}(p^{\prime})$ (6) $\displaystyle\braket{e,p_{x},p_{y},p_{z}=p+p^{\prime}}{\rho}{e,p_{x},p_{y},p_{z}=p+p^{\prime}}\,.$ Here $\mathcal{N}(p^{\prime})=\frac{3}{4\hbar k}(1-p^{\prime 2}/\hbar^{2}k^{2})$ results from the classical dipole radiation pattern Castin _et al._ (1989) of the $\pi$ transition. With all the atoms initially at the ground state $\ket{g}$ with a Maxwell-Boltzmann distribution of temperature $T$, we numerically integrate the equations (5) to get the system evolution. The solution of the off-diagonal elements $\rho_{eg}(p)$ results in the susceptibility $\chi(p)\propto n\rho_{eg}(p)$ with the atomic density $n$. The absorption profile is then calculated by tracing the imaginary part of the susceptibility over all momenta, i.e. $\sum_{p}\mathrm{Im}\chi(p)$, and then integrating over the interaction duration. For the solid curves in Figs. 3(a), we fit the experimental data to the calculated profiles with the maximum integrated OD and the peak position as the only free parameters. Both the lineshape asymmetry and the shift of the absorption peak at high saturation can be reproduced very well by Eq. (5) including the momentum transfer due to the photon-scattering events. While the model predicts a significant shift of the absorption peak, its position is still used as a free parameter in the fits to account for the possible deviation between the measurements and the calculations, as discussed in more details in Fig. 4. One can gain further insight into the photon-recoil effects by considering the quasi-steady solution of the off-diagonal element in Eq. (5), $\mathrm{Im\rho_{eg}(p)}\propto\Delta\rho(p)=\rho_{gg}(p)-\rho_{ee}(p)$. We show from Fig. 3(b) to 3(d) the calculated distribution of the population difference $\Delta\rho(p)$ at two saturation parameters of $s\approx 0.09$ (blue curves) and $s=35.8$ (red curves) after 200-$\mu$s atom-light interaction time (about $10/\Gamma$, the imaging pulse length in this measurement), when the probe laser is detuned by $-5\Gamma,0,+5\Gamma$ from left to right. At the low saturation ($s\approx 0.09$), $\Delta\rho(p)$ is only slightly modified compared to the initial Maxwell-Boltzmann distribution (black dot-dashed lines), remaining almost Gaussian even after long interaction time, such that the convolution between the velocity-dependent Lorentzian profile. The momentum distribution results in a lineshape nearly the Voigt one, as the blue curve seen in Fig. 3(a). When highly saturated ($s=35.8$), however, the $\Delta\rho(p)$ distribution is strongly modified and depleted near the resonant momentum (marked by vertical dashed lines) where the Doppler shift compensates the bare imaging detuning. In Fig. 3(b) with a detuning of $-5\Gamma$, the distribution maintains a Gaussian shape with the center shifted by $\sim 1.2\hbar k$ after 200 $\mu$s. While at a detuning of $+5\Gamma$ in Fig. 3(d), two peaks appear on the opposite sides of the resonant momentum. Such a strong dependence on the detuning leads to the observed asymmetric lineshape and the peak shift. The effects of the photon recoil can also be revealed by studying the time evolution of the momentum distribution. In Figs. 3(b-d) the $\Delta\rho(p)$ at $s=35.8$ after 100-$\mu$s interaction (red dash-dot curves) are shown as a comparison to the 200-$\mu$s case. Small but clear differences of $\Delta\rho(p)$ are observed for all three detunings indicating that the momentum distribution undergoes some time evolution, which may result in a time-dependent absorption lineshape. This is actually demonstrated in Fig. 4 by comparing the saturation-dependent shift of the absorption peak position for the 100- and 200-$\mu$s imaging durations. The peak position is shifted towards the positive detuning when increasing the imaging intensity and such a shift becomes larger in the case of a longer exposure, i.e. more photons are scattered. The solid curves represent the calculated results without any free parameters, while the solid dots are from fits with the peak position and height as the free fitting parameters [see Fig. 3(a)]. Overall, the fitted shifts agree well with the calculations without free parameters, while deviations are seen for some points coming from fluctuations of experimental conditions like laser power and atom number, as well as the low SNR for large saturation parameters. ## IV The lensing effect Figure 5: Absorption spectrum with $s=17$ at two different atomic densities of $8.9\times 10^{10}$ cm-3 (red circles) and $2.8\times 10^{11}$ cm-3 (blue diamonds). The red curve is a fit to the numerical solution of Eq. (5). We obtain negative peak ODs at some large positive detunings. In the right inset, the lower OD image measured at a large positive detuning with the high atomic density has a dark hole instead of a bright peak in the cloud center, caused by the lensing effect. At the large negative detuning, the dark position appears at the edges of the cloud (left inset). As a comparison, we also show an example of the OD images for the low-density case with a normal Gaussian distribution in the upper panel of the inset. As shown in Fig. 5, we have also experimentally observed another phenomenon in the absorption spectrum at high atomic densities, the so-called lensing effect which is well known in standard absorption imaging. The absorption spectra at two different atomic densities are compared at a saturation of $s=17$. In the low-density case ($n\sim 8.9\times 10^{10}$ cm-3, red dots in the figure), we find a similar asymmetry as that in Fig. 3(a) for the high saturation. With a 3-fold higher density ($n\sim 2.8\times 10^{11}$ cm-3, blue diamonds in the figure), a negative peak OD is obtained from the two-dimension Gaussian fit at some large positive detunings. Checking the OD images there (one example shown as the right inset in Fig. 5), a dark hole instead of a bright peak is seen at the central region of the atomic cloud for the large positive detuning, while at the negative one a dark edge is observed. This phenomonon is related to the microscopic lensing effect studied in e.g. Refs. Labeyrie _et al._ (2003); Wang and Saffman (2004); Labeyrie _et al._ (2007); Roof _et al._ (2015); Han _et al._ (2015); Noaman _et al._ (2018); Gilbert _et al._ (2018), where a spatial-dependent index of refraction leads to a focusing or defocusing effect on the imaging beam depending on the detuning. The observed lensing effect can be understood from the following equation for describing the phase shift of the imaging field in the transverse plane propagating through a cloud of two-level atoms Labeyrie _et al._ (2007), $d\phi(x,y)=-\sigma_{0}n(x,y,z)dz\frac{\delta/\Gamma}{1+4(\delta/\Gamma)^{2}+s(x,y)}\,.$ (7) Here $dz$ is the thickness of the atomic cloud along the light propagation direction, $\delta$ is the detuning, and $s(r)$ has a spatial dependence due to the intensity distribution of a Gaussian probe beam. Spatial inhomogeneity of the index of refraction can be induced by the spatial distribution of the atomic density, or the probe intensity, or both. For negative (positive) detuning, Eq. (7) leads to a focusing (defocusing) of the imaging beam. The observed lensing effect here mainly stems from the density inhomogeneity as indicated by the density-dependence (see Fig. 5) and the fact that the imaging beam is much larger than the atomic cloud ($\sim 200$ times). The lensing induced by such densitiy inhomogeneity was observed in both the weak- Roof _et al._ (2015) and strong-saturation Labeyrie _et al._ (2003); Wang and Saffman (2004); Labeyrie _et al._ (2007) regimes. The lensing effect shown in Fig. 5 with strong saturation is also observable in the weak-probe case in our experiment. However, to quantitatively explain our observation, detailed calculations on the light propagation are needed like in Refs. Han _et al._ (2015); Gilbert _et al._ (2018), even including the atom dipolar interactions or multiple scattering events (e.g. Bromley _et al._ (2016); Zhu _et al._ (2016); Chabé _et al._ (2014)), which is beyond the scope of this paper. ## V Conclusion In conclusion, we have studied both experimentally and theoretically the absorption spectrum of a narrow-line transition at 689 nm in an ultracold 88Sr gas. The atomic cloud temperature down to 1 $\mu$K can be inferred from the measured absorption lineshape at low probe saturations ($s\ll 1$) if the Doppler width dominates over other line-broadening effects. Information on the atom number can also be reliably extracted from the low-saturation absorption. In the strongly saturated regime, we observed the photon-recoil-induced asymmetry in the absorption spectrum, which can be described by two-level OBEs involving the photon recoils. We also showed a lensing effect when probing a high-density sample, which is due to the spatial-dependent dispersive response of the atomic cloud to the imaging field. It is of strong interest in studying further the weak-probe high-density regime because of the collective and cooperative effects that are predicted theoretically Bienaimé _et al._ (2013); Zhu _et al._ (2016); Kupriyanov _et al._ (2017); Bettles _et al._ (2020). The narrow-line absorption can also be employed as sensitive probe for other cold atom systems with similar narrow-line transitions, like, e.g., Yb. The good resolution also makes the narrow-line absorption applicable to detection of interactions in more complicated systems, e.g. the spatial correlation Günter _et al._ (2012) due to Rydberg blockade. ## Acknowledgements We acknowledge C. Qiao, L. Couturier, and I. Nosske for their contributions on setting up the experiment at the early stage of project. F.H acknowledges Yaxiong Liu for helpful discussions on numerical algorithms. M.W.’s research activities in China are supported by the 1000-Talent-Program. The work was supported by the National Natural Science Foundation of China (Grant Nos. 11574290 and 11604324) and Shanghai Natural Science Foundation (Grant No. 18ZR1443800). Y.H.J. also acknowledges support under Grant No. 11827806. ## Appendix The low-saturation ($s\ll 1$) OD spatial distribution is represented as, $\centering OD(x,y)=\int_{-\infty}^{+\infty}\sigma_{0}n(x,y)f(v)L(\delta,v,\Gamma)dv\@add@centering$ (8) where $L(\delta,v,\Gamma)=\frac{\Gamma^{2}/4}{(\delta-kv)^{2}+\Gamma^{2}/4}$ is the Lorentzian profile with $\delta$ the bare laser detuning, $\Gamma$ the natural linewidth, $k=2\pi/\lambda$ the laser wavenumber, and $v$ the atom velocity, $f(v)=\frac{1}{u\sqrt{\pi}}e^{-v^{2}/u^{2}}$ is the Gaussian velocity distribution with $u=\sqrt{2k_{B}T/m}$ the most probable speed. The Doppler width $v_{G}$ is related to $u$, $v_{G}=ku\sqrt{\ln 2}/\pi$. Then Eq. (8) reads $\displaystyle OD(x,y)$ $\displaystyle=\sigma_{0}n(x,y)\int_{-\infty}^{+\infty}\frac{1}{u\sqrt{\pi}}e^{-(v/u)^{2}}\frac{\Gamma^{2}/4}{(\delta- kv)^{2}+\Gamma^{2}/4}dv$ (9) $\displaystyle=\sigma_{0}n(x,y)\frac{\alpha^{2}}{\sqrt{\pi}}\int_{-\infty}^{+\infty}\frac{e^{-(x^{\prime}+\delta/ku)^{2}}}{x^{\prime 2}+\alpha^{2}}dx^{\prime}$ The substitution $x^{\prime}=kv-\delta$ is used in the second step. Here $\alpha=\frac{\sqrt{\ln 2}\Gamma}{2\pi v_{G}}$ represents the ratio between the natural linewidth and the Doppler width. At the on-resonance condition ($\delta=0$) we have the Eq. (1). The coefficients $C(\Gamma,v_{G})=\sqrt{\pi}\alpha e^{\alpha^{2}}\textrm{Erfc}(\alpha)$ for correcting the on-resonance absorption cross section are plotted in Fig. for the $5\mathrm{s^{2}}\,^{1}\textrm{S}_{0}-5\mathrm{s}5\mathrm{p}\,^{1}\textrm{P}_{1}$ (black dashed line), $5\mathrm{s^{2}}\,^{1}\textrm{S}_{0}-5\mathrm{s}5\mathrm{p}\,^{3}\textrm{P}_{1}$ (blue dashed line) transitions of 88Sr and the D2 transition of 87Rb (red dotted line) as a comparison. Figure 6: The coefficient for correcting the on-resonance absorption cross section due to the Doppler effect. The plotted temperature range is $0.01-10$ $\mu$K. Three atomic transitions are compared: the broad $5\mathrm{s^{2}}\,^{1}\textrm{S}_{0}-5\mathrm{s}5\mathrm{p}\,^{1}\textrm{P}_{1}$ (black dashed line) and narrow $5\mathrm{s^{2}}\,^{1}\textrm{S}_{0}-5\mathrm{s}5\mathrm{p}\,^{3}\textrm{P}_{1}$ (blue dashed line) transitions in 88Sr, and the D2 line of 87Rb (red dotted line). This coefficient is 1 for broad transitions ($\Gamma\gg 2\pi v_{G}$) and strongly modified for narrow ones ($\Gamma\lesssim 2\pi v_{G}$) in the ultracold range. ## References * Jun Ye (2008) H. K. Jun Ye, H. J. Kimble, Science (2008), 10.1126/science.1148259. * Ludlow _et al._ (2015) A. D. Ludlow, M. M. Boyd, J. Ye, E. Peik, and P. Schmidt, Reviews of Modern Physics 87, 637 (2015). * Safronova _et al._ (2018a) M. Safronova, D. Budker, D. DeMille, D. F. J. Kimball, A. Derevianko, and C. W. Clark, Reviews of Modern Physics 90 (2018a), 10.1103/RevModPhys.90.025008. * Safronova _et al._ (2018b) M. S. Safronova, S. G. Porsev, C. Sanner, and J. Ye, Physical Review Letters 120 (2018b), 10.1103/physrevlett.120.173001. * Kennedy _et al._ (2020) C. J. Kennedy, E. Oelker, J. M. Robinson, T. Bothwell, D. Kedar, W. R. Milner, G. E. Marti, A. Derevianko, and J. Ye, Physical Review Letters 125 (2020), 10.1103/physrevlett.125.201302. * Hu _et al._ (2017) L. Hu, N. Poli, L. Salvi, and G. M. Tino, Physical Review Letters 119 (2017), 10.1103/physrevlett.119.263601. * Hu _et al._ (2019) L. Hu, E. Wang, L. Salvi, J. N. Tinsley, G. M. Tino, and N. Poli, Classical and Quantum Gravity 37, 014001 (2019). * Rudolph _et al._ (2020) J. Rudolph, T. Wilkason, M. Nantel, H. Swan, C. M. Holland, Y. Jiang, B. E. Garber, S. P. Carman, and J. M. Hogan, Physical Review Letters 124 (2020), 10.1103/physrevlett.124.083604. * Ye _et al._ (1998) J. Ye, L.-S. Ma, and J. L. Hall, Journal of the Optical Society of America B 15, 6 (1998). * Christensen _et al._ (2015) B. T. R. Christensen, M. R. Henriksen, S. A. Schäffer, P. G. Westergaard, D. Tieri, J. Ye, M. J. Holland, and J. W. Thomsen, Physical Review A 92 (2015), 10.1103/physreva.92.053820. * Westergaard _et al._ (2015) P. G. Westergaard, B. T. Christensen, D. Tieri, R. Matin, J. Cooper, M. Holland, J. Ye, and J. W. Thomsen, Physical Review Letters 114 (2015), 10.1103/physrevlett.114.093002. * Dunning _et al._ (2016) F. B. Dunning, T. C. Killian, S. Yoshida, and J. Burgdörfer, Journal of Physics B: Atomic, Molecular and Optical Physics 49, 112003 (2016). * Madjarov _et al._ (2020) I. S. Madjarov, J. P. Covey, A. L. Shaw, J. Choi, A. Kale, A. Cooper, H. Pichler, V. Schkolnik, J. R. Williams, and M. Endres, Nature Physics (2020), 10.1038/s41567-020-0903-z. * Curtis _et al._ (2001) E. A. Curtis, C. W. Oates, and L. Hollberg, Phys. Rev. A 64, 031403 (2001). * Loftus _et al._ (2004) T. H. Loftus, T. Ido, A. D. Ludlow, M. M. Boyd, and J. Ye, Physical Review Letters 93 (2004), 10.1103/physrevlett.93.073003. * Guttridge _et al._ (2016) A. Guttridge, S. A. Hopkins, S. L. Kemp, D. Boddy, R. Freytag, M. P. A. Jones, M. R. Tarbutt, E. A. Hinds, and S. L. Cornish, Journal of Physics B: Atomic, Molecular and Optical Physics 49, 145006 (2016). * Stellmer _et al._ (2013) S. Stellmer, B. Pasquiou, R. Grimm, and F. Schreck, Physical Review Letters 110 (2013), 10.1103/physrevlett.110.263003. * Norcia _et al._ (2018) M. A. Norcia, A. W. Young, and A. M. Kaufman, Physical Review X 8, 041054 (2018). * Cooper _et al._ (2018) A. Cooper, J. P. Covey, I. S. Madjarov, S. G. Porsev, M. S. Safronova, and M. Endres, Physical Review X 8 (2018), 10.1103/physrevx.8.041055. * Saskin _et al._ (2019) S. Saskin, J. Wilson, B. Grinkemeyer, and J. Thompson, Physical Review Letters 122 (2019), 10.1103/physrevlett.122.143002. * Bromley _et al._ (2016) S. L. Bromley, B. Zhu, M. Bishof, X. Zhang, T. Bothwell, J. Schachenmayer, T. L. Nicholson, R. Kaiser, S. F. Yelin, M. D. Lukin, A. M. Rey, and J. Ye, Nature Communications 7 (2016), 10.1038/ncomms11039. * Ferrari _et al._ (2003) G. Ferrari, P. Cancio, R. Drullinger, G. Giusfredi, N. Poli, M. Prevedelli, C. Toninelli, and G. M. Tino, Physical Review Letters 91 (2003), 10.1103/physrevlett.91.243002. * Ido _et al._ (2005) T. Ido, T. H. Loftus, M. M. Boyd, A. D. Ludlow, K. W. Holman, and J. Ye, Physical Review Letters 94 (2005), 10.1103/physrevlett.94.153001. * Schmitt _et al._ (2013) M. Schmitt, E. A. L. Henn, J. Billy, H. Kadau, T. Maier, A. Griesmaier, and T. Pfau, Opt. Lett. 38, 637 (2013). * Ketterle _et al._ (1999) W. Ketterle, D. S. Durfee, and D. Stamper-Kurn, arXiv preprint cond-mat/9904034 (1999). * Ketterle and Zwierlein (2008) W. Ketterle and M. W. Zwierlein, arXiv preprint arXiv:0801.2500 (2008). * Oates _et al._ (2005) C. Oates, G. Wilpers, and L. Hollberg, Physical Review A 71 (2005), 10.1103/physreva.71.023404. * Grimm and Mlynek (1988) R. Grimm and J. Mlynek, Physical Review Letters 61, 2308 (1988). * Stellmer _et al._ (2011) S. Stellmer, R. Grimm, and F. Schreck, Physical Review A 84 (2011), 10.1103/physreva.84.043611. * Nosske _et al._ (2017) I. Nosske, L. Couturier, F. Hu, C. Tan, C. Qiao, J. Blume, Y. H. Jiang, P. Chen, and M. Weidemüller, Physical Review A 96 (2017), 10.1103/physreva.96.053415. * Qiao _et al._ (2019) C. Qiao, C. Z. Tan, F. C. Hu, L. Couturier, I. Nosske, P. Chen, Y. H. Jiang, B. Zhu, and M. Weidemüller, Applied Physics B 125 (2019), 10.1007/s00340-019-7328-3. * Lewandowski _et al._ (2003) H. J. Lewandowski, D. Harber, D. L. Whitaker, and E. A. Cornell, Journal of low temperature physics 132, 309 (2003). * Foot (2004) C. Foot, _Atomic Physics_ (Oxford University Press, 2004). * Horikoshi _et al._ (2017) M. Horikoshi, A. Ito, T. Ikemachi, Y. Aratake, M. Kuwata-Gonokami, and M. Koashi, Journal of the Physical Society of Japan 86, 104301 (2017). * Stenholm (1978) S. Stenholm, Applied Physics 15, 287 (1978). * Castin _et al._ (1989) Y. Castin, H. Wallis, and J. Dalibard, Journal of the Optical Society of America B 6, 2046 (1989). * Labeyrie _et al._ (2003) G. Labeyrie, T. Ackemann, B. Klappauf, M. Pesch, G. Lippi, and R. Kaiser, The European Physical Journal D-Atomic, Molecular, Optical and Plasma Physics 22, 473 (2003). * Wang and Saffman (2004) Y. Wang and M. Saffman, Phys. Rev. A 70, 013801 (2004). * Labeyrie _et al._ (2007) G. Labeyrie, G. Gattobigio, T. Chanelière, G. Lippi, T. Ackemann, and R. Kaiser, The European Physical Journal D 41, 337 (2007). * Roof _et al._ (2015) S. Roof, K. Kemp, M. Havey, I. M. Sokolov, and D. V. Kupriyanov, Opt. Lett. 40, 1137 (2015). * Han _et al._ (2015) J. Han, T. Vogt, M. Manjappa, R. Guo, M. Kiffner, and W. Li, Physical Review A 92 (2015), 10.1103/physreva.92.063824. * Noaman _et al._ (2018) M. Noaman, M. Langbecker, and P. Windpassinger, Opt. Lett. 43, 3925 (2018). * Gilbert _et al._ (2018) J. R. Gilbert, C. P. Roberts, and J. L. Roberts, J. Opt. Soc. Am. B 35, 718 (2018). * Zhu _et al._ (2016) B. Zhu, J. Cooper, J. Ye, and A. M. Rey, Physical Review A 94 (2016), 10.1103/physreva.94.023612. * Chabé _et al._ (2014) J. Chabé, M.-T. Rouabah, L. Bellando, T. Bienaimé, N. Piovella, R. Bachelard, and R. Kaiser, Physical Review A 89 (2014), 10.1103/physreva.89.043833. * Bienaimé _et al._ (2013) T. Bienaimé, R. Bachelard, N. Piovella, and R. Kaiser, Fortschritte der Physik 61, 377 (2013). * Kupriyanov _et al._ (2017) D. Kupriyanov, I. Sokolov, and M. Havey, Physics Reports 671, 1 (2017). * Bettles _et al._ (2020) R. J. Bettles, M. D. Lee, S. A. Gardiner, and J. Ruostekoski, Communications Physics 3, 1 (2020). * Günter _et al._ (2012) G. Günter, M. R. de Saint-Vincent, H. Schempp, C. S. Hofmann, S. Whitlock, and M. Weidemüller, Physical Review Letters 108 (2012), 10.1103/physrevlett.108.013002.
# A Generative Model of Galactic Dust Emission Using Variational Inference Ben Thorne,1 Lloyd Knox,1 and Karthik Prabhu1 1Department of Physics, University of California, One Shields Avenue, Davis, CA 95616, USA E-mail<EMAIL_ADDRESS> (Accepted XXX. Received YYY; in original form ZZZ) ###### Abstract Emission from the interstellar medium can be a significant contaminant of measurements of the intensity and polarization of the cosmic microwave background (CMB). For planning CMB observations, and for optimizing foreground-cleaning algorithms, a description of the statistical properties of such emission can be helpful. Here we examine a machine learning approach to inferring the statistical properties of dust from either observational data or physics-based simulations. In particular, we apply a type of neural network called a Variational Auto Encoder (VAE) to maps of the intensity of emission from interstellar dust as inferred from Planck sky maps and demonstrate its ability to a) simulate new samples with similar summary statistics as the training set, b) provide fits to emission maps withheld from the training set, and c) produce constrained realizations. We find VAEs are easier to train than another popular architecture: that of Generative Adversarial Networks (GANs), and are better-suited for use in Bayesian inference. ###### keywords: cosmology: cosmic microwave background – ISM: general – methods:statistical ††pubyear: 2021††pagerange: A Generative Model of Galactic Dust Emission Using Variational Inference–A Generative Model of Galactic Dust Emission Using Variational Inference ## 1 Introduction Among the many research enterprises stimulated by the detection of large-scale anisotropies in the cosmic microwave background (CMB) by the COsmic Background Explorer (COBE) with its Differential Microwave Radiometer (Smoot et al., 1992), is the hunt for signatures of primordial gravitational waves (PGW). To date, only upper limits have been set, most commonly expressed as limits on the ratio of primordial tensor perturbation power to scalar perturbation power, $r$. Soon after the COBE detection it was realized that reliably detecting levels below $r\simeq 0.1$ could not be done with temperature anisotropies alone (Knox & Turner, 1994), and that proceeding further would require highly sensitive measurements of the polarization of the CMB on angular scales of about a degree, or larger (Kamionkowski et al., 1997; Seljak & Zaldarriaga, 1997). Polarized emission from the interstellar medium of the Milky Way, in the cleanest parts of the sky at the cleanest observing frequencies, is comparable to the cosmic microwave background signal generated by PGWs if the PGW signal is near the current 95% confidence upper limit of $r<0.06$ (BICEP2 Collaboration et al., 2018). So-called Stage III CMB experiments, such as the Simons Observatory (Ade et al., 2019), and BICEP Array (Hui et al., 2018) combined with SPT-3G (Benson et al., 2014) are designed to have sufficient sensitivity and systematic error control to tighten the 95% confidence upper limits by a factor of about 20. The Stage IV experiments LiteBIRD and CMB-S4 are targeting upper limits factors of 2 and 5 times more stringent still, respectively. Thus we are rapidly moving into a regime where the foreground contamination is up to two orders of magnitude larger111This is for fluctuation power. The rms level of contamination in the map is up to one order of magnitude larger than the signal of interest. than the signal of interest. The most exciting possibility is that there will be a detection of PGW, as opposed to improved upper limits. A detection claim would essentially be a claim that there is power remaining in the map that cannot be explained as a residual instrumental systematic or residual foreground emission. Detection, therefore, requires not only foreground cleaning, but the capability to quantify the probability distribution of residual foreground power. Such capability is hampered by our lack of prior knowledge of the probability distribution of the non-Gaussian and non-isotropic galactic foreground emission. The state of the art in analysis of such observations either implicitly or explicitly has the galactic emission, or their residuals, modeled as Gaussian isotropic fields (Planck Collaboration et al., 2020; Aiola et al., 2020; BICEP2 Collaboration et al., 2018). They are modeled as such not because they are, but strictly for convenience. At the very least, we need sufficient simulations of galactic emission to test such algorithms for bias. A more ambitious objective is to abandon assumptions of Gaussianity and isotropy altogether, and perform a complete Bayesian analysis with incorporation of an appropriate prior for the spatial distribution of interstellar emission. Groundbreaking progress toward such a Bayesian analysis has been made recently, with the development of analysis methodologies by Millea et al. (2020a), and the recent application to real data (Millea et al., 2020b). The analysis framework in Millea et al. (2020a) was developed for “de-lensing” of the CMB; i.e., taking into account the impact of gravitational lensing on the statistical properties of CMB polarization. Although it has not been applied to multi-frequency data, or used for foreground cleaning, at a conceptual level the framework can be straightforwardly extended to analysis of foreground-contaminated multi-frequency data. Although this extension could be implemented with isotropic Gaussian priors for foreground emission, it also presents the opportunity to incorporate more realistic priors – priors that more accurately reflect what we know about such emission from other data, or from physics-based simulations. We are thus interested in both creating simulated maps of galactic emission with the appropriate statistical properties for testing analysis algorithms to be used on real data, and also in learning, from other data and perhaps physical modeling (e.g. MHD simulations of the interstellar medium (Kim et al., 2019)) the statistical properties of maps of galactic emission for use in Bayesian inference engines. Here we report on progress toward accomplishing both of these tasks with the use of neural networks. Aylor et al. (2019) studied the use of generative adversarial networks (GANs) for learning how to simulate new emission maps with statistic properties similar to those from a training set, whilst Krachmalnicoff & Puglisi (2020) trained to simulate non-Gaussian small-scale polarized dust emission. Here we present a similar study, this time using a different neural network architecture and training program, that of variational auto encoders (VAEs). VAEs and GANs are examples of deep generative models. These models have had recent success in accurately modeling complicated, high-dimensional, datasets, and generating realistic novel samples (Razavi et al., 2019; van den Oord et al., 2016b; Brock et al., 2018). Generative models can be divided into two main categories: likelihood-based models that seek to optimize the log likelihood of the data, these include the VAE (Kingma & Welling, 2013; Jimenez Rezende et al., 2014), _flow_ based methods (Dinh et al., 2014, 2016; Jimenez Rezende & Mohamed, 2015; Kingma & Dhariwal, 2018), and _autoregressive_ models (van den Oord et al., 2016a); and implicit models, such as GANs (Goodfellow et al., 2014), which train a generator and discriminator in an adversarial game scenario. There are many trade-offs to consider when selecting a likelihood- based approach (Kingma & Dhariwal, 2018), but here we choose to explore the use of VAEs due to their simplicity and computational scalability to higher resolution datasets. We find some advantages of VAEs over GANs. The adversarial training process does not produce an explicit inference model, and it is hard to consistently compare model performance against some test set. Furthermore, it is also a common problem that samples from GANs do not represent the full diversity of the underlying distribution (Grover et al., 2017). In contrast, VAEs optimize the log likelihood of the data. This means both that it is possible to directly compare models, and trained models should support the entire dataset, which is crucial when applying a trained model to real data. VAEs also tend to be easier to train in that training success is more stable to variation of hyperparameters. As a downside, VAEs are well known for loss of resolution. We see this in our results and discuss adaptations one could make to avoid this degradation of angular resolution. Although our work is motivated by the PGW-driven desire to understand the statistical properties of polarized foreground emission, in this paper, as was the case in Aylor et al. (2019), we restrict ourselves to intensity. Observations of polarized dust emission with high signal-to-noise over a large fraction of sky do not currently exist, which precludes the training of similar models on real data. However, in ongoing work, we are exploring the use of magnetohydrodynamical (MHD) (Kim et al., 2019) simulations to train generative models of polarized emission. In this scenario a trained model would provide a ‘compression’ of the information available in MHD simulations into a single statistical model, which could then be used either in inference, or to augment real low-resolution observations with physically-motivated small-scale realizations. The rest of this paper is structured as follows. In Section 2 we introduce variational autoencoders, and the objective for their optimization. We then describe the network architecture we used, the training dataset we produced to train the network, and how hyperparameter values were set. In Section 3 we present the results of applying the trained VAE to test set images. Finally, in Section 4 we summarize our findings and discuss areas of current and future work. ## 2 Variational Autoencoders In this Section we will introduce the idea of variational autoencoders, the specific model we implement, and the details of how we train that model. Our goal here is to take a set of images of thermal emission from interstellar dust $\mathbf{x}^{(i)}=(x_{1}^{(i)},\dots,x_{N}^{(i)})\in\mathbb{R}^{N}$, and infer from them an underlying distribution, $p(x)$ from which they could have been drawn, using the techniques of _generative modeling_. Variational autoencoders are a type of generative machine learning model, which provide a framework by which we may infer the parameters of a joint distribution over our original data, and some _latent variables_ , $\mathbf{z}$, representing the unobserved part of the model. We can factorize the joint distribution of the data and latent variables into two terms representing the generative process of the data, and the latent space, responsible for the variance in the observed data: $p(\mathbf{x},\mathbf{z})=\underbrace{p(\mathbf{x}|\mathbf{z})}_{{\rm Generative}}\underbrace{p(z)}_{{\rm Variance}}.$ (1) The VAE approach is to model the conditional distribution with an appropriate family of functions with some unknown weights, $\theta$: $p_{\theta}(\mathbf{x}|\mathbf{z})\approx p(\mathbf{x}|\mathbf{z})$. This conditional model encodes the generative process by which $\mathbf{x}$ depends on the latent set of variables $\mathbf{z}$. The choice of $p(\mathbf{z})$ can then be a simple, perhaps Gaussian, prior probability distribution $p(\mathbf{z})$, which encodes the dataset variation in a simple latent space. This can be seen as a type of regularization by which we separate out different sources of variation within the dataset, a process that is quite natural for physical processes, and often makes the resulting model interpretable. The goal of training is thus to find a transformation that delivers an acceptable approximation $p_{\theta}(\mathbf{x})\approx p(\mathbf{x})$, that is optimal (in some sense), given the training set data. Toward that end we consider the parametrized joint distribution of $\mathbf{x}$ and $\mathbf{z}$: $p_{\theta}(\mathbf{x},\mathbf{z})=p_{\theta}(\mathbf{x}|\mathbf{z})p(\mathbf{z}),$ (2) which leads to our object of interest via marginalization over $z$: $p_{\theta}(\mathbf{x})=\int d\mathbf{z}~{}p_{\theta}(\mathbf{x},\mathbf{z}).$ (3) Our tasks are thus to choose a parameterization – this is referred to as a choice of _architecture_ – and then find a means of optimizing these parameters $\theta$ with resepect to a chosen _objective_ , via a process referred to as _training_. ### 2.1 Objective In principle we could determine $\theta$ by maximizing the training set’s joint likelihood $\Pi_{i}p_{\theta}(\mathbf{x}^{i})$. In practice, however, this would involve evaluating the integral in Equation 3 for each datapoint individually, which is intractable for even moderately high-dimensional latent spaces. The VAE framework provides an objective function that bounds the maximum likelihood value, and is computationally tractable. Let a dataset $\mathcal{D}$ be made up of samples $\mathbf{x}^{(i)}=(x_{1}^{(i)},\dots,x_{N}^{(i)})\in\mathbb{R}^{N}$, which we will assume to be independent and identically distributed samples from some true underlying distribution $p_{\mathcal{D}}(\mathbf{x})$. Absent an analytical model for $p_{\mathcal{D}}(\mathbf{x})$, we can instead take it to be a member of an expressive family of functions parametrized by $\bm{\theta}$: $p_{\mathcal{D}}(\mathbf{x})=p_{\bm{\theta}}(\mathbf{x})$. This can be done by introducing an unobserved set of latent variables, $\mathbf{z}=(z_{1},\dots,z_{d})\in\mathbb{R}^{d}$, and considering the joint distribution $p(\mathbf{x},\mathbf{z})$. This joint distribution is specified by: the prior over the latent space, $p(\mathbf{z})$, which is assumed to be some simple distribution (typically Gaussian); and the conditional distribution $p(\mathbf{x}|\mathbf{z})$, which is intended to represent most of the complexity in the true underlying distribution $p_{\mathcal{D}}(\mathbf{x})$. We model this distribution as a neural network with weights $\theta$: $p_{\theta}(\mathbf{x}|\mathbf{z})$. The marginal likelihood is then: $p_{\theta}(\mathbf{x})=\int d\mathbf{z}~{}p(\mathbf{z})p_{\theta}(\mathbf{x}|\mathbf{z})=\mathbb{E}_{p(\mathbf{z})}\left[p_{\theta}(\mathbf{x}|\mathbf{z})\right],$ (4) where we have introduced the notation $\mathbb{E}_{Y}[h(y)]$ to indicate the expectation of the function $h(y)$ with respect to the distribution $y\sim Y$. In principle, we could determine the conditional model by fixing $\theta$ to a value that maximizes the marginal likelihood. In practice, however, the integral in Equation 4 is intractable, due to the dimensionality of the latent space, and in any case would require a per-datapoint optimization process. As a result, the posterior $p_{\theta}(\mathbf{z}|\mathbf{x})=p_{\theta}(\mathbf{z},\mathbf{x})/p_{\theta}(\mathbf{x})$ is also intractable. We make progress by introducing a second approximation, this time to the posterior: $q_{\phi}(\mathbf{z}|\mathbf{x})\approx p_{\theta}(\mathbf{z}|\mathbf{x})$, where $q_{\phi}(\mathbf{z}|\mathbf{x})$ is often referred to as an _inference_ network. For any choice of $q_{\phi}(\mathbf{z}|\mathbf{x})$, including any choice of its weights $\phi$, we can write the log likelihood of the data as: $\log~{}p_{\theta}(\mathbf{x})=\mathbb{E}_{q_{\phi}(\mathbf{z}|\mathbf{x})}\left[\log~{}p_{\theta}(\mathbf{x})\right].$ (5) Applying the chain rule of probability: $p_{\theta}(\mathbf{x},\mathbf{z})=p_{\theta}(\mathbf{z})p_{\theta}(\mathbf{x}|\mathbf{z})$, and inserting an identity, this can be split into two terms: $\log p_{\theta}(\mathbf{x})=\mathbb{L}_{\theta,\phi}(\mathbf{x})+\mathbb{D}_{\rm KL}(q_{\phi}(\mathbf{z}|\mathbf{x})||p_{\theta}(\mathbf{z}|\mathbf{x})),$ (6) where $\mathbb{L}_{\theta,\phi}$ is referred to as the _evidence lower bound_ (ELBO): $\mathbb{L}_{\theta,\phi}(\mathbf{x})\equiv\mathbb{E}_{q_{\phi}(\mathbf{z}|\mathbf{x})}\left[\log\left[\frac{p_{\theta}(\mathbf{x},\mathbf{z})}{q_{\phi}(\mathbf{z}|\mathbf{x})}\right]\right],$ (7) and the second term is the Kullback-Leibler (KL) divergence: $\mathbb{D}_{\rm KL}(q_{\phi}(\mathbf{z}|\mathbf{x})||p_{\theta}(\mathbf{z}|\mathbf{x}))=\mathbb{E}_{q_{\phi}(\mathbf{z}|\mathbf{x})}\left[\log\left[\frac{q_{\phi}(\mathbf{z}|\mathbf{x})}{p_{\theta}(\mathbf{z}|\mathbf{x})}\right]\right],$ (8) which is a measure of the ‘distance’ between two distributions, and is always positive. From Equation 6 we see that the bound $\mathbb{L}_{\theta,\phi}(\mathbf{x})$ will become tightest when $\mathbb{D}_{\rm KL}(q_{\phi}(\mathbf{z}|\mathbf{x})||p_{\theta}(\mathbf{z}|\mathbf{x}))\rightarrow 0$, such that our approximation to the posterior, $q_{\phi}(\mathbf{z}|\mathbf{x})\approx p_{\theta}(\mathbf{z}|\mathbf{x})$, becomes exact. However, due to the presence of the $p_{\theta}(\mathbf{z}|\mathbf{x})$ term, $\mathbb{D}_{\rm KL}(q_{\phi}(\mathbf{z}|\mathbf{x})||p_{\theta}(\mathbf{z}|\mathbf{x}))$ can not be evaluated directly, and so we are not able to directly optimize the likelihood in Equation 6. Instead, we seek to maximize the evidence lower bound, thereby achieving an ‘optimum’ set of weights $\theta,~{}\phi$. The evidence lower bound and its gradient with respect to $\theta$ can be computed straightforwardly. The gradients with respect to $\phi$ appear more problematic, since the expectation we are calculating is taken over a distribution parametrized by $\phi$. The typical Monte Carlo estimates of this expectation, and its derivatives, are unbiased, but tend to have a high variance, often making the training process unstable. Through a reparametrization presented in Kingma & Welling (2013), it is possible to rewrite this expectation such that the source of randomness is not dependent on $\phi$, and gradients with respect to $\phi$ may be calculated with standard Monte Carlo techniques. We are therefore able to optimize $\mathbb{L}_{\theta,\phi}(\mathbf{x})$ by stochastic gradient descent, and approximately optimize the marginal log likelihood. ### 2.2 Architecture In this section we describe the architecture of the networks $p_{\theta}(\mathbf{x}|\mathbf{z})$ and $q_{\phi}(\mathbf{z}|\mathbf{x})$, and the latent prior $p(\mathbf{z})$. We adopt a convolutional architecture for both the encoder and decoder network. #### 2.2.1 Latent Space We choose to use a $d$-dimensional latent space, with a multivariate normal prior, $\mathbf{z}\sim\mathcal{N}(0,\mymathbb{1}^{d\times d})$. #### 2.2.2 Encoder The encoder maps input images $\mathbf{x}\in\mathbb{R}^{256\times 256}$ to latent space distribution parameters, $\mathbf{[}\bm{\mu}^{d},\bm{\sigma}^{d}]\in\mathbb{R}^{2d}$. It is worth emphasizing the point that, since we are modelling the distribution $p(\mathbf{z}|\mathbf{x})$, the output of the encoder is not a single point in the latent parameter space, but rather a distribution, parametrized by the mean and variance $\mathbf{[}\bm{\mu}^{d},\bm{\sigma}^{d}]$. The mapping from image to latent space parameters requires both a dimensionality reduction, and a reshaping. We achieve these goals by using a _convolutional neural network_. In the following we will describe the precise network that we implemented, using the language of neural networks. For details on the motivation for these choices, and their technical meaning, we refer to introductory texts on machine learning and convolutional neural networks such as Goodfellow et al. (2016) The encoder reduces the dimension of the input image by applying a series of strided convolutions with a rectified linear unit activation function, and then flattens the image for input to a final dense layer connected to the output latent space distribution parameters. Each convolution is characterized by a kernel shape with a number of pixels, $k_{i}$, where $i$ indicates the layer, and a stride length, which we set to 2. The values $k_{i}$ are set during the hyperaparameter optimization stage described in Section 2.3.3. We apply a batch normalization with momentum parameter equal to 0.9 after each convolution. This regularizes the weights, and leads to more stable training. A summary of the encoder model is given in Table 1. #### 2.2.3 Decoder The decoder is essentially the reverse process to the encoder, mapping a latent vector $\mathbf{z}\in\mathbb{R}^{d}$ to an image $\mathbf{x}\in\mathbb{R}^{256\times 256}$. We denote a decoder $g$, with weights $\phi$ as $g_{\phi}:\mathbf{z}\rightarrow\mathbf{x}$. The primary difference to the structure of the encoder is that we use transverse convolutions as opposed to convolutions, in order to increase the size of each dimension. A summary of the decoder model is given in Table 2. Layer | Layer Output Shape | Hyperparameters ---|---|--- Input | (256, 256, 1) | Conv2D | (128, 128, 256) | stride=2 ReLu | (128, 128, 256) | BatchNorm | (128, 128, 256) | momentum=0.9 Conv2D | (64, 64, 128) | stride=2 ReLu | (64, 64, 128) | BatchNorm | (64, 64, 128) | momentum=0.9 Conv2D | (32, 32, 64) | stride=2 ReLu | (32, 32, 64) | BatchNorm | (32, 32, 64) | momentum=0.9 Dense | (1024) | Dense | (512) | Table 1: This table shows the structure of the encoder network, $q_{\phi}(\mathbf{z}|\mathbf{x})$. Layer | Layer Output Shape | Hyperparameters ---|---|--- Input | (256, 1) | Dense | (8192) | Reshape | (16, 16, 32) | BatchNorm | (16, 16, 32) | momentum=0.9 TransposeConv2D | (32, 32, 128) | stride=2 ReLu | (32, 32, 128) | BatchNorm | (32, 32, 128) | momentum=0.9 TransposeConv2D | (64, 64, 64) | stride=2 ReLu | (64, 64, 64) | BatchNorm | (64, 64, 64) | momentum=0.9 TransposeConv2D | (128, 128, 32) | stride=2 ReLu | (128, 128, 32) | BatchNorm | (128, 128, 32) | momentum=0.9 TransposeConv2D | (256, 256, 16) | stride=2 ReLu | (256, 256, 16) | BatchNorm | (256, 256, 16) | momentum=0.9 TransposeConv2D | (256, 256, 1) | stride=1 Table 2: This table shows the structure of the decoder network, $p_{\theta}(\mathbf{x}|\mathbf{z})$. ### 2.3 Training In this section we detail the process by which we optimize the weights of the VAE model described in Section 2.2 with respect to the ELBO objective introduced in Section 2.1. The training process requires us to specify the training dataset, $\mathcal{D}$, the training _strategy_ by which we make updates to the weights $\theta,~{}\phi$, and the process of hyperparameter optimization by which we make concrete selections of meta parameters of the model (such as kernel shapes and training parameters). #### 2.3.1 Data Machine learning techniques are notoriously data-hungry, and will perform best for larger datasets. Standard computer vision datasets on which algorithms are tested (e.g. ImageNet (Russakovsky et al., 2015)) contain tens of thousands, sometimes millions, of images. However, we have only one sky from which to obtain observations of Galactic dust. As such, we are forced to partition the sky into patches, which we treat as separate images in the training process. In order to obtain $\sim 1000$’s of images, the natural linear scale of an individual patch is $\sim 10^{\circ}$. Such a small patch size has the advantage that we are then justified in projecting the cutouts onto the flat sky, and applying standard machine learning techniques to the resulting two- dimensional images, sidestepping the issue of defining neural networks that operate on spherical images (for such implementations see Perraudin et al. (2019); Krachmalnicoff & Tomasi (2019)). We use the _P_ lanck GNILC-separated thermal dust intensity map at 545 GHz 222http://pla.esac.esa.int/pla/aio/product-action?MAP.MAP_ID=COM_CompMap_Dust- GNILC-F545_2048_R2.00.fits, which we download from the Planck Legacy Archive. In order to extract cutout images from this map we follow a similar procedure to Aylor et al. (2019). We mask the Galactic plane by excluding all regions at latitudes below $15^{\circ}$. Then we lay down a set of centroids $(l_{i+1},b_{i+1})=(l_{i}+s,b_{i}+s/\cos(l_{i}))$, where $s$ is a step size parameter, and $s/\cos(l_{i})$ is a step between longitudes for a given latitude, which ensures the same angular separation in the latitudinal direction. Each centroid is then rotated to the equator, and an $8^{\circ}\times 8^{\circ}$ square region around the centroid is projected onto a cartesian grid with 256 pixels along each size. For $s=4^{\circ}$, this results in a dataset, $\mathcal{D}$, of 2254 maps. We then shuffle and split $\mathcal{D}$ into three groups: a 70% training set, $\mathbf{x}^{\rm train}$, a 15% validation set, $\mathbf{x}^{\rm val}$, and a 15% test set, $\mathbf{x}^{\rm test}$. In order to artificially increase the diversity of images in our limited sample we employ two standard data augmentation techniques. During the data preprocessing stage of training, we randomly flip each image along the horizontal and vertical directions, and rotate each image by an integer multiple of $90^{\circ}$. These transformations are not invariant under convolution; however, these would constitute perfectly realistic foreground images. #### 2.3.2 Strategy Here we discuss the training strategy used to learn the weights $\theta,\phi$. As discussed in Section 2, to train a VAE we maximize the lower bound on the log likelihood of the data given in Equation 7 with respect to the weights $\theta,\phi$. In practice, at each step we compute a Monte Carlo estimate of this quantity: $\mathbb{E}_{q_{\phi}(\mathbf{z}|\mathbf{x})}\left[\frac{p_{\theta}(\mathbf{x},\mathbf{z})}{q_{\phi}(\mathbf{z}|\mathbf{x})}\right]\approx\log p_{\theta}(\mathbf{x}|\mathbf{z})+\log p(\mathbf{z})-\log q_{\phi}(\mathbf{z}|\mathbf{x})$ (9) where $\mathbf{x}$ on the RHS is now a minibatch of the data, the size of which is a hyperparameter of the training process. The analysis we present in Section 2.3.3 shows that a batch size of 8 is preferred. For each batch we then calculate the gradients of this quantity with respect to the weights $\theta,\phi$ and backpropagate the errors through the network, adjusting $\theta,\phi$ in accordance with the learning schedule. For this schedule we used the Adam optimizer with hyperparameters determined through the optimization process described in Section 2.3.3. The training was performed by passing over the entire dataset 100 times, and in each pass splitting the data into batches of 8 images. To guard against overfitting we evaluated $\mathbb{L}_{\theta,\phi}(\mathbf{x}^{\rm train})$ and $\mathbb{L}_{\theta,\phi}(\mathbf{x}^{\rm val})$ every five epochs and checked for divergence between these quantities at late epochs. If the network had begun to overfit on the training data, its predictions for the validation set would deteriorate, which would be reflected in a worsening $\mathbb{L}_{\theta,\phi}(\mathbf{x}^{\rm val})$. We found that the $\mathbb{L}_{\theta,\phi}(\mathbf{x}^{\rm train})$ plateaued after 50 epochs, and saw no divergence between $\mathbb{L}_{\theta,\phi}(\mathbf{x}^{\rm train})$ and $\mathbb{L}_{\theta,\phi}(\mathbf{x}^{\rm val})$ after training for an additional 50 epochs. Models were built using the Tensorflow software package (Abadi et al., 2015), and trained using a Tesla V100 GPU on the Cori supercomputer at NERSC. #### 2.3.3 Hyperparameter Optimization In this section we provide motivation for our selection of the model hyperparameters. It is not possible to optimize model hyperparameters such as batch size, or model architecture, using the same stochastic gradient descent technique that is used to optimize model weights and biases. Instead, a limited number of hyperparameter combinations can be trained, and the corresponding model that achieves the best loss after a certain amount of training time, or certain number of epochs, is used. The space of hyperparameters is high-dimensional, and so can not be uniformly densely sampled due to computational cost. Instead, we employed a Bayesian optimization approach in which a few random combinations of hyperparameters are chosen, and trained for 20 epochs each. From this set of hyperparameters, a Gaussian process (GP) model of the loss as a function of hyperparameters is built. From this GP model, new trial candidates are selected, and trained, with the resulting loss then being incorporated into the GP weights. We allowed this process to continue for 100 different trials, and used the hyperparameters that achieved the lowest loss after twenty epochs of training. ## 3 Results ### 3.1 Reconstructions In this section we present reconstructions of test set images, and compare their pixel value distribution and power spectra. For a given image, $\mathbf{x}_{\rm test}$, we can sample the posterior as $\mathbf{z}_{\rm test}^{(i)}\sim q_{\phi}(\mathbf{z}|\mathbf{x})$, and push these through the decoder to get a reconstructed image $\mathbf{x}^{(i)}_{\rm test}=\mathbf{g}_{\theta}(\mathbf{z}^{(i)}_{\rm test})$. To summarize the distribution of reconstructed images, we draw $L$ samples and calculate their average: $\tilde{\mathbf{x}}\approx\frac{1}{L}\sum_{l=1}^{L}\mathbf{g}_{\theta}(\mathbf{z}_{\rm test}^{(l)}).$ (10) For the remainder of this section, a ‘reconstruction’ refers to the calculation of Equation 10 with $L=100$. For a given reconstruction, we can straightforwardly calculate two statistics: i) the histogram of its pixel values and ii) the power spectrum. We calculate the histogram of pixel values in 20 bins from -3 to 5, and normalize the count such that the area under the histogram is equal to unity. To calculate the power spectrum we apply a cosine apodization with a characteristic scale of one degree to the image, such that it smoothly tapers to zero at the edge of the map. We then calculate the mode coupling matrix for this mask, and calculate the uncoupled power spectrum using the NaMaster code (Alonso et al., 2019). For reasons that will become clear later we are primarily interested in comparing ranges of multipoles in the signal-dominated regime, well within the resolution limit of the original maps, and so we do not make any efforts to noise debias or account for the beam present in the original maps. First, we present the reconstructions of three randomly-selected test set images, and show the resulting maps, along with the residuals, in Figure 1. We can see that the network does very well in reconstructing the large-scale features in these test-set maps, and the visual quality is sufficient to appear ‘real’, if lower-resolution. Features are well recovered up to $\sim$degree scales, with features below that scale being smoothed out by the calculation of the expectation in Equation 10. The residuals shown in the bottom row of Figure 1 are well behaved and do not show any strong biases correlated with features in the map. Figure 1: This figure shows the reconstruction of three randomly-selected images from the test set, not used during the training or validation of the network. The top row are the original images, the second row are the reconstructions. and the third row are the residuals of the reconstructions. The reconstructions clearly lose small-scale details, but but manage to recover the large scale variations well. In Figure 2 we take a single randomly-selected test set image, and show its reconstruction, the pixel value histograms of each image, and their power spectra. As was the case for the three examples shown in Figure 1, there is excellent visual agreement between the original image and its reconstruction. This is enforced by the excellent agreement between the distribution of pixel values in the two images, shown in the bottom left panel of Figure 2. The reconstructed power spectrum in the bottom right panel of Figure 2 also shows excellent agreement up to $\ell\sim 400$, and suppression of power in the reconstructed image going to smaller scales, consistent with the visual blurriness of the reconstructed image. Figure 2: _Top left_ : a randomly-selected test set image, $\mathbf{x}$. _Top right_ : the reconstruction of the test set image, $\tilde{\mathbf{x}}$, as computed using Equation 10. _Bottom left_ : kernel density estimate of the distribution of pixel values of the original image, and its reconstruction. _Bottom right_ : the log power spectra of the test set image and its reconstruction. Note that since the test set images are standardized, these quantities are unitless. In order to compare reconstructions for the whole test set, we now calculate the pixel value distribution and power spectrum for each of the 339 images in the test set and their reconstructions. In order to represent the distribution of pixel value histograms across this test set, we calculate the quartiles and median in each bin, across the test set. In Figure 3 we plot the $25^{\rm th}$ percentile, median, and $75^{\rm th}$ percentile as a function of bin center, for both the original test set images, and their reconstructions. There is excellent agreement between the two sets of images, with no evidence of any aggregate bias in the reconstructions. Figure 3: In this figure we compare the pixel value distributions of the 339 test set images (black), and their reconstructions (green). We calculate quantiles across the test set, and plot the $25^{\rm th}$ and $75^{\rm th}$ quartiles (the dashed lines), and the median as functions of pixel value (the solid lines). In Figure 4 we compare the power spectra of all test set images and their reconstructions. Figure 4 shows that the same behavior as was seen in Figure 2 is displayed for the entire test set. Spectra are generally well recovered for $\ell<400$, with power being increasingly suppressed for $\ell>400$, relative to the real image power spectra. Figure 4: In this figure we compare the power spectra of the 339 test set images (black) and their reconstructions (green). Each power spectrum is plotted as an individual line. Here, we are encountering a known issue with VAEs: reconstructed images are often blurry (Kingma & Dhariwal, 2018; Kingma et al., 2016; Kingma & Welling, 2019). The blurriness can be understood by considering the objective function in Equation 7, and inspecting the term $\mathbb{E}_{q_{\phi}(\mathbf{z}|\mathbf{x})}\left[p_{\theta}(\mathbf{x},\mathbf{z})\right]$. Since this expectation is taken with respect to the distribution $q_{\phi}(\mathbf{z}|\mathbf{x})$, it will strongly penalize points $(\mathbf{x},\mathbf{z})$ that are likely under $q_{\phi}$, but unlikely under $p_{\theta}$. On the other hand, points that are likely under $p_{\theta}$, but are not present in the empirical data distribution, will suffer a much smaller penalty. The result is that, if the model is not sufficiently flexible to fit the data distribution exactly, it will compensate by widening the support of $p_{\theta}(\mathbf{x},\mathbf{z})$ beyond what is present in the data distribution, inflating the variance of $p_{\theta}(\mathbf{x}|\mathbf{z})$. Since we have assumed a Gaussian distribution for the decoder model that is independent from pixel to pixel, and given that the signal in the training images is red-tilted (as is the case for most natural images containing extended recognizable structures), the increased variance leads to a degradation of small-scale features through the averaging process of Equation 10 (Zhao et al., 2017). A corollary of the extended support of $p_{\theta}(\mathbf{x},\mathbf{z})$ is that sampling the prior in order to generate novel images will not necessarily produce realistic samples (Kingma & Welling, 2019). One way in which the flexibility of VAEs may be enhanced is through the use of _normalizing flows_ (Jimenez Rezende & Mohamed, 2015). As the name suggests, the idea here is to start with a simple distribution, such as a multivariate normal, and ‘stack’ layers of invertible transformations, such that the output may be significantly more complex. There are certain requirements placed on these transformations such that they remain computationally efficient, for example they must have tractable Jacobians (Jimenez Rezende & Mohamed, 2015). Expanding the VAE model presented here by introducing normalizing flows could be expected to improve both the reconstruction quality, and the quality of novel samples, and is the subject of current work. ### 3.2 Interpolation in the latent space As a means of investigating the structure of the encoding that has been learned, we study the ‘interpolation’ between real images, $\mathbf{x}_{1}$ and $\mathbf{x}_{2}$, by performing the interpolation between their latent encodings, $\mathbf{z}_{1}$ and , $\mathbf{z}_{2}$. From the smooth nature of the changes in the resulting continuum of maps we will see that smooth variations in the latent space result in smooth variations in the map space. This study also demonstrates the ability of the VAE approach to generate novel foreground images by restricting to a region of the latent space close to the encodings of real maps, therefore avoiding the spurious regions of $(\mathbf{x},\mathbf{z})$ that could be obtained by sampling from an ill- fitted prior, as discussed at the end of Section 3.1. The probability mass in high-dimensional distributions tends to concentrate in a shell relatively far from the modal probability density. Therefore, traversing the latent space in a straight line (in the Euclidean sense), does not necessarily pass through areas of high probability mass. In order to keep the interpolated points within areas of high probability mass, we interpolate from $\mathbf{z}_{1}$ to $\mathbf{z}_{2}$ using spherical trajectories that traverse great circles in the latent space, as the distance from the origin smoothly changes from $|\mathbf{z_{1}}|$ to $|\mathbf{z_{2}}|$. Specifically, we follow this continuous trajectory parametrized by some factor $\lambda$: $\mathbf{z}_{1,2}(\lambda)=\frac{\sin((1-\lambda)\theta)}{\sin\theta}\mathbf{z}_{1}+\frac{\sin(\lambda\theta)}{\sin\theta}\mathbf{z}_{2},$ (11) where $\cos(\theta)=\hat{\mathbf{z}}_{1}\cdot\hat{\mathbf{z}}_{2}$. We then take $N$ points along this line corresponding to $\lambda=[1/(N+1),2/(N+1),\dots,N/(N+1)]$, and decode to obtain the corresponding map $\mathbf{x}_{1,2}(\lambda)=g_{\phi}(\mathbf{z}_{1,2}(\lambda))$. Figure 5: This figure presents synthetic images generated by interpolating between real images, $\mathbf{x}_{1}$ and $\mathbf{x}_{2}$, shown in the top left and bottom right panels respectively. The interpolation is carried out in the latent space using Equation 11, and is parametrized by a continuous variable $\lambda$. The intermediate panels show the interpolation evaluated at $N=10$ points along the trajectory. Figure 5 shows the smooth transition in image space between the two real images (the top left panel and the bottom right panel) randomly selected from the test set, calculated using the interpolation described above. Features, such as the strong filamentary structures in the center of the image, transition smoothly in and out of the image, demonstrating that small perturbations in the latent space result in small perturbations in decoded images. ### 3.3 Data Imputation In this section we consider a possible application of our trained model to the reconstruction of corrupted data. During the analysis of CMB data there are many possible reasons that data may be incomplete, from masking of point sources, to corruption by uncontrolled systematics. The task of _inpainting_ these regions is simple when the missing emission is well described by Gaussian statistics, as is the case for the CMB (Bucher & Louis, 2012). The lack of a similarly simple approach for the non-Gaussian foreground signal means that previous efforts have relied on empirically-validated, simple, algorithms, such as diffusive filling (Bucher et al., 2016). Future surveys will have ever-lower noise floors, and so will be increasingly contaminated by point-sources, even in polarization. The aggressive masking required in this regime could lead to the failure of simple foreground inpainting techniques (Puglisi & Bai, 2020). The statistical foreground model presented here allows us to take a Bayesian approach to foreground inpainting, in which we may compute a posterior distribution for the missing data, conditioned on the observed data (Böhm et al., 2019). This has the advantage of conserving the foregrounds’ statistical properties, whilst also taking into account all of the contextual information in the image, unlike methods such as diffusive inpainting. In the rest of this section we will present a toy model for corrupted data, and show that we are able to perform inpainting by optimizing the posterior distribution in the latent space. Representing the contamination as a linear operator $\mathsf{A}$, we can write down a model for the observed data $\mathbf{d}$: $\mathbf{d}=\mathsf{A}\mathbf{x}+\mathbf{n}$, where $\mathbf{n}$ is a possible noise term. The posterior distribution of $\mathbf{z}$ is given by Bayes’ theorem: $\log p(\mathbf{z}|\mathbf{d})=\log p(\mathbf{z})+\log p_{\theta}(\mathbf{d}|\mathbf{z})-\log p(\mathbf{d}).$ (12) For a given statistical model of the noise, we have a complete description of the term $\log p(\mathbf{d}|\mathbf{z})$, and we can work with the posterior distribution in the latent space. As a concrete example we will consider the case of a binary $N\times N$ masking operator, $\mathsf{A}$, with elements equal to one (zero) where pixels are (un)observed. To form simulated ‘corrupted’ images, we take random images from the test dataset, apply $\mathsf{A}$, and add white Gaussian noise $\mathbf{n}$, characterized by a pixel standard deviation $\sigma$: $\mathbf{d}_{\rm test}=\mathsf{A}\mathbf{x}_{\rm test}+\mathbf{n}$. The posterior distribution in the latent space is then: $-2\log p(\mathbf{z}|\mathbf{d}_{\rm test})\propto\mathbf{z}^{T}\mathbf{z}+\frac{\bm{\mu}_{\theta}(\mathbf{z})^{T}\bm{\mu}_{\theta}(\mathbf{z})}{\sigma^{2}},$ (13) where we have written the residual vector as $\bm{\mu}_{\theta}(\mathbf{z})=\mathsf{A}\mathbf{g}_{\theta}(\mathbf{z})-\mathbf{d}_{\rm test}$. Fully sampling Equation 13 can be computationally expensive due to the dimensionality of $\mathbf{z}$, and is made more challenging by the possibility of $\log p(\mathbf{z}|\mathbf{d}_{\rm test})$ being multi-modal. For these reasons, applying standard Markov Chain Monte Carlo techniques can often fail to fully explore the posterior (Böhm et al., 2019), and we leave a sampling approach for future work, here taking only a single representative sample by maximizing $\hat{\mathbf{z}}_{\rm test}=\operatorname*{argmax}_{\mathbf{z}}\log p(\mathbf{z}|\mathbf{d}_{\rm test})$. In the following we will take $\mathsf{A}$ to be a masking operator that applies a binary mask to a map. However, as long as a forward model for the corruption operation can be written down (e.g. a Gaussian convolution), the same technique could be applied. We take three randomly selected test set images, $\mathbf{x}_{1},\mathbf{x}_{2},\mathbf{x}_{3}$, and apply three different binary masks, $\mathsf{A}_{1},\mathsf{A}_{2},\mathsf{A}_{3}$. To each corrupted image, we add a white noise realization with a pixel standard deviation of 0.2. For each corrupted, noisy image, we then maximize the posterior in Equation 13 to find $\mathbf{z}_{i}^{\rm MAP}$ using the LBFGS algorithm. In Figure 6 we show the randomly selected test set images in the first row, the corrupted images in the second row, and the reconstructed map $g(\mathbf{z}_{i}^{\rm MAP})$ in the third row. We also calculate the pixel value histograms and power spectra of the input and reconstructed maps and show these in the bottom two rows of Figure 6. Figure 6: This figure shows three randomly-selected test set images, $\mathbf{x}_{1,2,3}$ in the top row. As described in Section 3.3, these images are corrupted with a binary mask $\mathsf{A}_{1,2,3}$ and white noise. The corrupted images are shown in the second row. The third row shows the reconstructed images obtained by maximizing the latent space posterior in Equation 13 for each of the three corrupted images, and decoding the resulting points in the latent space. The fourth and fifth rows show the pixel value histograms and power spectra of the original and reconstructed maps. One can see from Figure 6 that all the images are well reconstructed, and there is no visible effect of the masking remaining in the reconstructions. Comparing the regions in the first and third rows corresponding to the masked areas, we see that the network does not reproduce the exact features in the masked region, for any of the $\mathbf{x}_{i}$, as expected. However, the network does reconstruct plausible inpaintings, with the correct statistics, given the context in the rest of the image. For example, the reconstruction $g_{\phi}(\mathbf{z}_{2}^{\rm MAP})$ does not replicate the true high- intensity filamentary structure in the input image, $\mathbf{x}_{2}$, which would be impossible. However, it does recognize from the context that intensity is increasing towards the masked area in the bottom left of the image, and populates that area with high-variance, high-intensity features. Correspondingly, such high-intensity features are not seen in the reconstructed regions of $g_{\phi}(\mathbf{z}_{1,3}^{\rm MAP})$, which correspond to relatively low-emission regions. The pixel value histograms and power spectra in the last two rows of Figure 6 show similar behavior. We see good agreement between the original and reconstructed histograms and powerspectra for both the $\mathbf{x}_{1}$ and $\mathbf{x}_{3}$ maps, up to the suppression at $\ell>400$ common to all reconstructions. On the other hand, we see a disagreement between the original and reconstructed statistics of $\mathbf{x}_{2}$, due to the higher variance associated with the filled-in region. These results show that the network has learned generalizable information about foreground behavior, and is able to inpaint novel foreground emission with correct statistical properties, based on the context of an image. The forward model used in this inpainting process can be easily extended to maps with multiple masks and different types of filtering and noise found in real data. ## 4 Discussion and Conclusions In this paper we have presented a new application of VAEs to images of Galactic thermal dust emission. Using a training set extracted from Planck observations of thermal dust emission, this technique allowed us to learn a transformation from a space of uncorrelated latent variables with a multivariate normal prior, to the space of possible dust maps. The training process was validated by computing and comparing summary statistics, including the distribution of pixel values, and power spectra of reconstructed maps, on a test set withheld during the training process. The applicability of the trained model was also demonstrated by reconstructing data corrupted by noise and masking. This was the first use of a trained generative dust model to perform Bayesian inference, and demonstrates the applicability of this approach in the simulation of foreground images, and the Bayesian modeling of polarized CMB data. The usefulness of this model is currently limited by the flexibility of the posterior, and its ability to fit the true underlying posterior. As was discussed in Section 3.1, this has two main consequences: i) a naïve sampling of the prior is not guaranteed to produce realistic samples, ii) reconstructed images are blurry, limiting accuracy to degree scales. Both of these issues may be tackled by increasing the expressiveness of the model (Kingma & Welling, 2019), which we plan to do by introducing a normalizing flow to link the prior and latent space (Kingma et al., 2016). As discussed in the Section 1, our main goal is to model polarized dust emission. We attempted a similar analysis to that presented here by repeating the training procedure on a network that accepted an additional ‘channel’ as input, representing a tuple of Stokes $Q$ and $U$ parameters, rather than only Stokes $I$, and using the Planck 353 GHz polarization observations to form a training set. We found that the network was not able to learn any meaningful information from this setup, consistent with what similar analyses have found (Petroff et al., 2020). In order to extend our analysis to polarization, we are therefore exploring the use of MHD simulations (Kim et al., 2019) as a training set. Kim et al. (2019) have demonstrated that simulations of a multiphase, turbulent, magnetized ISM produce synthetic observations of the ISM with statistics (such as the ratio of $E$ power to $B$ power, and the tilt of the $EE$ and $BB$ power spectra) matching those of real skies. Our initial results have shown that this is a promising alternative to the use of real data in training generative networks. ## Acknowledgements We would like to acknowledge useful conversations with Ethan Anderes and Kevin Aylor in the preparation of this work. This work was supported by an XSEDE start up allocation, PHY180022. This work was supported in part by the National Science Foundation via awards OPP-1852617 and AST-1836010. We also acknowledge the use of the Perlmutter preparedness GPU allocation on the Cori super computer at NERSC. ## Data Availability The data used in this study is available on the Planck Legacy Archive at the URL: http://pla.esac.esa.int/pla/aio/product- action?MAP.MAP_ID=COM_CompMap_Dust-GNILC-F545_2048_R2.00.fits ## References * Abadi et al. (2015) Abadi M., et al., 2015, TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems, https://www.tensorflow.org/ * Ade et al. (2019) Ade P., et al., 2019, J. Cosmology Astropart. Phys., 2019, 056 * Aiola et al. (2020) Aiola S., et al., 2020, arXiv e-prints, p. arXiv:2007.07288 * Alonso et al. (2019) Alonso D., Sanchez J., Slosar A., LSST Dark Energy Science Collaboration 2019, MNRAS, 484, 4127 * Aylor et al. (2019) Aylor K., Haq M., Knox L., Hezaveh Y., Perreault-Levasseur L., 2019, arXiv e-prints, p. arXiv:1909.06467 * BICEP2 Collaboration et al. (2018) BICEP2 Collaboration et al., 2018, Phys. Rev. Lett., 121, 221301 * Benson et al. (2014) Benson B. A., et al., 2014, in Holland W. S., Zmuidzinas J., eds, Vol. 9153, Millimeter, Submillimeter, and Far-Infrared Detectors and Instrumentation for Astronomy VII. SPIE, pp 552 -- 572, doi:10.1117/12.2057305, https://doi.org/10.1117/12.2057305 * Böhm et al. (2019) Böhm V., Lanusse F., Seljak U., 2019, arXiv e-prints, p. arXiv:1910.10046 * Brock et al. (2018) Brock A., Donahue J., Simonyan K., 2018, arXiv e-prints, p. arXiv:1809.11096 * Bucher & Louis (2012) Bucher M., Louis T., 2012, MNRAS, 424, 1694 * Bucher et al. (2016) Bucher M., Racine B., van Tent B., 2016, Journal of Cosmology and Astroparticle Physics, 2016, 055 * Dinh et al. (2014) Dinh L., Krueger D., Bengio Y., 2014, arXiv e-prints, p. arXiv:1410.8516 * Dinh et al. (2016) Dinh L., Sohl-Dickstein J., Bengio S., 2016, arXiv e-prints, p. arXiv:1605.08803 * Goodfellow et al. (2014) Goodfellow I. J., Pouget-Abadie J., Mirza M., Xu B., Warde-Farley D., Ozair S., Courville A., Bengio Y., 2014, arXiv e-prints, p. arXiv:1406.2661 * Goodfellow et al. (2016) Goodfellow I., Bengio Y., Courville A., 2016, Deep Learning. MIT Press * Grover et al. (2017) Grover A., Dhar M., Ermon S., 2017, arXiv e-prints, p. arXiv:1705.08868 * Hui et al. (2018) Hui H., et al., 2018, in Zmuidzinas J., Gao J.-R., eds, Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series Vol. 10708, Millimeter, Submillimeter, and Far-Infrared Detectors and Instrumentation for Astronomy IX. p. 1070807 (arXiv:1808.00568), doi:10.1117/12.2311725 * Jimenez Rezende & Mohamed (2015) Jimenez Rezende D., Mohamed S., 2015, arXiv e-prints, p. arXiv:1505.05770 * Jimenez Rezende et al. (2014) Jimenez Rezende D., Mohamed S., Wierstra D., 2014, arXiv e-prints, p. arXiv:1401.4082 * Kamionkowski et al. (1997) Kamionkowski M., Kosowsky A., Stebbins A., 1997, Phys. Rev. Lett., 78, 2058 * Kim et al. (2019) Kim C.-G., Choi S. K., Flauger R., 2019, ApJ, 880, 106 * Kingma & Dhariwal (2018) Kingma D. P., Dhariwal P., 2018, arXiv e-prints, p. arXiv:1807.03039 * Kingma & Welling (2013) Kingma D. P., Welling M., 2013, arXiv e-prints, p. arXiv:1312.6114 * Kingma & Welling (2019) Kingma D. P., Welling M., 2019, arXiv e-prints, p. arXiv:1906.02691 * Kingma et al. (2016) Kingma D. P., Salimans T., Jozefowicz R., Chen X., Sutskever I., Welling M., 2016, arXiv e-prints, p. arXiv:1606.04934 * Knox & Turner (1994) Knox L., Turner M. S., 1994, Phys. Rev. Lett., 73, 3347 * Krachmalnicoff & Puglisi (2020) Krachmalnicoff N., Puglisi G., 2020, arXiv e-prints, p. arXiv:2011.02221 * Krachmalnicoff & Tomasi (2019) Krachmalnicoff N., Tomasi M., 2019, A&A, 628, A129 * Millea et al. (2020a) Millea M., Anderes E., Wandelt B. D., 2020a, arXiv e-prints, p. arXiv:2002.00965 * Millea et al. (2020b) Millea M., et al., 2020b, arXiv e-prints, p. arXiv:2012.01709 * Perraudin et al. (2019) Perraudin N., Defferrard M., Kacprzak T., Sgier R., 2019, Astronomy and Computing, 27, 130 * Petroff et al. (2020) Petroff M. A., Addison G. E., Bennett C. L., Weiland J. L., 2020, arXiv e-prints, p. arXiv:2004.11507 * Planck Collaboration et al. (2020) Planck Collaboration et al., 2020, A&A, 641, A6 * Puglisi & Bai (2020) Puglisi G., Bai X., 2020, arXiv e-prints, p. arXiv:2003.13691 * Razavi et al. (2019) Razavi A., van den Oord A., Vinyals O., 2019, arXiv e-prints, p. arXiv:1906.00446 * Russakovsky et al. (2015) Russakovsky O., et al., 2015, International Journal of Computer Vision (IJCV), 115, 211 * Seljak & Zaldarriaga (1997) Seljak U., Zaldarriaga M., 1997, Phys. Rev. Lett., 78, 2054 * Smoot et al. (1992) Smoot G. F., et al., 1992, ApJ, 396, L1 * Zhao et al. (2017) Zhao S., Song J., Ermon S., 2017, arXiv e-prints, p. arXiv:1702.08658 * van den Oord et al. (2016a) van den Oord A., Kalchbrenner N., Kavukcuoglu K., 2016a, arXiv e-prints, p. arXiv:1601.06759 * van den Oord et al. (2016b) van den Oord A., et al., 2016b, arXiv e-prints, p. arXiv:1609.03499
###### Proof. .: 5pt ###### Transporting a prediction model for use in a new target population ###### Abstract We consider methods for transporting a prediction model and assessing its performance for use in a new target population, when outcome and covariate data for model development are available from a simple random sample from the source population, but only covariate data are available from a simple random sample from the target population. We discuss how to tailor the prediction model for use in the target population, how to assess model performance (e.g., by estimating the target population mean squared error), and how to perform model and tuning parameter selection. We provide identifiability results for measures of performance in the target population for a potentially misspecified prediction model under a sampling design where the source and the target population samples are obtained separately. We also introduce the concept of prediction error modifiers that can be used to reason about tailoring measures of model performance to the target population. We illustrate the methods using simulated data. Keywords: transportability, generalizability, model performance, prediction error modifier, covariate-shift, domain adaptation ## Introduction Users of prediction models typically want to obtain predictions in a specific target population. For example, a healthcare system may want to deploy a clinical risk prediction model [1] to identify individuals at high risk for adverse outcomes among all patients receiving care. Prediction models are often built using data from source populations represented in prospective epidemiological cohorts, confirmatory randomized trials [2], or administrative databases [3]. In most cases, the data from the source population that are used for developing the prediction model cannot be treated as a random sample from the target population where the model will be deployed because the two populations have different data distributions. Consequently, a model developed using the data from the source population may not be applicable to the target population and model performance estimated using data from the source population may not reflect performance in the target population. Consider a setup where outcome and covariate data are available from a sample of the source population and only covariate data are available from a sample of the target population. For example, covariate data from the target population may be obtained from administrative databases, but outcome data may be unavailable (e.g., when outcome ascertainment requires specialized assessments) or insufficient (e.g., when the number of outcome events is small due to incomplete followup). In this setup, developing and assessing the performance of a prediction model for the target population is not possible using standard methods because of the complete lack of outcome data from the target population; using data from the source population can be an attractive alternative. Yet, as noted above, directly applying a prediction model developed in data from the source population to the target population, or treating model performance measures (e.g., mean squared prediction error) estimated in the source data as reflective of performance in the target population may be inappropriate when the two populations have different data distributions. Thus, investigators are faced with two transportability tasks: (1) tailoring a prediction model for use in a target population when relying on outcome data from the source population; and (2) assessing the performance of the model in that target population. These two transportability tasks have received attention in the computer science literature on covariate shift and domain adaptation [4, 5, 6, 7, 8, 9, 10, 11, 12]. In epidemiology, however, the transportability of prediction models has been treated heuristically and commonly used methods do not have well-understood statistical behavior. The related problem of transporting inferences about treatment effects to a target population has received more attention [13, 14, 15, 16], but there are important differences between transportability of treatment effects and prediction models in terms of the parameters being estimated and the methods used for estimation. Here, we examine the conditions that allow transporting prediction models from the source population to the target population. We discuss the implications of these conditions both for tailoring the models for use in the target population and for assessing model performance in that context. We show that many popular measures of model performance can be identified and estimated using covariate and outcome data from the source population and just covariate data from the target population under both nested and non-nested sampling designs, without the strong assumption that the prediction model is correctly specified. We discuss the relevance of our results when using modern model- building approaches such as cross-validation-based model selection. We introduce the concept of prediction error modifiers, which is useful for reasoning about transportability of measures of model performance to the target population. Last, we illustrate the methods using simulated data. ## Sampling design and identifiability conditions Let $Y$ be the outcome of interest and $X$ a covariate vector. We assume that outcome and covariate information is obtained from a simple random sample from the source population $\\{(X_{i},Y_{i}):i=1,\ldots,n_{\text{\tiny source}}\\}$. Furthermore, covariate information is obtained from a simple random sample from the target population, $\\{X_{i}:i=1,\ldots,n_{\text{\tiny target}}\\}$; no outcome information is available from the target population. This “non-nested” sampling design [17, 18], where the samples from the target and source population are obtained separately, is the one most commonly used in studies examining the performance of a prediction model in a new target population. For that reason, we will present results for non-nested designs in some detail, before considering nested designs, where the source population is a subset of a larger population that represents the target population. Let $S$ be an indicator for the population from which data are obtained, with $S=1$ for the source population and $S=0$ for the target population, and denote $n=n_{\text{\tiny source}}+n_{\text{\tiny target}}$ as the sample size of the composite dataset consisting of the data from the source and target population samples. This composite dataset is randomly split into a training set and a test set. The training set is used to build a prediction model for the expectation of the outcome conditional on covariates in the source population, $\operatorname{\textnormal{\mbox{E}}}[Y|X,S=1]$, and then, the test set is used to evaluate model performance. We use $g_{\beta}(X)$ to denote the posited parametric model, indexed by the parameter $\beta$, and $g_{\widehat{\beta}}(X)$ to denote the “fitted” model with estimated parameter $\widehat{\beta}$. We use $f(\cdot)$ to generically denote densities. We assume the following identifiability conditions: 1. A1. Conditional independence of the outcome $Y$ and the data source $S$. For every $x$ with positive density in the target population, $f(X=x,S=0)>0$, $f(Y|X=x,S=1)=f(Y|X=x,S=0).$ Informally, this condition means that the relationship between $Y$ and $X$ is the same in the source population and the target population and it implies that the conditional expectation of $Y$ given $X$ is the same in the two populations, $\operatorname{\textnormal{\mbox{E}}}[Y|X,S=1]=\operatorname{\textnormal{\mbox{E}}}[Y|X,S=0]$. 2. A2. Positivity. For every $x$ such that $f(X=x,S=0)\neq 0$, $\Pr[S=1|X=x]>0$. Informally, this condition means that every covariate pattern in the target population can occur in the source data, as sample size goes to infinity. Next, we discuss how, under assumptions A1 and A2, the prediction model can be tailored for use in the target population and how we can assess model performance in the target population. ## Tailoring the model to the target population Recall that $g_{\beta}(X)$ is a model for $\operatorname{\textnormal{\mbox{E}}}[Y|X,S=1]$. Suppose that the parameter $\beta$ takes values in the space $\mathcal{B}$. We say that the model is correctly specified if there exists a $\beta_{0}\in\mathcal{B}$ such that $g_{\beta_{0}}(X)=\operatorname{\textnormal{\mbox{E}}}[Y|X,S=1]$ [19]. Tailoring the fitted model $g_{\widehat{\beta}}(X)$ for use in the target population depends on whether the posited model $g_{\beta}(X)$ is correctly specified. ##### Correctly specified model: Suppose that the model $g_{\beta}(X)$ is correctly specified and thus we can construct a model-based estimator $g_{\widehat{\beta}}(X)$ that consistently estimates $\operatorname{\textnormal{\mbox{E}}}[Y|X,S=1]$. Under condition A1, a consistent estimator for $\operatorname{\textnormal{\mbox{E}}}[Y|X,S=1]$ is also consistent for $\operatorname{\textnormal{\mbox{E}}}[Y|X,S=0]$ (because the two expectations are equal when condition A1 holds). Moreover, when the model for the conditional expectation is parametric (as we have assumed up to now) and the parameter $\beta$ is estimated using maximum likelihood methods, then the unweighted maximum likelihood estimator $\widehat{\beta}$ estimated using only the source data training set is optimal in terms of having the smallest asymptotic variance [20, 21]. ##### Missspecified model: Now, suppose, as is more likely to be the case, that the model $g_{\beta}(X)$ is misspecified. In that case, theoretical work on the behavior of weighted maximum likelihood estimators for $\beta$ under covariate shift [21] shows that the maximum likelihood estimator estimated using only source population data is no longer optimal, in the sense of minimizing the Kullback-Leibler divergence between the estimated and true conditional density of the outcome given covariates. Instead, the Kullback-Leibler divergence is minimized by using a weighted maximum likelihood estimator with weights set equal to the ratio of the densities in the target and source populations, that is, $f(X|S=0)/f(X|S=1)$. In applied work, the density ratio is typically unknown and needs to be estimated using the data, but direct estimation of density ratios is challenging, particularly when $X$ is high-dimensional [22]. Instead, we can use the fact that the density ratio is, up to a proportionality constant, equal to the inverse of the odds of being from the source population, $\dfrac{f(X|S=0)}{f(X|S=1)}\propto\dfrac{\Pr[S=0|X]}{\Pr[S=1|X]},$ to replace density ratio weights with inverse-odds weights and obtain an optimal estimator of the model, tailored for use in the target population. The inverse-odds weights can be obtained by estimating the probability of an observation being from the source population conditional on covariates – a task for which many practical methods are available for high-dimensional data [23]. A reasonable approach for tailoring a potentially misspecified prediction model for use in the target population could proceed in three steps. Fist, estimate the probability of being from the source population, using training data from the source population and target population. Second, use the estimated probabilities to construct inverse-odds of participation weights for observations in the training set from the source population. Third, apply the weights from the second step to estimate the prediction model using all observations in the training set from the source population. One difficulty with the above procedure is that, in non-nested designs, the sample from the source population and the sample from the target population are obtained separately, with sampling fractions from the corresponding underlying populations that are unknown by the investigators and unlikely to be equal. When that is the case, the probabilities $\Pr[S=0|X]$ and $\Pr[S=1|X]$ in the inverse-odds weights are not identifiable from the observed data [18, 24] (i.e., cannot be estimated using the observed data). Although the inverse-odds weights are not identifiable, in Appendix A.1 we show that, up to an unknown proportionality constant, they are equal to the inverse-odds of participation weights _in the training set_ , $\dfrac{\Pr[S=0|X]}{\Pr[S=1|X]}\propto\dfrac{\Pr[S=0|X,D_{\text{\tiny train}}=1]}{\Pr[S=1|X,D_{\text{\tiny train}}=1]},$ (1) where $D_{\text{\tiny train}}$ is an indicator if data from an observation is in the training set and used to estimate the inverse-odds weights. It follows that we can use inverse-odds weights estimated in the training set, when estimating $\beta$ with the weighted maximum likelihood estimator. ## Assessing model performance in the target population We now turn our attention to assessing model performance in the target population. For concreteness, we focus on model assessment using the squared error loss function and on identifying and estimating its expectation, that is, the mean squared error (MSE), in the target population. The squared error loss $(Y-g_{\widehat{\beta}}(X))^{2}$ quantifies the discrepancy between the (observable) outcome $Y$ and the model-derived prediction $g_{\widehat{\beta}}(X)$ in terms of the square of their difference. The MSE in the target population is defined as $\psi_{\widehat{\beta}}=\operatorname{\textnormal{\mbox{E}}}[(Y-g_{\widehat{\beta}}(X))^{2}|S=0].$ In the main text of this paper, we focus on the MSE because it is a commonly used measure of model performance. Our results, however, readily extend to other measures of performance. In Appendix A.1, we provide identifiability results for general loss function-based measures of model performance. ### Prediction error modifiers To help explain why model performance measures need to be tailored for use in the target population, we introduce the term “prediction error modifier” to describe a covariate that, for a given prediction model, is associated with prediction error as assessed with some specific measure of model performance. Slightly more formally and using the squared error loss as an example, we say that the random variable $Z$ is a prediction error modifier, for the model $g_{\widehat{\beta}}(X)$, with respect to MSE in the source population, if the conditional expectation $\operatorname{\textnormal{\mbox{E}}}[(Y-g_{\widehat{\beta}}(X))^{2}|Z=z,S=1]$ varies as a function of $z$. Several parametric or non-parametric methods are available to examine whether $\operatorname{\textnormal{\mbox{E}}}[(Y-g_{\widehat{\beta}}(X))^{2}|Z,S=1]$ is a constant [25]. The prediction error modifier $Z$ can contain all the covariates in $X$ or only a subset of them. When the distribution of prediction error modifiers differs between the source and target populations, measures of model performance estimated using data from the source population are unlikely to be applicable in the target population, in the sense that the performance of the model in the source data may be very different (either better or worse) compared to performance of the same model in the target population. Large differences in performance measures between the source and target population can occur even if the true outcome model in the two populations is the same (i.e., even if condition A1 holds) because most common measures of model performance average (marginalize) prediction errors over the data distribution of the target population, and the covariate distribution of the target population can be different from the distribution in the source population. Figure 1 shows an example of a prediction error modifier that is differently distributed between the source and target population resulting in an MSE in the target population that is higher than the MSE in the source population; as the covariate vector in the example is one dimensional $X$ and $Z$ are equal. In the middle panel of Figure 1 we plot the inverse-odds weights as a function of the prediction error modifier $X$; in the bottom panel we plot the conditional squared errors as a function of $X$. Because both the conditional squared errors and the inverse-odds weights (and therefore the probability of being from the target population) increase as $X$ increases, the target population MSE (which is equal to the expectation of the squared errors) is larger than the source population MSE. Hence, directly using the source population MSE in the context of the target population would lead to over- optimism about model performance. ### Assessing model performance in the target population In our setup, where outcome information is only available from the sample of the source population, we need to account for differences in the data distribution between the source population and the target population to assess model performance in the target population. Proposition 1 in Appendix A.1 shows that, under the setup described previously and conditions A1 and A2, $\psi_{\widehat{\beta}}$ is identifiable using source and target population data through the expression $\psi_{\widehat{\beta}}=\operatorname{\textnormal{\mbox{E}}}[\operatorname{\textnormal{\mbox{E}}}[(Y-g_{\widehat{\beta}}(X))^{2}|X,S=1,D_{\text{\tiny test}}=1]|S=0,D_{\text{\tiny test}}=1],$ or equivalently using an inverse-odds weighting expression $\psi_{\widehat{\beta}}=\frac{1}{\Pr[S=0|D_{\text{\tiny test}}=1]}\operatorname{\textnormal{\mbox{E}}}\left[\frac{I(S=1)\Pr[S=0|X,D_{\text{\tiny test}}=1]}{\Pr[S=1|X,D_{\text{\tiny test}}=1]}(Y-g_{\widehat{\beta}}(X))^{2}\bigg{|}D_{\text{\tiny test}}=1\right].$ (2) Here, $D_{\text{\tiny test}}$ is an indicator for whether an observation is in the source or target test data. The identifiability result in expression (2) suggests the following inverse- odds weighting estimator [26, 21] for the target population MSE: $\widehat{\psi}_{\widehat{\beta}}=\frac{\sum\limits_{i=1}^{n}I(S_{i}=1,D_{\text{\tiny test},i}=1)\widehat{o}(X_{i})\left(Y_{i}-g_{\widehat{\beta}}(X_{i})\right)^{2}}{\sum\limits_{i=1}^{n}I(S_{i}=0,D_{\text{\tiny test},i}=1)},$ (3) where $\widehat{o}(X)$ is an estimator for the inverse-odds weights in the test set, $\dfrac{\Pr[S=0|X,D_{\text{\tiny test}}=1]}{\Pr[S=1|X,D_{\text{\tiny test}}=1]}$. To ensure independence between the data used to train the model and the data used to evaluate the model, we propose to use inverse-odds weights estimated using the training set for model building and inverse-odds weights estimated using the test set for estimating model performance. An important feature of our result is that it does not require the prediction model to be correctly specified, that is, we do not assume that $g_{\widehat{\beta}}(X)$ converges to the true conditional expectation of the outcome in the source population, $\operatorname{\textnormal{\mbox{E}}}[Y|X,S=1]$. This implies that model performance measures in the target population are identifiable and estimable, both for misspecified and correctly specified models. Informally, our identifiability results require the existence of a common underlying model for the source and target population (condition A1), but they do not require the (much less plausible) assumption that investigators can correctly specify that model. So far we have focused on the scenario where the prediction model is built using the training data and is evaluated using the test data, and where the entire composite dataset (formed by appending data from the source and target population) is split into a test and a training set that are used for model estimation and assessment. In some cases an established model is available (e.g., one developed using external data) and the goal of the analysis is limited to assessing model performance in the target population. In that case, no data from the source or target population need to be used for model development and all available data can be used to evaluate model performance and treated as a part of the “test set”. We should note here that provided the prediction model is correctly specified, exchangeability in mean over $S$, that is $\operatorname{\textnormal{\mbox{E}}}[Y|X,S=1]=\operatorname{\textnormal{\mbox{E}}}[Y|X,S=0]$, is sufficient for the parameter $\beta$ to be identifiable using data from the source population alone. Exchangeability in mean over $S$ is a weaker condition than condition A1; that is, condition A1 implies exchangeability in mean, but the converse is not true. Exchangeability in mean, however, is not sufficient for transporting measures of model performance, such as the MSE. In Appendix C we give an example of a setting where exchangeability in mean holds but it is not sufficient to identify the target population MSE. ## Model and tuning parameter selection Up to now we have proceeded as if the source population data in the training set are used to estimate parameters of a pre-specified parametric model, without employing any form of model selection (e.g., variable choice or other specification search) or tuning parameter selection. Yet, when developing prediction models, analysts often select between multiple different models and statistical learning algorithms usually have one or more tuning parameters. Importantly, data-driven methods for model and tuning parameter selection, such as cross-validation-based procedures, rely on optimizing some measure of model performance, such as the MSE. Consider, for instance, tuning parameter selection using $K$-fold cross- validation. In such an analysis, we split the data into $K$ mutually exclusive subsets (“folds”) and for each value of the tuning parameter we build the model with the selected tuning parameter value on $K-1$ of the folds and estimate a measure of model performance on the fold that is not used for model building. This process is repeated where each of the $K$ folds is left out of the model building process, resulting in $K$ estimates of model performance. The final cross-validated estimator of model performance associated with the tuning parameter value is the average of the $K$ estimators. The cross- validated value of the tuning parameter is selected as the value of the tuning parameter that optimizes the cross-validated estimator of model performance. Clearly, data-driven model and tuning parameter selection relies on estimating measures of model performance. Furthermore, tailoring the cross-validated model for use in the target population and tuning parameter selection to improve model performance for use in the target population require incorporating the results from the two preceding sections to account for differences in the distribution of covariates between the source and target population. Specifically, when prediction error modifiers have a different distribution in the source and the target population, cross-validated measures of model performance calculated using the source population data are biased estimators of model performance in the target population. Inverse-odds weighting estimators can adjust for that bias and failing to adjust for this bias when performing cross-validation is likely to lead to sub-optimal model or tuning parameter selection in the context of the target population. ## Illustration using simulated data In this section we use simulated data to illustrate (i) the performance of correctly and incorrectly specified prediction models when used with or without inverse-odds of participation weights; (ii) the potential for bias resulting from the naive (unweighted) MSE estimator that uses only the source population data to estimate the target population MSE; and (iii) the ability to adjust for that bias using the inverse-odds weighting estimator. ##### Data generation: We simulated the outcome using the linear model $Y=1+X+0.5X^{2}+\varepsilon$, where $\varepsilon\sim\mathcal{N}(0,X)$ and $X\sim Uniform(0,10)$. Under this model, the errors are heteroscedastic because the error variance directly depends on the covariate $X$. We simulated participation in the source data using a logistic regression model $\ln\left(\frac{\Pr[S=1|X]}{1-\Pr[S=1|X]}\right)=1.5-0.3X$. We set the total sample size to $1000$ and the source and target population data were randomly split in a 1:1 ratio into a training and a test set. Under this data generating mechanism, the target population MSE is larger than the source population MSE and both conditions A1 and A2 are satisfied. We considered two prediction models, a correctly specified linear regression model that included main effects of $X$ and $X^{2}$ and a misspecified linear regression model that only included the main effect of $X$. We also considered two approaches for estimating each posited prediction model: ordinary least squares regression (unweighted, OLS) and weighted least squares regression (WLS) where the weights were equal to the inverse of estimated odds of participation in the source data training set. We estimated the inverse-odds of participation in the training set, $\Pr[S=0|X,D_{\text{\tiny train}}=1]/\Pr[S=1|X,D_{\text{\tiny train}}=1]$, using a correctly specified logistic regression model for $\Pr[S=1|X,D_{\text{\tiny train}}=1]$. Figure 2 highlights the relationship between the correct model, and the large-sample limits of the weighted and unweighted misspecified models. For the inverse- odds weighting estimator $\widehat{\psi}_{\widehat{\beta}}$, we estimated the odds weights $\widehat{o}(X)$ in the test set by fitting a correctly specified logistic regression model for $\Pr[S=1|X,D_{\text{\tiny test}}=1]$ using the test set data. ##### Simulation results: The results from $10,000$ runs of the simulation are presented in Table 1. For both OLS and WLS estimation of the prediction model, the correctly specified model resulted in smaller average target population and source population MSE estimates compared with the misspecified model. When comparing the performance of OLS and WLS estimation of the prediction model in the target population OLS performed slightly better than WLS when the model was correctly specified (average MSE of $45.8$ vs. $46.2$). When the prediction model was incorrectly specified, OLS performed worse than WLS (average MSE of $66.3$ vs. $58.0$). The last column in the Table shows that the average of the inverse-odds weighting MSE estimator across the simulations was very close to the true target population MSE (obtained via numerical methods) for all combinations of model specifications and use of weights. In all scenarios of this simulation, the source population MSE estimator was substantially lower than the target population MSE. Hence, using the estimated source population MSE as an estimator for the target population MSE would lead to substantial underestimation of the MSE (i.e., showing model performance to be better than it is in the context of the target population). In contrast, the inverse-odds weighting estimator would give an accurate assessment of model performance in the target population. ## Nested designs Thus far, we have focused on the non-nested sampling design. Nested sampling designs are an alternative approach where the source population is a subset of the target population of interest [16, 18, 27]. Examples of such nested designs arise when the sample from the source population, from which outcome information is available, can be embedded within a larger cohort (e.g., via record linkage techniques) that can be viewed as representing the target population. Our results can be applied, with minor modifications, to nested designs. In Appendix B, we prove an identification result for nested designs and provide an estimator for loss-based measure of target population model performance. ## Discussion We considered transporting prediction models to a different population than was used for original model development, when outcome and covariate data are available on a simple random sample from the source population and covariate information is available on a simple random sample from the target population. We described the adjustments needed when the covariate distribution differs between the source and target population and provided identification results. We discussed how to tailor the prediction model to the target population and how to calculate measures of model performance in the context of the target population, without requiring the prediction model to be correctly specified. We also examined tailoring data-driven model and tuning parameter selection to the target population. The key insight is that most measures of model performance average over the covariate distribution and, as a result, estimators of these measures obtained in data from the source population will typically be biased for the corresponding measures in the target population, when the covariate distribution differs between the two populations. To simplify the exposition, throughout this paper we have assumed that the covariates needed to satisfy the conditional independence condition (A1) are the same as the covariates used in the prediction model. In practice, the set of covariates needed to satisfy condition A1 may be much larger than the set of covariates that are practically useful to include in the prediction model. The identifiability results in our paper can be easily modified to allow for the two sets of covariates to be different. ## References * [1] Ewout W Steyerberg et al. Clinical prediction models. Springer, 2019. * [2] Romin Pajouheshnia, Rolf HH Groenwold, Linda M Peelen, Johannes B Reitsma, and Karel GM Moons. When and how to use data from randomised trials to develop or validate prognostic models. BMJ, 365, 2019. * [3] Benjamin A Goldstein, Ann Marie Navar, Michael J Pencina, and John Ioannidis. Opportunities and challenges in developing risk prediction models with electronic health records data: a systematic review. Journal of the American Medical Informatics Association, 24(1):198–208, 2017. * [4] Steffen Bickel, Michael Brückner, and Tobias Scheffer. Discriminative learning for differing training and test distributions. In Proceedings of the 24th International Conference on Machine Learning, pages 81–88, 2007. * [5] Masashi Sugiyama, Matthias Krauledat, and Klaus-Robert MÞller. Covariate shift adaptation by importance weighted cross validation. Journal of Machine Learning Research, 8(May):985–1005, 2007. * [6] Sinno Jialin Pan, Ivor W Tsang, James T Kwok, and Qiang Yang. Domain adaptation via transfer component analysis. IEEE Transactions on Neural Networks, 22(2):199–210, 2010. * [7] Bin Cao, Xiaochuan Ni, Jian-Tao Sun, Gang Wang, and Qiang Yang. Distance metric learning under covariate shift. In Twenty-Second International Joint Conference on Artificial Intelligence, 2011. * [8] Masashi Sugiyama and Motoaki Kawanabe. Machine learning in non-stationary environments: introduction to covariate shift adaptation. MIT press, 2012. * [9] Wouter M Kouw and Marco Loog. An introduction to domain adaptation and transfer learning. arXiv preprint arXiv:1812.11806, 2018. * [10] Sentao Chen and Xiaowei Yang. Tailoring density ratio weight for covariate shift adaptation. Neurocomputing, 333:135–144, 2019. * [11] Masato Ishii, Takashi Takenouchi, and Masashi Sugiyama. Partially zero-shot domain adaptation from incomplete target data with missing classes. In The IEEE Winter Conference on Applications of Computer Vision, pages 3052–3060, 2020. * [12] Abhirup Datta, Jacob Fiksel, Agbessi Amouzou, and Scott L Zeger. Regularized bayesian transfer learning for population-level etiological distributions. Biostatistics, 2020. * [13] Stephen R Cole and Elizabeth A Stuart. Generalizing evidence from randomized clinical trials to target populations: the actg 320 trial. American Journal of Epidemiology, 172(1):107–115, 2010. * [14] Kara E Rudolph and Mark J van der Laan. Robust estimation of encouragement-design intervention effects transported across sites. Journal of the Royal Statistical Society. Series B, Statistical Methodology, 79(5):1509, 2017. * [15] Issa J Dahabreh, Sarah E Robertson, Jon A Steingrimsson, Elizabeth A Stuart, and Miguel A Hernán. Extending inferences from a randomized trial to a new target population. Statistics in Medicine, 39(14):1999–2014, 2020. * [16] Issa J Dahabreh, Sarah E Robertson, Eric J Tchetgen, Elizabeth A Stuart, and Miguel A Hernán. Generalizing causal inferences from individuals in randomized trials to all trial-eligible individuals. Biometrics, 75(2):685–694, 2019. * [17] Issa J Dahabreh and Miguel A Hernán. Extending inferences from a randomized trial to a target population. European Journal of Epidemiology, 34(8):719–722, 2019. * [18] Issa J Dahabreh, Sebastien JP Haneuse, James M Robins, Sarah E Robertson, Ashley L Buchanan, Elisabeth A Stuart, and Miguel A Hernán. Study designs for extending causal inferences from a randomized trial to a target population. arXiv preprint arXiv:1905.07764, 2019. * [19] Jeffrey M Wooldridge. Econometric analysis of cross section and panel data. MIT press, 2010. * [20] Guido W Imbens and Tony Lancaster. Efficient estimation and stratified sampling. Journal of Econometrics, 74(2):289–318, 1996. * [21] Hidetoshi Shimodaira. Improving predictive inference under covariate shift by weighting the log-likelihood function. Journal of Statistical Planning and Inference, 90(2):227–244, 2000\. * [22] Masashi Sugiyama, Taiji Suzuki, and Takafumi Kanamori. Density ratio estimation in machine learning. Cambridge University Press, 2012. * [23] Trevor Hastie, Robert Tibshirani, and Jerome Friedman. The elements of statistical learning: data mining, inference, and prediction. Springer Science &amp; Business Media, 2009. * [24] Issa J Dahabreh, James M Robins, and Miguel A Hernán. Benchmarking observational methods by comparing randomized trials and their emulations. Epidemiology, 31(5):614–619, 2020. * [25] Alex Luedtke, Marco Carone, and Mark J van der Laan. An omnibus non-parametric test of equality in distribution for unknown functions. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 81(1):75–99, 2019. * [26] Bianca Zadrozny. Learning and evaluating classifiers under sample selection bias. In Proceedings of the twenty-first international conference on Machine learning, page 114, 2004. * [27] Yi Lu, Daniel O Scharfstein, Maria M Brooks, Kevin Quach, and Edward H Kennedy. Causal inference for comprehensive cohort studies. arXiv preprint arXiv:1910.03531, 2019. ## Figures Figure 1: An example of a prediction error modifier, $X$. The top panel shows a scatter-plot of the data (including the unobserved target population outcomes) and the solid black line is the true conditional expectation function $\operatorname{\textnormal{\mbox{E}}}[Y|X,S=1]$. The middle panel shows the inverse-odds weights (IOW) as a function of $X$ and the bottom panel shows the conditional mean squared error (MSE) as a function of $X$. In these artificial data, larger values of $X$ have higher probability of being from the target population, $S=0$ (corresponding to lower odds of being from the source population and higher inverse-odds weights) and higher MSE. Hence, $X$ is a prediction error modifier that is differentially distributed between the source and the target population. This leads to the source population MSE being smaller than the target population MSE (0.47 versus 0.74). Figure 2: An example of simulated data used to illustrate transportability of prediction models. The solid curve is $\operatorname{\textnormal{\mbox{E}}}[Y|X,S=1]$, the dashed line is the large-sample limit when estimating the misspecified model without weighting, and the dotted line is the large-sample limit when estimating the misspecified model using inverse-odds weights. The weighted estimation gives more influence to observations with higher values of $X$, compared to unweighted estimation, because higher values of $X$ are associated with higher odds of a sampled observation being from the target population (i.e., lower odds of being from the source population, corresponding to higher inverse-odds weights). This is seen in the figure as for high values of $X$ the weighted model better approximates $\operatorname{\textnormal{\mbox{E}}}[Y|X,S=1]$ compared to the unweighted model, but the opposite is true for smaller values of $X$. ## Table Table 1: Target population mean squared error (MSE), average of the source data MSE estimators, and the estimators for the target population MSE that weight observations by the inverse-odds of being from the source population. | Model specification, --- estimation approach | True target --- population MSE | Average of unweighted --- MSE estimator | Average of weighted --- MSE estimator | Correctly specified, --- OLS 45.8 | 22.5 | 45.8 | Incorrectly specified, --- OLS 66.3 | 34.5 | 66.3 | Correctly specified, --- WLS 46.2 | 22.8 | 46.2 | Incorrectly specified, --- WLS 58.0 | 43.6 | 57.9 Correctly specified and incorrectly specified refers to the specification of the posited prediction model. OLS = model estimation using ordinary least squares regression (unweighted); WLS = model estimation using weighted least squares regression with weights equal to the inverse of the odds of being from the source population. Weighted MSE estimator results were obtained using the estimator in equation (3). Results were averaged over $10,000$ simulations. The true target population MSE was obtained using numerical methods. ## Appendix A Proofs of key results ### A.1 Identifiability for non-nested designs #### Proof of identifiability of target population MSE We will provide the identifiability result for a general loss function $L(Y,g_{\widehat{\beta}}(X))$. Many common performance measures, including the mean squared error, absolute error, and the Brier score, are special cases of expected loss functions. We define $D_{\text{ \tiny test}}$ as an indicator if an observation is in the source or target test data. ###### Proposition 1. Under conditions A1 and A2 and when the source and target data are obtained by separate simple random sampling of the corresponding underlying populations, with potentially unknown sampling probabilities, then the target population MSE, $\psi_{\widehat{\beta}}$, is identifiable as $\psi_{\widehat{\beta}}=\operatorname{\textnormal{\mbox{E}}}[\operatorname{\textnormal{\mbox{E}}}[(Y-g_{\widehat{\beta}}(X))^{2}|X,S=1,D_{\emph{\tiny test}}=1]|S=0,D_{\emph{\tiny test}}=1];$ (A.1) or, using an inverse-odds weighting representation, $\psi_{\widehat{\beta}}=\frac{1}{\Pr[S=0|D_{\emph{\tiny test}}=1]}\operatorname{\textnormal{\mbox{E}}}\left[\frac{I(S=1)\Pr[S=0|X,D_{\emph{\tiny test}}=1]}{\Pr[S=1|X,D_{\emph{\tiny test}}=1]}(Y-g_{\widehat{\beta}}(X))^{2}\bigg{|}D_{\emph{\tiny test}}=1\right].$ (A.2) All quantities in expressions (A.1) and (A.2) condition on $D_{\emph{\tiny test}}=1$ and can therefore be calculated using the available test data. ###### Proof. For the first representation we have $\displaystyle\psi_{\widehat{\beta}}$ $\displaystyle=\operatorname{\textnormal{\mbox{E}}}[L(Y,g_{\widehat{\beta}}(X))|S=0]$ $\displaystyle=\operatorname{\textnormal{\mbox{E}}}\left[\operatorname{\textnormal{\mbox{E}}}[L(Y,g_{\widehat{\beta}}(X))|X,S=0]|S=0\right]$ $\displaystyle=\operatorname{\textnormal{\mbox{E}}}\left[\int L(y,g_{\widehat{\beta}}(X))dF(y|X,S=0)\bigg{|}S=0\right]$ $\displaystyle=\operatorname{\textnormal{\mbox{E}}}\left[\int L(y,g_{\widehat{\beta}}(X))dF(y|X,S=1)\bigg{|}S=0\right]$ $\displaystyle=\operatorname{\textnormal{\mbox{E}}}\left[\operatorname{\textnormal{\mbox{E}}}[L(Y,g_{\widehat{\beta}}(X))|X,S=1]|S=0\right],$ where the first equality follows from the definition of $\psi_{\widehat{\beta}}$, the second from the law of iterated expectations, the third from the definition of conditional expectation, and the fourth from identifiability condition A1. All expectations conditional on $(X,S=1)$ in the above formula are well defined by the positivity condition A2. Rewrite $\displaystyle\psi_{\widehat{\beta}}$ $\displaystyle=\operatorname{\textnormal{\mbox{E}}}[\operatorname{\textnormal{\mbox{E}}}[L(Y,g_{\widehat{\beta}}(X))|X,S=1]|S=0]$ $\displaystyle=\int\operatorname{\textnormal{\mbox{E}}}[L(Y-g_{\widehat{\beta}}(X))|X=x,S=1]dF(x|S=0).$ The conditional expectation $\operatorname{\textnormal{\mbox{E}}}[L(Y,g_{\widehat{\beta}}(X))|X=x,S=1]$ is identifiable because, under the non-nested sampling design, data are available from a random sample of observations from the source population ($S=1$). Furthermore, the conditional distribution $F(x|S=0)$ is also identifiable because, under the non-nested sampling design, data are available from a random sample of observations from the target population ($S=0$). More formally, the random sampling ensures that $\psi_{\widehat{\beta}}=\operatorname{\textnormal{\mbox{E}}}[\operatorname{\textnormal{\mbox{E}}}[L(Y,g_{\widehat{\beta}}(X))|X,S=1,D_{\text{\tiny test}}=1]|S=0,D_{\text{\tiny test}}=1].$ For the inverse-odds weighting representation $\displaystyle\psi_{\widehat{\beta}}$ $\displaystyle=\operatorname{\textnormal{\mbox{E}}}[\operatorname{\textnormal{\mbox{E}}}[L(Y,g_{\widehat{\beta}}(X))|X,S=1,D_{\text{\tiny test}}=1]|S=0,D_{\text{\tiny test}}=1]$ $\displaystyle=\operatorname{\textnormal{\mbox{E}}}\left[\operatorname{\textnormal{\mbox{E}}}\left[\frac{I(S=1)}{\Pr[S=1|X,D_{\text{\tiny test}}=1]}L(Y,g_{\widehat{\beta}}(X))\bigg{|}X,D_{\text{\tiny test}}=1\right]\bigg{|}S=0,D_{\text{\tiny test}}=1\right]$ $\displaystyle=\frac{1}{\Pr[S=0|D_{\text{\tiny test}}=1]}\operatorname{\textnormal{\mbox{E}}}\left[I(S=0)\operatorname{\textnormal{\mbox{E}}}\left[\frac{I(S=1)}{\Pr[S=1|X,D_{\text{\tiny test}}=1]}L(Y,g_{\widehat{\beta}}(X))\bigg{|}X,D_{\text{\tiny test}}=1\right]\Bigg{|}D_{\text{\tiny test}}=1\right]$ $\displaystyle=\frac{1}{\Pr[S=0|D_{\text{\tiny test}}=1]}\operatorname{\textnormal{\mbox{E}}}\left[\operatorname{\textnormal{\mbox{E}}}\left[\frac{I(S=1)\Pr[S=0|X,D_{\text{\tiny test}}=1]}{\Pr[S=1|X,D_{\text{\tiny test}}=1]}L(Y,g_{\widehat{\beta}}(X))\bigg{|}X,D_{\text{\tiny test}}=1\right]\bigg{|}D_{\text{\tiny test}}=1\right]$ $\displaystyle=\frac{1}{\Pr[S=0|D_{\text{\tiny test}}=1]}\operatorname{\textnormal{\mbox{E}}}\left[\frac{I(S=1)\Pr[S=0|X,D_{\text{\tiny test}}=1]}{\Pr[S=1|X,D_{\text{\tiny test}}=1]}L(Y,g_{\widehat{\beta}}(X))\bigg{|}D_{\text{\tiny test}}=1\right].$ For the fourth equality we have used that $\displaystyle\operatorname{\textnormal{\mbox{E}}}\left[I(S=0)\operatorname{\textnormal{\mbox{E}}}\left[\frac{I(S=1)}{\Pr[S=1|X,D_{\text{\tiny test}}=1]}L(Y,g_{\widehat{\beta}}(X))\Bigg{|}X,D_{\text{\tiny test}}=1\right]\Bigg{|}D_{\text{\tiny test}}=1\right]$ $\displaystyle=\operatorname{\textnormal{\mbox{E}}}\left[\operatorname{\textnormal{\mbox{E}}}\left[I(S=0)\operatorname{\textnormal{\mbox{E}}}\left[\frac{I(S=1)}{\Pr[S=1|X,D_{\text{\tiny test}}=1]}L(Y,g_{\widehat{\beta}}(X))\Bigg{|}X,D_{\text{\tiny test}}=1\right]\Bigg{|}X,D_{\text{\tiny test}}=1\right]\Bigg{|}D_{\text{\tiny test}}=1\right]$ $\displaystyle=\operatorname{\textnormal{\mbox{E}}}\left[\operatorname{\textnormal{\mbox{E}}}\left[\frac{I(S=1)}{\Pr[S=1|X,D_{\text{\tiny test}}=1]}L(Y,g_{\widehat{\beta}}(X))\Bigg{|}X,D_{\text{\tiny test}}=1\right]\operatorname{\textnormal{\mbox{E}}}[I(S=0)|X,D_{\text{\tiny test}}=1]\Bigg{|}D_{\text{\tiny test}}=1\right]$ $\displaystyle=\operatorname{\textnormal{\mbox{E}}}\left[\operatorname{\textnormal{\mbox{E}}}\left[\frac{I(S=1)\Pr[S=0|X,D_{\text{\tiny test}}=1]}{\Pr[S=1|X,D_{\text{\tiny test}}=1]}L(Y,g_{\widehat{\beta}}(X))\Bigg{|}X,D_{\text{\tiny test}}=1\right]\Bigg{|}D_{\text{\tiny test}}=1\right]$ All of the quantities in $\frac{1}{\Pr[S=0|D_{\text{\tiny test}}=1]}\operatorname{\textnormal{\mbox{E}}}\left[\frac{I(S=1)\Pr[S=0|X,D_{\text{\tiny test}}=1]}{\Pr[S=1|X,D_{\text{\tiny test}}=1]}L(Y,g_{\widehat{\beta}}(X))\bigg{|}D_{\text{\tiny test}}=1\right].$ condition on $D_{\text{\tiny test}}=1$ and are therefore identifiable using the observed data. ∎ #### Proof of identifiability of inverse-odd weights Let $D_{\text{\tiny train}}$ be an indicator if data from an observation is in the training set and used to estimate the inverse-odds weights. The sampling design assumes that $\Pr[D_{\text{\tiny train}}=1|X,S=1]=a$ for some potentially unknown constant $a>0$; and $\Pr[D_{\text{\tiny train}}=1|X,S=0]=b$ for some potentially unknown constant $b>0$. By the random formation of the test and the training set, the inverse-odds weights in the test and the training set are equal. But, to ensure independence between the data used to train the model and the data used to evaluate the model we propose to use inverse-odds weights estimated using the training set for model building and the inverse-odds weights estimated using the test set for estimating model performance. #### Proof of expression 1 from the main text Recall that the sampling design assumes that $\Pr[D_{\text{\tiny train}}=1|X,S=1]=a$ for some potentially unknown constant $a>0$ and $\Pr[D_{\text{\tiny train}}=1|X,S=0]=b$ for some potentially unknown constant $b>0$. Using that, we have $\displaystyle\frac{\Pr[S=0|X,D_{\text{\tiny train}}=1]}{\Pr[S=1|X,D_{\text{\tiny train}}=1]}$ $\displaystyle=\frac{\Pr[S=0,D_{\text{\tiny train}}=1|X]}{\Pr[S=1,D_{\text{\tiny train}}=1|X]}$ $\displaystyle=\frac{\Pr[S=0|X]}{\Pr[S=1|X]}\times\frac{\Pr[D_{\text{\tiny train}}=1|X,S=0]}{\Pr[D_{\text{\tiny train}}=1|X,S=1]}$ $\displaystyle=\frac{\Pr[S=0|X]}{\Pr[S=1|X]}\times\frac{\Pr[D_{\text{\tiny train}}=1|S=0]}{\Pr[D_{\text{\tiny train}}=1|S=1]}$ $\displaystyle=\frac{\Pr[S=0|X]}{\Pr[S=1|X]}\times\frac{b}{a}$ $\displaystyle\propto\frac{\Pr[S=0|X]}{\Pr[S=1|X]}.$ ∎ ## Appendix B Identification and estimation in nested designs Consider a nested design where the source population is a subset of a larger target population of interest. We assume that covariate data, $X$, are available on all target population members, but outcome data, $Y$, are only available on everyone in the source population. The data is assumed to be realizations of $\\{(X_{i},S_{i},S_{i}\times Y_{i},i=1,\ldots,n\\},$ where $n$ is the total number of observations (i.e., the total number of individuals in a cohort representing the target population and in which the sample from the source population is nested) and $S$ is the indicator of an observation coming from the source population ($S=1$ for observations in the source population and $S=0$ for observations not in the source population). For nested designs the target parameter is defined as $\phi_{\widehat{\beta}}=\operatorname{\textnormal{\mbox{E}}}\left[L(Y,g_{\widehat{\beta}}(X))\right].$ We introduce the following modified identifiability conditions: 1. B1. For every $x$ such that $f(X=x)\neq 0$, $f(Y|X=x,S=1)=f(Y|X=x).$ 2. B2. For every $x$ such that $f(X=x)\neq 0$, $\Pr[S=1|X=x]>0$ . ###### Proposition 2. Under conditions B1 and B2, $\phi_{\widehat{\beta}}$ can be written as the observed data functional $\phi_{\widehat{\beta}}=\operatorname{\textnormal{\mbox{E}}}\left[\operatorname{\textnormal{\mbox{E}}}\left[L(Y,g_{\widehat{\beta}}(X))\big{|}X,S=1,D_{\text{\tiny test}}=1\right]D_{\text{\tiny test}}=1\right].$ (A.3) Or using the inverse probability weighting representation $\phi_{\widehat{\beta}}=\operatorname{\textnormal{\mbox{E}}}\left[\frac{I(S=1)}{\Pr[S=1|X,D_{\text{\tiny test}}=1]}L(Y,g_{\widehat{\beta}}(X))\Bigg{|}D_{\text{\tiny test}}=1\right].$ (A.4) #### Proof of Proposition 2: We have $\displaystyle\phi_{\widehat{\beta}}$ $\displaystyle=\operatorname{\textnormal{\mbox{E}}}\left[L(Y,g_{\widehat{\beta}}(X))\right]$ $\displaystyle=\operatorname{\textnormal{\mbox{E}}}\left[\operatorname{\textnormal{\mbox{E}}}\left[L(Y,g_{\widehat{\beta}}(X))\big{|}X\right]\right]$ $\displaystyle=\operatorname{\textnormal{\mbox{E}}}\left[\operatorname{\textnormal{\mbox{E}}}\left[L(Y,g_{\widehat{\beta}}(X))\big{|}X,S=1\right]\right]$ $\displaystyle=\operatorname{\textnormal{\mbox{E}}}\left[\operatorname{\textnormal{\mbox{E}}}\left[L(Y,g_{\widehat{\beta}}(X))\big{|}X,S=1,D_{\text{\tiny test}}=1\right]\Big{|}D_{\text{\tiny test}}=1\right].$ For the inverse probability weighting representation $\displaystyle\phi_{\widehat{\beta}}$ $\displaystyle=\operatorname{\textnormal{\mbox{E}}}\left[\operatorname{\textnormal{\mbox{E}}}\left[L(Y,g_{\widehat{\beta}}(X))\big{|}X,S=1,D_{\text{\tiny test}}=1\right]\Bigg{|}D_{\text{\tiny test}}=1\right]$ $\displaystyle=\operatorname{\textnormal{\mbox{E}}}\left[\operatorname{\textnormal{\mbox{E}}}\left[\frac{I(S=1)}{\Pr[S=1|X,D_{\text{\tiny test}}=1]}L(Y,g_{\widehat{\beta}}(X))\bigg{|}X,D_{\text{\tiny test}}=1\right]\Bigg{|}D_{\text{\tiny test}}=1\right]$ $\displaystyle=\operatorname{\textnormal{\mbox{E}}}\left[\frac{I(S=1)}{\Pr[S=1|X,D_{\text{\tiny test}}=1]}L(Y,g_{\widehat{\beta}}(X))\Bigg{|}D_{\text{\tiny test}}=1\right]$ $\displaystyle=\frac{1}{\Pr[D_{\text{\tiny test}}=1]}\operatorname{\textnormal{\mbox{E}}}\left[\frac{I(S=1,D_{\text{\tiny test}}=1)}{\Pr[S=1|X,D_{\text{\tiny test}}=1]}L(Y,g_{\widehat{\beta}}(X))\right],$ which establishes the identifiability of $\phi_{\widehat{\beta}}$. ∎ Using plug-in estimators into identifiability expression (A.4) gives the inverse probability weighting estimator for nested designs. That is, $\widehat{\phi}_{\widehat{\beta}}=\frac{\sum_{i=1}^{n}\frac{I(S_{i}=1,D_{\text{\tiny test},i}=1)}{\widehat{p}(X_{i})}L(Y_{i},g_{\widehat{\beta}}(X_{i}))}{\sum_{i=1}^{n}I(D_{\text{\tiny test},i}=1)},$ where $\widehat{p}(X)$ is an estimator for $\Pr[S=1|X,D_{\text{\tiny test}}=1]$. ## Appendix C Inverse-odds weighting estimators can be biased under mean exchangeability For correctly specified prediction models, exchangeability in mean over $S$, that is $\operatorname{\textnormal{\mbox{E}}}[Y|X,S=1]=\operatorname{\textnormal{\mbox{E}}}[Y|X,S=0]$, is sufficient for the parameter $\beta$ to be identifiable using data from the source population alone. Exchangeability in mean over $S$ is a weaker condition than condition A1; that is, condition A1 implies exchangeability in mean, but the converse is not true. Exchangeability in mean, however, is insufficient for transportability of the MSE. This can be seen in Figure 3 where $\operatorname{\textnormal{\mbox{E}}}[Y|X,S=1]=\operatorname{\textnormal{\mbox{E}}}[Y|X,S=0]$ but $\mbox{Var}[Y|X,S=1]\neq\mbox{Var}[Y|X,S=0]$ (and thus assumption A1 does not hold). As the conditional variance is different between the two populations, standardizing to the target population covariate distribution is not sufficient to transport the MSE to the target population. If the outcome is binary, condition A1 can be written as $\Pr[Y=1|X=x,S=1]=\Pr[Y=1|X=x,S=0]$, so for binary outcomes distributional independence over $S$ is equivalent to exchangeability in mean over $S$. Figure 3: An example of a setting where condition A1 does not hold. Here, $\operatorname{\textnormal{\mbox{E}}}[Y|X,S=1]=\operatorname{\textnormal{\mbox{E}}}[Y|X,S=0]$ (the black line is the true conditional mean for both populations), but $\mbox{Var}[Y|X,S=1]<\mbox{Var}[Y|X,S=0]$ for all values of $X$. In this case, estimators of model performance measures that use weights equal to the inverse-odds of being from the source population (e.g., the MSE estimator in the main text of the paper) will be biased.
# GymD2D: A Device-to-Device Underlay Cellular Offload Evaluation Platform David Cotton 0000-0002-8817-3736 School of Electrical and Data Engineering University of Technology Sydney Sydney, Australia <EMAIL_ADDRESS>Zenon Chaczko School of Electrical and Data Engineering University of Technology Sydney Sydney, Australia <EMAIL_ADDRESS> ###### Abstract Cellular offloading in device-to-device communication is a challenging optimisation problem in which the improved allocation of radio resources can increase spectral efficiency, energy efficiency, throughout and reduce latency. The academic community have explored different optimisation methods on these problems and initial results are encouraging. However, there exists significant friction in the lack of a simple, configurable, open-source framework for cellular offload research. Prior research utilises a variety of network simulators and system models, making it difficult to compare results. In this paper we present GymD2D, a framework for experimentation with physical layer resource allocation problems in device-to-device communication. GymD2D allows users to simulate a variety of cellular offload scenarios and to extend its behaviour to meet their research needs. GymD2D provides researchers an evaluation platform to compare, share and build upon previous research. We evaluated GymD2D with state-of-the-art deep reinforcement learning and demonstrate these algorithms provide significant efficiency improvements. ###### Index Terms: device-to-device (D2D) communication, cellular offload, resource allocation, radio resource management, network simulator, deep reinforcement learning, OpenAI Gym ††publicationid: pubid: ©2021 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. ## I Introduction Wireless data use is increasing rapidly presenting a challenge for cellular service providers. The explosion of video and smart device traffic has intensified demand for limited cellular resources [1]. A multi-faceted approach to developing next generation networks include: unlocking millimeter- wave frequencies, increasing cell density, multiple-input multiple-output, improved information encoding, and the focus of this research—smarter coordination [2]. Device-to-device (D2D) communication is a broad set of protocols for ad-hoc, peer-to-peer wireless communication for cellular or Internet connected wireless devices. In contrast to normal cellular operation, user equipment (UE) utilising D2D mode communicate directly with one another instead of through base stations and connected networks [3]. D2D has been proposed for a variety of use cases such as: public safety communications, communications relay, localised services, Internet of things and cellular offload [4]. Figure 1: Example of D2D cellular offloading with underlay networks. This example demonstrates the cellular offload optimisation problem. Depicted is a BS, CUE and 2 DUE pairs. The BS has 2 RBs available to allocate. The BS has assigned the CUE RB1 for downlink and has the option of either assigning pair A RB1 and pair B RB2 or vice versa. If the BS assigns RB1 to pair A it is likely that due to their proximity to the CUE, significant interference would occur. As pair B is situated relatively distant to the CUE, if pair B transmitted with a lower power level, their interference on the CUE would be negligible. Therefore the optimal solution is to assign RB1 to pair B and RB2 to pair A. In this research we are interested in cellular offload, a mechanism to divert cellular traffic through alternative channels to improve network efficiency. This can be achieved by communicating out of band, such as over WiFi, or inband via overlay or underlay networking. Underlay cellular offload is a form of opportunistic spectrum access in which D2D UE (DUE) act as secondary users sharing radio resources with primary cellular UE (CUE) [5]. In underlay networking, DUE are responsible for managing their interference with CUE so as to avoid degrading primary network performance. Radio resource management (RRM) is the system wide management of radio resources, such as transmit power and resource blocks (RB), across a wireless network to manage interference and utilise resources as efficiently as possible. In cellular systems, which are limited by co-channel interference, improved resource allocation can increase spectral efficiency, energy efficiency, throughput and reduce latency. We show a simplified example of the resource allocation problem in Figure 1. A challenge facing researchers developing D2D algorithms is the need for improved tooling, reliable measurements and established benchmarks. An open- source evaluation platform allows researchers to compare results, share algorithms and build upon previous research. In this paper we present GymD2D, a network simulator and evaluation platform for RRM in D2D underlay cellular offload. GymD2D provides convenient abstractions to aid researchers in quick prototyping of resource allocation algorithms. The toolkit allows users to programmatically configure the environment to simulate a wide variety of scenarios. GymD2D has been designed with extensibility as a core design principle, allowing users to override and extend its behaviour to meet their research needs. It has been developed in the Python programming language to allow users to leverage its extensive ecosystem of scientific computing packages. The open-source nature of GymD2D centralises development effort and avoids the redundant work of individual researchers creating their own simulators. This puts more eyes on bug fixing, provides a more stable platform and increases confidence in reported empirical results. GymD2D reduces entry barriers for junior researchers, helps researchers from other disciplines to cross-pollinate ideas easier and more generally increases participation. Our software package is provided to the community under a MIT licence at https://github.com/davidcotton/gym-d2d. ## II Background ### II-A Device-to-device communication In this section we provide an overview of the D2D RRM literature to situate the reader as to the requirements of the platform. Firstly, we highlight the most common optimisation problems. Secondly, we analyse key differences across simulators, paying special attention to simplifying assumptions frequently observed. Thirdly, we survey the optimisation algorithms used for resource allocation. Finally, we outline limitations of existing research, providing direction for the simulation requirements of future work. #### II-A1 Optimisation Problems RRM is an optimisation problem where the objective is to utilise radio resources as efficiently as possible. D2D RRM has been proposed for improving spectral efficiency, energy efficiency and quality of service. In these uses cases, the objective can be to optimise data rates, throughput, capacity, signal to interference noise ratio (SINR), power consumption, energy efficiency or latency [4, 6]. D2D systems can be centrally managed by the network operator, manage interference autonomously or use a hybrid control mode which aims to combine the benefits of both. The choice of control mode limits the applicability of certain algorithms which may only be feasible in centrally managed paradigms. #### II-A2 Simulation models D2D cellular offload typically investigates networks using orthogonal frequency division multiple access (OFDMA), communicating on licensed bands using underlay networking. The most common scenario is a single macro base stations (MBS) surrounded by many randomly positioned CUEs and DUEs. It is generally assumed that cellular systems are under full load and each RB is allocated to a CUE. In the literature, simulations vary in scope from 2–30 RBs, 2–30 CUEs and 2–60 DUEs, while MBS operate with a cell radius of 20–500m. It is frequently assumed that DUE are already paired and operate in a range between 10–30m apart. Typically, omni-directional antenna and isotropic propagation are utilised. Path loss is commonly modeled using log-distance models, with or without shadowing. #### II-A3 Optimisation Algorithms A wide variety of optimisation methods have been investigated on a range of D2D RRM problems. Initially, D2D radio resources were proposed to be managed using existing cellular uplink power control mechanisms [7]. Consequently, it was identified that resources could be more efficiently allocated with the use of mathematical optimisation [8]. However, due to the computational complexity of these methods and the millisecond timescales involved, it may not feasible to solve to optimality. This can be addressed with the use of greedy heuristic algorithms which reduce computational complexity at the cost of global optimality [9]. Alternatively, resource allocation can be optimised graph theoretically [10], game theoretically [11], with evolutionary algorithms [12] or reinforcement learning (RL) [13]. More recently, deep reinforcement learning (DRL), a subfield of RL which uses deep neural networks to represent policy and/or value functions has demonstrated promising results [14, 15]. DRL is well suited for many D2D RRM problems as neural networks provide rich approximations, scale well and generalise to unseen data. #### II-A4 Research limitations A common limitation observed in D2D cellular offload research is BSs not enforcing uplink power control for CUE, who transmit at maximum power, a very energy inefficient approach. Another research challenge is accounting for large SINR increases on the primary network, such as could significantly impact primary network throughput or drive up CUE transmit power levels. Resource allocations algorithms need to demonstrate their effectiveness in larger search spaces that more closely reflect real world demands. Iterative learning algorithms need to be capable of generalising to out of training distribution data and be robust under diverse propagation conditions. Lastly, in our opinion, one of the greatest limitations of existing research is the lack of established benchmarks and comparison with other algorithms. ### II-B OpenAI Gym OpenAI Gym is an open-source software toolkit for reinforcement learning (RL) [16]. Gym provides an abstraction layer that enables a variety of tasks, known as environments in RL parlance, to be wrapped to present a consistent interface. The abstraction provided by Gym allows the easy interchange of algorithms and environments. This makes it is easy to test how a single algorithm generalises across a diverse set of environments or to benchmark different algorithms on a given environment. The simplicity and flexibility Gym offers has proved very popular and has lead to it becoming the de facto environment format in RL. While Gym was designed for RL research, the application programming interface (API) it provides makes it easy to apply many other algorithms types. ### II-C Network simulation One of the most widely used network simulators in education and research is ns-3. Ns-3 is an open-source, modular, discrete-event simulator for wired and wireless networks. It provides the full TCP/IP stack and wireless propagation modelling. Another popular alternative with comparable features is OMNeT++, while there exists similar commercial tools such as NetSim and MATLAB. Ns-3 has been incorporated into an OpenAI Gym environment under the ns3-gym project [17]. ## III Design Principles The design of GymD2D has been inspired by the authors experience developing and comparing reinforcement learning algorithms. In our experience the following design principles stimulate experimentation and the sharing of ideas. * • Simple: Easy to get started with, the framework should allow researchers to be productive quickly. * • Configurable: The framework should be easily configured to meet the broad range of D2D cellular offload use cases. Configurability allows researchers to programmatically test algorithm generalisation and scalability. * • Extensible: The framework should allow users to extend the system’s behaviour to meet their needs. The nature of research dictates a stream of new ideas we can’t anticipate but we can provide researchers the flexibility to adapt. * • Scalable: The framework should be performant and easily parallelisable. Developing new algorithms requires significant experimentation and reducing the time spent waiting for results is important for productivity. Some algorithms, such as policy gradient DRL, require parallel environments to function. Real world solutions are often a combination of both algorithmic and architectural components. * • Reproducible: Experiments should be easily repeatable. To build confidence in our deductions, it is important that we can reperform experiments to ensure the observed outcomes were not statistical anomalies. Reproducibility allows researchers to share their contributions with community more easily. ## IV System Design ### IV-A System model GymD2D is designed to simulate physical layer resource allocation problems in D2D underlay cellular offload. The framework abstracts away data link and above layers, D2D session establishment and management concerns. The system model contains a single MBS $b$, a set of $M$ CUEs $\mathcal{M}=\\{1,\dots,M\\}$ and a set of $N$ DUE pairs $\mathcal{N}=\\{1,\dots,N\\}$, that reside within the coverage area of the cell. We denote the $m^{th}$ CUE $C_{m}$, the $n^{th}$ DUE pair $D_{n}$ and the transmitter and receiver of pair $D_{n}$ by $D_{n}^{t}$ and $D_{n}^{r}$ respectively. The system employs OFDMA, with a set of $K$ RBs $k\in\mathcal{K}$ are available for allocation. An assumption is made that all devices are equipped with omni-directional antenna and transmit isotropically. Accordingly, the network resides within a circular cell of radius $R$, with the MBS located in the centre at position $(0,0)$. The simulation environment contains no obstructions or outside interference. D2D communicate one-to-one and D2D relay is not supported. We denote the effective isotropic radiated power (EIRP) $P$ of BSs, CUEs and DUEs as $P^{b}$, $P^{c}$, $P^{d}$ respectively. The EIRP of a BS is calculated, $P^{b}=P_{tx}-10log_{10}s+g_{ant}-l_{ix}-l_{cb}+g_{amp}$ (1) and the EIRP of CUE and DUE, $P^{c}=P^{d}=P_{tx}-10log_{10}s+g_{ant}-l_{ix}-l_{bd}$ (2) where $P_{tx}$ is the transmission power level, $s$ is the number of subcarriers, $g_{ant}$ is the transmitting antenna gain, $l_{ix}$ is the interference margin loss to approximate noise from surrounding cells, $l_{bd}$ is body loss to approximate attenuation caused by the user, $l_{cb}$ is cable loss, and $g_{amp}$ is amplifier gain. We denote the received signal level $R$ from transmitter $i$ at receiver $j$ of BS, CUEs and DUEs as $R^{b}_{i,j}$, $R^{c}_{i,j}$, $R^{d}_{i,j}$. The received signal level of BS as, $R^{b}_{i,j}=P_{i}-PL_{i,j}+g_{ant}-l_{cb}+g_{amp}$ (3) and the received signal level of CUE or DUE, $R^{c}_{i,j}=R^{d}_{i,j}=P_{i}-PL_{i,j}+g_{ant}-l_{bd}$ (4) where $P_{i}$ is the EIRP from transmitter $i$ and $PL_{i,j}$ is the path loss of the chosen path loss model between $i$ and $j$. We assume D2D transmissions are synchronised to cellular transmissions and occupy the same $K$ orthogonal resources. During both uplink and downlink, co- channel interference is calculated for each receiver sharing RB $k$. GymD2D considers co-channel interference between: * • D2D to cellular, interference from secondary DUE on the primary cellular network, * • cellular to D2D, interference from CUE or BS to DUE, and * • D2D to D2D, the interference between DUE pairs sharing a RB. Accordingly, we model the instantaneous SINR $\xi$ of receiver $j$ from transmitter $i$ on RB $k$, $\xi_{i,j,k}=\frac{R_{i,j}}{\sum_{n\in\mathcal{T}_{k},n\neq i}R_{n,j}+\sigma^{2}}$ (5) where $\mathcal{T}_{k}$ is the set of transmitters allocated to RB $k$ and $\sigma^{2}$ is additive white Gaussian noise (AWGN). The capacity of channel $C_{i,j}$ can be calculated using the SINR $\xi_{i,j}$, $C_{i,j}[Mbps]=Blog_{2}(1+\xi_{i,j})$ (6) where $B$ is the channel bandwidth in MHz. ### IV-B Path loss models GymD2D contains several of the most common path loss models and makes it easy for users to implement their own custom models. By default, GymD2D uses the simplest model, free space path loss (FSPL), $FSPL(f,d)[dB]=10nlog_{10}\Big{(}\frac{4\pi fd}{c}\bigg{)}$ (7) where $n=2$ is the path loss exponent (PLE) in free space, $f$ is the carrier frequency in Hz, $d$ is the distance between the transmitter and receiver and $c$ is the speed of light in m/s. To simulate obstructed propagation environments it can be useful to model fading effects as random processes. One such model is the log-distance with shadowing path loss model, which is included in GymD2D. The log-distance path loss model extends FSPL to mimic random shadowing effects, such as caused by buildings, with a log-normal distribution, $PL^{LD}(f,d)[dB]=FSPL(f,d_{0})+10nlog_{10}\frac{d}{d_{0}}+\chi_{\sigma}$ (8) where $d_{0}$ is an arbitrary close-in reference distance, typically 1–100m and $\chi_{\sigma}$ is a zero-mean Gaussian with standard deviation $\sigma$ in dB. Empirical measurements have shown values of $n=2.7\text{ to }3.5$ to be suitable to model urban environments [18]. ### IV-C Architecture GymD2D consists of two main components, a network simulator and a Gym environment. The network simulator models physical layer cellular networking. The Gym environment provides an abstraction layer to allow researchers to experiment with different simulation parameters and algorithms programmatically. Users supply RRM algorithms to manage the wireless devices under simulation. GymD2D outputs data on the state of the simulation to the user; to allow the effectiveness of RRM algorithms to be studied, such as through visualisation. A high level overview of the architecture of GymD2D is depicted in Figure 2. Figure 2: Proposed GymD2D architecture. GymD2D consists of a network simulator, wrapped by an OpenAI Gym environment. The user creates their own RRM algorithms to control wireless devices. ### IV-D Network simulator The network simulator models a single cellular cell which is populated with a collection of randomly placed CUEs and DUE pairs. It is a configurable component which can be customised to emulate a range of cellular offload scenarios. This includes the number and configuration of BSs, CUEs and DUEs and environmental parameters such as the available RBs, cell size and path loss model. The main components of the network simulator are: a collection of wireless devices (BSs, CUEs, DUEs), a path loss model and a traffic model as shown in the class diagram in Figure 3. Figure 3: Network simulator architecture. The main components of the network simulator are a collection of CUEs, DUEs and BS, the path loss model and the traffic model. Each simulation, the actions of BS and UEs within the cell can be generated internally by the traffic model or externally from a user defined RRM algorithm. A typical use case would be to use the internal traffic model to control BS and CUEs and the user RRM algorithm the DUEs. GymD2D uses a discrete-event simulation model. This method is congruent with the Gym API in which the incoming actions are the events and the Gym step() method calls equate to the system update intervals and model a single LTE or NR frame. At each step, each device may transmit, receive, or take no action. An action is tuple consisting of a transmitter, receiver, communication mode, RB and transmission power. The simulator consolidates the actions from both the traffic model and the RRM algorithm, then calculates the resulting propagation and interference. After calculating propagation, metrics on the state of the network, such as SINR and throughput, are output to the Gym environment. ### IV-E Gym environment The Gym environment has been designed to be configuration driven, to facilitate the programmatic scheduling and reproducibility of experiments. When instantiating a new Gym environment, configuration can be provided to specify the BSs, CUEs and DUEs that inhabit the simulation and the environmental conditions. The Gym environment provides RRM algorithms with an observation and action space. These provide a mapping to configure for the expected format of inputs and outputs. For example, when using a DRL algorithm, this would allow DRL to configure its neural networks for the shape of incoming observations and output actions of the correct dimension. At each step, the Gym environment receives actions from the RRM algorithm and converts them to a format suitable for the network simulator. Once a simulation step is complete, the environment uses the state of the simulator to create the observations and rewards RRM algorithms consume to make their decisions. ## V Evaluation ### V-A Methodology Figure 4: Evaluation results. To evaluate GymD2D we compared the performance of several state-of-the-art DRL algorithms in their efficiency allocating radio resources as D2D demand increased. Solid lines indicate the mean algorithm performance across ten trials with the shaded area the 95% confidence interval. The dashed red line indicates the baseline total system capacity without D2D communication. (a) The total system capacity of the DRL and random agent. (b) The total system capacity of just the DRL agents. (c) The total DUE capacity of the DRL agents. (d) The mean transmit power of all agents. We evaluated GymD2D with several leading DRL algorithms to determine their efficiency allocating radio resources as D2D demand increased. The objective was to maximise the total system capacity, that is the sum data rate of all CUE and DUE, calculated for each transmitter/receiver pair $i,j$ by $C_{i,j}[Mbps]=\begin{cases}Blog_{2}(1+\xi_{i,j})&\xi_{i,j}\geq\rho_{j}\\\ 0&\xi_{i,j}<\rho_{j}\\\ \end{cases}$ (9) where $B=0.18$ is the RB bandwidth in MHz and $\rho_{b}=-123.4$ and $\rho_{d}=-107.5$ is the receiver sensitivity of a BS and DUE respectively in dBm. Our evaluation simulated a single cell under full load. The scenario contained 25 RBs and CUEs, with each CUE allocated an individual RB. We employed a centrally managed control mode in which DUE communicated in the uplink frame, with the resource allocation managed by the network operator. Each RRM algorithm was evaluated with 10, 20, 30, 40 and 50 communicating D2D pairs. Algorithms were compared by training to convergence, then evaluating for 100 episodes. For each algorithm–D2D link density comparison, we conducted ten trials, retraining from scratch and evaluating, to account for variations in performance. Each episode lasted for ten steps or equivalently ten LTE/NR frames to simulate short bursts of traffic on a busy network. In each episode all CUE and DUE remained geographically fixed, but at the end of each episode, all CUE and DUE were randomly repositioned within the cell to simulate new devices accessing the network. Wireless propagation was modelled using the Log-Distance Shadowing model (8) with PLE $n=2.0$ and $\chi_{\sigma}=2.7$. The simulation parameters are detailed in Table I. TABLE I: Simulation parameters Parameter | Value ---|--- Cell radius | $500$ m Maximum D2D pair distance | $30$ m Carrier frequency | $2.1$ GHz RB bandwidth | $180$ kHz Number of RBs | $25$ Number of CUEs | $25$ Number of DUE pairs | $10,20,30,40,50$ CUE transmit power | $23$ dBm DUE min, max transmit power | $0,20$ dBm Path loss model | Log-Distance Shadowing Path loss exponent | $2.0$ Shadowing SD $\chi_{\sigma}$ | 2.7 We evaluated three DRL algorithms, Rainbow DQN [19], Discrete Soft Actor- Critic (SAC) [20], and Advantage Actor-Critic (A2C) [21], and a random agent baseline. All DRL algorithms used a fully connected neural network with two hidden layers trained using the Adam optimiser. Each hidden layer contained 128 units and used ReLU activation between layers. Policies used a reward discounting factor of $\gamma=0.9$. Our Rainbow DQN (Table II) implementation used distributional, dueling, double-Q and noisy networks with a prioritised replay buffer and single step returns. The discrete action space variant of the SAC (Table III) was used. A2C (Table IV) is the synchronous version of A3C and used Generalised Advantage Estimator (GAE) [22] with $\lambda=1.0$. TABLE II: Rainbow DQN hyperparameters Parameter | Value ---|--- Discounting factor $\gamma$ | $0.9$ Learning rate $\alpha$ | $5\cdot 10^{-4}$ Batch size | $32$ Online network update period | $4$ steps Learning start | $1,000$ steps Target network sync period | $500$ steps Distributional atoms | 51 Distributional bounds $v_{min}$, $v_{max}$ | [-1,10] Replay buffer capacity | $50,000$ Replay buffer prioritisation exponent $\omega$ | $0.6$ Replay buffer importance sampling $\beta$ | $0.6\rightarrow 0.4$ Importance sampling annealing | 20,000 steps TABLE III: SAC hyperparameters Parameter | Value ---|--- Discounting factor $\gamma$ | $0.9$ Learning rate $\alpha$ | $3\cdot 10^{-4}$ Batch size | $256$ Learning start | $1,500$ steps Target smoothing coefficient | 0.005 Replay buffer capacity | $50,000$ TABLE IV: A2C hyperparameters Parameter | Value ---|--- Discounting factor $\gamma$ | $0.9$ Learning rate $\alpha$ | $1\cdot 10^{-4}$ Rollout length | $10$ Entropy coefficient $\beta$ | $0.01$ GAE $\lambda$ | $1.0$ ### V-B Results The results of our evaluation can be seen in Figure 4. Baseline total system capacity measures the efficiency of the system without D2D communication. For our scenario this was 94.75 Mbps. We found that all the DRL algorithms achieved a similar level of performance, increasing system capacity over the baseline by more than 11%. Our results show that the system capacity continued to increase sublinearly as number of active D2D links demand grew. Conversely, the performance of the random agent shows that without careful resource allocation, the system capacity drops sharply. We found that despite allowing DUE to communicate up to half the power of CUE (20 vs. 23 dBm), they typically converged into operating ranges between 7 and 15 dBm. This resulted in a negligible decrease in the total CUE capacity, 1–2 Mbps or $\approx$1.84% below the baseline system capacity. This decrease was approximately constant across D2D density. ### V-C Discussion Inspecting the actions the DRL algorithms selected, we found that they converged to allocating all DUE onto one or two RBs. This is surprising as we had anticipated the DUE to be evenly distributed amongst all available RBs. Investigating further, we observed that over the course of a training run, the DQN converging from an even RB distribution to the focused allocation strategy. This behaviour developed in the later stages of training and only contributed modest increases to the system capacity. As expected, the optimal strategy for resource allocation was to assign DUE to share RBs with the most geographically distant CUE. When combined with the focused RB allocation described above, this typically resulted in the RRM algorithm choosing to allocate DUE to share with the one or two most isolated CUE. Despite the random UE positioning, the DRL agents were able to learn policies that generalised much better than we anticipated when using fully connected neural networks. We were also surprised how quickly agents adapted during an episode, improving their performance over the course of the ten-step episode. ## VI Conclusion In this research we have presented GymD2D, a network simulator and evaluation platform for RRM in D2D underlay cellular offload. GymD2D makes it easy for researchers to build, benchmark and share RRM algorithms and results. Our toolkit is designed to quickly prototype physical layer resource allocation algorithms, without the complexity of higher layer protocols. GymD2D is configurable and extensible, allowing it to be employed to simulate a range of D2D research needs. We have evaluated GymD2D with several leading DRL algorithms and demonstrated the performance gains of intelligent RRM, increasing system capacity by more than 11%. There was no clear winner amongst the DRL algorithms which performed similarly. The results also demonstrated that D2D cellular offload can significantly minimise its impact on primary networks. In the future we plan to increase the simulation complexity in GymD2D, adding more realistic modelling. Other interesting research challenge include investigating the impacts of CUE power control on cellular offload and supporting D2D relay. We continue to use GymD2D in ongoing research, developing methods for scaling up DRL based D2D RRM. ## References * [1] T. S. Rappaport, W. Roh, and K. Cheun, “Mobile’s millimeter-wave makeover,” _IEEE Spectrum_ , vol. 51, no. 9, pp. 34–58, 2014. * [2] A. Gupta and R. K. Jha, “A survey of 5g network: Architecture and emerging technologies,” _IEEE access_ , vol. 3, pp. 1206–1232, 2015. * [3] B. Kaufman and B. Aazhang, “Cellular networks with an overlaid device to device network,” in _2008 42nd Asilomar conference on signals, systems and computers_. IEEE, 2008, Conference Proceedings, pp. 1537–1541. * [4] A. Asadi, Q. Wang, and V. Mancuso, “A survey on device-to-device communication in cellular networks,” _IEEE Communications Surveys & Tutorials_, vol. 16, no. 4, pp. 1801–1819, 2014. * [5] P. Janis, V. Koivunen, C. Ribeiro, J. Korhonen, K. Doppler, and K. Hugl, “Interference-aware resource allocation for device-to-device radio underlaying cellular networks,” in _VTC Spring 2009-IEEE 69th Vehicular Technology Conference_. IEEE, 2009, Conference Proceedings, pp. 1–5. * [6] C. Chakraborty and J. J. Rodrigues, “A comprehensive review on device-to-device communication paradigm: Trends, challenges and applications,” _Wireless Personal Communications_ , pp. 1–23, 2020. * [7] P. Janis, C.-H. Yu, C. Ribeiro, C. Wijting, K. Hugl, O. Tirkkonen, and V. Koivunen, “Device-to-device communication underlaying cellular communications systems,” _Int’l J. of Communications, Network and System Sciences_ , vol. 2009, 2009. * [8] C.-H. Yu, O. Tirkkonen, K. Doppler, and C. Ribeiro, “Power optimization of device-to-device communication underlaying cellular communication,” in _2009 IEEE international conference on communications_. IEEE, 2009, Conference Proceedings, pp. 1–5. * [9] M. Zulhasnine, C. Huang, and A. Srinivasan, “Efficient resource allocation for device-to-device communication underlaying lte network,” in _2010 IEEE 6th International conference on wireless and mobile computing, networking and communications_. IEEE, 2010, Conference Proceedings, pp. 368–375. * [10] R. Zhang, X. Cheng, L. Yang, and B. Jiao, “Interference-aware graph based resource sharing for device-to-device communications underlaying cellular networks,” in _2013 IEEE wireless communications and networking conference (WCNC)_. IEEE, 2013, Conference Proceedings, pp. 140–145. * [11] C. Xu, L. Song, Z. Han, Q. Zhao, X. Wang, and B. Jiao, “Interference-aware resource allocation for device-to-device communications as an underlay using sequential second price auction,” in _2012 IEEE international conference on communications (ICC)_. IEEE, 2012, Conference Proceedings, pp. 445–449. * [12] L. Su, Y. Ji, P. Wang, and F. Liu, “Resource allocation using particle swarm optimization for d2d communication underlay of cellular networks,” in _2013 IEEE wireless communications and networking conference (WCNC)_. IEEE, 2013, Conference Proceedings, pp. 129–133. * [13] Y. Luo, Z. Shi, X. Zhou, Q. Liu, and Q. Yi, “Dynamic resource allocations based on q-learning for d2d communication in cellular networks,” in _2014 11th International Computer Conference on Wavelet Actiev Media Technology and Information Processing (ICCWAMTIP)_. IEEE, 2014, Conference Proceedings, pp. 385–388. * [14] Z. Li, C. Guo, and Y. Xuan, “A multi-agent deep reinforcement learning based spectrum allocation framework for d2d communications,” _arXiv preprint arXiv:1904.06615_ , 2019. * [15] J. Tan, L. Zhang, and Y.-C. Liang, “Deep reinforcement learning for channel selection and power control in d2d networks,” in _2019 IEEE Global Communications Conference (GLOBECOM)_. IEEE, 2019, Conference Proceedings, pp. 1–6. * [16] G. Brockman, V. Cheung, L. Pettersson, J. Schneider, J. Schulman, J. Tang, and W. Zaremba, “Openai gym,” _arXiv preprint arXiv:1606.01540_ , 2016. * [17] P. Gawłowicz and A. Zubow, “ns3-gym: Extending openai gym for networking research,” _arXiv preprint arXiv:1810.03943_ , 2018. * [18] T. S. Rappaport, _Wireless communications: principles and practice_. prentice hall PTR New Jersey, 1996, vol. 2. * [19] M. Hessel, J. Modayil, H. Van Hasselt, T. Schaul, G. Ostrovski, W. Dabney, D. Horgan, B. Piot, M. Azar, and D. Silver, “Rainbow: Combining improvements in deep reinforcement learning,” _arXiv preprint arXiv:1710.02298_ , 2017\. * [20] P. Christodoulou, “Soft actor-critic for discrete action settings,” _arXiv preprint arXiv:1910.07207_ , 2019. * [21] V. Mnih, A. P. Badia, M. Mirza, A. Graves, T. Lillicrap, T. Harley, D. Silver, and K. Kavukcuoglu, “Asynchronous methods for deep reinforcement learning,” in _International Conference on Machine Learning_ , 2016, Conference Proceedings, pp. 1928–1937. * [22] J. Schulman, P. Moritz, S. Levine, M. Jordan, and P. Abbeel, “High-dimensional continuous control using generalized advantage estimation,” _arXiv preprint arXiv:1506.02438_ , 2015.
# Arbitrary-Oriented Ship Detection through Center-Head Point Extraction Feng Zhang, Xueying Wang, Shilin Zhou, Yingqian Wang, Yi Hou This work was partially supported in part by the National Natural Science Foundation of China (Nos. 61903373, 61401474, 61921001).Feng Zhang, Xueying Wang, Shilin Zhou, Yingqian Wang, Yi Hou are with the College of Electronic Science and Technology, National University of Defense Technology (NUDT), P. R. China. Emails: {zhangfeng01, wangxueying, slzhou, wangyingqian16<EMAIL_ADDRESS>(Corresponding author: Xueying Wang) ###### Abstract Ship detection in remote sensing images plays a crucial role in various applications and has drawn increasing attention in recent years. However, existing arbitrary-oriented ship detection methods are generally developed on a set of predefined rotated anchor boxes. These predefined boxes not only lead to inaccurate angle predictions but also introduce extra hyper-parameters and high computational cost. Moreover, the prior knowledge of ship size has not been fully exploited by existing methods, which hinders the improvement of their detection accuracy. Aiming at solving the above issues, in this paper, we propose a _center-head point extraction based detector_ (named CHPDet) to achieve arbitrary-oriented ship detection in remote sensing images. Our CHPDet formulates arbitrary-oriented ships as rotated boxes with head points which are used to determine the direction. And rotated Gaussian kernel is used to map the annotations into target heatmaps. Keypoint estimation is performed to find the center of ships. Then, the size and head point of the ships are regressed. The orientation-invariant model (OIM) is also used to produce orientation-invariant feature maps. Finally, we use the target size as prior to finetune the results. Moreover, we introduce a new dataset for multi-class arbitrary-oriented ship detection in remote sensing images at a fixed ground sample distance (GSD) which is named FGSD2021. Experimental results on FGSD2021 and two other widely used data sets, i.e., HRSC2016, and UCAS-AOD demonstrate that our CHPDet achieves state-of-the-art performance and can well distinguish between bow and stern. Code and FGSD2021 dataset are available at https://github.com/zf020114/CHPDet. ###### Index Terms: Arbitrary-oriented ship detection, Remote sensing images, Keypoint estimation, Deep convolution neural networks Figure 1: Four different representations of the arbitrary-oriented ship and the disadvantage of the angle regression scheme. (a) Horizontal boxes parameterized by 4 tuples $(x_{min},y_{min},x_{max},y_{max})$. (b) Rotated box with the angle parameterized by 5 tuples $(x_{c},y_{c},w,h,\theta)$. (c) Rotated box with vertices $(a,b,c,d)$, parametrized by 8 tuples $(x_{a},y_{a},x_{b},y_{b},x_{c},y_{c},x_{d},y_{d})$. (d) Rotated box with head point which is parameterized by 6 tuples $(x_{c},y_{c},w,h,x_{h},y_{h})$. (e) A small angle disturbance will cause a large IoU decrease. (f) The angle is discontinous when reaches its range boundary. ## I Introduction Ship detection from high-resolution optical remote sensing images is widely applied in various tasks such as illegal smuggling, port management, and target reconnaissance. Recently, ship detection has received increasing attention and was widely investigated in the past decades [1, 2, 3, 4]. However, ship detection in remote sensing images is a highly challenging task due to the arbitrary orientations, densely-parking scenarios, and complex backgrounds [5, 6, 7]. To handle the multi-orientation issue, existing methods generally use a series of predefined anchors [8], which has the following shortcomings. _Inaccurate angle regression._ Fig. 1(a)-(d) illustrate four different representations of an arbitrary-oriented ship. Since ships in remote sensing images are generally in strips, the intersection over union (IoU) score is very sensitive to the angle of bounding boxes. As shown in Fig. 1(e), the ground truth box is the bounding box of a ship with an aspect ratio of 10:1. The red rotated box is generated by rotating the ground truth box with a small angle of $5^{\circ}$. It can be observed that such a small angle variation reduces the IoU between these two boxes to 0.63. Therefore, the anchor-based detectors which define the positive and negative anchors by IoU score usually suffer from an imbalance issue, and thus resulting in detection performance degeneration [9]. Moreover, the angle of the ship is a periodic function, and it is discontinuous at the boundary ($0^{\circ}$ or $180^{\circ}$), as shown in Fig. 1(f). This discontinuity will also cause performance degeneration. Figure 2: The overall framework of our arbitrary-oriented ship detection method. The dotted lines in the graph represent the same position on the feature maps. Feature maps are first generated by using a fully convolutional backbone network and orientation-invariant model (OIM). Afterward, the peaks of the feature map of center points are selected as center points. Then, the center points offset, object sizes, and head regression locations are regressed on the corresponding feature maps at the position of each center point. The potential head points are collected by extracting peaks with confidence scores larger than $0.1$ on the head feature map. The final head location is obtained by assigning each regressed location to its nearest potential head points and then add the head offset. _Excessive hyper-parameters and high computational cost._ Existing methods generally use oriented bounding boxes as anchors to handle rotated objects and thus introduce excessive hyper-parameters such as box sizes, aspect ratios, and orientation angles. Note that, these hyper-parameters have to be manually tuned for novel scenarios, which limits the generalization capability of these methods. Predefined anchor-based methods usually require a large number of anchor boxes. For example, in R2PN [10], 6 different orientations were used in rotated anchor boxes, and there are a total of 24 anchors at each pixel on its feature maps. A large number of anchor boxes introduce excessive computational cost when calculating IoU scores and executing the non-maximum suppression (NMS) algorithm. _Under-exploitation of prior information of ships._ Most previous ship detectors adopted the commonly-used rotation detection algorithms in the area of remote sensing and scene text detection, while overlook the unique characteristics of ships in remote sensing images. That is, the position of the bow is relatively obvious and a certain category of the ship in remote sensing images has a relatively fixed size range by normalizing the ground sample distance (GSD) of images. The size of the ship and the position of the ship’s head are important clues for detection. However, these prior informations have been under-exploited by previous ship detection algorithms. These methods only model the ships as rotated rectangles to regress the parameters and do not use the obvious bow point to determine the direction of the ship. Due to the limitation of the effective receptive field of the network, the appearance information near the central point is mainly used in target classification. Size regression and target classification are obtained independently by two parallel branches. Therefore, the size of the target can not effectively assist target classification. Motivated by the anchor-free detectors CenterNet [11] in natural scenes, in this paper, we propose a one-stage, anchor-free and NMS-free method for arbitrary-oriented ship detection in remote sensing images. We formulate ships as rotated boxes with a head point representing the direction. Specifically, orientation-invariant feature maps are first produced by an orientation- invariant model. Afterward, the peaks of the center feature map are selected as center points. Then, the offset, object sizes, and head positions are regressed on the corresponding feature maps at each center point. Finally, target size is used to adjust the classification score. The architecture of our CHPDet is shown in Fig. 2. The major contributions of this paper are summarized as follows. * • We develop a one-stage, anchor-free ship detector CHPDet, Specifically, we represent the ships using rotated boxes with a head point. This representation addresses the problem of angle periodicity by transforming the angle regression task into a keypoint estimation task. Moreover, our proposed method can expand the scope of angle to [$0^{\circ}$-$360^{\circ}$), and distinguish between bow and stern. * • We design rotated Gaussian kernel to map the annotations into target heatmaps, which can better adapting to the characteristics of the rotated target. * • We propose a module to refine the detection results based on prior information. Moreover, we proposed a new dataset named FGSD2021 for multi- class arbitrary-oriented ship detection in remote sensing images at fixed GSD. This dataset can facilitate the use of prior knowledge of ship size and promote the actual application for remote sensing ship detection. * • We introduce an orientation-invariant model (OIM) to generate orientation- invariant feature maps. Extensive experimental results on three datasets show that our CHPDet achieves state-of-the-art performance in both speed and accuracy, as shown in Fig. 3. Figure 3: Speed vs. accuracy on our proposed FGSD2021 dataset. The rest of this paper is organized as follows. In Section II, we briefly review the related work. In Section III, we introduce the proposed method in detail. Experimental results and analyses are presented in Section IV. Finally, we conclude this paper in Section V. ## II Related Work In this section, we briefly review the major works in horizontal object detection, rotated object detection, and remote sensing ship detection. ### II-A Horizontal Object Detection In recent years, deep convolutional neural networks (DCNN) have been developed as a powerful tool for feature representation learning [12, 13], and have achieved significant improvements in horizontal object detection [14]. Existing object detection methods generally represent objects as horizontal boxes, as shown in Fig. 1(a). According to different detection paradigms, deep learning-based object detection methods can be roughly divided into two-stage detectors, single-stage detectors, and multi-stage detectors. Two-stage detectors (e.g., RCNN [15], Fast-RCNN [16], Faster-RCNN [17], Mask-RCNN [18], R-FCN [19]) used a pre-processing approach to generate object proposals, and extract features from the generated proposals to predict the category. In contrast, one-stage detectors (e.g., YOLO [20, 21], SSD [22], RetinaNet [23]) do not have the pre-processing step and directly performed categorical prediction on the feature maps. Multi-stage detectors (e,g, cascade RCNN [24], HTC [25]) performed multiple classifications and regressions, resulting in notable accuracy improvements. In summary, two-stage and multi-stage detectors generally achieve better performance, but one-stage detectors are usually more time-efficient. Compared to the above-mentioned anchor-based methods, anchor-free methods [26] [11] can avoid the requirement of anchors and have become a new research focus in recent years. For example, CornerNet [26] detected objects at each position of the feature map using the top-left and bottom-right corner points. CenterNet [11] modeled an object as a center point and performed keypoint estimation to find center points and regressed the object size. FCOS [27] predicted four distances, a center score, and a classification score at each position of the feature map to detect objects. The above-mentioned approaches achieve significant improvement in general object detection tasks. However, these detectors can only generate horizontal bounding boxes, which limits their applicability. ### II-B Arbitrary-oriented object detection Arbitrary-oriented detectors are widely used in remote sensing and scene text images. Most of these detectors used rotated bounding boxes or quadrangles to represent multi-oriented objects, as shown in Fig. 1(b) (c). In RRPN [28], rotated region proposal network was proposed to improve the quality of the region proposals. In R2CNN [29], a horizontal region of interest (RoI) was generated to simultaneously predict the horizontal and rotated boxes. RoI- Trans [30] transformed a horizontal RoI into a rotated RoI (RRoI). In SCRDet [31] and RSDet [9], novel losses were employed to address the boundary problem for oriented bounding boxes. In R3Det [32], a refined single-stage rotated detector was proposed for the feature misalignment problem. In CSL [33] and DCL [34], angle regression was converted into a classification task to handle the boundary problem. In S2A-Net [35], a fully convolutional layer was proposed to align features to achieve better performance. The aforementioned methods need a set of anchor boxes for classification and regression. These anchors introduce excessive hyper-parameters which limit the generalization capability and introduce an excessive computational cost. At present, several anchor-free arbitrary-oriented detectors (e.g., O2D-Net [36] and X-LineNet [37]) are proposed to detect oriented objects by predicting a pair of intersecting lines. However, The features used in these methods are not rotation-invariant and the performance still lags behind that of the anchor- base detectors. ### II-C Ship detection in remote sensing images Different from other objects in remote sensing images, ships are in strips with a large aspect ratio. Generally, the outline of the ships is an approximate pentagon with two parallel long sides, and the position of the bow is relatively obvious. Consequently, a certain category of the ship in remote sensing images has a relatively fixed size range by normalizing the GSD of images. Traditional ship detectors generally used a coarse-to-fine framework with two stages including ship candidate generation and false alarm elimination. For example, Shi et al. [38] first generated ship candidates by considering ships as anomalies and then discriminated these candidates using the AdaBoost approach [39]. Yang et al. [40] proposed a saliency-based method to generate candidate regions, and used a support vector machine (SVM) to further classify these candidates. Liu et al [41, 42] introduced an RRoI pooling layer to extract features of rotated regions. In R2PN [10], a rotated region proposal network was proposed to generate arbitrary-proposals with ship orientation angle information. The above detectors are also based on a set of anchors and cannot fully exploit the prior information of ships. ## III Proposed Method In this section, the architecture of CHPDet is introduced in detail. Our method consists of 5 modules including an arbitrary-oriented ship representation module, a rotated Gaussian kernel module, a head point estimation module, an orientation-invariant module and a probability refinement module. All ships are represented by rotated boxes with a head point. We first detect centers of ships by extracting the peaks in heatmaps which are generated by rotated Gaussian kernels. Then, we locate the head points by two steps (directly regress from image features at the center location, and estimate head points from head heatmaps). We also extract orientation-invariant feature maps by the orientation-invariant model (OIM) to increased consistency between targets and corresponding features. Finally, we refine the detection results based on the prior information. The overall framework of CHPDet is shown in Fig. 2. Figure 4: A schematic diagram of map a rotated bounding box to a rotated Gaussian distribution. Figure 5: A visualization of (a) center heatmap, (b) head heatmap. In center and head heatmaps, different colors represent different categories. ### III-A Arbitrary-oriented ship representation As shown in Fig. 1, the widely-used horizontal bounding boxes cannot be directly applied to the arbitrary-oriented ship detection task since excessive redundant background area is included. Moreover, since the arbitrary-oriented ships generally have a large aspect ratio and park densely, the NMS algorithm using a horizontal bounding box tends to produce missing detection. To this end, many methods represent ships as rotated bounding boxes, and these boxes are parameterized with 5 tuples $(c_{x},c_{y},w,h,\theta)$, where $(x,y)$ is the coordinate of the center of the rotated bounding box, $w$ and $h$ are the width and length of the ship, respectively. The angle $\theta\in[0^{\circ},180^{\circ})$ is the orientation of the long side with respect to the y-axis. This representation can result in the regression inconsistency issue near the boundary case. Recently, some detectors represent objects by four clockwise vertices, which are parameterized by 8 tuples $(x_{a},y_{a},x_{b},y_{b},x_{c},y_{c},x_{d},y_{d})$. This representation can also introduce regression inconsistency due to the order of the four corner points. To avoid the afore-mentioned inconsistency problem, we represent ships as two points and their corresponding size, which are parameterized by 6 tuples $(x_{c},y_{c},w,h,x_{h},y_{h})$. $(x_{c},y_{c})$ is the coordinate of the center of the rotated bounding box, $w$ and $h$ are the width and length of the ship, $(x_{h},y_{h})$ is the coordinate of the head point of the ship. The direction of the ship is determined by connecting the center and the bow. This representation of ships converts discontinuous angle regression to continuous keypoint estimation. This representation also extends the range of angle representation to $[0^{\circ},360^{\circ})$ and enables the network to distinguish between bow and stern. ### III-B Rotated Gaussian Kernel Our detectors uses center heatmaps to classify and locate ships simultaneously. To adapt to the characteristics of the rotated target, we use the rotated Gaussian kernel (see Fig. 4) to map the annotations to target heatmaps in the training stage. Specifically, given $m^{th}$ annotated box $\left(x,y,w,h,\theta\right)$ belongs to $c_{m}^{th}$ category, it is linearly mapped to the feature map scale. Then, 2D Gaussian distribution $\mathcal{N}(\mathbf{m},\mathbf{\Sigma})$ is adopted to produce target heatmap $\textbf{C}\in\mathbb{R}^{\frac{W}{s}\times\frac{H}{s}\times C}$. Here, $m=(x,y)$ represents the probability density function of the rotated Gaussian distribution, and the probability density function can be calculated according to covariance matrix Eq. 1. $\displaystyle\Sigma^{1/2}$ $\displaystyle=\mathbf{RSR}^{\top}$ (1) $\displaystyle=\left(\begin{array}[]{cc}\cos\theta&-\sin\theta\\\ \sin\theta&\cos\theta\end{array}\right)\left(\begin{array}[]{cc}\sigma_{x}&0\\\ 0&\sigma_{y}\end{array}\right)\left(\begin{array}[]{cc}\cos\theta&\sin\theta\\\ -\sin\theta&\cos\theta\end{array}\right)$ $\displaystyle=\left(\begin{array}[]{cc}\sigma_{x}\cos^{2}\theta+\sigma_{y}\sin^{2}\theta&\left(\sigma_{x}-\sigma_{y}\right)\cos\theta\sin\theta\\\ \left(\sigma_{x}-\sigma_{y}\right)\cos\theta\sin\theta&\sigma_{x}\cos^{2}\theta+\sigma_{y}\sin^{2}\theta\end{array}\right),$ where $s$ is a downsampling stride and $\sigma_{x}=\alpha\frac{\sigma_{p}\times w}{\sqrt{w\times h}}$, $\sigma_{y}=\alpha\frac{\sigma_{p}\times h}{\sqrt{w\times h}}$, $\sigma_{p}$ is a size-adaptive standard deviation [11]. $\alpha$ is set to 1.2 in our implementation, and it’s not carefully selected. Fig. 4 is a schematic diagram of mapping a rotated bounding box to a rotated Gaussian distribution. If two Gaussian kernels belong to the same category with an overlap region, we take the maximum value at each pixel of the feature map. $\hat{\textbf{C}}\in\mathbb{R}^{\frac{W}{s}\times\frac{H}{s}\times C}$ is a prediction on feature maps produced by the backbones. Fig. 5(a) shows a visualization of the center heatmaps. We extract locations with values larger or equal to their 8-connected neighbors as detected center points. The value of the peak point is set as a confidence measurement, and the coordinates in the feature map are used as an index to get other attributes. Therefore, the accurate location of the center point on the feature map is the key part of the whole detection. The peaks of the Gaussian kernel, also the centers of rotated box, are treated as the positive samples while any other pixels are treated as the negative samples, which may cause a huge imbalance between positive and negative samples. To handle the imbalance issue, we use the variant focal loss as [23, 11]: $\mathcal{L}_{c}=\frac{-1}{N}\left\\{\begin{array}[]{cl}\sum_{xyc}\left(1-\hat{\textbf{C}}_{xyc}\right)^{\gamma}\log\left(\hat{\textbf{C}}_{xyc}\right)&\text{ if }\textbf{C}(xyc)=1\\\ \sum_{xyc}\left(1-\textbf{C}_{xyc}\right)^{\beta}\left(\hat{\textbf{C}}_{xyc}\right)^{\gamma}\\\ \log\left(1-\hat{\textbf{C}}_{xyc}\right)&\text{ otherwise }\end{array}\right.$ (2) where $\gamma$ and $\beta$ are the hyper-parameters of the focal loss, $N$ is the number of objects in image $I$ which is used to normalize all positive focal loss instances to $1$. We set $\gamma=2$ and $\beta=4$ in our experiments empirically as in [26]. To reduce the quantization error caused by the output stride, we produce local offset feature maps $\textbf{O}\in\mathbb{R}^{\frac{W}{S}\times\frac{H}{S}\times 2}$. Suppose that ${c}=\left\\{\left(\hat{x}_{k},\hat{y}_{k}\right)\right\\}_{k=1}^{n}$ is the set of detected center points, center point location is given by an integer coordinates $c_{k}=(\hat{x_{i}},\hat{y_{i}})$ on feature map C. For each predicted center point $c_{k}$, let the value on the offset feature maps $f_{k}=(\delta\hat{x}_{k},\delta\hat{y}_{k})$ be the offset of center point $c_{k}$. The final center point location of class $c$ is $\hat{center_{c}}=\left\\{\left(\hat{x_{k}}+\delta\hat{x}_{k},\hat{y_{k}}+\delta\hat{y}_{k}\right)\right\\}_{k=1}^{n}$. Note that, all classes share the same offset predictions to reduce the computational complexity. The offset is optimized with an L1 loss. This supervision is performed on all center point. $\mathcal{L}_{\text{co}}=\frac{1}{N}\sum_{k=1}^{N}\left|\textbf{O}{{c_{k}}}-\left(\frac{\rm{center}_{k}}{S}-c_{k}\right)\right|.$ (3) The regression of the size of objects is similar to that of local offset. ### III-C Head Point estimation We perform two steps for better head points estimation. #### III-C1 Regression-based head point estimation Let $\rm{head}_{k}$$=(h_{x},h_{y})$ be the $k^{th}$ head point,we directly regress to the offsets $(\varDelta\hat{x}_{k},\varDelta\hat{y}_{k})$ on feature map $\textbf{R}\in\mathbb{R}^{\frac{W}{S}\times\frac{H}{S}\times 2}$ at each predicted center point $c_{k}\in\hat{center}$. The regression-based head point is $\left\\{\left(\hat{x_{k}}+\varDelta\hat{x}_{k},\hat{y_{k}}+\varDelta\hat{y}_{k}\right)\right\\}_{k=1}^{n}$, where $\left(\varDelta\hat{x}_{i},\varDelta\hat{y}_{i}\right)$ is the head point regression, and an L1 loss is used to optimized head regression feature maps. $\mathcal{L}_{hr}=\frac{1}{N}\sum_{k=1}^{N}\left|\textbf{R}_{c_{k}}-h_{k}\right|.$ (4) #### III-C2 Bottom-up head point estimation We use standard bottom-up multi-human pose estimation [43] to refine the head points. A target map $\textbf{H}\in\mathbb{R}^{\frac{W}{s}\times\frac{H}{s}\times 1}$ is computed as described in Section III-B. A low-resolution equation is $\rm\tilde{head}=\left\lfloor\frac{head}{s}\right\rfloor$. Head point heatmap $\textbf{E}\in\mathbb{R}^{\frac{W}{S}\times\frac{H}{S}\times 1}$ and local offset heatmap $\textbf{HO}\in\mathbb{R}^{\frac{W}{S}\times\frac{H}{S}\times 2}$ are head maps produced by the backbones. These two head maps are trained with variant focal loss and an L1 loss. $\mathcal{L}_{he}=\frac{-1}{N}\sum_{xy}\left\\{\begin{array}[]{cl}\left(1-\textbf{E}_{xy}\right)^{\gamma}\log\left(\textbf{E}_{xy}\right)&\text{ if }\textbf{H}_{xy}=1\\\ \left(1-\textbf{H}_{xy}\right)^{\beta}\left(\textbf{E}_{xy}\right)^{\gamma}\\\ \log\left(1-\textbf{E}_{xy}\right)&\text{ otherwise }\end{array}\right.$ (5) $\mathcal{L}_{ho}=\frac{1}{N}\sum_{k=1}^{N}\left|\textbf{HO}_{c_{k}}-\left(\frac{\rm{head_{k}}}{S}-\tilde{head}\right)\right|.$ (6) The bottom-up head point estimation is the same as the center point detection. Note that, in center point detection, each category has a center points heat map, while in head points estimation, all categories share one head points heatmap. We extract all peak point locations $\hat{\rm{head}}=\left\\{\tilde{l}_{i}\right\\}_{i=1}$ with a confidence $\textbf{HO}_{x,y}>0.1$ as a potential head points set, and refine the potential head point locations by adding the offset ${(\xi_{x},\xi_{y})}$. Fig. 5(b) visualizes the head points heatmap. We introduce a set of weighted factors to balance the contribution of these parts, and set $\lambda_{o}=1$, $\lambda_{s}=0.1$, $\lambda_{\rm{hr}}=1$, $\lambda_{\rm{he}}=1$, and $\lambda_{\rm{ho}}=1$ in all our experiments. We set $\lambda_{s}=0.1$ since the scale of the loss is ranged from $0$ to the output size $h/S$. The overall training loss is $\displaystyle\mathcal{L}=$ $\displaystyle\mathcal{L}_{c}+\lambda_{o}\mathcal{L}_{o}+\lambda_{s}\mathcal{L}_{s}+\lambda_{\rm{hr}}\mathcal{L}_{\rm{hr}}+\lambda_{\rm{he}}\mathcal{L}_{\rm{he}}+\lambda_{\rm{ho}}\mathcal{L}_{\rm{ho}}.$ (7) In the testing phase, we first extracted the center points on the output center heatmaps C for each category. We used a $3\times 3$ max-pooling layer to get the peak points and selected the top 100 peaks as potential center points. Each center point location is represented as an integer coordinates $\hat{c}=(\hat{x},\hat{y})$. Take out the offsets $(\delta_{x},\delta_{y})$, size $(w,h)$, and head points regression $\left(\varDelta_{x},\varDelta_{y}\right)$ on the corresponding feature map at the location of center points. We also picked all head peak points on the output center heatmaps E with a scores $\hat{\rm{head}}\in(x,y),if~{}\textbf{E}_{x,y}>0.1$, and then assigned each regressed location ${\rm{head}_{r}=\left(\hat{x}+\varDelta{x},\hat{y}+\varDelta{y}\right)}$ to its closest detected keypoint $\arg\min_{l\in\rm{head}_{r}}\left(l-\hat{\rm{head}}\right)^{2}$ as the head point $(\hat{h_{x}},\hat{h_{y}})$, then we add the head point offset $(\xi_{x},\xi_{y})$ to refine the head point estimation. Finally, we get the rotated boxes ${(\hat{x}+\delta_{x},\hat{y}+\delta_{y},w,h,\hat{h_{x}}+\xi_{x},\hat{h_{y}}+\xi_{y})}$. We use the line connecting the center point and the head point to determine the orientation of targets. Figure 6: A visualization of ship probability density map. In the ship probability density map, $l_{a}$ represents the mean length of category $a$, $l$ represents the length of the detected ship. The red area is the probability that the target belongs to category $a$. ### III-D Orientation-Invariant Model Let $\textbf{I}\in\mathbb{R}^{W\times H\times 3}$ be an input image with width $W$ and height $H$, the feature map generated from backbone is $\textbf{F}\in\mathbb{R}^{\frac{W}{s}\times\frac{H}{s}\times K}$, where $S$ is the output stride, $C$ is the output feature channels. In this paper, we set the default stride value to $S=4$ and feature channels to $K=64$. The feature generated from these backbones is not rotation-invariant [44], while ships in remote sensing images are distributed with arbitrary orientations. To alleviate the inconsistency, we introduce an orientation- invariant model (OIM) which consists of two modules: active rotating filters (ARF) and oriented response pooling (ORPooling) [44]. We first use active ARF to explicitly encode the orientation information. An ARF is a $k\times k\times N$ filter that actively rotates $N-1$ times during convolution to produce a feature map with $N$ orientation channels. For a feature map M and an ARF $\mathcal{F}$, the $i^{th}$ filter $\mathbf{I}^{(i)}$, $i\in[1,N-1]$, is obtained by clockwise rotating $\mathcal{F}$ by $\frac{2\pi n}{N}$(N is set to 8 by default) , and can be computed as $\mathbf{I}^{(i)}=\sum_{n=0}^{N-1}\mathcal{F}_{\theta_{i}}^{(n)}\cdot\mathbf{M}^{(n)},\theta_{i}=i\frac{2\pi}{N},i=0,\ldots,N-1$ (8) where $\mathcal{F}_{\theta_{i}}$ is the clockwise $\theta_{i}$-rotated version of $\mathcal{F}$, $\mathcal{F}_{i}^{(n)}$ and $\textbf{M}^{(n)}$ are the $n^{th}$ orientation channel of $\mathcal{F}_{i}$ and M respectively. The ARF captures image response in $N$ directions and explicitly encodes its location and orientation into a single feature map with $N$ orientation channels. To reduce computational complexity, we use the combination of small $3\times 3$ filters and an $8$ orientation channels in our experiments. Feature maps captured by ARF are not rotation-invariant as orientation information are encoded instead of being discarded. Then ORPooling is used to extract orientation-invariant feature. It is simply achieved by choosing the orientation channel with the strongest response as the output feature $\textbf{I}\in\mathbb{R}^{\frac{W}{s}\times\frac{H}{s}\times K}$. That is, $\hat{\mathbf{I}}=\max\\{\mathbf{I}^{(n)}\\},0<n<N-1.$ (9) Since ORPooling is introduced to extract the maximum response value for all ARF, the target features of different orientations at this location are identical. Based on the rotation invariance feature, six kinds of feature maps are got by convolution layers respectively. Moreover, OIM only introduces one convolution layer with a small number of parameters, which has little effect on the speed of training and inferencing. The rotation-invariant feature is very important for detecting arbitrary oriented objects, which enhances the consistency of the feature. Our detectors extract locations with local maximum as detected center points, so at the object center, the rotation-invariant feature of arbitrary oriented objects are identical, which increases the generalization ability of the network. Otherwise, more parameters are needed to encode the orientation information. ### III-E Refine probability according to size By normalizing the GSD of remote sensing images, objects of the same size on the ground have the same size in all images. The size of the target is an important clue to identify the target because a certain type of target in remote sensing images usually has a relatively fixed size range. We propose an approach to adjust the confidence score of targets according to the prior knowledge of ship size. As shown in Fig. 5(d), suppose that the category of the detected box is $a$, the original confidence score is $s_{a}$, assume that the length of the detected ship obeys a normal distribution, the mean and standard deviation of the length of category $a$ are $L_{a}$, $\delta_{a}$. Then the probability of the target belonging to $a$ is $p_{a}$, i.e. $p_{a}=\frac{2}{\delta_{a}\sqrt{2\pi}}\int_{-\infty}^{-|l-la|}\exp\left(-\frac{(x-la)^{2}}{2\delta_{a}^{2}}\right)dx.$ (10) To reduce hyper-parameters, we assume that the standard deviation is proportional to the mean $\delta_{a}=L_{a}\times\lambda$ for all categories of ships. We multiply the two probabilities to obtain the final detection confidence, $\hat{p_{a}}=p_{a}\times s_{a}$. ## IV Experiments We evaluate our method on our FGSD2021 dataset, the public HRSC2016 [45] and UCAS-AOD [46] dataset. In this section, we first introduce the datasets and implementation details, then perform ablation studies and compare our network to several state-of-the-art methods. Figure 7: Example images from the proposed FGSD2021 dataset. 20 categories are chosen and annotated in our dataset, including Aircraft carriers, Wasp-class, Tarawa-class, Austin-class, Whidbey-island-class, San-Antonio-class, Newport- class, Ticonderoga-class, Arleigh-Burke-class, Perry-class, Lewis and Clark- class, Supply-class, Henry J. Kaiser-class, Bob Hope-Class, Mercy-class, Freedom-class, Independence-class, Avenger-class, submarine, and others. ### IV-A Datasets #### IV-A1 HRSC2016 The HRSC2016 dataset [45] is a challenging dataset for ship detection in remote sensing images, which collected six famous harbors on Google Earth. The training, validation, and test sets include 436 images with 1207 samples, 181 with 541 samples, and 444 images with 1228 samples, respectively. The image size of this dataset ranges from $300\times 300$ to $1500\times 900$. This dataset includes three levels of tasks (i.e., L1, L2, and L3), and these three tasks contain 1 class, 4 classes, and 19 classes, respectively. Besides, the head point of ships is given in this dataset. Following [28] [35] [32], we evaluate our method on task L1. We used the training and validation set in the training phase and evaluated the detection performance on the test set. #### IV-A2 FGSD2021 Existing ship datasets HRSC2016 have the following shortcomings. First, the GSD is unknown, so we cannot get the size of objects in the image by the actual size on the ground. Second, the size of the image is very small which is inconsistent with the actual remote sensing image detection task. To solve these problems, we propose a new ship detection dataset FGSD2021 which has a fixed GSD. Our dataset is developed by collecting high-resolution satellite images from publicly available Google Earth, which covers some famous ports such as Dandiego, Kitsap-Bremerton, Norfolk, Pearl Harbor, and Yokosuka. We usually obtain multiple images of the same port on different days, and there are also some images from the HRSC2016 dataset. We collected 636 images with a normalized GSD, 1 meter per pixel. The images in our dataset are very large, usually, one image covers a whole port. The width of images is ranged from 157 to 7789 pixels, and the average width is 1202 pixels, the height is ranged from 224 to 6506 pixels, and the average height is 1205 pixels. Our FGSD2021 dataset is divided into 424 training images and 212 test images. The training set is used in the training phase. The detection performance of the proposed method is evaluated on the test set. FGSD2021 including 5274 labeled targets and 20 categories are chosen and annotated. We use the labelimg2111https://github.com/chinakook/labelImg2 tools to label the ship, the angle range is $[0^{\circ},360^{\circ})$, and the main direction is the direction of the bow. Some examples of annotated patches are shown in Fig. 7. #### IV-A3 UCAS-AOD The UCAS-AOD dataset [46] contains 1510 aerial images of about $659\times 1280$ pixels and 14596 instances of two categories including plane and car. The angle range of target in this dataset is $[0^{\circ},180^{\circ})$, so we manually marked the direction of the head. We randomly sampled 1132 images for training and 378 images for testing. All images were cropped into patches of size $672\times 672$. ### IV-B Implementation Details Our network was implemented in PyTorch on a PC with Intel Core i7-8700K CPU, NVIDIA RTX 2080Ti GPU. We used the Adam method [47] as the optimizer, and the initial learning rate was set to $2.5\times 10^{-4}$. We trained our network for 140 epochs with a learning rate being dropped at 90 epochs. During the training phase, we used random rotation, random flipping, and color jittering for data augmentation. To maintain the GSD of the image, we cropped all images into $1024\times 1024$ slices with a stride of 820, resized them to $512\times 512$. We merged the detection results of all the slices to restore the detecting results on the original image. Finally, we applied rotated-non- maximum-suppression (RNMS) with an IoU threshold of 0.15 to discard repetitive detections. The speed of the proposed network was measured on a single NVIDIA RTX 2080Ti GPU. Several different backbones (e.g., deep layer aggregation (DLA) [48] and hourglass network (Hourglass) [49]) can be used to extract features from images. We followed CenterNet [11] to enhance DLA by replacing ordinary convolutions with deformable convolutions and add a 256 channel $3\times 3$ convolutional layer before the output head. The hourglass network consists of two sequential hourglass modules. Each hourglass module includes 5 pairs of down and up convolutional networks with skip connections. This network generally yields better keypoint estimation performance [26]. ### IV-C Evaluation Metrics The IoU between oriented boxes is used to distinguish detection results. The mean average precision (mAP) and head direction accuracy are used to evaluate the performance of arbitrary-Oriented detectors. #### IV-C1 IoU The IoU is the result of dividing the overlapping area by the union area of two boxes. We adopted the evaluation approach in DOTA [50] to get the IoU. If the IoU between a detection box and a ground-truth is higher than a threshold, the detection box is marked as true-positive (TP), otherwise false-positive (FP). If a ground-truth box has no matching detections, it is marked as false negative (FN). #### IV-C2 mAP The precision and recall are calculate by $\text{precision }=\frac{\text{ TP }}{\mathrm{TP}+\mathrm{FP}}$, $\text{recall}=\frac{\text{TP}}{\mathrm{TP}+\mathrm{FN}}$. We first set a set of thresholds, and then we get a corresponding maximum precision for each recall threshold. AP is the average of these precisions. The mean average precision (mAP) is the mean of APs over all classes. The mAP0.5-mAP0.8 is computed under the IoU threshold of 0.5-0.8 respectively. PASCAL VOC2007 metric is used to compute the mAP in all of our experiments. #### IV-C3 Head direction accuracy The prediction angle range of the previous algorithm is 0∘-180∘, which cannot distinguish between the bow and stern of the ship. The mAP base on the IoU between two rotated boxes is taken as the only evaluation criterion, which cannot reflect the accuracy of the bow direction. To solve this problem, we define bow direction accuracy as an additional evaluation. That is the proportion of the ships whose angle difference from the ground-truth less than 10 degrees in all TPs. ### IV-D Ablation Study In this subsection, we present ablation experiments to investigate our models. #### IV-D1 CenterNet as baseline As an anchor-free detector, CenterNet performs keypoint estimation to find the center point and regresses the object size at each center point position. To carry out arbitrary-oriented ship detection, we add an extra branch to predict the angle as a baseline which is named CenterNet-Rbb. CenterNet-Rbb uses a DLA34 as the backbone, and presents ships as rotated boxes with angle, and uses the L1 loss function to optimized angle regression feature maps. We set weighted factor $\lambda_{angle}=0.1$ to balance the contribution of these parts since the scale of the loss is ranged from $0$ to $180$. As shown in Table I, CenterNet-Rbb achieves an mAP of 70.52% which demonstrates that our baseline achieves competitive performance. TABLE I: Results achieved on FGSD2021 with different ablation versions. ‘Baseline’ represents adding a branch to predict the angle based on CenterNet. ‘Head Point’ represents replacing the angle prediction branch to head point estimation module. ‘Rotate kernel’ represents generating center heatmap by rotated kernel in training. ‘OIM’ represents add orientation-invariant model behand the backbone. ‘Extra convolution’ represents replacing the OIM with two extra convolution layers. ‘Refine probability’ represents using the prior size information to adjust the confidence score of the detected boxes. | baseline | Different Settings of CHPDet | ---|---|---|--- Head Point | | $\checkmark$ | $\checkmark$ | $\checkmark$ | $\checkmark$ | $\checkmark$ Rotate kernel | | | $\checkmark$ | $\checkmark$ | $\checkmark$ | $\checkmark$ OIM | | | | $\checkmark$ | | $\checkmark$ Extra convolution | | | | | $\checkmark$ | Refine Probability | | | | | | $\checkmark$ mAP | 70.52 | 82.96 | 83.56 | 86.61 | 82.66 | 87.91 | | | | | | TABLE II: Performance of CHEDet achieved on FGSD2021 with different variance coefficient $\lambda$. ‘without refine’ represents using the original confidence without refinement. ‘Ground truth class’ represents using ground truth class label to eliminate the misclassification. Backbone | Image Size | coefficient $\lambda$ | without refine | Ground truth class ---|---|---|---|--- 0.1 | 0.2 | 0.3 | 0.4 | 0.5 | 0.6 | 0.7 | 0.8 DLA34 | $512\times 512$ | 87.40 | 87.91 | 87.39 | 87.45 | 87.17 | 87.20 | 87.15 | 87.10 | 86.61 | 89.33 DLA34 | $1024\times 1024$ | 86.37 | 87.84 | 89.28 | 88.17 | 88.68 | 88.85 | 88.47 | 88.50 | 88.39 | 89.74 | | | | | | | | | | | TABLE III: Detection accuracy on different types of ships and overall performance with the state-of-the-art methods on FGSD. The short names for categories are defined as (abbreviation-full name): Air - Aircraft carriers, Was - Wasp class, Tar - Tarawa class, Aus - Austin class, Whi - Whidbey Island class, San -San Antonio class, New - Newport class, Tic - Ticonderoga class, Bur- Arleigh Burke class, Per - Perry class, Lew -Lewis and Clark class, Sup - Supply class, Kai - Henry J. Kaiser class, Hop - Bob Hope Class, Mer - Mercy class, Fre - Freedom class, Ind - Independence class, Ave - Avenger class, Sub - Submarine and Oth - Other. CHPDet† means CHPDet trained and detected with $1024\times 1024$ image size. Method | Air | Was | Tar | Aus | Whi | San | New | Tic | Bur | Per | Lew | Sup | Kai | Hop | Mer | Fre | Ind | Ave | Sub | Oth | mAP(07) ---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|--- R2CNN [29] | 89.9 | 80.9 | 80.5 | 79.4 | 87.0 | 87.8 | 44.2 | 89.0 | 89.6 | 79.5 | 80.4 | 47.7 | 81.5 | 87.4 | 100 | 82.4 | 100 | 66.4 | 50.9 | 57.2 | 78.09 Retinanet-Rbb [23] | 89.7 | 89.2 | 78.2 | 87.3 | 77.0 | 86.9 | 62.7 | 81.5 | 83.3 | 70.6 | 46.8 | 69.9 | 80.2 | 83.1 | 100 | 80.6 | 89.7 | 61.5 | 42.5 | 9.1 | 73.49 ROI-Trans[30] | 90.9 | 88.6 | 87.2 | 89.5 | 78.5 | 88.8 | 81.8 | 89.6 | 89.8 | 90.4 | 71.7 | 74.7 | 73.7 | 81.6 | 78.6 | 100 | 75.6 | 78.4 | 68.0 | 66.9 | 83.48 SCRDet [31] | 77.3 | 90.4 | 87.4 | 89.8 | 78.8 | 90.9 | 54.5 | 88.3 | 89.6 | 74.9 | 68.4 | 59.2 | 90.4 | 77.2 | 81.8 | 73.9 | 100 | 43.9 | 43.8 | 57.1 | 75.90 CSL [33] | 89.7 | 81.3 | 77.2 | 80.2 | 71.4 | 77.2 | 52.7 | 87.7 | 87.7 | 74.2 | 57.1 | 97.2 | 77.6 | 80.5 | 100 | 72.7 | 100 | 32.6 | 37.0 | 40.7 | 73.73 DCL [34] | 89.9 | 81.4 | 78.6 | 80.7 | 78.0 | 87.9 | 49.8 | 78.7 | 87.2 | 76.1 | 60.6 | 76.9 | 90.4 | 80.0 | 78.8 | 77.9 | 100 | 37.1 | 31.2 | 45.6 | 73.34 R3Det[32] | 90.9 | 80.9 | 81.5 | 90.1 | 79.3 | 87.5 | 29.5 | 77.4 | 89.4 | 69.7 | 59.9 | 67.3 | 80.7 | 76.8 | 72.7 | 83.3 | 90.9 | 38.4 | 23.1 | 40.0 | 70.47 RSDet[9] | 89.8 | 80.4 | 75.8 | 77.3 | 78.6 | 88.8 | 26.1 | 84.7 | 87.6 | 75.2 | 55.1 | 74.4 | 89.7 | 89.3 | 100 | 86.4 | 100 | 27.6 | 37.6 | 50.6 | 73.74 S2A-Net[35] | 90.9 | 81.4 | 73.3 | 89.1 | 80.9 | 89.9 | 81.2 | 89.2 | 90.7 | 88.9 | 60.5 | 75.9 | 81.6 | 89.2 | 100 | 68.6 | 90.9 | 61.3 | 55.7 | 64.7 | 80.19 ReDet[51] | 90.9 | 90.6 | 80.3 | 81.5 | 89.3 | 88.4 | 81.8 | 88.8 | 90.3 | 90.5 | 78.1 | 76.0 | 90.7 | 87.0 | 98.2 | 84.4 | 90.9 | 74.6 | 85.3 | 71.2 | 85.44 Oriented R-CNN[52] | 90.9 | 89.7 | 81.5 | 81.1 | 79.6 | 88.2 | 98.9 | 89.8 | 90.6 | 87.8 | 60.4 | 73.9 | 81.8 | 86.7 | 100 | 60.0 | 100 | 79.4 | 66.9 | 63.7 | 82.54 BBAVectors[53] | 99.5 | 90.9 | 75.9 | 94.3 | 90.9 | 52.9 | 88.5 | 90.0 | 80.4 | 72.2 | 76.9 | 88.2 | 99.6 | 100 | 94.0 | 100 | 74.5 | 58.9 | 63.1 | 81.1 | 83.59 DARDet[54] | 90.9 | 89.2 | 69.7 | 89.6 | 88.0 | 81.4 | 90.3 | 89.5 | 90.5 | 79.7 | 62.5 | 87.9 | 90.2 | 89.2 | 100 | 68.9 | 81.8 | 66.3 | 44.3 | 56.2 | 80.31 CenterNet-Rbb[11] | 67.2 | 77.9 | 79.2 | 75.5 | 66.8 | 79.8 | 76.8 | 83.1 | 89.0 | 77.7 | 54.5 | 72.6 | 77.4 | 100 | 100 | 60.8 | 74.8 | 46.5 | 44.1 | 6.8 | 70.52 CHPDet-DLA34 | 90.9 | 90.4 | 89.6 | 89.3 | 89.6 | 99.1 | 99.4 | 90.2 | 90.2 | 90.3 | 70.7 | 87.9 | 89.2 | 96.5 | 100 | 85.1 | 100 | 84.4 | 68.5 | 56.9 | 87.91 CHPDet-DLA34† | 90.9 | 90.2 | 90.9 | 90.3 | 89.3 | 89.2 | 98.9 | 90.2 | 90.2 | 90.2 | 72.2 | 96.5 | 90.7 | 95.3 | 100 | 95.2 | 90.9 | 86.4 | 85.9 | 62.4 | 89.29 | | | | | | | | | | | | | | | | | | | | | #### IV-D2 Effectiveness of the head point estimation When we replace the angle prediction branch with the head point estimation module, the overall mAP is improved from 70.52% to 82.96%. It is a significant improvement, which fully demonstrates the effectiveness of the head point estimation approach. This improvement mainly comes from two aspects. First, the algorithm makes full use of the prior knowledge of the bow point and improves the accuracy of angle regression. Second, since multi-task learning is performed, bow detection increases the supervision information and improves the accuracy of other tasks. To further verify the promoting effect of head point estimation for center point detection and size detection, we set all angles of ground-truth and the detected box to 0∘. Compared with the CenterNet-Rbb, The mAP of CHPDet has risen from 84.4% to 88.0%. This shows that the head point estimation is equivalent to multi-task joint training. It gives more supervision to the network and improves the performance of the network. Besides, the head point estimation only introduces 3 additional channels feature maps and 0.7 ms speed latency. #### IV-D3 Effectiveness of the rotated Gaussian kernel Our detector uses the rotated Gaussian kernel to map the annotations to target heatmaps and achieves an improvement of 0.6% in terms of nomal Gaussian kernels. This implies that rotated Gaussian kernel is a better representation for OBB in the aerial images. The rotated Gaussian kernel can adjust its shape and direction according to the shape of the target and reduce the influence of positioning error on the detection results. As shown in Fig. 4, the rotated Gaussian kernel has the maximum error in the long axis direction, so in the detection process, the center point has a large error on the long axis. Because the error of the center point in the long axis has the least influence on the IoU, the rotated Gaussian kernel can reduce the influence of positioning error on the detection results, and vice versa. Note that, rotated Gaussian kernel does not introduce any additional parameters, and they do not increase training and inferencing time. Consequently, it is a completely cost-free module. #### IV-D4 Effectiveness of the orientation-invariant model We add an orientation-invariant model (OIM) at the end of the backbone and keep other settings unchanged to validate its effectiveness. As shown in Table I, compared with the standard backbone, the backbone with the orientation- invariant model improves mAP by about 3 percentages to 86.61%, while only introduces 2.6 ms speed latency. To further verify the effectiveness of the OIM structure, we replace the OIM with two convolution layers. Compared with the standard backbone, the backbone with two extra convolution layers model drops the performance to 82.66%. It is proved that the performance improvement does not come from the improvement of the number of parameters. We argue that the standard backbones are not rotation-invariant, and the corresponding features are rotation-sensitive. Consequently, OIM increases the consistency between targets and corresponding features. It not only improves the accuracy of angle prediction, but also improves the accuracy of center point detection and size regression. #### IV-D5 Effectiveness of the Refine probability model In the FGSD2021 dataset, the actual length of each category is determined. For example, the length of the Ticonderoga-class cruiser is 172.8 meters. In our designed network, the prior knowledge of ship length is used to refine the confidence of the detected ships belonging to a certain category. Table I shows the mAP values of different ablation versions on the test set. It can be observed that the baseline model achieves the lowest mAP. When the prior size information is incorporated, the performance has been improved. The accuracy improvement on low-resolution images is more obvious, e.g., from 86.61% to 87.91%, an increase of 1.3% in mAP. It demonstrates that the prior size information can improve classification accuracy. We set a variance coefficient to adjust the influence of size on probability. Consequently, we use the length of this type of ship $l_{a}$ multiplied by a coefficient $r$ as the mean square error of this type $\delta_{a}$, $\delta_{a}=l_{a}\times r$. The variance coefficient will affect classification accuracy. When the coefficient is large, the probability difference between different categories will be smaller, and the influence of the size on the confidence of the category will be smaller, and vice versa. As can be observed in Table II, when the coefficient is small, it is equivalent to use size as the main information to classify objects. Accuracy increases gradually as the coefficient increases, and when the coefficient is larger than 0.2, the coefficient has little impact on the accuracy. When we treat all categories as one category and remove the category influence on the detection results, the mAP is $89.33$%, and $89.74$%, respectively. At the same time, by incorporating prior information to adjust the classification confidence, the detection accuracy under 20 categories with an input image of size 1024x1024 achieved an mAP of $89.28$% which shows that after incorporating the prior information, almost all categories are classified correctly. TABLE IV: Detection performance on the FGSD2021 at different IoU thresholds and the accuracy of bow direction. BDA presents bow direction accuracy Method | Backbone | Image Size | mAP0.5 | mAP0.6 | mAP0.7 | mAP0.8 | BDA | FPS ---|---|---|---|---|---|---|---|--- R2CNN[29] | Resnet50 | $512\times 512$ | 78.09 | 75.03 | 64.83 | 36.41 | _ | 10.3 Retinanet-Rbb[23] | Resnet50 | $512\times 512$ | 73.49 | 69.17 | 62.82 | 45.00 | _ | 35.6 RoI-Trans[30] | Resnet50 | $512\times 512$ | 83.48 | 82.63 | 80.35 | 65.18 | _ | 19.2 SCRDet[31] | Resnet50 | $512\times 512$ | 75.90 | 70.98 | 61.82 | 35.12 | _ | 9.2 CSL[33] | Resnet50 | $512\times 512$ | 73.73 | 69.71 | 60.25 | 34.93 | _ | 10.4 DCL[34] | Resnet50 | $512\times 512$ | 73.34 | 69.19 | 57.80 | 28.54 | _ | 10.0 R3Det[32] | Resnet50 | $512\times 512$ | 70.47 | 68.32 | 57.17 | 27.44 | _ | 14.0 RSDet[9] | Resnet50 | $512\times 512$ | 73.74 | 69.55 | 61.52 | 35.83 | _ | 15.4 S2A-Net[35] | Resnet50 | $512\times 512$ | 80.19 | 79.58 | 75.65 | 58.82 | _ | 33.1 ReDet[51] | ReResnet50 | $512\times 512$ | 85.44 | 84.65 | 80.24 | 67.94 | _ | 13.8 Oriented R-CNN[52] | Resnet50 | $512\times 512$ | 82.54 | 81.32 | 78.53 | 64.87 | _ | 27.4 BBAVectors[53] | Resnet50 | $512\times 512$ | 83.59 | 82.74 | 78.55 | 62.48 | _ | 18.5 DARDet[54] | Resnet50 | $512\times 512$ | 80.31 | 79.62 | 74.77 | 59.21 | _ | 31.9 CenterNet-Rbb[11] | DLA34 | $512\times 512$ | 70.52 | 69.34 | 65.52 | 45.33 | _ | 48.5 CHPDet(ours) | DLA34 | $512\times 512$ | 87.91 | 87.15 | 83.69 | 71.24 | 97.84 | 41.7 CHPDet(ours) | DLA34 | $1024\times 1024$ | 89.29 | 88.98 | 86.57 | 73.56 | 98.39 | 15.4 | | | | | | | | Figure 8: Comparison of the detection results in FGSD2021 with different methods. The first column is the ground truth, and the second to the last columns are the results of Retinanet-Rbb [23], ROI-Trans [30], SCRDet [31] , S2A-Net [35], and CHPDet (ours), respectively. Different color of rotated boxes represents a different type of ships. The pink point represents the head point. TABLE V: Detection accuracy on the HRSC2016 dataset, 07 means using the 2007 evaluation metric. Method | Backbone | mAP(07) ---|---|--- R2CNN [29] | Resnet101 | 73.07 RRPN [28] | Resnet101 | 79.08 R2PN[10] | VGG16 | 79.6 ROI-trans[30] | Resnet101 | 86.20 Gliding Vertex[55] | Resnet101 | 88.20 BBAVectors[53] | Resnet101 | 88.6 R3Det [32] | Resnet101 | 89.26 FPN-CSL[33] | Resnet101 | 89.62 R3Det-DCL[34] | Resnet101 | 89.46 DAL[56] | Resnet101 | 89.77 R3Det-GWD [57] | Resnet101 | 89.85 RSDet [9] | ResNet152 | 86.5 FR-Est [58] | Resnet101 | 89.7 S2A-Net [35] | Resnet101 | 90.2 Oriented RepPoints[59] | Resnet50 | 90.38 ReDet[51] | ReResnet50 | 90.46 Oriented R-CNN[52] | Resnet101 | 90.50 DARDet [54] | Resnet50 | 90.37 CHPDet(ours) | DLA34 | 88.81 CHPDet(ours) | Hourglass104 | 90.55 | | TABLE VI: Detection accuracy on the UCAS-AOD dataset. Method | Backbone | car | airplane | mAP(07) ---|---|---|---|--- YOLOv3 [60] | Darknet53 | 74.63 | 89.52 | 82.08 RetinaNet [23] | Resnet101 | 84.64 | 90.51 | 87.57 FR-O[50] | Resnet101 | 86.87 | 89.86 | 88.36 ROI-trans[30] | Resnet101 | 87.99 | 89.90 | 88.95 FPN-CSL[33] | Resnet101 | 88.09 | 90.38 | 89.23 R3Det-DCL[34] | Resnet101 | 88.15 | 90.57 | 89.36 DAL[56] | Resnet101 | 89.25 | 90.49 | 89.87 CHPDet(ours) | DLA34 | 88.58 | 90.64 | 89.61 CHPDet(ours) | Hourglass104 | 89.18 | 90.81 | 90.00 | | | | #### IV-D6 Bow direction accuracy It can be seen from Table III that the bow direction accuracy of our CHPDet is up to 97.84, 98.14, and 98.39, respectively. This shows that almost all bow directions of ships are correct. As shown in Fig. 9, the pink dots represent the correct head point and the green dots represent the wrong head point. Our detection algorithm can well detect the bow direction of all types of ships, including aircraft carriers, amphibious ships. Only a small number of ships or submarines whose bow and stern are similar from a bird-view perspective, the bow direction will be opposite. ### IV-E Comparison with other methods In this section, we compare our method with other representative ship detectors including RetinaNet-Rbb [23] ROI-trans [30]222https://github.com/dingjiansw101/AerialDetection/, R2CNN [29], CSL [33], DCL [34], RSDet [9], SCRDet [31]333https://github.com/yangxue0827/RotationDetection, and S2A-Net [35]444https://github.com/csuhan/s2anet on three benchmark datasets including FGSD2021, HRSC2016 [45] and UCAS-AOD [46]. To achieve fair comparison, we used the default settings of the original codes on the DOTA dataset including the same data augmentation strategy, and the number of training epochs. Figure 9: Some bow direction detection result of CHPDet. The pink dots represent the correct head point and the green dots represent the wrong head point. #### IV-E1 Results on FGSD2021 We evaluate CHPDet on the FGSD2021 dataset and compare our method with other rotation detection methods. It can be seen from Table III that CHPDet achieves $87.91\%$ mAP at the speed of $41.7$ FPS, which surpass the other compared methods. Compared with the general rotation detection methods RoI-Trans [30] and S2A-Net [35], our proposed method achieves a remarkable improvement by 4.5%, 7.7% in mAP and 19.3, 8.6 in FPS. When higher resolution images are used, the accuracy can be improved to $89.29\%$. This confirms that our method achieves a large superiority in terms of accuracy and speed. To further verify the accuracy of the prediction, we gradually increase the IoU threshold. As can be seen from Table IV, when the IoU threshold is gradually increased, the performance of other detectors dropped significantly, and the decline of our detector is relatively small. When the IoU threshold was increased to $0.8$, the mAP of our CHPDet remained at $71.24$. This shows that our detector can get higher quality rotated boxes than other algorithms. Fig. 8 shows a visual comparison of the detection results of Retinanet-Rbb [23], ROI-Trans [30], SCRDet [31], S2A-Net [35], and our method. As shown in the first row, all the other methods have misclassification or false alarms, S2A-Net [35] has an inaccurate angle prediction, while our method precisely detects them. For the densely parking scene in the second row, all the compared detectors lost at least two submarines, and our method is not influenced by the densely parking scene. The last row of Fig. 8 is a harbor with a complex background. Note that, two ships are not in the water but on the dry dock. ROI-trans [30] and S2A-Net [35] miss the targets, SCRDet [31] has an inaccurate bounding box. Compared to these four methods, our method can better detect the ships in the complex background and is more robust for challenging situations. This improvement mainly comes from three aspects. First, the algorithm makes full use of the prior knowledge of the bow point and improves the accuracy of directional regression. Second, since multi-task learning is performed, bow detection increases the supervision information and improves the accuracy of other tasks. Last, the prior knowledge of ship length is used to refine the confidence of the detected ships belonging to a certain category. The usage of the prior knowledge of ships introduces significant performance improvements. Figure 10: Sample object detection results of our proposed CHPDet on HRSC2016 dataset. Figure 11: Sample object detection results of our proposed CHPDet on UCAS-AOD dataset. #### IV-E2 Results on HRSC2016 The HRSC2016 dataset contains plenty of ships with arbitrary orientations. we evaluate our method on task L1 which contains 1 class and report the results with VOC2007 metric. To demonstrate the performance of our detector, we compare it with other state-of-the-art methods, i.e., ReDet [51], Oriented R-CNN [52], and Oriented RepPoints [59]. The overall comparison performance is reported in Table V. Our method achieves the best performance over all the compared methods, at an accuracy of $90.55\%$. To further show the performance of CHPDet, the detection results are visualized in Fig. 10. As shown in the first two columns, the densely parked ships can be detected well. In the last two columns, there is a lot of background around ships, which is a huge challenge for detectors. The results indicate that our proposed method can avoid false alarms in complex background. #### IV-E3 Results on UCAS-AOD The UCAS-AOD dataset contains a large mount of cars and planes, which are often overwhelmed by a complex background in aerial images. For a fair comparison, we only report the results under VOC2007 metric. Table VI shows the results with the recent methods on the UCAS-AOD dataset. It can be seen that our proposed method achieves the best performance (with an mAP of 90.00%). The CHPDet, which uses a larger output resolution (output stride of 4) compared to traditional object detectors (output stride of 8) and presents ship as the center and head points, can capture abundant information of small objects. Fig. 11 gives some example detection results on the UCAS-AOD dataset. We find that CHPDet performs well in a variety of challenging scenes, which demonstrates the generalization capability of the detector. ## V Conclusion Our proposed approach converts discontinuous angle regression to continuous keypoint estimation by formulating ships as rotated boxes with a head point representing the direction. This design can incorporate the prior knowledge of the bow point, which not only improves the detection performance, but also expands the scope of predicted angle to $[0^{\circ}-360^{\circ})$. Our method can distinguish between bow and stern. CHPDet has simple structure. It has only one positive sample per annotation and simply extracts local peaks in the keypoint heatmap. It does not need Non-Maximum Suppression (NMS). This design ensures high time efficiency. The prior knowledge of ship length is also incorporated to refine the confidence of the detected ships belonging to a certain category. Although our method achieves encouraging results on ship detection from remote sensing images, our method can not be directly used in normal object detection datasets in aerial images such as DOTA [50]. That is because, CHPDet needs more accurate annotations which mark the direction of the target head in the range of $360^{\circ}$. CHPDet is several times faster than most detectors in inference, but it suffers from a long training time. For future work, we will address this issue by encoding more training samples from annotated boxes. In this paper, we proposed a one-stage anchor-free detection framework to detect arbitrary-oriented ships from remote sensing images by making full use of the prior of ships. Our method detects ships by extracting the center, head of ships, and regresses the size of ships at each center point with rotation- invariant features. And we refine the detection results based on the prior information. And we refine the detection results based on the prior information. CHPDet avoids complex anchor design and computing relative to the anchor-based methods and can accurately predict angles in a large range [$0^{\circ}$-$360^{\circ}$). Experimental results demonstrate that our method achieves better accuracy and efficiency as compared with other ship detectors. ## References * [1] S. He, H. Zou, Y. Wang, R. Li, F. Cheng, X. Cao, and M. Li, “Enhancing mid–low-resolution ship detection with high-resolution feature distillation,” _IEEE Geoscience and Remote Sensing Letters_ , 2021. * [2] B. Li, Y. Guo, J. Yang, L. Wang, Y. Wang, and W. An, “Gated recurrent multiattention network for vhr remote sensing image classification,” _IEEE Transactions on Geoscience and Remote Sensing_ , 2021. * [3] Z. Deng, H. Sun, S. Zhou, and J. Zhao, “Learning deep ship detector in sar images from scratch,” _IEEE Transactions on Geoscience and Remote Sensing_ , vol. 57, no. 6, pp. 4021–4039, 2019. * [4] Z. Deng, H. Sun, S. Zhou, J. Zhao, L. Lei, and H. Zou, “Multi-scale object detection in remote sensing imagery with convolutional neural networks,” _ISPRS Journal of Photogrammetry and Remote Sensing_ , vol. 145, pp. 3–22, 2018. * [5] G. Cheng and J. Han, “A survey on object detection in optical remote sensing images,” _ISPRS Journal of Photogrammetry and Remote Sensing_ , vol. 117, pp. 11–28, 2016. * [6] X. Sun, P. Wang, C. Wang, Y. Liu, and K. Fu, “Pbnet: Part-based convolutional neural network for complex composite object detection in remote sensing imagery,” _ISPRS Journal of Photogrammetry and Remote Sensing_ , vol. 173, pp. 50–65, 2021. * [7] Q. He, X. Sun, Z. Yan, and K. Fu, “Dabnet: Deformable contextual and boundary-weighted network for cloud detection in remote sensing images,” _IEEE Transactions on Geoscience and Remote Sensing_ , 2021. * [8] M. Li, W. Guo, Z. Zhang, W. Yu, and T. Zhang, “Rotated region based fully convolutional network for ship detection,” in _IGARSS 2018-2018 IEEE International Geoscience and Remote Sensing Symposium_. IEEE, 2018, pp. 673–676. * [9] W. Qian, X. Yang, S. Peng, Y. Guo, and C. Yan, “Learning modulated loss for rotated object detection,” _arXiv preprint arXiv:1911.08299_ , 2019. * [10] Z. Zhang, W. Guo, S. Zhu, and W. Yu, “Toward arbitrary-oriented ship detection with rotated region proposal and discrimination networks,” _IEEE Geoscience and Remote Sensing Letters_ , vol. 15, no. 11, pp. 1745–1749, 2018\. * [11] X. Zhou, D. Wang, and P. Krähenbühl, “Objects as points,” _arXiv preprint arXiv:1904.07850_ , 2019. * [12] S. Liu, Q. Du, X. Tong, A. Samat, and L. Bruzzone, “Unsupervised change detection in multispectral remote sensing images via spectral-spatial band expansion,” _IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing_ , vol. 12, no. 9, pp. 3578–3587, 2019. * [13] D. S. Maia, M.-T. Pham, E. Aptoula, F. Guiotte, and S. Lefèvre, “Classification of remote sensing data with morphological attributes profiles: a decade of advances,” _IEEE Geoscience and Remote Sensing Magazine_ , 2021. * [14] L. Liu, W. Ouyang, X. Wang, P. Fieguth, J. Chen, X. Liu, and M. Pietikäinen, “Deep learning for generic object detection: A survey,” _International Journal of Computer Vision_ , vol. 128, no. 2, pp. 261–318, 2020. * [15] R. B. Girshick, J. Donahue, T. Darrell, and J. Malik, “Rich feature hierarchies for accurate object detection and semantic segmentation,” in _2014 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2014, Columbus, OH, USA, June 23-28, 2014_. IEEE Computer Society, 2014, pp. 580–587. * [16] R. B. Girshick, “Fast R-CNN,” in _2015 IEEE International Conference on Computer Vision, ICCV 2015, Santiago, Chile, December 7-13, 2015_. IEEE Computer Society, 2015, pp. 1440–1448. * [17] S. Ren, K. He, R. B. Girshick, and J. Sun, “Faster R-CNN: towards real-time object detection with region proposal networks,” in _Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada_ , C. Cortes, N. D. Lawrence, D. D. Lee, M. Sugiyama, and R. Garnett, Eds., 2015, pp. 91–99. * [18] K. He, G. Gkioxari, P. Dollár, and R. B. Girshick, “Mask R-CNN,” in _IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017_. IEEE Computer Society, 2017, pp. 2980–2988. * [19] J. Dai, Y. Li, K. He, and J. Sun, “R-FCN: object detection via region-based fully convolutional networks,” in _Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain_ , D. D. Lee, M. Sugiyama, U. von Luxburg, I. Guyon, and R. Garnett, Eds., 2016, pp. 379–387. * [20] J. Redmon, S. K. Divvala, R. B. Girshick, and A. Farhadi, “You only look once: Unified, real-time object detection,” in _2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016_. IEEE Computer Society, 2016, pp. 779–788. * [21] J. Redmon and A. Farhadi, “YOLO9000: better, faster, stronger,” in _2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017_. IEEE Computer Society, 2017, pp. 6517–6525. * [22] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y. Fu, and A. C. Berg, “Ssd: Single shot multibox detector,” in _European Conference on Computer Vision_. Springer, 2016, pp. 21–37. * [23] T. Lin, P. Goyal, R. B. Girshick, K. He, and P. Dollár, “Focal loss for dense object detection,” in _IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017_. IEEE Computer Society, 2017, pp. 2999–3007. * [24] Z. Cai and N. Vasconcelos, “Cascade R-CNN: delving into high quality object detection,” in _2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018_. IEEE Computer Society, 2018, pp. 6154–6162. * [25] K. Chen, J. Pang, J. Wang, Y. Xiong, X. Li, S. Sun, W. Feng, Z. Liu, J. Shi, W. Ouyang, C. C. Loy, and D. Lin, “Hybrid task cascade for instance segmentation,” in _IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019_. Computer Vision Foundation / IEEE, 2019, pp. 4974–4983. * [26] H. Law and J. Deng, “Cornernet: Detecting objects as paired keypoints,” in _Proceedings of the European Conference on Computer Vision (ECCV)_ , 2018, pp. 734–750. * [27] Z. Tian, C. Shen, H. Chen, and T. He, “FCOS: fully convolutional one-stage object detection,” in _2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019_. IEEE, 2019, pp. 9626–9635. * [28] J. Ma, W. Shao, Y. Hao, W. Li, W. Hong, Y. Zheng, and X. Xue, “Arbitrary-oriented scene text detection via rotation proposals,” _IEEE Transactions on Multimedia_ , vol. PP, no. 99, p. 1, 2017. * [29] Y. Jiang, X. Zhu, X. Wang, S. Yang, W. Li, H. Wang, P. Fu, and Z. Luo, “R2cnn: Rotational region cnn for arbitrarily-oriented scene text detection,” in _2018 24th International Conference on Pattern Recognition (ICPR)_. IEEE, 2018, pp. 3610–3615. * [30] J. Ding, N. Xue, Y. Long, G. Xia, and Q. Lu, “Learning roi transformer for detecting oriented objects in aerial images,” _arXiv: Computer Vision and Pattern Recognition_ , 2018. * [31] X. Yang, J. Yang, Y. Zhang, T. Zhang, Z. Guo, X. Sun, and K. Fu, “Scrdet: Towards more robust detection for small, cluttered and rotated objects,” in _2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019_. IEEE, 2019, pp. 8231–8240. * [32] X. Yang, Q. Liu, J. Yan, A. Li, Z. Zhang, and G. Yu, “R3det: Refined single-stage detector with feature refinement for rotating object,” _arXiv preprint arXiv:1908.05612_ , 2019. * [33] X. Yang and J. Yan, “Arbitrary-oriented object detection with circular smooth label,” pp. 677–694, 2020. * [34] X. Yang, L. Hou, Y. Zhou, W. Wang, and J. Yan, “Dense label encoding for boundary discontinuity free rotation detection,” pp. 15 819–15 829, 2021. * [35] J. Han, J. Ding, J. Li, and G.-S. Xia, “Align deep features for oriented object detection,” _IEEE Transactions on Geoscience and Remote Sensing_ , 2021. * [36] H. Wei, Y. Zhang, Z. Chang, H. Li, H. Wang, and X. Sun, “Oriented objects as pairs of middle lines,” _ISPRS Journal of Photogrammetry and Remote Sensing_ , vol. 169, pp. 268–279, 2020. * [37] H. Wei, Y. Zhang, B. Wang, Y. Yang, H. Li, and H. Wang, “X-linenet: Detecting aircraft in remote sensing images by a pair of intersecting line segments,” _IEEE Transactions on Geoscience and Remote Sensing_ , 2020. * [38] Z. Shi, X. Yu, Z. Jiang, and B. Li, “Ship detection in high-resolution optical imagery based on anomaly detector and local shape feature,” _IEEE Transactions on Geoscience and Remote Sensing_ , vol. 52, no. 8, pp. 4511–4523, 2013. * [39] Y. Freund and R. E. Schapire, “A decision-theoretic generalization of on-line learning and an application to boosting,” _Journal of Computer and System Sciences_ , vol. 55, no. 1, pp. 119–139, 1997. * [40] F. Yang, Q. Xu, and B. Li, “Ship detection from optical satellite images based on saliency segmentation and structure-lbp feature,” _IEEE Geoscience and Remote Sensing Letters_ , vol. 14, no. 5, pp. 602–606, 2017. * [41] Z. Liu, H. Wang, L. Weng, and Y. Yang, “Ship rotated bounding box space for ship extraction from high-resolution optical satellite images with complex backgrounds,” _IEEE Geoscience and Remote Sensing Letters_ , vol. 13, no. 8, pp. 1074–1078, 2017. * [42] Z. Liu, J. Hu, L. Weng, and Y. Yang, “Rotated region based cnn for ship detection,” in _IEEE International Conference on Image Processing_ , 2018\. * [43] Z. Cao, T. Simon, S. Wei, and Y. Sheikh, “Realtime multi-person 2d pose estimation using part affinity fields,” in _2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017_. IEEE Computer Society, 2017, pp. 1302–1310. * [44] Y. Zhou, Q. Ye, Q. Qiu, and J. Jiao, “Oriented response networks,” in _2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017_. IEEE Computer Society, 2017, pp. 4961–4970. * [45] Z. Liu, L. Yuan, L. Weng, and Y. Yang, “A high resolution optical satellite image dataset for ship recognition and some new baselines,” in _International Conference on Pattern Recognition Applications and Methods_ , vol. 2. SCITEPRESS, 2017, pp. 324–331. * [46] C. Li, C. Xu, Z. Cui, D. Wang, T. Zhang, and J. Yang, “Feature-attentioned object detection in remote sensing imagery,” in _2019 IEEE International Conference on Image Processing (ICIP)_. IEEE, 2019, pp. 3886–3890. * [47] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” in _3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings_ , Y. Bengio and Y. LeCun, Eds., 2015. * [48] F. Yu, D. Wang, E. Shelhamer, and T. Darrell, “Deep layer aggregation,” in _2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018_. IEEE Computer Society, 2018, pp. 2403–2412. * [49] A. Newell, K. Yang, and J. Deng, “Stacked hourglass networks for human pose estimation,” pp. 483–499, 2016. * [50] G. Xia, X. Bai, J. Ding, Z. Zhu, S. J. Belongie, J. Luo, M. Datcu, M. Pelillo, and L. Zhang, “DOTA: A large-scale dataset for object detection in aerial images,” in _2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018_. IEEE Computer Society, 2018, pp. 3974–3983. * [51] J. Han, J. Ding, N. Xue, and G.-S. Xia, “Redet: A rotation-equivariant detector for aerial object detection,” in _Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition_ , 2021, pp. 2786–2795. * [52] X. Xie, G. Cheng, J. Wang, X. Yao, and J. Han, “Oriented r-cnn for object detection,” _arXiv preprint arXiv:2108.05699_ , 2021. * [53] J. Yi, P. Wu, B. Liu, Q. Huang, H. Qu, and D. Metaxas, “Oriented object detection in aerial images with box boundary-aware vectors,” in _Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision_ , 2021, pp. 2150–2159. * [54] F. Zhang, X. Wang, S. Zhou, and Y. Wang, “Dardet: A dense anchor-free rotated object detector in aerial images,” _arXiv preprint arXiv:2110.01025_ , 2021\. * [55] Y. Xu, M. Fu, Q. Wang, Y. Wang, K. Chen, G.-S. Xia, and X. Bai, “Gliding vertex on the horizontal bounding box for multi-oriented object detection,” _IEEE Transactions on Pattern Analysis and Machine Intelligence_ , vol. 43, no. 4, pp. 1452–1459, 2020. * [56] Q. Ming, Z. Zhou, L. Miao, H. Zhang, and L. Li, “Dynamic anchor learning for arbitrary-orientedd object detection,” _arXiv preprint arXiv:2012.04150_ , vol. 1, no. 2, p. 6, 2020. * [57] X. Yang, J. Yan, Q. Ming, W. Wang, X. Zhang, and Q. Tian, “Rethinking rotated object detection with gaussian wasserstein distance loss,” _arXiv preprint arXiv:2101.11952_ , 2021. * [58] K. Fu, Z. Chang, Y. Zhang, and X. Sun, “Point-based estimator for arbitrary-oriented object detection in aerial images,” _IEEE Transactions on Geoscience and Remote Sensing_ , 2020. * [59] W. Li and J. Zhu, “Oriented reppoints for aerial object detection,” _arXiv preprint arXiv:2105.11111_ , 2021. * [60] J. Redmon and A. Farhadi, “Yolov3: An incremental improvement,” _arXiv preprint arXiv:1804.02767_ , 2018. | Feng Zhang received the B.E. degree in electronic information engineering from Harbin Institute of Technology(HIT), Harbin, China, in 2009, and the M.E. degree in information and communication engineering from National University of Defense Technology (NUDT), Changsha, China, in 2011. He is currently pursuing a Ph.D. degree from the College of Electronic Science and Technology, NUDT. His research interests focus on include remote sensing image processing, pattern recognition, and computer vision. ---|--- | Xueying Wang received the B.S. degree in electronic information engineering from Beihang University, Beijing, China, in 2009, the M.S. and Ph.D. degrees in electronic science and technology from the National University of Defense Technology, Changsha, China, in 2011 and 2016. He is currently an Assistant Professor with the College of Electrical Science, National University of Defense Technology. His research interests include remote sensing image processing, pattern recognition. ---|--- | Shilin Zhou received the B.S., M.S., and Ph.D. degrees in electrical engineering from Hunan University, Changsha, China, in 1994, 1996, and 2000, respectively. He is currently a Full Professor with the College of Electrical Science, National University of Defense Technology, Changsha. He has authored or co-authored over 100 referred papers. His research interests include image processing and pattern recognition. ---|--- | Yingqian Wang received the B.E. degree in electrical engineering from Shandong University (SDU), Jinan, China, in 2016, and the M.E. degree in information and communication engineering from National University of Defense Technology (NUDT), Changsha, China, in 2018. He is currently pursuing a Ph.D. degree from the College of Electronic Science and Technology, NUDT. He has authored several papers in journals and conferences such as TPAMI, TIP, CVPR, and ECCV. His research interests focus on low-level vision, particularly on light field imaging and image super-resolution. ---|--- | Yi Hou received the B.S. degree from Wuhan University, China, and the M.S. as well as a Ph.D. degree from the National University of Defense Technology, China. He held a visiting position with the Department of Computing Science, University of Alberta, Canada, from 2014 to 2016. His main research interests include robot visual SLAM, visual place recognition, time series classification, signal processing, computer vision, deep learning, pattern recognition, and image processing. ---|---
# Boost-S: Gradient Boosted Trees for Spatial Data and Its Application to FDG-PET Imaging Data Reza Iranzad Department of Industrial Engineering, University of Arkansas Xiao Liu Department of Industrial Engineering, University of Arkansas W. Art Chaovalitwongse Department of Industrial Engineering, University of Arkansas Daniel S. Hippe Department of Radiology, University of Washington Shouyi Wang Department of Industrial, Manufacturing & Systems Engineering, University of Texas at Arlington Jie Han Department of Industrial, Manufacturing & Systems Engineering, University of Texas at Arlington Phawis Thammasorn Department of Industrial Engineering, University of Arkansas Chunyan Duan Department of Mechanical Engineering, Tongji University Jing Zeng Department of Radiation Oncology, University of Washington Stephen R. Bowen Department of Radiology, University of Washington Department of Radiation Oncology, University of Washington ###### Abstract Boosting Trees are one of the most successful statistical learning approaches that involve sequentially growing an ensemble of simple regression trees (i.e., “weak learners”). This paper proposes a new gradient Boosted Trees algorithm for Spatial Data (Boost-S) for spatially correlated data with covariate information. Boost-S integrates the spatial correlation structure into the classical framework of gradient boosted trees. Each tree is grown by solving a regularized optimization problem, where the objective function involves two penalty terms on tree complexity and takes into account the underlying spatial correlation. A computationally-efficient algorithm is proposed to obtain the ensemble trees. The proposed Boost-S is applied to the spatially-correlated FDG-PET (fluorodeoxyglucose-positron emission tomography) imaging data collected during cancer chemoradiotherapy. Our numerical investigations successfully demonstrate the advantages of the proposed Boost-S over existing approaches for this particular application. Key words: Gradient Boosted Trees, Spatial Statistics, FDG-PET, Chemoradiotherapy ## Introduction Spatial data refer to an important type of data which arise in a spatial area and are often correlated in space. Capturing such correlation is the centerpiece of statistical analysis of spatial data (Cressie and Wikle, 2011b; Schabenberger and Gotway, 2005). Applications of statistical spatial data modeling can be found in a spectrum of scientific and engineering applications ranging from energy (Ezzat et al., 2019), reliability (Liu et al., 2018c; Fang et al., 2019), quality and manufacturing (Zang and Qiu, 2019; Wang et al., 2016; Yue et al., 2020), environmental and natural process (Guinness and Stein, 2013; Liu et al., 2018a, 2020), medical informatics (Yao et al., 2017; Yan et al., 2019), etc. ### The Problem Statement In this paper, we are concerned with the modeling problem: $Y(\bm{s})=Z(\bm{x}_{\bm{s}})+\varepsilon(\bm{s}),\quad\quad\bm{s}\in\mathcal{S}$ (1) where $Y(\bm{s})$ represents the observation at a spatial location $\bm{s}\in\mathcal{S}$, $\bm{x}_{\bm{s}}$ is a vector that collects the available features at $\bm{s}$, $Z(\bm{x})\equiv\mathbb{E}(Y;\bm{x})$ is the mean-value function given $\bm{x}$, and $\varepsilon(\bm{s})$ is an isotropic and weakly stationary spatial noise process with zero-mean and a covariance $C(h)$, where $C(h)=\text{cov}(Y(\bm{s}),Y(\bm{s}^{\prime}))$ with $h$ being some distance measure between $\bm{s}$ and $\bm{s}^{\prime}$; for example, the Euclidean distance. The goal of this paper is to tackle the modeling problem (1) by devising a new additive-tree-based method that approximates $Z(\bm{x})$ as follows: $Z(\bm{x})\approx\phi(\bm{x})=\sum_{k=1}^{K}f_{k}(\bm{x}),\quad\quad f\in\mathcal{F}$ (2) where $\\{f_{k}(\bm{x})\\}_{k=1}^{K}$ represents an ensemble of binary decision trees, and $\mathcal{F}$ represents the tree space. Apparently, the problem hinges on how the ensemble of trees $\\{f_{k}(\bm{x})\\}_{k=1}^{K}$ can be grown, while taking into account the important spatial correlation of $Y(\bm{s})$. In particular, we exploit the idea of gradient boosting which involves sequentially growing an ensemble of simple regression trees. Mathematically, each tree is added to the ensemble by solving an optimization problem with a carefully chosen objective function. This goal boils down to three fundamental tasks to be addressed in this paper: (i) formulate a (regularized) optimization problem that balances the complexity of individual trees and takes into account the spatial correlation associated with $Y(\bm{s})$; (ii) devise an algorithm that solves the regularized optimization problem in a computationally efficient manner (so that trees can be added sequentially to the ensemble); and (iii) validate the performance of the proposed method on real datasets. ### A Motivating Application A motivating application is first presented. Fluorodeoxyglucose-Positron Emission Tomography (FDG-PET) has been widely used in cancer diagnosis and treatment to detect metabolically active malignant lesions, and plays a critical role in quantitatively assessing and monitoring tumor response to treatment. For illustrative purposes, Figure 1 shows the Standardized Uptake Values (SUV) obtained from the FDG-PET imaging of a patient. In this figure, the top row shows the baseline image taken before the radiotherapy (i.e., Pre- RT), while the bottom row shows the FDG PET/CT imaging 3 weeks after radiotherapy (i.e., Mid-RT image). In this case, it is possible to observe the shrinkage of the tumor, indicating the effectiveness of treatment. Hence, the difference between the Mid-RT and Pre-RT images can be naturally used to quantify the tumor’s spatial response to treatment. Figure 1 also suggests that a tumor typically presents spatially-correlated and spatially- heterogeneous responses. Some areas of a tumor may respond well to treatment while some areas appear to be less responsive. The importance of capturing such spatially-varying and spatially-correlated responses has been discussed in Bowen et al. (2019). Figure 1: An illustrative example of spatially-correlated Pre-RT and Mid-RT FDG-PET images Figure 2: 3D Pre-RT and Mid-RT SUV images for two patients A tumor is a three-dimensional object. Figure 2 shows the SUV in three- dimensional spaces for another two patients. The top row shows the Pre-RT data, while the bottom row shows the Mid-RT data collected 3 weeks after the radiotherapy. It is seen that, * • The overall SUV level decreases 3 weeks after the radiotherapy; * • The change in SUV (i.e., tumor’s response to radiotherapy) varies over space. In particular, the SUV level gradually decreases from the centroid to the surface of a tumor. Hence, if we let $Y(\bm{s})$ in (1) represent the change in SUV level at location $\bm{s}$ within the spatial domain $\mathcal{S}$ occupied by the tumor, then, a statistical spatial model is needed that enables clinicians to understand how $Y(\bm{s})$ depends on a vector of features (including geometric features, therapy dosage, and so on) for treatment optimization and control. However, in practice, the complex relationship between the response $Y$ and features $\bm{x}$ can hardly be directly specified. The inevitably nonlinear and interaction effects motivate us to investigate the non- parametric tree-based methods in this paper. More detailed discussions and rationales are provided in the next section. ### Literature Review and Contributions The pioneering work of statistical modeling of spatial data can be found in Banerjee et al. (2004), Schabenberger and Gotway (2005) and Cressie and Wikle (2011a). The mainstay approach, also known as the geostatistical paradigm, models spatial processes by random fields with fully specified covariance functions. A linear relationship between covariates and process mean is often assumed, and the model parameters can be obtained through Generalized Least Squares or Maximum Likelihood Estimation. The covariance structures are typically derived from moments of multivariate probability distributions and motivated by considerations of mathematical convenience (e.g., stationary, isotropic, space-time separable, etc.). Hence, there have been prolonged research interests to provide flexible and effective ways to construct non- stationary covariance functions (Cressie and Huang, 1999; Gneiting, 2002; Fuentes et al., 2005; Gneiting et al., 2006; Ghosh et al., 2010; Reich et al., 2011; Lenzi et al., 2019). The geostatistical modeling paradigm, which heavily relies on random fields, becomes less practical for large problems and approximations are commonly used, such as the Gaussian Markov Random Fields representation (Lindgren and Rue, 2011), Nearest-Neighbor Gaussian Process (Datta et al., 2016; Banerjee, 2017), kernel convolution (Higdon, 1998), low- rank representation (Cressie and Johannesson, 2002; Nychka and Wikle, 2002; Banerjee et al., 2008), the approximation of likelihood functions (Stein et al., 2004; Fuentes, 2007; Guinness and Fuentes, 2015), Bayesian inference for latent Gaussian models based on the integrated nested Laplace approximations (Rue et al., 2009; R-INLA, 2019), Lagrangian spatio-temporal covariance function (Gneiting et al., 2006), matrix-free state-space model (Mondal and Wang, 2019), Vecchia approximations of Gaussian processes (Katzfuss et al., 2020), as well as the multi-resolution approximation (M-RA) of Gaussian processes observed at irregular spatial locations (Katzfuss, 2017). Other spatial models have also been proposed in the literature. Notably, the Markov Random Fields (MRF) model focuses on the (conditional) distribution of the quantity at a particular spatial location given the neighboring observations, such as the auto Poisson model (Besag, 1974), Conditional Autoregressive Model (Carlin and Banerjee, 2003; Liu et al., 2018b), etc. The Spatial Moving Average (SMA) approach models a spatial process through a process convolution with a convolution kernel (Higdon, 1998; Brown et al., 2000; Liu et al., 2016). There is also a large body of literature focusing on spatio-temporal data. For example, the Hierarchical Dynamical Spatio-Temporal Models (DSTM) (Wikle and Cressie, 1999; Berliner, 2003; Cressie and Wikle, 2011a; Stroud et al., 2010; Katzfuss et al., 2020), and the SPDE-based modeling approach that aims to integrate governing physics into statistical spatio-temporal models (Brown et al., 2000; Hooten and Wikle, 2008; Stroud et al., 2010; Sigrist et al., 2015; Liu et al., 2020). A summary of the latest advances in the spatial modeling with SPDE can be found in Cressie and Wikle (2011a) and Krainski et al. (2019). For spatial models in the form of (1), one challenge arising from practice is to specify the relationship between the covariates $\bm{x}$ and response $Y$, which can rarely be adequately captured by linear models. For the FDG-PET imaging data presented in Section 1.2, for example, both non-linear and interaction effects are expected between tumor’s response and covariates (such as treatment, geometric features of the tumor, etc.). Hence, non-parametric approaches, especially the additive-tree-based approaches, provide some major modeling advantages. Constructing a tree does not require parametric assumptions on the complex relationship between features and event processes. An individual tree performs a partition of the feature space. For each sub feature space, a predicted value is found for the individuals over that sub feature space. A sum-of-trees model consists of multivariate components that effectively handle the complex interaction effects among features (Chipman et al., 2010). Feature selection is also possible under the framework of additive-tree-based models (Hastie et al., 2009; Liu and Pan, 2020). Among the additive-tree-based methods, gradient boosted trees have become one of the most successful statistical learning approaches over the last two decades, generating 17 winning solutions among 29 Kaggle challenges in 2015 (Chen and Guestrin, 2016). Schapire (1999) introduced the first applicable Boosting method. The main idea of gradient boosting hinges on fitting a sequence of correlated “weaker learners” (such as simple trees). Each tree explains only a small amount of variation not captured by previous trees (Hastie et al., 2009; Chipman et al., 2010). However, many existing boosting trees, such as XGBoost, do not consider the possible spatial correlation when they are applied to spatial data. To our best knowledge, Sigrist (2020) recently proposed the only boosted-trees-based approach for Gaussian Process and mixed effects model (which captures correlated errors). Such a method minimizes the negative log-likelihood at each tree node splitting, and is available in the R package, GPBoost. The package also provides a range of regularization and tuning parameter options. In our paper, on the other hand, regularizations have been directly added to the objective function in order to control the complexity of individual trees (i.e., number of leaves and leaf weights), leading to a regularized optimziation problem at each tree node splitting. The regularization terms are motivated by the fundamental idea behind boosting trees which involves a sequence of correlated simple trees. This idea has been adopted by XGBoost (Chen and Guestrin, 2016), while an alternative approach adopted by the Bayesian Additive Regression Trees (BART) involves assigning prior distributions on parameters charactering the tree structure (Chipman et al., 2010). Hence, the main contribution of the paper is to propose a computationally-efficient gradient boosting method for growing the ensemble trees $\\{f_{k}(\bm{x})\\}_{k=1}^{K}$ in (2) for spatially- correlated data. Each tree is grown by solving a regularized optimization problem, where the objective function involves regularizations on tree complexity and takes into account the underlying spatial correlation. The proposed algorithm is referred to as Boost-S, which stands for Gradient Boosted Trees for Spatial Data with covariate information (Boost-S). Boost-S integrates the spatial correlation structure into the classical framework of gradient boosted trees. In Statistics, Ordinary Least Squares is extended to Generalized Least Squares for correlated data. An analogous notion can be formulated here when extending the classical framework of gradient boosted trees to spatially-correlated data, giving rise to the proposed Boost-S. The rest of this paper is structured as follows: Section 2 presents the technical details of the Boost-S algorithm. The applications and numerical illustrations of Boost-S are presented in Section 3. Section 4 concludes the paper. ## Boost-S: Technical Details This section provides the technical details behind Boost-S. Suppose that data are collected from a number of $n$ spatial locations, $\bm{s}_{1},\bm{s}_{2},...,\bm{s}_{n}$. At each location, we observe a response $y$ and a $m$-dimensional feature vector $\bm{x}=(x_{1},x_{2},...,x_{m})^{T}$. From (1) and (2), Boost-S aims to construct an ensemble of binary trees, $\\{f_{k}(\bm{x})\\}_{k=1}^{K}$, using gradient boosting such that $Y(\bm{s})=\sum_{k=1}^{K}f_{k}(\bm{x})+\varepsilon(\bm{s}),\quad\quad\bm{s}\in\\{\bm{s}_{1},\bm{s}_{2},...,\bm{s}_{n}\\}.$ (3) Let $\bm{Y}=(Y(\bm{s}_{1}),Y(\bm{s}_{2}),...,Y(\bm{s}_{n}))^{T}$ be a multivariate random vector representing the responses from the $n$ spatial locations, and let $\bm{f}^{(k)}$ be a vector of predicted values at the $n$ spatial locations generated from the $k$th tree ($k\geq 0$), we re-write (3) as $\bm{Y}=\sum_{k=0}^{K}\bm{f}^{(k)}+\bm{\varepsilon},\quad\quad\bm{\varepsilon}\sim\mathcal{N}(\bm{0},\bm{\Sigma}_{\bm{\theta}}).$ (4) where $\bm{f}^{(0)}$ is a vector of zeros corresponding to the initial condition when no tree has been grown. ### A Regularized Problem This subsection presents the detailed tree structures and formulates a regularized optimization problem that leads to a sequence of ensemble trees. In (3), each tree $f_{k}$ is a Classification and Regression Tree (CART) that resides in a binary tree space $\mathcal{F}=\left\\{f(\bm{x})=w_{q(\bm{x})}\right\\}$ (5) where $q:\mathcal{R}^{p}\rightarrow T$ and $w\in\mathcal{R}^{T}$. Here, $T$ represents the number of tree leaves (i.e., terminal nodes), $w$ is the value on a tree leaf (i.e., leaf weight), and $q$ determines the tree structure (i.e., a mapping that links a feature vector $\bm{x}$ to a tree leaf). Suppose that a number of $k-1$ trees have been grown ($k\geq 1$). Then, the (ensemble) predicted values at the $n$ spatial locations are given by $\hat{\bm{y}}^{(k-1)}=\sum_{j=0}^{k-1}\bm{f}^{(j)}$ from the $k-1$ trees. An immediate next step is to construct the $k$th tree and add the new tree to the ensemble such that $\hat{\bm{y}}^{(k)}=\hat{\bm{y}}^{(k-1)}+\bm{f}^{(k)}=\sum_{j=0}^{k}\bm{f}^{(j)}$. For binary trees, this involves finding the optimal split features as well as the split points for the $k$th tree. This task can be formulated as a regularized optimization problem: $\min_{\bm{f}^{(k)}}\left\\{\ell(\hat{\bm{y}}^{(k-1)}+\bm{f}^{(k)})+\Omega(\bm{f}^{(k)})\right\\}$ (6) where $\ell$ is a loss function that depends on the output of the $k$th tree, and the regularization $\Omega$ is given by: $\Omega(\bm{f})=\gamma T+\frac{1}{2}\lambda\|\bm{w}\|^{2}.$ (7) The first term, $\gamma T$, regularizes the depth of the tree (by penalizing the total number of leaves), while the second term is used to regularize the contribution of tree $k$ to the ensemble predictions (by penalizing the weights on tree leaves). Recall that, the fundamental idea behind boosting trees is to construct a sequence of correlated “weaker learners” (i.e., simple trees), where each “weaker learner” is added to explain the unexplained variation by existing trees in the ensemble (Chipman et al., 2010). Hence, the regularization (7) effectively controls the complexity of individual trees. In fact, it is worth noting that the regularization also helps to prevent the well-known overfitting issue of boosting trees. When a sufficient number of trees have been included in the ensemble, the penalty of adding one more tree may dominate the benefit (of adding more trees), which stops the algorithm from growing more trees. Because the multivariate response $\bm{Y}$ is spatially correlated with the covariance matrix $\bm{\Sigma}_{\bm{\theta}}$, a sensible choice for the loss function $\ell$ is the squared Mahalanobis length of the residual vector: $\begin{split}\ell(\hat{\bm{y}}^{(k-1)}+\bm{f}^{(k)})&=\ell(\hat{\bm{y}}^{(k)})\\\ &\equiv(\bm{y}-\hat{\bm{y}}^{(k)})^{T}\bm{\Sigma}^{-1}_{\theta}(\bm{y}-\hat{\bm{y}}^{(k)})\end{split}$ (8) and (6) can thus be written as $\min_{\bm{f}^{(k)}}\left\\{(\bm{y}-\hat{\bm{y}}^{(k)})^{T}\bm{\Sigma}^{-1}_{\theta}(\bm{y}-\hat{\bm{y}}^{(k)})+\gamma T+\frac{1}{2}\lambda\|\bm{w}\|^{2}\right\\}.$ (9) Note that, (9) above extends the classical XGBoost which does not consider the correlation among the elements of $\bm{Y}$ (Chen and Guestrin, 2016). The extension made by this paper is in analogy to the extension from Ordinary Least Squares to Generalized Least Squares. However, such an extension requires new algorithms for the problem (9) to be efficiently solved. The regularized optimization above is a formidable combinatorial optimization problem which can hardly be directly solved. Hence, we approximate the objective function $\ell(\hat{\bm{y}}^{(k)})+\Omega(\bm{f}^{(k)})$ by a second-order multivariate Taylor expansion: $\begin{split}\ell(\hat{\bm{y}}^{(k)})&+\Omega(\bm{f}^{(k)})\\\ &\approx\ell(\hat{\bm{y}}^{(k-1)})+\bm{g}^{T}\bm{f}^{(k)}+\frac{1}{2}(\bm{f}^{(k)})^{T}\bm{H}\bm{f}^{(k)}+\Omega(\bm{f}^{(k)})\end{split}$ (10) where $\bm{g}$ is the column gradient vector of the loss function with its $i$th element given by $g_{i}=\partial\ell(\hat{\bm{y}}^{(k-1)})/\partial\hat{y}_{i}^{(k-1)}$, and $\bm{H}$ is the Hessian matrix with its $(i,j)$th entry being given by $h_{i,j}=\partial^{2}\ell(\hat{\bm{y}}^{(k-1)})/\partial\hat{y}_{i}^{(k-1)}\partial\hat{y}_{j}^{(k-1)}$. Because the first term on the right-hand-side of (10) is a constant, it is sufficient to minimize the sum of the remaining three terms: $L^{(k)}=\bm{g}^{T}\bm{f}^{(k)}+\frac{1}{2}(\bm{f}^{(k)})^{T}\bm{H}\bm{f}^{(k)}+\Omega(\bm{f}^{(k)}).$ (11) Note that, for any given tree structure, it is possible to define a set $I_{p}=\\{i\mid q(x_{i})=p\\}$ that consists of all samples that fall into leaf $p$. Then, we let $\bm{g}_{p}$ be a column vector that only retains the elements in $\bm{g}$ corresponding to samples in $I_{p}$, and similarly, let $\bm{H}_{p,q}$ be a matrix by only keeping the rows and columns of $\bm{H}$ corresponding to samples in $I_{p}$ and $I_{q}$, respectively. Figure 3 provides an illustration of how $\bm{g}_{p}$ and $\bm{H}_{p,q}$ are constructed. Consider a simple example where $n=7$ (i.e., only 7 samples are available), then, the dimensions of the gradient vector $\bm{g}$ and the Hessian matrix $\bm{H}$ are $7\times 1$ and $7\times 7$, respectively. Suppose that $I_{p}={1,2,6}$ and $I_{q}={3,4}$. Then, the vector $\bm{g}_{p}$ consists of the 1st, 2nd and the 6th element in $\bm{g}$, and the matrix $\bm{H}_{p,q}$ is a $3\times 2$ matrix that retains the entries $\bm{H}$ located at the intersections of rows 1, 2 and 6, and columns 3 and 4. Figure 3: An illustration of the vector $\bm{g}_{p}$ and the matrix $\bm{H}_{p,q}$ Then, for any given tree structure $q(\bm{x})$, we re-write (11) as follows: $\displaystyle L^{(k)}=\sum_{p=1}^{T}\bm{g}^{T}_{p}\bm{w}_{p}+\frac{1}{2}\left\\{\sum_{p\in T}\bm{w}^{T}_{p}\bm{H}_{p,p}\bm{w}_{p}\right\\}$ (12) $\displaystyle+\frac{1}{2}\left\\{\sum_{(p,q)\in C_{2}^{T}}\bm{w}^{T}_{p}\bm{H}_{p,q}\bm{w}_{q}\right\\}+\Omega(\bm{f}^{(k)})$ where $C_{2}^{T}$ denotes the combination (i.e., the number of 2-combinations from a given set of $T$ elements), $\bm{w}_{p}=w_{p}\bm{1}_{|I_{p}|}$ with $w_{p}$ and $\bm{1}_{|I_{p}|}$ respectively being the weight on tree leaf $p$ and a column vector of ones of dimension $|I_{p}|$, and $|\cdot|$ represents the cardinality of a set. Substituting (7) into (12) yields $\begin{split}L^{(k)}=&\gamma T+\sum_{p=1}^{T}\left\\{w_{p}\sum_{i\in I_{p}}g_{i}+\frac{1}{2}\left[\lambda+\sum_{(i,j)\in I_{p}}h_{i,j}\right]w_{p}^{2}\right\\}\\\ &+\frac{1}{2}\sum_{p=1}^{T}\left\\{\sum_{q=1;q\neq p}^{T}\left[\lambda+\sum_{i\in I_{p};j\in I_{q}}h_{i,j}\right]w_{p}w_{q}\right\\}.\end{split}$ (13) Taking the partial derivatives of (13) with respect to $\\{w_{p}\\}_{p=1}^{T}$ leads to a system of equations, which provides the key computational advantage: given any tree structure, the optimal weights $\bm{w}=\\{w_{1},w_{2},...,w_{T}\\}$ can be quickly found by solving the linear system: $\Xi\bm{w}=-\tilde{\bm{g}}$ (14) where $\Xi$ is a $T\times T$ matrix with its $p$th row given by $\frac{1}{2}(\lambda+\sum_{i\in I_{p};j\in I_{1}}h_{i,1}),\frac{1}{2}(\lambda+\sum_{i\in I_{p};j\in I_{2}}h_{i,2}),...,(\lambda+\sum_{(i,j)\in I_{p}}h_{i,j}),...,\frac{1}{2}(\lambda+\sum_{i\in I_{p};j\in I_{T}}h_{i,T})$ (15) and $\tilde{\bm{g}}$ is a $T\times 1$ vector with its $p$th element given by $\sum_{i\in I_{p}}g_{i}$. Obtaining the linear system (14) plays an extremely important role in searching for the optimal tree: given any candidate tree structure, it is possible to quickly and accurately find the optimal weights $\bm{w}$ on the leaf nodes by solving (14) using least squares, i.e., no numerical search is required. By substituting the optimal $\bm{w}$ into (11) immediately yields the value of the objective function $L^{(k)}$. ### The Sequential Update of the Unknown Covairance Matrix Constructing the linear system (14) and evaluating the objective function $L^{(k)}$ require a known covariance matrix of the errors, $\bm{\Sigma}_{\bm{\theta}}$. However, $\bm{\Sigma}_{\bm{\theta}}$ is not known and needs to be estimated before the first tree ($k=1$), or any subsequent tree ($k>1$), can be constructed. In this section, we describe how $\bm{\Sigma}_{\bm{\theta}}$ can be consistently estimated before the $k$th tree is constructed, given the outputs from the first $k-1$ trees. Suppose that $k-1$ trees have been constructed ($k\geq 1$), and let $\bm{r}^{(k-1)}=\bm{Y}-\sum\bm{f}^{(k-1)}$ be the residual vector. It follows from (4) that $\bm{r}^{(k-1)}$ is Gaussian with the covariance $\bm{\Sigma}_{\bm{\theta}}$. Note that, the mean of $\bm{r}^{(k-1)}$ may not even be close to zero when $k$ is small, i.e., when there are not sufficient trees in the ensemble to well capture the mean of $\bm{Y}$. In this case, we model $\bm{r}^{(k-1)}$ by a Locally Weighted Mixture of Linear Regressions (LWMLR) (Stroud et al., 2001): $\bm{r}^{(k-1)}=\sum_{j=1}^{J}\pi^{T}_{j}(\bm{s})\bm{k}_{j}(\bm{s})\bm{\beta}_{j}+\varepsilon(\bm{s}),\quad\quad k\geq 1$ (16) where $\bm{k}_{j}(\bm{s})=\\{k_{j1}(\bm{s}),...,k_{jq}(\bm{s})\\}$ is a set of spatial basis functions, $\bm{\beta}_{j}=(\beta_{j1},...,\beta_{jq})$ is a vector of unknown coefficients, $\pi_{j}(\bm{s})$ is a Gaussian kernel given as follows: $\pi_{j}(\bm{s})\propto|\bm{V}_{j}|^{-1/2}\exp\left\\{-\frac{1}{2}(\bm{s}-\bm{\mu}_{j})^{T}\bm{V}_{j}^{-1}(\bm{s}-\bm{\mu}_{j})\right\\}$ (17) Note that, we may re-write (16) as a linear model, $\bm{r}^{(k-1)}=\bm{X}\bm{B}+\bm{\varepsilon}$, where $\bm{X}=(\text{diag}(\pi_{1})\bm{X}_{1},...,\text{diag}(\pi_{J})\bm{X}_{J})$, $\bm{X}_{j}=(\bm{k}_{j}^{T}(\bm{s}_{1}),\bm{k}_{j}^{T}(\bm{s}_{2}),...,\bm{k}_{J}^{T}(\bm{s}_{J}))^{T}$, and $\bm{B}=(\bm{\beta}_{1},\bm{\beta}_{2},...,\bm{\beta}_{J})^{T}$. Then, it is possible to obtain a consistent estimate of $\bm{\Sigma}_{\bm{\theta}}$, $\hat{\bm{\Sigma}}_{\bm{\theta}}^{(k-1)}$, using the Feasible Generalized Least Square (FGLS), before the $k$th tree can be constructed. ### The Boost-S Algorithm Following the discussions above, Figure 4 provides a high-level illustration of the flow of the Boost-S algorithm. At the initialization stage, we obtain the initial estimate of the covariance matrix $\hat{\bm{\Sigma}}_{\bm{\theta}}^{(0)}$. Then, by solving the regularized optimization problem (6), a new tree $k$ is constructed and added to the ensemble. Next, the estimate of the covariance matrix is updated $\hat{\bm{\Sigma}}_{\bm{\theta}}^{(k)}$ before tree $k+1$ can be constructed. The steps are repeated until $K$ trees have been grown in the ensemble. The Boost-S algorithm is formalized by Algorithm 1. Set the values for $\lambda$, $\gamma$ and $K$ Let k = 0, $\bm{f}^{(0)}=\bm{0}$ and $\bm{r}^{(0)}=\bm{y}$ Obtain the initial estimate, $\hat{\bm{\Sigma}}_{\bm{\theta}}^{(0)}$, from (16) using FGLS for _k=1,…,K_ do Grow tree $k$ by repeating the following steps: (i) Given the current tree topology, generate a set of all possible new tree structures by splitting a tree node based on candidate split variables and candidate split values. (ii) For each new tree structure, obtain the weights $\bm{\omega}$ on the leaf nodes by solving the linear system $\Xi\bm{w}=-\tilde{\bm{g}}$, and evaluate the objective function $L^{(k)}$ in (11) for the new topology. (iii) If there exists at least one new tree structure that further reduces the objective function (over the existing tree structure), retain the new topology that generates the greatest reduction and go to (i); otherwise, terminate the tree growing process for tree $k$, and go to the next step. (iv) Update the estimate $\hat{\bm{\Sigma}}_{\bm{\theta}}^{(k)}$using FGLS. end for Algorithm 1 Boost-S: Gradient Boosted Trees for Spatial Data Figure 4: An illustration of the flow of the Boost-S algorithm ## Illustration of Boost-S on the FDG-PET Imaging Data This section re-visits the motivating application presented in Section 1.2. We apply Boost-S to real datasets and compare the performance of Boost-S to that of existing approaches. ### Data The data used in this section are obtained from 25 patients diagnosed with locally advanced and surgically unresctable non-small cell lung cancer (NSCLC) who enrolled onto the FLARE-RT clinical trial (NCT02773238). For each patient, this clinical trial data set contains geographic features, dosage, Pre-RT and Mid-RT SUV levels. As discussed in Section 1.2, the goal is to model the difference between the Pre-RT SUV (before treatment) and Mid-RT SUV (during third week of treatment course), which helps clinicians further optimize or control the treatment plans. Let $Y(\bm{s})$ represent the ratio between Mid-RT SUV and Pre-RT SUV at location $\bm{s}$ (a voxel in the image) within the spatial domain $\mathcal{S}$ occupied by the tumor, i.e., $Y(\bm{s})=\frac{\text{Mid-RT SUV}}{\text{Pre-RT SUV}}.$ (18) If the treatment is effective, the Mid-RT SUV is expected to be lower than the Pre-RT SUV level. Hence, a lower ratio indicates more effective treatment. ### Application of Boost-S We first demonstrate the application of Boost-S on the data collected from one of the 25 patients. The PET scan for this patient has 3110 voxels (i.e., the number of spatial locations). We randomly split the data into two parts: 15% for training while 85% for testing. Such a low training-testing ratio is chosen to demonstrate the out-of-sample prediction capability of Boost-S constructed from a relatively small training dataset. To obtain the initial estimate of the covariance matrix $\hat{\bm{\Sigma}}_{\bm{\theta}}^{(0)}$, the FGLS is used to solve the linear model (16) in Algorithm 1. Before the FGLS is performed, one needs to first choose a parametric spatial covariance function $c(\cdot)$ of the process $\varepsilon$ in (16). Figure 5 shows both the empirical semivariance and the fitted semivariance using FGLS assuming the Gaussian covariance function. The sill, range and nugget effect can be clearly seen from Figure 5, and the Gaussian covariance function appears to be an appropriate choice. Figure 5: Exploratory analysis on the covariance structure: plot of the empirical and fitted semivariance using FGLS Algorithm 1 requires the tuning parameters $\lambda$ and $\gamma$ to be pre- specified. The choice of these two parameters affects both the depth and contribution of individual trees to the ensemble prediction, which in turn influences the total number of trees in the ensemble. A common strategy suggests that we explore the suitable values for $\lambda$ and $\gamma$ while leaving $K$ as the primary parameter (Hastie et al., 2009). Although it is theoretically possible to perform a grid search for the best combinations of $\lambda$ and $\gamma$ on a two-dimensional space, such an approach may not be practical nor necessary in practice when it is computationally intensive to run Boost-S on big datasets. Hence, we resort to a powerful tool in computer experiments—the space-filling designs (Joseph, 2016). The idea of space- filling designs is to have points everywhere in the experimental region with as few gaps as possible, which serves our purpose very well. Figure 6 shows the Maximum Projection Latin Hypertube Design (MaxProLHD, Joseph et al. (2015)) of 16 runs with different combinations of $\lambda$ and $\gamma$, where the experimental ranges for these two parameters are respectively $[0,0.1]$ and $[0,10]$. For each design, Figure 7 shows the box plot of the number of tree leaves per tree in an ensemble for each combination of $\lambda$ and $\gamma$. Since the key idea behind boosting trees is that each individual tree needs to be kept simple (Hastie et al., 2009), we identify that Designs #7, #8 and #9 provide the most suitable combinations of $\lambda$ and $\gamma$. From Figure 6, these three design points are adjacent to each other, indicating that the appropriate choices for $\lambda$ and $\gamma$ are approximately within $[0.025,0.075]$ and $[4,5.5]$. A refined search in a much smaller experimental region yields an appropriate combination of $\lambda=0.05$ and $\gamma=4.25$ (Design “$*$” in Figure 6). Design “$*$” is between Designs #7 and #8, and the 25th, 50th and 75th empirical quartiles of the number of tree leaves are 6, 8 and 10 as shown in Figure 7. Figure 6: Maximum projection design of 16 runs with different combinations of $\lambda$ and $\gamma$. Design “$*$” yields an appropriate combination such that $\lambda=0.05$ and $\gamma=4.25$. Figure 7: Boxplot of the number of tree leaves per tree in an ensemble for the 16 candidate designs. Design “$*$” indicates the chosen combination such that $\lambda=0.05$ and $\gamma=4.25$. After $\lambda$ and $\gamma$ have been appropriately chosen, 49 trees are constructed to form the ensemble predictor using Algorithm 1. Figure 8 (top panel) shows that the Mahalanobis distance (i.e., the objective function) decreases as more trees have been included into the ensemble, indicating that the algorithm is working as expected. Figure 8 (bottom panel) shows the number of tree leaves for individual trees in this ensemble. It is interesting to note that the algorithm no longer splits the (root) tree node after 37 trees have been grown. In addition, the outputs of these one-node trees are all zeros, indicating that all trees after tree 37 are completely redundant. This is precisely due to the regularization $\gamma T$ and $\frac{1}{2}\lambda\|\bm{w}\|^{2}$ in (6). The gain in $\ell$ is outweighted by the loss caused by $\Omega$ if one more tree is added. Figure 8: Top panel: the Mahalanobis distance (i.e., objective function) decreases as the number of trees grows; Bottom panel: the number of tree leaves of individual trees. Applying the constructed ensemble trees to the testing dataset, Figure 9 shows the out-of-sample Root-Mean-Square-Error (RMSE) and Mean-Gross-Error (MGE). We see that both performance metrics decrease as more trees are included and stabilize approximately after 30 to 40 trees have been grown. As discussed above, this is precisely due to the fact that all trees after #37 are redundant with zero output. Figure 10 shows the (out-of-sample) predictions against actual observations of the SUV level at differnt voxels. The figure shows that the proposed Boost-S accurately predicts the SUV levels given the covariate information. Figure 9: Out-of-sample model performance in terms of RMSE and MGE Figure 10: Out-of-sample predictions against actual observations Figure 11: MidPET images on the $x-y$ plane for different values of $z$. Rows 1 and 3: observed images; Row 2 and 4: resconstructed images. Furthermore, we reconstruct the Mid-RT SUV (3 weeks after treatment) by the proposed Boost-S, and compare the actual and predicted Mid-RT SUV using slice plots (Figure 11). Because a tumor is a 3D object with three coordinates, $x$, $y$, and $z$, the MidPET images are shown on the $x-y$ plane for given values of $z$. We see that the ensemble trees are capable of predicting the Mid-RT SUV for the entire tumor body. ### Comparison Study We compare the predictive capability of Boost-S to five other methods using the data collected from 25 patients. The five other methods included in the comparison study are: Random Forests (RF) (Breiman, 2001), Extreme Gradient Boosting Trees (XGBoost) (Chen and Guestrin, 2016), non-parametric cubic splines, universal kriging with a linear spatial trend, and multiple linear regression without considering the spatial correlation. For each patient, 15% of the observations are randomly chosen as the training dataset, and all 6 models are constructed to generate the out-of-sample predictions using the testing dataset. Such a low training-testing ratio is used to test the predictive capabilities of these methods. Tables 1, 2 and 3 respectively show the Mean Gross Error (MGE), Relative Error (RE) and Rooted Mean Squared Error (RMSE) of the out-of-sample predictions for 25 patients using 6 different methods. It is interesting to see that * • The proposed Boost-S yields the lowest MGE for 19 out of the 25 patients, while RF performs the best for the remaining 6 patients; * • The proposed Boost-S yields the lowest RE for 18 out of the 25 patients, while RF performs the best for the remaining 7 patients; * • The proposed Boost-S yields the lowest RMSE for 17 out of the 25 patients, while RF performs the best for the remaining 8 patients; Hence, we conclude that, the proposed Boost-S provides the best performance for majority of the patients in terms of all three performance measures, although no method uniformly outperforms others (which is realistic and expected given the well-known modeling power of RF and XGBoost). The reason why Boost-S outperforms other additive-tree-based methods, i.e., RF and XGBoost, is because of its capability of accounting for the spatial correlation among observations. The reason why Boost-S outperforms universal Kriging and multiple linear regression is due to the advantages of non- parametric tree-based method in capturing complex non-linear and interaction effects between features and responses. The observation demonstrates the effectiveness of Boost-S as a useful extension to existing ensemble tree-based methods by accounting for the spatial correlation among observations. Table 1: Mean Gross Error of the out-of-sample predictions for 25 patients using 6 different methods (note that: rows 1 to 25 respectively show the results corresponding to the 25 patients, while the last row shows the p-value of the one-side paired Wilcoxon test) Boost-S | RF | XGBoost | Cubic Splines | Universal Kriging | Linear Regression ---|---|---|---|---|--- 8.24 | 6.78 | 7.31 | 8.45 | 10.32 | 9.01 10.68 | 9.61 | 10.55 | 11.57 | 13.41 | 13.79 5.84 | 6.22 | 6.32 | 7.14 | 10.37 | 9.08 4.30 | 4.37 | 4.68 | 6.19 | 6.76 | 6.20 3.15 | 4.45 | 3.97 | 4.70 | 10.04 | 8.58 1.95 | 2.40 | 2.51 | 3.07 | 5.42 | 4.25 5.20 | 5.40 | 5.67 | 5.47 | 13.30 | 10.34 5.04 | 5.30 | 5.93 | 9.00 | 10.39 | 9.88 8.59 | 9.28 | 9.50 | 9.22 | 12.55 | 11.79 2.37 | 2.79 | 3.17 | 4.46 | 6.70 | 5.97 9.08 | 8.63 | 8.91 | 11.16 | 10.79 | 11.22 1.44 | 1.74 | 2.08 | 2.64 | 4.00 | 2.99 1.65 | 1.91 | 2.08 | 3.61 | 4.36 | 3.69 0.74 | 1.32 | 1.35 | 1.58 | 2.48 | 1.83 10.19 | 9.77 | 10.59 | 10.98 | 15.03 | 13.71 4.17 | 4.37 | 4.25 | 7.79 | 9.64 | 7.89 2.71 | 3.20 | 3.78 | 4.22 | 7.15 | 6.47 6.53 | 7.06 | 8.27 | 8.41 | 13.43 | 12.18 13.10 | 10.62 | 11.36 | 12.57 | 13.77 | 14.23 3.51 | 3.96 | 4.06 | 4.44 | 9.30 | 6.85 1.22 | 1.24 | 1.57 | 1.58 | 2.55 | 2.03 4.17 | 4.82 | 5.01 | 7.69 | 11.25 | 8.71 5.47 | 6.62 | 6.43 | 10.65 | 18.82 | 12.05 3.87 | 3.63 | 4.13 | 5.08 | 5.52 | 5.47 1.86 | 2.42 | 2.58 | 3.57 | 5.60 | 4.19 N.A. | 0.04 | 0.001 | $<10^{-6}$ | $<10^{-7}$ | $<10^{-7}$ Table 2: Relative Error (in %) of the out-of-sample predictions for 25 patients using 6 different methods (note that: rows 1 to 25 respectively show the results corresponding to the 25 patients, while the last row shows the p-value of the one-side paired Wilcoxon test) Boost-S | RF | XGBoost | Cubic Splines | Universal Kriging | Linear Regression ---|---|---|---|---|--- 9.67 | 8.01 | 8.58 | 9.91 | 12.29 | 10.68 56.62 | 51.01 | 55.37 | 60.03 | 60.47 | 67.27 11.04 | 12.52 | 11.93 | 13.72 | 18.40 | 17.01 5.03 | 5.34 | 5.53 | 7.60 | 8.36 | 7.61 7.30 | 9.41 | 8.26 | 10.77 | 20.25 | 19.07 4.42 | 5.73 | 5.65 | 7.29 | 12.52 | 9.72 8.92 | 9.41 | 9.52 | 9.72 | 20.53 | 18.06 23.25 | 22.32 | 25.52 | 43.34 | 43.51 | 43.78 13.62 | 15.45 | 15.18 | 14.45 | 20.04 | 18.91 2.28 | 2.74 | 3.07 | 4.26 | 6.48 | 5.77 11.44 | 10.78 | 10.94 | 14.42 | 13.41 | 14.39 1.09 | 1.32 | 1.55 | 1.99 | 3.06 | 2.25 3.66 | 4.25 | 4.55 | 8.00 | 9.70 | 8.20 1.49 | 2.75 | 2.72 | 3.17 | 4.70 | 3.59 21.40 | 19.97 | 21.21 | 23.26 | 32.02 | 29.69 8.36 | 8.67 | 8.13 | 15.27 | 20.18 | 15.49 3.90 | 4.68 | 5.34 | 6.05 | 9.93 | 9.51 7.46 | 8.79 | 9.97 | 9.99 | 17.54 | 15.31 21.70 | 17.32 | 18.19 | 20.08 | 21.71 | 23.53 4.02 | 4.58 | 4.60 | 5.05 | 10.15 | 7.95 1.91 | 1.94 | 2.43 | 2.53 | 4.08 | 3.24 6.34 | 7.18 | 7.61 | 11.55 | 17.54 | 13.37 8.10 | 9.55 | 8.73 | 16.69 | 25.45 | 18.96 5.24 | 5.01 | 5.63 | 7.10 | 7.85 | 7.61 5.07 | 6.59 | 6.86 | 9.72 | 14.92 | 11.26 N.A. | 0.09 | 0.004 | $<10^{-6}$ | $<10^{-7}$ | $<10^{-7}$ Table 3: Rooted Mean Squared Error of the out-of-sample predictions for 25 patients using 6 different methods (note that: rows 1 to 25 respectively show the results corresponding to the 25 patients, while the last row shows the p-value of the one-side paired Wilcoxon test) Boost-S | RF | XGBoost | Cubic Splines | Universal Kriging | Linear Regression ---|---|---|---|---|--- 11.05 | 9.01 | 9.66 | 10.89 | 13.00 | 11.66 16.46 | 14.77 | 16.22 | 16.14 | 19.66 | 19.42 7.73 | 7.83 | 8.10 | 9.17 | 12.89 | 11.61 5.74 | 5.51 | 5.89 | 7.73 | 8.28 | 7.74 4.13 | 5.97 | 5.41 | 6.00 | 12.83 | 10.64 2.68 | 3.11 | 3.24 | 3.89 | 6.68 | 5.36 7.03 | 6.68 | 7.13 | 6.98 | 17.10 | 12.64 7.62 | 7.88 | 8.58 | 11.50 | 14.09 | 13.07 11.15 | 11.69 | 12.02 | 11.80 | 15.57 | 15.04 3.21 | 3.80 | 4.32 | 5.97 | 8.28 | 7.66 12.52 | 12.06 | 12.41 | 14.47 | 14.68 | 14.87 1.89 | 2.20 | 2.64 | 3.64 | 4.75 | 4.02 2.20 | 2.46 | 2.66 | 4.56 | 5.10 | 4.68 1.04 | 1.69 | 1.72 | 1.95 | 3.21 | 2.37 14.17 | 13.19 | 14.17 | 14.29 | 18.66 | 17.32 5.77 | 5.94 | 5.71 | 10.18 | 12.27 | 10.31 3.48 | 4.15 | 4.73 | 5.30 | 8.99 | 8.28 8.72 | 8.98 | 10.50 | 10.83 | 16.69 | 15.26 19.26 | 15.23 | 16.30 | 17.69 | 20.01 | 19.80 5.02 | 5.21 | 5.29 | 5.94 | 11.18 | 8.33 1.72 | 1.82 | 2.22 | 2.09 | 3.29 | 2.72 5.52 | 6.61 | 6.59 | 9.75 | 13.63 | 11.06 7.36 | 8.73 | 8.68 | 13.29 | 23.28 | 15.15 5.10 | 4.78 | 5.41 | 6.63 | 7.07 | 7.07 2.55 | 3.16 | 3.34 | 4.67 | 7.14 | 5.39 N.A. | 0.19 | 0.003 | $<10^{-4}$ | $<10^{-7}$ | $<10^{-7}$ Figures 14, 12 and 13 respectively show the MGE, RE and RMSE of the out-of- sample predictions for the 25 patients. Such visualizations provide a more holistic perspective on the performance of the six candidate methods. Figure 12: Mean Gross Error (MGE) of the out-of-sample predictions for 25 patients using 6 different methods Figure 13: Relative Error (RE) of the out-of-sample predictions for 25 patients using 6 different methods Figure 14: Rooted Mean Squared Error (RMSE) of the out-of-sample predictions for 25 patients using 6 different methods ## Conclusion This paper proposed a new gradient Boosted Trees algorithm for Spatial Data with covariate information (Boost-S). It has been shown that the Boost-S successfully integrates spatial correlation into the classical framework of gradient boosted trees. A computationally-efficient algorithm as well as the technical details have been presented. The Boost-S algorithm grows individual trees by solving a regularized optimization problem, where the objective function involves two penalty terms on tree complexity and takes into account the underlying spatial correlation. The advantages of the proposed Boost-S, over five other commonly used approaches, have been demonstrated using real datasets involving the spatially-correlated FDG-PET imaging data collected during cancer chemoradiotherapy. ## Acknowledgment This investigation was supported in part by National Institutes of Health grant R01CA204301. ## References * Banerjee (2017) Banerjee, S. (2017), “High-Dimensional Bayesian Geostatistics,” Bayesian Analysis, 12, 583–614. * Banerjee et al. (2004) Banerjee, S., Carlin, B. P., and Gelfand, A. E. (2004), Hierarchical Modeling and Analysis for Spatial Data, 2nd ed., Boca Raton, Florida: CRC Press. * Banerjee et al. (2008) Banerjee, S., Gelfand, A. E., Finley, A. O., and Sang, H. (2008), “Gaussian Predictive Proess Models for Large Spatial Data Sets,” Journal of the Royal Statistical Society: Series B, 70, 825–848. * Berliner (2003) Berliner, L. M. (2003), “Physical-Statistical Modeling in Geophysics,” Journal of Geophysical Research-Atmospheres, 108, 3–10. * Besag (1974) Besag, J. E. (1974), “Spatial Interaction and the Statistical Analysis of Lattice Systems,” Journal of the Royal Statistical Society, B, 36, 192–225. * Bowen et al. (2019) Bowen, S. R., Hippe, D. S., Chaovalitwongse, W. A., Duan, C., Thammasorn, P., Liu, X., Miyaoka, R. S., Vesselle, H. J., Kinahan, P. E., Rengan, R., et al. (2019), “Voxel Forecast for Precision Oncology: predicting spatially variant and multiscale cancer therapy response on longitudinal quantitative molecular imaging,” Clinical Cancer Research, 25, 5027–5037. * Breiman (2001) Breiman, L. (2001), “Random forests,” Machine learning, 45, 5–32. * Brown et al. (2000) Brown, P. E., Karesen, K. F., Roberts, G. O., and Tonellato, S. (2000), “Blur-Generated Non-Separable Space-Time Models,” Journal of the Royal Statistical Society: Series B, 62, 847–860. * Carlin and Banerjee (2003) Carlin, B. P. and Banerjee, S. (2003), Hierarchicla Multivariate CAR Models for SpatioTemporally Correlated Survival Data (with discussion), Oxford: Oxford University Press. * Chen and Guestrin (2016) Chen, T. and Guestrin, C. (2016), “XGBoost: A Scalable Tree Boosting System,” in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. * Chipman et al. (2010) Chipman, H. A., George, E. I., and McCulloch, R. E. (2010), “Bart: Bayesian Additive Regression Trees,” The Annals of Applied Statistics, 4, 266–298. * Cressie and Huang (1999) Cressie, N. and Huang, H. C. (1999), “Classes of Nonseparable, Spatio-Temporal Stationary Covariance Functions,” Journal of the American Statistical Association, 94, 1330–1340. * Cressie and Johannesson (2002) Cressie, N. and Johannesson, G. (2002), “Fixed Rank Kriging for Very Large Spatial Data Sets,” Journal of the Royal Statistical Society: Series B, 70, 209–226. * Cressie and Wikle (2011a) Cressie, N. and Wikle, C. (2011a), Statistics for Spatio-Temporal Data, Hoboken, New Jersey: John Wiley & Sons. * Cressie and Wikle (2011b) Cressie, N. and Wikle, C. K. (2011b), Statistics for spatio-temporal data, John Wiley & Sons. * Datta et al. (2016) Datta, A., Banerjee, S., Finley, A. O., and Gelfand, A. E. (2016), “Hierarchical Nearest-Neighbor Gaussian Process Models for Large Geostatistical Datasets,” Journal of the American Statistical Association, 111, 800–812. * Ezzat et al. (2019) Ezzat, A., Jun, M., and Ding, Y. (2019), “Spatio-temporal short-term wind forecast: A calibrated regime-switching method,” The Annals of Applied Statistics, 1484–1510. * Fang et al. (2019) Fang, X., Paynabar, K., and Gebraeel, N. (2019), “Image-Based Prognostics Using Penalized Tensor Regression,” Technometrics, 61, 369–384. * Fuentes (2007) Fuentes, M. (2007), “Approximate Likelihood for Large Irregularly Spaced Spatial Data,” Journal of the American Statistical Association, 102, 321–331. * Fuentes et al. (2005) Fuentes, M., Chen, L., Davis, J. M., and Lackmann, G. M. (2005), “Modeling and Predicting Complex Space-Time Structures and Patterns of Coastal Wind Fields,” Environmetrics, 16, 449–464. * Ghosh et al. (2010) Ghosh, S. K., Bhave, P. E., Davis, J. M., and Lee, H. (2010), “Spatio-Temporal Analysis of Total Nitrate Concentrations using Dynamic Statistical Models,” Journal of the American Statistical Association, 105, 538–551. * Gneiting (2002) Gneiting, T. (2002), “Nonseparable, Stationary Covariance Functions for Space-Time Data,” Journal of the American Statistical Association, 97, 590–600. * Gneiting et al. (2006) Gneiting, T., Genton, M. G., and Guttorp, P. (2006), “Geostatistical space-time models, stationarity, separability, and full symmetry.” in Statistical Methods for Spatio-Temporal Systems, eds. Finkenstadt, B., Held, L., and Isham, V., Boca Raton: Chapman & Hall, pp. 151–175. * Guinness and Fuentes (2015) Guinness, J. and Fuentes, M. (2015), “Likelihood Approximations for Big Nonstationary Spatial-Temporal Lattice Data,” Statistica Sinica. * Guinness and Stein (2013) Guinness, J. and Stein, M. (2013), “Interpolation of Nonstationary High Frequency Spatial-Temporal Temperature Data,” Annals of Applied Statistics, 7, 1684––1708. * Hastie et al. (2009) Hastie, T., Tibshirani, R., and Friedman, J. (2009), The Elements of Statistical Learning, 2nd Edition, New York: Springer. * Higdon (1998) Higdon, D. (1998), “A Process-Convolution Approach to Modeling Temperatures in the North Atlantic Ocean,” Environmental Ecology Statistics, 5, 173–190. * Hooten and Wikle (2008) Hooten, M. B. and Wikle, C. K. (2008), “A hierarchical Bayesian non-linear spatio-temporal model for the spread of invasive species with appliation to the Eurasian Collared-Dove,” Environmental and Ecological Statistics, 15, 59–70. * Joseph (2016) Joseph, V. R. (2016), “Space-filling designs for computer experiments: A review,” Quality Engineering, 28, 28–35. * Joseph et al. (2015) Joseph, V. R., Gul, E., and Ba, S. (2015), “Maximum Projection Designs for Computer Experiments,” Biometrika, 102, 371–380. * Katzfuss (2017) Katzfuss, M. (2017), “A multi-resolution approximation for massive spatial datasets,” Journal of the American Statistical Association, 112, 201–214. * Katzfuss et al. (2020) Katzfuss, M., Stroud, J. R., and Wikle, C. K. (2020), “Ensemble Kalman methods for high-dimensional hierarchical dynamic space-time models,” Journal of the American Statistical Association, 115, 866–885. * Krainski et al. (2019) Krainski, E. T., Gomez-Rubio, V., Bakka, H., Lenzi, A., Castro-Camilo, D., Simpson, D., Lindgren, F., and Rue, H. (2019), Advanced Spatial Modeling with Stochastic Partial Differential Equations Using R and INLA, Boca Raton: Chapman and Hall/CRC. * Lenzi et al. (2019) Lenzi, A., Castruccio, S., Rue, H., and Genton, M. G. (2019), “Improving Bayesian Local Spatial Models in Large Data Sets,” arXiv:1907.06932. * Lindgren and Rue (2011) Lindgren, F. and Rue, H. (2011), “An Explicit Link between Gaussian Fields and Gaussian Markov Random Fields: the Stochastic Partial Differntial Equation Approach,” Journal of the Royal Statistical Society: Series B, 73, 423–498. * Liu et al. (2018a) Liu, X., Gopal, V., and Kalagnanam, J. (2018a), “A Spatio-Temporal Modeling Framework for Weather Radar Image Data in Tropical Southeast Asia,” The Annals of Applied Statistics, 12, 378–407. * Liu et al. (2018b) — (2018b), “A Spatio-Temporal Modeling Framework for Weather Radar Image Data in Tropical Southeast Asia,” The Annals of Applied Statistics, 12, 378–407. * Liu and Pan (2020) Liu, X. and Pan, R. (2020), “Analysis of Large Heterogeneous Repairable System Reliability Data with Static System Attributes and Dynamic Sensor Measurement in Big Data Environment,” Technometrics, 62, 206–222. * Liu et al. (2016) Liu, X., Yeo, K. M., Hwang, Y. D., Singh, J., and Kalagnanam, J. (2016), “A Statistical Modeling Approach for Air Quality Data Based on Physical Dispersion Processes and Its Application to Ozone Modeling,” The Annals of Applied Statistics, 10, 756–785. * Liu et al. (2018c) Liu, X., Yeo, K. M., and Kalagnanam, J. (2018c), “A Statistical Modeling Approach for Spatio-Temporal Degradation Data,” Journal of Quality Technology, 50, 166–182. * Liu et al. (2020) Liu, X., Yeo, K. M., and Lu, S. Y. (2020), “Statistical Modeling for Spatio-Temporal Data from Stochastic Convection-Diffusion Processes,” Journal of the American Statistical Association, to appear, arXiv:1910.10375. * Mondal and Wang (2019) Mondal, D. and Wang, C. (2019), “A matrix-free method for spatial-temporal Gaussian state-space models,” Statstica Sinica (to appear), 29, 2205–2227. * Nychka and Wikle (2002) Nychka, D. and Wikle, C. Royle, J. A. (2002), “Multiresolution Models for Nonstationary Spatial Covariance Functions,” Statistical Modeling, 2, 315–331. * R-INLA (2019) R-INLA (2019), The R-INLA Project, http://www.r-inla.org/. * Reich et al. (2011) Reich, B. J., Eidsvik, J. Guindani, M., Nail, A. J., and Schmidt, A. M. (2011), “A Class of Covariate-Dependent Spatiotemporal Covariance Functions for the Analysis of Daily Ozone Concentration,” The Annals of Applied Statistics, 5, 2425–2447. * Rue et al. (2009) Rue, H., Martino, S., and Chopin, N. (2009), “Approximate Bayesian inference for latent Gaussian models by using integrated nested Laplace approximations,” Journal of the Royal Statistical Society, B. * Schabenberger and Gotway (2005) Schabenberger, O. and Gotway, C. A. (2005), Statistical Methods for Spatial Data Analysis, Boca Raton, Florida: Chapman & Hall/CRC. * Schapire (1999) Schapire, R. E. (1999), “A Brief Introduction to Boosting,” in Proceedings of the Sixteenth International Joint Conference on Artificial Intelligence. * Sigrist (2020) Sigrist, F. (2020), “Gaussian Process Boosting,” arXiv:2004.02653v2. * Sigrist et al. (2015) Sigrist, F., Kunsch, H. R., and Stahel, W. A. (2015), “Stochastic Partial Differential Equation based Modelling of Large Space-Time Data Sets,” Journal of the Royal Statistical Society: Series B, 77, 3–33. * Stein et al. (2004) Stein, M. L., Chi, Z., and Welty, L. J. (2004), “Approximating Likelihoods for Large Spatial Data Sets,” Journal of the Royal Statistical Society: Series B, 66, 275–296. * Stroud et al. (2001) Stroud, J. R., Muller, P., and Sanso, B. (2001), “Dynamic Models for Spatiotemporal Data,” Journal of the Royal Statistical Society: Series B, 63, 673–689. * Stroud et al. (2010) Stroud, J. R., Stein, M. L., L. B. M., Schwab, D. J., and Beletsky, D. (2010), “An Ensemble Kalman Filter and Smoother for Satellite Data Assimilation,” Journal of American Statistical Association, 105, 978–990. * Wang et al. (2016) Wang, K., Jiang, W., and Li, B. (2016), “A Spatial Variable Selection Method for Monitoring Product Surface,” International Journal of Production Research, 54, 4161–4181. * Wikle and Cressie (1999) Wikle, C. K. and Cressie, N. (1999), “A Dimension-Reduced Approach to Space-Time Kalman Filtering,” Biometrika, 86, 815–829. * Yan et al. (2019) Yan, H., Zhao, X., Hu, Z., and Du, D. (2019), “Physics-based Deep Spatio-temporal Metamodeling for Cardiac Electrical Conduction Simulation,” . * Yao et al. (2017) Yao, B., Zhu, R., and Yang, H. (2017), “Characterizing the location and extent of myocardial infarctions with inverse ECG modeling and spatiotemporal regularization,” IEEE Journal of Biomedical and Health Informatics, 22, 1445–1455. * Yue et al. (2020) Yue, X., Wen, Y., Hunt, J. H., and Shi, J. (2020), “Active Learning for Gaussian Process considering Uncertainties, with an Application to Automatic Shape Control of Composite Fuselage,” arxiv: 2004.10931. * Zang and Qiu (2019) Zang, Y. and Qiu, P. (2019), “Phase I Monitoring of Spatial Surface Data from 3D Printing,” Technometrics, 60, 169–180.
# On Small-World Networks: Survey and Properties Analysis Alaa Eddin Alchalabi _Graduate School of Electronics and Computer Engineering_ Istanbul Sehir University, Istanbul, Turkey <EMAIL_ADDRESS> ###### Abstract Complex networks has been a hot topic of research over the past several years over-crossing many disciplines, starting from mathematics and computer science and ending by the social and biological sciences. Random graphs were studied to observe the qualitative features they have in common in planetary-scale data sets which helps us to project the insights proven to real-world networks. In this paper, We survey the particular case of small-world phenomena and decentralized search algorithms. We start by explaining the first empirical study for the “six degrees of separation” phenomenon in social networks; then we review some of the probabilistic network models based on this work, elaborating how these models tried to explain the phenomenon’s properties, and lastly, we review few of the recent empirical studies empowered by these models. Finally, some future works are proposed in this area of research. ###### Index Terms: Small-World, Complex Networks, Lattices and Random Graphs, Search Algorithms. ## I Introduction Recently, the study of complex networks has emerged in various range of disciplines and research areas. The World Wide Web has revolutionized the way we deal with everything one deals in daily life. Computer scientists were curious to find a way to handle the wheel of controlling the complexity and the enormous growth of the Internet. Social networks’ data scale is unpredictably uncontrollable by social scientists. The biological interactions in cell’s metabolism are expected to define its pathways and could provide insights to biologists [13]. The urge a new born science is needed in order to be able to manipulate networks before networks manipulate our needs [8]. The study of complex networks evolved since the study of randomly generated graphs by Erdos and Renyi [4], and the appearance of a large-scale network data had leashed tremendous work in multi-disciplinary areas including the real and the virtual world [13]. The efforts were put to describe the properties of random graphs in large networks which raised more and more technical questions to be answered. To mimic real-networks, a randomly produced stylized network model is adopted in order to generalize the resulting conclusions and properties onto real-networks. Simple models fails to capture the complexity of a realistic network’s structure and features offering a strong mathematical basis which futures investigations can be build upon. In the next sections of this paper, we survey the “small-world phenomenon” and few related problems. We start with the famous psychologist Stanley Milgram’s social experiment, that captures the main aspects of the phenomenon [11], [14]; we review few of the models based on random graphs that tries to explain the problem [7], [9], [12], [15], [16]; and then we mention recent work that has applied the traditional insights of these models on large data sets extracted from famous web applications [2], [10]. Lastly, some suggested further extensions to small-world networks are discussed, along with some future works followed by their relevance to this field. ## II small-world phenomenon The small-world phenomenon has recently been the hot topic of both theoretical and practical research, and it has been given huge attention from most, if not all, multi-disciplinary researchers. The term “small-world”, linked by all means to the “short chains of acquaintances”, or the “six degrees or separation” [5][6][16], refers to the human social network’s graph; where nodes replaces people, and edges between two nodes mimic if the two corresponding persons know each other on a first-name basis [8]. The graph is described to be a “small-world” because of the fact that any two random pairs of nodes are separated by relatively a small number of nodes, generally less than 6. Although the first-name basis rule is a bit naive for an edge definition, the resulted graph behaves as a real-world network. Small-world networks are of great importance because of adoption to the limitations of either of the end extreme networks types; random networks and regular lattices. Small-world networks proved their ability to be used as frameworks for the study of interaction networks of complex systems [1]. The most important key of the small-world study is to prove the hypothesis that assumes the qualitatively shared structure among a variety of networks across variant fields. A common property arises in large networks which is the existence of short paths between most of the nodes pairs although nodes in network have a high degree of clustering. Nodes can also be reached and navigated with no need of universal understanding of the whole network. Such properties contributed in describing large-scale social networks behavior, and additionally, they gave important insights to create applications to the internal design of decentralized peer-to-peer systems. ### II-A Milgram’s Experiment Stanley Milgram, the famous social psychologist, made an experiment to measure people connectivity in the USA in the 1960s and to test the small-world property [11][14]. The experiment questions the probability of an arbitrarily two selected individuals from a large data set to know each other in person. A target person was selected in the state of Massachusetts who was a Boston stockbroker, and 296 arbitrarily selected individuals as “starting persons” from Nebraska and Boston were asked to generate acquaintance chains to the stockbroker in Boston. The selected group were given a document of the described study with the target’s name. The senders were asked to choose a recipient which they think that he/she will contribute to carry the message to the target person in the shortest way possible. The concept of “roster” was introduced to prevent the message goes back to a previous sender “loop” and to track the number of node the message reached. The results of the experiment were quite astonishing. Sixty-four chains made their way to the target person. The mean number of the intermediaries was 5.2 with median of 6. Boston starting chains exhibits shorter range chains than Nebraska’s chains. Additional experiments by Korte and Milgram proved that these numbers are quite stable [14]. Some comments on Milgram’s experiments exhibit the inability of this model to be generalized to larger networks. Varying the information about the target person might affect the decisions taken by senders, and here we meant psychological and sociological factors take place. ## III Small-world based empirical models ### III-A Watts and Strogatz’s Model Watts and Strogatz came up with a model that aims to explain the small-world property. After Bollobas and de la Vega [3] introduced the theorem which proves the logarithmic property in the path length with respect to the number of nodes _O(log n)_ in small-world networks, Watts and Strogatz felt that there was something missing in the theorem. The model proposed considered small-world networks to be highly-structured with relatively a few number of random links added within. Long-range connections in this model plays a crucial rule in creating short paths through all nodes [15]. The model adopts the idea of rewiring edges between nodes with a certain probability to generate a graph. The probability allows the change between regular networks (p=0) and random networks (p=1). The model starts by generating a ring of n connected nodes (average degree of k). Then, the rewiring of each edge with the probability p and the landing node is also chosen randomly. The clustering coefficient _$(C_{p})$_ is a measure which reflects the fraction of the connected neighbours to a node in a graph compared to all possible connections of the neighboured averaged over all nodes [15]. The results derived for a regular graph (p = 0) a highly clustered (_$C\sim$_ 3/4) and path length _$L\sim n/2k$_ where _$n >k>ln(n)>1$_ should be chosen. For random graphs (p = 1) the resulted network is poorly with a low clustering coefficient (_$C\sim k/n$_) and a small path length _$L\sim ln(n)/ln(k)$_. Their research also included three empirical real-world examples of small- world models, and their main finding was that the average path length of the chosen example was slightly higher than the random model, while the clustering coefficient had clearly a much higher value than the random model. Using their results, they reasoned how infectious disease spread rapidly in small-world societies. Some drawbacks of Watts and Strgatz’s model is that it cannot be generalized to all small-world models. Some extended works by other scientists tried to fill in gaps. ### III-B Classes of Small-World Networks Due to the limited vision of the Watts and Strogatz model, new explanation was needed. Trying to look at the dilemma from another prospective, Amaral et al. tried to classify small-world networks to three classes reporting an empirical study of real-world networks [1]. The study covers mainly the statistical properties of real-world networks, and it was enough to prove the existence of three classes of small-world networks: Scale-free, broad-scale, and single scale[1]. #### III-B1 Scale-free The networks which is characterized by a vertex connectivity distribution which decays as a power law. #### III-B2 Broad-scale The networks characterized by a connectivity distribution that has a power-low region and followed by a sharp cutt-off. #### III-B3 Single-scale The networks characterized by a connectivity distribution with a fast decaying tail. The research also gave an answer to why such taxonomy exist, and they reasoned that by mentioning two types of factors. The first factor was the aging of the vertices, which in time old nodes will stop being as effective in the network, and an example of that can be the actors network. The second type of factor was the cost of adding new links to the vertices which limited by the vertex capacity. An example of this can be the airports map where placing too many vertices are pricey and not practical. ### III-C Kleinbergs’s Algorithmic Prospective Kleinberg’s way of explaining small-world properties was a bit close to Watts and Strogatz but with slight differences [7]. Kleinberg used a _n x n_ grid of nodes to represent the network, and to add the small-world flavour, a number of long-range connection edges were added and not rewired. After adding the edges, the probability of connecting two random vertices (v,w) is proportional to _$1/d(v,w)^{q}$_ where q is the clustering coefficient [9]. Kleinberg came up with theorems to quantify the decentralized algorithms’ delivery time which generalized the results in [3] by Bollobás and de la Vega of the logarithmic behavior of short pathways in networks. He proved that the time needed is not always logarithmic but it depends on other parameters. A new parameter was introduced ($\alpha$ >=0) that controls the long-range connections. Interestingly, the delivery time varies depends on $\alpha$ as follows: #### III-C1 For 0 <$\alpha$ <2 the delivery time of any decentralized algorithm in the grid-based model is $\Omega$ ($n^{(2-a)/3}$). #### III-C2 For $\alpha$ = 2 the delivery time of any decentralized algorithm in the grid-based model is O($log^{2}n$). #### III-C3 For $\alpha$ >2 the delivery time of any decentralized algorithm in the grid-based model is $\Omega$ ($n^{(a-2)/(a-1)}$) [7]. ## IV Recent real-world Empirical experiments ### IV-A Dodds, Muhammad, and Watts Experiment Dodds et al. tried to mimic Milgram’s experiment with the electronic messaging systems. Around 60,000 randomly selected email users attempted to reach 18 target persons in 13 different countries. The findings were quite unexpected. Successful social chains passed through intermediate to weak strength ties [12]. This finding proves that the highly connected hubs’ effect is negligible. The attrition of message chains showed that messages could make it to the target in a median of five to seven. The cool fact about attrition rate was the constancy of its value for a certain period of time. The 384 completed chains (out of 24,163) had an average chain length of 4.05. This number was considered misleading by the authors, which made them evaluate the experiment using new metrics. The general results showed that the the network structure alone is not enough to interpret the network. Actions and perceptions of the individuals are big contributors. ### IV-B Leskovec et al. on a Large Instant-Messaging Network Leskovec et al. presented a study in 2003 which captured a month of communication activities within the Microsoft Messenger instant-messaging system [10]. The data set contained about 30 million conversations between 240 million people, and a graph was constructed containing 180 million nodes and 1.3 billion undirected edges. The network represents accounts that were active during June 2006. The resulted average path length among users was 6.6 with 6 being the median. The results showed that users with similar age, language, and location tend to communicate with each other. Users with different genders tend to communicate more and also for longer conversations [10]. Conversations tends to decrease with the increase in the distance. However, communication chains through relatively long distances tend to carry longer conversations [10]. ### IV-C Bakhshandeh’s Degrees of Separation in Twitter Bakhshandeh et al. did an interesting analysis to identify the degree separation between two Twitter users. They used a new search technique to provide near-optimal solutions more optimized than greedy approaches [2]. The average separation path length was 3.43 between any two random Twitter users which required 67 requests from Twitter, while the near-optimal was 3.88 using only 13.3 requests on average. Surprisingly, Twitter’s 3.43 degree of separation is small and the reason they have claimed was the indicative of changing social norms in the modern connected world. ## V Further Extensions and future works There is no doubt that small-world networks are still and will still be a hot research topic due to its nature. In this section, we would like to propose some ideas for future extensions which might propose solutions for unanswered or vaguely answered questions. Introducing machine learning techniques to small-world networks, in my opinion, is a good idea. Constructing networks should be smart enough in order to be controllable not only interpretable. Small-world networks could be build to mimic the brain neural-map which might give us more insight on how the human brain works. ML techniques can be also used to conserve the “six degrees of separation rule” or even to break it which completely depends on the application. Introducing local reference nodes in such networks could be a new idea to be implemented. Reference nodes could have some regional knowledge about the surrounding nodes. They can control the “hubs” and determine how new links can be distributed among reference nodes. We can think of routers as examples. The uniqueness of the node is somehow unrealistic for some applications, and that shows the urge of introducing a new concept. ## VI Conclusion At this paper, we discussed the famous phenomenon of small-world networks and its importance in various areas. Few of the small-world driven models were surveyed. Then recent real-world experiments in the context of complex networks were mentioned. Later, further extensions and future works were proposed. In the future, we will try to implement that the suggested ideas practically on a given data set. By taking into account their bros and cons, the ideas will be later evaluated against the other state-of-art implementations. ## References * [1] Amaral, Luıs A. Nunes, et al. ”Classes of small-world networks.” Proceedings of the national academy of sciences 97.21 (2000): 11149-11152. * [2] Bakhshandeh, Reza, et al. ”Degrees of separation in social networks.” Fourth Annual Symposium on Combinatorial Search. 2011. * [3] Bollobas, B., de la Vega,W. F., The diameter of random regular graphs. (1982), 125–134. * [4] Erdos, P., and Renyi, A., On the Evolution of Random Graphs. Magyar Tud. Akad. Mat. Kutató Int. Közl. 5 (1960), 17–61. * [5] Fass, Craig, Mike Ginelli, and Brian Turtle. Six Degrees of Kevin Bacon. Plume Books, 1996. * [6] Guare, John. Six degrees of separation: A play. Vintage, 1990. * [7] J. Kleinberg. The small-world phenomenon: An algorithmic perspective. Proc. 32nd ACM Symposium on Theory of Computing, 2000. * [8] J. Kleinberg. Complex Networks and Decentralized Search Algorithms. Proceedings of the International Congress of Mathematicians (ICM), 2006. * [9] Kleinberg, J. Navigation in a Small World. Nature 406(2000), 845. * [10] Leskovec, Jure, and Eric Horvitz. ”Planetary-scale views on a large instant-messaging network.” Proceedings of the 17th international conference on World Wide Web. ACM, 2008. * [11] Milgram, Stanley. ”The small world problem.” Psychology today 2.1 (1967): 60-67. * [12] Peter Sheridan Dodds, Roby Muhamad, Duncan J. Watts. An Experimental Study of Search in Global Social Networks. Science 301(2003), 827. * [13] Strogatz, Steven H. ”Exploring complex networks.” Nature 410.6825 (2001): 268-276. * [14] Travers, Jeffrey, and Stanley Milgram. ”An experimental study of the small world problem.” Sociometry (1969): 425-443. * [15] Watts, D. J. and S. H. Strogatz. Collective dynamics of ’small-world’ networks. Nature 393:440-42(1998). * [16] Watts, Duncan J. Six degrees: The science of a connected age. WW Norton and Company, 2004. | John Doe Hello, here is some text without a meaning. This text should show what a printed text will look like at this place. If you read this text, you will get no information. Really? Is there no information? Is there a difference between this text and some nonsense like “Huardest gefburn”? Kjift – not at all! A blind text like this gives you information about the selected font, how the letters are written and an impression of the look. This text should contain all letters of the alphabet and it should be written in of the original language. There is no need for special content, but the length of words should match the language. ---|---
# Hidden-charm pentaquarks with triple strangeness due to the $\Omega_{c}^{(*)}\bar{D}_{s}^{(*)}$ interactions Fu-Lai Wang1,2<EMAIL_ADDRESS>Xin-Dian Yang1,2<EMAIL_ADDRESS>Rui Chen4,5 chen<EMAIL_ADDRESS>Xiang Liu1,2,3111Corresponding author <EMAIL_ADDRESS>1School of Physical Science and Technology, Lanzhou University, Lanzhou 730000, China 2Research Center for Hadron and CSR Physics, Lanzhou University and Institute of Modern Physics of CAS, Lanzhou 730000, China 3Lanzhou Center for Theoretical Physics, Key Laboratory of Theoretical Physics of Gansu Province, and Frontiers Science Center for Rare Isotopes, Lanzhou University, Lanzhou 730000, China 4Center of High Energy Physics, Peking University, Beijing 100871, China 5School of Physics and State Key Laboratory of Nuclear Physics and Technology, Peking University, Beijing 100871, China ###### Abstract Motivated by the successful interpretation of these observed $P_{c}$ and $P_{cs}$ states under the meson-baryon molecular picture, we systematically investigate the possible hidden-charm molecular pentaquark states with triple strangeness which is due to the $\Omega_{c}^{(*)}\bar{D}_{s}^{(*)}$ interactions. We perform a dynamical calculation of the possible hidden-charm molecular pentaquarks with triple strangeness by the one-boson-exchange model, where the $S$-$D$ wave mixing effect and the coupled channel effect are taken into account in our calculation. Our results suggest that the $\Omega_{c}\bar{D}_{s}^{*}$ state with $J^{P}={3}/{2}^{-}$ and the $\Omega_{c}^{*}\bar{D}_{s}^{*}$ state with $J^{P}={5}/{2}^{-}$ can be recommended as the candidates of the hidden-charm molecular pentaquark with triple strangeness. Furthermore, we discuss the two-body hidden-charm strong decay behaviors of these possible hidden-charm molecular pentaquarks with triple strangeness by adopting the quark-interchange model. These predictions are expected to be tested at the LHCb, which can be as a potential research issue with more accumulated experimental data in near future. ## I Introduction As is well known, the study of the matter spectrum is an important way to explore the relevant matter structures and the involved interaction properties. In the hadron physics, since the discovery of the $X(3872)$ by the Belle Collaboration Choi:2003ue , a series of exotic states has been observed benefiting from the accumulation of more and more experimental data with high precision, and the exotic hadrons have stimulated extensive studies in the past two decades (see the review articles Chen:2016qju ; Liu:2019zoy ; Olsen:2017bmm ; Guo:2017jvc ; Liu:2013waa ; Hosaka:2016pey ; Brambilla:2019esw for learning the relevant processes). Exploring these exotic hadronic states not only gives new insights for revealing the hadron structures, but also provides useful hints to deepening our understanding of the nonperturbative behavior of the quantum chromodynamics (QCD) in the low energy regions. In fact, investigating the pentaquark states has been a long history, which can be tracked back to the birth of the quark model GellMann:1964nj ; Zweig:1981pd . Among exotic hadronic states, the hidden-charm molecular pentaquarks have attracted much attention as early as 2010 Li:2014gra ; Karliner:2015ina ; Wu:2010jy ; Wang:2011rga ; Yang:2011wz ; Wu:2012md ; Chen:2015loa and become a hot topic with the discovery of the $P_{c}(4380)$ and $P_{c}(4450)$ in the $\Lambda_{b}\to J/\psi pK$ process by the LHCb Collaboration Aaij:2015tga . In 2019, there was a new progress about the observation of three narrow structures [$P_{c}(4312)$, $P_{c}(4440)$, and $P_{c}(4457)$] by revisiting the process $\Lambda_{b}\to J/\psi pK$ based on more collected data Aaij:2019vzc , and they are just below the corresponding thresholds of the $S$-wave charmed baryon and $S$-wave anticharmed meson. This provides strong evidence to support the existence of the hidden-charm meson- baryon molecular states. More recently, the LHCb Collaboration reported a possible hidden-charm pentaquark with strangeness $P_{cs}(4459)$ Aaij:2020gdg , and this structure can be assigned as the $\Xi_{c}\bar{D}^{*}$ molecular state Chen:2016ryt ; Wu:2010vk ; Hofmann:2005sw ; Anisovich:2015zqa ; Wang:2015wsa ; Feijoo:2015kts ; Lu:2016roh ; Xiao:2019gjd ; Shen:2020gpw ; Chen:2015sxa ; Zhang:2020cdi ; Wang:2019nvm ; Chen:2020uif ; Peng:2020hql ; 1830432 ; 1830426 ; Liu:2020hcv ; 1839195 . Facing the present status of exploring the hidden-charm molecular pentaquarks Chen:2016qju ; Liu:2019zoy ; Olsen:2017bmm ; Guo:2017jvc , we naturally propose a meaningful question: why are we interested in the hidden-charm molecular pentaquark states? The hidden-charm pentaquark states are relatively easy to produce via the bottom baryon weak decays in the experimental facilities Aaij:2019vzc ; Aaij:2020gdg , and the hidden-charm quantum number is a crucial condition for the existence of the hadronic molecules Li:2014gra ; Karliner:2015ina . In addition, it is worth indicating that the heavy hadrons are more likely to generate the bound states due to the relatively small kinetic terms, and the interactions between the charmed baryon and the anticharmed meson may be mediated by exchanging a series of allowed light mesons Chen:2016qju ; Liu:2019zoy . Indeed, these announced hidden-charm pentaquark states have a ($c\bar{c}$) pair Chen:2016qju ; Liu:2019zoy ; Olsen:2017bmm ; Guo:2017jvc . Based on the present research progress on the hidden-charm pentaquarks Chen:2016qju ; Liu:2019zoy ; Olsen:2017bmm ; Guo:2017jvc , the theorists should pay more attention to making the reliable prediction of various types of the hidden-charm molecular pentaquarks and give more abundant suggestions to searching for the hidden-charm molecular pentaquarks accessible at the forthcoming experiment. Generally speaking, there are two important approaches to construct the family of the hidden-charm molecular pentaquark states which is very special in the hadron spectroscopy. Firstly, we propose that there may exist a series of hidden-charm molecular pentaquarks with the different strangeness. Secondly, we also have enough reason to believe that there may exist more hidden-charm molecular pentaquark states with higher mass. In fact, we already studied the $\Xi_{c}^{(\prime,*)}\bar{D}_{s}^{(*)}$ systems with double strangeness Wang:2020bjt and the $\mathcal{B}_{c}^{(*)}\bar{T}$ systems with $\mathcal{B}_{c}^{(*)}=\Lambda_{c}/\Sigma_{c}^{(*)}$ and $\bar{T}=\bar{D}_{1}/\bar{D}_{2}^{*}$ Wang:2019nwt , and predicted a series of possible candidates of the hidden-charm molecular pentaquarks. In fact, the triple-strangeness hidden-charm pentaquarks may be regarded as systems that can be used to reveal the binding mechanism and the importance of the scalar- meson exchange as they are not expected to exist in the treatment of Ref. 1839195 . Thus, we investigate the possible hidden-charm molecular pentaquarks with triple strangeness from the $\Omega_{c}^{(*)}\bar{D}_{s}^{(*)}$ interactions, which will be a main task of the present work. In the present work, we perform a dynamical calculation with the possible hidden-charm molecular pentaquark states with triple strangeness by adopting the one-boson-exchange (OBE) model Chen:2016qju ; Liu:2019zoy , which involves the interactions between an $S$-wave charmed baryon $\Omega_{c}^{(*)}$ and an $S$-wave anticharmed-strange meson $\bar{D}_{s}^{(*)}$. In concrete calculation, the $S$-$D$ wave mixing effect and the coupled channel effect are taken into account. Furthermore, we study the two-body hidden-charm strong decay behaviors of these possible hidden-charm molecular pentaquarks. Here, we adopt the quark-interchange model to estimate the transition amplitudes for the decay widths Barnes:1991em ; Barnes:1999hs ; Barnes:2000hu , which is widely used to give the decay information of the exotic hadronic states during the last few decades Wang:2018pwi ; Wang:2019spc ; Xiao:2019spy ; Wang:2020prk ; Hilbert:2007hc . We hope that the present investigation is a key step to complement the family of the hidden-charm molecular pentaquark state and may provide crucial information of searching for possible hidden-charm molecular pentaquarks with triple strangeness. With higher statistic data accumulation at Run III of the LHC and after High-Luminosity-LHC upgrade Bediaga:2018lhg , it is highly probable that these possible hidden-charm molecular pentaquarks with triple strangeness can be detected at the LHCb Collaboration in the near future, which will be full of opportunities and challenges. The remainder of this paper is organized as follows. In Sec. II, we introduce how to deduce the effective potentials and present the bound state properties of these investigated $\Omega_{c}^{(*)}\bar{D}_{s}^{(*)}$ systems. We present the quark-interchange model and the two-body strong decay behaviors of these possible molecular pentaquarks in Sec. III. Finally, a short summary follows in Sec. IV. ## II The $\Omega_{c}^{(*)}\bar{D}_{s}^{(*)}$ interactions ### II.1 OBE effective potentials In the present work, we study the interactions between an $S$-wave charmed baryon $\Omega_{c}^{(*)}$ and an $S$-wave anticharmed-strange meson $\bar{D}_{s}^{(*)}$. Here, we adopt the OBE model Chen:2016qju ; Liu:2019zoy , and consider the effective potentials from the $f_{0}(980)$, $\eta$, and $\phi$ exchanges. In particular, we need to emphasize that the light scalar meson $f_{0}(980)$ exchange provides effective interaction for these investigated systems, and we do not consider the $\sigma$ and $a_{0}(980)$ exchanges in our calculation, since the $\sigma$ is usually considered as a meson with only up and down quarks and the $a_{0}(980)$ is the light isovector scalar meson. In this subsection, we construct the relevant wave functions and effective Lagrangians, and deduce the OBE effective potentials in the coordinate space for all of the investigated systems. Firstly, we introduce the flavor and spin-orbital wave functions involved in our calculation. For the $\Omega_{c}^{(*)}\bar{D}_{s}^{(*)}$ systems, the flavor wave function $|I,I_{3}\rangle$ is quite simple and reads as $|0,0\rangle=|\Omega_{c}^{(*)0}{D}_{s}^{(*)-}\rangle$, where $I$ and $I_{3}$ are the isospin and its third component of the discussed system. In addition, the spin-orbital wave functions $|{}^{2S+1}L_{J}\rangle$ for these investigated $\Omega_{c}^{(*)}\bar{D}_{s}^{(*)}$ systems are explicitly written as $\displaystyle\left|\Omega_{c}\bar{D}_{s}\left({}^{2S+1}L_{J}\right)\right\rangle$ $\displaystyle=$ $\displaystyle\sum_{m,m_{L}}C^{J,M}_{\frac{1}{2}m,Lm_{L}}\chi_{\frac{1}{2}m}\left|Y_{L,m_{L}}\right\rangle,$ $\displaystyle\left|\Omega_{c}^{*}\bar{D}_{s}\left({}^{2S+1}L_{J}\right)\right\rangle$ $\displaystyle=$ $\displaystyle\sum_{m,m_{L}}C^{J,M}_{\frac{3}{2}m,Lm_{L}}\Phi_{\frac{3}{2}m}\left|Y_{L,m_{L}}\right\rangle,$ $\displaystyle\left|\Omega_{c}\bar{D}_{s}^{*}\left({}^{2S+1}L_{J}\right)\right\rangle$ $\displaystyle=$ $\displaystyle\sum_{m,m^{\prime},m_{S},m_{L}}C^{S,m_{S}}_{\frac{1}{2}m,1m^{\prime}}C^{J,M}_{Sm_{S},Lm_{L}}\chi_{\frac{1}{2}m}\epsilon_{m^{\prime}}^{\mu}\left|Y_{L,m_{L}}\right\rangle,$ $\displaystyle\left|\Omega_{c}^{*}\bar{D}_{s}^{*}\left({}^{2S+1}L_{J}\right)\right\rangle$ $\displaystyle=$ $\displaystyle\sum_{m,m^{\prime},m_{S},m_{L}}C^{S,m_{S}}_{\frac{3}{2}m,1m^{\prime}}C^{J,M}_{Sm_{S},Lm_{L}}\Phi_{\frac{3}{2}m}\epsilon_{m^{\prime}}^{\mu}\left|Y_{L,m_{L}}\right\rangle.$ In the above expressions, $S$, $L$, and $J$ denote the spin, orbit angular momentum, and total angular momentum for the discussed system, respectively. The constant $C^{e,f}_{ab,cd}$ is the Clebsch-Gordan coefficient, and $|Y_{L,m_{L}}\rangle$ is the spherical harmonics function. In the static limit, the polarization vector $\epsilon_{m}^{\mu}\,(m=0,\,\pm 1)$ with the spin-1 field can be expressed as $\epsilon_{0}^{\mu}=\left(0,0,0,-1\right)$ and $\epsilon_{\pm}^{\mu}=\left(0,\,\pm 1,\,i,\,0\right)/\sqrt{2}$. $\chi_{\frac{1}{2}m}$ stands for the spin wave function of the charmed baryon with spin $S={1}/{2}$, and the polarization tensor $\Phi_{\frac{3}{2}m}$ of the charmed baryon with spin quantum number $S={3}/{2}$ can be written in a general form, i.e., $\displaystyle\Phi_{\frac{3}{2}m}=\sum_{m_{1},m_{2}}C^{\frac{3}{2},m}_{\frac{1}{2}m_{1},1m_{2}}\chi_{\frac{1}{2}m_{1}}\epsilon_{m_{2}}^{\mu}.$ (2.2) In order to write out the relevant scattering amplitudes quantitatively, we usually adopt the effective Lagrangian approach. To be convenient, we construct two types of super-fields $\mathcal{S}_{\mu}$ and $H^{(\overline{Q})}_{a}$ via the heavy quark limit Wise:1992hn . The superfield $\mathcal{S}_{\mu}$ is expressed as a combination of the charmed baryons $\mathcal{B}_{6}$ with $J^{P}=1/2^{+}$ and $\mathcal{B}^{*}_{6}$ with $J^{P}=3/2^{+}$ in the $6_{F}$ flavor representation Chen:2017xat , and the superfield $H^{(\overline{Q})}_{a}$ includes the anticharmed-strange vector meson $\bar{D}^{*}_{s}$ with $J^{P}=1^{-}$ and the pseudoscalar meson $\bar{D}_{s}$ with $J^{P}=0^{-}$ Ding:2008gr . The general expressions of the super-fields $\mathcal{S}_{\mu}$ and $H^{(\overline{Q})}_{a}$ can be given by $\displaystyle\mathcal{S}_{\mu}$ $\displaystyle=$ $\displaystyle-\sqrt{\frac{1}{3}}(\gamma_{\mu}+v_{\mu})\gamma^{5}\mathcal{B}_{6}+\mathcal{B}_{6\mu}^{*},$ $\displaystyle H^{(\overline{Q})}_{a}$ $\displaystyle=$ $\displaystyle\left(\bar{D}^{*(\overline{Q})\mu}_{a}\gamma_{\mu}-\bar{D}^{(\overline{Q})}_{a}\gamma_{5}\right)\frac{1-/\\!\\!\\!v}{2}.$ (2.3) Here, $v_{\mu}=(1,\bm{0})$ is the four velocity under the nonrelativistic approximation. With the above preparation, we construct the relevant effective Lagrangians to describe the interactions among the heavy hadrons $\mathcal{B}_{6}^{(*)}/\bar{D}_{s}^{(*)}$ and the light scalar, pseudoscalar, or vector mesons as Ding:2008gr ; Chen:2017xat $\displaystyle\mathcal{L}_{\mathcal{B}^{(*)}_{6}}$ $\displaystyle=$ $\displaystyle l_{S}\langle\bar{\mathcal{S}}_{\mu}f_{0}\mathcal{S}^{\mu}\rangle-\frac{3}{2}g_{1}\varepsilon^{\mu\nu\lambda\kappa}v_{\kappa}\langle\bar{\mathcal{S}}_{\mu}{\mathcal{A}}_{\nu}\mathcal{S}_{\lambda}\rangle$ $\displaystyle+i\beta_{S}\langle\bar{\mathcal{S}}_{\mu}v_{\alpha}\left(\mathcal{V}^{\alpha}-\rho^{\alpha}\right)\mathcal{S}^{\mu}\rangle+\lambda_{S}\langle\bar{\mathcal{S}}_{\mu}F^{\mu\nu}(\rho)\mathcal{S}_{\nu}\rangle,$ $\displaystyle\mathcal{L}_{H}$ $\displaystyle=$ $\displaystyle g_{S}\langle\bar{H}^{(\overline{Q})}_{a}f_{0}H^{(\overline{Q})}_{a}\rangle+ig\langle\bar{H}^{(\overline{Q})}_{a}\gamma_{\mu}{\mathcal{A}}_{ab}^{\mu}\gamma_{5}H^{(\overline{Q})}_{b}\rangle$ (2.4) $\displaystyle-i\beta\langle\bar{H}^{(\overline{Q})}_{a}v_{\mu}\left(\mathcal{V}^{\mu}-\rho^{\mu}\right)_{ab}H^{(\overline{Q})}_{b}\rangle$ $\displaystyle+i\lambda\langle\bar{H}^{(\overline{Q})}_{a}\sigma_{\mu\nu}F^{\mu\nu}(\rho)_{ab}H^{(\overline{Q})}_{b}\rangle,$ which satisfy the requirement of the heavy quark symmetry, the chiral symmetry, and the hidden local symmetry Casalbuoni:1992gi ; Casalbuoni:1996pg ; Yan:1992gz ; Harada:2003jx ; Bando:1987br . The axial current $\mathcal{A}_{\mu}$ and the vector current ${\cal V}_{\mu}$ can be defined as ${\mathcal{A}}_{\mu}=\left(\xi^{\dagger}\partial_{\mu}\xi-\xi\partial_{\mu}\xi^{\dagger}\right)/2$ and ${\mathcal{V}}_{\mu}=\left(\xi^{\dagger}\partial_{\mu}\xi+\xi\partial_{\mu}\xi^{\dagger}\right)/2$, respectively. Here, the pseudo-Goldstone field can be written as $\xi=\exp(i\mathbb{P}/f_{\pi})$, where $f_{\pi}$ is the pion decay constant. In the above formulas, the vector meson field $\rho_{\mu}$ and its strength tensor $F_{\mu\nu}(\rho)$ are $\rho_{\mu}=i{g_{V}}\mathbb{V}_{\mu}/{\sqrt{2}}$ and $F_{\mu\nu}(\rho)=\partial_{\mu}\rho_{\nu}-\partial_{\nu}\rho_{\mu}+[\rho_{\mu},\rho_{\nu}]$, respectively. Here, $\mathcal{B}_{6}^{(*)}$, $\mathbb{V}_{\mu}$, and ${\mathbb{P}}$ are the matrices of the charmed baryon in the $6_{F}$ flavor representation, light vector meson, and light pseudoscalar meson, which can be written as $\displaystyle\left.\begin{array}[]{c}\mathcal{B}_{6}^{(*)}=\left(\begin{array}[]{ccc}\Sigma_{c}^{{(*)}++}&\frac{\Sigma_{c}^{{(*)}+}}{\sqrt{2}}&\frac{\Xi_{c}^{(^{\prime},*)+}}{\sqrt{2}}\\\ \frac{\Sigma_{c}^{{(*)}+}}{\sqrt{2}}&\Sigma_{c}^{{(*)}0}&\frac{\Xi_{c}^{(^{\prime},*)0}}{\sqrt{2}}\\\ \frac{\Xi_{c}^{(^{\prime},*)+}}{\sqrt{2}}&\frac{\Xi_{c}^{(^{\prime},*)0}}{\sqrt{2}}&\Omega_{c}^{(*)0}\end{array}\right),\\\ {\mathbb{V}}_{\mu}={\left(\begin{array}[]{ccc}\frac{\rho^{0}}{\sqrt{2}}+\frac{\omega}{\sqrt{2}}&\rho^{+}&K^{*+}\\\ \rho^{-}&-\frac{\rho^{0}}{\sqrt{2}}+\frac{\omega}{\sqrt{2}}&K^{*0}\\\ K^{*-}&\bar{K}^{*0}&\phi\end{array}\right)}_{\mu},\\\ {\mathbb{P}}={\left(\begin{array}[]{ccc}\frac{\pi^{0}}{\sqrt{2}}+\frac{\eta}{\sqrt{6}}&\pi^{+}&K^{+}\\\ \pi^{-}&-\frac{\pi^{0}}{\sqrt{2}}+\frac{\eta}{\sqrt{6}}&K^{0}\\\ K^{-}&\bar{K}^{0}&-\sqrt{\frac{2}{3}}\eta\end{array}\right)},\end{array}\right.$ (2.17) respectively. By expanding the compact effective Lagrangians to the leading order of the pseudo-Goldstone field $\xi$, we can further obtain the concrete effective Lagrangians. The effective Lagrangians for $\mathcal{B}_{6}^{(*)}$ and the light mesons are expressed as $\displaystyle\mathcal{L}_{\mathcal{B}_{6}^{(*)}\mathcal{B}_{6}^{(*)}f_{0}}$ $\displaystyle=$ $\displaystyle- l_{S}\langle\bar{\mathcal{B}}_{6}f_{0}\mathcal{B}_{6}\rangle+l_{S}\langle\bar{\mathcal{B}}_{6\mu}^{*}f_{0}\mathcal{B}_{6}^{*\mu}\rangle$ (2.18) $\displaystyle-\frac{l_{S}}{\sqrt{3}}\langle\bar{\mathcal{B}}_{6\mu}^{*}f_{0}\left(\gamma^{\mu}+v^{\mu}\right)\gamma^{5}\mathcal{B}_{6}\rangle+h.c.,$ $\displaystyle\mathcal{L}_{\mathcal{B}_{6}^{(*)}\mathcal{B}_{6}^{(*)}\mathbb{P}}$ $\displaystyle=$ $\displaystyle i\frac{g_{1}}{2f_{\pi}}\varepsilon^{\mu\nu\lambda\kappa}v_{\kappa}\langle\bar{\mathcal{B}}_{6}\gamma_{\mu}\gamma_{\lambda}\partial_{\nu}\mathbb{P}\mathcal{B}_{6}\rangle$ (2.19) $\displaystyle-i\frac{3g_{1}}{2f_{\pi}}\varepsilon^{\mu\nu\lambda\kappa}v_{\kappa}\langle\bar{\mathcal{B}}_{6\mu}^{*}\partial_{\nu}\mathbb{P}\mathcal{B}_{6\lambda}^{*}\rangle$ $\displaystyle+i\frac{\sqrt{3}g_{1}}{2f_{\pi}}v_{\kappa}\varepsilon^{\mu\nu\lambda\kappa}\langle\bar{\mathcal{B}}_{6\mu}^{*}\partial_{\nu}\mathbb{P}{\gamma_{\lambda}\gamma^{5}}\mathcal{B}_{6}\rangle+h.c.,$ $\displaystyle\mathcal{L}_{\mathcal{B}_{6}^{(*)}\mathcal{B}_{6}^{(*)}\mathbb{V}}$ $\displaystyle=$ $\displaystyle-\frac{\beta_{S}g_{V}}{\sqrt{2}}\langle\bar{\mathcal{B}}_{6}v\cdot\mathbb{V}\mathcal{B}_{6}\rangle$ (2.20) $\displaystyle-i\frac{\lambda_{S}g_{V}}{3\sqrt{2}}\langle\bar{\mathcal{B}}_{6}\gamma_{\mu}\gamma_{\nu}\left(\partial^{\mu}\mathbb{V}^{\nu}-\partial^{\nu}\mathbb{V}^{\mu}\right)\mathcal{B}_{6}\rangle$ $\displaystyle-\frac{\beta_{S}g_{V}}{\sqrt{6}}\langle\bar{\mathcal{B}}_{6\mu}^{*}v\cdot\mathbb{V}\left(\gamma^{\mu}+v^{\mu}\right)\gamma^{5}\mathcal{B}_{6}\rangle$ $\displaystyle-i\frac{\lambda_{S}g_{V}}{\sqrt{6}}\langle\bar{\mathcal{B}}_{6\mu}^{*}\left(\partial^{\mu}\mathbb{V}^{\nu}-\partial^{\nu}\mathbb{V}^{\mu}\right)\left(\gamma_{\nu}+v_{\nu}\right)\gamma^{5}\mathcal{B}_{6}\rangle$ $\displaystyle+\frac{\beta_{S}g_{V}}{\sqrt{2}}\langle\bar{\mathcal{B}}_{6\mu}^{*}v\cdot{V}\mathcal{B}_{6}^{*\mu}\rangle$ $\displaystyle+i\frac{\lambda_{S}g_{V}}{\sqrt{2}}\langle\bar{\mathcal{B}}_{6\mu}^{*}\left(\partial^{\mu}\mathbb{V}^{\nu}-\partial^{\nu}\mathbb{V}^{\mu}\right)\mathcal{B}_{6\nu}^{*}\rangle+h.c.,$ and the effective Lagrangians to describe the $S$-wave anticharmed-strange mesons $\bar{D}_{s}^{(*)}$ and the light scalar, pseudoscalar, or vector mesons are $\displaystyle\mathcal{L}_{{\bar{D}}^{(*)}{\bar{D}}^{(*)}f_{0}}$ $\displaystyle=$ $\displaystyle-2g_{S}{\bar{D}}_{a}{\bar{D}}_{a}^{{\dagger}}f_{0}+2g_{S}{\bar{D}}_{a\mu}^{*}{\bar{D}}_{a}^{*\mu{\dagger}}f_{0},$ (2.21) $\displaystyle\mathcal{L}_{{\bar{D}}^{(*)}{\bar{D}}^{(*)}\mathbb{P}}$ $\displaystyle=$ $\displaystyle\frac{2ig}{f_{\pi}}v^{\alpha}\varepsilon_{\alpha\mu\nu\lambda}{\bar{D}}_{a}^{*\mu{\dagger}}{\bar{D}}_{b}^{*\lambda}\partial^{\nu}{\mathbb{P}}_{ab}$ (2.22) $\displaystyle+\frac{2g}{f_{\pi}}\left({\bar{D}}_{a}^{*\mu{\dagger}}{\bar{D}}_{b}+{\bar{D}}_{a}^{{\dagger}}{\bar{D}}_{b}^{*\mu}\right)\partial_{\mu}{\mathbb{P}}_{ab},$ $\displaystyle\mathcal{L}_{{\bar{D}}^{(*)}{\bar{D}}^{(*)}\mathbb{V}}$ $\displaystyle=$ $\displaystyle\sqrt{2}\beta g_{V}{\bar{D}}_{a}{\bar{D}}_{b}^{{\dagger}}v\cdot\mathbb{V}_{ab}-\sqrt{2}\beta g_{V}{\bar{D}}_{a\mu}^{*}{\bar{D}}_{b}^{*\mu{\dagger}}v\cdot\mathbb{V}_{ab}$ $\displaystyle-2\sqrt{2}i\lambda g_{V}{\bar{D}}_{a}^{*\mu{\dagger}}{\bar{D}}_{b}^{*\nu}\left(\partial_{\mu}\mathbb{V}_{\nu}-\partial_{\nu}\mathbb{V}_{\mu}\right)_{ab}$ $\displaystyle-2\sqrt{2}\lambda g_{V}v^{\lambda}\varepsilon_{\lambda\mu\alpha\beta}\left({\bar{D}}_{a}^{*\mu{\dagger}}{\bar{D}}_{b}+{\bar{D}}_{a}^{{\dagger}}{\bar{D}}_{b}^{*\mu}\right)\partial^{\alpha}\mathbb{V}^{\beta}_{ab}.$ In the above effective Lagrangians, the coupling constants can be either extracted from the experimental data or calculated by the theoretical models, and the signs of these coupling constants are fixed via the quark model Riska:2000gd . The values of these coupling constants are $l_{S}=6.20$, $g_{S}=0.76$,222In this work, we consider the contribution from light scalar meson $f_{0}(980)$ exchange. Here, the corresponding coupling constant involved in effective Lagrangians [Eq. (2.18) and Eq. (2.21)] is approximately taken as the same as that for the case of light scalar $\sigma$. $g_{1}=0.94$, $g=0.59$, $f_{\pi}=132~{}\rm{MeV}$, $\beta_{S}g_{V}=10.14$, $\beta g_{V}=-5.25$, $\lambda_{S}g_{V}=19.2~{}\rm{GeV}^{-1}$, and $\lambda g_{V}=-3.27~{}\rm{GeV}^{-1}$ Chen:2019asm , which are widely used to discuss the hadronic molecular states Wang:2020bjt ; Chen:2017xat ; Wang:2019nwt ; Chen:2019asm ; He:2015cea ; He:2019ify ; Chen:2018pzd . In particular, we need to emphasize that these input coupling constants can well reproduce the masses of the $P_{c}(4312)$, $P_{c}(4440)$, and $P_{c}(4457)$ Aaij:2019vzc under the meson-baryon molecular picture when adopting the OBE model Chen:2019asm ; He:2019ify . We follow the standard strategy to deduce the effective potentials in the coordinate space in Refs. Wang:2020dya ; Wang:2019nwt ; Wang:2019aoc , which is a lengthy and tedious calculation. In Fig. 1, we present the relevant Feynman diagram for the $\Omega_{c}^{(*)}\bar{D}_{s}^{(*)}\to\Omega_{c}^{(*)}\bar{D}_{s}^{(*)}$ scattering processes. Figure 1: Relevant Feynman diagram for the $\Omega_{c}^{(*)}\bar{D}_{s}^{(*)}\to\Omega_{c}^{(*)}\bar{D}_{s}^{(*)}$ scattering processes. At the hadronic level, we firstly write out the scattering amplitude $\mathcal{M}(h_{1}h_{2}\to h_{3}h_{4})$ of the scattering process $h_{1}h_{2}\to h_{3}h_{4}$ by considering the effective Lagrangian approach. And then, the effective potential in momentum space $\mathcal{V}^{h_{1}h_{2}\to h_{3}h_{4}}(\bm{q})$ can be related to the scattering amplitude $\mathcal{M}(h_{1}h_{2}\to h_{3}h_{4})$ with the help of the Breit approximation Breit:1929zz ; Breit:1930zza and the nonrelativistic normalization, i.e., $\displaystyle\mathcal{V}_{E}^{h_{1}h_{2}\to h_{3}h_{4}}(\bm{q})$ $\displaystyle=$ $\displaystyle-\frac{\mathcal{M}(h_{1}h_{2}\to h_{3}h_{4})}{\sqrt{\prod_{i}2m_{i}\prod_{f}2m_{f}}},$ (2.24) where $m_{i}$ and $m_{f}$ are the masses of the initial states $(h_{1},\,h_{2})$ and final states $(h_{3},\,h_{4})$, respectively. By performing the Fourier transformation, the effective potential in the coordinate space $\mathcal{V}^{h_{1}h_{2}\to h_{3}h_{4}}(\bm{r})$ can be deduced $\displaystyle\mathcal{V}^{h_{1}h_{2}\to h_{3}h_{4}}_{E}(\bm{r})=\int\frac{d^{3}\bm{q}}{(2\pi)^{3}}e^{i\bm{q}\cdot\bm{r}}\mathcal{V}_{E}^{h_{1}h_{2}\to h_{3}h_{4}}(\bm{q})\mathcal{F}^{2}(q^{2},m_{E}^{2}).$ In order to reflect the finite size effect of the discussed hadrons and compensate the off-shell effect of the exchanged light mesons Wang:2020dya , we need to introduce the monopole form factor $\mathcal{F}(q^{2},m_{E}^{2})=(\Lambda^{2}-m_{E}^{2})/(\Lambda^{2}-q^{2})$ in the interaction vertex Tornqvist:1993ng ; Tornqvist:1993vu . Here, $\Lambda$, $m_{E}$, and $q$ are the cutoff parameter, the mass, and the four momentum of the exchanged light meson, respectively. In addition, we also need a series of normalization relations for the heavy hadrons $D_{s}$, $D_{s}^{*}$, $\Omega_{c}$, and $\Omega_{c}^{*}$, i.e., $\displaystyle\langle 0|D_{s}|c\bar{s}\left(0^{-}\right)\rangle$ $\displaystyle=$ $\displaystyle\sqrt{M_{D_{s}}},$ (2.26) $\displaystyle\langle 0|D_{s}^{*\mu}|c\bar{s}\left(1^{-}\right)\rangle$ $\displaystyle=$ $\displaystyle\sqrt{M_{D_{s}^{*}}}\epsilon^{\mu},$ (2.27) $\displaystyle\langle 0|\Omega_{c}|css\left({1}/{2}^{+}\right)\rangle$ $\displaystyle=$ $\displaystyle\sqrt{2M_{\Omega_{c}}}{\left(\chi_{\frac{1}{2}m},\frac{\bf{\sigma}\cdot\bf{p}}{2M_{\Omega_{c}}}\chi_{\frac{1}{2}m}\right)^{T}},$ (2.28) $\displaystyle\langle 0|\Omega_{c}^{*\mu}|css\left({3}/{2}^{+}\right)\rangle$ $\displaystyle=$ $\displaystyle\sum_{m_{1},m_{2}}C_{1/2,m_{1};1,m_{2}}^{3/2,m_{1}+m_{2}}\sqrt{2M_{\Omega_{c}^{*}}}$ (2.29) $\displaystyle\times\left(\chi_{\frac{1}{2}m_{1}},\frac{\bf{\sigma}\cdot\bf{p}}{2M_{\Omega_{c}^{*}}}\chi_{\frac{1}{2}m_{1}}\right)^{T}\epsilon^{\mu}_{m_{2}}.$ With the above preparation, we can deduce the OBE effective potentials in the coordinate space for all of the investigated processes, which are collected in the A. ### II.2 Finding bound state solutions for discussed systems Now, we attempt to find the loosely bound state solutions of these discussed $\Omega_{c}^{(*)}\bar{D}_{s}^{(*)}$ systems by solving the coupled channel Schrödinger equation, i.e., $\displaystyle-\frac{1}{2\mu}\left(\nabla^{2}-\frac{\ell(\ell+1)}{r^{2}}\right)\psi(r)+V(r)\psi(r)=E\psi(r)$ (2.30) with $\nabla^{2}=\frac{1}{r^{2}}\frac{\partial}{\partial r}r^{2}\frac{\partial}{\partial r}$, where $\mu=\frac{m_{1}m_{2}}{m_{1}+m_{2}}$ is the reduced mass for the discussed system. The bound state solutions include the binding energy $E$, the root- mean-square radius $r_{\rm RMS}$, and the probability of the individual channel $P_{i}$, which provides us with valuable information to analyze whether the loosely bound state exists. In this work, we are interested in the $S$-wave $\Omega_{c}^{(*)}\bar{D}_{s}^{(*)}$ systems since there exists the repulsive centrifugal potential for the higher partial wave states $\ell\geqslant 1$. In our calculation, the masses of these involved hadrons are $m_{f_{0}}=990.00$ MeV, $m_{\eta}=547.85$ MeV, $m_{\phi}=1019.46$ MeV, $m_{D_{s}}=1968.34$ MeV, $m_{D_{s}^{*}}=2112.20$ MeV, $m_{\Omega_{c}}=2695.20$ MeV, and $m_{\Omega_{c}^{*}}=2765.90$ MeV, which are taken from the Particle Data Group (PDG) Zyla:2020zbs . As the remaining phenomenological parameter, we take the cutoff value from 1.00 to 4.00 GeV. Usually, a loosely bound state with the cutoff parameter closed to 1.00 GeV can be suggested as the possible hadronic molecular candidate according to the experience of the deuteron Tornqvist:1993ng ; Tornqvist:1993vu ; Wang:2019nwt ; Chen:2017jjn . For an ideal hadronic molecular candidate, the reasonable binding energy should be at most tens of MeV, and the typical size should be larger than the size of all the included component hadrons Chen:2017xat . In addition, the $S$-$D$ wave mixing effect is considered in this work, which plays an important role to modify the bound state properties of the deuteron Wang:2019nwt . The relevant channels $|{}^{2S+1}L_{J}\rangle$ are summarized in Table 1. Table 1: The relevant channels $|{}^{2S+1}L_{J}\rangle$ involved in our calculation. Here, “$...$” means that the $S$-wave component for the corresponding channel does not exist. $J^{P}$ | $\Omega_{c}\bar{D}_{s}$ | $\Omega_{c}^{*}\bar{D}_{s}$ | $\Omega_{c}\bar{D}_{s}^{*}$ | $\Omega_{c}^{*}\bar{D}^{*}_{s}$ ---|---|---|---|--- ${1}/{2}^{-}$ | $|{}^{2}\mathbb{S}_{1/2}\rangle$ | $...$ | $|{}^{2}\mathbb{S}_{1/2}\rangle/|{}^{4}\mathbb{D}_{1/2}\rangle$ | $|{}^{2}\mathbb{S}_{1/2}\rangle/|{}^{4,6}\mathbb{D}_{1/2}\rangle$ ${3}/{2}^{-}$ | $...$ | $|{}^{4}\mathbb{S}_{3/2}\rangle/|{}^{4}\mathbb{D}_{3/2}\rangle$ | $|{}^{4}\mathbb{S}_{3/2}\rangle/|{}^{2,4}\mathbb{D}_{3/2}\rangle$ | $|{}^{4}\mathbb{S}_{3/2}\rangle/|{}^{2,4,6}\mathbb{D}_{3/2}\rangle$ ${5}/{2}^{-}$ | $...$ | $...$ | $...$ | $|{}^{6}\mathbb{S}_{5/2}\rangle/|{}^{2,4,6}\mathbb{D}_{5/2}\rangle$ Before performing numerical calculation, we analyze the OBE effective potentials for these discussed $\Omega_{c}^{(*)}\bar{D}_{s}^{(*)}$ systems as below: * • For the $\Omega_{c}\bar{D}_{s}$ and $\Omega_{c}^{*}\bar{D}_{s}$ systems, only the $f_{0}$ and $\phi$ exchange interactions are allowed. Meanwhile, the tensor force from the $S$-$D$ wave mixing effect disappears in the effective potentials, and thus the contribution of the $S$-$D$ wave mixing effect does not affect the bound state properties of the $\Omega_{c}\bar{D}_{s}$ and $\Omega_{c}^{*}\bar{D}_{s}$ systems. * • For the $\Omega_{c}\bar{D}_{s}^{*}$ and $\Omega_{c}^{*}\bar{D}_{s}^{*}$ systems, in addition to the $f_{0}$ and $\phi$ exchange interactions, the $\eta$ exchange interaction and the $S$-$D$ wave mixing effect need to be taken into account. #### II.2.1 The $\Omega_{c}\bar{D}_{s}$ and $\Omega_{c}^{*}\bar{D}_{s}$ systems For the $S$-wave $\Omega_{c}\bar{D}_{s}$ state with $J^{P}={1}/{2}^{-}$ and the $S$-wave $\Omega_{c}^{*}\bar{D}_{s}$ state with $J^{P}={3}/{2}^{-}$, we fail to find their bound state solutions by varying the cutoff parameter in the range of $1.00$-$4.00~{}{\rm GeV}$ with the single channel analysis. Nevertheless, we can further take into account the coupled channel effect. In the coupled channel analysis, the binding energy of the bound state is determined by the lowest mass threshold among various investigated channels Chen:2017xat . For the $S$-wave $\Omega_{c}\bar{D}_{s}$ state with $J^{P}={1}/{2}^{-}$, we consider the coupled channel effect from the $\Omega_{c}\bar{D}_{s}$, $\Omega_{c}\bar{D}_{s}^{*}$, and $\Omega_{c}^{*}\bar{D}_{s}^{*}$ channels. In Table 2, we present the obtained bound state solutions by performing the coupled channel analysis. When we set the cutoff parameter $\Lambda$ around 2.92 GeV, the loosely bound state solutions can be obtained, and the $\Omega_{c}\bar{D}_{s}$ channel is dominant with almost 90% probabilities. Since the cutoff parameter $\Lambda$ is obviously different from 1.00 GeV Tornqvist:1993ng ; Tornqvist:1993vu , the $S$-wave $\Omega_{c}\bar{D}_{s}$ state with $J^{P}={1}/{2}^{-}$ is not priority for recommending the hadronic molecular candidate. Table 2: Bound state solutions of the $S$-wave $\Omega_{c}\bar{D}_{s}$ state with $J^{P}={1}/{2}^{-}$ by performing coupled channel analysis. Here, the cutoff parameter $\Lambda$, binding energy $E$, and root-mean-square radius $r_{RMS}$ are in units of $\rm{GeV}$, $\rm{MeV}$, and $\rm{fm}$, respectively. $\Lambda$ | $E$ | $r_{\rm RMS}$ | P($\Omega_{c}\bar{D}_{s}/\Omega_{c}\bar{D}_{s}^{*}/\Omega_{c}^{*}\bar{D}_{s}^{*}$) ---|---|---|--- 2.92 | $-3.71$ | 1.26 | 92.92/4.77/2.31 2.93 | $-12.65$ | 0.64 | 90.11/6.66/3.23 In Table 3, we list the bound state solutions of the $S$-wave $\Omega_{c}^{*}\bar{D}_{s}$ state with $J^{P}={3}/{2}^{-}$ with the coupled channel analysis. Our numerical results show that the bound state solutions can be obtained by choosing the cutoff parameter $\Lambda$ around 1.78 GeV or even larger, and the $\Omega_{c}\bar{D}_{s}^{*}$ system is the dominant channel with the probabilities over 80%. However, we find the size ($r_{\rm RMS}\sim 0.33~{}{\rm{fm}}$) of this bound state is too small,333 We notice that the obtained values of $r_{RMS}$ are too small, which is due to the fact that this sysmtem is dominated by the $\Omega_{c}\bar{D}_{s}^{*}$ channel as shown in the last column of Table 3. which is not consistent with a loosely molecular state picture Chen:2017xat . Thus, we tentatively exclude the possibility of the existence of the $S$-wave $\Omega_{c}^{*}\bar{D}_{s}$ molecular state with $J^{P}={3}/{2}^{-}$. Table 3: Bound state solutions of the $S$-wave $\Omega_{c}^{*}\bar{D}_{s}$ state with $J^{P}={3}/{2}^{-}$ when the coupled channel effect is introduced. The units are the same as Table 2. $\Lambda$ | $E$ | $r_{\rm RMS}$ | P($\Omega_{c}^{*}\bar{D}_{s}/\Omega_{c}\bar{D}_{s}^{*}/\Omega_{c}^{*}\bar{D}_{s}^{*}$) ---|---|---|--- 1.78 | $-6.15$ | 0.33 | 0.01/86.64/13.36 1.79 | $-17.41$ | 0.32 | 0.01/86.37/13.63 #### II.2.2 The $\Omega_{c}\bar{D}_{s}^{*}$ and $\Omega_{c}^{*}\bar{D}_{s}^{*}$ systems For the $S$-wave $\Omega_{c}\bar{D}_{s}^{*}$ system, the relevant numerical results are collected in Table 4. For $J^{P}={1}/{2}^{-}$, there do not exist bound states until we increase the cutoff parameter to be around 4.00 GeV, even if we consider the coupled channel effect. Thus, we conclude that our quantitative analysis does not support the existence of the $S$-wave $\Omega_{c}\bar{D}_{s}^{*}$ molecular state with $J^{P}={1}/{2}^{-}$. Table 4: Bound state solutions of the $S$-wave $\Omega_{c}\bar{D}_{s}^{*}$ system. The units are the same as Table 2. Effect | Single channel | $S$-$D$ wave mixing effect | Coupled channel ---|---|---|--- $J^{P}$ | $\Lambda$ | $E$ | $r_{\rm RMS}$ | $\Lambda$ | $E$ | $r_{\rm RMS}$ | P(${}^{4}\mathbb{S}_{\frac{3}{2}}/{}^{2}\mathbb{D}_{\frac{3}{2}}/{}^{4}\mathbb{D}_{\frac{3}{2}})$ | $\Lambda$ | $E$ | $r_{\rm RMS}$ | P($\Omega_{c}\bar{D}_{s}^{*}/\Omega_{c}^{*}\bar{D}_{s}^{*}$) ${3}/{2}^{-}$ | 1.96 | $-0.19$ | 4.76 | 1.96 | $-0.33$ | 4.14 | 99.94/0.01/0.05 | 1.67 | $-1.36$ | 2.27 | 95.84/4.16 1.98 | $-5.36$ | 1.09 | 1.98 | $-5.71$ | 1.06 | 99.92/0.02/0.06 | 1.69 | $-8.38$ | 0.90 | 92.17/7.83 1.99 | $-9.44$ | 0.82 | 1.99 | $-9.84$ | 0.81 | 99.93/0.02/0.05 | 1.70 | $-13.35$ | 0.72 | 91.00/9.00 For the $S$-wave $\Omega_{c}\bar{D}_{s}^{*}$ state with $J^{P}={3}/{2}^{-}$, we notice that the effective potentials from the $f_{0}$, $\eta$, and $\phi$ exchanges provide the attractive forces, and there exist the bound state solutions with the cutoff parameter around 1.96 GeV by performing the single channel analysis. More interestingly, the bound state properties will change accordingly after including the coupled channels $\Omega_{c}\bar{D}_{s}^{*}$ and $\Omega_{c}^{*}\bar{D}_{s}^{*}$, where we can obtain the loosely bound state solutions when the cutoff parameter $\Lambda$ around 1.67 GeV. Moreover, this bound state is mainly composed of the $\Omega_{c}\bar{D}_{s}^{*}$ channel with the probabilities over 90%. Based on our numerical results, the $S$-wave $\Omega_{c}\bar{D}_{s}^{*}$ state with $J^{P}={3}/{2}^{-}$ can be recommended as a good candidate of the hidden-charm molecular pentaquark with triple strangeness. Comparing the numerical results, it is obvious that the $D$-wave probabilities are less than 1% and the $S$-$D$ mixing effect can be ignored in forming the $S$-wave $\Omega_{c}\bar{D}_{s}^{*}$ bound states, but the coupled channel effect is obvious in generating the $S$-wave $\Omega_{c}\bar{D}_{s}^{*}$ bound states, especially for the $S$-wave $\Omega_{c}\bar{D}_{s}^{*}$ molecular candidate with $J^{P}={3}/{2}^{-}$. For the $S$-wave $\Omega_{c}^{*}\bar{D}_{s}^{*}$ system, the bound state properties are collected in Table 5. Here, we still scan the $\Lambda$ parameter range from 1.00 GeV to 4.00 GeV. For $J^{P}=1/2^{-}$, the binding energy is a few MeV and the root-mean-square radii are around 1.00 fm with the cutoff parameter $\Lambda$ larger than 3.59 GeV when only considering the $S$-wave channel, and we can also obtain the bound state solutions when the cutoff value $\Lambda$ is lowered down 3.51 GeV after adding the contribution of the $D$-wave channels. Because the obtained cutoff parameter $\Lambda$ is far away from 1.00 GeV Tornqvist:1993ng ; Tornqvist:1993vu , our numerical results disfavor the existence of the molecular candidate for the $S$-wave $\Omega_{c}^{*}\bar{D}_{s}^{*}$ state with $J^{P}=1/2^{-}$. For $J^{P}=3/2^{-}$, there do not exist the bound state solutions when the cutoff parameter varies from 1.00 GeV to 4.00 GeV. This situation does not change when the $S$-$D$ wave mixing effect is considered. Thus, we can exclude the $S$-wave $\Omega_{c}^{*}\bar{D}_{s}^{*}$ state with $J^{P}=3/2^{-}$ as the hadronic molecular candidate. For $J^{P}=5/2^{-}$, we notice that the total effective potentials due to the $f_{0}$, $\eta$, and $\phi$ exchanges are attractive. We can obtain the loosely bound state solutions by taking the cutoff value around 1.64 GeV when only considering the contribution of the $S$-wave channel, and the bound state solutions also can be found with the cutoff parameter around 1.64 GeV after considering the $S$-$D$ wave mixing effect. As a result, the $S$-wave $\Omega_{c}^{*}\bar{D}_{s}^{*}$ state with $J^{P}=5/2^{-}$ can be regarded as the hidden-charm molecular pentaquark candidate with triple strangeness. Table 5: Bound state solutions of the $S$-wave $\Omega_{c}^{*}\bar{D}_{s}^{*}$ system. The units are the same as Table 2. Effect | Single channel | $S$-$D$ wave mixing effect ---|---|--- $J^{P}$ | $\Lambda$ | $E$ | $r_{\rm RMS}$ | $\Lambda$ | $E$ | $r_{\rm RMS}$ | P(${}^{2}\mathbb{S}_{\frac{1}{2}}/{}^{4}\mathbb{D}_{\frac{1}{2}}/{}^{6}\mathbb{D}_{\frac{1}{2}})$ ${1}/{2}^{-}$ | $3.59$ | $-0.27$ | $4.96$ | 3.51 | $-0.29$ | 4.89 | 99.97/0.02/0.01 $3.80$ | $-1.18$ | $2.96$ | 3.76 | $-1.74$ | 2.52 | 99.92/0.05/0.03 $4.00$ | $-2.63$ | $2.11$ | 4.00 | $-4.35$ | 1.72 | 99.87/0.08/0.05 $J^{P}$ | $\Lambda$ | $E$ | $r_{\rm RMS}$ | $\Lambda$ | $E$ | $r_{\rm RMS}$ | P(${}^{6}\mathbb{S}_{\frac{5}{2}}/{}^{2}\mathbb{D}_{\frac{5}{2}}/{}^{4}\mathbb{D}_{\frac{5}{2}}/{}^{6}\mathbb{D}_{\frac{5}{2}})$ ${5}/{2}^{-}$ | 1.64 | $-0.31$ | 4.27 | 1.64 | $-0.80$ | 2.97 | 99.81/0.02/0.01/0.15 1.66 | $-4.93$ | 1.19 | 1.66 | $-5.81$ | 1.11 | 99.76/0.03/0.01/0.20 1.68 | $-13.13$ | 0.74 | 1.67 | $-9.55$ | 0.87 | 99.77/0.03/0.01/0.19 To summarize, we predict two types of hidden-charm molecular pentaquark states with triple strangeness, i.e., the $S$-wave $\Omega_{c}\bar{D}_{s}^{*}$ molecular state with $J^{P}={3}/{2}^{-}$ and the $S$-wave $\Omega_{c}^{*}\bar{D}_{s}^{*}$ molecular state with $J^{P}=5/2^{-}$. Here, we want to indicate that the effective potentials from the $\phi$ and $\eta$ exchanges are attractive for the $\Omega_{c}\bar{D}_{s}^{*}$ system with $J^{P}={3}/{2}^{-}$ and the $\Omega_{c}^{*}\bar{D}_{s}^{*}$ system with $J^{P}=5/2^{-}$, which is due to the contributions from the ${\bf q}^{2}$ terms in the deduced effective potentials. In fact, this issue has been discussed in Ref. 1839195 . ## III Decay behaviors of these possible $\Omega_{c}^{(*)}\bar{D}_{s}^{*}$ molecular states In order to further reveal the inner structures and properties of the possible hidden-charm molecular pentaquarks with triple strangeness, we calculate the strong decay behaviors of these possible molecular candidates. In this work, we discuss the hidden-charm decay mode, the corresponding final states including the $\eta_{c}(1S)\Omega$ and $J/\psi\Omega$. Different with the binding of the possible hidden-charm molecular pentaquarks with triple strangeness, the interactions in the very short range distance contribute to the hidden-charm decay processes. Thus, the quark-interchange model Barnes:1991em ; Barnes:1999hs can be as a reasonable theoretical framework. ### III.1 The quark-interchange model When using the quark-interchange model to estimate the transition amplitudes in calculating the decay widths, we usually adopt the nonrelativistic quark model to describe the quark-quark interaction Wang:2019spc ; Xiao:2019spy , which is expressed as Wong:2001td $V_{ij}(q^{2})=\frac{\lambda_{i}}{2}\cdot\frac{\lambda_{j}}{2}\left(\frac{4\pi\alpha_{s}}{q^{2}}+\frac{6\pi b}{q^{4}}-\frac{8\pi\alpha_{s}}{3m_{i}m_{j}}e^{-{\frac{q^{2}}{4\sigma^{2}}}}{\bf{s}}_{i}\cdot{\bf{s}}_{j}\right),$ (3.1) where $\lambda_{i}(\lambda_{j})$, $m_{i}(m_{j})$, and ${\bf{s}}_{i}({\bf{s}}_{j})$ represent the color factor, the mass, and the spin operator of the interacting quarks, respectively. $\alpha_{s}$ denotes the running coupling constant, which reads as Wong:2001td $\alpha_{s}(Q^{2})=\frac{12\pi}{\left(32-2n_{f}\right){\rm ln}\left(A+\frac{Q^{2}}{B^{2}}\right)},$ (3.2) where $Q^{2}$ is the square of the invariant mass of the interacting quarks, and the relevant parameters Wong:2001td in Eqs. (3.1) and (3.2) are collected in Table 6. Table 6: The parameters of the nonrelativistic quark model Wong:2001td and the oscillating parameters of the Gaussian function Wang:2019spc . Quark model | $b~{}(\rm{GeV}^{2})$ | $\sigma~{}(\rm{GeV})$ | $A$ ---|---|---|--- 0.180 | 0.897 | 10 $B$ (GeV) | $m_{s}~{}(\rm{GeV})$ | $m_{c}~{}(\rm{GeV})$ 0.310 | 0.575 | 1.776 Oscillating parameters | $\beta_{D^{\ast}_{s}}~{}(\rm{GeV})$ | $\beta_{\eta_{c}}~{}(\rm{GeV})$ | $\beta_{J/\psi}~{}(\rm{GeV})$ 0.562 | 0.838 | 0.729 $\alpha_{\lambda\Omega}~{}(\rm{GeV})$ | $\alpha_{\rho\Omega}~{}(\rm{GeV})$ | $\alpha_{\lambda\Omega_{c}}~{}(\rm{GeV})$ 0.466 | 0.407 | 0.583 $\alpha_{\rho\Omega_{c}}~{}(\rm{GeV})$ | $\alpha_{\lambda\Omega^{\ast}_{c}}~{}(\rm{GeV})$ | $\alpha_{\rho\Omega^{\ast}_{c}}~{}(\rm{GeV})$ 0.444 | 0.537 | 0.423 To get the transition amplitudes within the quark-interchange model, we take the same convention as the previous work Wang:2019spc ; Xiao:2019spy ; Wang:2020prk . The transition amplitude for the process $A(css)+B(s\bar{c})\to C(sss)+D(c\bar{c})$ can be decomposed as four processes in the hadronic molecular picture, which are illustrated in Fig. 2. Figure 2: Quark-interchange diagrams for the process $A(css)+B(s\bar{c})\to C(sss)+D(c\bar{c})$ in the hadronic molecular picture. The Hamiltonian of the initial hidden-charm molecular pentaquark state can be written as Wang:2019spc $H_{\rm{Initial}}=H^{0}_{A}+H^{0}_{B}+V_{AB},$ (3.3) where $H^{0}_{A}$ and $H^{0}_{B}$ are the Hamiltonian of the free baryon A and meson B, and $V_{AB}$ denotes the interaction between the baryon A and the meson B. Furthermore, we define the color wave function $\omega_{\rm{color}}$, the flavor wave function $\chi_{\rm{flavor}}$, the spin wave function $\chi_{\rm{spin}}$, and the momentum space wave function $\phi(\bf{p})$, respectively. Thus, the total wave function can be expressed as $\psi_{\rm{total}}=\omega_{\rm{color}}\chi_{\rm{flavor}}\chi_{\rm{spin}}\phi(\bf{p}).$ (3.4) In this work, we take the Gaussian functions to approximate the momentum space wave functions for the baryon, meson, and molecule. The more explicit forms of the relevant Gaussian function can be found in Ref. Wang:2019spc , and the oscillating parameters of the meson and baryon are estimated by fitting their mass spectrum in the Godfrey-Isgur model Godfrey:1985xj , which are listed in Table 6. For an $S$-wave loosely bound state composed of two hadrons A and B, the oscillating parameter $\beta$ can be related to the mass of the molecular state $m$, i.e., $\beta=\sqrt{3\mu(m_{A}+m_{B}-m)}$ with $\mu=\frac{m_{A}m_{B}}{m_{A}+m_{B}}$ Weinberg:1962hj ; Weinberg:1963zza ; Guo:2017jvc . And then, the $T$-matrix $T_{fi}$ represents the relevant effective potential in the quark-interchange diagrams, which can be factorized as $T_{fi}=I_{\rm{color}}I_{\rm{flavor}}I_{\rm{spin}}I_{\rm{space}},$ (3.5) where $I_{i}$ with the subscripts color, flavor, spin, and space stand for the corresponding factors, and the calculation details of these factors $I_{i}\,(i=\rm{color},\,\rm{flavor},\,\rm{spin},\,\rm{space})$ are referred to in Ref. Wang:2019spc . For the two-body strong decay widths of these discussed molecular candidates, they can be explicitly expressed as $\displaystyle\Gamma=\frac{|{\bf{P}}_{C}|}{32\pi^{2}m^{2}(2J+1)}\int d\Omega|\mathcal{M}|^{2}.$ (3.6) In the above expression, ${\bf{P}}_{C}$, $m$, and $\mathcal{M}$ stand for the momentum of the final state, the mass of the molecular state, and the transition amplitude of the discussed process, respectively. Here, we want to emphasize that there exists a relation of the transition amplitude $\mathcal{M}$ and the $T$-matrix $T_{fi}$, i.e., $\displaystyle\mathcal{M}=-(2\pi)^{\frac{3}{2}}\sqrt{2m2E_{C}2E_{D}}T_{fi},$ (3.7) where $E_{C}$ and $E_{D}$ are the energies of the final states C and D, respectively. Through the above preparation, we can calculate the two-body hidden-charm strong decay widths of these proposed $\Omega_{c}^{(*)}\bar{D}_{s}^{*}$ molecular states. ### III.2 Two-body hidden-charm strong decay widths of these proposed $\Omega_{c}^{(*)}\bar{D}_{s}^{*}$ molecular states In the above section, our results suggest that the $S$-wave $\Omega_{c}\bar{D}_{s}^{*}$ state with $J^{P}={3}/{2}^{-}$ and the $S$-wave $\Omega_{c}^{*}\bar{D}_{s}^{*}$ state with $J^{P}=5/2^{-}$ can be regarded as the hidden-charm molecular pentaquark candidates with triple strangeness. Thus, we will study the two-body strong decay property of these possible hidden-charm molecular pentaquarks with triple strangeness, which provides valuable information to search for these proposed molecular candidates in experiment. In this work, we focus on the two-body hidden-charm strong decay channels for these predicted hidden-charm molecular pentaquarks with triple strangeness. For the $S$-wave $\Omega_{c}\bar{D}_{s}^{*}$ molecular state with $J^{P}={3}/{2}^{-}$, it can decay into the $J/\psi\,\Omega$ and $\eta_{c}\,\Omega$ channels through the $S$-wave interaction. For the $S$-wave $\Omega_{c}^{*}\bar{D}_{s}^{*}$ molecular state with $J^{P}={5}/{2}^{-}$, we only take into account the $J/\psi\,\Omega$ decay channel via the $S$-wave coupling, while the $\eta_{c}\,\Omega$ channel is suppressed since it is a $D$-wave decay Wang:2019spc . In order to intuitively clarify the uncertainty of the binding energies, we present the binding energies dependence of the decay widths for the $S$-wave $\Omega_{c}\bar{D}_{s}^{*}$ molecular state with $J^{P}={3}/{2}^{-}$ and the $S$-wave $\Omega_{c}^{*}\bar{D}_{s}^{*}$ molecular state with $J^{P}={5}/{2}^{-}$ in Fig. 3. As stressed in Sec. II, the hadronic molecule is a loosely bound state Chen:2017xat , so the binding energies of these hidden-charm molecular pentaquarks with triple strangeness change from $-20$ to $-1$ MeV in calculating the decay widths. With increasing the absolute values of the binding energy, the decay widths become larger, which is consistent with other theoretical calculations Chen:2017xat ; Lin:2017mtz ; Lin:2018kcc ; Lin:2018nqd ; Shen:2019evi ; Lin:2019qiv ; Lin:2019tex ; Dong:2019ofp ; Dong:2020rgs ; Xiao:2019mvs ; Wu:2018xaa ; Chen:2017abq ; Xiao:2016mho . Figure 3: The binding energies dependence of the decay widths for the $S$-wave $\Omega_{c}\bar{D}_{s}^{*}$ molecular state with $J^{P}={3}/{2}^{-}$ and the $S$-wave $\Omega_{c}^{*}\bar{D}_{s}^{*}$ molecular state with $J^{P}={5}/{2}^{-}$. As illustrated in Fig. 3, when the binding energies are taken as $-15$ MeV with typical values, the dominant decay channel is the $J/\psi\,\Omega$ around one MeV for the $S$-wave $\Omega_{c}\bar{D}_{s}^{*}$ molecular state with $J^{P}={3}/{2}^{-}$, and the decay width of the $J/\psi\,\Omega$ channel is predicted to be around several MeV for the $S$-wave $\Omega_{c}^{*}\bar{D}_{s}^{*}$ molecule with $J^{P}={5}/{2}^{-}$. Thus, the $J/\psi\,\Omega$ should be the promising channel to observe the $S$-wave $\Omega_{c}\bar{D}_{s}^{*}$ molecular state with $J^{P}={3}/{2}^{-}$ and the $S$-wave $\Omega_{c}^{*}\bar{D}_{s}^{*}$ molecular state with $J^{P}={5}/{2}^{-}$. Meanwhile, it is interesting to note that the $S$-wave $\Omega_{c}\bar{D}_{s}^{*}$ molecular state with $J^{P}={3}/{2}^{-}$ prefers to decay into the $J/\psi\,\Omega$ channel, but the decay width of the $\eta_{c}\,\Omega$ channel is comparable to the $J/\psi\,\Omega$ channel, which indicates that the $S$-wave $\Omega_{c}\bar{D}_{s}^{*}$ molecule with $J^{P}={3}/{2}^{-}$ can be detected in the $\eta_{c}\,\Omega$ channel in future experiment. In the heavy quark symmetry, the relative partial decay branch ratio between the $\eta_{c}(1S)\Omega$ and $J/\psi\Omega$ for the $\Omega_{c}\bar{D}_{s}^{*}$ state with $J^{P}=3/2^{-}$ can be estimated as $\displaystyle\mathcal{R}_{\text{HQS}}=\frac{\Gamma(\Omega_{c}\bar{D}_{s}^{*}\to\eta_{c}(1S)\Omega)}{\Gamma(\Omega_{c}\bar{D}_{s}^{*}\to J/\psi\Omega)}=0.6,$ (3.8) since the relative momentum in the $\eta_{c}(1S)\Omega$ channel is larger than that in the $J/\psi\Omega$ channel, $\mathcal{R}(E)$ should be a little larger than $\mathcal{R}_{\text{HQS}}=0.6$, where $E$ is the binding energy. In our calculation, we obtain $\displaystyle\mathcal{R}(-5~{}\text{MeV})=\mathcal{R}(-10~{}\text{MeV})=0.62,\quad\mathcal{R}(-15~{}\text{MeV})=0.67.$ Obviously, our results are consistent with the estimation in the heavy quark limit. ## IV Summary Searching for exotic hadronic state is an interesting and important research topic of hadron physics. With accumulation of experimental data, the LHCb observed three narrow $P_{c}(4312)$, $P_{c}(4440)$, and $P_{c}(4457)$ in 2019 Aaij:2019vzc , and found the evidence of the $P_{cs}(4459)$ as a hidden-charm pentaquark with strangeness Aaij:2020gdg . These progresses make us have reason to believe that there should exist a zoo of the hidden-charm molecular pentaquark. At present, the hidden-charm molecular pentaquark with triple strangeness is still missing, which inspires our interest in exploring how to find these intriguing hidden-charm molecular pentaquark states with triple strangeness. Mass spectrum information is crucial to searching for them. In this work, we perform the dynamical calculation of the possible hidden-charm molecular pentaquark states with triple strangeness from the $\Omega_{c}^{(*)}\bar{D}_{s}^{(*)}$ interactions, where the effective potentials can be obtained by the OBE model. By finding bound state solutions of these discussed systems, we find that the most promising hidden-charm molecular pentaquarks with triple strangeness are the $S$-wave $\Omega_{c}\bar{D}_{s}^{*}$ state with $J^{P}={3}/{2}^{-}$ and the $S$-wave $\Omega_{c}^{*}\bar{D}_{s}^{*}$ state with $J^{P}=5/2^{-}$. Besides mass spectrum study, we also discuss their two-body hidden-charm strong decay behaviors within the quark-interchange model. In concrete calculation, we mainly focus on the $J/\psi\,\Omega$ and $\eta_{c}\,\Omega$ decay modes for the predicted $S$-wave $\Omega_{c}\bar{D}_{s}^{*}$ molecule with $J^{P}={3}/{2}^{-}$ and the $J/\psi\,\Omega$ decay channel for the predicted $S$-wave $\Omega_{c}^{*}\bar{D}_{s}^{*}$ molecule with $J^{P}={5}/{2}^{-}$. In the following years, the LHCb Collaboration will collect more experimental data at Run III and upgrade the High-Luminosity-LHC Bediaga:2018lhg . Experimental searches for these predicted hidden-charm molecular pentaquarks with triple strangeness are an area full of opportunities and challenges in future experiments. ## ACKNOWLEDGMENTS We would like to thank Z. W. Liu, G. J. Wang, L. Y. Xiao, and S. Q. Luo for very helpful discussions. This work is supported by the China National Funds for Distinguished Young Scientists under Grant No. 11825503, National Key Research and Development Program of China under Contract No. 2020YFA0406400, the 111 Project under Grant No. B20063, and the National Natural Science Foundation of China under Grant No. 12047501. R. C. is supported by the National Postdoctoral Program for Innovative Talent. ## Appendix A Relevant subpotentials Through the standard strategy Wang:2020dya ; Wang:2019nwt ; Wang:2019aoc , we can derive the effective potentials in the coordinate space for these investigated $\Omega_{c}^{(*)}\bar{D}_{s}^{(*)}$ systems, i.e., $\displaystyle\mathcal{V}^{\Omega_{c}\bar{D}_{s}\rightarrow\Omega_{c}\bar{D}_{s}}$ $\displaystyle=$ $\displaystyle-AY_{f_{0}}-\frac{C}{2}Y_{\phi},$ (1.1) $\displaystyle\mathcal{V}^{\Omega_{c}^{*}\bar{D}_{s}\rightarrow\Omega_{c}^{*}\bar{D}_{s}}$ $\displaystyle=$ $\displaystyle-A\mathcal{A}_{1}Y_{f_{0}}-\frac{C}{2}\mathcal{A}_{1}Y_{\phi},$ (1.2) $\displaystyle\mathcal{V}^{\Omega_{c}\bar{D}_{s}^{*}\rightarrow\Omega_{c}\bar{D}_{s}^{*}}$ $\displaystyle=$ $\displaystyle-A\mathcal{A}_{2}Y_{f_{0}}+\frac{2B}{9}\left[\mathcal{A}_{3}\mathcal{O}_{r}+\mathcal{A}_{4}\mathcal{P}_{r}\right]Y_{\eta}$ (1.3) $\displaystyle-\frac{C}{2}\mathcal{A}_{2}Y_{\phi}-\frac{2D}{9}\left[2\mathcal{A}_{3}\mathcal{O}_{r}-\mathcal{A}_{4}\mathcal{P}_{r}\right]Y_{\phi},$ $\displaystyle\mathcal{V}^{\Omega_{c}^{*}\bar{D}_{s}^{*}\rightarrow\Omega_{c}^{*}\bar{D}_{s}^{*}}$ $\displaystyle=$ $\displaystyle-A\mathcal{A}_{5}Y_{f_{0}}-\frac{B}{3}\left[\mathcal{A}_{6}\mathcal{O}_{r}+\mathcal{A}_{7}\mathcal{P}_{r}\right]Y_{\eta}$ (1.4) $\displaystyle-\frac{C}{2}\mathcal{A}_{5}Y_{\phi}+\frac{D}{3}\left[2\mathcal{A}_{6}\mathcal{O}_{r}-\mathcal{A}_{7}\mathcal{P}_{r}\right]Y_{\phi},$ $\displaystyle\mathcal{V}^{\Omega_{c}\bar{D}_{s}\rightarrow\Omega_{c}^{*}\bar{D}_{s}}$ $\displaystyle=$ $\displaystyle\frac{A}{\sqrt{3}}\mathcal{A}_{8}Y_{f_{0}1}+\frac{C}{2\sqrt{3}}\mathcal{A}_{8}Y_{\phi 1},$ (1.5) $\displaystyle\mathcal{V}^{\Omega_{c}\bar{D}_{s}\rightarrow\Omega_{c}\bar{D}_{s}^{*}}$ $\displaystyle=$ $\displaystyle\frac{2B}{9}\left[\mathcal{A}_{9}\mathcal{O}_{r}+\mathcal{A}_{10}\mathcal{P}_{r}\right]Y_{\eta 2}$ (1.6) $\displaystyle+\frac{2D}{9}\left[2\mathcal{A}_{9}\mathcal{O}_{r}-\mathcal{A}_{10}\mathcal{P}_{r}\right]Y_{\phi 2},$ $\displaystyle\mathcal{V}^{\Omega_{c}\bar{D}_{s}\rightarrow\Omega_{c}^{*}\bar{D}_{s}^{*}}$ $\displaystyle=$ $\displaystyle-\frac{B}{3\sqrt{3}}\left[\mathcal{A}_{11}\mathcal{O}_{r}+\mathcal{A}_{12}\mathcal{P}_{r}\right]Y_{\eta 3}$ (1.7) $\displaystyle-\frac{D}{3\sqrt{3}}\left[2\mathcal{A}_{11}\mathcal{O}_{r}-\mathcal{A}_{12}\mathcal{P}_{r}\right]Y_{\phi 3},$ $\displaystyle\mathcal{V}^{\Omega_{c}^{*}\bar{D}_{s}\rightarrow\Omega_{c}\bar{D}_{s}^{*}}$ $\displaystyle=$ $\displaystyle\frac{B}{3\sqrt{3}}\left[\mathcal{A}_{13}\mathcal{O}_{r}+\mathcal{A}_{14}\mathcal{P}_{r}\right]Y_{\eta 4}$ (1.8) $\displaystyle+\frac{D}{3\sqrt{3}}\left[2\mathcal{A}_{13}\mathcal{O}_{r}-\mathcal{A}_{14}\mathcal{P}_{r}\right]Y_{\phi 4},$ $\displaystyle\mathcal{V}^{\Omega_{c}^{*}\bar{D}_{s}\rightarrow\Omega_{c}^{*}\bar{D}_{s}^{*}}$ $\displaystyle=$ $\displaystyle\frac{B}{3}\left[\mathcal{A}_{15}\mathcal{O}_{r}+\mathcal{A}_{16}\mathcal{P}_{r}\right]Y_{\eta 5}$ (1.9) $\displaystyle+\frac{D}{3}\left[2\mathcal{A}_{15}\mathcal{O}_{r}-\mathcal{A}_{16}\mathcal{P}_{r}\right]Y_{\phi 5},$ $\displaystyle\mathcal{V}^{\Omega_{c}\bar{D}_{s}^{*}\rightarrow\Omega_{c}^{*}\bar{D}_{s}^{*}}$ $\displaystyle=$ $\displaystyle\frac{A}{\sqrt{3}}\mathcal{A}_{17}Y_{f_{0}6}+\frac{B}{3\sqrt{3}}\left[\mathcal{A}_{18}\mathcal{O}_{r}+\mathcal{A}_{19}\mathcal{P}_{r}\right]Y_{\eta 6}$ $\displaystyle+\frac{C\mathcal{A}_{17}}{2\sqrt{3}}Y_{\phi 6}-\frac{D}{3\sqrt{3}}\left[2\mathcal{A}_{18}\mathcal{O}_{r}-\mathcal{A}_{19}\mathcal{P}_{r}\right]Y_{\phi 6}.$ Here, $\mathcal{O}_{r}=\frac{1}{r^{2}}\frac{\partial}{\partial r}r^{2}\frac{\partial}{\partial r}$ and $\mathcal{P}_{r}=r\frac{\partial}{\partial r}\frac{1}{r}\frac{\partial}{\partial r}$. Additionally, we also define several variables, which include $A=l_{S}g_{S}$, $B=g_{1}g/f_{\pi}^{2}$, $C=\beta_{S}\beta g_{V}^{2}$, and $D=\lambda_{S}\lambda g_{V}^{2}$. The function $Y_{i}$ can be defined as $\displaystyle Y_{i}=\dfrac{e^{-m_{i}r}-e^{-\Lambda_{i}r}}{4\pi r}-\dfrac{\Lambda_{i}^{2}-m_{i}^{2}}{8\pi\Lambda_{i}}e^{-\Lambda_{i}r}.$ (1.11) Here, $m_{i}=\sqrt{m^{2}-q_{i}^{2}}$ and $\Lambda_{i}=\sqrt{\Lambda^{2}-q_{i}^{2}}$. Variables $q_{i}\,(i=1\,,...,\,6)$ are defined as $q_{1}=0.04$ GeV, $q_{2}=0.06$ GeV, $q_{3}=0.02$ GeV, $q_{4}=0.10$ GeV, $q_{5}=0.06$ GeV, and $q_{6}=0.04$ GeV. In the above effective potentials, we also introduce several operators, i.e., $\displaystyle\mathcal{A}_{1}$ $\displaystyle=$ $\displaystyle\sum_{a,b,m,n}C^{\frac{3}{2},a+b}_{\frac{1}{2}a,1b}C^{\frac{3}{2},m+n}_{\frac{1}{2}m,1n}\chi^{\dagger a}_{3}\left({\bm{\epsilon}^{\dagger b}_{3}}\cdot{\bm{\epsilon}^{n}_{1}}\right)\chi^{m}_{1},$ $\displaystyle\mathcal{A}_{2}$ $\displaystyle=$ $\displaystyle\chi^{\dagger}_{3}\left({\bm{\epsilon}^{\dagger}_{4}}\cdot{\bm{\epsilon}_{2}}\right)\chi_{1},$ $\displaystyle\mathcal{A}_{3}$ $\displaystyle=$ $\displaystyle\chi^{\dagger}_{3}\left[{\bm{\sigma}}\cdot\left(i{\bm{\epsilon}_{2}}\times{\bm{\epsilon}^{\dagger}_{4}}\right)\right]\chi_{1},$ $\displaystyle\mathcal{A}_{4}$ $\displaystyle=$ $\displaystyle\chi^{\dagger}_{3}T({\bm{\sigma}},i{\bm{\epsilon}_{2}}\times{\bm{\epsilon}^{\dagger}_{4}})\chi_{1},$ $\displaystyle\mathcal{A}_{5}$ $\displaystyle=$ $\displaystyle\sum_{a,b,m,n}C^{\frac{3}{2},a+b}_{\frac{1}{2}a,1b}C^{\frac{3}{2},m+n}_{\frac{1}{2}m,1n}\chi^{\dagger a}_{3}\left({\bm{\epsilon}^{n}_{1}}\cdot{\bm{\epsilon}^{\dagger b}_{3}}\right)\left({\bm{\epsilon}_{2}}\cdot{\bm{\epsilon}^{\dagger}_{4}}\right)\chi^{m}_{1},$ $\displaystyle\mathcal{A}_{6}$ $\displaystyle=$ $\displaystyle\sum_{a,b,m,n}C^{\frac{3}{2},a+b}_{\frac{1}{2}a,1b}C^{\frac{3}{2},m+n}_{\frac{1}{2}m,1n}\chi^{\dagger a}_{3}\left({\bm{\epsilon}^{n}_{1}}\times{\bm{\epsilon}^{\dagger b}_{3}}\right)\cdot\left({\bm{\epsilon}_{2}}\times{\bm{\epsilon}^{\dagger}_{4}}\right)\chi^{m}_{1},$ $\displaystyle\mathcal{A}_{7}$ $\displaystyle=$ $\displaystyle\sum_{a,b,m,n}C^{\frac{3}{2},a+b}_{\frac{1}{2}a,1b}C^{\frac{3}{2},m+n}_{\frac{1}{2}m,1n}\chi^{\dagger a}_{3}T({\bm{\epsilon}^{n}_{1}}\times{\bm{\epsilon}^{\dagger b}_{3}},{\bm{\epsilon}_{2}}\times{\bm{\epsilon}^{\dagger}_{4}})\chi^{m}_{1},$ $\displaystyle\mathcal{A}_{8}$ $\displaystyle=$ $\displaystyle\sum_{a,b}C^{\frac{3}{2},a+b}_{\frac{1}{2}a,1b}\chi^{\dagger a}_{3}\left({\bm{\epsilon}^{\dagger b}_{3}}\cdot{\bm{\sigma}}\right)\chi_{1},$ $\displaystyle\mathcal{A}_{9}$ $\displaystyle=$ $\displaystyle\chi^{\dagger}_{3}\left({\bm{\sigma}}\cdot{\bm{\epsilon}^{\dagger}_{4}}\right)\chi_{1},$ $\displaystyle\mathcal{A}_{10}$ $\displaystyle=$ $\displaystyle\chi^{\dagger}_{3}T({\bm{\sigma}},{\bm{\epsilon}^{\dagger}_{4}})\chi_{1},$ $\displaystyle\mathcal{A}_{11}$ $\displaystyle=$ $\displaystyle\sum_{a,b}C^{\frac{3}{2},a+b}_{\frac{1}{2}a,1b}\chi^{\dagger a}_{3}\left[{\bm{\epsilon}^{\dagger}_{4}}\cdot\left(i{\bm{\sigma}}\times{\bm{\epsilon}^{\dagger b}_{3}}\right)\right]\chi_{1},$ $\displaystyle\mathcal{A}_{12}$ $\displaystyle=$ $\displaystyle\sum_{a,b}C^{\frac{3}{2},a+b}_{\frac{1}{2}a,1b}\chi^{\dagger a}_{3}T({\bm{\epsilon}^{\dagger}_{4}},i{\bm{\sigma}}\times{\bm{\epsilon}^{\dagger b}_{3}})\chi_{1},$ $\displaystyle\mathcal{A}_{13}$ $\displaystyle=$ $\displaystyle\sum_{a,b}C^{\frac{3}{2},a+b}_{\frac{1}{2}a,1b}\chi^{\dagger}_{3}\left[{\bm{\epsilon}^{\dagger}_{4}}\cdot\left(i{\bm{\sigma}}\times{\bm{\epsilon}^{b}_{1}}\right)\right]\chi^{a}_{1},$ $\displaystyle\mathcal{A}_{14}$ $\displaystyle=$ $\displaystyle\sum_{a,b}C^{\frac{3}{2},a+b}_{\frac{1}{2}a,1b}\chi^{\dagger}_{3}T({\bm{\epsilon}^{\dagger}_{4}},i{\bm{\sigma}}\times{\bm{\epsilon}^{b}_{1}})\chi^{a}_{1},$ $\displaystyle\mathcal{A}_{15}$ $\displaystyle=$ $\displaystyle\sum_{a,b,m,n}C^{\frac{3}{2},a+b}_{\frac{1}{2}a,1b}C^{\frac{3}{2},m+n}_{\frac{1}{2}m,1n}\chi^{\dagger a}_{3}\left[{\bm{\epsilon}^{\dagger}_{4}}\cdot\left(i{\bm{\epsilon}^{n}_{1}}\times{\bm{\epsilon}^{\dagger b}_{3}}\right)\right]\chi^{m}_{1},$ $\displaystyle\mathcal{A}_{16}$ $\displaystyle=$ $\displaystyle\sum_{a,b,m,n}C^{\frac{3}{2},a+b}_{\frac{1}{2}a,1b}C^{\frac{3}{2},m+n}_{\frac{1}{2}m,1n}\chi^{\dagger a}_{3}T({\bm{\epsilon}^{\dagger}_{4}},i{\bm{\epsilon}^{n}_{1}}\times{\bm{\epsilon}^{\dagger b}_{3}})\chi^{m}_{1},$ $\displaystyle\mathcal{A}_{17}$ $\displaystyle=$ $\displaystyle\sum_{a,b}C^{\frac{3}{2},a+b}_{\frac{1}{2}a,1b}\chi^{\dagger a}_{3}\left({\bm{\sigma}}\cdot{\bm{\epsilon}^{\dagger b}_{3}}\right)\left({\bm{\epsilon}_{2}}\cdot{\bm{\epsilon}^{\dagger}_{4}}\right)\chi_{1},$ $\displaystyle\mathcal{A}_{18}$ $\displaystyle=$ $\displaystyle\sum_{a,b}C^{\frac{3}{2},a+b}_{\frac{1}{2}a,1b}\chi^{\dagger a}_{3}\left({\bm{\sigma}}\times{\bm{\epsilon}^{\dagger b}_{3}}\right)\cdot\left({\bm{\epsilon}_{2}}\times{\bm{\epsilon}^{\dagger}_{4}}\right)\chi_{1},$ $\displaystyle\mathcal{A}_{19}$ $\displaystyle=$ $\displaystyle\sum_{a,b}C^{\frac{3}{2},a+b}_{\frac{1}{2}a,1b}\chi^{\dagger a}_{3}T({\bm{\sigma}}\times{\bm{\epsilon}^{\dagger b}_{3}},{\bm{\epsilon}_{2}}\times{\bm{\epsilon}^{\dagger}_{4}})\chi_{1}.$ (1.12) Here, $T({\bm{x}},{\bm{y}})=3\left(\hat{\bm{r}}\cdot{\bm{x}}\right)\left(\hat{\bm{r}}\cdot{\bm{y}}\right)-{\bm{x}}\cdot{\bm{y}}$ is the tensor force operator. In Table 7, we collect the numerical matrix elements $\langle f|\mathcal{A}_{k}|i\rangle\,(k=1\,,...,\,7)$ with the $S$-$D$ wave mixing effect analysis. Of course, the relevant numerical matrix elements $\langle f|\mathcal{A}_{k}|i\rangle\,(k=8\,,...,\,19)$ will be involved in the coupled channel analysis. For the coupled channel analysis with $J=1/2$, we have $\mathcal{A}_{9}=\sqrt{3}$, $\mathcal{A}_{11}=\sqrt{2}$, $\mathcal{A}_{18}=-\sqrt{{2}/{3}}$, and $\mathcal{A}_{k}=0\,(k=10,\,12,\,17,\,19)$. And, there exists $\mathcal{A}_{13}=1$, $\mathcal{A}_{15}=\sqrt{{5}/{3}}$, $\mathcal{A}_{18}=-\sqrt{{5}/{3}}$, and $\mathcal{A}_{k}=0\,(k=14,\,16,\,17,\,19)$ for the coupled channel analysis with $J=3/2$. Table 7: The numerical matrix elements $\langle f|\mathcal{A}_{k}|i\rangle\,(k=1\,,...,\,7)$ with the $S$-$D$ wave mixing effect analysis. Matrix elements | $J=1/2$ | $J=3/2$ | $J=5/2$ ---|---|---|--- $\langle\Omega_{c}^{*}\bar{D}_{s}|\mathcal{A}_{1}|\Omega_{c}^{*}\bar{D}_{s}\rangle$ | $/$ | diag(1,1) | $/$ $\langle\Omega_{c}\bar{D}_{s}^{*}|\mathcal{A}_{2}|\Omega_{c}\bar{D}_{s}^{*}\rangle$ | diag(1,1) | diag(1,1,1) | $/$ $\langle\Omega_{c}\bar{D}_{s}^{*}|\mathcal{A}_{3}|\Omega_{c}\bar{D}_{s}^{*}\rangle$ | diag($-2$,$1$) | diag($1$,$-2$,$1$) | $/$ $\langle\Omega_{c}\bar{D}_{s}^{*}|\mathcal{A}_{4}|\Omega_{c}\bar{D}_{s}^{*}\rangle$ | $\left(\begin{array}[]{cc}0&-\sqrt{2}\\\ -\sqrt{2}&-2\end{array}\right)$ | $\left(\begin{array}[]{ccc}0&1&2\\\ 1&0&-1\\\ 2&-1&0\end{array}\right)$ | $/$ $\langle\Omega_{c}^{*}\bar{D}_{s}^{*}|\mathcal{A}_{5}|\Omega_{c}^{*}\bar{D}_{s}^{*}\rangle$ | diag(1,1,1) | diag(1,1,1,1) | diag(1,1,1,1) $\langle\Omega_{c}^{*}\bar{D}_{s}^{*}|\mathcal{A}_{6}|\Omega_{c}^{*}\bar{D}_{s}^{*}\rangle$ | diag($\frac{5}{3}$,$\frac{2}{3}$,$-1$) | diag($\frac{2}{3}$,$\frac{5}{3}$,$\frac{2}{3}$,$-1$) | diag($-1$,$\frac{5}{3}$,$\frac{2}{3}$,$-1$) $\langle\Omega_{c}^{*}\bar{D}_{s}^{*}|\mathcal{A}_{7}|\Omega_{c}^{*}\bar{D}_{s}^{*}\rangle$ | $\left(\begin{array}[]{ccc}0&-\frac{7}{3\sqrt{5}}&\frac{2}{\sqrt{5}}\\\ -\frac{7}{3\sqrt{5}}&\frac{16}{15}&-\frac{1}{5}\\\ \frac{2}{\sqrt{5}}&-\frac{1}{5}&\frac{8}{5}\end{array}\right)$ | $\left(\begin{array}[]{cccc}0&\frac{7}{3\sqrt{10}}&-\frac{16}{15}&-\frac{\sqrt{7}}{5\sqrt{2}}\\\ \frac{7}{3\sqrt{10}}&0&-\frac{7}{3\sqrt{10}}&-\frac{2}{\sqrt{35}}\\\ -\frac{16}{15}&-\frac{7}{3\sqrt{10}}&0&-\frac{1}{\sqrt{14}}\\\ -\frac{\sqrt{7}}{5\sqrt{2}}&-\frac{2}{\sqrt{35}}&-\frac{1}{\sqrt{14}}&\frac{4}{7}\end{array}\right)$ | $\left(\begin{array}[]{cccc}0&\frac{2}{\sqrt{15}}&\frac{\sqrt{7}}{5\sqrt{3}}&-\frac{2\sqrt{14}}{5}\\\ \frac{2}{\sqrt{15}}&0&\frac{\sqrt{7}}{3\sqrt{5}}&-\frac{4\sqrt{2}}{\sqrt{105}}\\\ \frac{\sqrt{7}}{5\sqrt{3}}&\frac{\sqrt{7}}{3\sqrt{5}}&-\frac{16}{21}&-\frac{\sqrt{2}}{7\sqrt{3}}\\\ -\frac{2\sqrt{14}}{5}&-\frac{4\sqrt{2}}{\sqrt{105}}&-\frac{\sqrt{2}}{7\sqrt{3}}&-\frac{4}{7}\end{array}\right)$ ## References * (1) S. K. Choi et al. (Belle Collaboration), Observation of a Narrow Charmonium-Like State in Exclusive $B^{\pm}\to K^{\pm}\pi^{+}\pi^{-}J/\psi$ Decays, Phys. Rev. Lett. 91, 262001 (2003). * (2) H. X. Chen, W. Chen, X. Liu, and S. L. Zhu, The hidden-charm pentaquark and tetraquark states, Phys. Rep. 639, 1 (2016). * (3) Y. R. Liu, H. X. Chen, W. Chen, X. Liu, and S. L. Zhu, Pentaquark and tetraquark states, Prog. Part. Nucl. Phys. 107, 237 (2019). * (4) S. L. Olsen, T. Skwarnicki, and D. Zieminska, Nonstandard heavy mesons and baryons: Experimental evidence, Rev. Mod. Phys. 90, 015003 (2018). * (5) F. K. Guo, C. Hanhart, U. G. Mei$\ss$ner, Q. Wang, Q. Zhao, and B. S. Zou, Hadronic molecules, Rev. Mod. Phys. 90, 015004 (2018). * (6) X. Liu, An overview of $XYZ$ new particles, Chin. Sci. Bull. 59, 3815 (2014). * (7) A. Hosaka, T. Iijima, K. Miyabayashi, Y. Sakai, and S. Yasui, Exotic hadrons with heavy flavors: $X$, $Y$, $Z$, and related states, Prog. Theor. Exp. Phys. 2016, 062C01 (2016). * (8) N. Brambilla, S. Eidelman, C. Hanhart, A. Nefediev, C. P. Shen, C. E. Thomas, A. Vairo, and C. Z. Yuan, The $XYZ$ states: Experimental and theoretical status and perspectives, Phys. Rep. 873, 1 (2020). * (9) M. Gell-Mann, A schematic model of baryons and mesons, Phys. Lett. 8, 214 (1964). * (10) G. Zweig, An SU(3) model for strong interaction symmetry and its breaking. Version 1, CERN, Report No. CERN-TH-401, 1964. * (11) J. J. Wu, R. Molina, E. Oset, and B. S. Zou, Prediction of Narrow $N^{*}$ and $\Lambda^{*}$ Resonances with Hidden Charm Above 4 GeV, Phys. Rev. Lett. 105, 232001 (2010). * (12) W. L. Wang, F. Huang, Z. Y. Zhang, and B. S. Zou, $\Sigma_{c}\bar{D}$ and $\Lambda_{c}\bar{D}$ states in a chiral quark model, Phys. Rev. C 84, 015203 (2011). * (13) Z. C. Yang, Z. F. Sun, J. He, X. Liu, and S. L. Zhu, The possible hidden-charm molecular baryons composed of anticharmed meson and charmed baryon, Chin. Phys. C 36, 6 (2012). * (14) J. J. Wu, T.-S. H. Lee, and B. S. Zou, Nucleon resonances with hidden charm in coupled-channel Models, Phys. Rev. C 85, 044002 (2012). * (15) X. Q. Li and X. Liu, A possible global group structure for exotic states, Eur. Phys. J. C 74, 3198 (2014). * (16) R. Chen, X. Liu, X. Q. Li, and S. L. Zhu, Identifying Exotic Hidden-Charm Pentaquarks, Phys. Rev. Lett. 115, 132002 (2015). * (17) M. Karliner and J. L. Rosner, New Exotic Meson and Baryon Resonances from Doubly-Heavy Hadronic Molecules, Phys. Rev. Lett. 115, 122001 (2015). * (18) R. Aaij et al. (LHCb Collaboration), Observation of $J/\psi$ Resonances Consistent with Pentaquark States in $\Lambda_{b}^{0}\rightarrow J/\psi K^{-}p$ Decays, Phys. Rev. Lett. 115, 072001 (2015). * (19) R. Aaij et al. (LHCb Collaboration), Observation of a Narrow Pentaquark State, $P_{c}(4312)^{+}$, and of Two-Peak Structure of the $P_{c}(4450)^{+}$, Phys. Rev. Lett. 122, 222001 (2019). * (20) R. Aaij et al. (LHCb Collaboration), Evidence of a $J/\psi\varLambda$ structure and observation of excited $\Xi^{-}$ states in the $\Xi_{b}^{-}\to J/\psi\varLambda K^{-}$ decay, arXiv:2012.10380. * (21) J. Hofmann and M. F. M. Lutz, Coupled-channel study of crypto-exotic baryons with charm, Nucl. Phys. A 763, 90 (2005). * (22) J. J. Wu, R. Molina, E. Oset and B. S. Zou, Dynamically generated $N^{*}$ and $\Lambda^{*}$ resonances in the hidden charm sector around 4.3 GeV, Phys. Rev. C 84, 015202 (2011). * (23) V. V. Anisovich, M. A. Matveev, J. Nyiri, A. V. Sarantsev, and A. N. Semenova, Nonstrange and strange pentaquarks with hidden charm, Int. J. Mod. Phys. A 30, 1550190 (2015). * (24) Z. G. Wang, Analysis of the ${\frac{1}{2}}^{\pm}$ pentaquark states in the diquark-diquark-antiquark model with QCD sum rules, Eur. Phys. J. C 76, 142 (2016). * (25) A. Feijoo, V. K. Magas, A. Ramos, and E. Oset, A hidden-charm $S=-1$ pentaquark from the decay of $\Lambda_{b}$ into $J/\psi\eta\Lambda$ states, Eur. Phys. J. C 76, no. 8, 446 (2016). * (26) J. X. Lu, E. Wang, J. J. Xie, L. S. Geng, and E. Oset, The $\Lambda_{b}\rightarrow J/\psi K^{0}\Lambda$ reaction and a hidden-charm pentaquark state with strangeness, Phys. Rev. D 93, 094009 (2016). * (27) H. X. Chen, L. S. Geng, W. H. Liang, E. Oset, E. Wang, and J. J. Xie, Looking for a hidden-charm pentaquark state with strangeness $S=-1$ from $\Xi^{-}_{b}$ decay into $J/\psi K^{-}\Lambda$, Phys. Rev. C 93, 065203 (2016). * (28) R. Chen, J. He, and X. Liu, Possible strange hidden-charm pentaquarks from $\Sigma_{c}^{(*)}\bar{D}_{s}^{*}$ and $\Xi^{(^{\prime},*)}_{c}\bar{D}^{*}$ interactions, Chin. Phys. C 41, 103105 (2017). * (29) C. W. Xiao, J. Nieves, and E. Oset, Prediction of hidden charm strange molecular baryon states with heavy quark spin symmetry, Phys. Lett. B 799, 135051 (2019). * (30) C. W. Shen, H. J. Jing, F. K. Guo, and J. J. Wu, Exploring possible triangle singularities in the $\Xi^{-}_{b}\to K^{-}J/\psi\Lambda$ decay, Symmetry 12, 1611 (2020). * (31) B. Wang, L. Meng, and S. L. Zhu, Spectrum of the strange hidden charm molecular pentaquarks in chiral effective field theory, Phys. Rev. D 101, 034018 (2020). * (32) Q. Zhang, B. R. He, and J. L. Ping, Pentaquarks with the $qqs\bar{Q}Q$ configuration in the Chiral Quark Model, arXiv:2006.01042. * (33) H. X. Chen, W. Chen, X. Liu, and X. H. Liu, Establishing the first hidden-charm pentaquark with strangeness, arXiv:2011.01079. * (34) F. Z. Peng, M. J. Yan, M. Sánchez Sánchez, and M. P. Valderrama, The $P_{cs}(4459)$ pentaquark from a combined effective field theory and phenomenological perspectives, arXiv:2011.01915. * (35) R. Chen, Can the newly $P_{cs}(4459)$ be a strange hidden-charm $\Xi_{c}\bar{D}^{*}$ molecular pentaquarks?, arXiv:2011.07214. * (36) H. X. Chen, Hidden-charm pentaquark states through the current algebra: From their productions to decays, arXiv:2011.07187. * (37) M. Z. Liu, Y. W. Pan, and L. S. Geng, Can discovery of hidden charm strange pentaquark states help determine the spins of $P_{c}(4440)$ and $P_{c}(4457)$ ?, Phys. Rev. D 103, 034003 (2021). * (38) X. K. Dong, F. K. Guo, and B. S. Zou, A survey of heavy-antiheavy hadronic molecules, arXiv:2101.01021. * (39) F. L. Wang, R. Chen, and X. Liu, Prediction of hidden-charm pentaquarks with double strangeness, Phys. Rev. D 103, 034014 (2021). * (40) F. L. Wang, R. Chen, Z. W. Liu, and X. Liu, Probing new types of $P_{c}$ states inspired by the interaction between $S$-wave charmed baryon and anti-charmed meson in a $\bar{T}$ doublet, Phys. Rev. C 101, 025201 (2020). * (41) T. Barnes and E. S. Swanson, A Diagrammatic approach to meson meson scattering in the nonrelativistic quark potential model, Phys. Rev. D 46, 131 (1992). * (42) T. Barnes, N. Black, D. J. Dean, and E. S. Swanson, BB intermeson potentials in the quark model, Phys. Rev. C 60, 045202 (1999). * (43) T. Barnes, N. Black, and E. S. Swanson, Meson meson scattering in the quark model: Spin dependence and exotic channels, Phys. Rev. C 63, 025204 (2001). * (44) J. P. Hilbert, N. Black, T. Barnes, and E. S. Swanson, Charmonium-nucleon dissociation cross sections in the quark model, Phys. Rev. C 75, 064907 (2007). * (45) G. J. Wang, X. H. Liu, L. Ma, X. Liu, X. L. Chen, W. Z. Deng, and S. L. Zhu, The strong decay patterns of $Z_{c}$ and $Z_{b}$ states in the relativized quark model, Eur. Phys. J. C 79, 567 (2019). * (46) G. J. Wang, L. Y. Xiao, R. Chen, X. H. Liu, X. Liu, and S. L. Zhu, Probing hidden-charm decay properties of $P_{c}$ states in a molecular scenario, Phys. Rev. D 102, 036012 (2020). * (47) L. Y. Xiao, G. J. Wang, and S. L. Zhu, Hidden-charm strong decays of the $Z_{c}$ states, Phys. Rev. D 101, 054001 (2020). * (48) G. J. Wang, L. Meng, L. Y. Xiao, M. Oka, and S. L. Zhu, Mass spectrum and strong decays of tetraquark $\bar{c}\bar{s}qq$ states, arXiv:2010.09395. * (49) R. Aaij et al. (LHCb Collaboration), Physics case for an LHCb Upgrade II-Opportunities in flavour physics, and beyond, in the HL-LHC era, arXiv:1808.08865. * (50) M. B. Wise, Chiral perturbation theory for hadrons containing a heavy quark, Phys. Rev. D 45, R2188 (1992). * (51) R. Chen, A. Hosaka, and X. Liu, Searching for possible $\Omega_{c}$-like molecular states from meson-baryon interaction, Phys. Rev. D 97, 036016 (2018). * (52) G. J. Ding, Are $Y(4260)$ and $Z_{2}^{+}$(4250) ${\rm D_{1}D}$ or ${\rm D_{0}D^{*}}$ hadronic molecules? Phys. Rev. D 79, 014001 (2009). * (53) R. Casalbuoni, A. Deandrea, N. Di Bartolomeo, R. Gatto, F. Feruglio, and G. Nardulli, Light vector resonances in the effective chiral Lagrangian for heavy mesons, Phys. Lett. B 292, 371 (1992). * (54) R. Casalbuoni, A. Deandrea, N. Di Bartolomeo, R. Gatto, F. Feruglio, and G. Nardulli, Phenomenology of heavy meson chiral Lagrangians, Phys. Rep. 281, 145 (1997). * (55) T. M. Yan, H. Y. Cheng, C. Y. Cheung, G. L. Lin, Y. C. Lin, and H. L. Yu, Heavy quark symmetry and chiral dynamics, Phys. Rev. D 46, 1148 (1992); [Phys. Rev. D 55, 5851E (1997)]. * (56) M. Bando, T. Kugo, and K. Yamawaki, Nonlinear realization and hidden local symmetries, Phys. Rep. 164, 217 (1988). * (57) M. Harada and K. Yamawaki, Hidden local symmetry at loop: A new perspective of composite gauge boson and chiral phase transition, Phys. Rep. 381, 1 (2003). * (58) D. O. Riska and G. E. Brown, Nucleon resonance transition couplings to vector mesons, Nucl. Phys. A 679, 577 (2001). * (59) R. Chen, Z. F. Sun, X. Liu, and S. L. Zhu, Strong LHCb evidence supporting the existence of the hidden-charm molecular pentaquarks, Phys. Rev. D 100, 011502 (2019). * (60) J. He, Study of $P_{c}(4457)$, $P_{c}(4440)$, and $P_{c}(4312)$ in a quasipotential Bethe-Salpeter equation approach, Eur. Phys. J. C 79, 393 (2019). * (61) J. He, $\bar{D}\Sigma^{*}_{c}$ and $\bar{D}^{*}\Sigma_{c}$ interactions and the LHCb hidden-charmed pentaquarks, Phys. Lett. B 753, 547 (2016). * (62) R. Chen, F. L. Wang, A. Hosaka, and X. Liu, Exotic triple-charm deuteronlike hexaquarks, Phys. Rev. D 97, 114011 (2018). * (63) F. L. Wang, R. Chen, Z. W. Liu, and X. Liu, Possible triple-charm molecular pentaquarks from $\Xi_{cc}D_{1}/\Xi_{cc}D_{2}^{*}$ interactions, Phys. Rev. D 99, 054021 (2019). * (64) F. L. Wang and X. Liu, Exotic double-charm molecular states with hidden or open strangeness and around $4.5\sim 4.7$ GeV, Phys. Rev. D 102, 094006 (2020). * (65) G. Breit, The effect of retardation on the interaction of two electrons, Phys. Rev. 34, 553 (1929). * (66) G. Breit, The fine structure of HE as a test of the spin interactions of two electrons, Phys. Rev. 36, 383 (1930). * (67) N. A. Tornqvist, From the deuteron to deusons, an analysis of deuteronlike meson-meson bound states, Z. Phys. C 61, 525 (1994). * (68) N. A. Tornqvist, On deusons or deuteron-like meson-meson bound states, Nuovo Cimento Soc. Ital. Fis. 107A, 2471 (1994). * (69) P. A. Zyla et al. (Particle Data Group), Review of particle physics, PTEP 2020, no.8, 083C01 (2020). * (70) R. Chen, A. Hosaka, and X. Liu, Prediction of triple-charm molecular pentaquarks, Phys. Rev. D 96, 114030 (2017). * (71) C. Y. Wong, E. S. Swanson, and T. Barnes, Heavy quarkonium dissociation cross-sections in relativistic heavy ion collisions, Phys. Rev. C 65, 014903 (2002) Erratum: [Phys. Rev. C 66, 029901 (2002)]. * (72) S. Godfrey and N. Isgur, Mesons in a relativized quark model with chromodynamics, Phys. Rev. D 32, 189 (1985). * (73) S. Weinberg, Elementary particle theory of composite particles, Phys. Rev. 130, 776 (1963). * (74) S. Weinberg, Quasiparticles and the Born series, Phys. Rev. 131, 440 (1963). * (75) Y. H. Lin, C. W. Shen, F. K. Guo, and B. S. Zou, Decay behaviors of the $P_{c}$ hadronic molecules, Phys. Rev. D 95, 114017 (2017). * (76) Y. H. Lin, C. W. Shen, and B. S. Zou, Decay behavior of the strange and beauty partners of $P_{c}$ hadronic molecules, Nucl. Phys. A980, 21-31 (2018). * (77) Y. H. Lin and B. S. Zou, Hadronic molecular assignment for the newly observed $\Omega^{*}$ state, Phys. Rev. D 98, 056013 (2018). * (78) C. W. Shen, J. J. Wu, and B. S. Zou, Decay behaviors of possible $\Lambda_{c\bar{c}}$ states in hadronic molecule pictures, Phys. Rev. D 100, 056006 (2019). * (79) Y. H. Lin and B. S. Zou, Strong decays of the latest LHCb pentaquark candidates in hadronic molecule pictures, Phys. Rev. D 100, 056005 (2019). * (80) Y. H. Lin, F. Wang, and B. S. Zou, Reanalysis of the newly observed $\Omega^{*}$ state in a hadronic molecule model, Phys. Rev. D 102, 074025 (2020). * (81) X. K. Dong, Y. H. Lin, and B. S. Zou, Prediction of an exotic state around 4240 MeV with $J^{PC}=1^{-+}$ as $C$-parity partner of $Y(4260)$ in molecular picture, Phys. Rev. D 101, 076003 (2020). * (82) X. K. Dong and B. S. Zou, Prediction of possible $DK_{1}$ bound states, arXiv:2009.11619. * (83) D. Y. Chen, C. J. Xiao, and J. He, Hidden-charm decays of $Y(4390)$ in a hadronic molecular scenario, Phys. Rev. D 96, 054017 (2017). * (84) C. J. Xiao and D. Y. Chen, Possible $B^{(\ast)}\bar{K}$ hadronic molecule state, Eur. Phys. J. A 53, 127 (2017). * (85) C. J. Xiao, Y. Huang, Y. B. Dong, L. S. Geng, and D. Y. Chen, Exploring the molecular scenario of $P_{c}(4312)$ , $P_{c}(4440)$, and $P_{c}(4457)$, Phys. Rev. D 100, 014022 (2019). * (86) Q. Wu, D. Y. Chen, and F. K. Guo, Production of the $Z_{b}^{(\prime)}$ states from the $\Upsilon(5S,6S)$ decays, Phys. Rev. D 99, 034022 (2019).
Recent advances in meta-learning has led to remarkable performances on several few-shot learning benchmarks. However, such success often ignores the similarity between training and testing tasks, resulting in a potential bias evaluation. We, therefore, propose a generative approach based on a variant of Latent Dirichlet Allocation to analyse task similarity to optimise and better understand the performance of meta-learning. We demonstrate that the proposed method can provide an insightful evaluation for meta-learning algorithms on two few-shot classification benchmarks that matches common intuition: the more similar the higher performance. Based on this similarity measure, we propose a task-selection strategy for meta-learning and show that it can produce more accurate classification results than methods that randomly select training tasks. § INTRODUCTION The vast development in machine learning has enabled possibilities to solve increasingly complex applications. Such complexity require high capacity models, which in turn need a massive amount of annotated data for training, resulting in an arduous, costly and even infeasible annotation process. This has, therefore, motivated the research of novel learning approaches, generally known as transfer learning, that exploit past experience (in the form of models learned from other training tasks) to quickly learn a new task using relatively small training sets. Transfer-learning, and in particular, meta-learning, has recently achieved state-of-the-art results in several few-shot learning benchmarks <cit.>. Such success depends not only on the effectiveness of transfer learning algorithms, but also on the similarity between training and testing tasks <cit.>. More specifically, the larger the subset of training tasks that are similar to the testing tasks, the higher the classification accuracy on those testing tasks. However, meta-learning methods are assessed without taking into account such observation, which can bias the meta-learning classification results depending on the policy for selecting training and testing tasks. In this paper, we propose a generative approach based on a continuous version of Latent Dirichlet Co-Clustering to model classification tasks. The resulting model represents tasks in a latent task-theme simplex, and hence, allows to quantitatively measure their similarity. The proposed similarity measure enables the possibility of selecting the most related tasks from the training set for the meta-learning of a novel testing task. We empirically demonstrate that the proposed task selection strategy outperforms the one that randomly selects training tasks across several meta-learning methods. § RELATED WORK With this paper, we target an improved understanding of meta-learning algorithms <cit.>, which can allow us to improve their current performance. Although meta-learning has progressed steadily with many remarkable achievements, it has been reported that there is a large variance of performance among testing tasks <cit.>. This observation suggests that not all testing tasks are equally related to training tasks. Task hardness which is based on the cosine similarity between the embedding of labelled and unlabelled data is, therefore, proposed to better justify the performance of meta-learning methods <cit.>. This, however, quantifies only the similarity between samples within a task without investigating the similarity between tasks. Task similarity has been intensively studied in the field of multi-task learning. Some remarkable works include task-clustering using k-nearest neighbours <cit.>, modelling common prior between tasks as a mixture of distributions <cit.> with the extension using Dirichlet Process <cit.>, applying a convex formulation to either cluster <cit.> or learning task relationship through task covariance matrices <cit.>. Other approaches try to provide theoretical guarantees when learning the similarity or relationship between tasks <cit.>. Following a similar approach, an extensive experiment was carried out on 26 computer-vision tasks to determine the correlation between those tasks, also known as taskonomy <cit.>. Some recent works <cit.> take a slightly different approach by investigating the correlation of the label distribution between those tasks of interest. One commonality of those studies is their reliance on a discriminative approach, where the similarity of task-specific classifiers are used to quantify task relatedness. In addition, most of those works focus more on the conventional machine learning setting, which requires a sufficient number of labelled data on the novel tasks to perform transfer learning. In contrast, our proposal follows a generative approach which does not depend on any task-specific classifier. Our approach can also work in the few-shot setting, where only a few labelled data points from the targeted tasks are available. Another work that is slightly related to task similarity is Task2Vec <cit.>, which employs Fisher information matrix of an external network, known as probe network, to model visual tasks as fixed vectors in an embedding space, allowing to analyse and calculate task similarity. However, its application is still limited due to the need of an external network pre-trained to perform specific tasks on some standard visual data sets. Our work is also related to finite mixture models <cit.>, such as the Latent Dirichlet Allocation (LDA) <cit.>, in topic modelling which analyses and summarises text data, or in computer vision <cit.>. LDA assumes that each document within a given corpus can be represented as a finite mixture model, where its components are the latent topics shared across all documents. Training an LDA model or its variants on a large text corpus is challenging, so several approximate inference techniques have been proposed, ranging from mean-field variational inference (VI) <cit.>, collapsed Gibbs' sampling <cit.> and collapsed VI <cit.>. Furthermore, several online inference methods have been developed to increase the training efficiency for large corpora <cit.>. Our work is slightly different from the inference for conventional LDA models, where we perform online learning for Latent Dirichlet Co-clustering <cit.> – a variant of LDA – that includes the information of paragraphs into the model. In addition, our approach considers word as continuous data, instead of the discrete data represented by a bag-of-word vector generally used in topic modelling. § METHOD [scale=0.875, rounded corners=2pt, every node/.style=scale=0.875, minimum size=1.1cm] Directed acyclic graph represents the continuous LDCC that models classification tasks as a finite mixture of Gaussian distributions. To relate image classification to topic modelling, we consider a task as a document, a class as a paragraph, and an image as a word. Given these analogies, we employ the Latent Dirichlet Co-clustering (LDCC) <cit.> – a variant of LDA – to model classification tasks. The LDCC extends the conventional LDA to a hierarchical structure by including the information about paragraphs, or in our case, data classes, into the model. Since the data in classification is assumed to be continuous, the categorical word-topic distribution in the original LDCC model is replaced by a Gaussian image-theme distribution. Each classification task can be modelled as a mixture of \(L\) task-themes (corresponding to document topic in LDCC), where each task-theme is a summary of many finite mixtures of \(K\) image-themes. We can, therefore, utilise this representation, and in particular the task-theme mixture parameter to quantify the similarity between tasks. We assume that there are \(M\) classification tasks, where each task consists of \(C\) classes, and each class has \(N\) images (i.e., using meta-learning nomenclature, this represents \(M\) \(C\)-way \(N\)-shot classification tasks). For simplicity, \(C\) and \(N\) are assumed to be fixed across all tasks, but the extension of varying \(C\) and \(N\) is trivial and can be implemented straightforwardly. The process to generate classification tasks from an \(L\)-task-\(K\)-image theme model shown in  <ref> can be presented as follows: * Initialise means and precision matrices of \(K\) Gaussian image-theme \(\{\bm{\mu}_{k}, \bm{\Lambda}_{k}\}_{k=1}^{K}\), where \(\bm{\mu}_{k} \in \mathbb{R}^{D}\), and \(\bm{\Lambda}_{k} \in \mathbb{R}^{D \times D}\) is positive definite matrix * For task \(d\)-th in the collection of \(M\) tasks: * Choose a task-theme mixture: \(\bm{\phi}_{d}~\sim~\mathrm{Dirichlet}_{L} \left( \bm{\phi}; \bm{\delta}\right) \) * For the \(c\)-th class in the \(d\)-th task: * Choose a task-theme assignment: \(\mathbf{y}_{dc}~\sim~\mathrm{Categorical}(\mathbf{y}; \bm{\phi}_{d})\) * Choose an image-theme mixture: \(\bm{\theta}_{dc} \sim \mathrm{Dirichlet}_{K} \left( \bm{\theta}; \bm{\alpha}_{l} \right) \), where \({y_{dcl} = 1}\) * For image \(n\)-th in class \(c\)-th of task \(d\)-th: * Choose an image-theme assignment: \(\mathbf{z}_{dcn}~\sim~\mathrm{Categorical} \left( \mathbf{z}; \bm{\theta}_{dc} \right)\) * Choose an image: \(\mathbf{x}_{dcn} \sim \mathcal{N}\left(\mathbf{x}; \bm{\mu}_{k}, \bm{\Lambda}_{k}^{-1}\right) \), where: \(z_{dcnk} = 1\). If the \(K\) Gaussian image-themes \( \{ (\bm{\mu}_k, \bm{\Lambda}_k) \} _{k=1}^K\), and the Dirichlet concentration \(\{ \bm{\alpha} \}_{l=1}^{L}\) for each class are known, we can infer the mixture parameter \(\bm{\phi}_{d}\) based on the observed images \(\mathbf{x}_{d}\) of any arbitrary task \(d\)-th to represent that task in the latent task-theme simplex. This representation enables the possibility of performing further analysis, such as measuring distances between tasks. Hence, our objective is to learn these parameters from the \(M\) given classification tasks. In short, our objective is to maximise log-likelihood: \begin{equation} \max_{\bm{\mu}, \bm{\Sigma}, \bm{\alpha}} \ln p(\mathbf{x} | \bm{\mu}, \bm{\Sigma}, \bm{\alpha}). \label{eq:mle} \end{equation} Due to the complexity of the graphical model with latent variables as shown in  <ref>, the inference for the likelihood in (<ref>) is intractable, and therefore, the estimation must rely on approximate inference. Current methods to approximate the posterior of LDA-based models fall into two main categories: sampling <cit.> and optimisation <cit.>. Each approach has strengths and weaknesses, where the choice mostly depends on the application of interest. For the problem of task similarity where \(M\) is very large, the optimisation approach, and in particular, the mean-field VI, is preferable due to its efficiency and scalability to large data sets. In this paper, VI is used to infer the parameters of interest. The log-likelihood of interest can be lower-bounded by Jensen's inequality. The lower-bound is often known as evidence lower-bound (ELBO) and can be expressed as: \begin{equation} \begin{aligned}[b] \mathsf{L} & = \mathbb{E}_{q} \left[ \ln p(\mathbf{x}, \bm{\phi}, \mathbf{y}, \bm{\theta}, \mathbf{z} | \bm{\delta}, \bm{\alpha}, \bm{\mu}, \bm{\Sigma}) \right] - \mathbb{E}_{q} \left[ q(\bm{\phi}, \mathbf{y}, \bm{\theta}, \mathbf{z}) \right]. \end{aligned} \label{eq:elbo} \end{equation} Following the conventional variational inference for LDA <cit.>, we choose a fully factorised variational distribution \(q\) as our variational posterior: \begin{equation} \begin{aligned}[b] q(\bm{\phi}, \mathbf{y}, \bm{\theta}, \mathbf{z}) & = \prod_{d=1}^{M} q(\bm{\phi}_{d}; \bm{\lambda}_{d}) \prod_{c=1}^{C} q(\mathbf{y}_{dc}; \bm{\eta}_{dc}) \, q(\bm{\theta}_{dc}; \bm{\gamma}_{dc}) \prod_{n=1}^{N} q(\mathbf{z}_{dcn}; \mathbf{r}_{dcn}), \end{aligned} \label{eq:q} \end{equation} \begin{align*} q(\bm{\phi}_{d}; \bm{\lambda}_{d}) = \mathrm{Dirichlet}_{L} \left(\bm{\phi}_{d}; \bm{\lambda}_{d} \right) & \qquad q(\mathbf{y}_{dc}; \bm{\eta}_{dc}) = \mathrm{Categorical}\left(\mathbf{y}_{dc}; \bm{\eta}_{dc}\right) \\ q(\bm{\theta}_{dc}; \bm{\gamma}_{dc}) = \mathrm{Dirichlet}_{K} \left( \bm{\theta}_{dc}; \bm{\gamma}_{dc} \right) & \qquad q(\mathbf{z}_{dcn}; \mathbf{r}_{dcn}) = \mathrm{Categorical} \left(\mathbf{z}_{dcn}; \mathbf{r}_{dcn} \right). \end{align*} Given the variational distribution \(q\) defined in Eq. (<ref>), we can rewrite the ELBO as: \begin{equation} \begin{aligned} \mathsf{L} & = \mathbb{E}_{q} \left[ \ln p(\mathbf{x} | \mathbf{z}, \bm{\mu}, \bm{\Sigma}) + \ln p(\mathbf{z} | \bm{\theta}) + \ln p(\bm{\theta} | \mathbf{y}, \bm{\alpha}) + \textcolor{violet}{\ln p(\mathbf{y} | \bm{\phi})} + \textcolor{violet}{\ln p(\bm{\phi} | \bm{\delta})} \right.\\ & \qquad \left. - \ln q(\mathbf{z}) - \ln q(\bm{\theta}) - \textcolor{violet}{\ln q(\mathbf{y})} - \textcolor{violet}{\ln q(\bm{\phi})} \right]. \end{aligned} \label{eq:elbo_factorised} \end{equation} Comparing to the conventional LDA <cit.>, the ELBO in Eq. (<ref>) contains 4 extra terms highlighted in violet. The presence of those terms are due to the hierarchical structure of LDCC that takes the factor of classes (analogous to paragraphs) into the model. Instead of maximising likelihood, we maximise its lower-bound, resulting in an alternative objective function: \begin{equation} \max_{\bm{\mu}, \bm{\Sigma}, \bm{\alpha}} \, \, \max_{\mathbf{r}, \bm{\gamma}, \bm{\eta}, \bm{\lambda}} \mathsf{L}. \end{equation} Given the usage of prior conjugate, all of the terms in the ELBO can be evaluated straightforwardly (please refer to  <ref>). The optimisation is based on gradient, and performed in two steps, resulting in a process analogous to the expectation-maximisation (EM) algorithm. In the E-step, the task-specific variational-parameters \(\mathbf{r}, \bm{\gamma}, \bm{\eta}\) and \(\bm{\lambda}\) are iteratively updated, while holding the meta-parameters \(\bm{\mu}, \bm{\Sigma}, \bm{\alpha}\) fixed. In the M-step, the meta-parameters are updated using the values of the task-specific variational-parameters obtained in the E-step. The inference for the meta image-themes are similar to the estimation of Gaussian mixture model <cit.>. Please refer to  <ref> for more details. Conventionally, the iterative updates in the E-step and M-step require a full pass through the entire collection of tasks. This is, however, very slow and even infeasible since \(M\) is often in the magnitude of millions. We, therefore, propose an online VI inspired by the online learning for LDA <cit.> to infer the image-themes. When the \(d\)-th task is observed, we perform EM to obtain the task-specific image-themes (denoted by a tilde on top of variables) that are locally optimal for that task. The meta image-themes of interest are then updated as a weighted average between their previous values and the task-specific values: \begin{equation} \bm{\mu} \gets (1 - \rho_{d}) \bm{\mu} + \rho_{d} \Tilde{\bm{\mu}}, \quad \bm{\Sigma} \gets (1 - \rho_{d}) \bm{\Sigma} + \rho_{d} \Tilde{\bm{\Sigma}}, \quad \bm{\alpha} \gets \bm{\alpha} - \rho_{d} \mathbf{H}^{-1} \mathbf{g}, \label{eq:online_update} \end{equation} where \(\rho_{d} = (\tau_{0} + d)^{-\tau_{1}}\) with \(\tau_{0} \ge 0\) and \(\tau_{1} \in (0.5, 1]\) <cit.>, and \(\mathbf{g}\) is the gradient of \(\mathsf{L}\) w.r.t. \(\bm{\alpha}\), and \(\mathbf{H}\) is the Hessian matrix. Please refer to  <ref> for the details of the online learning algorithm. Also, instead of updating the image-themes when observing a single task, we use multiple or a mini-batch of tasks to reduce noise. The mini-batch version requires a slight modification, where we calculate the average of all task-specific image-themes for the tasks in the same mini-batch, and use that as the task-specific value to update the corresponding meta image-theme. Given the image-themes \(\{\bm{\mu}_{k}, \bm{\Sigma}_{k} \}_{k=1}^{K} \) and the Dirichlet parameter \(\{ \bm{\alpha}_{l} \}_{l=1}^{L}\), we can represent a task by its variational Dirichlet posterior of the task-topic mixing coefficients \(q(\bm{\phi}_{d}; \bm{\lambda}_{d})\) in the latent task-theme simplex. This new representation of classification tasks has two advantages comparing to the recently proposed task representation Task2Vec <cit.>: (i) it does not need any pre-trained networks, and (ii) the use of probability distribution, instead of a single value vector as in Task2Vec, allowing to include modelling uncertainty when representing tasks. In addition, we can utilise this representation to quantitatively analyse the similarity between two tasks through a divergence between \(q(\bm{\phi}_{d}; \bm{\lambda}_{d})\). Commonly, symmetric distances, such as Jensen-Shannon divergence, Hellinger distance, or earth's mover distance are employed to calculate the divergence between distributions. However, it is argued that similarity should be represented as an asymmetric measure <cit.>. This is reasonable in the context of transfer learning, since knowledge gained from learning a difficult task might significantly facilitate the learning of an easy task, but the reverse might not always have the same level of effectiveness. In light of asymmetric distance, we decide to use Kullback-Leibler (KL) divergence, denoted as \(D_{\mathrm{KL}}[. \Vert .]\). As \(D_{\mathrm{KL}} \left[ P \Vert Q \right]\) is defined as the information lost when using a code optimised for \(Q\) to encode the samples of \(P\), we, therefore, calculate \(D_{\mathrm{KL}} \left[ q(\bm{\phi}_{d}; \bm{\lambda}_{M + 1}) \Vert q(\bm{\phi}_{d}; \bm{\lambda}_{d}) \right]\), where \(d \in \{1, \ldots, M\}\), to assess how the \(d\)-th training task differs from the learning of the novel \((M + 1)\)-th task. Correlation Diagram We define a correlation diagram as a qualitative measure that represents visually the performance effectiveness for meta-learning algorithms. The diagram plots the expected classification accuracy as a function of KL divergence between testing and training tasks. Intuitively, the closer a testing task is from the training tasks, the higher the performance. Hence, we can use our proposed correlation diagram to qualitatively compare different meta-learning methods. A correlation diagram can be constructed by first calculating the average distance between each testing task, denoted as \(M + i\) subscript with \(i \in \mathbb{N}\), to all training tasks: \begin{equation*} \overline{D}_{M + i} = \frac{1}{M} \sum_{d=1}^{M} D_{\mathrm{KL}} \left[q(\bm{\phi}; \bm{\lambda}_{M + i}) \Vert q(\bm{\phi}; \bm{\lambda}_{d}) \right]. \end{equation*} The obtained average distances are then grouped into \(J\) interval bins, each of size \(\triangle_{J}~=~\max_{i} \overline{D}_{M + i} /J\). Let \(B_{j}\) with \(j \in \{1, \ldots, J\}\) be the set of testing tasks that have their average KL distances falling within the interval \(I_{j} = \left((j - 1) \triangle_{J}, j \triangle_{J} \right]\). The distance of bin \(B_{j}\) is defined as: \begin{equation*} d(B_{j}) = \frac{1}{|B_{j}|} \sum_{i \in B_{j}} \overline{D}_{M + i}. \end{equation*} Next, a model trained on the training tasks is employed to evaluate the prediction accuracy \(a^{(v)}_{i}\) on all the testing tasks to obtain the accuracy for bin \( B_{j} \): \begin{equation*} a(B_{j}) = \frac{1}{|B_{j}|} \sum_{i \in B_{j}} a^{(v)}_{i}. \end{equation*} Finally, plotting \(d(B_{j})\) against \(a(B_{j})\) gives the desired correlation diagram (e.g.,  <ref>). § EXPERIMENTS We carry out two experiments – correlation diagram and task selection – to demonstrate the capability of the proposed approach. We evaluate the proposed approach on \(n\)-way classification tasks formed from two separated data sets: Omniglot <cit.> and mini-ImageNet <cit.>. In this setting, a testing task is represented by a \(k\)-shot labelled data without the availability of unlabelled data following the transductive learning setting <cit.>. We evaluate the performance on several meta-learning algorithms, such as MAML <cit.>, Prototypical Networks <cit.>, Amortised Meta-learner (ABML) <cit.>, BMAML <cit.> and VAMPIRE <cit.>, to verify the distance-performance correlation using our proposed method. For Omniglot, we follow the pre-processing steps as in few-shot image classification without any data augmentation, and use the standard train-test split in the original paper to prevent information leakage. For mini-ImageNet, we follow the common train-test split with 80 classes for training and 20 classes for testing <cit.>. Since the dimension of raw images in mini-ImageNet is large, we employ the 640-dimensional features extracted from a wide-residual-network <cit.> to ease the calculation. We follow Algorithm <ref> in  <ref> to obtain the posterior of the image-theme using tasks in training set. We use \(L = 4\) task-themes and \(K = 8\) image-themes for both data sets. The Dirichlet distribution for task-theme mixture, \(\mathrm{Dirchlet}_{L}(\bm{\phi}_{d} | \bm{\delta}\), is chosen to be symmetric with \(\delta = 0.5\). The parameter inference, or training, is carried out with 16 images per class while varying the number of classes between 5 to 10 to fit into the memory of a Nvidia 1080 Ti GPU. The inference of the variational parameter \(\bm{\lambda}\) is done on all available labelled data in each class (\(20\) for Omniglot and \(600\) for mini-ImageNet). Note that this is used for the correlation diagram demonstration. For the task selection, this number matches the number of shots in the few-shot learning setting[Implementation can be found at <https://github.com/cnguyen10/similarity_classification_tasks>]. For the evaluation on meta-learning algorithms, we use a similar 4 convolutional module network to train on Omniglot <cit.>, while using a fully connected network with 1 hidden layer consisting of 128 units to train on the extracted features of mini-ImageNet <cit.>. Note that the numbers of tasks formed from the two data sets are very large. For Omniglot, approximate \(6.8 \times 10^{12}\) and \(10^{12}\) unique tasks can be generated from the training and testing sets, respectively. For mini-ImageNet, these numbers are slightly more manageable with about \(2.4 \times 10^{6}\) unique tasks for training, and \(15,504\) tasks for testing. To reduce the computation and facilitate the analysis, we randomly select 1 million Omniglot tasks for training, and \(20,000\) tasks for testing. For mini-ImageNet, we select 2 million tasks for training and \(15,504\) tasks for testing. §.§ Correlation Diagram Correlation diagram plots the average accuracy predicted by meta-learning algorithms as a function of the average KL divergence of each task in the testing set to all tasks in the training set on the 5-way 1-shot setting. To construct the correlation diagram, we train a continuous LDCC on \(n\)-way 16-shot setting (\(n\) varies from 5 to 10), and then infer the variational parameter \(\bm{\lambda}\) of the task-theme mixture \(\bm{\phi}\). The inferred \(\lambda\) is used to calculate the KL divergence distance between testing tasks to all training tasks. Note that the continuous LDCC is only trained on the training tasks. We then separately evaluate the performance of different meta-learning algorithms on the same 5-way 1-shot setting, and plot the correlation diagram in  <ref>. The results of the performance versus the task distance (or similarity) agree well with the common intuition: the testing tasks closer to the training tasks have higher prediction accuracy. Note that this observation is consistent across several meta-learning methods. It is also interesting to notice that some methods are more robust than others with respect to the dissimilarity between training and testing tasks. §.§ Task Selection The prediction accuracy of several meta-learning methods on 5-way 5-shot mini-ImageNet testing tasks when training tasks are pro-actively selected outperforms the un-selective approaches, and slightly better than Task2Vec. The error bars on the un-selective cases represent the 95% confident intervals calculated on the 50 trials of random task selection. MAML 5-way ProtoNet 5-way MAML 10-way ProtoNet 10-way The proposed task-selective approach outperforms the randomly chosen training tasks, and shows slightly better results than Task2Vec when varying the number of classes within a classification task as well as the number of training tasks. We show that when there is a constraint on the number of training tasks, selecting tasks based on the proposed similarity outperforms the un-selective one that randomly selects training tasks. To demonstrate, we assume that one can pick a small number of mini-ImageNet tasks from the whole training set to train a meta-learning model, and evaluate on all tasks in the testing set. In the selective case, we use the LDCC model trained on all training tasks to infer the variational mixture parameters \(\bm{\lambda}\) for all training and testing tasks. We then pick the training tasks that are closest to all the testing tasks using the proposed KL divergence, and use them to train a meta-learning model. In the un-selective case, we randomly select the same number of training tasks without measuring any similarity. We also include Task2Vec as a baseline for the selective case to compare with our proposed approach. As the experiment is based on extracted features of mini-ImageNet, it is difficult to adapt to some common pre-trained networks, which is used as a probe network in Task2Vec. To work around, we use MAML to train a fully-connected network with three hidden layers on the training set under 5-way 5-shot setting, and use the feature extractor (excluding the last layer) of this network as the probe network for Task2Vec. This modelling approach results in a 3-D Task2Vec representation which is the same dimension as \(\bm{\phi}_{d}\) in the continuous LDCC, and hence, can be compared fairly. In addition, we directly calculate the diagonal of Fisher information matrix of the probe network without using the proposed approximation in Task2Vec to reduce the complexity of hyper-parameter tuning. <ref> shows the accuracy results tested on \(15,504\) mini-ImageNet testing tasks on the 5-way 5-shot setting for models trained on 1,000 training tasks. We also report the 95% confident interval for the case of random task selection. Statistically, meta-learning methods trained on tasks selected from our proposed solution outperform the un-selective cases, and slightly better than Task2Vec, especially for the probabilistic meta-learning methods such as BMAML, ABML and VAMPIRE. To study the effects induced by the number of training tasks, and the number of ways within each task, we run an extensive experiment with a similar 5-shot setting, but varying the number of training tasks and ways, and plot the results in  <ref>. In general, the proposed approach out-performs the un-selective approach, and is slightly better than Task2Vec. Despite promising results, there are some limitations of our proposed task selection. The proposed approach requires a sufficient number of labelled data in the testing tasks. More specifically, we need 5 labelled images per class, so that the trained LDCC model can correctly infer \(\bm{\lambda}\). Further reduction in the number of labelled data in the test set might result in a poor estimation of \(\bm{\lambda}\), hindering the task selection process. This is a well-known issue in LDA and its variations, which do not work well for short texts. Nevertheless, the assumption of 5-shot setting, which shows a promising result for task selection, is still reasonable in many few-shot learning applications. § CONCLUSION We propose a generative approach based on the continuous LDCC adopted in topic modelling to model classification tasks. Under this modelling approach, a classification task can be expressed as a finite mixture model of Gaussian distributions, whose components are shared across all tasks. This new representation of classification tasks allows one to quantify the similarity between tasks through the asymmetric KL divergence. We also introduce a task selection strategy based on the proposed task similarity, and demonstrate its superiority in meta-learning comparing to the conventional approach where training tasks are randomly selected. § ACKNOWLEDGEMENT This work was supported with supercomputing resources provided by the Phoenix HPC service at the University of Adelaide. § BROADER IMPACT The proposed approach is helpful in transfer-learning tasks, especially when the amount of training data for the testing task is limited. By representing tasks in the topic space, the proposed approach allows to assess task similarity and provide insightful understanding when transfer-learning will be effective. This has the benefit of saving costs on data collection and annotation for the testing task. However, the trade-off is related to the computational cost involved in the training of the LDCC model for computing the task-to-task similarities. Missing 'biblatex' package The bibliography requires the 'biblatex' package. booktitleIEEE International Conference on Computer Vision titleTASK2VEC: Task embedding for meta-learning journaltitleJournal of Machine Learning Research titleTask clustering and gating for Bayesian multitask learning titlePattern Recognition and Machine Learning journaltitleJournal of Machine Learning Research titleLatent Dirichlet allocation booktitleArtificial Intelligence and Statistics titleOnline inference of topics with latent Dirichlet allocation booktitleInternational Conference on Learning Representations titleA closer look at few-shot classification journaltitlearXiv preprint arXiv:2003.04390 titleA new meta-baseline for few-shot learning booktitleInternational Conference on Learning Representations titleA baseline for few-shot image classification booktitleInternational Conference on Machine Learning titleModel-Agnostic Meta-Learning for Fast Adaptation of Deep Networks booktitleInternational Conference on Knowledge Discovery and Data Mining (ACM SIGKDD) titleStochastic collapsed variational Bayesian inference for latent Dirichlet allocation journaltitleNational Academy of Sciences titleFinding scientific topics booktitleAdvances in Neural Information Processing Systems titleOnline learning for latent Dirichlet allocation booktitleAdvances in Neural Information Processing Systems titleClustered multi-task learning: A convex formulation American Association for the Advancement of Science titleHuman-level concept learning through probabilistic program induction booktitleInternational Conference on Computer Vision and Pattern Recognition titleA bayesian hierarchical model for learning natural scene categories Technical report, MIT titleEstimating a Dirichlet distribution booktitleIEEE Winter Conference on Applications of Computer Vision titleUncertainty in model-agnostic meta-learning using variational inference booktitleInternational Conference on Machine Learning titleLEEP: A New Measure to Evaluate Transferability of Learned Representations Genetics Soc America titleInference of population structure using multilocus genotype data booktitleInternational Conference on Learning Representations titleAmortized Bayesian meta-learning booktitleInternational Conference on Learning Representations titleOptimization as a model for few-shot learning booktitleInternational Conference on Learning Representations titleMeta-learning with latent embedding optimization booktitleInternational Conference on Data Mining titleLatent Dirichlet co-clustering booktitleInternational Joint Conference on Artificial Intelligence titleA Principled Approach for Learning Task Similarity in Multitask Learning booktitleAdvances in Neural Information Processing Systems titlePrototypical networks for few-shot learning booktitleAdvances in Neural Information Processing Systems titleA collapsed variational Bayesian inference algorithm for latent Dirichlet allocation booktitleInternational Conference on Machine Learning titleDiscovering structure in multiple learning tasks: The TC algorithm booktitleInternational Conference on Computer Vision titleTransferability and hardness of supervised classification tasks American Psychological Association journaltitlePsychological review titleFeatures of similarity. booktitleAdvances in Neural Information Processing Systems titleMatching networks for one shot learning journaltitleJournal of Machine Learning Research titleMulti-task learning for classification with Dirichlet process priors booktitleAdvances in Neural Information Processing Systems titleBayesian Model-Agnostic Meta-Learning booktitleIEEE Conference on Computer Vision and Pattern Recognition titleTaskonomy: Disentangling task transfer learning booktitleConference on Uncertainty in Artificial Intelligence titleA convex formulation for learning task relationships in multi-task learning Missing 'biblatex' package The bibliography requires the 'biblatex' package. booktitleIEEE International Conference on Computer Vision titleTASK2VEC: Task embedding for meta-learning journaltitleJournal of Machine Learning Research titleTask clustering and gating for Bayesian multitask learning titlePattern Recognition and Machine Learning journaltitleJournal of Machine Learning Research titleLatent Dirichlet allocation booktitleArtificial Intelligence and Statistics titleOnline inference of topics with latent Dirichlet allocation booktitleInternational Conference on Learning Representations titleA closer look at few-shot classification journaltitlearXiv preprint arXiv:2003.04390 titleA new meta-baseline for few-shot learning booktitleInternational Conference on Learning Representations titleA baseline for few-shot image classification booktitleInternational Conference on Machine Learning titleModel-Agnostic Meta-Learning for Fast Adaptation of Deep Networks booktitleInternational Conference on Knowledge Discovery and Data Mining (ACM SIGKDD) titleStochastic collapsed variational Bayesian inference for latent Dirichlet allocation journaltitleNational Academy of Sciences titleFinding scientific topics booktitleAdvances in Neural Information Processing Systems titleOnline learning for latent Dirichlet allocation booktitleAdvances in Neural Information Processing Systems titleClustered multi-task learning: A convex formulation American Association for the Advancement of Science titleHuman-level concept learning through probabilistic program induction booktitleInternational Conference on Computer Vision and Pattern Recognition titleA bayesian hierarchical model for learning natural scene categories Technical report, MIT titleEstimating a Dirichlet distribution booktitleIEEE Winter Conference on Applications of Computer Vision titleUncertainty in model-agnostic meta-learning using variational inference booktitleInternational Conference on Machine Learning titleLEEP: A New Measure to Evaluate Transferability of Learned Representations Genetics Soc America titleInference of population structure using multilocus genotype data booktitleInternational Conference on Learning Representations titleAmortized Bayesian meta-learning booktitleInternational Conference on Learning Representations titleOptimization as a model for few-shot learning booktitleInternational Conference on Learning Representations titleMeta-learning with latent embedding optimization booktitleInternational Conference on Data Mining titleLatent Dirichlet co-clustering booktitleInternational Joint Conference on Artificial Intelligence titleA Principled Approach for Learning Task Similarity in Multitask Learning booktitleAdvances in Neural Information Processing Systems titlePrototypical networks for few-shot learning booktitleAdvances in Neural Information Processing Systems titleA collapsed variational Bayesian inference algorithm for latent Dirichlet allocation booktitleInternational Conference on Machine Learning titleDiscovering structure in multiple learning tasks: The TC algorithm booktitleInternational Conference on Computer Vision titleTransferability and hardness of supervised classification tasks American Psychological Association journaltitlePsychological review titleFeatures of similarity. booktitleAdvances in Neural Information Processing Systems titleMatching networks for one shot learning journaltitleJournal of Machine Learning Research titleMulti-task learning for classification with Dirichlet process priors booktitleAdvances in Neural Information Processing Systems titleBayesian Model-Agnostic Meta-Learning booktitleIEEE Conference on Computer Vision and Pattern Recognition titleTaskonomy: Disentangling task transfer learning booktitleConference on Uncertainty in Artificial Intelligence titleA convex formulation for learning task relationships in multi-task learning
# Joint Coreference Resolution and Character Linking for Multiparty Conversation Jiaxin Bai1, Hongming Zhang1, Yangqiu Song1, and Kun Xu2 1CSE, HKUST 2 Tencent AI Lab {jbai, hzhangal<EMAIL_ADDRESS><EMAIL_ADDRESS> ###### Abstract Character linking, the task of linking mentioned people in conversations to the real world, is crucial for understanding the conversations. For the efficiency of communication, humans often choose to use pronouns (e.g., “she”) or normal phrases (e.g., “that girl”) rather than named entities (e.g., “Rachel”) in the spoken language, which makes linking those mentions to real people a much more challenging than a regular entity linking task. To address this challenge, we propose to incorporate the richer context from the coreference relations among different mentions to help the linking. On the other hand, considering that finding coreference clusters itself is not a trivial task and could benefit from the global character information, we propose to jointly solve these two tasks. Specifically, we propose C2, the joint learning model of Coreference resolution and Character linking. The experimental results demonstrate that C2 can significantly outperform previous works on both tasks. Further analyses are conducted to analyze the contribution of all modules in the proposed model and the effect of all hyper- parameters. ## 1 Introduction Understanding conversations has long been one of the ultimate goals of the natural language processing community, and a critical step towards that is grounding all mentioned people to the real world. If we can achieve that, we can leverage our knowledge about these people (e.g., things that happened to them before) to better understand the conversation. On the other hand, we can also aggregate the conversation information back to our understanding about these people, which can be used for understanding future conversations that involve the same people. To simulate the real conversations and investigate the possibility for models to ground mentioned people, the character linking task was proposed Chen and Choi (2016). Specifically, it uses the transcripts of TV shows (i.e., Friends) as the conversations and asks the models to ground all person mentions to characters. Figure 1: The composition of the mentions in conversations for character grounding. Over 88% of the mentions are not named entities, which brings exceptional challenges when linking those to character entities. Even though the character linking task can be viewed as a special case of the entity linking task, it is more challenging than the ordinary entity linking task for various reasons. First, the ordinary entity linking task often aims at linking named entities to external knowledge bases such as Wikipedia, where rich information (e.g., definitions) are available. However, for the character linking task, we do not have the support of such rich knowledge base and all we have are the names of these characters and simple properties (e.g., gender) about these characters. Second, the mentions in the ordinary entity linking are mostly concepts and entities, but not pronouns. However, as shown in Figure 1, 88% of the character mentions are pronouns (e.g., “he”) or personal nouns (e.g., “that guy”) while only 12% are named entities. Figure 2: Coreference clusters can help to connect the whole conversation to provide a richer context for each mention such that we can better link them to Paul. Meanwhile, the character Pual can also provide global information to help resolve the coreference. Considering that pronouns have relatively weak semantics by themselves, to effectively ground mentions to the correct characters, we need to fully utilize the context information of the whole conversation rather than just the local context they appear in. One potential solution is using the coreference relations among different mentions as the bridge to connect the richer context. One example is shown in Figure 2. It is difficult to directly link the highlighted mentions to the character Paul based on their local context because the local context of each mention can only provide a single piece of information about its referent, e.g., “person is a male” or “the person works with Monica.” Given the coreference cluster, the mentions refer to the same person, and the pieces of information are put together to jointly determining the referent. As a result, it is easier for a model to do character linking with resolved coreference. Similar observations are also made in Chen et al. (2017). At the same time, we also noticed that coreference resolution, especially those involving pronouns, is also not trivial. As shown by the recent literature on the coreference resolution task Lee et al. (2018); Kantor and Globerson (2019), the task is still challenging for current models and the key challenge is how to utilize the global information about entities. And that is exactly what the character linking model can provide. For example, in Figure 2, it is difficult for a coreference model to correctly resolve the last mention he in the utterance given by Ross based on its local context, because another major male character (Joey) joins the conversation, which can distract and mislead the coreference model. However, if the model knows the mention he links to the character Paul and Paul works with Monica, it is easier to resolve he to some guy that Monica works with. Motivated by these observations, we propose to jointly train the Coreference resolution and Character linking tasks and name the joint model as C2. C2 adopts a transformer-based text encoder and includes a mention-level self- attention (MLSA) module that enables the model to do mention-level contextualization. Meanwhile, a joint loss function is designed and utilized so that both tasks can be jointly optimized. The experimental results demonstrate that C2 outperforms all previous work significantly on both tasks. Specifically, compared with the previous work Zhou and Choi (2018), C2 improves the performance by 15% and 26% on the coreference resolution and character linking tasks111The performance on the coreference resolution is evaluated based on the average F1 score of B3, CEAFϕ4, and BLANC. The performance on the character linking task is evaluated by the average F1 score of the micro and macro F1. respectively comparing to the previous state-of- the-art model ACNN Zhou and Choi (2018) . Further hyper-parameter and ablation studies testify the effectiveness of different components of C2 and the effect of all hyper-parameters. Our code is available at https://github.com/HKUST- KnowComp/C2. ## 2 Problem Formulations and Notations We first introduce the coreference resolution and character linking tasks as well as used notations. Given a conversation, which contains multiple utterances and $n$ character mentions $c_{1},c_{2},...,c_{n}$, and a pre- defined character set $\mathcal{Z}$, which contains $m$ characters $z_{1},z_{2},...,z_{m}$. The coreference resolution task is grouping all mentions to clusters such that all mentions in the same cluster refer to the same character. The character linking task is linking each mention to its corresponding character. ## 3 Model Figure 3: The coreference module and the linking module share the same mention representation $g^{(n)}$ as inputs. The mention representation $g^{(i)}$ are iteratively refined through the mention-level self-attention layers. The initial mention representations $g^{(0)}$ are the sum of text span representations from a pre-trained text encoder and corresponding speaker embeddings. In this section, we introduce the proposed C2 framework, which is illustrated in Figure 3. With the conversation and all mentions as input, we first encode them with a shared mention representation encoder module, which includes a pre-trained transformer text encoder and a mention-level self-attention (MLSA) module. After that, we make predictions for both tasks via two separate modules. In the end, a joint loss function is devised so that the model can be effectively trained on both tasks simultaneously. Details are as follows. ### 3.1 Mention Representation We use pre-trained language models Devlin et al. (2018); Joshi et al. (2019a) to obtain the contextualized representations for mentions. As speaker information is critical for the conversation understanding, we also include that information by appending speaker embeddings to each mention. As a result, the initial representation of mention $i$ is: $g_{i}^{(0)}=t_{start_{i}}+t_{end_{i}}+e_{speaker_{i}},$ (1) where $t_{start_{i}}$ and $t_{end_{i}}$ are the contextualized representation of the beginning and the end tokens of mention $i$, and the $e_{speaker_{i}}$ is the speaker embedding for the current speaker. Here, we omit the embeddings of inner tokens because their semantics has been effectively encoded via the language model. The speaker embeddings are randomly initialized before training. Sometimes the local context of a mention is not enough to make reasonable predictions, and it is observed that the co-occurred mentions can provide document-level context information. To refine the mention representations given the presence of other mentions in the document, we introduce the Mention-Level Self-Attention (MLSA) layer, which has $n$ layers of transformer encoder structure Vaswani et al. (2017) and is denoted as $T$. Formally, this iterative mention refinement process can be described by $\displaystyle g_{1}^{(i+1)},...,g_{k}^{(i+1)}=T(g_{1}^{(i)},...,g_{k}^{(i)}),$ (2) where $k$ indicates the number of mentions in a document, and the $g^{(i)}$ means the mention representation from the $i$-th layer of MLSA. Dataset | Episodes | Scenes | Utterances | Speakers | Mentions | Entities ---|---|---|---|---|---|--- TRN | 76 | 987 | 18,789 | 265 | 36,385 | 628 DEV | 8 | 122 | 2,142 | 48 | 3,932 | 102 TST | 13 | 192 | 3,597 | 91 | 7,050 | 165 Total | 97 | 1,301 | 24,528 | 331 | 47,367 | 781 Table 1: The detailed information about the datasets. For each season, the episode 1 to 19 are used for training, the episode 20 to 21 for development, and the remaining for testing. ### 3.2 Coreference Resolution Following the previous work Joshi et al. (2019a), we model the coreference resolution task as an antecedent finding problem. For each mention, we aim at finding one of the previous mentions that refer to the same person. If no such previous mention exists, it should be linked to the dummy mention $\varepsilon$. Thus the goal of a coreference model is to learn a distribution, $P(y_{i})$ over each antecedent for each mention $i$: $\displaystyle P(y_{i})=\frac{e^{s(i,y_{i})}}{\Sigma_{y^{\prime}\in\mathcal{Y}(i)}e^{s(i,y^{\prime})}},$ (3) where $s(i,j)$ is the score for the antecedent assignment of mention $i$ to $j$. The score $s(i,j)$ contains two parts: (1) the plausibility score of the mentions $s_{a}(i,j)$; (2) the mention score measuring the plausibility of being a proper mention $s_{m}(i)$. Formally, the $s(i,j)$ can be expressed by $\displaystyle s(i,j)$ $\displaystyle=s_{m}(i)+s_{m}(j)+s_{a}(i,j),$ (4) $\displaystyle s_{m}(i)$ $\displaystyle=FFNN_{m}(g_{i}^{(n)}),$ (5) $\displaystyle s_{a}(i,j)$ $\displaystyle=FFNN_{a}([g_{i}^{(n)},g_{j}^{(n)}]),$ (6) where $g^{(n)}$ stands for the last layer mention representation resulted from the MLSA and $FFNN$ indicates the feed-forward neural network. ### 3.3 Character Linking The character linking is formulated as a multi-class classification problem, following previous work Zhou and Choi (2018). Given the mention representations $g^{(n)}$, the linking can be done with a simple feed-forward network, denoted as $FFNN(\cdot)$. Specifically, the probability of character entity $z_{i}$ is linked with a given mention $i$ can be calculated by: $\displaystyle Q(z_{i})=Softmax(FFNN_{l}(g_{i}^{(n)}))_{z_{i}},$ (7) where the notation $(.)_{z}$ represents the $z$-th composition of a given vector. ### 3.4 Joint Learning To jointly optimize both coreference resolution and entity linking, we design a joint loss of both tasks. For coreference resolution, given the gold clusters, we minimize the negative log-likelihood of the possibility that each mention is linked to a gold antecedent. Then the coreference loss $L_{c}$ becomes $\displaystyle L_{c}=-\sum_{i=1}^{N}\log\sum_{y\in\mathcal{Y}(i)\cap GOLD(i)}P(y),$ (8) where the $GOLD(i)$ denotes the gold coreference cluster that mention $i$ belongs to. Similarly, for character linking, we minimize the negative log- likelihood of the joint probability for each mention being linked to the correct referent character: $\displaystyle L_{l}=-\sum_{i=1}^{N}\log Q(z_{i}).$ (9) Finally, the joint loss can be the arithmetic average of the coreference loss and linking loss: $\displaystyle L=\frac{1}{2}(L_{l}+L_{c}).$ (10) ## 4 Experiments In this section, we introduce the experimental details to demonstrate the effectiveness of C2. Model | B3 | CEAF$\phi 4$ | BLANC | Ave.F1 ---|---|---|---|--- Prec. | Rec. | F1 | Prec. | Rec. | F1 | Prec. | Rec. | F1 | ACNN | 84.30 | 71.90 | 77.60 | 54.50 | 71.80 | 62.00 | 84.30 | 80.40 | 82.10 | 73.96 (0.97) CorefQA (SpanBERT-Large) | 73.72 | 75.55 | 74.62 | 65.82 | 72.38 | 68.94 | 86.82 | 84.69 | 85.75 | 76.44 (0.20) C2F (BERT-Base) | 69.62 | 76.11 | 72.72 | 66.44 | 60.92 | 63.56 | 79.38 | 86.05 | 82.38 | 72.88 (0.23) C2F (BERT-Large) | 71.72 | 80.25 | 75.75 | 69.97 | 62.61 | 66.08 | 81.65 | 88.23 | 84.63 | 75.49 (0.18) C2F (SpanBERT-Base) | 72.49 | 77.88 | 75.08 | 66.00 | 64.23 | 65.10 | 81.60 | 87.43 | 84.27 | 74.81 (0.19) C2F (SpanBERT-Large) | 81.93 | 84.38 | 82.57 | 78.04 | 71.99 | 74.89 | 88.15 | 91.09 | 89.56 | 82.34 (0.17) C2 (BERT-Base) | 78.10 | 81.56 | 79.79 | 72.48 | 69.87 | 71.15 | 86.14 | 89.49 | 87.74 | 80.14 (0.21) C2 (BERT-Large) | 78.49 | 81.90 | 80.16 | 73.81 | 71.15 | 72.46 | 86.20 | 89.93 | 87.97 | 80.17 (0.23) C2 (SpanBERT-Base) | 81.18 | 83.59 | 82.36 | 73.64 | 73.09 | 73.36 | 88.06 | 91.04 | 89.49 | 81.74 (0.19) C2 (SpanBERT-Large) | 85.83 | 85.27 | 85.55 | 77.13 | 77.84 | 77.48 | 92.31 | 92.03 | 92.17 | 85.06 (0.16) Table 2: Experimental results on the coreference resolution task. The results are presented in a 2-digit decimal following previous work. Standard deviations of the average F1 scores are shown in brackets. ### 4.1 Data Description We use the latest released character identification V2.0222https://github.com/emorynlp/character-identification as the experimental dataset, and we follow the standard training, developing, and testing separation provided by the dataset. In the dataset, all mentions are annotated with their referent global entities. For example, in Figure 4, the mention I is assigned to ROSS, and the mentions mom and dad are assigned to JUDY and JACK respectively in the first utterance given by Ross. The gold coreference clusters are derived by grouping the mentions assigned to the same character entity. Statistically, the dataset includes four seasons of the TV show Friends, which contain 97 episodes, 1,301 scenes, and 24,528 utterances. In total, there are 47,367 mentions, which are assigned to 781 unique characters. The detailed statistics are shown in Table 1. Figure 4: The example annotations for character identification. The arrows in the figure are pointing from the character mentions to their referent character entities. ### 4.2 Baseline Methods The effectiveness of the joint learning model is evaluated on both the coreference resolution and character linking tasks. To fairly compare with existing models, only the singular mentions are used following the singular- only setting (S-only) in the previous work Zhou and Choi (2018). For the coreference resolution task, we compare with the following methods. * • ACNN: A CNN-based model Zhou and Choi (2018) coreference resolution model that can also produce the mention and mention-cluster embeddings at the same time. * • C2F: The end-to-end coarse-to-fine coreference model Joshi et al. (2019b) with BERT Devlin et al. (2018) or SpanBERT Joshi et al. (2019a) as the encoder. * • CorefQA: An approach that reformulates the coreference resolution problem as a question answering problem Wu et al. (2020) and being able to be benefited from fine-tuned question-answer text encoders. For the character linking task, we also include ACNN as a baseline method. Considering existing general entity linking models Kolitsas et al. (2018); van Hulst et al. (2020); Raiman and Raiman (2018); Onando Mulang et al. (2020) cannot be applied to the character linking problem because they are not designed to handle pronouns, we propose another text-span classification model with transformer encoder as another strong baseline for the character linking task. * • ACNN: A model that uses the mention and mention-cluster embeddings as input to do character linking Zhou and Choi (2018). * • BERT/SpanBERT: A text-span classification model consists of a transformer text encoder followed by a feed-forward network. ### 4.3 Evaluation Metrics We follow the previous work Zhou and Choi (2018) for the evaluation metrics. Specifically, for coreference resolution, three evaluation metrics, B3, CEAFϕ4, and BLANC, are used. The metrics are all proposed by the CoNNL’12 shared task Pradhan et al. (2012) to evaluate the output coreference cluster against the gold clusters. We follow Zhou and Choi (2018) to use BLANC Recasens and Hovy (2011) to replace MUC Vilain et al. (1995) because BLANC takes singletons into consideration but MUC does not. As for the character linking task, we use the Micro and Macro F1 scores to evaluate the multi-class classification performance. ### 4.4 Implementation Details In our experiments, we consider four different pre-trained language encoders: BERT-Base, BERT-Large, SpanBERT-Base, and SpanBERT-Large, and we use $n=2$ layers of the mention-level self-attention (MLSA). The feed-forward networks are implemented by two fully connected layers with ReLU activations. Following the previous work, Zhou and Choi (2018), the scene-level setting is used, where, each scene is regarded as a document for coreference resolution and linking. During the training, each mini-batch consists of segments obtained from a single document. The joint learning model is optimized with the Adam optimizer Kingma and Ba (2015) with an initial learning rate of 3e-5, and a warming-up rate of 10%. The model is set to be trained for 100 epochs with an early stop. All the experiments are repeated three times, and the average results are reported. Model | Ro | Ra | Ch | Mo | Jo | Ph | Em | Ri | Micro | Macro ---|---|---|---|---|---|---|---|---|---|--- ACNN | 78.3 | 86.5 | 78.8 | 81.7 | 78.3 | 88.8 | 69.2 | 83.9 | 73.7 (0.6) | 59.6 (2.3) BERT-Base | 87.4 | 89.9 | 86.6 | 88.2 | 87.1 | 91.1 | 94.3 | 62.4 | 84.0 (0.1) | 77.3 (0.2) BERT-Large | 88.2 | 89.9 | 87.9 | 88.8 | 87.7 | 93.1 | 93.5 | 68.0 | 84.8 (0.2) | 79.1 (0.2) SpanBERT-Base | 87.6 | 91.8 | 86.7 | 88.2 | 86.8 | 92.6 | 94.6 | 73.3 | 84.2 (0.1) | 77.3 (0.2) SpanBERT-Large | 90.9 | 92.8 | 88.3 | 90.3 | 90.2 | 94.3 | 94.6 | 71.7 | 85.5 (0.1) | 79.8 (0.2) C2 (BERT-Base) | 86.5 | 87.8 | 85.6 | 86.8 | 88.1 | 92.4 | 93.0 | 66.0 | 84.0 (0.1) | 78.6 (0.2) C2 (BERT-Large) | 85.9 | 90.0 | 87.3 | 86.9 | 87.2 | 93.0 | 96.1 | 66.0 | 84.9 (0.1) | 79.5 (0.2) C2(SpanBERT-Base) | 89.8 | 91.3 | 90.5 | 90.9 | 87.8 | 93.2 | 93.4 | 71.3 | 85.7 (0.1) | 81.0 (0.1) C2 (SpanBERT-Large) | 91.2 | 94.1 | 91.1 | 92.5 | 90.4 | 94.4 | 89.2 | 77.1 | 87.0 (0.1) | 81.1 (0.1) Table 3: Experimental results per character on the character linking. The results are presented in a 1-digit decimal following previous work. Standard deviations of the Micro and Macro F1 scores are shown in brackets. The names in the table are written in two-letter acronyms. Ro: Ross, Ra: Rachel, Ch: Chandler, Mo: Monica, Jo: Joey, Ph: Phoebe, Em: Emily, Ri: Richard ## 5 Results and Analysis In this section, we discuss the experimental results and present a detailed analysis. ### 5.1 Coreference Resolution Results The performances of coreference resolution models are shown in Table 2. C2 with SpanBERT-large achieves the best performance on all evaluation metrics. Comparing to the baseline ACNN model, which uses hand-crafted features, C2 uses a transformer to better encode the contextual information. Besides that, even though ACNN formulates the coreference resolution and character linking tasks in a pipe-line and uses the coreference resolution result to help character linking, the character linking result cannot be used to help to resolve coreference clusters. As a comparison, we treat both tasks jointly such that they can help each other. Currently, CorefQA is the best-performing general coreference resolution model on the OntoNotes dataset Pradhan et al. (2012). However, its performance is limited on the conversation dataset due to two reasons. First, different from the experimental setting of OntoNotes, the mentions in our experiment setting are gold mentions. Consequently, the flexible span predicting strategy of CorefQA loses its advantages because of the absence of the mention proposal stage. Second, the CorefQA leverages the fine-tuning on other question answering (QA) datasets and it is possible that the used QA dataset (i.e., SQuAD-2.0 Rajpurkar et al. (2018)) is more similar to OntoNotes rather than the used multiparty conversation dataset, which is typically much more informal. As a result, the effect of such fine-tuning process only works on OntoNotes. The coarse-to-fine (C2F) model Joshi et al. (2019b) with a transformer encoder was the previous state-of-the-art model on OntoNotes. Referring to Table 2, given the same text encoder, the proposed C2 model can constantly outperform the C2F model. These results further demonstrate that with the help of the proposed joint learning framework, the out-of-context character information can help achieve better mention representations so that the coreference models can resolve them more easily. ### 5.2 Character Linking Results As shown in Table 3, the proposed joint learning model also achieves the best performance on the character linking task and there are mainly two reasons for that. First, the contextualized mention representations obtained from pre- trained language encoders can better encode the context information than those representations used in ACNN. Second, with the help of coreference clusters, richer context about the whole conversation is encoded for each mention. For example, when using the same pre-trained language model as the encoder, C2 can always outperform the baseline classification model. These empirical results confirm that, though the BERT and SpanBERT can produce very good vector representation for the mentions based on the local context, the coreference clusters can still provide useful document-level contextual information for linking them to a global character entity. Figure 5: The x-axis is the number of MLSA layers used in the C2. The y-axes are the F1 scores on each metric for their corresponding tasks. The curves have general trends of going up, which indicates that the model performs better when there are more layers. ### 5.3 The Number of MLSA Layers Figure 6: Case study. All mentions that are linked to the same character and in the same coreference cluster are highlighted with the same color. The misclassified mention is marked with the red cross. Another contribution of the proposed C2 model is the proposed mention-level self-attention (MLSA) module, which helps iteratively refine the mention representations according to the other mentions co-occurred within the same document. In this section, to show its effect and the influence of iteration layers, we tried different layers and show their performances on the test set in Figure 5. We conducted the experiments with the SpanBERT-Base encoder and all other hyper-parameters are the same. The x-axis is the number of layers, and the y-axes are F1 scores of B3, CEAF, and BLANC for coreference resolution, the Macro and Micro F1 scores for character linking. From the results, we can see that with the increase of layer number from zero to five, the F1 scores on both tasks gradually increase. This trend demonstrates that the model can perform better on both tasks when there are more layers. Meanwhile, the marginal performance improvement of the MLSA layer is decreasing. This indicates that adding too many layers of MLSA may not further help improve the performance because enough context has been included. Considering the balance between performance and computational efficiency, we chose the iteration layers to be two in our current model based on similar observations made on the development set. | Coreference F1 | Linking F1 ---|---|--- Model | B3 | CEAF$\phi 4$ | BLANC | Micro | Macro C2 | 85.54 | 77.48 | 92,17 | 87.05 | 81.09 \- MLSA | 83.57 | 75.32 | 90.51 | 86.26 | 80.32 \- Linking | 83.50 | 76.10 | 90.08 | - | - \- Coref. | - | - | - | 86.94 | 79.58 Table 4: Three ablation studies are conducted concerning the MLSA layers, the coreference resolution module, and the character linking module. ### 5.4 Ablation Study In this section, we present the ablation study to clearly show the effect of different modules in the proposed framework C2 in Table 4. First, we try to remove the mention-level self-attention (MLSA) from our joint learning model and a clear performance drop is observed on both tasks. Specifically, the performance on coreference resolution is reduced by 1.21 on the average F1, and meanwhile, the macro-F1 and micro-F1 scores on character linking decreased by 0.77 and 0.79 respectively. The reduction reveals that the MLSA indeed helps achieve better mention representations with the help from co-occurred mentions. Second, we try to remove the coreference resolution and character linking modules. When the character linking module is removed, it is observed that the performance on coreference resolution decreased by 1.94 on the averaged F1 score. When the coreference module is removed, the performance of C2 on character linking dropped by 0.83 on the average of Micro and Macro F1 scores. These results prove that the modeling of coreference resolution and character linking can indeed help each other and improve the performance significantly, and the proposed joint learning framework can help to achieve that goal. ### 5.5 Case Study Besides the quantitative evaluation, in this section, we present the case study to qualitatively evaluate the strengths and weaknesses of the proposed C2 model. As shown in Figure 6, we randomly select an example from the development set to show the prediction results of the proposed model on both tasks. To illustrate the coreference resolution and character linking results from the C2 model, the mentions from the same coreference cluster are highlighted with the same color. Also, we use the same color to indicate to which character the mentions are referring. Meanwhile, the falsely predicted result is marked with a red cross. #### 5.5.1 Strengths For this example, the results on both tasks are consistent. The mentions that are linked to the same character entity are in the same coreference group and vice versa. Based on this observation and previous experimental results, it is more convincing that the proposed model can effectively solve the two problems at the same time. Besides that, we also notice that the model does not overfit the popular characters. It can correctly solve all the mentions referring to not only main characters, and also for the characters that only appear several times such as MAN 1. Last but not least, the proposed model can correctly resolve the mention to the correct antecedent even though there is a long distance between them in the conversation. For example, the mention me in utterance 14 can be correctly assigned to the mention you in utterance 2, though there are 11 utterances in between. It shows that by putting two tasks together, the proposed model can better utilize the whole conversation context. The only error made by the model is incorrectly classifying a mention and at the same time putting it into a wrong coreference cluster. #### 5.5.2 Weaknesses By analyzing the error case, it is noticed that the model may have trouble in handling the mentions that require common sense knowledge. Humans can successfully resolve the mention her to Danielle because they know Danielle is on the other side of the telephone, but Monica is in the house. As a result, Chandler can only deceive Danielle but not Monica. But the current model, which only relies on the context, cannot tell the difference. ### 5.6 Error Analysis We use the example in Figure 6 to emphasize the error analysis that compares the performance of our model and the baseline models. The details are as follows. In this example, the only mistake made by our model is related to common-sense knowledge, and the baseline models are also not able to make a correct prediction. For coreference resolution, 3 out of 25 mentions are put into a wrong cluster by the c2f baseline model. The baseline model failed to do long-distance antecedent assignments (e.g., the “me” in utterance 14). Meanwhile, our model is better in this case because it successfully predicts the antecedent of the mention “me”, even though its corresponding antecedent is far away in utterance 2. This example demonstrates the advantage that our joint model can use global information obtained from character linking to better resolve the co-referents that are far away from each other. For character linking, 2 out of 25 mentions are linked to the wrong characters by the baseline model. It is observed that the baseline model cannot consistently make correct linking predictions to less-appeared characters, for example, the “He” in utterance 6. In this case, our model performs better mainly because it can use the information gathered from the nearby co- referents to adjust its linking prediction, as its nearby co-referents are correctly linked to corresponding entities. ## 6 Related Works Coreference resolution is the task of grouping mentions to clusters such that all the mentions in the same cluster refer to the same real-world entity Pradhan et al. (2012); Zhang et al. (2019a, b); Yu et al. (2019). With the help of higher-order coreference resolution mechanism Lee et al. (2018) and strong pre-trained language models (e.g., SpanBERT Joshi et al. (2019b)), the end-to-end based coreference resolution systems have been achieving impressive performance on the standard evaluation dataset Pradhan et al. (2012). Recently, motivated by the success of the transfer learning, Wu et al. (2020) propose to model the coreference resolution task as a question answering problem. Through the careful fine-tuning on a high-quality QA dataset (i.e., SQUAD-2.0 Rajpurkar et al. (2018)), it achieves the state-of-the-art performance on the standard evaluation benchmark. However, as disclosed by Zhang et al. (2020), current systems are still not perfect. For example, they still cannot effectively handle pronouns, especially those in informal language usage scenarios like conversations. In this paper, we propose to leverage the out-of-context character information to help resolve the coreference relations with a joint learning model, which has been proven effective in the experiments. As a traditional NLP task, entity linking Mihalcea and Csomai (2007); Ji et al. (2015); Kolitsas et al. (2018); Raiman and Raiman (2018); Onando Mulang et al. (2020); van Hulst et al. (2020) aims at linking mentions in context to entities in the real world (typically in the format of knowledge graph). Typically, the mentions are named entities and the main challenge is the disambiguation. However, as a special case of the entity linking, the character linking task has its challenge that the majority of the mentions are pronouns. In the experiments, we have demonstrated that when the local context is not enough, the richer context information provided by the coreference clusters could be very helpful for linking mentions to the correct characters. In the NLP community, people have long been thinking that the coreference resolution task and entity linking should be able to help each other. For example, Ratinov and Roth (2012) show how to use knowledge from named-entity linking to improve the coreference resolution, but do not consider doing it in a joint learning approach. After that, Hajishirzi et al. (2013) demonstrate that the coreference resolution and entity linking are complementary in terms of reducing the errors in both tasks. Motivated by these observations, a joint model for coreference, typing, and linking is proposed Durrett and Klein (2014) to improve the performance on three tasks at the same time. Compared with previous works, the main contributions of this paper are two-fold: (1) we tackle the challenging character linking problem; (2) we design a novel mention representation encoding method, which has been shown effective on both the coreference resolution and character linking tasks. ## 7 Conclusion In this paper, we propose to solve the coreference resolution and character linking tasks jointly. The experimental results show that the proposed model C2 performs better than all previous models on both tasks. Detailed analysis is also conducted to show the contribution of different modules and the effect of the hyper-parameter. ## 8 Acknowledgements This paper was supported by the NSFC Grant U20B2053 from China, the Early Career Scheme (ECS, No. 26206717), the General Research Fund (GRF, No. 16211520), and the Research Impact Fund (RIF, No. R6020-19) from the Research Grants Council (RGC) of Hong Kong, with special thanks to the Tencent AI Lab Rhino-Bird Focused Research Program. ## References * Chen et al. (2017) Henry Y. Chen, Ethan Zhou, and Jinho D. Choi. 2017. Robust coreference resolution and entity linking on dialogues: Character identification on TV show transcripts. In _Proceedings of CoNLL)_ , pages 216–225. * Chen and Choi (2016) Yu-Hsin Chen and Jinho D. Choi. 2016. Character identification on multiparty conversation: Identifying mentions of characters in TV shows. In _Proceedings of SIGDIAL_ , pages 90–100. * Devlin et al. (2018) Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. _arXiv preprint arXiv:1810.04805_. * Durrett and Klein (2014) Greg Durrett and Dan Klein. 2014. A joint model for entity analysis: Coreference, typing, and linking. _Transactions of the association for computational linguistics_ , 2:477–490. * Hajishirzi et al. (2013) Hannaneh Hajishirzi, Leila Zilles, Daniel S. Weld, and Luke Zettlemoyer. 2013. Joint coreference resolution and named-entity linking with multi-pass sieves. In _Proceedings of the EMNLP 2013_ , pages 289–299. * van Hulst et al. (2020) Johannes M. van Hulst, Faegheh Hasibi, Koen Dercksen, Krisztian Balog, and Arjen P. de Vries. 2020. Rel: An entity linker standing on the shoulders of giants. _Proceedings of SIGIR 2020_. * Ji et al. (2015) Heng Ji, Joel Nothman, Ben Hachey, and Radu Florian. 2015. Overview of tac-kbp2015 tri-lingual entity discovery and linking. _Theory and Applications of Categories_. * Joshi et al. (2019a) Mandar Joshi, Danqi Chen, Y. Liu, Daniel S. Weld, L. Zettlemoyer, and Omer Levy. 2019a. Spanbert: Improving pre-training by representing and predicting spans. _Transactions of the Association for Computational Linguistics_ , 8:64–77. * Joshi et al. (2019b) Mandar Joshi, Omer Levy, Daniel S. Weld, and Luke Zettlemoyer. 2019b. BERT for coreference resolution: Baselines and analysis. In _Proceedings of EMNLP_. * Kantor and Globerson (2019) Ben Kantor and Amir Globerson. 2019. Coreference resolution with entity equalization. In _Proceedings of ACL 2019_ , pages 673–677. * Kingma and Ba (2015) Diederik P. Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In _Proceedings of ICLR 2015_. * Kolitsas et al. (2018) Nikolaos Kolitsas, Octavian-Eugen Ganea, and Thomas Hofmann. 2018. End-to-end neural entity linking. In _Proceedings of CoNNL 2018_ , pages 519–529. * Lee et al. (2018) Kenton Lee, Luheng He, and Luke Zettlemoyer. 2018. Higher-order coreference resolution with coarse-to-fine inference. In _Proceedings of NAACL (Short Papers)_ , pages 687–692. * Mihalcea and Csomai (2007) Rada Mihalcea and Andras Csomai. 2007. Wikify! linking documents to encyclopedic knowledge. CIKM ’07, page 233–242. * Onando Mulang et al. (2020) Isaiah Onando Mulang, Kuldeep Singh, Chaitali Prabhu, Abhishek Nadgeri, Johannes Hoffart, and Jens Lehmann. 2020. Evaluating the impact of knowledge graph context on entity disambiguation models. _arXiv e-prints_ , pages arXiv–2008. * Pradhan et al. (2012) Sameer Pradhan, Alessandro Moschitti, Nianwen Xue, Olga Uryupina, and Yuchen Zhang. 2012. Conll-2012 shared task: Modeling multilingual unrestricted coreference in ontonotes. pages 1–40. * Raiman and Raiman (2018) Jonathan Raiman and Olivier Raiman. 2018. Deeptype: multilingual entity linking by neural type system evolution. _arXiv preprint arXiv:1802.01021_. * Rajpurkar et al. (2018) Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. Know what you don’t know: Unanswerable questions for squad. In _Proceedings of ACL 2018_ , pages 784–789. * Ratinov and Roth (2012) Lev Ratinov and Dan Roth. 2012. Learning-based multi-sieve co-reference resolution with knowledge. In _Proceedings of EMNLP 2012_ , pages 1234–1244. * Recasens and Hovy (2011) Marta Recasens and Eduard Hovy. 2011. Blanc: Implementing the rand index for coreference evaluation. _Natural Language Engineering_ , 17:485 – 510. * Vaswani et al. (2017) Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, L. Kaiser, and Illia Polosukhin. 2017. Attention is all you need. _ArXiv_ , abs/1706.03762. * Vilain et al. (1995) Marc Vilain, John Burger, John Aberdeen, Dennis Connolly, and Lynette Hirschman. 1995. A model-theoretic coreference scoring scheme. _Proceedings of the 6th conference on message understanding_ , pages 45–52. * Wu et al. (2020) Wei Wu, Fei Wang, Arianna Yuan, Fei Wu, and Jiwei Li. 2020. CorefQA: Coreference resolution as query-based span prediction. In _Proceedings ACL 2020_ , pages 6953–6963. * Yu et al. (2019) Xintong Yu, Hongming Zhang, Yangqiu Song, Yan Song, and Changshui Zhang. 2019. What you see is what you get: Visual pronoun coreference resolution in dialogues. In _Proceedings of EMNLP-IJCNLP 2019_ , pages 5122–5131. * Zhang et al. (2019a) Hongming Zhang, Yan Song, and Yangqiu Song. 2019a. Incorporating context and external knowledge for pronoun coreference resolution. In _Proceedings of NAACL-HLT 2019_ , pages 872–881. * Zhang et al. (2019b) Hongming Zhang, Yan Song, Yangqiu Song, and Dong Yu. 2019b. Knowledge-aware pronoun coreference resolution. In _Proceedings of ACL 2019_ , pages 867–876. * Zhang et al. (2020) Hongming Zhang, Xinran Zhao, and Yangqiu Song. 2020. A brief survey and comparative study of recent development of pronoun coreference resolution. _CoRR_ , abs/2009.12721. * Zhou and Choi (2018) Ethan Zhou and Jinho D. Choi. 2018. They exist! introducing plural mentions to coreference resolution and entity linking. In _Proceedings of ICCL_ , pages 24–34.
# The cylindrical width of transitive sets Ashwin Sah , Mehtaab Sawhney and Yufei Zhao Department of Mathematics, Massachusetts Institute of Technology, Cambridge, MA 02139, USA <EMAIL_ADDRESS> ###### Abstract. We show that for every $1\leq k\leq d/(\log d)^{C}$, every finite transitive set of unit vectors in $\mathbb{R}^{d}$ lies within distance $O(1/\sqrt{\log(d/k)})$ of some codimension $k$ subspace, and this distance bound is best possible. This extends a result of Ben Green, who proved it for $k=1$. Sah and Sawhney were supported by NSF Graduate Research Fellowship Program DGE-1745302. Zhao was supported by NSF Award DMS-1764176, a Sloan Research Fellowship, and the MIT Solomon Buchsbaum Fund. ## 1\. Introduction The following counterintuitive fact was conjectured by the third author and proved by Green [4]. It says that every finite transitive subset of a high dimensional sphere is close to some hyperplane. Here a subset $X$ of a sphere in $\mathbb{R}^{d}$ is _transitive_ if for every $x,x^{\prime}\in X$, there is some $g\in\mathsf{O}(\mathbb{R}^{d})$ so that $gX=X$ and $gx=x^{\prime}$. We say that $X$ has _width_ at most $2r$ if it lies within distance $r$ of some hyperplane. The finiteness assumption is important since otherwise the whole sphere is a counterexample. ###### Theorem 1.1 (Green [4]). Let $X$ be a finite transitive subset of the unit sphere in $\mathbb{R}^{d}$. Then the width of $X$ is at most $O(1/\sqrt{\log d})$. Furthermore, this upper bound is best possible up to a constant factor. The bound in the theorem is tight since the set $X$ obtained by taking all permutations and coordinate-wise $\pm$ signings of the unit vector $(1,1/\sqrt{2},\ldots,1/\sqrt{d})/\sqrt{H_{d}}$, where $H_{d}=1+1/2+\cdots+1/d\sim\log d$, has width on the order of $1/\sqrt{\log d}$. Green’s proof uses a clever induction scheme along with sophisticated group theoretic arguments, including an application of the classification of finite simple groups. We generalize Green’s result by showing that a finite transitive set lies not only near some hyperplane, but in fact it lies near a subspace of codimension $k$, as long as $k$ is not too large. We say that $X\subset\mathbb{R}^{d}$ has _$k$ -cylindrical width_ at most $2r$ if $X$ lies within distance $r$ of some affine codimension $k$ subspace. The case $k=1$ corresponds to the usual notion of width. Our main result below implies that every finite transitive subset of the unit sphere in $\mathbb{R}^{d}$ has $k$-cylindrical width $O(1/\sqrt{\log(d/k)})$ as long as $k$ is not too large. ###### Theorem 1.2. There is an absolute constant $C>0$ so that the following holds. Let $1\leq k\leq d/(\log(3d))^{C}$. Let $X$ be a finite transitive subset of the unit sphere in $\mathbb{R}^{d}$. Then there is a real $k$-dimensional subspace $W$ such that $\sup_{\mathbf{x}\in X}\lVert\operatorname{proj}_{W}\mathbf{x}\rVert_{2}\lesssim\frac{1}{\sqrt{\log(d/k)}}.$ Here and throughout $a\lesssim b$ means that $a\leq C^{\prime}b$ for some absolute constant $C^{\prime}$. We write $\lVert\mathbf{x}\rVert_{2}$ for the usual Euclidean norm of a vector $\mathbf{x}$. Also $\operatorname{proj}_{W}$ is the orthogonal projection onto $W$. We deduce the above theorem from a complex version using a theorem on restricted invertibility (see Section 6). A transitive subset of the complex unit sphere is defined to be the orbit of a point under the action of some subgroup of the unitary group. ###### Theorem 1.3. There is an absolute constant $C>0$ so that the following holds. Let $1\leq k\leq d/(\log(3d))^{C}$. Let $X$ be a finite transitive subset of the unit sphere in $\mathbb{C}^{d}$. Then there is a complex $k$-dimensional subspace $W$ such that $\sup_{\mathbf{x}\in X}\lVert\operatorname{proj}_{W}\mathbf{x}\rVert_{2}\lesssim\frac{1}{\sqrt{\log(d/k)}}.$ We suspect that the $1\leq k\leq d/(\log(3d))^{C}$ hypothesis is unnecessary in both Theorems 1.2 and 1.3. ###### Conjecture 1.4. Let $1\leq k\leq d$. Let $X$ be a finite transitive subset of the unit sphere in $\mathbb{C}^{d}$. Then there is a complex $k$-dimensional subspace $W$ such that $\sup_{\mathbf{x}\in X}\lVert\operatorname{proj}_{W}\mathbf{x}\rVert_{2}\lesssim\frac{1}{\sqrt{\log(2d/k)}}.$ One particularly intriguing special case of 1.4 is that every finite transitive set of unit vectors in $\mathbb{R}^{d}$ has $k$-cylindrical width $o(1)$ for all $k=o(d)$. We prove a matching lower bound on the cylindrical radius (See Section 7 for proof.) ###### Theorem 1.5. Let $1\leq k\leq d$. There exists a transitive set $X$ in $\mathbb{R}^{d}$ such that for any (real or complex) $k$-dimensional subspace $W$ we have $\sup_{\mathbf{x}\in X}\lVert\operatorname{proj}_{W}\mathbf{x}\rVert_{2}\gtrsim\frac{1}{\sqrt{\log(2d/k)}}.$ We propose another closely related conjecture: every finite transitive set in $\mathbb{R}^{d}$ lies inside a small cube. ###### Conjecture 1.6. Let $X$ be a finite transitive subset of the unit sphere in $\mathbb{R}^{d}$ (or $\mathbb{C}^{d}$). Then there is a unitary basis $L$ such that $\sup_{\mathbf{x}\in X,\mathbf{v}\in L}|\langle\mathbf{v},\mathbf{x}\rangle|\lesssim\frac{1}{\sqrt{\log d}}.$ (1.1) Establishing an upper bound that decays to zero as $d\to\infty$ would already be interesting. Note that Theorem 1.3 implies the existence of a set $L$ of orthonormal vectors with $\lvert L\rvert\geq d^{0.99}$ so that 1.1 holds (and likely extendable to $\lvert L\rvert\geq d/(\log d)^{C}$ via our techniques). Proving either conjecture in full appears to require additional ideas. ###### Remark. Green’s proof [4] of Theorems 1.2 and 1.3 in the case $k=1$ contains two errors. The first error is due to a missing supremum inside the integral in the first and second lines of the last display equation in proof of Proposition 2.1 on page 560. The second error occurs at the final equality step of the top display equation on page 569, after right after (4.4); here an orthogonality relation was incorrectly applied as it requires an unjustified exchange of the integral and supremum. Our proof here corrects these errors. Green has also updated the arXiv version of his paper [4] incorporating these corrections. ## 2\. Proof strategy The subspace $W$ in Theorem 1.3 must vary according to the transitive set $X$. On other hand, the strategy is to construct a single probability distribution $\mu$ (depending only on the symmetry group $G\leqslant\mathsf{U}(\mathbb{C}^{d})$ but not on $X$) on the set $\operatorname{Gr}_{\mathbb{C}}(k,d)$ of $k$-dimensional subspaces of $\mathbb{C}^{d}$. This is an important idea introduced by Green (for $k=1$). ###### Definition 2.1. Let $1\leq k\leq d$. Let $f_{k}(d)$ be the smallest value so that for every finite $G\leqslant\mathsf{U}(\mathbb{C}^{d})$, there is a probability measure $\mu$ on $\operatorname{Gr}_{\mathbb{C}}(k,d)$ such that for all $\mathbf{v}\in\mathbb{S}(\mathbb{C}^{d})$, $\int\sup_{g\in G}\lVert\operatorname{proj}_{W}(g\mathbf{v})\rVert_{2}^{2}d\mu(W)\leq f_{k}(d)^{2}$ The values $f_{k}(d)$ are well defined since the space of probability measures $\mu$ in question is closed under weak limits. Our main result about $f_{k}(d)$ is stated below. ###### Theorem 2.2. If $k\leq d/(\log d)^{20}$, then $f_{k}(d)\lesssim\frac{1}{\sqrt{\log(d/k)}}.$ ###### Proof of Theorem 1.3 given Theorem 2.2. Let our transitive set $X$ be the orbit of $\mathbf{v}\in\mathbb{S}(\mathbb{C}^{d})$ under the action of the the finite subgroup $G\leqslant\mathsf{U}(\mathbb{C}^{d})$. By Theorem 2.2 and Definition 2.1, there is a measure $\mu$ on $\operatorname{Gr}_{\mathbb{C}}(k,d)$ such that $\int\sup_{g\in G}\lVert\operatorname{proj}_{W}(g\mathbf{v})\rVert_{2}^{2}d\mu(W)\leq f_{k}(d)^{2}.$ Therefore there is some $k$-dimensional subspace $W$ with $\sup_{g\in G}\lVert\operatorname{proj}_{W}(g\mathbf{v})\rVert_{2}\leq f_{k}(d)\lesssim\frac{1}{\sqrt{\log(d/k)}}.\qed$ To prove Theorem 2.2 , we will decompose $G$ to “smaller”, more restricted cases, namely irreducible and primitive representations. We will also need to consider permutation groups (for both the reduction step as well as the primitive case). ### 2.1. Preliminaries ###### Definition 2.3. We say that $G\leqslant\mathsf{U}(\mathbb{C}^{d})$ is _imprimitive_ if there is a _system of imprimitivity_ : a decomposition $\mathbb{C}^{d}=\bigoplus_{i=1}^{\ell}V_{i}$ with $0<\dim V_{i}<d$ such that for every $g\in G$ and $i\in[\ell]$ one has $gV_{i}=V_{j}$ for some $j\in[\ell]$. (The subspaces $V_{i}$ need not be orthogonal.) Otherwise we say that $G$ is _primitive_. ###### Remark. Both primitivity and irreducibility are properties of a representation, rather than intrinsic to a group. We identify $G\leqslant\mathsf{U}(\mathbb{C}^{d})$ with its natural representation on $\mathbb{C}^{d}$. It follows from Maschke’s theorem that primitive group representations are irreducible. ###### Definition 2.4. Given $\mathbf{v}=(v_{1},\ldots,v_{d})\in\mathbb{C}^{d}$, let $\mathbf{v}^{\succ}=(\lvert v_{\sigma(1)}\rvert,\ldots,\lvert v_{\sigma(d)}\rvert)\in\mathbb{R}^{d}$ where $\sigma$ is a permutation of $[d]$ so that $\lvert v_{\sigma(1)}\rvert\geq\cdots\geq\lvert v_{\sigma(d)}\rvert.$ We write $v_{i}^{\succ}$ for the $i$-th coordinate of $\mathbf{v}^{\succ}$. Let $\operatorname{Dom}(\mathbf{v})=\\{\mathbf{w}\in\mathbb{C}^{d}:w_{i}^{\succ}\leq v_{i}^{\succ}\text{for all }i\in[d]\\}.$ Let (here $\mathfrak{S}_{d}$ denotes the symmetric group) $\Gamma_{d}:=\mathfrak{S}_{d}\ltimes(\mathbb{S}^{1})^{d}\leq\mathsf{U}(\mathbb{C}^{d})$ be the group that acts on $\mathbb{C}^{d}$ be permuting its coordinates and multiplying individual coordinates by unit complex numbers. Then $\operatorname{Dom}(\mathbf{v})$ is the convex hull of the $\Gamma_{d}$-orbit of $\mathbf{v}$. We define some variants of $f_{k}(d)$ when the group $G$ is restricted to special types. ###### Definition 2.5. Given $k\in[d]$, let $f^{\operatorname{irred}}_{k}(d)$ (resp. $f^{\operatorname{prim}}_{k}(d)$) be the smallest value so that for every finite $G\leqslant\mathsf{U}(\mathbb{C}^{d})$ which is irreducible (resp. primitive), there is a probability measure $\mu$ on $\operatorname{Gr}_{\mathbb{C}}(k,d)$ such that for every $\mathbf{v}\in\mathbb{S}(\mathbb{C}^{d})$, $\int\sup_{g\in G}\lVert\operatorname{proj}_{W}(g\mathbf{v})\rVert_{2}^{2}d\mu(W)\leq f^{\operatorname{irred}}_{k}(d)^{2}\qquad\text{(resp. $f^{\operatorname{prim}}_{k}(d)^{2}$)}.$ The permutation action on $\mathbb{C}^{d}$ deserves special attention. ###### Definition 2.6. Let $f^{\operatorname{sym}}_{k}(d)$ be the smallest value so that there is a probability measure $\mu$ on $\operatorname{Gr}_{\mathbb{C}}(k,d)$ such that for every $\mathbf{v}\in\mathbb{S}(\mathbb{C}^{d})$, $\int\sup_{\mathbf{u}\in\operatorname{Dom}(\mathbf{v})}\lVert\operatorname{proj}_{W}(\mathbf{u})\rVert_{2}^{2}d\mu(W)\leq f^{\operatorname{sym}}_{k}(d)^{2}.$ Define $f^{\operatorname{alt}}_{k}(d)$ to be the same with the additional constraint that $\mu$ is supported on the set of $k$-dimensional subspaces of the hyperplane $x_{1}+\cdots+x_{d}=0$. We will often equivalently consider, instead of $\mu$ on $\operatorname{Gr}_{\mathbb{C}}(k,d)$, the corresponding measure $\mu^{\ast}$ on the complex Stiefel manifold $V_{k}(\mathbb{C}^{d})$, that is, $\mu^{\ast}$ is derived from $\mu$ by first sampling a $\mu$-random $k$-dimensional subspace $W$ of $\mathbf{C}^{d}$, and then outputting a uniformly sampled a unitary basis $(\mathbf{w}_{1},\ldots,\mathbf{w}_{k})$ of $W$. We have $\lVert\operatorname{proj}_{W}(\mathbf{u})\rVert_{2}^{2}=\sum_{\ell=1}^{k}\lvert\langle g\mathbf{v}_{1},\mathbf{w}_{\ell}\rangle\rvert^{2}$. ### 2.2. Reductions We first reduce the general problem to the irreducible case. ###### Proposition 2.7. If $1\leq k<\ell\leq d$ then $f_{k}(d)\leq\max\bigl{\\{}\sqrt{k/\ell},\sup_{d^{\prime}\geq d/(2\ell)}f^{\operatorname{irred}}_{\lceil 2kd^{\prime}/d\rceil}(d^{\prime})\bigr{\\}}.$ We then reduce the irreducible case to the primitive case and the alternating case. ###### Proposition 2.8. If $k\leq d/2$, then $f^{\operatorname{irred}}_{k}(d)\leq\max_{d_{1}d_{2}=d}\bigl{(}\min\bigl{\\{}f^{\operatorname{prim}}_{\lceil k/d_{1}\rceil}(d_{2}),f^{\operatorname{alt}}_{k}(d_{1})+\mathbbm{1}_{k\geq d_{1}}\bigr{\\}}\bigr{)}.$ The symmetric and alernating cases can be handled explicitly, yielding the following. ###### Proposition 2.9. If $k\leq d/(\log d)^{5}$, then $f_{k}^{\operatorname{sym}}(d)\leq f_{k}^{\operatorname{alt}}(d)\lesssim 1/\sqrt{\log(d/k)}.$ This leaves the primitive case, which we prove by invoking an group theoretic result proved by Green [4, Proposition 4.2] that allows us to once again reduce to the alternating case once again. ###### Proposition 2.10. There is an absolute constant $c>0$ such that for $k\leq cd/(\log d)^{4}$ we have $f_{k}^{\operatorname{prim}}(d)\lesssim\sup_{d^{\prime}\geq cd/(\log d)^{4}}f^{\operatorname{alt}}_{k}(d^{\prime}).$ ### 2.3. Putting everything together We are now in position to derive Theorem 2.2 using the preceding statements. ###### Proposition 2.11. If $k\leq 2d/(\log d)^{10}$ then $f^{\operatorname{prim}}_{k}(d)\lesssim 1/\sqrt{\log(d/k)}$. ###### Proof. Combine Propositions 2.9 and 2.10. ∎ ###### Proposition 2.12. If $k\leq d/(\log d)^{10}$ then $f^{\operatorname{irred}}_{k}(d)\lesssim 1/\sqrt{\log(d/k)}$. ###### Proof. By Proposition 2.8, we have $f^{\operatorname{irred}}_{k}(d)\leq\max_{d_{1}d_{2}=d}(\min(f^{\operatorname{prim}}_{\lceil k/d_{1}\rceil}(d_{2}),f^{\operatorname{alt}}_{k}(d_{1})+\mathbbm{1}_{k\geq d_{1}})).$ First consider the case $d_{1}\leq k$. We have $\lceil k/d_{1}\rceil\leq\frac{2d}{d_{1}(\log d)^{10}}\leq\frac{2d_{2}}{(\log d_{2})^{10}}.$ By Proposition 2.11, we have $f^{\operatorname{prim}}_{\lceil k/d_{1}\rceil}(d_{2})\lesssim\frac{1}{\sqrt{\log(d_{2}/\lceil k/d_{1}\rceil)}}\leq\frac{1}{\sqrt{\log(d/(2k))}}.$ Now consider the case $d_{1}>k$. Since $d_{2}(d_{1}/k)=d/k$, we have $\max\\{d_{1},d_{2}/k\\}\geq\sqrt{d/k}$. If $d_{2}\geq\sqrt{d/k}$, then $f^{\operatorname{prim}}_{\lceil k/d_{1}\rceil}(d_{2})=f^{\operatorname{prim}}_{1}(d_{2})\lesssim\frac{1}{\sqrt{\log d_{2}}}\lesssim\frac{1}{\sqrt{\log(d/k)}}.$ On the other hand, if $d_{1}/k\geq\sqrt{d/k}$, then $d_{1}/k\geq(\log d)^{5}$ so $k\leq\frac{d_{1}}{(\log d)^{5}}\leq\frac{d_{1}}{(\log d_{1})^{5}}.$ Hence Proposition 2.9 yields $f^{\operatorname{alt}}_{k}(d_{1})\lesssim\frac{1}{\sqrt{\log(d_{1}/k)}}\lesssim\frac{1}{\sqrt{\log(d/k)}}.$ Thus it follows that, for all $d_{1}d_{2}=d$, $\min(f^{\operatorname{prim}}_{\lceil k/d_{1}\rceil}(d_{2}),f^{\operatorname{alt}}_{k}(d_{1})+\mathbbm{1}_{k\geq d_{1}})\lesssim\frac{1}{\sqrt{\log(d/k)}},$ and the result follows. ∎ Now we show the main result assuming the above statements. ###### Proof of Theorem 2.2. Let $\ell=\lceil\sqrt{dk}\rceil\geq 2k$. We have $k/\ell\lesssim\sqrt{k/d}\lesssim\frac{1}{\sqrt{\log(d/k)}}.$ Also, if $d^{\prime}\geq d/(2\ell)$ then $\bigg{\lceil}\frac{2kd^{\prime}}{d}\bigg{\rceil}\leq\frac{d^{\prime}}{d/(2\ell)}\leq\frac{d^{\prime}}{(\log d)^{10}}\leq\frac{d^{\prime}}{(\log d^{\prime})^{10}}.$ By Proposition 2.12, we have $f^{\operatorname{irred}}_{\lceil 2kd^{\prime}/d\rceil}(d^{\prime})\lesssim\frac{1}{\sqrt{\log(d^{\prime}/\lceil 2kd^{\prime}/d\rceil)}}\lesssim\frac{1}{\sqrt{\log(d/(2\ell))}}\lesssim\frac{1}{\sqrt{\log(d/k)}}.$ Applying Proposition 2.7 to $k$ and $\ell=\lceil\sqrt{dk}\rceil$, we find $f_{k}(d)\leq\max(\sqrt{k/\ell},\sup_{d^{\prime}\geq d/(2\ell)}f^{\operatorname{irred}}_{\lceil 2kd^{\prime}/d\rceil}(d^{\prime}))\lesssim\frac{1}{\sqrt{\log(d/k)}}.\qed$ ### 2.4. Paper outline In Section 3, we prove the two key reductions, Propositions 2.7 and 2.8. In Section 4, we prove the key estimate for the symmetric and alternating cases, Proposition 2.9. In Section 5, we prove the primitive case, Proposition 2.10. Finally, in Section 6 we deduce a real version from the complex version, proving Theorem 1.2. In Section 7 we demonstrate optimality of our results by exhibiting the matching lower bound Theorem 1.5. ## 3\. Reduction to primitive representations We first reduce the general case to the alternating and irreducible cases. ###### Proof of Proposition 2.7. Consider $G\leqslant\mathsf{U}(\mathbb{C}^{d})$. By Maschke’s theorem, we can decompose into irreducible representations of $G$: $\mathbb{C}^{d}=\bigoplus_{j=1}^{m}V_{j}.$ Let $d_{j}=\dim V_{j}$. Let $J=\\{j\in[m]:d_{j}\geq d/(2\ell)\\}.$ First suppose $\sum_{j\in J}d_{j}\geq d/2$. Then in each such $V_{j}$, we consider the probability measure $\mu_{j}$ that witnesses $f^{\operatorname{irred}}_{\lceil 2kd_{j}/d\rceil}(d_{j})$ for the irreducible representation of $G$ on $V_{j}$. That is, $\mu_{j}$ samples a $\lceil 2kd_{j}/d\rceil$-dimensional subspace of $V_{j}$ and satisfies $\int\sup_{g\in G}\lVert\operatorname{proj}_{W}(g\mathbf{v})\rVert_{2}^{2}d\mu_{j}(W)\leq f^{\operatorname{irred}}_{\lceil 2kd^{\prime}/d\rceil}(d_{j})^{2}\lVert\mathbf{v}\rVert_{2}^{2}$ for each $\mathbf{v}\in V_{j}$. We define $\mu$ to be a uniformly random $k$-dimensional subspace of $\bigoplus_{j\in J}W_{j}$, where each $W_{j}$ is an independent $\mu_{j}$-random $\lceil 2kd_{j}/d\rceil$-dimensional subspace of $V_{j}$. (Note the $W_{j}$’s are orthogonal as the $V_{j}$’s are.) The total dimension of this direct sum is at least $k$, so $\mu$ is well-defined. Given $\mathbf{v}\in\mathbb{C}^{d}$, write $\mathbf{v}=\sum_{j=1}^{m}\mathbf{v}_{j}$ with $\mathbf{v}_{j}\in V_{j}$. We have $\displaystyle\int\sup_{g\in G}\lVert\operatorname{proj}_{W}(g\mathbf{v})\rVert_{2}^{2}d\mu(W)$ $\displaystyle\leq\int\sup_{g\in G}\lVert\operatorname{proj}_{\bigoplus_{j\in J}W_{j}}(g\mathbf{v})\rVert_{2}^{2}\prod_{j\in J}d\mu_{j}(W_{j})$ $\displaystyle\leq\sum_{j\in J}\int\sup_{g\in G}\lVert\operatorname{proj}_{W_{j}}(g\mathbf{v})\rVert_{2}^{2}d\mu_{j}(W_{j})$ $\displaystyle\leq\sum_{j\in J}f^{\operatorname{irred}}_{\lceil 2kd_{j}/d\rceil}(d_{j})^{2}\lVert\mathbf{v}_{j}\rVert_{2}^{2}$ $\displaystyle\leq\sup_{d^{\prime}\geq d/(2\ell)}f^{\operatorname{irred}}_{\lceil 2kd^{\prime}/d\rceil}(d^{\prime})^{2}\lVert\mathbf{v}\rVert_{2}^{2}$ by orthogonality of the $V_{j}$. Next suppose $\sum_{j\in J}d_{j}<d/2$. Then $|[m]\setminus J|\geq\ell$. Let $I$ be an $\ell$-element subset of $[m]\setminus J$. Choose arbitrary $\mathbf{w}_{j}\in\mathbb{S}(V_{j})\subseteq\mathbb{C}^{d}$ for $j\in I$, which are clearly orthogonal. Let $\mu$ be the probability measure on $k$-dimensional subspaces of $\mathbb{C}^{d}$ obtained by taking the span of $k$ uniform random elements in $\\{\mathbf{w}_{1},\ldots,\mathbf{w}_{\ell}\\}$. For each $g\in G$, write $\mathbf{u}_{g}=(\langle g\mathbf{v},\mathbf{w}_{1}\rangle,\ldots,\langle g\mathbf{v},\mathbf{w}_{\ell}\rangle)$ and $\mathbf{v}^{\prime}=(\lVert\operatorname{proj}_{V_{1}}\mathbf{v}\rVert_{2},\ldots,\lVert\operatorname{proj}_{V_{\ell}}\mathbf{v}\rVert_{2}).$ Given $S\subseteq[\ell]$, let $\operatorname{proj}_{S}$ take the projection of an $\ell$-dimensional vector down to that subset of coordinates. We have $\displaystyle\int\sup_{g\in G}\lVert\operatorname{proj}_{W}(g\mathbf{v})\rVert_{2}^{2}d\mu(W)$ $\displaystyle=\frac{1}{\binom{\ell}{k}}\sum_{S\in\binom{[\ell]}{k}}\sup_{g\in G}\lVert\operatorname{proj}_{S}(\mathbf{u}_{g})\rVert_{2}^{2}$ $\displaystyle\leq\frac{1}{\binom{\ell}{k}}\sum_{S\in\binom{[\ell]}{k}}\sum_{j\in S}(v_{j}^{\prime})^{2}$ $\displaystyle=\frac{k}{\ell}\sum_{j=1}^{\ell}(v_{j}^{\prime})^{2}\leq\frac{k}{\ell}\lVert\mathbf{v}\rVert_{2}^{2}.$ The first equality follows by the definition of $\mu$, the subsequent inequality follows by $|\langle g\mathbf{v},\mathbf{w}_{j}\rangle|\leq v_{j}^{\prime}$, and the last line is by direct computation and orthogonality of the $V_{j}$. ∎ We next reduce the irreducible case to the primitive case. We first collect a few facts proved in [4] regarding systems of imprimitivity. ###### Lemma 3.1 ([4, Section 2]). Let $G\leqslant\mathsf{U}(\mathbb{C}^{d})$ be irreducible but imprimitive. Consider a system imprimitivity $\mathbb{C}^{d}=\bigoplus_{j=1}^{d_{1}}V_{j}$ with $d_{1}$ maximal over all such systems of primitivity. Let $H=\\{g\in G:gV_{1}=V_{1}\\}$ and choose $\gamma_{1},\ldots,\gamma_{d_{1}}$ such that $\gamma_{j}V_{1}=V_{j}$. Then the following hold: 1. 1. The $V_{j}$ are orthogonal and have the same dimension, and $G$ acts transitively on them. 2. 2. $H$ has primitive action on $V_{1}$ (i.e. the representation of $H$ on $V_{1}$ is primitive). 3. 3. $\gamma_{1},\ldots,\gamma_{d_{1}}$ form a complete set of left coset representatives for $H$ in $G$. 4. 4. For each $g\in G$ there is $\sigma_{g}\in\mathfrak{S}_{d_{1}}$ so that $\gamma_{\sigma_{g}(j)}^{-1}g\gamma_{j}\in H$ for all $j\in[d_{1}]$ (i.e., $\sigma_{g}$ records how $g$ permutes $\\{V_{1},\ldots,V_{d_{1}}\\}$). Now we are ready to prove Proposition 2.8, which recall says that for all $k\leq d/2$, $f^{\operatorname{irred}}_{k}(d)\leq\max_{d_{1}d_{2}=d}\bigl{(}\min\bigl{\\{}f^{\operatorname{prim}}_{\lceil k/d_{1}\rceil}(d_{2}),f^{\operatorname{alt}}_{k}(d_{1})+\mathbbm{1}_{k\geq d_{1}}\bigr{\\}}\bigr{)}.$ ###### Proof of Proposition 2.8. Let $G\leqslant\mathsf{U}(\mathbb{C}^{d})$ be irreducible but imprimitive. Consider a system of imprimitivity $\mathbb{C}^{d}=\bigoplus_{j=1}^{d_{1}}V_{j}$ with $d_{1}$ maximal among all systems of imprimitivity. By Lemma 3.1, the spaces $V_{j}$ are orthogonal and all the $\dim V_{j}$ are equal. Let $d_{2}=\dim V_{1}$, so that $d_{1}d_{2}=d$. Furthermore, $H=\\{g\in G:gV_{1}=V_{1}\\}$ acts primitively on $V_{1}$, that $G$ acts transitively on the $V_{j}$, and that there are $\gamma_{1},\ldots,\gamma_{d_{1}}$ so that $\gamma_{j}V_{1}=V_{j}$ which form a complete set of left coset representatives for $H$ in $G$. For each $g\in G$ we have some $\sigma_{g}\in\mathfrak{S}_{d_{1}}$ so that $\gamma_{\sigma_{g}(j)}^{-1}g\gamma_{j}\in H$ for all $j\in[d_{1}]$. Define $h(g,j)=\gamma_{\sigma_{g}(j)}^{-1}g\gamma_{j}$. Let $\mathbf{v}\in\mathbb{C}^{d}$. There is a unique orthogonal decomposition $\mathbf{v}=\sum_{j=1}^{d_{1}}\gamma_{j}\mathbf{v}_{j}$ where $\mathbf{v}_{j}\in V_{1}$ for all $j\in[d_{1}]$. We have $g\mathbf{v}=\sum_{j=1}^{d_{1}}g\gamma_{j}\mathbf{v}_{j}=\sum_{j=1}^{d_{1}}\gamma_{j}h(g,\sigma_{g}^{-1}(j))\mathbf{v}_{\sigma_{g}^{-1}(j)}.$ Finally, if $\mathbf{w}=\sum_{j=1}^{d_{1}}\lambda_{j}\gamma_{j}\mathbf{x}$ for some $\bm{\lambda}=(\lambda_{1},\ldots,\lambda_{d_{1}})\in\mathbb{C}^{d_{1}}$ and $\mathbf{x}\in V_{1}$ then we see from the above and orthogonality that $\langle g\mathbf{v},\mathbf{w}\rangle=\sum_{j=1}^{d_{1}}\lambda_{j}\langle h(g,\sigma_{g}^{-1}(j))\mathbf{v}_{\sigma_{g}^{-1}(j)},\mathbf{x}\rangle.$ Now we return to the situation at hand: we need to choose a $k$-dimensional space with a good projection for our transitive set. Consider the map $\psi:V_{1}\times\mathbb{C}^{d_{1}}\to\mathbb{C}^{d}$ given by $\psi(\mathbf{x},\bm{\lambda})=\sum_{j=1}^{d_{1}}\lambda_{j}\gamma_{j}\mathbf{x}.$ It clearly maps the pair of unit spheres into the unit sphere. Given probability measures $\mu_{1}$ on $\operatorname{Gr}_{\mathbb{C}}(k_{1},V_{1})$ and $\mu_{2}$ on $\operatorname{Gr}_{\mathbb{C}}(k_{2},\mathbb{C}^{d_{1}})$, we define the pushforward measure $\mu$ on $\operatorname{Gr}_{\mathbb{C}}(k_{1}k_{2},d)$ by taking the image of these two subspaces under $\psi$. Equivalently, suppose $\mu_{1}^{\ast}$ samples a unitary basis $\mathbf{x}_{1},\ldots,\mathbf{x}_{k_{1}}$ of a subspace of $V_{1}$ and $\mu_{2}^{\ast}$ samples a unitary basis $\bm{\lambda}_{1},\ldots,\bm{\lambda}_{k_{2}}$ of a subspace of $\mathbb{C}^{d_{1}}$, then $\mu$ samples the subspace of $\mathbf{C}^{d}$ with basis $\\{\psi(\mathbf{x}_{i},\bm{\lambda}_{j}):i\in[k_{1}],j\in[k_{2}]\\}$. It is easy to check this basis is in fact unitary. Next, we choose $\mu_{1}$ and $\mu_{2}$ based on the sizes of $d_{1}$ and $d_{2}$. First let $k_{1}=\lceil k/d_{1}\rceil\leq d_{2}$ (as $k\leq d/2$) and $k_{2}=d_{1}$. We let $\mu_{1}$ be the measure guaranteed by Definition 2.1 so that $\int\sup_{h\in H}\lVert\operatorname{proj}_{W}(h\mathbf{u})\rVert_{2}^{2}d\mu_{1}(W)\leq f^{\operatorname{prim}}_{k_{1}}(d_{2})^{2}\lVert\mathbf{u}\rVert_{2}^{2}$ for all $\mathbf{u}\in V_{1}$ and let $\mu_{2}$ be the atom on the space $\mathbb{C}^{d_{1}}$ in $\operatorname{Gr}_{\mathbb{C}}(d_{1},d_{1})$. Let $\mu$ be the $\psi$-pushforward of $(\mu_{1},\mu_{2})$ as described earlier. We find $\displaystyle\int\sup_{g\in G}\lVert\operatorname{proj}_{W}(g\mathbf{v})\rVert_{2}^{2}d\mu(W)$ $\displaystyle=\int\sup_{g\in G}\sum_{\ell=1}^{k_{1}}\sum_{j=1}^{d_{1}}|\langle g\mathbf{v},\psi(\mathbf{x}_{\ell},\mathbf{e}_{j})\rangle|^{2}d\mu_{1}^{\ast}(\mathbf{x}_{1},\ldots,\mathbf{x}_{k_{1}})$ $\displaystyle=\int\sup_{g\in G}\sum_{\ell=1}^{k_{1}}\sum_{j=1}^{d_{1}}|\langle h(g,\sigma_{g}^{-1}(j))\mathbf{v}_{\sigma_{g}^{-1}(j)},\mathbf{x}_{\ell}\rangle|^{2}d\mu_{1}^{\ast}(\mathbf{x}_{1},\ldots,\mathbf{x}_{k_{1}})$ $\displaystyle=\int\sup_{g\in G}\sum_{\ell=1}^{k_{1}}\sum_{j=1}^{d_{1}}|\langle h(g,j)\mathbf{v}_{j},\mathbf{x}_{\ell}\rangle|^{2}d\mu_{1}^{\ast}(\mathbf{x}_{1},\ldots,\mathbf{x}_{k_{1}})$ $\displaystyle\leq\sum_{j=1}^{d_{1}}\int\sup_{g\in G}\sum_{\ell=1}^{k_{1}}|\langle h(g,j)\mathbf{v}_{j},\mathbf{x}_{\ell}\rangle|^{2}d\mu_{1}^{\ast}(\mathbf{x}_{1},\ldots,\mathbf{x}_{k_{1}})$ $\displaystyle\leq\sum_{j=1}^{d_{1}}\int\sup_{h\in H}\sum_{\ell=1}^{k_{1}}|\langle h\mathbf{v}_{j},\mathbf{x}_{\ell}\rangle|^{2}d\mu_{1}^{\ast}(\mathbf{x}_{1},\ldots,\mathbf{x}_{k_{1}})$ $\displaystyle\leq\sum_{j=1}^{d_{1}}f^{\operatorname{prim}}_{k_{1}}(d_{2})^{2}\lVert\mathbf{v}_{j}\rVert_{2}^{2}=f^{\operatorname{prim}}_{k_{1}}(d_{2})^{2}\lVert\mathbf{v}\rVert_{2}^{2}.$ The last equality is by orthogonality of $V_{1},\ldots,V_{d_{1}}$ and unitarity of $\gamma_{j}$ for $j\in[d_{1}]$. Now suppose that $k<d_{1}$. Let $k_{1}=1$ and $k_{2}=k$. Choose an arbitrary unit vector $\mathbf{x}\in V_{1}$ and $\mu_{1}$ be an atom on $\operatorname{Gr}_{\mathbb{C}}(1,V_{1})$ supported on the line $\mathbb{C}\mathbf{x}$. Let $\mu_{2}$ be guaranteed by Definition 2.6 so that $\int\sup_{\mathbf{u}\in\operatorname{Dom}(\mathbf{w})}\sum_{\ell=1}^{k}|\langle\mathbf{u},\bm{\lambda}_{\ell}\rangle|^{2}d\mu_{2}^{*}(\bm{\lambda}_{1},\ldots,\bm{\lambda}_{k})\leq f_{k}^{\operatorname{alt}}(d_{1})^{2}\lVert\mathbf{w}\rVert_{2}^{2}$ for all $\mathbf{w}\in V_{1}$. Let $\mu$ be the $\psi$-pushforward of $(\mu_{1},\mu_{2})$ as described earlier. We find $\displaystyle\int\sup_{g\in G}\lVert\operatorname{proj}_{W}(g\mathbf{v})\rVert_{2}^{2}d\mu(W)$ $\displaystyle=\int\sup_{g\in G}\sum_{\ell=1}^{k}|\langle g\mathbf{v},\mathbf{w}_{\ell}\rangle|^{2}d\mu^{\ast}(\mathbf{w}_{1},\ldots,\mathbf{w}_{k})$ $\displaystyle=\int\sup_{g\in G}\sum_{\ell=1}^{k}|\langle g\mathbf{v},\psi(\mathbf{x},\bm{\lambda}_{\ell})\rangle|^{2}d\mu_{2}^{\ast}(\bm{\lambda}_{1},\ldots,\bm{\lambda}_{k})$ $\displaystyle=\int\sup_{g\in G}\sum_{\ell=1}^{k}\bigg{|}\sum_{j=1}^{d_{1}}\lambda_{\ell,j}\langle h(g,\sigma_{g}^{-1}(j))\mathbf{v}_{\sigma_{g}^{-1}(j)},\mathbf{x}\rangle\bigg{|}^{2}d\mu_{2}^{\ast}(\bm{\lambda}_{1},\ldots,\bm{\lambda}_{k})$ $\displaystyle\leq\int\sup_{\mathbf{u}\in\operatorname{Dom}(\mathbf{y})}\sum_{\ell=1}^{k}|\langle\mathbf{u},\bm{\lambda}_{\ell}\rangle|^{2}d\mu_{2}^{\ast}(\bm{\lambda}_{1},\ldots,\bm{\lambda}_{k}),$ where $\mathbf{y}$ has coordinates $y_{j}=\sup_{h\in H}|\langle h\mathbf{v}_{j},\mathbf{x}\rangle|$ for $j\in[d_{1}]$. We immediately deduce $\displaystyle\int\sup_{g\in G}\lVert\operatorname{proj}_{W}(g\mathbf{v})\rVert_{2}^{2}d\mu(W)$ $\displaystyle\leq\int\sup_{\mathbf{u}\in\operatorname{Dom}(\mathbf{y})}\sum_{\ell=1}^{k}|\langle\mathbf{u},\bm{\lambda}_{\ell}\rangle|^{2}d\mu_{2}^{\ast}(\bm{\lambda}_{1},\ldots,\bm{\lambda}_{k})$ $\displaystyle\leq f_{k}^{\operatorname{alt}}(d_{1})^{2}\lVert\mathbf{y}\rVert_{2}^{2}\leq f_{k}^{\operatorname{alt}}(d_{1})^{2}\sum_{j=1}^{d_{1}}\lVert\mathbf{v}_{j}\rVert_{2}^{2}=f_{k}^{\operatorname{alt}}(d_{1})^{2}\lVert\mathbf{v}\rVert_{2}^{2}.$ Note that the above constructed measures in both cases are independent of $\mathbf{v}$. The second construction is only valid when $k<d_{1}$. Therefore since the $f$ values are clearly bounded by $1$, we have an upper bound of $f^{\operatorname{prim}}_{k}(d)\leq\max_{d_{1}d_{2}=d}(\min(f^{\operatorname{prim}}_{\lceil k/d_{1}\rceil}(d_{2}),f^{\operatorname{alt}}_{k}(d_{1})+\mathbbm{1}_{k\geq d_{1}})),$ as claimed. ∎ ## 4\. Permutation groups In this section, we establish upper bounds for $f^{\operatorname{sym}}_{k}(d)$ and $f^{\operatorname{alt}}_{k}(d)$, extending the previous construction [4, Section 3] for $k=1$. A useful high dimensional intuition is that, for small $k$, a random $k$-dimensional subspace of $\mathbb{R}^{d}$ has the property that _all_ its unit vectors have distribution of coordinate magnitudes similar to that of a random Gaussian vector. We first need the existence of a large dimension subspace of $\mathbb{R}^{d}$ with certain delocalization properties. We encode this through the following norm. ###### Definition 4.1. Given $\mathbf{v}\in\mathbb{R}^{d}$, let $\lVert\mathbf{v}\rVert_{T}^{2}=\sup_{\emptyset\subsetneq S\subseteq[d]}\log^{4}(2d/|S|)\sum_{j\in S}v_{j}^{2}$ and let $T^{\ast}=\\{\mathbf{t}\in\mathbb{R}^{d}:|\langle\mathbf{t},\mathbf{w}\rangle|\leq 1\text{ whenever }\lVert\mathbf{w}\rVert_{T}\leq 1\\}.$ ###### Remark. Note that $\lVert\cdot\rVert_{T}$ is a norm as it can be represented as a supremum of seminorms. Hence $\lVert\mathbf{w}\rVert_{T}=\sup_{t\in T^{\ast}}|\langle\mathbf{t},\mathbf{w}\rangle|.$ We next recall a classical lemma regarding the concentration of norms on Gaussian space (see e.g. [5]); we provide a short proof for convenience. ###### Lemma 4.2. There is an absolute constant $C>0$ so that for all $p\geq 1$, a Gaussian random vector $\mathbf{w}\sim\mathcal{N}(0,I_{d})$ satisfies $(\mathbb{E}_{w_{1}+\cdots+w_{d}=0}\lVert\mathbf{w}\rVert_{T}^{p})^{1/p}\leq(\mathbb{E}\lVert\mathbf{w}\rVert_{T}^{p})^{1/p}\leq\mathbb{E}\lVert\mathbf{w}\rVert_{T}+C\sqrt{p}\sup_{\mathbf{t}\in T^{\ast}}\lVert\mathbf{t}\rVert_{2}.$ ###### Proof. For first inequality note that $\mathbf{w}\sim\mathcal{N}(0,I_{d})$ can be written as $\mathbf{w^{\prime}}+G\mathbf{1}$ where $\mathbf{w^{\prime}}$ is drawn from $\mathcal{N}(0,I_{d})$ conditioned on having coordinate sum zero and $G\in\mathcal{N}(0,1)$ is independent of $\mathbf{w^{\prime}}$. Then by convexity note that $(\mathbb{E}\lVert\mathbf{w}\rVert_{T}^{p})^{1/p}=(\mathbb{E}\lVert\mathbf{w^{\prime}}+G\mathbf{1}\rVert_{T}^{p})^{1/p}\geq(\mathbb{E}_{\mathbf{w^{\prime}}}\lVert\mathbb{E}[\mathbf{w^{\prime}}+\mathbf{v^{\prime}}|\mathbf{w^{\prime}}]\rVert_{T}^{p})^{1/p}=(\mathbb{E}_{w_{1}+\cdots+w_{d}=0}\lVert\mathbf{w}\rVert_{T}^{p})^{1/p}.$ To prove the second inequality first note that $\lVert\mathbf{w}\rVert_{T}-\lVert\mathbf{v}\rVert_{T}\leq\lVert\mathbf{w}-\mathbf{v}\rVert_{T}=\sup_{\mathbf{t}\in T*}|\langle\mathbf{t},\mathbf{w}-\mathbf{v}\rangle|\leq\lVert\mathbf{w}-\mathbf{v}\rVert_{2}\sup_{t\in T*}\lVert\mathbf{t}\rVert_{2}.$ Therefore if $L=\sup_{\mathbf{t}\in T^{\ast}}\lVert\mathbf{t}\rVert_{2}$ then $\mathbf{w}\mapsto\lVert\mathbf{w}\rVert_{T}$ is an $L$-Lipschitz function with respect to Euclidean distance. Therefore by Gaussian concentration for Lipschitz functions (see e.g. [1, p. 125]) we have that $\mathbb{P}[|\lVert\mathbf{w}\rVert_{T}-\mathbb{E}[\lVert\mathbf{w}\rVert_{T}]|\geq t]\leq 2\exp(-ct^{2}/L^{2})$ where $c$ is an absolute constant. Using standard moment bounds for sub- Gaussian random variables (see e.g. [8, Proposition 2.5.2]), we find that $(\mathbb{E}|\lVert\mathbf{w}\rVert_{T}-\mathbb{E}\lVert\mathbf{w}\rVert_{T}|^{p})^{1/p}\leq C\sqrt{p}\sup_{\mathbf{t}\in T^{\ast}}\lVert\mathbf{t}\rVert_{2}$ for an absolute constant $C>0$. Finally, Minkowski’s inequality implies that $(\mathbb{E}\lVert\mathbf{w}\rVert_{T}^{p})^{1/p}\leq\mathbb{E}\lVert\mathbf{w}\rVert_{T}+(\mathbb{E}|\lVert\mathbf{w}\rVert_{T}-\mathbb{E}\lVert\mathbf{w}\rVert_{T}|^{p})^{1/p}$ and therefore the result follows. ∎ We now prove an upper bound for $\mathbb{E}[\lVert\mathbf{w}\rVert_{T}]$. ###### Lemma 4.3. A Gaussian random vector $\mathbf{w}\sim\mathcal{N}(0,I_{d})$ satisfies $\mathbb{E}\lVert\mathbf{w}\rVert_{T}\lesssim\sqrt{d}$. ###### Proof. Recall $\mathbf{w}_{i}^{\succ}$ from Definition 2.4. We have $\displaystyle\mathbb{E}(w_{i}^{\succ})^{2}$ $\displaystyle=\int_{0}^{\infty}\mathbb{P}[w_{i}^{\succ}\geq\sqrt{t}]dt\leq\int_{0}^{\infty}\min\bigg{(}1,\binom{d}{i}(2e^{-t/2})^{i}\bigg{)}dt$ $\displaystyle\leq\int_{0}^{\infty}\min(1,(2de^{1-t/2}/i)^{i})dt\lesssim\log(2d/i).$ Therefore $\displaystyle(\mathbb{E}\lVert\mathbf{w}\rVert_{T})^{2}$ $\displaystyle\leq\mathbb{E}\lVert\mathbf{w}\rVert_{T}^{2}\leq\sum_{i=1}^{d}\log^{4}(2d/i)(w_{i}^{\succ})^{2}\lesssim\sum_{i=1}^{d}\log^{5}(2d/i)$ $\displaystyle\leq d\int_{0}^{1}\log(2/x)^{5}~{}dx=d\int_{0}^{\infty}(y+\log 2)^{5}e^{-y}~{}dy\lesssim d.\qed$ We are in position to derive a high-probability version. ###### Lemma 4.4. With probability at least $1-\exp(-2d/(\log d)^{4})$, a standard Gaussian vector $\mathbf{w}\sim\mathcal{N}(0,I_{d})$ satisfies $\lVert\mathbf{w}\rVert_{T}\lesssim\sqrt{d}$. In fact, the same is true after conditioning $\mathbf{w}$ to have coordinate sum $0$. ###### Proof. Note that if $\mathbf{t}\in T^{\ast}$, then $\lVert\mathbf{t}\rVert_{2}^{2}=\lVert\mathbf{t}\rVert_{T}\bigg{|}\bigg{\langle}\mathbf{t},\frac{\mathbf{t}}{\lVert\mathbf{t}\rVert_{T}}\bigg{\rangle}\bigg{|}\leq\lVert\mathbf{t}\rVert_{T}\leq\log^{2}(2d)\lVert\mathbf{t}\rVert_{2}.$ Hence $\sup_{\mathbf{t}\in T^{\ast}}\lVert\mathbf{t}\rVert_{2}\leq\log^{2}(2d).$ To deduce the claimed bound, note that $\displaystyle\mathbb{P}[\lVert\mathbf{w}\rVert_{T}\geq K\sqrt{d}]$ $\displaystyle\leq(K\sqrt{d})^{-p}\mathbb{E}[\lVert\mathbf{w}\rVert_{T}^{p}]$ $\displaystyle\leq(K\sqrt{d})^{-p}(\mathbb{E}\lVert\mathbf{w}\rVert_{T}+C\sqrt{p}\sup_{\mathbf{t}\in T^{\ast}}\lVert\mathbf{t}\rVert_{2})^{p}$ $\displaystyle\leq(K\sqrt{d})^{-p}(C^{\prime}\sqrt{d}+C^{\prime}\sqrt{p}\log^{2}(2d))^{p}$ for appropriate absolute constants $C,C^{\prime}>0$, using Lemmas 4.2 and 4.3 and the above inequality. Letting $p=d/(\log d)^{4}$ and $K>0$ be a sufficiently large absolute constant yields $\mathbb{P}[\lVert\mathbf{w}\rVert_{T}\geq K\sqrt{d}]\leq\exp(-2p),$ as desired. The same holds is we condition on sum $0$, using the moment bound for the conditional variable derived in Lemma 4.2 instead. ∎ ###### Lemma 4.5. There is a $\lceil d/(\log d)^{4}\rceil$-dimensional subspace of the hyperplane $\mathbf{1}^{\perp}$ in $\mathbb{R}^{d}$ such that each of its unit vectors $\mathbf{v}$ satisfies $\lVert\mathbf{v}\rVert_{T}\lesssim 1.$ ###### Proof. We can assume $d$ is sufficiently large. Let $k=\lceil d/(\log d)^{4}\rceil$, and consider a uniformly random $k$-dimensional subspace $W$ of $\mathbf{1}^{\perp}$. Let $U$ be a $d\times k$ matrix whose columns form an orthonormal basis of $W$, chosen uniformly at random. By a standard volume packing argument (e.g., see [7, Lemma 4.3]), there exists $\mathcal{N}\subset\mathbb{S}(\mathbb{R}^{k})$ with $\lvert\mathcal{N}\rvert\leq 6^{k}$ such that for every $\mathbf{v}\in\mathbb{S}(\mathbb{R}^{k})$ there is $\mathbf{v}^{\prime}\in\mathcal{N}$ so that $\lVert\mathbf{v}-\mathbf{v}^{\prime}\rVert_{2}\leq 1/2$. Thus if $\mathbf{u}$ is a unit vector in the direction of $\mathbf{v}-\mathbf{v}^{\prime}$, we have $\lVert U\mathbf{v}\rVert_{T}\leq\lVert U\mathbf{v}^{\prime}\rVert_{T}+\lVert U(\mathbf{v}-\mathbf{v}^{\prime})\rVert_{T}\leq\lVert U\mathbf{v}^{\prime}\rVert_{T}+\frac{1}{2}\lVert U\mathbf{u}\rVert_{T}.$ We deduce $\sup_{\mathbf{v}\in\mathbb{S}(\mathbb{R}^{k})}\lVert U\mathbf{v}\rVert_{T}\leq\sup_{\mathbf{v}^{\prime}\in\mathcal{N}}\lVert U\mathbf{v}^{\prime}\rVert_{T}+\frac{1}{2}\sup_{\mathbf{u}\in\mathbb{S}(\mathbb{R}^{k})}\lVert U\mathbf{u}\rVert_{T}$ and thus $\sup_{\mathbf{v}\in\mathbb{S}(\mathbb{R}^{k})}\lVert U\mathbf{v}\rVert_{T}\leq 2\sup_{\mathbf{v}^{\prime}\in\mathcal{N}}\lVert U\mathbf{v}^{\prime}\rVert_{T}.$ Now fix some $\mathbf{v}\in\mathcal{N}$. Note the distribution of $U\mathbf{v}$ is uniform among unit vectors in $\mathbf{1}^{\perp}$ since $W$ was chosen uniformly. Now note that for any constant $C$ we have that $\mathbb{P}[\lVert U\mathbf{v}\rVert_{T}\geq C]=\mathbb{P}[\lVert\mathbf{G}/\lVert\mathbf{G}\rVert_{2}\rVert_{T}\geq C]$ where $\mathbf{G}\sim N(0,I_{d}-(\mathbf{1}^{T}\mathbf{1})/d)$. Now since $\mathbf{G}/\lVert\mathbf{G}\rVert_{2}$ and $\lVert\mathbf{G}\rVert_{2}$ are independent we have that $\displaystyle\mathbb{P}[\lVert\mathbf{G}/\lVert\mathbf{G}\rVert_{2}\rVert_{T}\geq C]$ $\displaystyle=\mathbb{P}[\lVert\mathbf{G}\rVert_{2}\leq 2\sqrt{d}]^{-1}\mathbb{P}[\lVert\mathbf{G}/\lVert\mathbf{G}\rVert_{2}\rVert_{T}\geq C\text{ and }\lVert\mathbf{G}\rVert_{2}\leq 2\sqrt{d}]$ $\displaystyle\leq 2\mathbb{P}[\lVert\mathbf{G}/\lVert\mathbf{G}\rVert_{2}\rVert_{T}\geq C\text{ and }\lVert\mathbf{G}\rVert_{2}\leq 2\sqrt{d}]$ $\displaystyle\leq 2\mathbb{P}[\lVert\mathbf{G}/\lVert\mathbf{G}\rVert_{2}\rVert_{T}\geq 2C\sqrt{d}].$ By Lemma 4.4, the last expression is at most $2\exp(-2d/(\log d)^{4})$. The result follows upon taking the union bound over at most $6^{k}$ vectors in $\mathcal{N}$, since $6<e^{2}$. ∎ Finally, we will need a form of Selberg’s inequality (see [3, Chapter 27, Theorem 1]). ###### Lemma 4.6. For $\mathbf{v}_{1},\ldots,\mathbf{v}_{m}\in\mathbb{C}^{d}$ we have that $\sup_{\mathbf{w}\in\mathbb{S}(\mathbb{C}^{d})}\sum_{i=1}^{m}|\langle\mathbf{w},\mathbf{v}_{i}\rangle|^{2}\leq\sup_{i\in[m]}\sum_{j=1}^{m}|\langle\mathbf{v}_{i},\mathbf{v}_{j}\rangle|.$ Now we prove Proposition 2.9, which recall says that for $k\leq d/(\log d)^{5}$, one has $f_{k}^{\operatorname{sym}}(d)\leq f_{k}^{\operatorname{alt}}(d)\lesssim 1/\sqrt{\log(d/k)}.$ The first inequality is immediate as the set of allowable $\mu$’s in the definition of $f_{k}^{\operatorname{alt}}$ is a subset of those of $f_{k}^{\operatorname{sym}}$. So we just need to prove the second inequality. ###### Proof of Proposition 2.9. Let $\mathbf{e}_{i}$ be the $i$-th coordinate vector. For each $j$ with $k\leq 2^{j}/(\log 2^{j})^{4}\leq d$, we apply Lemma 4.5 to the space $V_{j}=\operatorname{span}_{\mathbb{R}}\\{\mathbf{e}_{1},\ldots,\mathbf{e}_{2^{j}}\\}$. Here the $T$-norm is defined with respect to this $2^{j}$-dimensional space. In particular, there exists a $k$-dimensional (real) subspace of the orthogonal complement of $\mathbf{e}_{1}+\cdots+\mathbf{e}_{2^{j}}$ within $V_{j}$, call it $W_{j}$, so that every unit vector $\mathbf{u}\in W_{j}$ satisfies $\sum_{i\in S}u_{i}^{2}\lesssim\frac{1}{\log^{4}(2^{j+1}/|S|)}$ for every nonempty $S\subseteq[2^{j}]$. Let $V_{j}^{\prime}=\operatorname{span}_{\mathbb{C}}V_{j}$ and $W_{j}^{\prime}=\operatorname{span}_{\mathbb{C}}W_{j}$. We immediately deduce that every unit vector $\mathbf{u}\in W_{j}^{\prime}$ satisfies $\sum_{i\in S}|u_{i}|^{2}\lesssim\frac{1}{\log^{4}(2^{j+1}/|S|)}$ (4.1) since we can write it as $\mathbf{u}=\alpha\mathbf{u}_{r}+\beta\sqrt{-1}\mathbf{u}_{c}$ where $\mathbf{u}_{r},\mathbf{u}_{c}\in W_{j}$ are real unit vectors and $\alpha,\beta\in\mathbb{R}$ satisfy $\alpha^{2}+\beta^{2}=1$. Now we construct our random subspace as follows: let $W=W_{j}$ where $j$ is a random integer uniformly chosen from $J=\\{\lceil\log_{2}(2k\log^{4}d)\rceil,\ldots,\lfloor\log_{2}d\rfloor\\}.$ Let $\mu$ be the probability measure on $\operatorname{Gr}_{\mathbb{C}}(k,n)$ that gives $W$. For every $\mathbf{v}\in\mathbb{S}(\mathbb{C}^{d})$, we have $\sup_{\gamma\in\Gamma_{d}}\lVert\operatorname{proj}_{W}(\gamma\mathbf{v})\rVert_{2}=\sup_{\gamma\in\Gamma_{d}}\sup_{\mathbf{w}\in\mathbb{S}(W)}|\langle\gamma\mathbf{v},\mathbf{w}\rangle|=\sup_{\mathbf{w}\in\mathbb{S}(W)}\langle\mathbf{v}^{\succ},\mathbf{w}^{\succ}\rangle.$ Therefore $\int\sup_{\gamma\in\Gamma_{d}}\lVert\operatorname{proj}_{W}(\gamma\mathbf{v})\rVert_{2}^{2}d\mu(W)=\frac{1}{|J|}\sum_{j\in J}\sup_{\gamma\in\Gamma_{d}}\lVert\operatorname{proj}_{W_{j}}(\gamma\mathbf{v})\rVert_{2}^{2}=\frac{1}{|J|}\sum_{j\in J}\sup_{\mathbf{w}\in\mathbb{S}(W_{j})}\langle\mathbf{v}^{\succ},\mathbf{w}^{\succ}\rangle^{2}.$ Let $\mathbf{w}_{j}^{\prime}\in\mathbb{S}(W_{j})$ be such that $\sup_{\mathbf{w}\in\mathbb{S}(W_{j})}\langle\mathbf{v}^{\succ},\mathbf{w}^{\succ}\rangle^{2}=\langle\mathbf{v}^{\succ},(\mathbf{w}_{j}^{\prime})^{\succ}\rangle^{2},$ which exists by compactness. For $i,j\in J$ with $i\geq j$, we have $|\langle(\mathbf{w}_{i}^{\prime})^{\succ},(\mathbf{w}_{j}^{\prime})^{\succ}\rangle|\leq\lVert\operatorname{proj}_{V_{i}}((\mathbf{w}_{j}^{\prime})^{\succ})\rVert_{2}\lesssim\frac{1}{\log^{2}(2^{i+1}/2^{j})}.$ The first inequality follows from $\mathbf{w}_{j}^{\prime}\in V_{j}$, which implies $(\mathbf{w}_{j}^{\prime})^{\succ}\in V_{j}$. The second follows from 4.1 applied to $\mathbf{w}_{j}^{\prime}$ and $S$ a subset of $[2^{i}]$ composed of the $2^{j}$ largest magnitude coordinates of $\mathbf{w}_{j}^{\prime}$. Applying Lemma 4.6, we deduce $\displaystyle\int\sup_{\gamma\in\Gamma_{d}}\lVert\operatorname{proj}_{W}(\gamma\mathbf{v})\rVert_{2}^{2}d\mu(W)$ $\displaystyle=\frac{1}{|J|}\sum_{j\in J}\langle\mathbf{v}^{\succ},(\mathbf{w}_{j}^{\prime})^{\succ}\rangle^{2}\leq\sup_{i\in J}\frac{1}{|J|}\sum_{j\in J}|\langle(\mathbf{w}_{i}^{\prime})^{\succ},(\mathbf{w}_{j}^{\prime})^{\succ}\rangle|$ $\displaystyle\lesssim\frac{1}{|J|}\bigg{(}\sum_{j\in J,j\geq i}\frac{1}{\log^{2}(2^{j+1}/2^{i})}+\sum_{j\in J,j<i}\frac{1}{\log^{2}(2^{i+1}/2^{j})}\bigg{)}\lesssim\frac{1}{|J|}.$ This $\mu$ thus shows that $f_{k}^{\operatorname{alt}}(d)\lesssim 1/\sqrt{\log(d/k)}$. ∎ ## 5\. Primitive representations We now turn to the case of bounding $f_{k}^{\operatorname{prim}}(d)$. First, we show that if the group $G\leqslant\mathsf{U}(\mathbb{R}^{d})$ is sufficiently small, then a random basis achieves the necessary bound for $f_{k}^{\operatorname{prim}}(d)$. This is a minor modification of [4, Proposition 4.1]. ###### Proposition 5.1. Let $G\leqslant\mathsf{U}(\mathbb{C}^{d})$. Suppose that $[G:Z_{d}\cap G]\leq e^{d/\log d}$, where $Z_{d}:=\\{\lambda I_{d}:|\lambda|=1\\}$. Then for $k\in[d]$ there exists a probability measure $\mu$ on $\operatorname{Gr}_{\mathbb{C}}(k,d)$ such that $\int\sup_{g\in G}\lVert\operatorname{proj}_{W}(g\mathbf{v})\rVert_{2}^{2}d\mu(W)\lesssim\frac{1}{\log(2d/k)}\lVert\mathbf{v}\rVert_{2}^{2}$ for all $\mathbf{v}\in\mathbb{C}^{d}$. ###### Proof. We let $\mu$ be the uniform measure on $\operatorname{Gr}_{\mathbb{C}}(k,d)$. By scaling, we may assume that $\mathbf{v}$ is a unit vector. Furthermore let $W^{\prime}$ be the subspace generated by the first $k$ coordinate vectors $\mathbf{e}_{1},\ldots,\mathbf{e}_{k}$. Note that $\displaystyle\mathbb{P}_{W}\bigg{[}\sup_{g\in G}\lVert\operatorname{proj}_{W}(g\mathbf{v})\rVert_{2}\geq t\bigg{]}$ $\displaystyle\leq e^{d/\log d}\mathbb{P}_{W}[\lVert\operatorname{proj}_{W}(\mathbf{v})\rVert_{2}\geq t]$ $\displaystyle\leq e^{d/\log d}\mathbb{P}_{\mathbf{v}^{\prime}\in\mathbb{S}(\mathbb{C}^{d})}[\lVert\operatorname{proj}_{W^{\prime}}(\mathbf{v}^{\prime})\rVert_{2}\geq t]$ using a union bound and then orthogonal invariance. Now note that $\mathbb{E}[\lVert\operatorname{proj}_{W^{\prime}}(\mathbf{v}^{\prime})\rVert_{2}]^{2}\leq\mathbb{E}[\lVert\operatorname{proj}_{W^{\prime}}(\mathbf{v}^{\prime})\rVert_{2}^{2}]=k/d$ and that $\lVert\operatorname{proj}_{W^{\prime}}(\mathbf{v^{\prime}})\rVert$ is a $1$-Lipschitz function of $\mathbf{v}^{\prime}$. Therefore by Lévy concentration on the sphere we have that $\mathbb{P}_{\mathbf{v}^{\prime}\in\mathbb{S}(\mathbb{C}^{d})}[\lVert\operatorname{proj}_{W^{\prime}}(\mathbf{v}^{\prime})\rVert_{2}\geq\sqrt{k/d}+C/\sqrt{\log d}]\leq e^{-2d/\log d}$ for a suitably large absolute constant $C$. Finally, using $\sqrt{k/d}\lesssim 1/\sqrt{\log(2d/k)}$ and using the bound $\lVert\operatorname{proj}_{W^{\prime}}(\mathbf{v^{\prime}})\rVert_{2}\leq 1$, the desired result follows immediately. ∎ We need the following key group theoretic result from Green [4], which in turn builds on ideas from Collins’ work on optimal bounds for Jordan’s theorem [2]. Roughly, it says that if $[G:Z_{d}\cap G]$ is large then $G$ has a large normal alternating subgroup. The first part of the following theorem is [4, Proposition 4.2], while the rest is implicit in the proof of [4, Proposition 1.11]. ###### Theorem 5.2 ([4, Section 4]). Let $G\leqslant\mathsf{U}(\mathbb{C}^{d})$ be primitive and suppose that $[G:Z_{d}\cap G]\geq e^{d/\log d}$. If $d$ is sufficiently large then all of the following hold. 1. (1) $G$ has a normal subgroup isomorphic to the alternating group $A_{n}$ for some $n\gtrsim d/(\log d)^{4}$. 2. (2) $G$ has a subgroup of index at most $2$ of the form $A_{n}\times H$, with the same $n$. 3. (3) The resulting representation $\rho:A_{n}\times H\hookrightarrow G\hookrightarrow\mathsf{U}(\mathbb{C}^{d})$ decomposes into irreducible representations, at least one of which (call it $\rho_{1}$) is of the form $\rho_{1}\simeq\psi\otimes\psi^{\prime}$, where $\psi^{\prime}$ is an irreducible representation of $H$ and $\psi$ is the representation of $A_{n}$ acting via permutation of coordinates on $\\{\mathbf{z}\in\mathbb{C}^{n}:z_{1}+\cdots+z_{n}=0\\}$. We are now in position to prove Proposition 2.10, which recall says that there is an absolute constant $c>0$ such that for every $k\leq cd/(\log d)^{4}$ we have $f_{k}^{\operatorname{prim}}(d)\lesssim\sup_{d^{\prime}\geq cd/(\log d)^{4}}f^{\operatorname{alt}}_{k}(d^{\prime}).$ The proof mirrors that of [4, Proposition 1.11], but we correct an error of Green ([4, p. 20]) involving an incorrect orthogonality identity. This erroneous deduction is replaced by an argument which still allows one to reduce the primitive case to the alternating case. ###### Proof of Proposition 2.10. We may assume $d$ is sufficiently large. If $[G:Z_{d}\cap G]\leq e^{d/\log d}$, then the result follows by Proposition 5.1. So we can assume $[G:Z_{d}\cap G]\geq e^{d/\log d}$, and thus by Theorem 5.2, $G$ has a normal subgroup isomorphic to $A_{n}$ for some $n\gtrsim d/(\log d)^{4}$ and that $G$ has a subgroup of index at most $2$ which is of the form $A_{n}\times H$. If the index is $2$, let $\tau$ be the nontrivial right coset representative of $A_{n}\times H$ in $G$ (otherwise just let $\tau$ be the identity). Note that $\displaystyle\sup_{g\in G}\lVert\operatorname{proj}_{W}(g\mathbf{v})\rVert_{2}^{2}$ $\displaystyle\leq\sup_{g\in A_{n}\times H}\lVert\operatorname{proj}_{W}(g\mathbf{v})\rVert_{2}^{2}+\sup_{g\in A_{n}\times H}\lVert\operatorname{proj}_{W}(g\tau\mathbf{v})\rVert_{2}^{2},$ so it is easy to see that, up to losing a constant factor, we may reduce to studying groups of the form $G=A_{n}\times H$ where $n\gtrsim d/(\log d)^{4}$ (but note that the representation may no longer be primitive, or even irreducible). Now Theorem 5.2 shows that the representation $\rho:A_{n}\times H\to\mathsf{U}(\mathbb{C}^{d})$ coming from this setup has an irreducible component of the form $\rho_{1}\simeq\psi\otimes\psi^{\prime}$, where $\psi^{\prime}$ is an irreducible representation of $H$ and $\psi$ is the representation of $A_{n}$ acting via permutation of coordinates on $\\{\mathbf{z}\in\mathbb{C}^{n}:z_{1}+\cdots+z_{n}=0\\}$. Note that $\dim\rho_{1}\geq\dim\psi=n-1\gtrsim d/(\log d)^{4}$, so $\dim\rho_{1}\geq k$ provided that $c>0$ is sufficiently small. We will choose a $k$-dimensional subspace of the irreducible component $\rho_{1}$. We explicitly present this situation as follows. Let $V^{\prime}$ be the space acted on by $\psi^{\prime}$ (unitarily). Consider $V=\mathbf{1}^{\perp}\subseteq\mathbb{C}^{n}$, and consider the spaces $V\otimes V^{\prime}\subseteq\mathbb{C}^{n}\otimes V^{\prime}$, which has a natural unitary structure given by the tensor product. Note $\psi$ acts on $V$ by permutation of coordinates when represented in $\mathbb{C}^{n}$. Every vector in $V\otimes V^{\prime}$ is spanned by pure tensors $\mathbf{v}\otimes\mathbf{v}^{\prime}$ where $\mathbf{v}$ has zero coordinate sum, and $\rho_{1}((a,h))$ acts by $\psi(a)\otimes\psi^{\prime}(h)$ on pure tensors. In fact, we can extend this action to all of $\mathbb{C}^{n}\otimes V^{\prime}$ in the natural way (and the resulting representation is isomorphic to a direct sum of $\rho_{1}$ and $\operatorname{triv}_{A_{n}}\otimes\psi^{\prime}$). At this point, the analysis will be similar to that in the proof of Proposition 2.8. Let $\nu$ be the measure on $\operatorname{Gr}_{\mathbb{C}}(k,n)$ which is guaranteed by Definition 2.6 (so is supported on subspaces of $V\subseteq\mathbb{C}^{n}$) and consider the measure which is supported on a single atom in $\operatorname{Gr}_{\mathbb{C}}(1,V^{\prime})$ in the direction of a fixed unit vector $\mathbf{x}$. Let $\mu$ be the tensor of these two measures, i.e., if $\nu^{\ast}$ samples $k$ orthonormal (sum zero) vectors $\mathbf{u}_{1},\ldots,\mathbf{u}_{k}$ then we choose the subspace with basis $\mathbf{u}_{1}\otimes\mathbf{x},\ldots,\mathbf{u}_{k}\otimes\mathbf{x}$. Now consider some $\mathbf{v}$ in the space $V\otimes V^{\prime}\subseteq\mathbb{C}^{n}\otimes V^{\prime}$, and write it as $\mathbf{v}=\sum_{j=1}^{n}\mathbf{e}_{j}\otimes\mathbf{v}_{j}^{\prime}$ where the $\mathbf{e}_{j}$ is the $j$-th coordinate vector of $\mathbb{C}^{n}$. In fact, the $\mathbf{v}_{j}^{\prime}$ must add up to $\mathbf{0}\in V^{\prime}$. We see that $\lVert\mathbf{v}\rVert_{2}^{2}=\sum_{j=1}^{n}\lVert\mathbf{v}_{j}^{\prime}\rVert_{2}^{2}.$ We have $\displaystyle\int\sup_{g\in A_{n}\times H}$ $\displaystyle\lVert\operatorname{proj}_{W}(g\mathbf{v})\rVert_{2}^{2}d\mu(W)$ $\displaystyle=\int\sup_{a\in A_{n},h\in H}\sum_{\ell=1}^{k}\bigg{|}\bigg{\langle}\sum_{j=1}^{n}\psi(a)\mathbf{e}_{j}\otimes\psi^{\prime}(h)\mathbf{v}_{j}^{\prime},\mathbf{w}_{\ell}\bigg{\rangle}\bigg{|}^{2}d\mu^{\ast}(\mathbf{w}_{1},\ldots,\mathbf{w}_{k})$ $\displaystyle=\int\sup_{a\in A_{n},h\in H}\sum_{\ell=1}^{k}\bigg{|}\sum_{j=1}^{n}\langle\psi(a)\mathbf{e}_{j},\mathbf{u}_{\ell}\rangle\langle\psi^{\prime}(h)\mathbf{v}_{j}^{\prime},\mathbf{x}\rangle\bigg{|}^{2}d\nu^{\ast}(\mathbf{u}_{1},\ldots,\mathbf{u}_{k})$ $\displaystyle\leq\int\sup_{\mathbf{w}\in\operatorname{Dom}(\mathbf{y})}\sum_{\ell=1}^{k}|\langle\mathbf{w},\mathbf{u}_{\ell}\rangle|^{2}d\nu^{\ast}(\mathbf{u}_{1},\ldots,\mathbf{u}_{k})$ $\displaystyle\leq f^{\operatorname{alt}}_{k}(n)^{2}\lVert\mathbf{y}\rVert_{2}^{2},$ where $\mathbf{y}\in\mathbb{C}^{n}$ satisfies $y_{j}=\sup_{h\in H}|\langle\psi^{\prime}(h)\mathbf{v}_{j}^{\prime},\mathbf{x}\rangle|$. The first inequality follows by noting that $\langle\psi(a)\mathbf{e}_{j},\mathbf{u}_{\ell}\rangle$ as $j$ varies simply records the coordinates of $\mathbf{u}_{\ell}$ in some permutation, and by considering $\mathbf{w}=(w_{1},\ldots,w_{n})$ defined via $w_{j}=\langle\psi^{\prime}(h)\mathbf{v}_{j}^{\prime},\mathbf{x}\rangle$, which is clearly on $\operatorname{Dom}(\mathbf{y})$. Now we see $\int\sup_{g\in A_{n}\times H}\lVert\operatorname{proj}_{W}(g\mathbf{v})\rVert_{2}^{2}d\mu(W)\leq f^{\operatorname{alt}}_{k}(n)^{2}\lVert\mathbf{y}\rVert_{2}^{2}\leq f^{\operatorname{alt}}_{k}(n)^{2}\sum_{j=1}^{n}\lVert\mathbf{v}_{j}^{\prime}\rVert_{2}^{2}=f^{\operatorname{alt}}_{k}(n)^{2}\lVert\mathbf{v}\rVert_{2}^{2}.\qed$ This completes all the components of the proof of Theorem 1.3. ## 6\. Real subspaces We already proved Theorem 1.3, which finds a complex subspace. Now we use it to deduce Theorem 1.2, which gives a real subspace. We will apply the following version of the restricted invertibility theorem, which is a special case of [6, Theorem 6]. We write $s_{1}(M)\geq s_{2}(M)\geq\cdots$ for the singular values of a matrix $M$. ###### Theorem 6.1 ([6, Theorem 6]). Let $M$ be a real $2k\times 4k$ matrix of rank $2k$. There exists $S\subseteq[4k]$ with $|S|=k$ such that $M_{S}$, the restriction of $M$ to the columns $S$, satisfies $s_{k}(M_{S})\gtrsim\sqrt{\frac{\sum_{j=3k/2}^{4k}s_{j}(M)^{2}}{k}}.$ ###### Proof of Theorem 1.2. Let $2k\leq d/(\log d)^{C}$, where $C$ is as in Theorem 1.3. By embedding $X$ in $\mathbb{S}(\mathbb{C}^{d})$ and using Theorem 1.3 we can find a $2k$-dimensional complex subspace $W$ of $\mathbb{C}^{d}$ such that $\sup_{\mathbf{x}\in X}\lVert\operatorname{proj}_{W}\mathbf{x}\rVert_{2}\lesssim 1/\sqrt{\log(d/k)}.$ Let $\mathbf{v}_{1},\ldots,\mathbf{v}_{2k}$ be a unitary basis for the subspace $W$ and let the matrix with these columns be denoted by $B$. Now consider the matrix $M$ which has $4k$ columns which are $\operatorname{Re}\mathbf{v}_{1},\ldots,\operatorname{Re}\mathbf{v}_{2k}$ and $\operatorname{Im}\mathbf{v}_{1},\ldots,\operatorname{Im}\mathbf{v}_{2k}$. Note that $M$ has $s_{2k}(M)\geq 1/\sqrt{2}$ as any vectors in $\mathbb{C}^{4k}$ satisfying $iv_{j}=v_{j+2k}$ have $\lVert M\mathbf{v}\rVert=\lVert\mathbf{v}\rVert/\sqrt{2}$. Therefore by Theorem 6.1 one can select $k$ columns such that the matrix $N$ with those $k$ columns satisfies $s_{k}(N)\gtrsim 1.$ Now consider any unit vector $\mathbf{v}$ in the image of $N$. Such a vector can be represented as $\mathbf{v}=N\mathbf{w}$ where $\lVert\mathbf{w}\rVert\lesssim 1$. It therefore suffices to prove that $\sup_{\mathbf{x}\in X,\mathbf{w}\in\mathbb{S}(\mathbb{R}^{k})}|\langle N\mathbf{w},\mathbf{x}\rangle|\lesssim 1/\sqrt{\log(d/k)}.$ To see this separate $N$ into $N_{1}$ and $N_{2}$ where $N_{1}$ corresponds to columns chosen from the real parts of vectors $\mathbf{v}_{i}$ and the columns are chosen from the complex parts of $\mathbf{v}_{i}$. Let these have $\ell$ and $k-\ell$ columns respectively. Then $\displaystyle\sup_{\mathbf{x}\in X,\mathbf{w}\in\mathbb{S}(\mathbb{R}^{k})}|\langle N\mathbf{w},\mathbf{x}\rangle|$ $\displaystyle\leq\sup_{\mathbf{x}\in X,\mathbf{w}\in\mathbb{S}(\mathbb{R}^{\ell})}|\langle N_{1}\mathbf{w},\mathbf{x}\rangle|+\sup_{\mathbf{x}\in X,\mathbf{w}\in\mathbb{S}(\mathbb{R}^{k-\ell})}|\langle N_{2}\mathbf{w},\mathbf{x}\rangle|$ $\displaystyle\leq 2\sup_{\mathbf{x}\in X,\mathbf{w}\in\mathbb{S}(\mathbb{C}^{k})}|\langle B\mathbf{w},\mathbf{x}\rangle|$ $\displaystyle\lesssim 1/\sqrt{\log(d/k)}.\qed$ ## 7\. Lower Bound Finally, we show a lower bound of $\Omega(1/\sqrt{\log(2d/k)})$, which demonstrates optimality of our results. ###### Proof of Theorem 1.5. We prove the real case; an analogous proof works over $\mathbb{C}$ by considering a suitably fine discretization of $\Gamma_{d}$, or we can repeat the proof in Section 6 to transfer a lower bound from real to complex. The claim for $k=1$ was already proved in [4, _Sharpness_ after Theorem 1.3] (see the construction at the beginning of this article right after Theorem 1.1). The case $k=1$ implies the result also for $k\leq d^{1-c}$ for any constant $c$, since we can project from $W$ onto a arbitrary 1-dimensional subspace of $W$. So from now on assume $k\geq d^{1/2}$. Consider the action of $G=\mathfrak{S}_{d}\ltimes(\mathbb{Z}/2\mathbb{Z})^{d}$ on $\mathbb{R}^{d}$ by permutation and signing. Let $\mathbf{a}=\left(\frac{1}{\sqrt{\lfloor k/2\rfloor+1}},\ldots,\frac{1}{\sqrt{d}},0,\ldots,0\right).$ Let $X$ be the $G$-orbit of $\mathbf{a}/\lVert\mathbf{a}\rVert_{2}$. Let $W$ be a $k$-dimensional subspace of $\mathbb{R}^{d}$. We wish to show $\sup_{\mathbf{x}\in X}\lVert\operatorname{proj}_{W}\mathbf{x}\rVert_{2}\gtrsim 1/\sqrt{\log(2d/k)}$. Let $\mathbf{y}=(y_{1},\ldots,y_{d})$ a uniform random vector in $\mathbb{S}(W)$. Let $\sigma_{i}=(\mathbb{E}y_{i}^{2})^{1/2}$. We have $\sigma_{1}^{2}+\cdots+\sigma_{d}^{2}=\mathbb{E}[y_{1}^{2}+\cdots+y_{d}^{2}]=1$ (7.1) and $\sigma_{i}^{2}=\frac{1}{k}\lVert\operatorname{proj}_{W}(\mathbf{e}_{i})\rVert^{2}\leq\frac{1}{k}.$ (7.2) Without loss of generality, assume that $1/\sqrt{k}\geq\sigma_{1}\geq\cdots\geq\sigma_{d}\geq 0$, so that $\sigma_{i}\leq 1/\sqrt{i}$ for each $i$. We claim that $a_{i}\geq\sqrt{\frac{2}{3}}\sigma_{i}\qquad\text{ for all }1\leq i\leq d-k/2.$ Indeed, for $i\leq k$, we have $a_{i}\geq 1/\sqrt{3k/2}\geq\sqrt{3/2}\sigma_{i}$. For $k<i\leq d-\lfloor k/2\rfloor$, we have $a_{i}=1/\sqrt{\lfloor k/2\rfloor+i}\geq\sigma_{i}\sqrt{i/(\lfloor k/2\rfloor+i)}\geq\sqrt{2/3}\sigma_{i}$. We have $\mathbb{E}|y_{i}|\gtrsim(\mathbb{E}y_{i}^{2})^{1/2}=\sigma_{i}$ since $y_{i}$ is distributed as the first coordinate of a random point on $\sigma_{i}\sqrt{k}\cdot\mathbb{S}(\mathbb{R}^{k})$. Putting everything together, we have $\displaystyle\lVert\mathbf{a}\rVert_{2}\,\sup_{\mathbf{x}\in X}\lVert\operatorname{proj}_{W}\mathbf{x}\rVert_{2}$ $\displaystyle\geq\sup_{g\in G}\lVert\operatorname{proj}_{W}g\mathbf{a}\rVert\geq\mathbb{E}\sup_{g\in G}\langle\mathbf{a},g\mathbf{y}\rangle$ $\displaystyle\geq\mathbb{E}\sum_{1\leq i\leq d}a_{i}|y_{i}|\gtrsim\sum_{i=1}^{d}a_{i}\sigma_{i}\gtrsim\sum_{i=1}^{d-k/2}\sigma_{i}^{2}\geq\frac{1}{2},$ where the final step uses 7.1 and 7.2. Thus $\sup_{\mathbf{x}\in X}\lVert\operatorname{proj}_{W}\mathbf{x}\rVert_{2}\gtrsim\frac{1}{\lVert\mathbf{a}\rVert_{2}}\gtrsim 1/\sqrt{\log(2d/k)}.\qed$ ## References * [1] Stéphane Boucheron, Gábor Lugosi, and Pascal Massart, _Concentration inequalities_ , Oxford University Press, Oxford, 2013, A nonasymptotic theory of independence, With a foreword by Michel Ledoux. * [2] Michael J. Collins, _On Jordan’s theorem for complex linear groups_ , J. Group Theory 10 (2007), 411–423. * [3] Harold Davenport, _Multiplicative number theory_ , third ed., Graduate Texts in Mathematics, vol. 74, Springer-Verlag, New York, 2000, Revised and with a preface by Hugh L. Montgomery. * [4] Ben Green, _On the width of transitive sets: Bounds on matrix coefficients of finite groups_ , Duke Math. J. 169 (2020), 551–578. * [5] Michel Ledoux and Michel Talagrand, _Probability in Banach spaces_ , Ergebnisse der Mathematik und ihrer Grenzgebiete (3) [Results in Mathematics and Related Areas (3)], vol. 23, Springer-Verlag, Berlin, 1991, Isoperimetry and processes. * [6] Assaf Naor and Pierre Youssef, _Restricted invertibility revisited_ , A journey through discrete mathematics, Springer, Cham, 2017, pp. 657–691. * [7] Mark Rudelson, _Recent developments in non-asymptotic theory of random matrices_ , Modern aspects of random matrix theory, Proc. Sympos. Appl. Math., vol. 72, Amer. Math. Soc., Providence, RI, 2014, pp. 83–120. * [8] Roman Vershynin, _High-dimensional probability_ , Cambridge Series in Statistical and Probabilistic Mathematics, vol. 47, Cambridge University Press, Cambridge, 2018, An introduction with applications in data science, With a foreword by Sara van de Geer.
# Statistical guided-waves-based SHM via stochastic non-parametric time series models Ahmad Amer Intelligent Structural Systems Laboratory (ISSL) Department of Mechanical, Aerospace and Nuclear Engineering Rensselaer Polytechnic Institute, Troy, NY, USA Email<EMAIL_ADDRESS>Fotis Kopsaftopoulos111Corresponding author. Intelligent Structural Systems Laboratory (ISSL) Department of Mechanical, Aerospace and Nuclear Engineering Rensselaer Polytechnic Institute, Troy, NY, USA Email<EMAIL_ADDRESS> ###### Abstract Damage detection in active-sensing, guided-waves-based Structural Health Monitoring (SHM) has evolved through multiple eras of development during the past decades. Nevertheless, there still exists a number of challenges facing the current state-of-the-art approaches, both in the industry as well as in research and development, including low damage sensitivity, lack of robustness to uncertainties, need for user-defined thresholds, and non-uniform response across a sensor network. In this work, a novel statistical framework is proposed for active-sensing SHM based on the use of ultrasonic guided waves. This framework is based on stochastic non-parametric time series models and their corresponding statistical properties in order to readily provide healthy confidence bounds and enable accurate and robust damage detection via the use of appropriate statistical decision making tests. Three such methods and corresponding statistical quantities (test statistics) along with decision making schemes are formulated and experimentally assessed via the use of three coupons with different levels of complexity: an Al plate with a growing notch, a Carbon fiber-reinforced plastic (CFRP) plate with added weights to simulate local damages, and the CFRP panel used in the Open Guided Waves project [1], all fitted with piezoelectric transducers and a pitch-catch configuration. The performance of the proposed methods is compared to that of state-of-the-art time-domain damage indices (DIs). The results demonstrate the increased sensitivity and robustness of the proposed methods, with better tracking capability of damage evolution compared to conventional approaches, even for damage-non-intersecting actuator-sensor paths. In particular, the $Z$ statistic emerges as the best damage detection metric compared to conventional DIs, as well as the other proposed statistics. This is attributed to the incorporation of experimental uncertainty in defining the $Z$ statistic, which results in a both sensitive and robust approach for damage detection. Overall, the proposed statistical methods exhibit greater damage sensitivity across different components, with enhanced robustness to uncertainty, as well as user-friendly application. ###### Contents 1. 1 Introduction 2. 2 The Statistical Damage Detection Framework 1. 2.1 The General Framework 2. 2.2 Overview of Non-parametric Time Series Representations 3. 2.3 The Single-set $F$ Statistic Method 4. 2.4 The Multiple-set Modified $F_{m}$ Statistic Method 5. 2.5 The Multiple-set $Z$ Statistic Method 6. 2.6 Reference State-of-the-Art Damage Indices 3. 3 Results and Discussion 1. 3.1 Test Case I: Damage Detection in an Aluminum Plate 1. 3.1.1 Test Setup, Damage Types and Data Acquisition 2. 3.1.2 Damage Detection Results 2. 3.2 Test Case II: Damage Detection in CFRP Plate 1. 3.2.1 Test Setup, Damage Types and Data Acquisition 2. 3.2.2 Damage Detection Results 3. 3.3 Test Case III: Damage Detection in the Open Guided-Waves CFRP Panel 1. 3.3.1 Test Setup, Damage Types and Data Acquisition 2. 3.3.2 Damage Detection Results 4. 4 Conclusions 5. A Damage Detection Summary Results: Notched Al Coupon 6. B Damage Detection Summary Results: CFRP Coupon 7. C Damage Detection Summary Results: OGW CFRP Panel ## 1 Introduction In the near future, Structural Health Monitoring (SHM) systems will be capable of implementing all four levels of SHM, namely: damage detection, localization, quantification and remaining useful life estimation (prognosis) [2, 3, 4, 5], with sustainable levels of performance in complex components under varying operational and environmental conditions. In order to reach this milestone, a number of challenges facing current SHM techniques needs to be addressed. These challenges originate from the deterministic nature of the majority of the currently-employed approaches, i.e. they do not allow for the extraction of appropriate confidence intervals for damage detection, localization and quantification [6, 7, 8]. This leads to the characterization of those techniques as inefficient in the face of uncertainty, stochastic time-variant and non-linear structural responses [9, 10, 11], as well as incipient damage types and complex failure modes that can be easily masked by the effects of varying states [12, 13]. Thus, there lies a need for the development of SHM frameworks, where proper understanding, modeling, and analysis of stochastic structural responses under varying states and damage characteristics is achieved for clearing the road towards achieving the aforementioned ultimate goal of SHM systems. Towards this end, many researchers have proposed the use of statistical distributions of damage- related features in devising SHM metrics and corresponding probabilities (also known as statistical and/or probabilistic SHM); for instance see [14, 15, 16, 17, 18, 19, 7] for probabilistic damage detection and localization, as well as [14, 20, 21, 22, 8] for probabilistic damage quantification). These approaches promise wider applicability due to the direct extraction of confidence bounds for detection, localization, and quantification, as well as present an alternative approach for SHM reliability quantification without the need for further non-destructive testing, such as that required to obtain the Probability of Detection (POD) [23, 24, 4, 25, 26]. In a more specific context, as the most fundamental level of SHM [6, 14, 27], damage detection has received significant attention throughout the last two decades. Within the framework of active-sensing guided-waves-based SHM, one of the most widely-used techniques for damage detection (and quantification) is the concept of the Damage/Health Index/Indicator (D/HI) [28, 29], where some features of the signal for an unknown structural state is compared to that coming from the healthy structure [30, 31]. To this end, the most-widely used DI-based approaches are based on the time delay of specific mode wave packets in the acousto-ultrasound signal, the amplitude/magnitude of the signal, and the energy content of the signals, all used as the features to differentiate between a healthy and a damaged structure [32, 33, 30, 4, 31, 34]. These approaches, thereon denoted as conventional DI-based methods, have been used extensively in the literature due to their simplicity (no experience needed in interpretation of results) and the allowance of a damage/no-damage paradigm, which may facilitate the decision-making stage [32]. However, there exists a number of challenges facing these types of methods when it comes to damage detection. Namely, their deterministic nature and the time-varying and non- linear structural responses within a structure can limit the applicability of such methods [6, 9, 10, 11]. In addition, the effect of complex damage types and their stochastic evolution can be masked under varying operational and environmental states, further inhibiting the effectiveness of such DIs in damage detection [6, 12, 13]. Other issues related to the conventional DI- based approach include the need for user-defined damage thresholds for damage detection [35, 36] and the phenomenon of saturation [37]. As such, the challenges facing the aforementioned methods have been tackled throughout the literature using different approaches, with most of them based on advanced variants of conventional DIs. Although these endeavors span many strategies, the most common approaches either enhance current time-domain DIs (see for instance [4, 38]), use frequency-domain or mixed-domain DIs (see for instance [32, 39, 35]), or use advanced signal processing/modelling as a preliminary step before calculating DIs (see [40, 41, 42]). Another family of techniques is based on baseline-free approaches (see [36, 43, 44]), which themselves can be further categorized into a number of approaches as will be discussed shortly. Although the proposed enhancements exhibit better damage detection performance compared to conventional DIs, no single technique seems to collectively address the current drawbacks of conventional DIs. The following discussion briefly outlines selected studies from each family of approaches, highlighting the advantages and drawbacks of each family of techniques. In the context of enhanced time-domain approaches, Janapati et al. [4] proposed a DI that depends solely on guided-wave propagation signal normalization and applied it to many identical coupons in order to pinpoint the source of variability between seemingly-identical SHM systems, and thus better understand the effect of uncertainties on time-domain DIs. Su and coworkers [39] compared three different DIs for fatigue damage characterization within an active-sensing acousto-ultrasonic SHM framework: the traditional time-of-flight delay and energy dissipation indices, and a novel index utilizing non-linear features of guided-wave signals, such as the second harmonic generation. They observed higher sensitivities, as well as better damage evolution-tracking using nonlinear features compared to the traditional approaches relying on linear ones. In addition, they concluded that analyzing the time-frequency domain instead of the time domain alone enhances damage detection ability, especially for early-stage cracks. Building upon that mixed-domain approach, Jin et al. [32] used DIs in both the time and the frequency domains. In addition, in order to address the uncertainties in each individual path DI due to noise and varying conditions, they proposed an arithmetic fusion algorithm, where DIs based on amplitude and energy, both generated in the time and the frequency domains (a total of four DIs), are each summed over all the actuator-sensor path signals coming from a steel plate in order to “visualize” fatigue crack growth. Although these endeavors were capable in addressing certain challenges facing DIs, they are still not probabilistic in nature, and are thus still prone to error due to uncertainties. In order to address this, and building upon the fact that the frequency domain may offer a different representation of the signal dynamics, many researchers studied the effect of damage on the energy of wavelets (coming from some type of wavelet transformation of the signals) by using the concept of entropy [45]. Basically, the Shannon entropy [46] is calculated for windows/parts of the signal and the observed changes are related to damage, as appropriate. As this method is essentially based on the probability distribution of the energy of each wavelet, it is more effective in capturing uncertainties compared to conventional approaches. However, defining damage/healthy thresholds, as well as the analysis of changes in entropy, are not straight forward [47], and user expertise is required in order to properly detect damage. Another approach was adopted by Qiu and coworkers [35], where time-domain (as-received signal amplitude) and frequency-domain (frequency-response function) variants of the DI were used as features to develop a Gaussian mixture model for guided-wave signals in the healthy case. Then, upon the acquisition of new feature values, the model is migrated and a statistical technique is used to measure the differences between the baseline and the migrated models. Applying this approach to a real-life fuselage component with a developing fatigue crack, they observed the enhanced sensitivity and better damage evolution tracking. Most importantly, they also concluded the superiority of this technique owing to the lack of requiring user-experience in defining detection thresholds. One drawback to this approach, however, is the complexity in defining the original Gaussian mixture model, which requires many data sets, and multiple steps, including k-means clustering and expectation maximization algorithms. In another attempt to enhance the detection capability of DIs under uncertainty, several researchers proposed baseline-free techniques, which do not require the presence of pre-sampled reference/healthy signals to compare to. In many of these techniques, either an instantaneous baseline is acquired from an identical path to the one being investigated, or from the reciprocal of that path, or the signal is reversed and analyzed for mode conversions and deviations from the original one, amongst other methods [43]. Although these approaches appear to be more robust to varying conditions owing to the lack of pre-sampled baseline signals, most of them require knowledge of dispersion curves for the components being monitored, and dictate sophisticated actuation strategies. In addition, it has been shown that, depending on the actuator- sensor path from which the signal is coming, the evolution of the DIs can proceed in a manner uncorrelated with damage evolution [8, 32], which clearly limits the applicability of many of these techniques to specific sensor network designs and simple boundary conditions. Although there are recent studies on averting this latter drawback [48, 44], these approaches still require experience in defining detection thresholds. Although the highlighted approaches show promise for enhancing the damage detection capability of DIs, from the discussion above it can be concluded that there exist no approaches capable of overcoming the aforementioned challenges facing DIs with a user-friendly way that can be widely applicable to different sensor networks and component designs. Namely, there still lies the need to develop damage detection techniques that overcome the following challenges, which are faced by the current DI-based approaches: * • Complexity in defining/calculating the damage detection metric. * • Lack of a straightforward approach in defining appropriate statistical damage thresholds, i.e. the need for historical data and user experience with specific damage cases. * • Poor performance in complex damage cases and non-linear structural responses (e.g. complex composite parts). * • Lack of robustness towards uncertainties originating from operational or environmental sources. * • Failure to follow damage evolution in some cases, either entirely (e.g. damage non-intersecting signals [8]), or starting from a certain damage size (e.g. saturation phenomenon [37]). In order to overcome these challenges, the use of non-parametric time series (NP-TS) models within a statistical framework is proposed herein in order to tackle damage detection under uncertainty. NP-TS models have been widely used in damage detection via vibration-based SHM [49, 50, 51] due to their stochastic nature, which inherently accounts for uncertainty and allows for the extraction of theoretical and experimental confidence intervals, avoiding the need for user-defined thresholds, based on statistical decision making schemes. Also, NP-TS frequency-domain representations of system dynamics can exhibit increased sensitivity to damage and entertain simplicity of application requiring little-to-no user experience [49, pp. 212]. Finally, as will be shown herein, NP-TS models can prove superior to conventional DIs in following damage evolution. Thus, the use of stochastic NP-TS representations for damage detection has the potential to enhance the detection performance with straight-forward applicability due to the ability of extracting “inherent thresholds” from the SHM metrics themselves, as well as simplicity of application. In a previous preliminary study [7], the authors applied NP-TS representations and statistical hypothesis testing (SHT) to an Al coupon and a stiffened panel showing the extraction of estimation confidence intervals from the metrics being used, as well as the enhanced detection capability with these stochastic models compared to DI-based approaches. In the current study, work on NP-TS models for damage detection in active-sensing acousto-ultrasound SHM is significantly expanded, and their performance in detecting damage over three different structural coupons is compared and experimentally assessed with that of two state-of-the-art DIs from the literature [4, 35]. To the author’s best of knowledge, the NP-TS-based damage detection metrics used herein have not been proposed previously within the framework of active-sensing guided-waves- based SHM. The main novel aspects of this study include: * (a) The introduction of a novel data-based statistical damage detection framework based on NP-TS models in active-sensing guided-waves-based SHM, as well as the expansion of two previously proposed methods by the authors and co-workers [52, 49]. * (b) The application of the proposed methods in two composite panels with different types of simulated damage, as well as in a notched Al coupon. * (c) The extraction of statistical confidence intervals and the detection of damage via appropriate statistical hypothesis testing schemes, negating the requirement of user-defined thresholds. * (d) The proposal of a straight-forward damage detection method in active-sensing guided-waves-based SHM, with the advantage of enhanced detection capability over conventional DI-based approaches, without sacrificing simplicity. The remainder of this paper is organized as follows: Section 2 introduces the development of the statistical framework (Subsection 2.1), the theory of the utilized stochastic NP-TS representation (Subsection 2.2), the statistics used in this study for damage detection (Subsections 2.3, 2.4 and 2.5), as well as briefly presents the literature-based DIs used for comparison in this study (Subsection 2.6). Then, the experimental setup, the results and the discussions are presented for every coupon consecutively in Section 3. Finally, Section 4 concludes this study and proposes extra steps for enhancement of damage detection within active-sensing guided-wave SHM systems. ## 2 The Statistical Damage Detection Framework ### 2.1 The General Framework The use of statistical methods for damage detection and identification has been previously reported for vibration-based SHM [53, 54, 49], and only recently for active-sensing guided-wave SHM [7]. A typical statistical framework for damage detection, localization and quantification is shown in Figure 1 [7, 52]. In this framework, $x[t]$ and $y[t]$ are the individual actuation and response signals, respectively, for every structural case, respectively, indexed with discrete time $t$ ($t=1,2,\ldots,N$), which can be converted to continuous time through the transformation $(t-1)T_{s}$, where $T_{s}$ is the sampling time for the recorded signals. The subscripts ($o,A,B,\ldots,$ and $u$) indicate the healthy, damage $A,B,\ldots,$ and unknown cases, respectively. In this context, the damage cases labeled as ($A,B,\ldots$) can resemble different types, sizes or locations of damage. For each structural case, all actuation ($X$) and response ($Y$) signals can be presented as $Z=(X,Y)$, with $Z_{o},Z_{A},Z_{B},...,$ and $Z_{u}$ indicating the different cases as before. Figure 1: Framework for statistical time series methods for structural health monitoring [7, 52, 49]. As shown in Figure 1, the statistical time series framework consists of two phases, namely the baseline and inspection phase. In the baseline phase, NP-TS models, each producing a characteristic quantity $\widehat{Q}$, are identified and properly validated for the healthy ($\widehat{Q}_{0}$) time series signal, as well as, if available, different predefined damage cases ($\widehat{Q}_{A}$, $\widehat{Q}_{B},\ldots$). Then, during the inspection phase, the same NP-TS models are identified for the unknown ($\widehat{Q}_{u}$) state of the system. Next, damage detection is achieved through applying appropriate binary statistical hypothesis tests to assess the statistical deviation of the unknown quantity $\widehat{Q}_{u}$ from $\widehat{Q}_{0}$ corresponding to the healthy signal (damage detection). Based on data availability, the statistical similarity to one of the damage characteristic quantities $\widehat{Q}_{A}$, $\widehat{Q}_{B},\ldots$ with the baseline quantity $\widehat{Q}_{0}$ can enable statistical damage identification/classification. In the present study, this framework is only used for damage detection. ### 2.2 Overview of Non-parametric Time Series Representations Stochastic NP-TS representations utilize time-domain Auto-/Cross-Covariance Functions (A/CCF) and/or frequency-domain Power-/Cross-Spectral Densities (P/CSD) in order to model a dynamic stationary signal [55, Chapter 2, pp. 39]. As discussed above, frequency-domain models are used in this study. In this context, several estimators have been developed for the PSD (also referred to as Auto Spectral Density) of a sensor excitation and/or response signal, including the periodogram, the Thompson, the Blackman-Tukey, and the Bartlett- Welch (or simply Welch) estimators [56, Chapter 5, pp. 235]. As such estimators are random variables that represent the true PSD of a system, their corresponding statistical properties, such as the mean and variance, allow for the extraction of estimation confidence intervals that can be subsequently used to represent statistical damage thresholds. In this study, the Welch PSD estimate, which is a modified periodogram estimator using a series of overlapping windows [57, Chapter 4, pp. 76] is used for damage detection. For a time series signal $x[t]$, the frequency-domain ($\omega$) Welch PSD ($\widehat{S}_{xx}(\omega)$) is based on the averaging of multiple-windowed periodograms using properly-selected sample windows $w[t]$ with 50% overlap, and is calculated as follows [58, Chapter 8, pp. 418] (the hat indicates an estimated variable): $\widehat{S}_{xx}(\omega)=\frac{1}{KLUT}\sum_{i=0}^{K-1}\Bigl{|}T\sum_{t=0}^{L-1}w[t]\cdot\widehat{x}[t+iD]^{(-j2\pi\omega tT)}\Bigr{|}^{2}$ (1) with $U=\frac{1}{L}\sum_{t=0}^{L-1}w^{2}[t],\quad\widehat{x}[t]=x[t]-\widehat{\mu}_{x},\quad N=L+D(K-1)$ (2) and $N$, $L$, $K$, $D$, and $T$ being the total number of signal samples, the size of each window, the number of utilized windows, the number of overlapping data points in each window, and the time period of the signal, respectively. $\widehat{\mu}_{x}$ represents the mean of the time series. The estimation statistics, that is the mean and variance, of the Welch PSD can be described as follows in case the Bartlett window is used [58, Chapter 8, pp. 419]: $E\\{\widehat{S}_{xx}(\omega)\\}=\frac{1}{2\pi LU}{S}_{xx}(\omega)|W(\omega)|^{2}$ (3) $Var\\{\widehat{S}_{xx}(\omega)\\}\approx\frac{9}{16}\frac{L}{N}{S}^{2}_{xx}(\omega)$ (4) where $W(\omega)$ designates the Fourier transform of the window function. One of the main reasons behind the wide use of the Welch PSD estimator is that it is asymptotically unbiased and consistent [56]. In this study, the Welch PSD estimate is used in developing appropriate statistical quantities, also referred to as test statistics, and corresponding statistical hypothesis tests for damage detection, as described in the following section. ### 2.3 The Single-set $F$ Statistic Method Based on the PSD-based NP-TS method and the corresponding statistical hypothesis testing setup presented in [49, 52], damage can be detected by assessing changes in the Welch PSD of properly-determined wave packets/modes from an acousto-ultrasound time series signal. Thus, the characteristic quantity in this study is $Q={S}_{xx}(\omega)={S}(\omega)$. The main idea is based on the comparison of the Welch PSD of the response of the structure in an unknown state, $S_{u}(\omega)$, to that of the structure in its healthy state, $S_{o}(\omega)$. Damage detection can thus be tackled using the following SHT problem [7, 49]: $\begin{array}[]{llll}H_{o}&:&S_{u}(\omega)=S_{o}(\omega)&\text{(null hypothesis -- healthy structure )}\\\ H_{1}&:&S_{u}(\omega)\neq S_{o}(\omega)&\text{(alternative hypothesis -- damaged structure)}\end{array}$ (5) Again, due to the finite nature of the experimental time series, the true PSD values are unknown, and thus corresponding estimated quantities are utilized instead ($\widehat{S}$). It can be shown that the Welch PSD estimate will have the following property [57, Chapter 3, pp. 46]: $2K\widehat{S}(\omega)/{S}(\omega)\sim\chi^{2}(2K)$ (6) In the above expression, the factor of 2 comes from the fact that every periodogram used in averaging the Welch PSD has a real and a complex component. Consequently, a damage detection statistic following the $\mathcal{F}$ distribution with $(2K,2K)$ degrees of freedom can be developed as follows: $F=\frac{\widehat{S}_{o}(\omega)/S_{o}(\omega)}{\widehat{S}_{u}(\omega)/S_{u}(\omega)}\,\sim\,\mathcal{F}(2K,2K)$ (7) In the case of a healthy structure (null hypothesis), $S_{u}(\omega)$ and $S_{o}(\omega)$ coincide, thus: $\text{Under}\;H_{o}:\quad F=\frac{\widehat{S}_{o}(\omega)}{\widehat{S}_{u}(\omega)}\,\sim\,\mathcal{F}(2K,2K)$ (8) Thus, the above SHT decision-making process can be modified as follows: $\begin{array}[]{ccl}f_{\frac{\alpha}{2}}(2K,2K)\leq F=\frac{\widehat{S}_{o}(\omega)}{\widehat{S}_{u}(\omega)}\leq f_{1-\frac{\alpha}{2}}(2K,2K)\quad(\forall\;\omega)&\Longrightarrow&H_{o}\;\text{is accepted (healthy structure)}\\\ \text{Else}&\Longrightarrow&H_{1}\;\text{is accepted (damaged structure)}\\\ \end{array}$ (9) where $\alpha$ is the Type I error (false alarm) probability, $f_{\frac{\alpha}{2}}$, $f_{1-\frac{\alpha}{2}}$ designate the $\mathcal{F}$ distribution’s $\frac{\alpha}{2}$ and $1-\frac{\alpha}{2}$ critical points, respectively ($f_{\alpha}$ is defined such that Prob$(F\leq f_{\alpha})=\alpha$). ### 2.4 The Multiple-set Modified $F_{m}$ Statistic Method In many realistic cases, a single baseline signal may not be representative of the healthy structure, and the average of many signal realizations might be more meaningful. In that case, multiple closely-spaced (time-wise) response signal realizations available for each state of the component being monitored under nominally-constant environmental/operational conditions can be used to entail some experimental statistics in the estimation of the SHM metric being used. Towards this end, the sample expectation, that is, $E\\{\widehat{S}_{o}(\omega)\\}=\frac{1}{M}\sum_{h=1}^{M}\widehat{S}_{o}(\omega)\;$ (10) can be used in order to “expand” the baseline/healthy estimates for the structure being monitored. In the above expression, $M$ is the number of healthy data sets used in the estimation of the metric being used. Then, following the aforementioned property of PSD estimates in equation (6), the following expression can be developed: $2KME\\{\widehat{S}_{o}(\omega)\\}/{S}_{o}(\omega)\sim\chi^{2}(2KM)$ (11) As such, a modified $F$ statistic can be developed by replacing the Welch PSD estimate with the mean of all PSD estimates of $M$ number of time-series signals taken for the system under the baseline/healthy state: $F_{m}=\frac{E\\{\widehat{S}_{o}(\omega)\\}/S_{o}(\omega)}{\widehat{S}_{u}(\omega)/S_{u}(\omega)}\;\sim\;\mathcal{F}(2KM,2K)$ (12) Under the null hypothesis in equation (5), the $S_{o}(\omega)$ and $S_{u}(\omega)$ coincide: $\text{Under}\;H_{o}:\quad F_{m}=\frac{E\\{\widehat{S}_{o}(\omega)\\}}{\widehat{S}_{u}(\omega)}\;\sim\;\mathcal{F}(2KM,2K)$ (13) thus, the modified decision making scheme with the appropriate confidence levels, can be expressed as follows: $\begin{array}[]{ccl}f_{\frac{\alpha}{2}}(2KM,2K)\leq F_{m}=\frac{E\\{\widehat{S}_{o}(\omega)\\}}{\widehat{S}_{u}(\omega)}\leq f_{1-\frac{\alpha}{2}}(2KM,2K)\quad(\forall\;\omega)&\Longrightarrow&H_{o}\;\text{is accepted (healthy)}\\\ \text{Else}&\Longrightarrow&H_{1}\;\text{is accepted (damaged)}\\\ \end{array}$ (14) with $f_{\frac{\alpha}{2}}$, $f_{1-\frac{\alpha}{2}}$ designating the $\mathcal{F}$ distribution’s $\frac{\alpha}{2}$ and $1-\frac{\alpha}{2}$ critical points, respectively ($f_{\alpha}$ is defined such that Prob$(F_{m}\leq f_{\alpha})=\alpha$). ### 2.5 The Multiple-set $Z$ Statistic Method With the availability of a sufficiently-large number of data sets, that is a large $M$ in equation (10), $E\\{\widehat{S}_{o}(\omega)\\}$ would approach the true PSD, and according to the Central Limit Theorem (CLT) [59, Chapter 3, pp. 62], $E\\{\widehat{S}_{o}(\omega)\\}$ would also follow a normal distribution with the true PSD being the mean and $\sigma_{0}$ the variance. Utilizing these statistical phenomena, and based on the $Z$ statistic developed by Fassois and coworkers [49, 52] for the Frequency Response Function (FRF) of vibration-based SHM signals, a novel $Z$ statistic is proposed herein utilizing the Welch PSD estimate for many baseline active- sensing acousto-ultrasound SHM signals. The following SHT problem is posed for damage detection in this case: $\begin{array}[]{llll}H_{o}&:&S_{o}(\omega)-S_{u}(\omega)=0&\text{(null hypothesis -- healthy structure)}\\\ H_{1}&:&S_{o}(\omega)-S_{u}(\omega)\neq 0&\text{(alternative hypothesis -- damaged structure)}\end{array}$ (15) where both terms in the hypothesis above are the true values of the respective PSDs. As mentioned above, under the assumption of a large $M$ in equation (10) (many baseline signals used for expectation estimation), the first term in the hypothesis test above ($S_{o}(\omega)$) can be replaced by the expectation, which would be normally distributed as aforementioned. Additionally, assuming the availability of many data points (that is, a large $N$) used for PSD estimation, the second term in the formulation of the hypothesis test ($S_{u}(\omega)$) can be replaced by an estimate [57, Chapter 3, pp. 45], which will also follow a normal distribution due to the asymptotic properties of $\chi^{2}$-distributed estimates, as also dictated by the central limit theorem (CLT) [59, Chapter 3, pp. 62]. Under the null hypothesis, both of these terms would follow the same distribution since both would be coming from a healthy structural case. Thus, under the null hypothesis: $\begin{array}[]{llll}\text{Under}\;H_{o}&:&E\\{\widehat{S}_{o}(\omega)\\}-\widehat{S}_{u}(\omega)\;\sim\;\mathcal{N}(0,2\sigma_{o}^{2}(\omega))&\mbox{(null hypothesis -- healthy structure)}\\\ \end{array}$ (16) where $\sigma_{o}^{2}(\omega)$ can be estimated from the baseline phase, and can be assumed to have negligible variability if a large number of signals is used in estimating the value of the PSDs [49, 52]. Thus, by defining an appropriate type I error, or false alarm, probability ($\alpha$), the Welch PSD-based $Z$ statistic can be expressed as follows: $\begin{array}[]{ccl}Z=\frac{\mid E\\{\widehat{S}_{o}(\omega)\\}-\widehat{S}_{u}(\omega)\mid}{\sqrt{2\widehat{\sigma}_{o}^{2}(\omega)}}\leq Z_{1-\frac{\alpha}{2}}\quad(\forall\;\omega)&\Longrightarrow&H_{o}\;\text{is accepted (healthy structure)}\\\ \text{Else}&\Longrightarrow&H_{1}\;\text{is accepted (damaged structure)}\\\ \end{array}$ (17) with $Z_{1-\frac{\alpha}{2}}$ designating the standard Normal distribution’s $1-\frac{\alpha}{2}$ critical point. Table 1 summarizes all three statistics used in this study for damage detection. Table 1: Summary of the different damage detection statistics utilized in this study. Quantity | $F$ Statistic | $F_{m}$ Statistic | $Z$ Statistic ---|---|---|--- Property | $2K\widehat{S}(\omega)/{S}(\omega)\sim\chi^{2}(2K)$ | $2KME\\{\widehat{S}(\omega)\\}/{S}_{o}(\omega)\sim\chi^{2}(2KM)$ | $E\\{\widehat{S}(\omega)\\}-S(\omega)\;\sim\;\mathcal{N}(0,2\sigma_{o}^{2}(\omega))$ Test Statistic | $F=\frac{\widehat{S}_{o}(\omega)}{\widehat{S}_{u}(\omega)}$ | $F_{m}=\frac{E\\{\widehat{S}_{o}(\omega)\\}}{\widehat{S}_{u}(\omega)}$ | $Z=\frac{\mid E\\{\widehat{S}_{o}(\omega)\\}-\widehat{S}_{u}(\omega)\mid}{\sqrt{2\sigma_{o}^{2}(\omega)}}$ | $K$: Number of non-overlapping segments Comment | $M$: Number of available baseline data sets | $\widehat{S}(\omega)$: Welch PSD estimate; $\omega\;\epsilon[0,2\pi/T_{s}]$: frequency in radians per second ($T_{s}$ is the sampling time). It is worth noting here that, although there lies a difference between the statistical assumptions behind the formulations of the $F_{m}$ and the $Z$ statistics, both are applied to the same data sets in this study, and their results are compared with respect to which statistic achieves better detection capabilities. ### 2.6 Reference State-of-the-Art Damage Indices In this work, two time-domain damage indices are utilized as reference in order to compare between the performance of DIs and the performance of the NP- TS models proposed herein. The first DI was adopted from the work of Janapati et al. [4], which is characterized by high sensitivity to damage size and orientation, and low sensitivity to other variations such as adhesive thickness and the material properties of the structure, sensors, and adhesive. Given a baseline $y_{0}[t]$ and an unknown $y_{u}[t]$ signal indexed with normalized discrete time $t$ ($t=1,2,3,\ldots,N$ where $N$ is the number of data samples considered in the calculation of the DIs, which depends on the studied coupon as will be shown in Section 3), the formulation of that DI is as follows: $Y_{u}^{n}[t]=\frac{y_{u}[t]}{\sqrt{\sum_{t=1}^{N}{y^{2}_{u}[t]}}},\quad Y_{0}^{n}[t]=\frac{\sum_{t=1}^{N}{(y_{0}[t]\cdot Y_{u}^{n}[t])}}{y_{0}[t]\cdot\sum_{t=1}^{N}{y_{0}^{2}[t]}},\quad DI=\sum_{t=1}^{N}{(Y^{n}_{u}[t]-Y^{n}_{0}[t])}$ (18) In this notation, $Y^{n}_{u}[t]$ and $Y^{n}_{0}[t]$ are normalized unknown (inspection) and baseline signals, respectively. The second DI used in this study is the time-domain DI presented by Qiu et al. [35] and used in training their Gaussian mixture models due to its sensitivity to changes in wave form and time of flight. The formulation of that DI is as follows: $DI=1-\sqrt{\frac{(\sum_{t=1}^{N}{y_{0}[t]\cdot y_{u}[t])^{2}}}{\sum_{t=1}^{N}{y_{0}^{2}[t]}\cdot\sum_{t=1}^{N}{y^{2}_{u}[t]}}}$ (19) ## 3 Results and Discussion In this work, the comparison between state-of-the-art DIs and the proposed NP- TS approaches in damage detection was carried out over three components with different damage cases: a notched Al plate, a Carbon Fiber-Reinforced Plastic (CFRP) coupon with weights taped on the surface to simulate a crack, and the open-source data sets available on the Open Guided-Waves project’s website [1]. ### 3.1 Test Case I: Damage Detection in an Aluminum Plate #### 3.1.1 Test Setup, Damage Types and Data Acquisition This first coupon was a 6061 Aluminum $152.4\times 254$ mm ($6\times 10$ in) coupon ($2.36$ mm/$0.093$ in thick) (McMaster Carr) with a 12-mm (0.5-in) diameter hole in the middle, as shown in Figure 2. Using Hysol EA 9394 adhesive, the coupon was fitted with six single-PZT (Lead Zirconate Titanate) SMART Layers type PZT-5A (Acellent Technologies, Inc) as shown in Figure 2. The PZT sensors are $0.2$ mm ($0.00787$ in) in thickness and $3.175$ mm ($1/8$ in) in diameter. To simulate damage, using an end-mill and a $0.8128$-mm ($0.032$-in) hand saw, a notch was generated extending from the middle hole of the coupon with length varying between $2$ and $20$ mm, in $2$-mm increments. Actuation signals in the form of 5-peak tone bursts (5-cycle Hamming-filtered sine wave) having an amplitude of 90 V peak-to-peak and various center frequencies were generated in a pitch-catch configuration over each sensor consecutively. With a sampling rate of 24 MHz, data was collected using a ScanGenie III data acquisition system (Acellent Technologies, Inc). Preliminary analysis was conducted, and a center frequency of 250 kHz was chosen for the complete analysis presented in this study based upon the best separation between the first two wave packets in various signal paths. All data sets were exported to MATLAB for analysis.222Matlab version R2018a; function pwelch.m (window size: 100 for single wave packet analyses and 500 for the full signal/two-wave packet analyses; NFFT: 2000; Overlap: 50%). Table 2 summarizes the relevant experimental details for this coupon. Figure 2: The Al coupon used in this study shown here with a 20-mm notch (the largest damage size of this test case). The arrows indicate the paths used in the analysis presented herein. Table 2: Summary of experimental details for the Al coupon. Structural State | Number of Data Sets ---|--- Healthy | 20† 2-mm notch | 20 4-mm notch | 20 6-mm notch | 20 8-mm notch | 20 10-mm notch | 20 12-mm notch | 20 14-mm notch | 20 16-mm notch | 20 18-mm notch | 20 20-mm notch | 20 Sampling Frequency: $f_{s}=24$ MHz. Center frequency range: [$50:50:750$] kHz Number of samples per data set $N=8000$. †M=20 in equation (10). #### 3.1.2 Damage Detection Results In order to assess the performance of the proposed approach, a simple isotropic Al coupon was initially used. Figure 3 panels a and b, respectively, show one indicative full response signal and its corresponding first-arrival wave packet off of sensor 6 when sensor 2 was actuated (refer to Figure 2 for sensor numbering) under different notch sizes. Because this is a damage- intersecting path, a gradual decrease in signal amplitude, with a slight delay, can be observed with increasing notch size. This is expected since the notch scatters the wave, decreasing the amount of energy going through to sensor 6 as scattering increases [8]. Figure 3c shows the evolution of the two chosen state-of-the-art DIs for the first-arrival wave packet. As shown, although the DIs closely follow damage for notch sizes more than 8 mm, it might be difficult to detect damages up to 8 mm in size, given the proximity of the DI values for the healthy case and the damaged cases. Without prior experience with these types of materials/components, assigning a threshold between a healthy component and a damaged one might be challenging in that range of damages. As the length of the analyzed signal increases, the DIs become more sensitive to small damages as shown in Figure 3d. However, it can be observed that the DIs do not follow the increase in notch size uniformly even for a damage-intersecting path like path 2-6. Exploring a damage-non- intersecting path (Figure 4), the DIs fail to follow damage evolution to a greater extent, with fluctuations being observed as notch size increases. Such fluctuation in the DIs can be mistaken for a change in conditions surrounding the component, which would make the task of damage detection and threshold identification even more challenging. abcd Figure 3: Indicative signal from the Al coupon for signal path 2-6 (damage-intersecting) under different notch sizes: (a) full signal; (b) first- arrival wave packet; (c) single wave packet DIs – the dashed lines designate the upper and lower $95\%$ confidence bounds for the Janapati et al. (blue) and Qiu et al. (red) DIs; (d) two wave packet DIs. abcd Figure 4: Indicative signal from the Al coupon for signal path 6-3 (damage-non-intersecting) under different notch sizes: (a) full signal; (b) a single wave packet; (c) single- wave packet DIs – the dashed lines designate the upper and lower $95\%$ confidence bounds for the Janapati et al. (blue) and Qiu et al. (red) DIs; (d) two-wave packet DIs. Figure 5 presents indicative results of applying the proposed framework to the response signal from path 2-6 in the Al coupon (see Table A.1 in the Appendix for summary results). Figure 5a shows the evolution of the Welch PSD of the signals as notch size increases, with the red and the black dashed lines indicating the theoretical (estimation uncertainty) and the experimental 95% confidence intervals of the healthy PSD, respectively. The first thing to be observed in this figure is that, using the theoretical estimation confidence intervals, notch sizes more than 2 mm can be detected with 95% confidence, and all damage sizes are detected when the experimental 95% confidence levels are considered. Although the latter result is expected due to the nominally- controlled lab environment significantly inhibiting change in the Welch PSD over multiple healthy signals, the former observation shows the enhanced detection capability of the frequency-domain PSD compared to time-domain DIs for damages of this type in Al. Furthermore, in contrast to the DIs in Figure 3, the PSDs evolve uniformly with damage, which hints on the enhanced damage quantification capability of these techniques. Thus, the Welch PSD emerges as a better metric when it comes to damage detection and quantification for the case at hand. Applying the SHT frameworks developed in Section 2.1 (Figure 5 panels b-d), one can assess the difference between both approaches in a statistical way. As shown in Figure 5b, the $F$ statistic is capable of only detecting the last three damage cases (10-18 mm) with 95% confidence. Although this performance is somewhat similar to that of the DIs, an advantage in the proposed approach is the extraction of confidence intervals directly from the SHM metric being used, without the need for user experience for defining damage thresholds. The $F_{m}$ statistic (Figure 5c) does a slightly better job by detecting the 8-mm damage with 95% confidence, which is attributed to the inclusion of some experimental statistics into the definition of this metric. Examining the $Z$ statistic (Figure 5d), one can observe that all damage cases are detected with $95\%$ confidence. Furthermore, the effect of damage on the $Z$ statistic is again uniform, indicating the superior performance of this statistic in damage quantification compared to the conventional time-domain DI approach. Figure 6a shows indicative Welch PSD estimates for the first two wave packets (using a window size equal to the width of a single wave packet). As shown, although the peak amplitude of the PSD at the actuation frequency decreases, the detection performance remains the same as for a single wave packet. Exploring the three statistics proposed in this study (Figure 6 panels b-d), it can also be observed that the detection performance stays the same, with all damages being detected by the $Z$ statistic with 95% confidence. Table 3: The different parameters used in estimating the Welch PSD for the Al coupon data sets. Segment Length | $100$ ---|--- Window Type | Hamming Frequency Resolution | $\Delta f=12$ kHz Sampling Frequency | $24$ MHz Single Wave Packet Data Length | $N=500$ samples ($\sim 20$ $\mu s$) No of non-overlapping segments | $9$ Full Signal Length Data Length | $N=8000$ samples ($\sim 330$ $\mu s$) No of non-overlapping segments | $159$ abcd Figure 5: Indicative results from applying the proposed NP-TS approach to the first arrival wave packet from path 2-6 in the Al coupon under different damage sizes: (a) Welch PSD – the red and the black dashed lines indicate the theoretical (estimation uncertainty) and the experimental 95% confidence bounds of the healthy PSD, respectively; (b) $F$ statistic; (c) $F_{m}$ statistic; (d) $Z$ statistic. abcd Figure 6: Indicative results from applying the proposed NP-TS approach to the full signal from path 2-6 in the Al coupon under different damage sizes: (a) Welch PSD – the red and the black dashed lines indicate the theoretical (estimation uncertainty) and the experimental 95% confidence bounds of the healthy PSD, respectively; (b) $F$ statistic; (c) $F_{m}$ statistic; (d) $Z$ statistic. Moving on to the damage-non-intersecting path (path 6-3), Figure 7 shows indicative results for a single wave packet (see Table A.2 in the Appendix for summary results). As shown in Figure 7a, using the PSD’s theoretical estimation confidence intervals (red dashed lines), all damages are deemed healthy with 95% confidence. This is attributed to the wide nature of the estimation uncertainty when the PSD of a deterministic signal is being estimated, as is the case in this study. However, just like the DIs, all damages are detected with 95% confidence when the experimental uncertainty is being considered. Being based on the theoretical confidence intervals, both the $F$ and $F_{m}$ statistics also show all damages as healthy with 95% confidence (with the exception of the 14 mm case for the $F_{m}$ statistic), as shown in Figure 7 panels b and c. On the other hand, the $Z$ statistic (Figure 7d) detects all damage cases with 95% confidence because its formulation is based on the experimental uncertainty. The same trend can be observed when the first two wave packets are considered, as shown in Figure 8. A number of conclusions can be drawn from these observations when it comes to comparing the proposed statistics to the DIs. Firstly, for a notched Al coupon, the Welch PSD, the $F$ and $F_{m}$ statistics can be used as a preliminary step in differentiating between damage-intersecting and non- intersecting paths, in contrast to the DIs, which do not show a clear distinction. Secondly, the sensitivity of the $Z$ statistic seems to be the same as the DIs because of both being based on the experimental confidence intervals, an advantage that the $Z$ statistic has over the DI is the extraction of the confidence bounds based on the assumption of a normal distribution of the expectation of the signals’ PSDs. Thus, the extracted damage detection thresholds emerge from the formulation of the SHM metric itself, and don’t require prior experience with such materials and damages, or physics-based modelling. In contrast, the DIs require complex approaches in order to set accurate thresholds, and do not entail any theoretical distribution on the signals, from which thresholds can emerge naturally. abcd Figure 7: Indicative results from applying the proposed NP-TS approach to the first-arrival wave packet from path 6-3 in the Al coupon under different damage sizes: (a) Welch PSD – the red and the black dashed lines indicate the theoretical (estimation uncertainty) and the experimental 95% confidence bounds of the healthy PSD, respectively; (b) $F$ statistic; (c) $F_{m}$ statistic; (d) $Z$ statistic. abcd Figure 8: Indicative results from applying the proposed NP-TS approach to the full signal from path 6-3 in the Al coupon under different damage sizes: (a) Welch PSD – the red and the black dashed lines indicate the theoretical (estimation uncertainty) and the experimental 95% confidence bounds of the healthy PSD, respectively; (b) $F$ statistic; (c) $F_{m}$ statistic; (d) $Z$ statistic. ### 3.2 Test Case II: Damage Detection in CFRP Plate #### 3.2.1 Test Setup, Damage Types and Data Acquisition The second coupon used in this study was a CFRP coupon (ACP Composites,) having the same dimensions as the Al coupon, with multiple $0/90$ unidirectional CF plies. This coupon was also fitted with 6 single-PZT SMART Layers type PZT-5A (Acellent Technologies, Inc) as shown in Figure 9. Damage was simulated by attaching 1-6 three-gm weights to the surface of the coupon next to each other using tacky tape. The same actuation and data acquisition properties were used for this coupon as with the Al one. Also, similar to the case of the Al coupon, the actuation center frequency of 250 kHz was chosen for the analysis presented herein. Tables 4 and 5 summarize the experimental details for this coupon. Figure 9: The CFRP coupon used in this study shown here with 6 weights as simulated damage (the largest damage size of this test case). The arrows indicate the paths used in the analysis presented herein. Table 4: Details of the first experimental data set for the CFRP coupon. Structural State | Number of Data Sets | Total Added Weight† (g) ---|---|--- Healthy | 20†† | 0 1 Steel weight | 1 | $3$ 2 Steel weights | 1 | $6$ 3 Steel weights | 1 | $9$ 4 Steel weights | 1 | $12$ 5 Steel weights | 1 | $15$ 6 Steel weights | 1 | $18$ Sampling Frequency: $f_{s}=24$ MHz. Center frequency range: [$50:50:750$] kHz. Number of samples per data set $N=8000$. †Weight of tacky tape not considered here. ††M=20 in equation (10). Table 5: Details of the second experimental data set for the CFRP coupon. Structural State | Number of Data Sets | Total Added Weight† (g) ---|---|--- Healthy | 20†† | 0 1 Steel weight | 20 | $3$ 2 Steel weights | 20 | $6$ 3 Steel weights | 20 | $9$ 4 Steel weights | 20 | $12$ 5 Steel weights | 20 | $15$ 6 Steel weights | 20 | $18$ Sampling Frequency: $f_{s}=24$ MHz. Center frequency range: [$50:50:750$] kHz. Number of samples per data set $N=8000$. †Weight of tacky tape not considered here. ††M=20 in equation (10). #### 3.2.2 Damage Detection Results Figure 10 panels a and b present, respectively, the signals and the first discernible wave packet, obtained at sensor 4 when sensor 3 was actuated under different damage sizes, where damage size here indicates the number of taped weights. Figure 10 panels c and d show the DIs for a single, and double wave packet lengths, respectively. As shown, because a CFRP coupon exhibits more non-linearity compared to an Al coupon, the single-wave packet DIs completely fail to follow damage evolution and can further only detect the last damage case (6 weights) within the 95% experimental healthy confidence intervals for both DI formulations as shown in Figure 10c. This performance is slightly enhanced when analyzing two wave packet lengths (Figure 10d). Figure 11 shows the same 4 plots for a damage non-intersecting path (path 1-4). As shown panel c, the performance of the DIs substantially deteriorates, with a decrease in the value of both DIs with increasing simulated damage size up to 4 weights. A similar trend is observed in Figure 11d when considering two wave packet lengths for the analysis. Furthermore, although some damage cases fall outside the 95% confidence bounds (4 weights for single-wave packet DIs, and 5 and 6 weights for two-wave packet DIs), the general trend is a reduction in the values of the DIs, which can again be easily mistaken with changing environmental or operational conditions over an otherwise healthy component. Thus, in terms of damage detection, the DIs offer poor performance for the CFRP coupon with the simulated damage used in this study. abcd Figure 10: Indicative signal from path 3-4 in the CFRP coupon under different simulated damage sizes (number of attached weights): (a) full signal; (b) single wave packet; (c) single-wave packet DIs – the dashed lines designate the upper and lower $95\%$ confidence bounds for the Janapati et al. (blue) and Qiu et al. (red) DIs; (d) DIs for double the wave packet length. abcd Figure 11: Indicative signal from path 1-4 in the CFRP coupon under different simulated damage sizes (number of attached weights): (a) full signal; (b) single wave packet; (c) single wave packet DIs – the dashed lines designate the healthy upper and lower $95\%$ confidence bounds for the Janapati et al. (blue) and Qiu et al. (red) DIs; (d) DIs for double the wave packet length. Figure 12a shows the estimated Welch PSD for a single wave packet from path 3-4 under different damage cases. Note that, as mentioned, due to the nature of the actuation signal, the theoretical confidence intervals are too wide to detect any of the simulated damages, and thus they are not shown here. Figure 12b shows the $Z$ statistic for that path, from which it can be concluded that the cases of 3-6 weights are all damage cases with 95% confidence. Thus, the $Z$ statistic surpasses the DIs in detection performance for this damage- intersecting path. Examining a longer signal length for the analysis, it can be seen that both the Welch PSD estimate and the $Z$ statistic (Figure 12 panels c and d, respectively) show more sensitivity to damage, with the former detecting damage sizes as small as 3 weights, and the latter detecting ones as small as 2 weights. Moving onto the damage non-intersecting path (1-4), it can be seen that the Welch PSD estimates (Figure 13 panels a and c) fail to detect almost any of the damages with 95% confidence levels. The same can be said for the $Z$ statistics, as shown in Figure 13 panels b and d) , with the exception of detecting the 2- and 4-weight cases for the single-wave packet $Z$ statistics. This reduction in sensitivity for damage-non-intersecting paths can be attributed to the effect of damage on the signal, as well as the relatively wide variability in the baseline signal amplitude, which in turn leads to widening the 95% healthy confidence bounds. abcd Figure 12: Indicative results from applying the proposed NP-TS approach to the signals from path 3-4 in the CFRP coupon under different simulated damage sizes (number of attached weights): (a) Welch PSD for single wave packet – the black dashed lines indicate the experimental 95% confidence bounds of the healthy PSD; (b) $Z$ statistic for single wave packet; (c) Welch PSD for double the wave packet lengths; (d) $Z$ statistic for double the wave packet lengths. abcd Figure 13: Indicative results from applying the proposed NP-TS approach to the signals from path 1-4 in the CFRP coupon under different simulated damage sizes (number of attached weights): (a) Welch PSD for single wave packet – the black dashed lines indicate the experimental 95% confidence bounds of the healthy PSD; (b) $Z$ statistic for single wave packet; (c) Welch PSD for full signal; (d) $Z$ statistic for full signal. For this reason, another experiment was taken out where the baseline data acquisition process was more restrictive (lab was empty). Table 5 presents the details of this second experiment. Figure 14 shows the the DI plots for the damage-intersecting (panels a and b) and the damage-non-intersecting (panels c and d) paths of this new data set. For reference, these plots respectively correspond to the ones in Figure 10 panels c and d, and Figure 11 panels c and d. As shown, although both this and the original data sets were all acquired off of the same coupon with the same temperature setting, even in a lab environment, baseline variability between different data sets can be significant. As shown in all panels of Figure 14, the spread in the values of the healthy DIs is smaller in this new data set, which allows for good detection performance for the DIs. Exploring the $Z$ statistics (Figure 15), one can observe the enhanced detection performance here too, given the narrower experimental confidence bounds in the new data set. Although both the DIs and the $Z$ statistics almost consistently follow damage size evolution for the damage-intersecting path (path 3-4), the $Z$ statistic shows better detection capability for the damage-non-intersecting path,detecting all damages with 95% confidence. abcd Figure 14: Damage Index results for the second acquired CFRP coupon data set shown in Table 5: (a) single-wave packet DIs for path 3-4; (b) double-wave packet DIs for path 3-4; (c) single-wave packet DIs for path 1-4; (d) double-wave packet DIs for path 1-4. In all plots, the dashed red lines are the healthy $95\%$ confidence bounds for the Janapati et al. (blue) and Qiu et al. (red) DIs. abcd Figure 15: Indicative $Z$ statistic results for the second acquired CFRP coupon data set shown in Table 5: (a) single-wave packet $Z$ statistics for path 3-4; (b) double-wave packet DIs for path 3-4; (c) single-wave packet $Z$ statistics for path 1-4; (d) double-wave packet $Z$ statistics for path 1-4. In all plots, the dashed red lines are the healthy $95\%$ confidence bounds. Table 6: Parameters used in estimating the Welch PSD for the CFRP coupon data sets. Segment Length | $100$ ---|--- Window Type | Hamming Frequency Resolution | $\Delta f=12$ kHz Sampling Frequency | $24$ MHz Single Wave Packet Data Length | $N=500$ samples ($\sim 20$ $\mu s$) No of non-overlapping segments | $9$ Full Signal Length Data Length | $N=8000$ samples ($\sim 330$ $\mu s$) No of non-overlapping segments | $159$ In order to assess the performance of all 4 metrics (the DI, $F$, $F_{m}$, and $Z$ statistics) at different alpha (false alarm levels) i.e. at different confidence intervals, the corresponding Receiver Operating Characteristics (ROC) curves were explored for different signal lengths (also, see Tables B.4 and B.3 in the Appendix for summary results). Figure 16a shows the ROC for the 4 metrics at alpha levels ranging from 1E-6 to 1, as applied to a single wave packet off of the damage-intersecting path 3-4. As shown, because this is a damage-intersecting path, 3 out of the 4 metrics exhibit perfect detection performance with an area under the ROC curve equal to 1. Also, although the performance of the $F_{m}$ statistic doesn’t seem to be better than the worst statistical estimator (the dashed line), the $F$ statistic shows optimal performance as the alpha levels change, in contrast to its weak detection performance at an alpha level of 0.05, as mentioned in the discussion of Figure 12. Moving onto two wave packets of the same path (Figure 16b, one can observe that the $Z$ statistic outperforms all other metrics in damage detection. In addition, the $F$ statistic outperforms the DI metric, which hints on the advantages of using frequency-domain approaches and statistical hypothesis tests instead of time-domain approaches. For the damage-non- intersecting path 1-4, it can be observed that the $Z$ statistic outperforms the DI for both: a single- (Figure 16c) and two- (Figure 16d) wave packet lengths. Thus, it can be concluded that the $Z$ statistic emerges as the best damage detection statistic in this study for the CFRP coupon investigated herein. abcd Figure 16: Receiver Operating Characteristics (ROC) plots comparing the different damage detection methods for the new data set of the CFRP coupon under the effect of the first simulated damage (1 weight): (a) path 3-4 wave packet; (b) path 3-4 full signal; (c) path 1-4 wave packet; (d) path 1-4 full signal. In all subplots, 15 out of 20 healthy signals were used for calculating the mean in estimating the $F_{m}$ and the $Z$ statistics. ### 3.3 Test Case III: Damage Detection in the Open Guided-Waves CFRP Panel #### 3.3.1 Test Setup, Damage Types and Data Acquisition The third test case used in this study was the CFRP panel utilized in the Open Guided-Waves project [1], which had a quasi-isotropic construction with layup $[45/0/-45/90/-45/0/-45/90]_{S}$. The panel had the dimensions of $500\times 500$ mm ($19.69\times 19.69$ in), and a thickness of 2 mm ($0.079$ in). During the fabrication process of the panel, 12 PZT sensors, 5 mm ($0.2$ in) in diameter and $0.2$ mm ($0.0079$ in) in thickness, were co-bonded on the panel. To simulate damage, a 10-mm diameter, $2.35$-mm-thick ($0.0925$ in) Al disk ($0.5$ g) was consecutively attached using tacky tape on 28 different locations on the panel grouped into seven groups. Figure 17a shows a schematic of the CFRP panel, where the sensor and damage locations are shown. Also, the inset in Figure 17a shows the simulated damage on one of the locations. Each sensor was consecutively actuated using a 5-peak tone burst signal (5-cycle Hanning-filtered sine wave) having an amplitude of $\pm 100$ V. Response signals were sampled over the remaining sensors at a sampling rate of $10$ MHz. Three sets of 20 baseline (healthy) signal realizations per sensor were recorded. After acquiring the first baseline set (the first 20 healthy signals), a single signal per sensor was recorded for each damage location, with the weight on locations D$1-11$. After that, the second baseline set was acquired (healthy signals $21-40$), followed by recording a single damage signal per sensor per weight location for locations D$12-20$. Finally, signals were recorded with the weight at locations D$21-28$ after the third baseline set was acquired (healthy signals $41-60$). This resulted in $60$ baseline realizations and $28$ damage signals per sensor (one signal for each damage location for each sensor.) Other data sets were also recorded that are not used in this study. Table 7 summarizes the experimental details of this panel, and the readers are directed to the original study [1] for more information on test setup. For ease of comparison of the proposed damage detection methods, only the response signals for the actuation with $260$ kHz center frequency were chosen for analysis in this study. In the present study, signals from simulated damages in the same damage group (see Figure 17a and Table 7) were treated as different realizations of single damage in the vicinity of that group on the CFRP panel. In addition, for all the detection metrics in this study, each damage group was analyzed against its corresponding baseline data set only (the healthy data set immediately preceding that damage group in the data acquisition process) for accuracy. Actuator-sensor path 3-12 was used to demonstrate the performance of the different detection techniques proposed herein. As such, damage groups $2$ and $3$ (see Table 7) were considered as different realizations of signals for two damages intersected by the path (with the first baseline set used for comparison). On the other hand, damage groups $7$ and $8$ were treated as different realizations of signals for two damages not intersected by the signal path (with the third baseline set used for comparison). Table 7: Summary of experimental details for the CFRP panel [1]. Structural State† | Number of Data Sets | Data Set Label ---|---|--- Healthy (weight unattached) | 20†† | First Baseline Set Weight on D1-4 | 4 | Damage Group 1 Weight on D5-8 | 4 | Damage Group 2 Weight on D9-11 | 3 | Damage Group 3 Healthy (weight unattached) | 20†† | Second Baseline Set Weight on D12 | 1 | Damage Group 4 Weight on D13-16 | 4 | Damage Group 5 Weight on D17-20 | 4 | Damage Group 6 Healthy (weight unattached) | 20†† | Third Baseline Set Weight on D21-24 | 4 | Damage Group 7 Weight on D25-28 | 4 | Damage Group 8 Sampling Frequency: $f_{s}=10$ MHz. Center frequency range: [$40:20:260$] kHz. Number of samples per data set $N=13106$. †Weight was attached to one location (e.g. D4) at a time within each damage group. ††M=20 in equation (10) #### 3.3.2 Damage Detection Results Figure 17 panels b and c show the complete signal and the first-arrival wave packet, respectively, for signal path 3-12 on the CFRP panel when the Al weight was on locations 5-11 (damage-intersecting case). It is worth noting that, examining other paths (not shown here), the packet shown in panel c is actually a combination of two wave packets merged together, as can also be observed from the number of cycles in the shown packet. Even though this limits the analysis to only this single wave structure, this path was still chosen because it directly intersects (or does not intersect) almost complete damage location groups, which allows for the analysis of detection performance for both types of paths. Figure 17d shows the values of the DI formulated by Janapati et al. [4] for the first 20 baseline signals and the signals corresponding to damage locations 5-11. As shown, within the $95$% healthy confidence bounds (dashed blue lines), only damages at locations 5-7 are detected, while the rest of the damage cases are considered healthy with $95$% confidence. Noting that the healthy signals in each baseline set were taken under controlled temperatures, it can be again concluded that the DI lacks robustness to uncertainties even in controlled environments, where the values of the DI still fluctuate even for the healthy case, producing wide healthy bounds that affect detection performance. Figure 18 shows the same set of figures for the third baseline set (healthy signals 41-60), and the signals corresponding to damage locations 21-28 (damage-non-intersecting case.) As shown, the DI fails to detect any of the damage cases with $95$% confidence. abcd Figure 17: (a) A schematic of the CFRP panel used in the Open Guided Waves open source data project [1] with all the simulated damage locations. The inset shows a snapshot of part of the actual panel with damage location markings and the Al weight used to simulate damage on one of the locations. The arrow indicates the path used in the analysis presented herein; (b) indicative signals from path 3-12 for the healthy case, as well as when the Al weight (simulated damage) was on locations D5-10 (damage-intersecting case); (c) the first arrival wave packets; (d) the Janapati et al. DI for the first arrival wave packets – the dashed blue lines are the upper and lower $95\%$ confidence bounds for the Janapati et al. DI as applied to the DI values of corresponding 20-baseline signal data set. abc Figure 18: (a) Indicative signals from path 3-12 for the healthy case, as well as when the Al weight (simulated damage) was on locations D21-28 (damage-non-intersecting case); (b) the first arrival wave packets; (c) the Janapati et al. DI for the first arrival wave packets - the dashed blue lines are the upper and lower $95\%$ confidence bounds for the Janapati et al. DI as applied to the DI values of corresponding 20-baseline signal data set. Examining the Welch PSD estimates for the damage-intersecting case, Figure 19a shows that all damage cases are detected with $95$% confidence. This immediately shows the advantage of this frequency-domain metric when it comes to damage detection. Figure 19b shows the $Z$ statistic for that case, where again all damage cases are detected with $95$% confidence. Also, it can be observed that there are no false alarms in this case, whereas there was at least one false alarm event with the DI (see Figure 17c). Figure 19c presents the Welch PSD estimate for the damage-non-intersecting set of signals. As shown, at least 5 out of the 8 damage cases were detected with $95$% accuracy. Examining the $Z$ statistic, it can be observed that only one damage case is detected with $95$% confidence, while the rest are deemed healthy. It is worth noting that neither of the other two statistics (the $F$ and $F_{m}$ statistics) detected any of the damage cases with the set confidence bounds for the damage-intersecting and non-intersecting cases. Again, this called upon the exploration of the effect of different confidence intervals (manifested in different alpha false alarm levels in the statistical hypothesis tests) in order to conclusively assess the performance of the different detectors proposed herein. Table 8: The Welch PSD estimation parameters for the OGW coupon data sets. Segment Length | $40$ ---|--- Window Type | Hamming Frequency Resolution | $\Delta f=5$ kHz Sampling Frequency | $10$ MHz Single Wave Packet Data Length | $N=360$ samples ($\sim 40$ $\mu s$) No of non-overlapping segments | $9$ abcd Figure 19: The results of applying the proposed NP-TS approach to the signals from path 3-12 in the Open Guided-Waves CFRP panel under different simulated damage locations: (a) Welch PSD for D5-10 (damage-intersecting case) - the black dashed lines indicate the theoretical (estimation uncertainty) and the experimental 95% confidence bounds of the healthy PSD, respectively; (b) $Z$ Statistic for D5-10 (damage-intersecting case); (c) Welch PSD for D21-28 (damage-non-intersecting case); (d) $Z$ Statistic for D21-28 (damage-non- intersecting case) Figure 20 panels a and b show the ROC plots for the damage-intersecting case and the damage-non-intersecting case, respectively. In constructing each plot, detection statistics from all corresponding damage locations were used, and only corresponding baseline groups were considered in each case. As shown in the damage-intersecting case (panel a), the $Z$, $F$, and $F_{m}$ statistics all outperform the DI in overall detection performance, with larger areas under the curves. For the damage-non-intersecting case (panel b), although the $F$, $F_{m}$ statistics and the DI show similar performance, the $Z$ statistic surpasses all of them, even for low alpha levels (wider confidence bounds). Tables C.5 and C.6 in the Appendix also present summary detection results. All of these results show the superiority of the $Z$ statistic when it comes to damage detection. ab Figure 20: ROC plots comparing the different damage detection methods for the Open Guided Waves project’s CFRP panel (path 3-12): (a) the Al weight (simulated damage) on D5-11 (damage-intersecting case); (d) the Al weight (simulated damage) on D21-28 (damage-non-intersecting case). In both subplots, 15 out of 20 healthy signals were used for calculating the mean in estimating the $F_{m}$ and the $Z$ statistics. ## 4 Conclusions In this study, three frequency-domain damage detection metrics based on stochastic non-parametric time series representations were developed and compared with state-of-the-art damage indices as applied to three different test cases: a notched Al plate, a CFRP coupon with stacked weights, and the CFRP panel with different weight locations used in the Open Guided-Waves project [1]. It was shown that, although the DIs can accurately detect damage and follow damage evolution in the isotropic Al coupon case, it fails to do either in the CFRP coupon for $95$% healthy confidence bounds. In addition, it also shows poor detection performance for the different damage cases of the CFRP panel at the same confidence levels. Examining the $F$ and $F_{m}$ statistics, because their detection thresholds are either solely dependent on the theoretical estimation confidence bounds of the Welch PSD estimator ($F$ statistic), or dependent on the theoretical estimation intervals with the incorporation of some experimental statistics ($F_{m}$ statistics), their detection performance at $95$% confidence can, in some cases, be even worse than the DIs. However, for different confidence levels, both, especially the $F$ statistic, exhibit a detection performance more or less similar to that of the DIs, as shown in the different Receiver Operating Characteristics plots in this study. On the other hand, the $Z$ statistic outperforms all other detectors used in this study for all three test cases, for both: damage- intersecting and non-intersecting paths. In addition, it also better follows the evolution of damage for the Al and CFRP coupons in the damage-intersecting case compared to the DIs, which hints on its damage quantification capabilities. Overall, it can be concluded from this study that, for the three test cases studied herein, methods based on frequency-domain non-parametric statistical time series models show greater sensitivity to damage, even when used to analyze damage-non-intersecting signals, compared to time-domain DI-based approaches, especially in materials exhibiting non-linearities and anisotropic behaviour such as composites. This was clearly demonstrated when constructing $95$% healthy confidence bounds accounting for the same experimental uncertainties in both approaches. In addition, the proposed approaches show increased robustness to uncertainty with less fluctuation in the values of the metrics for the healthy test cases compared to the time-domain-based DIs. Thus, non-parametric time series representations emerge as sources of constructing accurate and robust metrics that promise enhancement in damage detection performance for SHM systems. ## Acknowledgment This work is carried out at the Rensselaer Polytechnic Institute under the Vertical Lift Research Center of Excellence (VLRCOE) Program, grant number W911W61120012, with Dr. Mahendra Bhagwat and Dr. William Lewis as Technical Monitors. ## References * [1] Moll, J., Kathol, J., Fritzen, C.-P., Moix-Bonet, M., Rennoch, M., Koerdt, M., Herrmann, A. S., Sause, M. G., and Bach, M., “Open Guided Waves: online platform for ultrasonic guided wave measurements,” Structural Health Monitoring, Vol. 18, 2019, pp. 1–12. * [2] Qiu, L., Liu, M., Qing, X., and Yuan, S., “A quantitative multidamage monitoring method for large-scale complex composite,” Structural Health Monitoring, Vol. 12, No. 3, 2013, pp. 183–196. * [3] Romano, F., Ciminello, M., Sorrentino, A., and Mercurio, U., “Application of structural health monitoring techniques to composite wing panels,” Journal of Composite Materials, Vol. 53, No. 25, 2019, pp. 3515–3533. * [4] Janapati, V., Kopsaftopoulos, F., Li, F., Lee, S., and Chang, F.-K., “Damage detection sensitivity characterization of acousto-ultrasound-based structural health monitoring techniques,” Structural Health Monitoring, Vol. 15, No. 2, 2016, pp. 143–161. * [5] Das, S. and Saha, P., “A review of some advanced sensors used for health diagnosis of civil engineering structures,” Measurement, Vol. 129, 2018, pp. 68–90. * [6] Farrar, C. R. and Worden, K., “An introduction to Structural Health Monitoring,” The Royal Society – Philosophical Transactions: Mathematical, Physical and Engineering Sciences, Vol. 365, 2007, pp. 303–315. * [7] Amer, A. and Kopsaftopoulos, F. P., “Probabilistic active sensing acousto-ultrasound SHM based on non-parametric stochastic representations,” Proceedings of the Vertical Flight Society 75th Annual Forum & Technology Display, Philadelphia, PA, USA, May 2019. * [8] Amer, A. and Kopsaftopoulos, F. P., “Probabilistic Damage Quantification via the Integration of Non-parametric Time-series and Gaussian Process Regression Models,” Proceedings of the 12th International Workshop on Structural Health Monitoring (IWSHM 2019), Palo Alto, CA, USA, September 2019. * [9] Kopsaftopoulos, F., Nardari, R., Li, Y.-H., and Chang, F.-K., “A stochastic global identification framework for aerospace structures operating under varying flight states,” Mechanical Systems and Signal Processing, Vol. 98, 2018, pp. 425–447. * [10] Spiridonakos, M. and Fassois, S., “Non-stationary random vibration modelling and analysis via functional series time-dependent ARMA (FS-TARMA) models–A critical survey,” Mechanical Systems and Signal Processing, Vol. 47, No. 1-2, 2014, pp. 175–224. * [11] Zhang, Q. W., “Statistical damage identification for bridges using ambient vibration data,” Computers and Structures, Vol. 85, 2007, pp. 476–485. * [12] Ahmed, S. and Kopsaftopoulos, F. P., “Uncertainty quantification of guided waves propagation for active sensing structural health monitoring,” Proceedings of the Vertical Flight Society 75th Annual Forum & Technology Display, Philadelphia, PA, USA, May 2019. * [13] Ahmed, S. and Kopsaftopoulos, F. P., “Investigation of broadband high-frequency stochastic actuation for active-sensing SHM under varying temperature,” Proceedings of the 12th International Workshop on Structural Health Monitoring (IWSHM 2019), Palo Alto, CA, USA, September 2019\. * [14] Zhao, J., Gao, H. D., Chang, G. F., Ayhan, B., Yan, F., Kwan, C., and Rose, J. L., “Active health monitoring of an aircraft wing with embedded piezoelectric sensor/actuator network: I. Defect detection, localization and growth monitoring,” Smart Materials and Structures, Vol. 16, No. 4, 2007, pp. 1208–1217. * [15] Flynn, E. B., Todd, M. D., Wilcox, P. D., Drinkwater, B. W., Croxford, A. J., and Kessler, S., “Maximum-likelihood estimation of damage location in guided-wave structural health monitoring,” Proceedings of The Royal Society A, Burlington, VT, Vol. 467, No. 2133, 2011, pp. 2575–2596. * [16] Todd, M. D., Flynn, E. B., Wilcox, P. D., Drinkwater, B. W., Croxford, A. J., and Kessler, S., “Ultrasonic wave-based defect localization using probabilistic modeling,” American Institute of Physics Conference Proceedings, Burlington, VT, May 2012. * [17] Haynes, C., Todd, M., Flynn, E., and Croxford, A., “Statistically-based damage detection in geometrically-complex structures using ultrasonic interrogation,” Structural Health Monitoring, Vol. 12, No. 2, 2012, pp. 141–152. * [18] Ng, C.-T., “On the selection of advanced signal processing techniques for guided wave damage identification using a statistical approach,” Engineering Structures, Vol. 67, 2014, pp. 50–60. * [19] Mujica, L. E., Ruiz, M., Pozo, F., Rodellar, J., and Güemes, A., “A structural damage detection indicator based on principal component analysis and statistical hypothesis testing,” Smart Materials and Structures, Vol. 23, No. 2, dec 2013, pp. 025014. * [20] Peng, T., Saxena, A., Goebel, K., Xiang, Y., Sankarararman, S., and Liu, Y., “A novel Bayesian imaging method for probabilistic delamination detection of composite materials,” Smart Materials and Structures, Vol. 22, 2013, pp. 125019–125028. * [21] Yang, J., He, J., Guan, X., Wang, D., Chen, H., Zhang, W., and Liu, Y., “A probabilistic crack size quantification method using in-situ Lamb wave test and Bayesian updating,” Mechanical Systems and Signal Processing, Vol. 78, 2016, pp. 118–133. * [22] He, J., Ran, Y., Liu, B., Yang, J., and Guan, X., “A Lamb wave based fatigue crack length estimation method using finite element simulations,” The 9th International Symposium on NDT in Aerospace, Xiamen, China, November 2017. * [23] MIL-HDBK-1823A, “MIL-HDBK-1823A,” Nondestructive Evaluation System Reliability Assessment, Department of Defense, April 2009. * [24] Gallina, A., Packo, P., and Ambrozinski, L., “Model Assisted Probability of Detection in Structural Health Monitoring,” Advanced Structural Damage Detection: From Theory to Engineering Applications, edited by T. Stepinski, T. Uhl, and W. Staszewski, John Wiley and Sons, Ltd., 2013, pp. 382–407. * [25] Jarmer, G. and Kessler, S. S., “Application of Model Assisted Probability of Detection (MAPOD) to a Guided Wave SHM System,” Structural Health Monitoring 2017: Real-Time Material State Awareness and Data-Driven Safety Assurance– Proceedings of the 12th International Workshop on Structural Health Monitoring (IWSHM 2017), edited by F.-K. Chang and F. Kopsaftopoulos, Stanford University, USA, 2017. * [26] Moriot, J., Quagebeur, N., Duff, A. L., and Masson, P., “A model-based approach for statistical assessment of detection and localization performance of guided wave–based imaging techniques,” Structural Health Monitoring, Vol. 17, No. 6, 2017, pp. 1460–1472. * [27] Giurgiutiu, V., “Flutter prediction for flight/wind-tunnel flutter test under atmospheric turbulence excitation,” Journal of Intelligent Materials Systems and Structures, Vol. 16, No. 4, 2005, pp. 291–305. * [28] Ihn, J. and Chang, F.-K., “Detection and monitoring of hidden fatigue crack growth using a built-in piezoelectric sensor/actuator network, Part I: Diagnostics,” Smart Materials and Structures, Vol. 13, 2004, pp. 609–620. * [29] Ihn, J. and Chang, F.-K., “Detection and monitoring of hidden fatigue crack growth using a built-in piezoelectric sensor/actuator network, Part II: Validation through riveted joints and repair patches,” Smart Materials and Structures, Vol. 13, 2004, pp. 621–630. * [30] Ihn, J. and Chang, F.-K., “Pitch-catch active sensing methods in structural health monitoring for aircraft structures,” Structural Health Monitoring, Vol. 7, No. 1, 2008, pp. 5–19. * [31] Giurgiutiu, V., “Piezoelectric Wafer Active Sensors for Structural Health Monitoring of Composite Structures Using Tuned Guided Waves,” Journal of Engineering Materials and Technology, Vol. 133, No. 4, 2011, pp. 041012. * [32] Jin, H., Yan, J., Li, W., and Qing, X., “Monitoring of fatigue crack propagation by damage index of ultrasonic guided waves calculated by various acoustic features,” Applied Sciences, Vol. 9, 2019, pp. 4254. * [33] Xu, B., Zhang, T., Song, G., and Gu, H., “Active interface debonding detection of a concrete-filled steel tube with piezoelectric technologies using wavelet packet analysis,” Mechanical Systems and Signal Processing, Vol. 36, 2013, pp. 7–17. * [34] Nasrollahi, A., Deng, W., Ma, Z., and Rizzo, P., “Multimodal structural health monitoring based on active and passive sensing,” Structural Health Monitoring, Vol. 17, No. 2, 2018, pp. 395–409. * [35] Qiu, L., Yuan, S., Bao, Q., Mei, H., and Ren, Y., “Crack propagation monitoring in a full-scale aircraft fatigue test based on guided wave Gaussian mixture model,” Smart Materials and Structures, Vol. 25, 2016, pp. 055048. * [36] Wang, F., Huo, L., and Song, G., “A piezoelectric active sensing method for quantitative monitoring of bolt loosening using energy dissipation caused by tangential damping based on the fractal contact theory,” Smart Materials and Structures, Vol. 27, 2018, pp. 015023. * [37] Castro, E., Moreno-Garcia, P., and Gallego, A., “Damage Detection in CFRP Plates Using Spectral Entropy,” Shock and Vibration, 2014, pp. 1–8. * [38] An, Y.-K., Giurgiutiu, V., and Sohn, H., “Integrated impedance and guided-wave-based damage detection,” Mechanical Systems and Signal Processing, Vol. 28, 2012, pp. 50–62. * [39] Su, Z., Zhou, C., Hong, M., Cheng, L., Wang, Q., and Qing, X., “Acousto-ultrasonics-based fatigue damage characterization: linear versus nonlinear signal features,” Mechanical Systems and Signal Processing, Vol. 45, 2014, pp. 225–239. * [40] Su, Z. and Ye, L., “Lamb wave-based quantitative identification of delamination in CF/EP composite structures using artificial neural algorithm,” Composite Structures, Vol. 66, 2004, pp. 627–637. * [41] Song, G., Gu, H., and Mo, Y.-L., “Smart aggregates: multi-functional sensors for concrete structure —a tutorial and a review,” Smart Materials and Structures, Vol. 17, 2008, pp. 033001. * [42] Tibaduiza, D. A., Mujica, L. E., Rodellar, J., and Güemes, A., “Structural damage detection using principal component analysis and damage indices,” Journal of Intelligent Material Systems and Structures, Vol. 27, No. 2, 2016, pp. 233–248. * [43] Lize, E., Rebillat, M., Mechbal, N., and Bolzmacher, C., “Optimal dual-PZT and network design for baseline-free SHM of complex anisotropic composite structures,” Smart Materials and Structures, Vol. 27, 2018, pp. 115018. * [44] Hua, J., Cao, X., Yi, Y., and Lin, J., “Time-frequency damage index of broadband lamb wave for corrosion inspection,” Journal of Sound and Vibration, Vol. 464, 2020, pp. 114985. * [45] Ibanez, F., Baltazar, A., and Mijarez, R., “Detection of damage in multiwire cables based on wavelet entropy evolution,” Smart Materials and Structures, Vol. 24, 2015, pp. 085036. * [46] Shannon, C., “A mathematical theory of communication,” Bell System Technology Journal, Vol. 27, 1948, pp. 379–423. * [47] Rojas, E., Baltazar, A., and Loh, K. J., “Damage detection using the signal entropy of an ultrasonic sensor network,” Smart Materials and Structures, Vol. 24, 2015, pp. 075008. * [48] Qiu, J., Li, F., Abbas, S., and Zhu, Y., “A baseline-free damage detection approach based on distance compensation of guided waves,” Journal of Low Frequency, Vibration and Active Control, Vol. 38, 2019, pp. 1132–1148. * [49] Kopsaftopoulos, F. P. and Fassois, S. D., “Vibration based health monitoring for a lightweight truss structure: experimental assessment of several statistical time series methods,” Mechanical Systems and Signal Processing, Vol. 24, 2010, pp. 1977–1997. * [50] Kopsaftopoulos, F. P. and Fassois, S. D., “A Functional Model Based Statistical Time Series Method for Vibration Based Damage Detection, Localization, and Magnitude Estimation,” Mechanical Systems and Signal Processing, Vol. 39, 2013, pp. 143–161. * [51] Kopsaftopoulos, F. P. and Fassois, S. D., “Identification of Stochastic Systems Under Multiple Operating Conditions: The Vector-dependent Functionally Pooled (VFP) Parametrization,” under preparation for publication, 2016. * [52] Fassois, S. D. and Kopsaftopoulos, F. P., “Statistical Time Series Methods for Vibration Based Structural Health Monitoring,” New Trends in Structural Health Monitoring, edited by W. Ostachowicz and A. Guemes, chap. 4, Springer, 2013, pp. 209–264. * [53] Kopsaftopoulos, F. P. and Fassois, S. D., “Experimental assessment of time series methods for structural health monitoring (SHM),” Proceedings of the 4th European Workshop on Structural Health Monitoring (EWSHM), Cracow, Poland, 2008. * [54] Kopsaftopoulos, F. P. and Fassois, S. D., “Vibration Based Health Monitoring for a Thin Aluminum Plate – A Comparative Assessment of Statistical Time Series Methods,” Proceedings of the 5th European Workshop on Structural Health Monitoring (EWSHM), Sorrento, Italy, 2010. * [55] Box, G. E. P., Jenkins, G. M., and Reinsel, G. C., Time Series Analysis: Forecasting & Control, Prentice Hall: Englewood Cliffs, NJ, 3rd ed., 1994\. * [56] Manolakis, D., Ingle, V. K., and Kogon, S. M., Statistical and Adaptive Signal Processing: Spectral Estimation, Signal Modeling, Adaptive Filtering and Array Processing, Artech House, New York, 2005. * [57] Kay, S. M., Modern Spectral Estimation: Theory and Application, Prentice Hall: New Jersey, 1988. * [58] Hayes, M. H., Statistical Digital Signal Processing and Modelling, John Wiley and Sons, New York, 1996. * [59] Bendat, J. S. and Piersol, A. G., Random Data: Analysis and Measurement Procedures, Wiley-Interscience: New York, 3rd ed., 2000. ## Appendix A Damage Detection Summary Results: Notched Al Coupon Table A.1: Damage detection summary results at an $\alpha$ value of $95\%$ for path 2-6 (single wave packet) in the Al plate (damage presented in units of mm). Method | False | Missed Damage ($\%$) ---|---|--- | Alarms ($\%$) | 2 | 4 | 6 | 8 | 10 | 12 | 14 | 16 | 18 | 20 DI† [4] | 6 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 $F$ Statistic† | 0 | 100 | 100 | 100 | 100 | 0 | 0 | 0 | 0 | 0 | 0 $F_{m}$ Statistic†† | 0 | 100 | 100 | 100 | 0 | 0 | 0 | 0 | 0 | 0 | 0 $Z$ Statistic†† | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 False alarms presented as percentage of 400 test cases for the DI and $F$ statistic, | and as percentage of 20 data sets for the $F_{m}$ and $Z$ statistics. Missed damages presented as percentage of 20 test cases. † All 20 baseline data sets were used as reference signals consecutively. †† 15 out of 20 baseline data sets were used to calculate the baseline mean. Table A.2: Damage detection summary results at an $\alpha$ value of $95\%$ for path 6-3 (single wave packet) in the Al plate (damage presented in units of mm). Method | False | Missed Damage ($\%$) ---|---|--- | Alarms ($\%$) | 2 | 4 | 6 | 8 | 10 | 12 | 14 | 16 | 18 | 20 DI† [4] | 5.5 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 $F$ Statistic† | 0 | 100 | 100 | 100 | 100 | 100 | 100 | 100 | 100 | 100 | 100 $F_{m}$ Statistic†† | 0 | 100 | 100 | 100 | 100 | 100 | 100 | 0 | 0 | 100 | 100 $Z$ Statistic†† | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 False alarms presented as percentage of 400 test cases for the DI and $F$ statistic, | and as percentage of 20 data sets for the $F_{m}$ and $Z$ statistics. Missed damage presented as percentage of 20 test cases. † All 20 baseline data sets were used as reference signals consecutively. †† 15 out of 20 baseline data sets were used to calculate the baseline mean. abcd Figure A.1: Receiver Operating Characteristics (ROC) plots comparing the different damage detection methods for the notched Al coupon with a notch size of 2 mm: (a) path 2-6 wave packet; (b) path 2-6 full signal; (c) path 6-3 wave packet; (d) path 6-3 full signal. ## Appendix B Damage Detection Summary Results: CFRP Coupon Table B.3: Damage detection summary results at multiple $\alpha$ values for path 1-4 (single wave packet) in the CFRP plate. Method | False | Missed damage ($\%$) ---|---|--- | alarms ($\%$) | 1 Weight | 2 Weights | 3 Weights | 4 Weights | 5 Weights | 6 Weights DIa† [4] | 5.25 | 99.75 | 84.5 | 94.5 | 99.25 | 93.75 | 53.5 $F$ Statisticb† | 25 | 65 | 5 | 5 | 0 | 0 | 0 $F_{m}$ Statisticc†† | 25 | 40 | 10 | 60 | 0 | 0 | 0 $Z$ Statistica†† | 0 | 25 | 10 | 30 | 10 | 0 | 0 False alarms presented as percentage of 20 test cases. Missed damages presented as percentage of 20 test cases. a $\alpha=95\%$.; b $\alpha=1\%$; c $\alpha=10\%$ † All 20 baseline data sets were used as reference signals consecutively. †† 15 out of 20 baseline data sets were used to calculate the baseline mean. Table B.4: Damage detection summary results at multiple $\alpha$ values for path 3-4 (single wave packet) in the CFRP plate. Method | False | Missed damage ($\%$) ---|---|--- | alarms ($\%$) | 1 Weight | 2 Weights | 3 Weights | 4 Weights | 5 Weights | 6 Weights DIa† [4] | 5.25 | 13.25 | 0 | 0 | 0 | 0 | 0 $F$ Statisticb† | 95 | 0 | 5 | 0 | 0 | 0 | 0 $F_{m}$ Statisticc†† | 30 | 60 | 0 | 0 | 0 | 0 | 0 $Z$ Statistica†† | 0 | 0 | 0 | 0 | 0 | 0 | 0 False alarms presented as percentage of 20 test cases. Missed damages presented as percentage of 20 test cases. a $\alpha=95\%$.; b $\alpha=1\%$; c $\alpha=10\%$ † All 20 baseline data sets were used as reference signals consecutively. †† 15 out of 20 baseline data sets were used to calculate the baseline mean. ## Appendix C Damage Detection Summary Results: OGW CFRP Panel Table C.5: Damage detection summary results at multiple $\alpha$ values for path 3-12 (damage-intersecting case) in the CFRP panel [1]. Method | False | Missed damage ($\%$) ---|---|--- | alarms ($\%$) | D5/6/7/8 | D9/10/11 DIa† [4] | 7.5 | 51.25 | 100 $F$ Statisticb† | 0 | 75 | 100 $F_{m}$ Statisticb†† | 0 | 75 | 33 $Z$ Statistica†† | 0 | 0 | 0 False alarms presented as percentage of 20 test cases. | Missed damages presented as percentage of all test cases per damage group. | a $\alpha=95\%$.; b $\alpha=80\%$ | † All 20 baseline data sets were used as reference signals consecutively. | †† 15 out of 20 baseline data sets were used to calculate the baseline mean. | Table C.6: Damage detection summary results at an $\alpha$ value of $95\%$ for path 3-12 (damage-non-intersecting case) in the CFRP panel [1]. Method | False | Missed damage ($\%$) ---|---|--- | alarms ($\%$) | D21/22/23/24 | D25/26/27/28 DIa† [4] | 7.5 | 100 | 93.75 $F$ Statisticb† | 0 | 100 | 75 $F_{m}$ Statisticb†† | 0 | 100 | 75 $Z$ Statistica†† | 0 | 50 | 75 False alarms presented as percentage of 20 test cases. | Missed damages presented as percentage of all test cases per damage group. | a $\alpha=95\%$.; b $\alpha=70\%$ | † All 20 baseline data sets were used as reference signals consecutively. | †† 15 out of 20 baseline data sets were used to calculate the baseline mean. |
# Impurity induced quantum chaos for an ultracold bosonic ensemble in a double-well Jie Chen<EMAIL_ADDRESS>Zentrum für Optische Quantentechnologien, Fachbereich Physik, Universität Hamburg, Luruper Chaussee 149, 22761 Hamburg, Germany Kevin Keiler Zentrum für Optische Quantentechnologien, Fachbereich Physik, Universität Hamburg, Luruper Chaussee 149, 22761 Hamburg, Germany Gao Xianlong Department of Physics, Zhejiang Normal University, Jinhua 321004, China Peter Schmelcher Zentrum für Optische Quantentechnologien, Fachbereich Physik, Universität Hamburg, Luruper Chaussee 149, 22761 Hamburg, Germany The Hamburg Centre for Ultrafast Imaging, Universität Hamburg, Luruper Chaussee 149, 22761 Hamburg, Germany ###### Abstract We demonstrate that an ultracold many-body bosonic ensemble confined in a one- dimensional (1D) double-well (DW) potential can exhibit chaotic dynamics due to the presence of a single impurity. The non-equilibrium dynamics is triggered by a quench of the impurity-Bose interaction and is illustrated via the evolution of the population imbalance for the bosons between the two wells. While the increase of the post-quench interaction strength always facilitates the irregular motion for the bosonic population imbalance, it becomes regular again when the impurity is initially populated in the highly excited states. Such an integrability to chaos (ITC) transition is fully captured by the transient dynamics of the corresponding linear entanglement entropy, whose infinite-time averaged value additionally characterizes the edge of the chaos and implies the existence of an effective Bose-Bose attraction induced by the impurity. In order to elucidate the physical origin for the observed ITC transition, we perform a detailed spectral analysis for the mixture with respect to both the energy spectrum as well as the eigenstates. Specifically, two distinguished spectral behaviors upon a variation of the interspecies interaction strength are observed. While the avoided level-crossings take place in the low-energy spectrum, the energy levels in the high-energy spectrum possess a band-like structure and are equidistant within each band. This leads to a significant delocalization of the low-lying eigenvectors which, in turn, accounts for the chaotic nature of the bosonic dynamics. By contrast, those highly excited states bear a high resemblance to the non-interacting integrable basis, which explains for the recovery of the integrability for the bosonic species. Finally, we discuss the induced Bose-Bose attraction as well as its impact on the bosonic dynamics. ## I Introduction Trapping of an ultracold many-body bosonic ensemble in a one-dimensional (1D) double-well (DW) potential constitutes a prototype system for the investigations of the correlated quantum dynamics DW_exp_1 ; DW_exp_2 ; DW_exp_3 . Such a system represents a bosonic Josephson junction (BJJ), an atomic analogy of the Josephson effect initially predicted for Cooper pair tunneling through two weakly linked superconductors BJJ_1 ; BJJ_2 . Owing to the unprecedented controllability of the trapping geometries as well as the atomic interaction strengths cold_atom_rev , studies of the BJJ unveil various intriguing phenomena which are not accessible for conventional superconducting systems BJJ_Rabi_1 ; BJJ_Rabi_2 ; BJJ_Rabi_3 ; BJJ_Frag_1 ; BJJ_Frag_2 ; BJJ_Squeeze_1 ; BJJ_Squeeze_2 . Examples are the Josephson oscillations BJJ_Rabi_1 ; BJJ_Rabi_2 ; BJJ_Rabi_3 , fragmentations BJJ_Frag_1 ; BJJ_Frag_2 , macroscopic quantum self trapping DW_exp_3 ; BJJ_Rabi_1 ; BJJ_Rabi_2 , collapse and revival sequences BJJ_Rabi_3 as well as the atomic squeezing state BJJ_Squeeze_1 ; BJJ_Squeeze_2 . Under the explicit time-dependent driving forces, the BJJ can alternatively turn into the quantum kicked top (QKT), a famous platform for the investigations of quantum chaos as well as the classical-quantum correspondence QKT_1 ; QKT_2 ; QKT_3 ; QKT_4 ; QKT_5 ; QKT_6 ; QKT_7 ; QKT_8 ; QKT_9 ; QKT_10 ; QKT_11 ; QKT_12 ; QKT_13 . To date, related studies include the spectral statistics QKT_2 , the entanglement entropy production QKT_3 ; QKT_4 ; QKT_5 ; QKT_6 ; QKT_7 ; QKT_8 ; QKT_9 ; QKT_10 , the quantum decoherence and quantum correlations QKT_11 ; QKT_12 as well as the border between regular and chaotic dynamics QKT_13 . Moreover, by viewing the QKT as a collective $N$-qubit system, the effects of the quantum chaos on the digital quantum simulations have also been detailed discussed recently QKT_14 ; QKT_15 . On the other hand, stimulated by the experimental progresses on few-body ensembles few_exp1 ; few_exp2 ; few_exp3 ; few_exp4 ; few_exp5 ; few_exp6 , significant theoretical effort also focuses on the 1D few-body atomic systems few_gs_1 ; few_gs_2 ; few_gs_3 ; few_gs_4 ; few_gs_5 ; few_gs_6 ; few_gs_7 ; few_quench_1 ; few_quench_2 ; few_quench_3 ; few_bf_SC1 ; few_bf_SC2 , revealing for example the ground state few_gs_1 ; few_gs_2 ; few_gs_3 ; few_gs_4 ; few_gs_5 ; few_gs_6 ; few_gs_7 ; few_bf_SC1 ; few_bf_SC2 as well as the dynamical properties few_quench_1 ; few_quench_2 ; few_quench_3 , which pave the way for the studies of the binary mixtures with large particle number imbalance. Such hybridized systems are deeply related to the polaron physics polaron_1 ; polaron_2 ; polaron_3 as well as the open quantum systems OQS_1 and are particularly interesting owing to the fact that one subsystem is in the deep quantum regime while the other one can more or less be described by the semi-classical physics. Note, however, that while most of the discussions focus on impacts on the minority species from the majority bath, studies which alternatively explore the feedback to the majority species due to the presence of the minority one are still rare. In the present paper, we investigate a binary ultracold atomic mixture made of a single impurity and a non-interacting many-body bosonic ensemble that are confined within a 1D DW potential. Unlike most of the previous studies where the focuses are put on the weak-interacting regime, rendering the impurity being restricted into the lowest two modes of the DW potential Impurity_BH_1 ; Impurity_BH_2 ; Impurity_BH_3 ; Impurity_BH_4 ; Impurity_BH_5 ; Impurity_BH_6 , our discussions are not restricted to such a scenario. Specifically, we study the onset of the chaos for the majority bosonic species due to the presence of the impurity and put particular emphasis on the its dynamical response upon a sudden quench of the impurity-Bose interaction strength. As an exemplary observable, we monitor the quantum evolution of the population imbalance for the bosons between the two wells starting from a balanced particle population. While the increase of the post-quench interaction strength always facilitates a chaotic motion for the bosonic population imbalance, it becomes regular again when the impurity initially is prepared in the highly excited states. In order to characterize such an integrability to chaos (ITC) transition, we employ the linear entanglement entropy as a signature of quantum chaos, which alternatively measures the decoherence for the bosonic species. Depending on the degree of chaos, the transient dynamics of the corresponding linear entanglement entropy can behave as either a rapid growth or a slow variation with increasing time, whereas, its infinite-time averaged value, in addition, captures the edge of quantum chaos, i.e., the border between the integrable and the chaotic regions in the corresponding classical phase space. Furthermore, by computing the infinite-time averaged values of the linear entanglement entropy for various initial conditions, we find a striking resemblance between its profile and a classical phase space with attractive Bose-Bose interaction, which implies the existence of an attractive interaction among the bosons induced by the impurity. In order to elucidate the physical origin for the above observed ITC transition, we perform a detailed spectral analysis with respect to both the energy spectrum as well as the eigenstates of the mixture. Two distinguished spectral behaviors upon a variation of the interspecies interaction strength are observed. While the avoided level-crossings take place in the low-energy spectrum, the energy levels in the high-energy spectrum possess a band-like structure and are equidistant within each band. Consequently, this results in a significant delocalization for those low-lying eigenstates which, in turn, accounts for the chaotic nature of the bosonic non-equilibrium dynamics. Remarkably, those highly excited states bear a striking resemblance to the non-interacting integrable basis, which explains the recovery of the integrability for the bosonic species. Finally, we also discuss the induced Bose-Bose attraction and its impact on the bosonic dynamics. This paper is organized as follows. In Sec. II, we introduce our setup including the Hamiltonian, the initial conditions as well as the quantities of interests. In Sec. III, we present our main observation: the ITC transition for the bosonic species. In Sec. IV, we perform a detailed spectral analysis for the mixture with respect to both the energy spectrum as well as the eigenstates, so as to elucidate the physical origin for the above observed ITC transition. Finally, our conclusions and outlook are provided in Sec. V. ## II Setup ### II.1 Hamiltonian and angular-momentum representation The Hamiltonian of our 1D ultracold impurity-Bose mixture is given by $\hat{H}=\hat{H}_{I}+\hat{H}_{B}+\hat{H}_{IB}$, where $\displaystyle\hat{H}_{\sigma}$ $\displaystyle=\int dx_{\sigma}~{}\hat{\psi}^{\dagger}_{\sigma}(x_{\sigma})\textit{h}_{\sigma}(x_{\sigma})\hat{\psi}_{\sigma}(x_{\sigma}),$ $\displaystyle\hat{H}_{IB}$ $\displaystyle={g_{IB}}\int dx~{}\hat{\psi}^{\dagger}_{I}(x)\hat{\psi}^{\dagger}_{B}(x)\hat{\psi}_{B}(x)\hat{\psi}_{I}(x),$ (1) and $\textit{h}_{\sigma}(x_{\sigma})=-\frac{\hbar^{2}}{2m_{\sigma}}\frac{\partial^{2}}{\partial x_{\sigma}^{2}}+V_{DW}(x_{\sigma})$ is the single-particle Hamiltonian for the $\sigma=I(B)$ species being confined within a 1D symmetric DW potential $V_{DW}(x_{\sigma})=a_{\sigma}(x_{\sigma}^{2}-b_{\sigma}^{2})^{2}$. For simplicity, we consider the atoms for both species are of the same mass ($m_{I}=m_{B}=m$) and are trapped by the same potential geometry, i.e., $a_{I}=a_{B}=a_{DW}$ and $b_{I}=b_{B}=b_{DW}$. $\hat{\psi}_{\sigma}^{\dagger}(x_{\sigma})$ [$\hat{\psi}_{\sigma}(x_{\sigma})$] is the field operator that creates (annihilates) a $\sigma$-species particle at position $x_{\sigma}$. Moreover, we neglect the interactions among the bosons and assume the impurity-Bose interaction is of zero range and can be modeled by a contact potential of strength Feshbach_1 ; Feshbach_2 ; Feshbach_3 ; few_quench_3 $g_{IB}=\frac{2\hbar^{2}a_{3D}}{\mu a_{\bot}^{2}}[1-C\frac{a_{3D}}{a_{\bot}}]^{-1}.$ (2) Here $a_{3D}$ is the 3D impurity-Bose $s$-wave scattering length and $C\approx 1.4603$ is a constant. The parameter $a_{\bot}=\sqrt{\hbar/\mu\omega_{\bot}}$ describes the transverse confinement with $\mu=m/2$ being the reduced mass and we assume the transverse trapping frequency $\omega_{\bot}$ to be equal for both species. In the following discussions, we rescale the Hamiltonian of the mixture $\hat{H}$ for the units of the energy, length and time as $\eta=\hbar\omega_{\bot}$, $\xi=\sqrt{\hbar/m\omega_{\bot}}$ and $\tau=1/\omega_{\bot}$, respectively. We focus on the repulsive interaction regime, i.e., $g_{IB}\geqslant 0$ and set $a_{DW}=0.5$, $b_{DW}=1.5$, such that the lowest two single-particle energy levels are well separated from the others [see Fig. 1 (a), the spatial geometry of $V_{DW}(x)$ (black dashed line) as well as the lowest six single-particle energy levels (grey solid lines)]. Throughout this work, we explore a binary mixture made of a single impurity and 100 bosons ($N_{I}=1$, $N_{B}=100$), and focus on the dynamical response for the majority bosonic species upon a sudden quench of the impurity-Bose interaction strength (see below). Let us note that such a 1D mixture is experimentally accessible by imposing strong transverse and weak longitudinal confinement for a binary e.g., Bose-Fermi mixture with two different kinds of atoms mixture_exp_bf_1 ; mixture_exp_bf_2 or a Bose-Bose mixture made of the same atoms with two different hyperfine states mixture_exp_bb_1 ; mixture_exp_bb_2 . The DW potential can also be readily constructed by imposing a 1D optical lattice on top of a harmonic trap DW_exp_3 ; BJJ_2 . Moreover, the contact interaction strength $g_{IB}$ can be controlled experimentally by tuning the $s$-wave scattering lengths via Feshbach or confinement-induced resonances Feshbach_1 ; Feshbach_2 ; Feshbach_3 . Noticing further that the bosonic species is confined within a tight DW potential with $\delta_{1}\gg\delta_{0}$ [c.f. Fig. 1 (a)], here $\delta_{i}$ denotes the energy difference between the $i$-th and the $(i+1)$-th single- particle eigenstates. We adopt the two-mode approximation $\hat{\psi}_{B}(x)=u_{L}(x)\hat{b}_{L}+u_{R}(x)\hat{b}_{R},$ (3) with $u_{L,R}(x)$ being the Wannier-like states localized in the left and right well, respectively. This leads to the low-energy effective Hamiltonian for the bosonic species $\hat{H}_{B}=-J_{0}(\hat{b}^{\dagger}_{L}\hat{b}_{R}+\hat{b}^{\dagger}_{R}\hat{b}_{L}),$ (4) corresponding to the two-site Bose-Hubbard (BH) model with $J_{0}=0.071$ being the hopping amplitude. Before proceeding, it is instructive to express the above BH Hamiltonian in the angular-momentum representation. To see this, we introduce three angular- momentum operators BJJ_Rabi_3 ; BJJ_4 $\displaystyle\hat{J}_{x}$ $\displaystyle=\frac{1}{2}(\hat{b}_{L}^{\dagger}\hat{b}_{R}+\hat{b}_{R}^{\dagger}\hat{b}_{L}),~{}~{}~{}\hat{J}_{y}=-\frac{i}{2}(\hat{b}_{L}^{\dagger}\hat{b}_{R}-\hat{b}_{R}^{\dagger}\hat{b}_{L}),$ $\displaystyle\hat{J}_{z}$ $\displaystyle=\frac{1}{2}(\hat{b}_{L}^{\dagger}\hat{b}_{L}-\hat{b}_{R}^{\dagger}\hat{b}_{R}),$ (5) obeying the SU(2) commutation relation $[\hat{J}_{\alpha},\hat{J}_{\beta}]=i\epsilon_{\alpha\beta\gamma}\hat{J}_{\gamma}$. The BH Hamiltonian in Eq. (4) thus can be rewritten as $\hat{H}_{B}=-2J_{0}\hat{J}_{x},$ (6) which describes the angular momentum precession of a single particle whose spatial degrees of freedom (DOFs) are frozen. According to definitions for $\hat{J}_{x}$ and $\hat{J}_{z}$ in Eq. (5), we note that the kinetic energy in the BH model as well as the population imbalance for the bosons between the two wells are in analogy to the magnetizations of this single particle along the $x$ and the $z$ axes. Moreover, the particle number conservation in the Hamiltonian (4) corresponds to the angular momentum conservation $\hat{J}^{2}=\hat{J}_{x}^{2}+\hat{J}_{y}^{2}+\hat{J}_{z}^{2}=\frac{N_{B}}{2}(\frac{N_{B}}{2}+1)$ (7) for the Hamiltonian (6). For the case $g_{IB}=0$, the angular momentum dynamics can be simply integrated out from the corresponding Heisenberg equations of motion, in which $\displaystyle\hat{J}_{y}(t)$ $\displaystyle=\hat{J}_{z}(0)\text{cos}(2J_{0}t)-\hat{J}_{y}(0)\text{sin}(2J_{0}t),$ $\displaystyle\hat{J}_{z}(t)$ $\displaystyle=\hat{J}_{y}(0)\text{cos}(2J_{0}t)+\hat{J}_{z}(0)\text{sin}(2J_{0}t),$ (8) being the harmonic oscillations with the frequency $\omega_{0}=2J_{0}$ and $\hat{J}_{x}(t)=\hat{J}_{x}(0)$ is time-independent since $[\hat{J}_{x},\hat{H}_{B}]=0$. Further introducing the normalized vector $\hat{\vec{S}}(t)=\hat{S}_{x}(t)\vec{i}+\hat{S}_{y}(t)\vec{j}+\hat{S}_{z}(t)\vec{k}$ with $\hat{S}_{\gamma}(t)=\hat{J}_{\gamma}(t)/J$ for $\gamma=x,y,z$ and $J=N_{B}/2$, together with the fact that $\sum_{\gamma=x,y,z}\langle\hat{S}_{\gamma}\rangle^{2}(t)=\sum_{\gamma=x,y,z}\langle\hat{S}_{\gamma}\rangle^{2}(0),$ (9) one can readily show that the motion of the vector $\hat{\vec{S}}(t)$ always lies on the Bloch sphere with unit radius if, in addition, we choose the initial state as the atomic coherent state (ACS) (see below). ### II.2 Classical dynamics The above angular momentum dynamics can alternatively be understood in a classical manner. As we will show below, the periodic motions for $\hat{J}_{y}(t)$ and $\hat{J}_{z}(t)$ [equivalently $\hat{S}_{y}(t)$ and $\hat{S}_{z}(t)$] correspond to the periodic oscillation of a classical non- rigid pendulum around its equilibrium position, while the conservation of $\hat{J}_{x}(t)$ [$\hat{S}_{x}(t)$] relates to the energy conservation of this pendulum. To this end, we first adopt the mean-field approximation as $\hat{b}_{\beta}=b_{\beta}$ ($\beta=L,R$) with $b_{\beta}$ being a $c$-number GPE_1 . The quantum operators $\hat{S}_{x}$, $\hat{S}_{y}$ and $\hat{S}_{z}$ then should be rewritten as $\displaystyle S_{x}$ $\displaystyle=\frac{1}{2J}(b_{L}^{\ast}b_{R}+b_{R}^{\ast}b_{L}),~{}~{}~{}S_{y}=-\frac{i}{2J}(b_{L}^{\ast}b_{R}-b_{R}^{\ast}b_{L}),$ $\displaystyle S_{z}$ $\displaystyle=\frac{1}{2J}(b_{L}^{\ast}b_{L}-b_{R}^{\ast}b_{R}).$ (10) Employing the phase-density representation for $b_{\beta}$ as $b_{\beta}=\sqrt{N_{\beta}^{B}}e^{i\theta_{\beta}}$ and further introducing the two conjugate variables $Z=(N_{L}^{B}-N_{R}^{B})/N_{B},~{}~{}~{}~{}~{}~{}\varphi=\theta_{R}-\theta_{L},$ (11) representing the relative population imbalance between the two wells and the relative phase difference, respectively, we arrive at $S_{x}=\sqrt{1-Z^{2}}\text{cos}\varphi,~{}~{}~{}S_{y}=\sqrt{1-Z^{2}}\text{sin}\varphi,~{}~{}~{}S_{z}=Z,$ (12) whose dynamics are governed by the Hamiltonian $H_{cl}=-J_{0}\sqrt{1-Z^{2}}\text{cos}\varphi,$ (13) which, as aforementioned, describes a non-rigid pendulum with angular momentum $Z$ whose length is proportional to $\sqrt{1-Z^{2}}$ BJJ_Rabi_1 ; BJJ_Rabi_2 ; BJJ_Rabi_3 ; BJJ_Driven_1 . Comparing the Eq. (12) to the Eq. (13), we note that $S_{y}$ and $S_{z}$, being the classical counterpart of the quantum operators $\hat{S}_{y}$ and $\hat{S}_{z}$, now represent the horizontal displacement and the angular momentum of this classical pendulum, while the $S_{x}$ ($\hat{S}_{x}$) proportions to its total energy which is conserved during the dynamics. In this way, an one-to-one correspondence between the quantum and classical dynamics is established in which the periodic motions for $\hat{S}_{y}(t)$ and $\hat{S}_{z}(t)$ are mapped to the periodic oscillations for this classical pendulum around its equilibrium position. Since our focus is put on the dynamics of the population imbalance of the bosons, we compare the quantum evolution $\hat{S}_{z}(t)$ for the case $g_{IB}=0$ to the classical dynamics $Z(t)$ in Fig. 1 (b) and no discrepancies are observed among them. Hence, for the case $g_{IB}=0$, we will always refer the classical $Z(t)$ dynamics as the quantum $S_{z}(t)$ evolutions. However, it should also be emphasized that the agreement between $\hat{S}_{z}(t)$ and $Z(t)$ takes place only for this non- interacting case. For $g_{IB}>0$, on one side, the mixture has no classical mapping, on the other side, the quantum correlations among the bosons come into play, and, as a result, one can witness even a completely different quantum dynamics as compared to the classical one, albeit the fact that the bare Bose-Bose interaction always vanishes (see below). The above classical interpretation provides us not only with a vivid picture for visualizing the quantum dynamics in a classical manner, but also with the profound physical insights with respect to its overall dynamical properties. In particular, the periodic motions for $\hat{S}_{y}(t)$ and $\hat{S}_{z}(t)$ obtained from the quantum simulations are a direct consequence of the integrability of the classical Hamiltonian $H_{cl}$. Owing to the energy conservation for the case $g_{IB}=0$, $H_{cl}$ is completely integrable with all the corresponding classical trajectories, characterized by $[Z(t),\varphi(t)]$, being periodic in time PS . Such an integrability is also transparently shown in the classical phase space [see Fig. 1 (c)]. Depending on the initial condition, two distinguished types of motions are clearly observed: a periodic trajectory orbiting around the fix point either located at $(Z=0,\varphi=0)$ or $(Z=0,\varphi=\pi)$, referred as the zero- and the $\pi$-phase mode for a 1D BJJ BJJ_2 . Figure 1: (Color online) (a) Single-particle spectrum for the double-well potential, in which the gray horizontal lines denote the lowest six energy levels and the blue (red) arrows represent possible transitions that reverse (preserve) the spatial parity of the impurity. (b) Real-time dynamics for the bosonic population imbalance $S_{z}(t)$ for the initial state $|\Psi(0)\rangle=|\phi_{0}\rangle\otimes|\pi/2,\pi/4\rangle$ and for the case $g_{IB}=0$ (red solid line), together with the classical $Z(t)$ dynamics starting from the phase point $(Z=0,\varphi=\pi/4)$ (blue dashed line). (c) Classical phase space for $J_{0}=0.071$. The red dot denotes the phase space point ($Z=0,\varphi=\pi/4$) corresponding to the ACS $|\theta,\varphi\rangle=|\pi/2,\pi/4\rangle$. ### II.3 Breaking of the integrability In contrast to the above integrable limit, the presence of the impurity-Bose interaction leads to the energy transport between the two species and, hence, breaks the integrability for the bosonic species. In order to elaborate on this process in more detail, we decompose the interspecies interaction into various impurity-boson pair excitations $\hat{H}_{IB}=\sum_{i,j=0}^{\infty}\sum_{\alpha,\beta=L,R}U_{ij\alpha\beta}\hat{a}_{i}^{\dagger}\hat{a}_{j}\hat{b}_{\alpha}^{\dagger}\hat{b}_{\beta},$ (14) with $U_{ij\alpha\beta}=g_{IB}\int dx~{}\phi_{i}(x)\phi_{j}(x)u_{\alpha}(x)u_{\beta}(x)$ and $\\{\phi_{i}(x)\\}$ being the single-particle basis for the DW potential. Moreover, $u_{L/R}(x)$, being the above mentioned localized Wannier-like states, are constructed via a linear superposition of the lowest two eigenstates $\phi_{0}(x)$ and $\phi_{1}(x)$. Note that Eq. (14) is obtained by means of an expansion of the field operator for the impurity $\hat{\psi}_{I}(x)=\sum_{i=0}^{\infty}\phi_{i}(x)\hat{a}_{i}$, meanwhile, by employing the two-mode approximation in Eq. (3) for the bosonic species. Besides, all the eigenstate wavefunctions $\\{\phi_{i}(x)\\}$ are chosen to be real due to the preserved time-reversal symmetry in the single-particle Hamiltonian $\textit{h}_{\sigma}$. Next, we group different pair excitations with respect to their bosonic indices as $\displaystyle\hat{H}_{IB}$ $\displaystyle=\left[\sum_{i,j=0}^{\infty}U_{ijLR}\hat{a}_{i}^{\dagger}\hat{a}_{j}\hat{b}_{L}^{\dagger}\hat{b}_{R}+\sum_{i,j=0}^{\infty}U_{ijRL}\hat{a}_{i}^{\dagger}\hat{a}_{j}\hat{b}_{R}^{\dagger}\hat{b}_{L}\right]$ $\displaystyle+\left[\sum_{i,j=0}^{\infty}U_{ijLL}\hat{a}_{i}^{\dagger}\hat{a}_{j}\hat{b}_{L}^{\dagger}\hat{b}_{L}+\sum_{i,j=0}^{\infty}U_{ijRR}\hat{a}_{i}^{\dagger}\hat{a}_{j}\hat{b}_{R}^{\dagger}\hat{b}_{R}\right].$ By noticing the fact that $U_{ijLR}=U_{ijRL},~{}~{}U_{ijLL}=\eta U_{ijRR}$ (16) with $\eta=1$ ($\eta=-1$) for $n_{e,o}=|i-j|$ being an even (odd) number, together with the definitions given in Eq. (5), we finally arrive at $\displaystyle\hat{H}_{IB}$ $\displaystyle=\left(\sum_{i,j=0}^{\infty}U_{ij}^{(1)}\hat{a}_{i}^{\dagger}\hat{a}_{j}\right)2\hat{J}_{x}+\left(\sum_{|i-j|=n_{e}}^{\infty}U_{ij}^{(2)}\hat{a}_{i}^{\dagger}\hat{a}_{j}\right)\hat{N}_{B}$ $\displaystyle+\left(\sum_{|i-j|=n_{o}}^{\infty}U_{ij}^{(3)}\hat{a}_{i}^{\dagger}\hat{a}_{j}\right)2\hat{J}_{z}$ $\displaystyle=\hat{H}_{IB}^{(1)}+\hat{H}_{IB}^{(2)}+\hat{H}_{IB}^{(3)}.$ (17) Here $U_{ij}^{(1)}=U_{ijLR}=U_{ijRL}$, $U_{ij}^{(2)}=U_{ijLL}=U_{ijRR}$ and $U_{ij}^{(3)}=U_{ijLL}=-U_{ijRR}$. Let us emphasize that the Eq. (16) relies on the fact that the DW potential is spatially symmetric, as a result, all its single-particle eigenstates $\\{\phi_{i}\\}$ respect the spatial parity symmetry. Equation (17) transparently elaborates how the interspecies interaction $\hat{H}_{IB}$ breaks the integrability for the Hamiltonian $\hat{H}_{B}$. Since both $\hat{H}_{IB}^{(1)}$ and $\hat{H}_{IB}^{(2)}$ commute with $\hat{H}_{B}$ [c.f. Eq. (6)], it is the non-commutativity between $\hat{H}_{IB}^{(3)}$ and $\hat{H}_{B}$ that results in the energy non- conservation for the bosonic species, and breaks its integrability for $g_{IB}=0$. Further inspecting the $\hat{H}_{IB}^{(3)}$ term in more detail, we notice that it corresponds to all the different single-particle excitations that reverse the impurity’s spatial parity [see Fig. 1 (a) for a schematic illustration]. With this, we conclude that those parity non-conservation transitions of the impurity leads to the integrability breaking for the majority bosonic species. ### II.4 Initial condition We prepare our impurity-Bose mixture initially as $|\Psi(0)\rangle=|\phi_{n}\rangle\otimes|\theta,\varphi\rangle$, being a product state between the two species. Here $|\phi_{n}\rangle$ is the $n$-th single-particle eigenstate for the impurity and $|\theta,\varphi\rangle$ denotes the ACS given by ACS_1 ; ACS_2 $\displaystyle|\theta,\varphi\rangle$ $\displaystyle=\frac{1}{\sqrt{N_{B}!}}\left[\text{cos}(\frac{\theta}{2})\hat{b}^{\dagger}_{L}+\text{sin}(\frac{\theta}{2})e^{i\varphi}\hat{b}^{\dagger}_{R}\right]^{N_{B}}~{}|vac\rangle$ $\displaystyle=\sum_{N^{B}_{L}=0}^{N_{B}}\left(\begin{array}[]{c}N_{B}\\\ N^{B}_{L}\end{array}\right)^{1/2}\text{cos}^{N^{B}_{L}}(\theta/2)~{}\text{sin}^{N^{B}_{R}}(\theta/2)~{}e^{iN^{B}_{R}\varphi}~{}|N^{B}_{L},N^{B}_{R}\rangle,$ (20) which is the linear superposition of all the number states $\\{|N^{B}_{L},N^{B}_{R}\rangle\\}$ and fulfills the completeness relation $(N_{B}+1)\int\frac{d\Omega}{4\pi}|\theta,\varphi\rangle\langle\theta,\varphi|=1$ (21) with $d\Omega=\text{sin}\theta d\theta d\varphi$ being the volume element. Physically, the ACS $|\theta,\varphi\rangle$ corresponds to the classical state $(Z,\varphi)$ in such a way that $\text{cos}\theta=(N^{B}_{L}-N^{B}_{R})/N_{B}=Z$ controls the initial population difference for the bosons and $\varphi$, possessing the same meaning with its classical counterpart, determines the phase difference between the two wells BJJ_4 . For a given ACS $|\theta,\varphi\rangle$, the mean values for the angular-momentum operators introduced in Eq. (5) are BJJ_Rabi_3 $\langle\hat{S}_{x}\rangle=\text{sin}\theta\text{cos}\varphi,~{}~{}~{}\langle\hat{S}_{y}\rangle=\text{sin}\theta\text{sin}\varphi,~{}~{}~{}\langle\hat{S}_{z}\rangle=\text{cos}\theta,$ (22) which satisfies the normalization condition $\langle\hat{S}_{x}\rangle^{2}+\langle\hat{S}_{y}\rangle^{2}+\langle\hat{S}_{z}\rangle^{2}=1$. Together with the Eqs. (8) and (9), we conclude that, for the case $g_{IB}=0$, the motion of the $\hat{\vec{S}}(t)$ vector starting from an arbitrary ACS always lies on a Bloch sphere with unit radius. Even for the case $g_{IB}>0$, where the vector $\hat{\vec{S}}(t)$ can jump out of the Bloch sphere significantly, the use of the ACS still allows us to visualize the quantum trajectory in a classical manner (see below), which simplifies the analysis of the complex quantum dynamics to a large extent, meanwhile, provides insights for the classical-quantum correspondence. Finally, let us note that the ACS has been implemented in recent ultracold experiments in a controllable manner. Tuning a two-photon transition between two hyperfine states of ${}^{87}\textrm{Rb}$ atoms, allows for preparing an ACS with arbitrary $|\theta,\varphi\rangle$ ACS_3 ; ACS_4 . In this paper, we aim at exploring the dynamical response of the majority bosonic species to the presence of the impurity. To this end, we quench at $t=0$ the impurity-Bose interaction strength from initial $g_{IB}=0$ to some finite value $g_{IB}>0$, and monitor the quantum evolution of the bosonic population imbalance starting from a balanced population. While the initial state for the mixture is $|\Psi(0)\rangle=|\phi_{n}\rangle\otimes|\theta,\varphi\rangle$, without other specifications, we always choose the bosonic part being $|\theta,\varphi\rangle=|\pi/2,\pi/4\rangle$. The corresponding $S_{z}(t)$ dynamics for this initial ACS and for the case $g_{IB}=0$ has been detailed discussed above and is presented in Fig. 1 (b) (red solid line). Furthermore, we also consider the scenarios for various initial impurity states $|\phi_{n}\rangle$, so as to explore its impact on the bosonic dynamics. ## III Bosonic ITC transition ### III.1 Onset of quantum chaos Let us first focus on the case where the impurity is initially prepared in its ground state. The many-body initial state for the mixture is then given by $|\Psi(0)\rangle=|\phi_{0}\rangle\otimes|\pi/2,\pi/4\rangle$. Fig. 2 depicts the real-time population imbalance for the bosonic species $S_{z}(t)$ for various fixed postquench impurity-Bose interaction strengths $g_{IB}=0.01$ [Fig. 2 (a)], $g_{IB}=0.1$ [Fig. 2 (b)] and $g_{IB}=1.0$ [Fig. 2 (c)], together with the classical $Z(t)$ dynamics (all blue dashed lines) which, as aforementioned, equivalents to the $S_{z}(t)$ for $g_{IB}=0$. For a weak impurity-Bose interaction ($g_{IB}=0.01$), the $S_{z}(t)$ dynamics is only slightly perturbed by the presence of the impurity, as a result, it leads to the small deviations of the population imbalance between the quantum and the classical simulations [c.f. Fig. 2 (a), red solid line and blue dashed line]. For a larger time scale ($t>5000$), a “collapse-and-revival” behavior for $S_{z}(t)$ is observed (result is not shown here), manifesting its near integrability in this weak interacting regime BJJ_Rabi_3 ; BJJ_Driven_1 . Further increasing the interaction strength, the quantum $S_{z}(t)$ evolution becomes much more complicated and large discrepancies between $S_{z}(t)$ and ${Z}(t)$ are observed with respect to both the oscillation amplitude and the frequencies. For the case $g_{IB}=1.0$, the quantum $S_{z}(t)$ dynamics finally becomes completely irregular [c.f. Fig. 2 (c), red solid line], signifying the onset of quantum chaos for the bosonic species. In order to diagnose such an ITC transition, meanwhile, to quantify the degree of the above observed quantum chaos, we employ the linear entanglement entropy (EE) $S_{L}=1-\text{tr}\hat{\rho}_{1B}^{2}$ (23) for the bosonic species, which represents the bipartite entropy between the single boson and the $N_{B}-1$ bosons after tracing out the impurity QKT_7 ; QKT_8 . Here $\hat{\rho}_{1B}$ stands for the reduced one-body density matrix for the bosonic species dma1_1 ; dma1_2 ; BJJ_4 . Before proceeding, let us point out the reason for not using the spectral statistics as an indicator for the quantum chaos. Similar to the situation for a single particle in a 1D harmonic trap, the single DOF of the Hamiltonian $\hat{H}_{B}$ for a fixed particle number violates the Berry-Tabor conjecture, which states that the energy level spacing distribution follows the universal Poisson form for an integrable system level_1 ; level_2 ; level_3 . As a result, the variation of the level distribution for our mixture upon the increase of $g_{IB}$ can behave largely different as compared to other systems level_3 , and hence, it is insufficient to capture the quantum chaos. Upon a spectral decomposition of the reduced density matrix $\hat{\rho}_{1B}$, $S_{L}$ in Eq. (23) can be expressed, with respect to the natural populations $\\{n_{1},n_{2}\\}$, as $S_{L}=1-\sum_{i=1}^{2}n^{2}_{i}$. In this way, the linear EE alternatively measures the degree of the decoherence for the bosonic species. Note that the two-mode expansion employed in the Eq. (3) renders the single-particle Hamiltonian $\textit{h}(x)$ being restricted to a two-dimensional Hilbert space and thus gives rise to only two natural populations obtained from the spectral decomposition BJJ_4 . For the case where all the bosons reside in the same single-particle state, the bosonic species is of complete coherence, as a result, we have $S_{L}=0$. By contrast, for the case of maximal decoherence we have $n_{1}=n_{2}=1/2$, which gives rise to the upper bound for the linear EE as $S_{L}=1/2$. The linear EE has been extensively used in the QKT systems as a signature of the quantum chaos QKT_7 ; QKT_8 . Depending on whether the corresponding classical trajectory is regular or chaotic, the linear EE behaves as either as rapidly growing or a slowly varying in a short time (referred to as the Ehrenfest time). On the other hand, the infinite-time averaged values of the linear EE for various initial ACSs additionally characterize the edge of the quantum chaos, denoted as the border between the integrable and the chaotic region in the corresponding classical phase space QKT_7 ; QKT_8 . Fig. 3 (a) reports the transient dynamics of the linear EE for the cases examined in Fig. 2 (a-c). At short times ($t<200$), the $S_{L}(t)$ evolution for a stronger interaction exhibits a more rapid growth as compared to the cases with a smaller $g_{IB}$. This is particularly obvious for the case $g_{IB}=1.0$, where we observe the linear EE surges to the value $S_{L}=0.38$ at $t=10$, while it only reaches to $S_{L}=0.02$ ($S_{L}=0.0007$) for the case $g_{IB}=0.1$ ($g_{IB}=0.01$). With this knowledge, we conclude that the different transient dynamical behaviors of the linear EE fully capture the ITC transition that is observed in the dynamics of the bosonic population imbalance. Besides, we shall also note that the linear EE for $t=0$ trivially vanishes since all the bosons are initially condensed into the same single- particle state [c.f. Eq. (20)]. Having investigated the transient dynamics of the linear EE for a specific ACS, let us now explore its asymptotic behaviors with respect to different ACSs, which, as aforementioned, characterize the edge of the quantum chaos. To this end, we compute the infinite-time averaged value of the linear EE (ITEE) for the initial state $|\Psi(0)\rangle=|\phi_{0}\rangle\otimes|\theta,\varphi\rangle$, $\overline{S}_{L}(\theta,\varphi)=lim_{T\rightarrow\infty}~{}\frac{1}{T}\int_{0}^{T}dt~{}S_{L}(t).$ (24) Note that, the impurity initially always occupies the ground state $|\phi_{0}\rangle$ and in our practical numerical simulations the time average is performed up to $t=10^{4}$, being much larger than any other time scales involved in the dynamics. Before proceeding, let us point out the geometrical interpretation of the ITEE value. To show it, we first of all rewrite the linear EE in Eq. (23) for time $t$ as QKT_7 ; QKT_8 $S_{L}(t)=\frac{1}{2}\left[1-\sum_{\gamma=x,y,z}\langle\hat{S}_{\gamma}\rangle^{2}(t)\right],$ (25) where we have used the relation $\hat{\rho}_{1B}=\frac{1}{2}\left[1+\sum_{\gamma=x,y,z}\langle\hat{S}_{\gamma}\rangle\hat{\sigma}_{\gamma}\right],$ (26) with $\\{\hat{\sigma}_{\gamma}\\}$ being the Pauli matrices. Since $S_{L}(t)$ is proportional to the instant distance of the vector $\hat{\vec{S}}(t)$ to the Bloch sphere, $\overline{S}_{L}$ thus measures its averaged distance for the entire dynamics. From the results in the QKT systems QKT_7 ; QKT_8 , we note that there exists a clear correspondence between the ITEE values and the classical phase space structure. Regions of low ITEE correspond to regular trajectories, while regions of high EE correspond to the chaotic trajectories. Moreover, a sudden change of the ITEE value takes place as one crosses the border between the integrable and the chaotic region, which, as aforementioned, characterizes the edge of the quantum chaos. Fig. 3 (b) depicts the computed ITEE values for various ACSs for the case $g_{IB}=1.0$. Note that, we have rescaled the $\theta$ axis to $\text{cos}\theta$ since $\text{cos}\theta=Z$ [see discussions in Sec. II.4]. Varying the initial ACS, the ITEE value varies accordingly. In particular, regions close to ($\text{cos}\theta=0,\varphi=\pi$) and ($\text{cos}\theta=\pm 0.8,\varphi=0,2\pi$) possess significant low ITEE values as compared to the other places. Such a $\overline{S}_{L}(\theta,\varphi)$ profile significantly deviates from the structure of the non-interacting classical phase space. Instead, it bears a striking resemblance to the phase space with an attractive Bose-Bose interaction with the positions for those fixed points precisely match with those low ITEE regions [c.f. Fig. 3 (c), red stars]. Hence, we note that it indicates an effective Bose-Bose attraction existing among the bosons. In Sec. IV.3, we will discuss this induced interaction in detail as well as its impact on the bosonic dynamics. Figure 2: (Color online) Time evolution of the bosonic population imbalance $S_{z}(t)$ (red solid lines) for the initial state $|\Psi(0)\rangle=|\phi_{0}\rangle\otimes|\pi/2,\pi/4\rangle$ and for various fixed impurity-Bose interaction strengths, in which (a) $g_{IB}=0.01$, (b) $g_{IB}=0.1$ and (c) $g_{IB}=1.0$. For comparisons, the classical $Z(t)$ dynamcis is depicted as well (all blue dashed lines). Figure 3: (Color online) (a) The linear EE evolutions for the initial state $|\Psi(0)\rangle=|\phi_{0}\rangle\otimes|\pi/2,\pi/4\rangle$ and for the post- quench interaction strengths $g_{IB}=0.01$ (red solid line), $g_{IB}=0.1$ (green solid line) and $g_{IB}=1.0$ (blue solid line). (b) Infinite-time averaged values for the linear EE for $g_{IB}=1.0$ and for various ACSs. (c) A typical classical phase space for the BJJ with an attractive on-site interaction, where the red stars denote the corresponding classical fixed points. ### III.2 Recovery of the integrability In this section, we investigate the scenario where the impurity is initially pumped into a highly excited state. The out-of-equilibrium dynamics again is triggered by a sudden quench of the impurity-Bose interaction strength. Here, our main aim is to show that the integrability of the bosonic species is recovered by means of preparing the impurity in a highly excited state. The initial condition of the impurity, therefore, provides an additional DOF for controlling the ITC transition of the majority bosonic species. Here, we note that the employed notion of “integrability” specifically refers to how close the bosonic dynamics in the interacting cases ($g_{IB}>0$) is to the one in the non-interacting integrable case ($g_{IB}=0$), which is different from the commonly used context in which it is uniquely associated to the system’s Hamiltonian. For an illustrative purpose, we consider the impurity is initially at $|\phi_{150}\rangle$, being the 150-th excited state, and focus on the case for the post-quench interaction strength $g_{IB}=1.0$. The many-body state for $t=0$ is again given by $|\Psi(0)\rangle=|\phi_{150}\rangle\otimes|\pi/2,\pi/4\rangle$. The corresponding quantum evolution of the bosonic population imbalance $S_{z}(t)$ is depicted in Fig. 4 (a) (red solid line). As compared to the classical $Z(t)$ dynamics [Fig. 4 (a), blue dashed line], we find a good agreement between them with negligible discrepancies. Interestingly, these discrepancies are even much smaller than the ones between $S_{z}(t)$ and $Z(t)$ for the case $g_{IB}=0.01$ [c.f. Fig. 2 (a)]. Besides, we also note that the negligible increment of the corresponding linear EE in the course of the dynamics alternatively signifies the recovery of the integrability for the bosonic species [c.f. Fig. 4 (b)]. Figure 4: (Color online) Time evolution of the bosonic population imbalance $S_{z}(t)$ for $g_{IB}=1.0$ and for the initial state $|\Psi(0)\rangle=|\phi_{150}\rangle\otimes|\pi/2,\pi/4\rangle$ (red solid line), together with the classical $Z(t)$ dynamics (blue dashed line) which corresponds to the $S_{z}(t)$ dynamics for $g_{IB}=0$. (b) The evolution of the linear EE for the corresponding case. ## IV Spectral analysis and induced interaction In order to shed light on the physics for the above-analyzed bosonic dynamics, hereafter, we perform a detailed spectral analysis for the mixture with respect to both the energy spectrum and the eigenstates via a numerically exact diagonalization (ED). In particular, we would like to unveil the physical origin for the observed ITC transition for the bosonic species manifested by the corresponding dynamics of the population imbalance. Moreover, we will discuss the presented Bose-Bose attraction induced by the impurity as well as its impact on the bosonic dynamics. ### IV.1 Spectral structure Let us begin with the case for $g_{IB}=0$. In the absence of the interspecies interaction, the two species are completely decoupled. As a result, the eigenenergy of the mixture is trivially given by $E=\epsilon_{k}+\epsilon^{B}_{l}$ with $\epsilon_{k}$ and $\epsilon^{B}_{l}$ being the $k$-th and $l$-th eigenvalue for the subsystem Hamiltonians $\hat{H}_{I}$ and $\hat{H}_{B}$, respectively. Owing to the neglected Bose- Bose interaction, the many-body spectrum for $\hat{H}_{B}$ is always equidistant with the energy difference $2J_{0}$ between the two successive levels, which accounts for the harmonic oscillation of the $S_{z}(t)$ dynamics for the case $g_{IB}=0$ [c.f. Fig. 1 (b)]. As for the impurity, due to the rapid growth of the energy difference between two successive eigenstates, the single-particle spectrum is inhomogeneous in which the high-energy part is much more sparse as compared to the low-energy one [c.f. Fig. 1 (a)]. An important consequence for such a spectral structure on the mixture’s many-body spectrum is the following. For $\delta_{i}>\Delta_{B}$, with $\delta_{i}=\epsilon_{i+1}-\epsilon_{i}$ being the energy difference between the $i$-th and the $(i+1)$-th single-particle eigenstates for the DW potential (see also the discussions in Sec. II.1), and $\Delta_{B}$ representing the width of the spectrum for the Hamiltonian $\hat{H}_{B}$, a band-like structure is naturally formed in the high-energy part of the many-body spectrum with the band gap being $\delta_{i}-\Delta_{B}$, meanwhile, the energy levels within each band are equidistant. This simple picture, however, ceases to be valid upon the variation of the impurity-Bose interaction. Indeed, the inclusion of the interspecies interaction introduces additional coupling between the two subsystems and, as a result, our spectral analysis needs to be performed with respect to the complete mixture. Figure 5 showcases the many-body spectrum as a function of the interspecies interaction strength $g_{IB}$. Owing to the preserved spatial parity symmetry in the Hamiltonian $\hat{H}$, we present here only half of the spectrum which corresponds to the even parity eigenstates. With the increase of $g_{IB}$, the low-energy spectrum shows many avoided-crossings among the energy levels, which is in sharp contrast to the high-energy spectrum where only a linear growth of their values is observed [c.f. Figs. 5 (a) and (c)]. Moreover, for the high-energy spectrum, features like the band-like structure as well as the equidistant energy levels within each band that are present in the non-interacting limit are retained in the interacting cases as well. The above two distinguished spectral behaviors can roughly be understood via the structure of the impurity’s single-particle spectrum [c.f. Fig. 1 (a)]. Owing to the large energy separations among those highly excited states, the transitions for the impurity among those states are significantly prohibited. From a many-body perspective, the resulting high-energy effective Hamiltonian of the mixture reads $\hat{H}^{\prime}=\hat{H}_{I}+\hat{H}_{B}+\hat{H}_{IB}^{\prime}$, with $\displaystyle\hat{H}_{I}$ $\displaystyle=\sum_{i\gg 1}\epsilon_{i}\hat{a}_{i}^{\dagger}\hat{a}_{i},~{}~{}~{}~{}~{}~{}~{}~{}~{}~{}\hat{H}_{B}=-2J_{0}\hat{J}_{x},$ $\displaystyle\hat{H}_{IB}^{\prime}$ $\displaystyle\approx\sum_{i\gg 1}2U_{i}^{(1)}\hat{J}_{x}+U_{i}^{(2)}\hat{N}_{B}\approx\sum_{i\gg 1}U_{i}^{(2)}\hat{N}_{B}.$ (27) Here $U_{i}^{(1)}=U_{iiLR}=U_{iiRL}$, $U_{i}^{(2)}=U_{iiLL}=U_{iiRR}$ and we notice that $U_{i}^{(1)}=g_{IB}\int dx~{}\phi_{i}(x)\phi_{i}(x)u_{L}(x)u_{R}(x)\approx 0$, due to the negligible spatial overlap between the two localized states $u_{L}(x)$ and $u_{R}(x)$. Before proceeding, we note the validity condition for the above high-energy effective Hamiltonian as: $\delta_{i}\gg\epsilon_{IB}$ and $\delta_{i}\gg\Delta_{B}$ with $\epsilon_{IB}$ being the interspecies interaction energy per particle. Eq. (27) explains the observed high-energy spectral behaviors as follows: since the interspecies interaction $\hat{H}_{IB}^{\prime}$ now becomes the “zero-point” energy of the mixture, the increment of the $g_{IB}$ thus only raises the energy level for those highly excited states. As a result, the band-like structure as well as equidistant nature that are formed in the non-interacting case are naturally preserved. In contrast, the densely distributed low-energy (single-particle) spectrum of the impurity facilitates the transitions among different (low-lying) many-body eigenstates caused by the interspecies interaction $\hat{H}_{IB}$ [c.f. Eq. (17)]. With increasing $g_{IB}$, this results in the above observed avoided level-crossings among the low-energy many-body spectrum QKT_1 . ### IV.2 Eigenstate delocalization The avoided level-crossings in the low-energy spectrum impact the characteristics of the corresponding eigenstates as well. Specifically, it results in a significant delocalization for those low-lying eigenvectors with respect to an integrable basis (see below), which, in turn, accounts for the chaotic nature of the bosonic non-equilibrium dynamics. To demonstrate this, we introduce the Shannon entropy $S^{S}_{j}=-\sum_{k}c_{j}^{k}\ln{c_{j}^{k}}$ (28) for a many-body eigenstate $|\Phi_{j}\rangle$ of the mixture as a measure of the delocalization IPR_1 ; IPR_2 . Here $c_{j}^{k}=|\langle\psi_{k}|\Phi_{j}\rangle|^{2}$ with $\\{|\psi_{k}\rangle\\}$ being the eigenbasis for the Hamiltonian $\hat{H}_{B}$ that are used as the “integrable basis”. The Shanon entropy thereby measures the number of this integrable basis vectors that contribute to each eigenstate. As a result, the lower the Shanon entropy value is the closer this eigenstate $|\Phi_{j}\rangle$ is to a non-interacting eigenvector. From the random matrix theory (RMT), for a chaotic system described by the gaussian orthogonal ensemble (GOE), the amplitudes $c_{j}^{k}$ are independent random variables and all eigenstates are completely delocalized QKT_1 . However, due to the spectral fluctuations the weights $\\{c_{j}^{k}\\}$ fluctuate around $1/D$, yielding the averaged value $S_{\text{GOE}}=\ln{(0.48D)}$ IPR_1 ; IPR_2 . Here, we refer to $D=N_{B}+1$ as the Hilbert space dimension for the bosonic species, which is different from the single-species cases IPR_1 ; IPR_2 . Figure 6 (a) presents the Shannon entropy of the many-body eigenstates as a function of their quantum numbers $j$ (sorted in the ascending order with respect to the energy) for the case $g_{IB}=1.0$. The distinguished localization nature between the low-lying and the highly excited eigenvectors are clearly exhibited. While those low-energy eigenvectors are delocalized with the corresponding Shannon entropy values close to the result from the GOE $S_{\text{GOE}}=3.8812$, for increasing $j$, a decrease of the $S^{S}_{j}$ value is clearly observed, indicating those high-energy eigenvectors are significantly localized. Thus, we may further conjecture that $S^{S}_{j}\rightarrow 0$ for $j\rightarrow\infty$. Physically, the avoided level-crossings in the low-energy spectrum results in a strong mixing of different eigenstates with respect to their physical properties. In this way, an eigenstate from the non-interacting basis can be largely delocalized after experiencing a serious of avoided level-crossings QKT_1 . On the other hand, the localized nature for those high-lying excited states can also be readily seen from the effective Hamiltonian in Eq. (27). Since here $\hat{H}_{IB}^{\prime}$ corresponds to the “zero-point ” energy of the mixture, it is not surprising that the interacting basis (eigenstates of the mixture for $g_{IB}>0$) is similar to the non-interacting integrable basis. Before proceeding, let us highlight that the degree of the localization for an eigenstate $|\Phi_{j}\rangle$ also reflects the degree of the encoded entanglement between the impurity and the majority bosons. To see this, we employ the von Neumann entropy for an eigenstate $|\Phi_{j}\rangle$ few_gs_7 ; Schmidt , $S^{V}_{j}=-\text{tr}(\hat{\rho}_{j}\ln{\hat{\rho}_{j}})$ (29) with $\hat{\rho}_{j}=|\Phi_{j}\rangle\langle\Phi_{j}|$ being the corresponding density matrix. For the case where the two species are non-entangled, the eigenstate $|\Phi_{j}\rangle$ is simply of a product form with respect to the wavefunctions of the two species. Correspondingly, it gives rise to the von Neumann entropy $S^{V}_{j}=0$. By contrast, any existing entanglement between the two species will lead to an increase of the von Neumann entropy, therefore, one can anticipate large $S^{V}_{j}$ values for those highly entangled eigenstates. The corresponding von Neumann entropies for various eigenstates for the case $g_{IB}=1.0$ are shown in Fig. 6 (b). As compared to the Fig. 6 (a), a striking resemblance between the $S^{V}_{j}$ and $S^{S}_{j}$ distributions are transparently observed, manifesting the existence of the correspondence between a delocalized (localized) eigenstate to a large (small) von Neumann entropy value. Based on this knowledge, we refer to the above eigenstate delocalization as the entanglement induced delocalization. Finally, let us discuss the impact of the eigenstate delocalization to the bosonic non-equilibrium dynamics. For the case $|\Psi(0)\rangle=|\phi_{0}\rangle\otimes|\pi/2,\pi/4\rangle$, the initial state is mainly a linear superposition of those low-lying eigenvectors for both $g_{IB}=0$ and $g_{IB}=1.0$ [c.f. Fig. 6 (c), the left part]. Owing to the delocalization nature for the eigenstates of the mixture for large interspecies interactions, the expansion coefficients $\\{A_{j}^{1}\\}$ for $g_{IB}=1.0$ are broadly distributed as compared to the ones ($\\{A_{j}^{0}\\}$) for $g_{IB}=0.0$, reflecting the fact that much more eigenstates are involved in the bosonic dynamics. Since the energy levels for the interacting case are no longer equidistant, it thus gives rise to the completely irregular behaviors for the above $S_{z}(t)$ dynamics [c.f. Fig. 2 (c)]. In contrast, those highly excited states in the interacting basis preserve the main features of the non-interacting basis, leaving a similar distribution of the corresponding expansion coefficients [c.f. Fig. 6 (c), the right part]. Together with the equidistant nature for those high-lying energy levels, it thereby accounts for the integrable $S_{z}(t)$ motion for the initial state $|\Psi(0)\rangle=|\phi_{150}\rangle\otimes|\pi/2,\pi/4\rangle$ and for the case $g_{IB}=1.0$. Figure 5: (Color online) Energy spectrum of the mixture as a function of impurity-Bose interaction strength $g_{IB}$. (a) High-energy part of the spectrum, (b) A zoom-in view of the high-energy spectrum, (c) Low-energy part of the spectrum, (d) A zoom-in view of the low-energy spectrum. Figure 6: (Color online) (a) Shannon entropy $S^{S}_{j}$ for the many-body eigenstates as a function of quantum number $j$ for the case $g_{IB}=1.0$. The red dashed line denotes the Shannon entropy from the GOE $S_{\text{GOE}}=3.8812$. (b) Von Neumann entropy $S^{V}_{j}$ for the eigenstates for the case $g_{IB}=1.0$. (c) Expansion coefficients $A_{j}=|\langle\Psi(0)|\Phi_{j}\rangle|^{2}$ with respect to eigenstates for initial states $|\Psi(0)\rangle=|\phi_{0}\rangle\otimes|\pi/2,\pi/4\rangle$ (left part) and $|\Psi(0)\rangle=|\phi_{150}\rangle\otimes|\pi/2,\pi/4\rangle$ (right part) and for the cases $g_{IB}=0.0$ (red solid line and are denoted as $A_{j}^{0}$) and $g_{IB}=1.0$ (blue dashed line and are denoted as $A_{j}^{1}$). ### IV.3 Induced Bose-Bose attraction The presence of the impurity not only brings the bosonic species into the chaotic regime, yielding an irregular behavior for the corresponding $S_{z}(t)$ motion, but also fundamentally changes its dynamical properties. As we will show below, the impurity effectively induces an attractive Bose-Bose interaction, which, in turn, leads to a completely different quantum trajectory as compared to the integrable case. To show it, we employ the time- averaged Husimi distribution (TAHD) TAHD_1 ; BJJ_4 ; QKT_7 $\overline{Q}_{H}(\theta,\varphi)=lim_{T\rightarrow\infty}~{}\frac{1}{T}\int_{0}^{T}Q_{H}(\theta,\varphi,t)dt,$ (30) with $Q_{H}(\theta,\varphi,t)=\frac{N_{B}+1}{4\pi}\langle\theta,\varphi|\hat{\rho}_{B}(t)|\theta,\varphi\rangle,$ (31) and $\hat{\rho}_{B}(t)$ being the reduced density matrix for the bosonic species after tracing out the impurity. According to the Eq. (21), $Q_{H}(\theta,\varphi,t)$ satisfies the normalization condition $\int Q_{H}(\theta,\varphi,t)d\Omega=1$. Physically, the TAHD represents the probability for the bosons to locate at a specific ACS $|\theta,\varphi\rangle$ averaged over the entire dynamics, which, with respect to its physical meaning, resembles to the probability density function (PDF) for a classical trajectory. In this sense, we note that the TAHD represents a quantum trajectory in an averaged manner. The computed TAHD for the initial state $|\Psi(0)\rangle=|\phi_{0}\rangle\otimes|\pi/2,\pi/4\rangle$ and for the case $g_{IB}=0$ is depicted in Fig. 7 (a), together with the classical trajectory governed by the Hamiltonian $H_{cl}$ and starting from the phase point $(Z=0,\varphi=\pi/4)$ (black solid line). Compared to the classical trajectory, we note that the TAHD profile fully captures its main characteristic with those high $\overline{Q}_{H}(\theta,\varphi)$ regions precisely matching the positions for this classical trajectory, which additionally manifests the agreement between the quantum $S_{z}(t)$ and classical $Z(t)$ dynamics for the case $g_{IB}=0$ [c.f. Fig. 1 (b)]. The TAHD for $g_{IB}=1.0$, however, deviates from the non-interacting case significantly and bears a striking resemblance to the classical trajectory corresponding to the BH Hamiltonian in Eq. (4) with an on-site attraction [c.f. Figs. 7 (b) and 3 (c)]. In this sense, we conjecture an effective Bose- Bose attraction is induced by the impurity in the dynamics which, in turn, alters the corresponding quantum trajectory. This expectation is indeed confirmed by analyzing the pair-correlation function few_gs_6 ; few_gs_7 ; GPE_1 $g_{2}(\alpha,\beta)=\frac{\rho_{2}^{B}(\alpha,\beta)}{\rho_{1}^{B}(\alpha)\rho_{1}^{B}(\beta)},$ (32) for the bosons, with $\rho_{2}^{B}(\alpha,\beta)$ and $\rho_{1}^{B}(\alpha)$ being the reduced two- and one-body density for the bosonic species and $\alpha,\beta=L,R$. Physically, $\rho_{2}^{B}(L,R)$ denotes a measure for the joint probability of finding one boson at the left well while the second is at the right well. Through the division by the one-body densities, the $g_{2}$ function excludes the impact of the inhomogeneous density distribution and thereby directly reveals the spatial two-particle correlations induced by the interaction few_gs_6 ; few_gs_7 . Based on this knowledge, let us first elaborate the $g_{2}$ function for the non-interacting case, which corresponds to the TAHD depicted in Fig. 7 (a). Since there is no interaction among the particles, all the bosons thus can independently hop between the two wells, hence, it always results in $g_{2}^{o}=g_{2}^{d}=1$, with $g_{2}^{o}=g_{2}(\alpha,\alpha)$ [$g_{2}^{d}=g_{2}(\alpha,\beta\neq\alpha)$] being the two-particle correlations within the same well (between the two wells). By contrast, the presence of the impurity-Bose interaction largely changes the above $g_{2}$ profile. As shown in Fig.7 (c), the $g_{2}$ function quickly deviates from the initial values $g_{2}^{o}=g_{2}^{d}=1$ to $g_{2}^{o}>1$ and $g_{2}^{d}<1$ for $t<5$ and persistently oscillate around their asymptotic values $g_{2}^{o}=1.28$ and $g_{2}^{d}=0.72$, respectively. Physically, such an evolution of the $g_{2}$ function indicates that the bosons are in favor of bunching together with a collective tunneling between the wells in the dynamics, which evidently manifests the existence of the Bose-Bose attraction induced by the impurity-Bose repulsion. Figure 7: (Color online) Time-averaged Husimi distribution for the initial state $|\Psi(0)\rangle=|\phi_{0}\rangle\otimes|\pi/2,\pi/4\rangle$ and for (a) $g_{IB}=0.0$ and (b) $g_{IB}=1.0$. Moreover, the black solid line in (a) denotes the classical trajectory starting from the phase point $(Z=0,\varphi=\pi/4)$. (c) The evolution of the pair-correlation function $g_{2}^{o}(t)$ (blue solid line) and $g_{2}^{d}(t)$ (red solid line) for the case examined in (b). ## V Conclusions and Outlook We have demonstrated that a non-interacting ultracold many-body bosonic ensemble confined in a 1D DW potential can exhibit a chaotic nature due to the presence of a single impurity. We trigger the non-equilibrium dynamics by means of a quench of the impurity-Bose interaction and monitor the evolution of the population imbalance for the bosons between the two wells. While the increase of the post-quench interaction strength always facilitates the chaotic motion for the bosonic population imbalance, it becomes regular again for the cases where the impurity is initially prepared in a highly excited state. Employing the linear entanglement entropy, it not only enables us to characterize such an ITC transition but also implies the existence of an effective Bose-Bose attraction in the dynamics induced by the impurity. In order to elucidate the physical origin for the above observed ITC transition, we perform a detailed spectral analysis for the mixture with respect to both the energy spectrum as well as the eigenstates. In particular, two distinguished spectral behaviors upon a variation of the interspecies interaction strength are observed: while the avoided level-crossings take place in the low-energy spectrum, the energy levels in the high-energy spectrum possess the main features of the integrable limit. Consequently, it results in a significant delocalization for the low-lying eigenvectors which, in turn, accounts for the chaotic nature of the bosonic dynamics. In contrast, those highly excited states bear a high resemblance to the non-interacting integrable basis, rendering the recovery of the integrability for the bosonic species. Finally, we discuss the induced Bose-Bose attraction as well as its impact on the bosonic dynamics. Possible future investigations include the impact on the bosonic dynamics with the inclusion of several additional impurities and/or the bare Bose-Bose repulsion. Since for the latter there exists a competition between the bare Bose-Bose repulsion and the induced attractive interaction, this may significantly affect the bosonic ITC transition. Another interesting perspective is the study of the chaotic dynamics for an atomic mixture consisting of atomic species with different masses. The impact of the higher bands of the DW potential, beyond the two-site BH description for the bosonic species, is also an interesting perspective. ###### Acknowledgements. The authors acknowledge fruitful discussions with A. Mukhopadhyay and X.-B. Wei. This work has been funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) - SFB 925 - project 170620586. K. K. gratefully acknowledges a scholarship of the Studienstiftung des deutschen Volkes. G. X. acknowledges support from the NSFC under Grants No. 11835011 and No. 11774316. ## References * (1) M. R. Andrews, C. G. Townsend, H.-J. Miesner, D. S. Durfee, D. M. Kurn, and W. Ketterle, Science 275, 637 (1997). * (2) A. Rohrl, M. Naraschewski, A. Schenzle, and H. Wallis, Phys. Rev. Lett. 78, 4143 (1997). * (3) M. Albiez, R. Gati, J. Fölling, S. Hunsmann, M. Cristiani, and M. K. Oberthaler, Phys. Rev. Lett. 95, 010402 (2005). * (4) B. D. Josephson, Phys. Lett. 1, 251 (1962). * (5) R. Gati and M. K. Oberthaler, J. Phys. B: At. Mol. Opt. Phys. 40 R61 (2007). * (6) W. D. Phillips, Rev. Mod. Phys. 70, 721 (1998); I. Bloch, J. Dalibard, and W. Zwerger, Rev. Mod. Phys. 80, 885 (2008). * (7) A. Smerzi, S. Fantoni, S. Giovanazzi, and S. R. Shenoy, Phys. Rev. Lett. 79, 4950 (1997). * (8) S. Raghavan, A. Smerzi, S. Fantoni, and S. R. Shenoy, Phys. Rev. A 59, 620 (1999). * (9) G. J. Milburn, J. Corney, E. M. Wright, and D. F. Walls, Phys. Rev. A 55, 4318 (1997). * (10) K. Sakmann, A. I. Streltsov, O. E. Alon, and L. S. Cederbaum, Phys. Rev. A 89, 023602 (2014). * (11) K. Sakmann, A. I. Streltsov, O. E. Alon, and L. S. Cederbaum, Phys. Rev. A 82, 013620 (2010). * (12) J. Estève, C. Gross, A. Weller, S. Giovanazzi, and M. K. Oberthaler, Nature (London) 455, 1216 (2008). * (13) B. Juliá-Díaz, T. Zibold, M. K. Oberthaler, M. Melé-Messeguer, J. Martorell, and A. Polls, Phys. Rev. A 86, 023615 (2012). * (14) F. Haake, Quantum Signatures of Chaos, (Springer, Berlin, Heidelberg, 2010). * (15) F. Haake, M. Kus and R. Scharf, Z. Phys. B 65, 381 (1987). * (16) M. Lombardi and A. Matzkin, Phys. Rev. E 83, 016207 (2011). * (17) R. Alicki, D. Makowiec, and W. Miklaszewski, Phys. Rev. Lett. 77, 838 (1996). * (18) Jayendra N. Bandyopadhyay and A. Lakshminarayan, Phys. Rev. E 69, 016201(2004). * (19) S. Chaudhury, A. Smith, B. E. Anderson, S. Ghose and P. S. Jessen, Nature 461, 768 (2009). * (20) A. Piga, M. Lewenstein, and J. Q. Quach, Phys. Rev. E 99, 032213 (2019). * (21) X. G. Wang, S. Ghose, B. C. Sanders, and B. B. Hu, Phys. Rev. E 70, 016217 (2004). * (22) C. Neill, P. Roushan, M. Fang, Y. Chen, M. Kolodrubetz, Z. Chen, A. Megrant, R. Barends, B. Campbell, B. Chiaro, A. Dunsworth, E. Jeffrey, J. Kelly, J. Mutus, P. J. J. O’Malley, C. Quintana, D. Sank, A. Vainsencher, J. Wenner, T. C. White, A. Polkovnikov, and J. M. Martinis, Nat. Phys. 12, 1037 (2016). * (23) J. B. Ruebeck, J. Lin, and A. K. Pattanayak, Phys. Rev. E 95, 062222 (2017). * (24) U. T. Bhosale and M. S. Santhanam, Phys. Rev. E 95, 012216 (2017). * (25) S. Ghose, R. Stock, P. Jessen, R. Lal, and A. Silberfarb, Phys. Rev. A 78, 042318 (2008). * (26) Y. S. Weinstein, S. Lloyd, and C. Tsallis, Phys. Rev. Lett. 89, 214101 (2002). * (27) M. Heyl, P. Hauke, and P. Zoller, Sci. Adv. 5, eaau8342 (2019). * (28) L. M. Sieberer, T. Olsacher, A. Elben, M. Heyl, P. Hauke, F. Haake, and P. Zoller, npj Quantum Information 5, 78 (2019). * (29) A. N. Wenz, G. Zürn, S. Murmann, I. Brouzos, T. Lompe, S. Jochim, Science 342, 457 (2013). * (30) G. Zürn, A. N. Wenz, S. Murmann, A. Bergschneider, T. Lompe, and S. Jochim, Phys. Rev. Lett. 111, 175302 (2013). * (31) S. Murmann, A. Bergschneider, V. M. Klinkhamer, G. Zürn, T. Lompe, S. Jochim, Phys. Rev. Lett. 114, 080402 (2015). * (32) F. Serwane, G. Zürn, T. Lompe, T. B. Ottenstein, A. N. Wenz, and S. Jochim, Science 332, 336 (2011). * (33) G. Zürn, F. Serwane, T. Lompe, A. N. Wenz, M. G. Ries, J. E. Bohn, and S. Jochim, Phys. Rev. Lett. 108, 075303 (2012). * (34) S. Murmann, F. Deuretzbacher, G. Zürn, J. Bjerlin, S. M. Reimann, L. Santos, T. Lompe, and S. Jochim, Phys. Rev. Lett. 115, 215301 (2015). * (35) A. S. Dehkharghani, F. F. Bellotti, and N. T. Zinner, J. Phys. B: At., Mol., Opt. Phys. 50, 144002 (2017). * (36) H. P. Hu, L. Pan, and S. Chen, Phys. Rev. A 93, 033636 (2016). * (37) A. S. Dehkharghani, A. G. Volosniev, and N. T. Zinner, J. Phys. B: At., Mol., Opt. Phys. 49, 085301 (2016). * (38) D. Pȩcak, A. S. Dehkharghani, N. T. Zinner, and T. Sowiński, Phys. Rev. A 95, 053632 (2017). * (39) K. Keiler, S Krönke and P. Schmelcher, New J. Phys. 20, 033030 (2018). * (40) J. Chen, J. M. Schurer, and P. Schmelcher, Phys. Rev. A 98, 023602 (2018). * (41) J. Chen, J. M. Schurer, and P. Schmelcher, Phys. Rev. Lett. 121, 043401 (2018). * (42) H. P. Hu, L. M. Guan, and S. Chen, New J. Phys. 18, 025009 (2016). * (43) B. Fang, P. Vignolo, M. Gattobigio, C. Miniatura, and A. Minguzzi, Phys. Rev. A 84, 023626 (2011). * (44) M. Pyzh, S Krönke, C. Weitenberg and P. Schmelcher, New J. Phys. 20, 015006 (2018). * (45) A. C. Pflanzer, S. Zöllner, and P. Schmelcher, Phys. Rev. A 81, 023612 (2010). * (46) A. C. Pflanzer, S. Zöllner, and P. Schmelcher, J. Phys. B: At., Mol., Opt. Phys. 42, 231002 (2009). * (47) L. D. Landau and S. I. Pekar, Zh. Eksp. Teor. Fiz. 18, 419 (1948). * (48) R. P. Feynman, Phys. Rev. 97, 660 (1955). * (49) A. S. Alexandrov and J. T. Devreese, Advances in Polaron Physics, (Springer-Verlag, Berlin, 2010). * (50) H. -P. Breuer and F. Petruccione, The Theory of Open Quantum Systems, (Oxford University Press, USA, 2007). * (51) M. Rinck and C. Bruder, Phys. Rev. A 83, 023608 (2011). * (52) F. Mulansky, J. Mumford, and D. H. J. O’Dell, Phys. Rev. A 84, 063602 (2011). * (53) J. Mumford and D. H. J. O’Dell, Phys. Rev. A 90, 063617 (2014). * (54) J. Mumford, J. Larson, and D. H. J. O’Dell, Phys. Rev. A 89, 023620 (2014). * (55) J. Mumford, W. Kirkby, and D. H. J. O’Dell, J. Phys. B: At. Mol. Opt. Phys. 53, 145301 (2020). * (56) J Mumford et al., J. Phys. B: At. Mol. Opt. Phys. 53, 145301 (2020). * (57) M. Olshanii, Phys. Rev. Lett. 81, 938 (1998). * (58) C. Chin, R. Grimm, P. Julienne, and E. Tiesinga, Rev. Mod. Phys. 82, 1225 (2010). * (59) T. Köhler, K. Góral, and P. S. Julienne, Rev. Mod. Phys. 78, 1311 (2006). * (60) A. G. Truscott, K. E. Strecker, W. I. McAlexander, G. B. Partridge, and R. G. Hulet, Science 291, 2570 (2001). * (61) F. Schreck, L. Khaykovich, K. L. Corwin, G. Ferrari, T. Bourdel, J. Cubizolles, and C. Salomon, Phys. Rev. Lett. 87, 080403 (2001). * (62) D. S. Hall, M. R. Matthews, J. R. Ensher, C. E. Wieman, and E. A. Cornell, Phys. Rev. Lett. 81, 1539 (1998). * (63) S. Tojo, Y. Taguchi, Y. Masuyama, T. Hayashi, H. Saito, and T. Hirano, Phys. Rev. A 82, 033609 (2010). * (64) J. Chen, A. K. Mukhopadhyay, and P. Schmelcher, Phys. Rev. A 102, 033302 (2020). * (65) C. J. Pethick and H. Smith, Bose-Einstein Condensation in Dilute Gases, (Cambridge University Press, New York, 2008). * (66) M. Holthaus and S. Stenholm, Eur. Phys. J. B 20, 451 (2001). * (67) M. Tabor, Chaos and Integrability in nonlinear Dynamics: An Introduction, (Wiley-Interscience, USA, 1989). * (68) F. T. Arecchi, Eric Courtens, R. Gilmore, and H. Thomas, Phys. Rev. A 6, 2211 (1972). * (69) J. M. Radcliffe, J. Phys. A: Gen. Phys. 4, 313 (1971). * (70) T. Zibold, E. Nicklas, C. Gross, and M. K. Oberthaler, Phys. Rev. Lett. 105, 204101 (2010). * (71) J. Tomkovic, W. Muessel, H. Strobel, S. Löck, P. Schlagheck, R. Ketzmerick, and M. K. Oberthaler, Phys. Rev. A 95, 011602(R) (2017). * (72) O. Penrose and L. Onsager, Phys. Rev. 104, 576 (1956). * (73) K. Sakmann, A. I. Streltsov, O. E. Alon, and L. S. Cederbaum, Phys. Rev. A 78, 023615 (2008). * (74) M. V. Berry and M. Tabor, Proc. R. Soc. Lond. A 356, 375 (1977). * (75) S. R. Dahmen et al., J. Stat. Mech., P10019 (2004). * (76) L. D’Alessio, Y. Kafri, A. Polkovnikov, and M. Rigol, Adv. Phys. 65, 239 (2016). * (77) L. F. Santos and M. Rigol, Phys. Rev. E 81, 036206 (2010). * (78) L. F. Santos and M. Rigol, Phys. Rev. E 82, 031130 (2010). * (79) A. Pathak, Elements of Quantum Computation and Quantum Communication, (Taylor & Francis, 2013). * (80) K. Husimi, Proc. Phys. Math. Soc. Jpn. 22, 264 (1940).
# Harvest: A Reliable and Energy Efficient Bulk Data Collection Service for Large Scale Wireless Sensor Networks Vinayak Naik Anish Arora BITS Pilani, Goa<EMAIL_ADDRESS>The Ohio State University <EMAIL_ADDRESS> ###### Abstract _We present a bulk data collection service, Harvest, for energy constrained wireless sensor nodes. To increase spatial reuse and thereby decrease latency, Harvest performs concurrent, pipelined exfiltration from multiple nodes to a base station. To this end, it uses a distance-k coloring of the nodes, notably with a constant number of colors, which yields a TDMA schedule whereby nodes can communicate concurrently with low packet losses due to collision. This coloring is based on a randomized CSMA approach which does not exploit location knowledge. Given a bounded degree of the network, each node waits only O $(1)$ time to obtain a unique color among its distance-k neighbors, in contrast to the traditional deterministic distributed distance-k vertex coloring wherein each node waits O$(\Delta^{2})$ time to obtain a color._ _Harvest offers the option of limiting memory use to only a small constant number of bytes or of improving latency with increased memory use; it can be used with or without additional mechanisms for reliability of message forwarding._ _We experimentally evaluate the performance of Harvest using 51 motes in the Kansei testbed. We also provide theoretical as well as TOSSIM-based comparison of Harvest with Straw, an extant data collection service implemented for TinyOS platforms that use one-node at a time exfiltration. For networks with more than 3-hops, Harvest reduces the latency by at least 33% as compared to that of Straw._ ## 1 Introduction Wireless sensor nodes often maintain logs of network, environment, middleware, and application behavior. Examples of logged information include link qualities, network routes, sensory data, mobility traces, exception reports, application statistics, etc. The collection of bulk data from a number of wireless sensor nodes is thus a frequent requirement for testers, operators, managers, modelers, and users. In this paper, we focus on the convergecast of the (potentially different) bulk data logged at a number of nodes to one “base station” node. We consider an “off-line” setting where no other data traffic is present on the network; this case arises when the bulk data collection can materially interfere with ongoing application traffic or when the size/generation-rate of the bulk data exceeds the effective communication capacity of the source nodes with respect to the base station. As networks scale to larger numbers of nodes and communication hops and as the bulk datum sizes grow, the reliability, energy-efficiency, and latency of the collection operation become key issues. While these issues have been well studied in the context of bulk data dissemination, they have received far less attention in the case of bulk data convergecast. Also, since network debugging and management are primary motivations for the collection service, it is desirable that this service have a small footprint in terms of instruction memory, data memory, message overhead, and wireless traffic, and to minimize its dependence on other network services, including localization and time synchronization. In this paper, we present and evaluate a protocol that meets these requirements; we call this protocol Harvest. Harvest achieves its reliability with two measures: First, it schedules the transmissions (of messages containing bulk datum pieces) so that message losses due to collision are reduced. We reduce the problem of computing a TDMA schedule that avoids hidden terminals to computing a distance-2 (henceforth D-2) vertex coloring; the color of a node decides the slot in which it can transmit. In the unit disk graph model, these two problems are equivalent [6]: this intuitively follows from the observation if two non-neighboring nodes $u$ and $v$ interfere with each other then there exists a node $w$ such that there are edges $(u,w)$ and $(w,v)$ in the unit disk graph. Second, Harvest uses acknowledgments and retransmissions at the MAC layer for recover from losses Harvest achieves its energy efficiency by avoiding energy-intensive flash operations: it performs at most one flash read and no flash write for any bulk datum piece, and stores the piece at nodes en route to the base station only in their RAM. Of course, avoiding message collision also yields energy efficiency. Harvest keeps the message overhead low (below 9 bytes per bulk data packet). To keep the number of control message tranmissions low, we present a distributed algorithm for its TDMA schedule computation that generates O(1) control message per node involved, which is significantly better than the traditional, distributed alternative, which incurs O$(\Delta^{2})$ per node involved, where $\Delta$ is the node degree. As we explain shortly, this improvement is enabled by computing vertex coloring with a constant number of colors that may be smaller than $\Delta$. As a result, not every node in the network is colored and so the TDMA schedule has to be computed in an ongoing manner, which in turn implies the control message savings is an ongoing one. Moreover, since an idle radio also consumes significant power (of the same order as that during message reception), Harvest schedules the switching off of the radio to save energy: asymptotically, a node keeps its radio on only for the time it is scheduled to send data on behalf of itself or someother node. Finally, Harvest achieves low latency in two ways. For one, it exploits spatial reuse. Instead of collecting data from only one node at a time, Harvest allows data collection from some constant number of nodes concurrently. For ease of exposition only, we let the concurrency constant be 2 henceforth (the protocol assumes that the user will specify this concurrency constant as a parameter; in fact, larger constants for dense networks yield lower latency). For nodes (including the base station) to concurrently receive data from 2 nodes, 4 colors are needed in the vertex coloring (one for the node, two for its allowed children, and one for its parent node). In other words, regardless of the density of the network, Harvest colors at most 4 nodes in the inteference region of any node at any time. Once a colored node has completed its transmissions, Harvest lets an uncolored node in its interference region to assume that color; this is the on-going aspect of the TDMA schedule computation. We validate that a concurrency constant of 2 yields 33% less latency for large networks —specifically networks larger than 3 hops— than the approach that collects data from one node at a time; the latter approach is adopted by the Straw protocol. This latency improvement is obtained even if we restrict Harvest to allow concurrent reception at only the base station and reception from at most one child node at every other colored node. And two, we ensure that the control algorithms used by Harvest have constant time complexities. In particular, our TDMA slot synchronization algorithm (which obviates the need for a time synchronization service) has a local convergence time of O$(1)$ as opposed to O$(\Delta^{2})+$O$(D)$ in traditional algorithms, given the bound on the node density. Contributions of the paper. 1. 1. We present a randomized distributed algorithm that assigns a constant number of colors in the D-2 region of every node (if such a coloring is possible) so that each node that gets a color waits O$(1)$ time until it gets a unique color; this contrasts with the O$(\Delta^{2})$ wait time of traditional deterministic D-2 coloring algorithms. This is achieved by executing our TDMA scheduling algorithm on top of a CSMA/CA-based MAC; thus, in the event that two nodes within reliable range of each other will start to contend on a color simulataneously (i.e. in the same slot), then due to CSMA/CA, one of the two nodes will back off and receive a packet from the other contending node, therby yielding that color. We note that the algorithm would work given (local or nonlocal) ways of calculating the interference regions of nodes other than our method for local computation of the D-2 neighborhood set. 2. 2. We present an algorithm that synchronizes its TDMA slots with that other colored nodes within O$(1)$ time of its being colored. Traditional deterministic TDMA slot synchronization algorithms have O$(Dia)$ convergence time, where $Dia$ is the diameter of the network. Intuitively, the reason for the O$(Dia)$ convergence time is that the maximum number of colors used the nodes in any interference region needs to be propagated to all the nodes after the D-2 coloring so that they can agree on the transmission period. Harvest achieves constant time converence because of its use of a pre-specified constant number of colors. 3. 3. We present a data collection algorithm that can use a small constant number (even 1) of packet-sized buffers, irrespective of the number of nodes and the bulk datum sizes. In effect, each node in Harvest can have at the most two children node that send data to it. Using 1 buffer let’s the node forward on behalf of only one child; using two let’s the node forward on behalf of both; and using more buffers helps in further reducing the latency. 4. 4. We evaluate the performance benefit of the spatial reuse and TDMA scheduling in Harvest, by providing a comparative performance with the Straw protocol, in terms of latency and energy improvement achieved via the former and the relative message overheads (Straw use 7 bytes per packet versus Harvest’s 9). Organization of the paper. In Section 2, we present the system model and problem statement in more detail. We describe the Harvest protocol and its TinyOS implementation in Section 3. We analyse the performance of the randomized TDMA scheduling algorithm and the data collection components of Harvest in Section 4. In Section 5, we overview the Straw protocol and compare its performance with that of Harvest, analytically and via TOSSIM simulations. We describe a number of extensions of Harvest and discuss its relevance for collecting on-line streaming data in Section 6. We discuss related work in Section 7 and make concluding remarks in Section 8. ## 2 The Bulk Data Convergecast Problem The system consists of $n,n>0,$ wireless sensor nodes, called motes, one of which is distinguished as a base station. We do not make assumptions of how the mote locations are spatially distributed, nor do we assume the availability of a location service. We do assume that each mote can communicate with the base station over zero or more communication hops and the degree $\Delta$ in the network is bounded. Some of the motes initially each have a bulk datum in their data store, whose size may vary from mote to mote and may exceed that of the mote’s RAM. For simplicity, we assume that each bulk datum resides in the nonvolatile store of its mote. Desired is a middleware service that upon initiation from the base station collects of all motes’ bulk data at the base station. In decreasing order of importance, the performance metrics of the service are: first, energy efficiency and reliability, and, second, latency, by which we mean the time taken at the base station from the initiation to the completion of the bulk data collection opreation. With respect to energy, Table 1 illustrates the energy cost in terms of current draw for common operations for the case of motes in the Mica-2/XSM family. Operation | Current Draw ---|--- CPU and Idle Radio | 8 mA Packet Reception | 7.03 mA Packet Transmission | 10.4 mA EEPROM Read | 6.2 mA EEPROM Write | 18.4 mA Table 1: Energy required by common operations The table suggests that given the aggregate current draw for an EEPROM read and write (24.6mA) is significant, and thus every addition of flash operations to the bulk convergecast forwarding process will almost double the current draw associated with the minimum aggregate current draw of radio receive, CPU, and radio transmission (25.43mA). It follows that minimizing EEPROM operations is desirable for the energy metric. The table also identifies the energy overhead associated with an idle radio. One implication is that a mote should sleep as soon as it has no data to send of its own or on behalf of other motes. With respect to reliability, we focus attention on obtaining high but not necessarily 100% reliable data collection. Unlike the dual problem of bulk data dissemination, where objectives such as mote reprogramming demand all-or- nothing delivery of bulk data, the use cases of bulk data convergecast can often tolerate low levels of unreliability. In designing our solution, we do not emphasize a particular selection of a link estimation technique or retrasmission mechanism. (Specifically, our experimental evaluation of our solution uses the WMEWMA link estimation approach of Woo and Culler [16] and 0-retransmissions, but these choices are not of central importance.) With respect to latency, we note that the problem statement does not emphasize the latency of collection from the perspective of individual motes. Had we considered the version of the problem where motes were continually generate and stream data to the base station, low jitter and comparable latency across the motes would have been desirable. We therefore regard these latter requirements as being optional, but not first order, for solving the problem. Finally, in designing our solution, we do not assume the availability of a time synchronization service. ## 3 The Harvest Protocol ### 3.1 The Components of Harvest In this section, we describe the three components of Harvest, viz. interference neighborhood discovery, randomized slot assignment and synchronization, and data collection. Neighborhood Discovery. Each node performs online link estimation to find out its 1-hop neighborhood set. For ease of exposition, we first assume that all the links are symmetric, i.e., the link quality between two nodes is same in both the directions. Therefore, it is sufficient to do link estimation in any one direction. However, the links in sensor network may not always be symmetric, so we will extend the link estimation in both the directions to deal with asymmetric links. A number of metrics can be used for this link estimation; for instance we may use the window mean with exponentially weighted moving average (WMEWMA) metric. This metric was has been used by MintRoute protocol [16] of Woo and Culler. There are two tuning parameters for WMEWMA-based link estimations, viz. $\alpha$ and $t$. The parameter $\alpha$ determines the size of the history used in link estimation and $t$ determines the rate at which link estimation is updated.(Experimental results in the literature show that the values 0.6 and 30 for $\alpha$ and $t$ respectively, provide stable and agile link estimation for Chipcon’s CC1000 radio. In particular, the settling time, which is the length of time for the estimator to converge within $\pm 10\%$ of the actual value and remain within the error bound.) Based on link estimation, path selection to the base station can be based again on a number of metrics studied in the literature, e.g., end-to-end path reliability, hop distance, end-to-end mac latency, etc;. for instance, we may use a combination of link quality and hop distance. In particular, we define the 1-hop neighborhood of a node $A$ to be the set of nodes that have WMEWMA value greater than or equal to 75 (which roughly implies a stable packet loss rate less than 10%). Among the 1-hop neighbors, node $A$ selects a node with the least hop distance to the base station as its parent. Using TOSSIM simulations, we find that the minimum WMEWMA link quality between two nodes at 2-hops from each other is 30 (which roughly implies the nodes can reliably sense each other’s carrier). Randomized Slot Assignment and Synchronization. As explained in Section 1, Harvest uses 4 colors in the entire network (i.e., two more colors than our de facto concurrency constant of 2). The TDMA scheduling divides time into intervals of length $T=4*t_{S}$, where $t_{S}$ is the duration of a timeslot. Note that the color assignment should be such two nodes with that are not within 2 hops of each other can use same color. Further, every node can have at the most two children. In each time period $T$, a node can forward only one packet (this could be its own packets or on behalf of one of its two children). The parent can signal which child should send a packet next by ordering the child IDs in the Harvest message, as we shall explain later. To begin the TDMA scheduling, the base station selects a color for itself and starts sending beacon messages in its timeslot. The 1-hop neighbors of base station randomly select an available color and start sending their payload. Every node’s packet contains the IDs of its 1-hop neighborhood transmitters. If node A hears its 1-hop neighbor transmitting in the same time-slot then one of two nodes gives up its color. The priority among the contending nodes is decided by considering which nodes was the first to select the color and then by the IDs of the nodes. Thus, priority is locally computed by looking at the sequence number of the messages and the unique IDs of the nodes. The underlying MAC layer in Harvest is CSMA/CA based. As a result, even if two nodes try to transmit in the same time slot, only one of them can succeed. We claim that this phenomenon applies to both scenarios, viz. when the two contending nodes are 1-hop neighbors of each other or 2-hops neighbors of each other. The nodes in the 1-hop neighborhood of the base station select a color and the wave propagates. After a node selects a color for itself or finds that there are no available colors in its D-2 neighborhood, it turns off the backoffs in the underlying CSMA/CA. Every node maintains a list of node IDs, which are using the 4 colors in its D-2 neighborhood, as a soft state. Whenever a node finishes its data transmission, it stops transmitting and its color is available for reuse. All the nodes in the D-2 neighborhood learn this information by the virtue of the soft state and enable backoff in the underlying CSMA/CA. And the process of randomized color selection repeats. The selection of 2 senders and the color assignment is a distributed operation. The operation is initiated from a single base station as opposed to the nodes in the network, since uniquely selecting nodes in a distributed manner would incur additional message overhead for coordination. The approach outperforms a centralized solution because in the latter a single node (such as the base station) would need to collect the entire topology information to compute disjoint paths between two nodes. Further, the operation would have to repeated whenever a new sender is selected. Data Forwarding Protocol. As soon as a node has selected its parent and uniquely selected a color in its D-2 neighborhood, it starts sending data packets to its parent in the corresponding timeslot. A parent can choose to receive packets from either of its children. If a parent node has 1 buffer space, then it can receive only 1 packet in the entire time period $T=4*t_{S}$ ($t_{S}$ is long enough to transmit a Harvest message with CSMA/CA backoffs disabled). In this csae, a parent node receives packets from a child in every alternate time period. The process of alternation ensures that the colors assigned to its children are not unassigned. This is a minor variation in the Harvest protocol as described above. Instead of one packet buffer at each node, more than one packet buffers can be allocated at each node. This will expedite the data collection. ### 3.2 Implementation Description In this section, we describe the implementation of Harvest in NesC under TinyOS 1.x release. Harvest has a single message structure; Figure 1 illustrates this structure, using numbers that denote field sizes in terms of bits. Figure 1: Harvest Message Structure The payload in each Harvest packet is 20 bytes (this is in contrast to Straw’s payload of 22 bytes). The color ID identifies one of the 4 colors used by the node. The # hops denotes the distance of the sender from the base station. This information is used by a node to select it parent, which has the minimum distance to the base station, among the set of nodes within 1-hop neighborhood. The child IDs are used identify the IDs of the sender’s children. A non-null value declares that the sender is available for forwarding. Also, the sender can use this field to declare its decision about the selected children in case multiple node are in contention for the selection. In particular, the ID in the first field among the two, should send packet in the next time period. The array of 4 node IDs denotes the IDs of the 4 nodes, in the sender’s D-2 neighborhood, which are currently using the 4 colors. The array is ordered in the increasing order of the color IDs. Every node copies the array received from its 1-hop neighborhood and maintains it as a soft state. If a color is not refreshed for a certain time, then the node assumes that the color is free and sends this information as part of its messages. The sequence number is a monotonically increasing number and denotes the sequence number of the packet. It is used in the calculation of WMEWMA link estimate and to select a unique node in the case of 2 nodes contending for the same color or same parent. The range of sequence numbers can be chosen depending upon the number of packets to be transmitted and also the number can be recycled to save space. There are no explicit sender ID and the destination ID fields. The sender ID can be retrieved from the message by reading the node ID at the location of sender’s color ID. The destination ID field is used from the TOS header in the TinyOS packet. Harvest uses promiscuous mode of transmission so that neighboring nodes can learn about the color allocation. But only the node identified as the destination node forwards the packet to the base station. The base station’s ID is 0. The receiver can identify whether the message is from the base station or not by looking at the sender ID. Harvest does not need an explicit tine synchronization service for its TDMA to function. Every packet contains the D-2 color of the sender. We use the synchronous reception property of the wireless medium to achieve time synchronization among the nodes [2]. In particular, when a node hears a packet from its parent, it synchronizes its time with that of its parent. Since base station is the root of the tree, all the time at all the nodes is synchronized to that of the single clock of the base station by virtue of induction. This synchronization scheme is also used in Sprinkler [8], which uses TDMA. ## 4 Performance Evaluation In this section, we evaluate the latency and number of packet transmissions for Harvest’s randomized slot assignment algorithm and data collection protocol. ### 4.1 Randomized Slot Assignment Performance As described earlier, Harvest uses an underlying CSMA/CA protocol for color selection. In particular, when a node receives a message, it finds out the available colors in its D-2 neighborhood from the received message. If one or more colors are available, the node randomly selects an available color and starts transmitting from the corresponding TDMA slot in next time interval. It is possible that two or more nodes can simultaneously select the same color and therefore their transmitted packets can collide with each other. We show here that in O$(1)$ time, a unique node will select a unique color in the node’s D-2 neighborhood. TinyOS uses a variant of non-persistent CSMS/CA protocol [15]. We briefly recall the definition of non-persistent CSMA/CA protocol [1]: 1. 1. A node senses channel before transmission. 2. 2. If the channel is free, it immediately transmits a frame; otherwise it waits for a random amount of time. 3. 3. After waiting, it repeats step 1. In the case of TinyOS, a node waits for a random amount of time before it senses the channel. This ensures that the transmissions are not synchronized. Because of the initial random wait, the throughput of the CSMA/CA in TinyOS is better than that of the classical non-persistent CSMA/CA. ###### Theorem 1 Given that the degree ($\Delta$) of network is bounded, Harvest takes O$(1)$ time for assigning a unique color to a node in its D-2 neighborhood. ###### Proof 4.2. Given that $\Delta$ is bounded, for non-persistent CSMA/CA, there exists a constant $\tau>0$ such that the probability of a frame transmission without collision is at least $\tau$ [1]. The same holds for the CSMA/CA in TinyOS which is variant of non-persistent CSMA/CA. Therefore, the expected time for a frame transmission without failure is also O$(1)$. In the event that two or more nodes select the same color and transmit in the same timeslot, in O$(1)$ time, a unique node will succeed in a transmission without failure. After one transmission without collision, all the nodes in the D-1 neighborhood will learn that the color is not available. Similarly, for the D-2 neighborhood, a unique color is selected in O$(1)$ time since the packet delivery rate between D-2 neighbors is non-zero. After one transmission without collision by the successful node, or via a neighbor of the successful node, the color assignment of the successful node get conveyed to its 2-hop neighborhood. The value of $\tau$ depends upon the range of values for random wait and $\Delta$. Instead of finding the value of $\tau$, we perform experiments to measure the convergence time of Harvest’s slot assignment for different values of $\Delta$. We use 51 XSM motes in an indoor testbed, Kansei. An XSM mote uses Chipcon’s CC1000 radio and is for the purposes of this experiment similar to a Mica-2 mote. We use the TinyOS 1.x release and the standard MAC that comes with the 1.x release. The topology of the network is as shown in the Figure 2. The motes are placed in grid with 3ft unit separation on the X and Y axes. We uses default power level and default frequency for transmission. The mote at location (0,0) is selected as the base station, as shown in Figure 2. Each slot is of the duration 31 msec, which is the minimum possible given that the radio transmission takes at 23 msec and the UART transmission takes at least 8 msec in TinyOS 1.x over XSM. Each node has a payload of 100 packets to be sent to the base station. Figure 2: Testbed Topolgy We measure the time required to collect all the data packets from the first mote after the start of the experiment as a function of the number of nodes. The time is sampled at a granularity of 30 times the time for a transmission period. The number of sampling periods required for the first node to complete data exfiltration denotes the convergence time of the color selection algorithm. As shown in Table 2, the convergence time has a variance of 1 sampling period, which is negligible. Hence, for the non-persistent CSMA/CA implementation in TinyOS 1.x release, the convergence time of Harvest’s randomized slot assignment is negligible for a $\Delta$ up to 51. # nodes | Convergence time ---|--- 6 | 8 12 | 8 18 | 8 22 | 8 31 | 7 42 | 8 51 | 9 Table 2: Scalability of color selection ### 4.2 Data Collection Protocol Latency. We define the total latency of Harvest data collection to be the duration between the moment that the base station receives the first data packet and the moment it receives the last data packet. Since the base station has 2 children and there are 4 timeslots per time period $T,T=4*t_{s}$, it receives 2 packets per time $T$. Therefore, for $n$ nodes and $M$ number of packets from each node, the time required to receive $n*M$ packets is $n*M*2*t_{s}$, which is O$(n*M)$. The time required to build the tree rooted at the base station is in the worst case O$(n)$. Note that the tree building is happening in parallel to the data collection. But for the worst case analysis, we can assume that the two processes happen sequentially. In that case, the total latency of Harvest is O$(n)$ \+ O$(n*M)$ = O$(n*M)$. Number of Transmissions. Let $h$ be the average height of a node in the routing tree. Therefore given $n*M$ packets, the total number of transmissions is O$((n*M*h)/2)$. ## 5 Performance Comparison ### 5.1 The Straw Protocol In this section, we compare the performance of Harvest with that of Straw [4]. Similar to Harvest, the objective of Straw protocol is to collect bulk data from all the nodes at the base station. Unlike Harvest, Straw collects data from one node at a time. For each node, the data collection is divided into two phases, viz. broadcast and collection. In case the collection phase loses packets, the two phases are repeated to recover from loss. (The broadcast command in the recovery phase contains the sequence numbers of the lost packets.) The overall goal of the protocol is to minimize latency and number of packet transmissions. The broadcast phase disseminates the ID of a selected node, from which data is to be collected. Following the broadcast phase, the selected node periodically sends packets to the base station. The route is selected using MintRoute protocol. For all nodes that are at a distance greater than 2-hops from the base station, the transmission period in Straw is $3*t_{h}$, where $t_{h}$ is the time required to traverse single hop. The number 3 is chosen to reduce the interference with data forwarding at an upstream node. If we color the nodes that transmit at the same time, then the coloring of transmitting nodes effectively yields a D-2 coloring. Note that the transmitting nodes induce a one dimensional graph, in other words, a single line (and hence the name “straw”). Due to the fact that each node sends packets at the period of $3*t_{h}$, the base station receives a packet after every $2*t_{h}$ time. For nodes at 1-hop and 2-hop distances from the base station, the transmission period is $t_{h}$ and $2*t_{h}$ respectively. The initial broadcast command sets up the colors on the linear path from a node to the base station. This corresponds to a deterministic slot assignment, as compared to the randomized slot assignment of Harvest. Further, a node from which data is collected, is selected by the base station as opposed to the local, distributed selection in Harvest. ### 5.2 Latency Comparison #### 5.2.1 Theoretical Comparison Straw uses a broadcast for slot assignment. In each broadcast phase, a node forwards the broadcast command once. For collecting data from $n$ nodes, Straw will therefore employ $n$ broadcast sessions, on average lasting for at least O($h$) time. Therefore, the total latency for assigning slots is O$(n*h)$ as compared to O$(n)$ for Harvest. In Straw’s data collection protocol, only the nodes on the path from the current sender to the base station are transmitting. The rest of the network is idle, in other words, spatial reuse is limited. If the rest of the network lies outside interference distance from the the transmitters, then an idle node from the rest of the network can send its data towards the base station. However, finding a nodes outside interference distance from the current transmitters could be impossible, especially near the base station. A solution is to increase the number of D-2 colors from 3. Instead of a linear structure, Harvest utilizes a tree structure to collect data packets. Given the concurrency constant of 2, Harvest uses a binary tree. Harvest uses 4 colors in order to ensure that the binary tree can be D-2 colored. In that case, the base station receives 2 packets every $4*t_{s}$, where $t_{s}$ is the duration of a timeslot. Therefore the rate of data collection at the base station is equal to a packet after every $t_{s}$ time. Note that we can utilize any m-ary tree and $C$ colors, and the resulting rate of data collection at the base station would be $m/C$. In Straw, the rate of data collection from the nodes at more than 2-hops from the base station is 1 packet per $3*t_{s}$. If we assume that the number of nodes at 1-hop and 2-hop distance from the base station is far less than that the total number of nodes $n$, the latency of data collection for Straw is $n*M*3*t_{s}$. Therefore, data collection of Harvest has 33.33% lower latency than that of Straw. Further, the overall order complexity of the latency of Straw is O$(n*h)$ \+ O$(n*M)$, which exceeds the O$(n*M)$ of harvest if O$(h)$ is greater than O$(M)$. #### 5.2.2 Simulation-based Comparison Figure 3: Network topology for simulation To validate the claimed improvement in latency, we perform simulations in TOSSIM. We setup a network of 20 non-base station motes and 1 base station node. As shown in Figure 3, we ensure that there are nodes at more than 2-hop distance from the base station. Also, the base station has more than 1 node at 1-hop distance. The QueuedSend buffer module at the TinyOS’s MAC layer uses explicit acknowledgment. In the case of unsuccessful transmission, a retransmission is attempted. However, the retransmission could happen in an incorrect timeslot, resulting in a collision. Therefore, we have disabled the MAC layer ACK in this simulation. However, we can still utilize the MAC-level ACK by channeling the ACK information to the Straw and Harvest protocol layer. Use of ACK will increase the reliability to data collection. All of the simulation are done under TOSSIM. This NesC-code simulator has an option to instantiate a link quality set given the node placement. The links qualities are based on some empirical measurements carried out for MICA-2 motes. The links qualities vary in spatial and temporal dimensions. Since the current implementation of Harvest assumes symmetric links and base its parent selection criteria by measuring link quality in one direction, we have pre- processed the link quality set so that all the links are symmetric. In our future work, we will refine the implementation to deal with asymmetric links by computing link quality in both directions. In particular, link quality from the child to parent will be computed by counting the number of successful and failed ACKs. We use our NesC implementation and we use the Straw code which has been available as part of a Golden Gate Bridge health monitoring project contribution folder under TinyOS 1.x release. We measure the rate at which data is collected at the base station. We observe that the rate of data collection is $1.67$ packets per $4*t_{s}$ for Harvest and $0.8$ packets per $3*t_{s}$ for Straw, as shown in Table 3. The rate of data collection is lower than the respective theoretical values due to the fact that ACKs are disabled. The observed latency gain under simulation is $36\%$, which is close to the theoretical value of $33.33\%$. Service | Theoretical | Simulation ---|---|--- Straw | 1 packets/$(3*t_{s})$ | 0.8 packets/$(3*t_{s})$ Harvest | 2 packets/$(4*t_{s})$ | 1.67 packets/$(3*t_{s})$ Table 3: Latency Comparison ### 5.3 Energy Comparison #### 5.3.1 Theoretical Comparison Straw uses a broadcast to disseminate the command to send the ID of a selected node. This is equivalent to the slot assignment in Harvest. In a broadcast phase, each node in the network forwards a newly heard packet exactly once. Therefore, the total number of transmissions in a broadcast phase are $n$, where $n$ is the total number of nodes in the network. Therefore, to collect data from $n$ nodes, the total number of packet transmissions are $n^{2}$. Also, each broadcast phase is followed by a reply from a selected node to the base station. The total number of transmissions, for each reply, is a function of number of hops from the selected node to the base station. In the worst case, the average path length in the network could be $n/2$. In that case, the total number of replies for all nodes is $n^{2}+n/2$, which is O$(n^{2})$. In Harvest, the control information pertaining to slot assignment is piggybacked on the data messages. Therefore, Harvest does not have packet transmissions for slot assignment. Hence, it saves O$(n^{2})$ number of packet transmissions as compared to Straw. We assume that Straw and Harvest both use the shortest path routes to transmit data packets to the base station. In that case, the total number of data packet transmissions for data collection purposes are the same for Straw and Harvest. In particular, this number is O$((n*M*h)/2)$. The total number of messages for Straw is O$(n^{2})$ \+ O$((n*M*h/2)$. #### 5.3.2 Simulation-based Comparison In reality, the radio behavior is more complex than that represented by the simplistic unit-disk radio model. Not only is packet delivery rate less than 100% but it also varies in space and time. Therefore, we conduct simulations over a multi-hop network to compare the number of packet transmissions for Straw and Harvest. We conduct simulations in the same network topology as used in Section 5.2.2. In future, we plan to compare results in a real sensor network. The number on top of each node, in Figure 4, illustrates the number of broadcast sessions required to reliably convey the command to each of the 20 nodes. The total number of broadcast sessions are 46. Given 20 nodes, Straw consumes 20 times 46, i.e. 920, more packet transmissions. Figure 4: Number of broadcast sessions for 21 nodes in Straw ## 6 Harvest Extensions and Discussion Duty Cycling of Radios. As we discussed in Section 2, an idle radio draws a significant amount of current and so energy efficiency is gained by letting idle nodes sleep. In Harvest, we achieve this as follows. When a node sees that no colors are available for itself in its interference region (i.e., its 2-hop neighborhood), it can switch off its radio until a color is expected to be available again. Given some knowledge of the number of packets to be transmitted that color and by observing the sequence number of the packet currently being transmitted for that color, a sleeping duration can be readily calculated. Furthermore, once a node is done with its role in the convergecast, it can switch off its radio permanently. A node is defined to be done with its transmissions after it has sent all of its packets and the packets of its children. Reliability when all Children Transmit Concurrently. As described in Section 3, the data collection protocol allows non-base station nodes to forward data from one or more of its children. (If more than one child can transmit, then the protocols maintains at least one buffer per child.) When only one child is allowed to transmit, the implicit acknowledgement scheme suffices for nodes to discover whether or not their transmissions were successfully received. When more than one child is allowed to transmit, using the implicit acknowledgement scheme implies either a delay in loss detection or a modification of the protocol to expose more node buffer information. One alternative in this case would be to use explicit acknowledgements. If we assume that explicit acknowledgements can be send immediately (or within some constant delay after message reception), then the tranmission slots can be extended to subsume both the transmission time and the acknowledgement time. Continuous Streaming of Data to the Base Station. Harvest collects data simultaneously from multiple nodes, as opposed to receiving data from only one node at a time. In this sense, the data received at the base station resembles a continuous stream of data from the network ordered in time. It is therefore conceivable to use Harvest as the basis for collecting in an on-line fashion continuous data streams from the network. Note that the description in Section 3 for the case where all nodes forward data from multiple children allows the possibility that the data from each child is forwarded in a round robin. More sophisticated rules for fair scheduling that consider the distance of the node from the base station can be defined to achieve global fairness, otherwise nodes near base station will contribute more packets as opposed to the ones farther from the base station. One extension of Harvest that we are presently studying is in the context of real-time wireless sensor network applications, such as visualizing link quality of the network in real time or viewing consistent global snapshots of the wireless sensor network. ## 7 Related Work TDMA and CSMA. Herman et al [3] have proposed a randomized TDMA algorithm that first forms clusters, each with a unique cluster-head. Each cluster-head then allocates colors to its children. Cluster-heads are ordered in a monotonically increasing order, so the color assignment occurs sequentially per that order. A similarity between this work and Harvest’s TDMA scheduling is the use of underlying CSMA/CA MAC layer to CSMA/CA to communicate control information pertaining to node coloring and TDMA scheduling. Z-MAC [9] uses both TDMA and CSMA/CA features in manner different from Harvest. Z-MAC is hybrid MAC that uses TDMA under high contention and CSMA under low contention, whereas data transmission in Harvest is always in TDMA mode. Kulkarni and Arumugam [7] describe TDMA based protocols that are optimized for convergecast, that work however assumes grid localization. RID [18] is a radio interference detection service that detects interference relations between nodes at run-time, and provides higher fidelity for collision avoidance when using TDMA. The RID approach would be a suitable candidate for enabling Harvest’s distributed coloring protocol. Convergecast routing. There is a rich body of work on convergecast routing for wireless sensor systems. Several protocols assume location information. Most of the others such as MintRoute [16], RMST [11], PSFQ [13], Drain [12] are, unlike Straw and Harvest, not optimized for the energy and latency requirements associated with the collection of payloads that can well be in the thousands of packets per node. For instance, MintRoute does not pipeline transmissions, which would yield higher latency for bulk data transport, and Drain is optimized for the case of a single packet payload per source mote. Reliability. The study of reliability in convergecast has often arisen in the context of concurrent event detections, which tend to occur in a bursty manner or with multiple sources are continuously/periodically generating packets (with low duty cycle). RBC [17] focuses on the former whereas the traffic models considered in CODA [14] and ESRT [10] focus on the latter. RBC deals with bursts by maintaining information about queue conditions of the neighboring node as well as number of times enqueued packets were retransmitted, which results in sizeable RAM usage. Also, the queue condition has to be transmitted in RBC, which results in sizeable communication overhead. The alternative approaches of packet retransmissions, of acknowledgements, of hop-by-hop recovery, as well as selecting alternative routes upon link failure are also relevant approaches for improving the reliability of Harvest in particular application contexts. The use of TDMA and receiver-driven flow control mitigate the consideration of congestion. Coding of bulk data is a relevant approach for tolerating packet loss in bulk convergecast. Kim et al [5] have considered the use of erasure codes. We have regarded this relevant consideration as being orthogonal to the pipelining and spatial reuse considerations of Harvest. ## 8 Conclusion We have presented a bulk data collection service, Harvest, for energy constrained wireless sensor nodes. Harvest assumes a bounded node density, i.e., degree $\Delta$. This assumption enable us to We assign distance-k (our exposition has used k=2) colors to nodes in O$(1)$ time by utilizing an underlying CSMA/CA MAC layer. We use a constant number of colors in the entire network, which enables the per node computation of its TDMA schedule to occur in O$(1)$ time. Harvest exploits the spatial parallelism in collecting data, thereby achieving a latency gain of at least 33% in large networks (i.e., networks with more than three hops) as compared to that of Straw. Harvest also avoids the O$(n^{2})$ number of broadcasted control transmissions used in Straw. Further, Harvest requires only O$(1)$ number of buffers at each node. Therefore, Harvest is suitable for large scale network of wireless sensor network. We provide theoretical bounds on the performance of Harvest and perform simulation results to validate the theoretical bound. We find that the spatial parallelism not only reduces latency, but also creates an opportunity to collect global data in a fair real-time manner. Our present work is studying extensions of Harvest for the case of on-line continous data streaming from the network to the base station. ## References * [1] D. Bertsekas and R. Gallager. Data Networks. Prentice Hall, Englewood Cliffs, NJ, 1987. * [2] J. Elson. Time Synchronization in Wireless Sensor Network. PhD thesis, UCLA, 2003. * [3] T. Herman and S. Tixeuil. A distributed tdma slot assignment algorithm for wireless sensor networks. In Algorithmic Aspects of Wireless Sensor Networks, pages 45–58, 2004. * [4] S. Kim. Wireless sensor networks for structural health monitoring. Master’s thesis, University of California at Berkeley, USA, 2005. * [5] S. Kim, R. Fonseca, and D. Culler. Reliable transfer on wireless sensor networks. In Annual IEEE Communications Society Conference on Sensor and Ad Hoc Communications and Networks, 2004. * [6] S. Krumke, M. Marathe, and S. Ravi. Models and approximation algorithms for channel assignment in radio networks. Wireless Networks, 7(6):575–584, 2001. * [7] S. Kulkarni and M. Arumugam. SS-TDMA: A Self-Stabilizing MAC for Sensor Networks, chapter In Sensor Network Operations. IEEE Press, 2005. * [8] V. Naik, A. Arora, P. Sinha, and H. Zhang. Sprinkler: A reliable and energy efficient data dissemination service for wireless embedded devices. In The 26th IEEE Real-Time Systems Symposium, December 2005. * [9] I. Rhee, A. Warrier, M. Aia, and J. Min. Z-mac: A hybrid mac for wireless sensor networks. In SenSys, pages 90–101, 2005. * [10] Y. Sankarasubramaniam, O. Akan, and I. Akyildiz. Esrt: Event-to-sink reliable transport in wireless sensor networks. In The ACM Symposium on Mobile Ad Hoc Networking and Computing (MobiHoc), 2003. * [11] F. Stann and J. Heidemann. Rmst: Reliable data transport in sensor networks. In The 1st IEEE Intl. Workshop on Sensor Network Protocols and Applications (SNPA), pages 102–112, 2003. * [12] G. Tolle and D. Culler. Design of an application-cooperative management system for wireless sensor networks. In Second European Workshop on Wireless Sensor Networks, 2005. * [13] C. Wan, A. Campbell, and L. Krishnamurthy. Psfq: A reliable transport protocol for wireless sensor networks. In WSNA ’02: Proceedings of the 1st ACM International Workshop on Wireless Sensor Networks and Applications, pages 1–11, New York, NY, USA, 2002. ACM Press. * [14] C. Wan, S. Eisenman, and A. Campbell. Coda: congestion detection and avoidance in sensor networks. In SenSys, pages 266–279, 2003. * [15] A. Woo and D. Culler. A transmission control scheme for media access in sensor networks. In ACM/IEEE International Conference on Mobile Computing and Networking (MobiCom), pages 221–235, 2001. * [16] A. Woo, T. Tong, and D. Culler. Taming the underlying challenges of reliable multihop routing in sensor networks. In SenSys ’03: Proceedings of the 1st international conference on Embedded networked sensor systems, pages 14–27, 2003. * [17] H. Zhang, A. Arora, Y. Choi, and M. Gouda. Reliable bursty convergecast in wireless sensor networks. In 6th ACM International Symposium on Mobile Ad Hoc Networking and Computing, 2005. * [18] G. Zhou, T. He, J. Stankovic, and T. Abdelzaher. Rid: radio interference detection in wireless sensor networks. In The 24th Annual Joint Conference of the IEEE Computer and Communications Societies (INFOCOM), pages 891– 901, 2005.
# Fine-Grained Named Entity Typing over Distantly Supervised Data via Refinement in Hyperbolic Space Muhammad Asif Ali,1 Yifang Sun,1 Bing Li,1 Wei Wang,1,2 1School of Computer Science and Engineering, UNSW, Australia 2College of Computer Science and Technology, DGUT, China <EMAIL_ADDRESS><EMAIL_ADDRESS> ###### Abstract Fine-Grained Named Entity Typing (FG-NET) aims at classifying the entity mentions into a wide range of entity types (usually hundreds) depending upon the context. While distant supervision is the most common way to acquire supervised training data, it brings in label noise, as it assigns type labels to the entity mentions irrespective of mentions’ context. In attempts to deal with the label noise, leading research on the FG-NET assumes that the fine- grained entity typing data possesses a euclidean nature, which restraints the ability of the existing models in combating the label noise. Given the fact that the fine-grained type hierarchy exhibits a hierarchal structure, it makes hyperbolic space a natural choice to model the FG-NET data. In this research, we propose FGNET-RH, a novel framework that benefits from the hyperbolic geometry in combination with the graph structures to perform entity typing in a performance-enhanced fashion. FGNET-RH initially uses LSTM networks to encode the mention in relation with its context, later it forms a graph to distill/refine the mention’s encodings in the hyperbolic space. Finally, the refined mention encoding is used for entity typing. Experimentation using different benchmark datasets shows that FGNET-RH improves the performance on FG-NET by up to 3.5% in terms of strict accuracy. Keywords — FG-NET, Hyperbolic Geometry, Distant Supervision, Graph Convolution ## 1 Introduction Named Entity Typing (NET) is a fundamental operation in natural language processing, it aims at assigning discrete type labels to the entity mentions in the text. It has immense applications, including: knowledge base construction [7]; information retrieval [12]; question answering [18]; relation extraction [27] etc. Traditional NET systems work with only a coarse set of type labels, e.g., organization, person, location, etc., which severely limit their potential in the down-streaming tasks. In recent past, the idea of NET is extended to Fine-Grained Named Entity Typing (FG-NET) that assigns a wide range of correlated entity types to the entity mentions [13]. Compared to NET, the FG-NET has shown a remarkable improvement in the sub-sequent applications. For example, Ling and Weld, [13] showed that FG-NET can boost the performance of the relation extraction by 93%. FG-NET encompasses hundreds of correlated entity types with little contextual differences, which makes it labour-intensive and error-prone to acquire manually labeled training data. Therefore, distant supervision is widely used to acquire training data for this task. Distant supervision relies on: (i) automated routines to detect the entity mention, and (ii) using type-hierarchy from existing knowledge-bases, e.g., Probase [24], to assign type labels to the entity mention. However, it assigns type-labels to the entity mention irrespective of the mention’s context, which results in label noise [20]. Examples in this regard are shown in Figure 1, where the distant supervision assigns labels: {person, author, president, actor, politician} to the entity mention: _“Donald Trump”_ , whereas, from contextual perspective, it should be labeled as: {person, president, politician} in S1, and {person, actor} in S2. Likewise, the entity mention: _“Vladimir Putin”_ should be labeled as: {person, author} and {person, athlete} in S3 and S4 respectively. This label noise in-turn propagates in the model learning and severely effects/limits the end-performance of the FG-NET systems. Figure 1: FG-NET training data acquired by distant supervision Earlier research on FG-NET either ignored the label noise [13], or applied some heuristics to prune the noisy labels [8]. Ren et al., [19] bifurcated the training data into clean and noisy data samples, and used different set of loss functions to model them. However, the modeling heuristics proposed by these models are not able to cope with the label noise, which limits the end- performance of the FG-NET systems relying on distant supervision. We, moreover, observe that these models are designed assuming a euclidean nature of the problem, which is inappropriate for FG-NET, as the fine-grained type hierarchy exhibit a hierarchical structure. Given that it is not possible to embed hierarchies in euclidean space [15], this assumption, in turn limits the ability of the existing models to: (i) effectively represent FG-NET data, (ii) cater label noise, and (iii) perform FG-NET classification task in a robust way. The inherent advantage of hyperbolic geometry to embed hierarchies is well- established in literature. It enforces the items on the top of the hierarchy to be placed close to the origin, and the items down in the hierarchy near infinity. This enables the embedding norm to cater to the depth in the hierarchy, and the distance between embeddings represent the similarity between the items. Thus the items sharing a parent node are close to each other in the embeddings space. This makes the hyperbolic space a perfect paradigm for embedding the distantly supervised FG-NET data, as it explicitly allows label-smoothing by sharing the contextual information across noisy entity mentions corresponding to the same type hierarchy, as shown in Figure 2 (b), for a 2D Poincaré Ball. For example, given the type hierarchy: _“Person”_ $\leftarrow$ _“Leader”_ $\leftarrow$ _“Politician”_ $\leftarrow$ _“President”_ , the hyperbolic embeddings, on contrary to the euclidean embeddings, offer a perfect geometry for the entity type _“President”_ to share and augments the context of _“Politician”_ , which in turn adds to the context of _“Leader”_ and _“Person”_ etc., shown in Figure 2 (a). We hypothesize that such hierarchically-organized contextually similar neighbours provide a robust platform for the end task, i.e., FG-NET over distantly supervised data, also discussed in detail in the section 4.5.1. Figure 2: (a) An illustration of how the entity type “President” shares the context of the entity type “Politician” which in turn shares the context of the entity-type “Leader” and so on; (b) Embedding FG-NET data in 2-D Poincare Ball, where each disjoint type may be embedded along a different direction Nevertheless, we propose Fine-Grained Entity Typing with Refinement in Hyperbolic space (FGNET-RH), shown in Figure 3. FGNET-RH follows a two-stage process, stage-I: encode the mention along with its context using multiple LSTM networks, stage-II: form a graph to refine mention’s encoding from stage-I by sharing contextual information in the hyperbolic space. In order to maximize the benefits of using the hyperbolic geometry in combination with the graph structure, FGNET-RH maps the mention encodings (from stage-I) to the hyperbolic space. And, performs all the operations: linear transformation, type-specific contextual aggregation etc., in the hyperbolic space, required for appropriate additive context-sharing along the type hierarchy to smoothen the noisy type-labels prior to the entity typing. The major contributions of FGNET-RH are enlisted as follows: 1. 1. FGNET-RH accommodates the benefits of: the graph structures and the hyperbolic geometry to perform fine-grained entity typing over distantly supervised noisy data in a robust fashion. 2. 2. FGNET-RH explicitly allows label-smoothing over the noisy training data by using graphs to combine the type-specific contextual information along the type-hierarchy in the hyperbolic space. 3. 3. Experimentation using two models of the hyperbolic space, i.e., the Hyperboloid and the Poincaré-Ball, shows that FGNET-RH outperforms the existing research by up to 3.5% in terms of strict accuracy. ## 2 Related Work Existing research on FG-NET can be bifurcated into two major categories: (i) traditional feature-based systems, and (ii) embedding models. Traditional feature-based systems rely on feature extraction, later using these features to train machine learning models for classification. Amongst them, Ling and Weld [13] developed FiGER, that uses hand-crafted features to develop a multi-label, multi-class perceptron classifier. Yosef et al., [29] developed HYENA, i.e., a hierarchical type classification model using hand- crafted features in combination with the SVM classifier. Gillick et al., [8] proposed context-dependent fine-grained typing using hand-crafted features along with logistic regression classifier. Shimaoka et al., [21] developed neural architecture for fine-grained entity typing using a combination of automated and hand-crafted features. Embedding models use widely available embedding resources with customized loss functions to form classification models. Yogatama et al., [28] used embeddings along with Weighted Approximate Rank Pairwise (WARP) loss. Ren et al., [19] proposed AFET that uses different set of loss functions to model the clean and the noisy entity mentions. Abhishek et al., [1] proposed end-to-end architecture to jointly embed the mention and the label embeddings. Xin et al., [25] used language models to compute the compatibility between the context and the entity type prior to entity typing. Choi et al., [4] proposed ultra-fine entity typing encompassing more than 10,000 entity types. They used crowd-sourced data along with the distantly supervised data for model training. Graph convolution networks are introduced in recent past in order to extend the concept of convolutions from regular-structured grids to graphs [11]. Ali et al., [2] proposed attentive convolutional network for fine-grained entity typing. Nickel et al., [15] illustrated the benefits of hyperbolic geometry for embedding the graph structured data. Chami et al., [3] combined graph convolutions with the hyperbolic geometry. López et al., [14] used hyperbolic geometry for ultra-fine entity typing. To the best of our knowledge, we are the first to explore the combined benefits of the graph convolution networks in relation with the hyperbolic geometry for FG-NET over distantly supervised noisy data. ## 3 Proposed Approach ### 3.1 Problem Definition In this paper, we build a multi-class, multi-label entity typing system using distantly supervised data to classify an entity mention into a set of fine- grained entity types. Specifically, we propose attentive type-specific contextual aggregation in the hyperbolic space to fine-tune the mention’s encodings learnt over noisy data prior to entity typing. We assume the availability of training corpus $C_{train}$ acquired via distant supervision, and manually labeled test corpus $C_{test}$. Each corpus $C$ (train/test) encompasses a set of sentences. For each sentence, the contextual token $\\{c_{i}\\}_{i=1}^{N}$, the mention spans $\\{m_{i}\\}_{i=1}^{N}$ (corresponding to the entity mentions), and the candidate type labels $\\{t_{i}\\}_{i=1}^{N}\in\\{0,1\\}^{T}$ ($T$-dimensional vector with $t_{i,x}=1$ if $x^{th}$ type corresponds to the true label and zero otherwise) have been priorly identified. The type labels are inferred from type hierarchy in the knowledge base $\psi$ with the schema $T_{\psi}$. Similar to Ren et al., [19], we bifurcate the training data $D_{tr}$ into clean $D_{tr\text{-}clean}$ and noisy $D_{tr\text{-}noisy}$, if the corresponding mention’s type-path follows a single path in the type-hierarchy $T_{\psi}$ or otherwise. Following the type-path in Figure 1 (ii), a mention with labels _{ person, author}_ will be considered as clean, whereas, a mention with labels _{ person, president, author}_ will be considered as noisy. ### 3.2 Overview Our proposed model, FGNET-RH, follows a two-step approach, labeled as stage-I and stage-II in the Figure 3. Stage-I follows text encoding pipeline to generate mention’s encoding in relation with its context. Stage-II is focused on label noise reduction, for this, we map the mention’s encoding (from stage-I) in the hyperbolic space and use a graph to share aggregated type- specific contextual information along the type-hierarchy in order to refine the mention encoding. Finally, the refined mention encoding is embedded along with the label encodings in the hyperbolic space for entity typing. Details of each stage are given in the following sub-sections. Figure 3: Proposed model, i.e., FGNET-RH, stage-I learns mention’s encodings based on local sentence-specific context, stage-II refines the encodings learnt in stage-I in the hyperbolic space. ### 3.3 Stage-I (Noisy Mention Encoding) Stage-I follows a standard text processing pipeline using multiple LSTM networks [9] to encode the entity mention in relation with its context. Individual components of stage-I are explained as follows: ##### Mention Encoding: We use LSTM network to encode the character sequence corresponding to the mention tokens. We use $\phi_{e}=[\overrightarrow{men}]\in\mathbf{R}^{e}$ to represent the encoded mention’s tokens. ##### Context Encoding: For context encoding, we use multiple Bi-LSTM networks to encode the tokens corresponding to the left and the right context of the entity mention. We use $\phi_{c_{l}}$ = $[\overleftarrow{c_{l}};\overrightarrow{c_{l}}]\in\mathbf{R}^{c}$ and $\phi_{c_{r}}$ = $[\overleftarrow{c_{r}};\overrightarrow{c_{r}}]\in\mathbf{R}^{c}$ to represent the encoded left and the right context respectively. ##### Position Encoding: For position encoding, we use LSTM network to encode the position of the left and the right contextual tokens. We use $\phi_{p_{l}}$ = $[\overleftarrow{l_{p}}]\in\mathbf{R}^{p}$ and ; $\phi_{p_{r}}=[\overrightarrow{r_{p}}]\in\mathbf{R}^{p}$ to represent the encoded position corresponding to the mention’s left and the right context. ##### Mention Encodings: Finally, we concatenate all the mention-specific encodings to get L-dimensional noisy encoding: $x_{m}\in\mathbf{R}^{L}$, where $L=e+2*c+2*p$. (3.1) $x_{m}=[\phi_{{p_{l}}};\phi_{c_{l}};\phi_{e};\phi_{c_{r}};\phi_{p_{r}}]$ ### 3.4 Stage-II (Fine-tuning the Mention Encodings) Stage-II is focused on alleviating the label noise. Underlying assumption in combating the label noise is that the contextually similar mentions should get similar type labels. For this, we form a graph to cluster contextually-similar mentions and employ hyperbolic geometry to share the contextual information along the type-hierarchy. As shown in Figure 3, the stage-II follows the following pipeline: 1. 1. Construct a graph such that contextually and semantically similar mentions end-up being the neighbors in the graph. 2. 2. Use exponential map to project the noisy mention encodings from stage-I to the hyperbolic space. 3. 3. In the hyperbolic space, use the corresponding exponential and logarithmic transformations to perform the core operations, i.e., (i) linear transformation, and (ii) contextual aggregation, required to fine-tune the encodings learnt in stage-I prior to entity typing. We work with two models in the hyperbolic space, i.e., the Hyperboloid $(\mathbb{H}^{d})$ and the Poincaré-Ball $(\mathbb{D}^{d})$. In the following sub-sections, we provide the mathematical formulation for the Hyperboloid model of the hyperbolic space. Similar formulation can be designed for the Poincaré-Ball model. #### 3.4.1 Hyperboloid Model $d$-dimensional hyperboloid model of the hyperbolic space (denoted by $\mathbb{H}^{d,K}$) is a space of constant negative curvature ${-1}/{K}$, with $\mathcal{T}_{\textbf{p}}\mathbb{H}^{d,K}$ as the euclidean tangent space at point p, such that: $\displaystyle\mathbb{H}^{d,K}$ $\displaystyle=\\{\textbf{p}\in\mathbb{R}^{d+1}:\langle\textbf{p},\textbf{p}\rangle=-K,p_{0}>0\\}$ (3.2) $\displaystyle\mathcal{T}_{\textbf{p}}\mathbb{H}^{d,K}$ $\displaystyle={\textbf{r}\in\mathbb{R}^{d+1}:\langle\textbf{r},\textbf{p}\rangle_{\mathcal{L}}=0}$ where $\langle,.,\rangle_{\mathcal{L}}:\mathbb{R}^{d+1}\times\mathbb{R}^{d+1}\rightarrow\mathbb{R}$ denotes the Minkowski inner product, with $\langle\textbf{p},\textbf{q}\rangle_{\mathcal{L}}=-p_{0}q_{0}+p_{1}q_{1}+...+p_{d}q_{d}$. ##### Geodesics and Distances: For two points p, $\textbf{q}\in\mathbb{H}^{d,K}$, the distance function between them is given by: (3.3) $\displaystyle d_{\mathcal{L}}^{K}(\textbf{p},\textbf{q})=\sqrt{K}\text{arccosh}(-\langle\textbf{p},\textbf{q}\rangle_{\mathcal{L}}/K)$ ##### Exponential and Logarithmic maps: We use exponential and logarithmic maps for mapping to and from the hyperbolic and the tangent space respectively. Formally, given a point $\textbf{p}\in\mathbb{H}^{d,K}$ and tangent vector $\textbf{t}\in\mathcal{T}_{\textbf{p}}\mathbb{H}^{d,K}$, the exponential map $\exp_{\textbf{p}}^{K}:\mathcal{T}_{\textbf{p}}\mathbb{H}^{d,K}\rightarrow\mathbb{H}^{d,K}$ assigns a point to t such that $\exp_{\textbf{p}}^{K}(\textbf{t})=\gamma(1)$, where $\gamma$ is the geodesic curve that satisfies $\gamma(0)=\textbf{p}$ and $\dot{\gamma}=\textbf{t}$. The logarithmic map $(\log^{K}_{\textbf{p}})$ being the bijective inverse maps a point in hyperbolic space to the tangent space at p. We use the following equations for the exponential and the logarithmic maps: (3.4) $\exp_{\textbf{p}}^{K}(\textbf{v})=\cosh(\frac{||\textbf{v}||_{\mathcal{L}}}{\sqrt{K}})\textbf{p}+\sqrt{K}\sinh(\frac{||\textbf{v}||_{\mathcal{L}}}{\sqrt{K}})\frac{\textbf{v}}{||\textbf{v}||_{\mathcal{L}}}$ (3.5) $\log^{K}_{\textbf{p}}(\textbf{q})=d_{\mathcal{L}}^{K}(\textbf{p},\textbf{q})\frac{\textbf{q}+\frac{1}{K}<\textbf{p},\textbf{q}>_{\mathcal{L}}\textbf{p}}{||\textbf{q}+\frac{1}{K}<\textbf{p},\textbf{q}>_{\mathcal{L}}\textbf{p}||_{\mathcal{L}}}$ #### 3.4.2 Graph Construction The end-goal of graph construction is to group the entity mentions in such a way that contextually similar mentions are clustered around each other by forming edges in the graph. Given the fact, the euclidean embeddings are better at capturing the semantic aspects of the text data [6], we opt to use deep contextualized embeddings in the euclidean domain [17] for the graph construction. For each entity type, we average out corresponding $1024d$ embeddings for all the mentions in the training corpus $C_{train}$, to learn prototype vectors for each entity type, i.e., $\\{prototype_{t}\\}_{t=1}^{T}$. Later, for each entity type $t$, we capture type-specific confident mention candidates $cand_{t}$, following the criterion: $cand_{t}=cand_{t}\cup men\text{ if }(cos(men,\\{Prototype_{t}\\})\geq\delta)$ $\forall men\in C;\forall{}t\in T$, where $\delta$ is a threshold. Finally, we form pairwise edges for all the mention candidates corresponding to each entity-type, i.e., $\\{cand\\}_{t=1}^{T}$, to construct the graph $G$, with adjacency matrix $A$. #### 3.4.3 Mapping Noisy Mention Encodings to the Hyperbolic space The mention encodings learnt in the stage-I are noisy, as they are learnt over distantly supervised data. These encodings lie in the euclidean space, and in order to refine them, we first map them to the hyperbolic space, where we may best exploit the fine-grained type hierarchy in relation with the type- specific context to fine-tune these encodings as an aggregate of contextually- similar neighbors. Formally, let $\mathbf{p}^{E}=X_{m}\in\mathbf{R}^{N\times L}$ be the matrix corresponding to the noisy mentions’ encodings in the euclidean domain. We consider $o=\\{\sqrt{K},0,...,0\\}$ as a reference point (origin) in a d-dimensional Hyperboloid with curvature $-1/K\,(\mathbb{H}^{d,K})$; $(0,\textbf{p}^{E})$ as a point in the tangent space $(\mathcal{T}\mathbb{H}^{d,K})$, and map it to $\textbf{p}^{H}\in\mathbb{H}^{d,K}$ using the exponential map given in Equation (3.4), as follows: $\displaystyle\textbf{p}^{H}$ $\displaystyle=\exp^{K}((0,\textbf{p}^{E}))$ $\displaystyle\exp^{K}((0,\textbf{p}^{E}))$ $\displaystyle=\Big{(}\sqrt{K}\cosh\Big{(}\frac{||\textbf{p}^{E}||_{2}}{\sqrt{K}}\Big{)},$ (3.6) $\displaystyle\sqrt{K}\sinh\Big{(}\frac{||\textbf{p}^{E}||_{2}}{\sqrt{K}}\Big{)}\frac{\textbf{p}^{E}}{||\textbf{p}^{E}||_{2}}\Big{)}$ #### 3.4.4 Linear Transformation In order to perform linear transformation operation on the noisy mention encodings, i.e., (i) multiplication by weight matrix W, and (ii) addition of bias vector b, we rely on the exponential and the logarithmic maps. For multiplication with the weight matrix, firstly, we apply logarithmic map on the encodings in the hyperbolic space, i.e., $\textbf{p}^{H}\in\mathbb{H}^{d,K}$, in order to project them to $\mathcal{T}\mathbb{H}^{d,K}$. This projection is then multiplied by the weight matrix $W$, and the resultant vectors are projected back to the manifold using the exponential map. For a manifold with curvature constant $K$, these operations can be summarized in the equation, given below: (3.7) $W\otimes\textbf{p}^{H}=\exp^{K}(W\log^{K}(\textbf{p}^{H}))$ For bias addition, we rely on parallel transport, let b be the bias vector in $\mathcal{T}\mathbb{H}^{d,K}$, we parallel transport b along the tangent space and finally map it to the manifold. Formally, let $\textbf{T}^{K}_{\textbf{o}\rightarrow\textbf{p}^{H}}$ represent the parallel transport of a vector from $\mathcal{T}_{\textbf{o}}\mathbb{H}^{d,K}$ to $\mathcal{T}_{\textbf{x}^{H}}\mathbb{H}^{d,K}$, we use the following equation for the bias addition: (3.8) $\textbf{p}^{H}\oplus\textbf{b}=\exp^{K}_{\textbf{x}^{H}}(\textbf{T}^{K}_{o\rightarrow\textbf{p}^{H}}(\textbf{b}))$ #### 3.4.5 Type-Specific Contextual Aggregation Aggregation is a crucial step for noise reduction in FG-NET, it helps to smoothen the type-label by refining/fine-tuning the noisy mention encodings by accumulating information from contextually similar neighbors lying at multiple hops. Given the graph $G$, with nodes $(V)$ being the entity mentions, we use the pairwise embedding vectors along the edges of the graph to compute the attention weights $\eta_{ij}=cos(men^{i},men^{j})\forall(i,j)\in V$. In order to perform the aggregation operation, we first use the logarithmic map to project the results of the linear transformation from hyperbolic space to the tangent space. Later, we use the neighboring information contained in $G$ to compute the refined mention encoding as attentive aggregate of the neighboring mentions. Finally, we map these results back to the manifold using the exponential map $\exp^{K}$. Our methodology for contextual aggregation is summarized in the following equation: (3.9) $AGG_{cxtx}(\textbf{p}^{H})_{i}=\exp^{K}_{\textbf{x}^{H}_{i}}\Big{(}\sum_{j\in\textit{N}(i)}(\widetilde{\eta_{ij}\odot A})\log^{K}(\textbf{p}^{H}_{j})\Big{)}$ where $\widetilde{\eta_{ij}\odot A}$ is the Hadamard product of the attention weights and the adjacency matrix $A$. It accommodates the degree of contextual similarity among the mention pairs in $G$. #### 3.4.6 Non-Linear Activation Contextually aggregated mention’s encoding is finally passed through a non- linear activation function $\sigma$ ($\mathsf{ReLU}$ in our case). For this, we follow similar steps, i.e., (i) map the encodings to the tangent space, (ii) apply the activation function in the tangent space, (iii) map the results back to the hyperbolic space using exponential map. These steps are summarized in the following equation: (3.10) $\sigma(\textbf{p}^{H})=\exp^{K}(\sigma(\log^{K}(\textbf{p}^{H})))$ ### 3.5 Complete Model We combine the above-mentioned steps to get the refined mention encodings at lth-layer $\textbf{z}_{out}^{l,H}$ as follows: $\displaystyle\textbf{p}^{l,H}$ $\displaystyle=W^{l}\otimes\textbf{p}^{l-1,H}\oplus\textbf{b}^{l}\text{;}$ (3.11) $\displaystyle\textbf{y}^{l,H}$ $\displaystyle=AGG_{cxtx}(\textbf{p}^{l,H})\text{;}\,\,\,\textbf{z}^{l,H}_{out}=\sigma(\textbf{y}^{l,H})$ Let $\textbf{z}_{out}^{l,H}\in\mathbb{H}^{d,K}$ correspond to the refined mentions’ encodings hierarchically organized in the hyperbolic space. We embed them along with the fine-grained type label encodings $\\{\phi_{t}\\}_{t=1}^{T}\in\mathbb{H}^{d}$. For that we learn a function $f(\textbf{z}^{l,H}_{out},\phi_{t})=\phi_{t}^{T}\times\textbf{z}^{l,H}+bias_{t}$, and separately learn the loss functions for the clean and the noisy mentions. ##### Loss for clean mentions: In order to model the clean entity mentions $D_{tr\text{-}clean}$, we use a margin-based loss to embed the refined mention encodings close to the true type labels ($T_{y}$), and push it away from the false type labels ($T_{y^{{}^{\prime}}}$). The loss function is summarized as follows: $\displaystyle L_{clean}$ $\displaystyle=\sum_{t\in T_{y}}\mathsf{ReLU}(1-f(\textbf{z}^{l,H}_{out},\phi_{t}))+$ (3.12) $\displaystyle\sum_{t^{{}^{\prime}}\in T_{y^{{}^{\prime}}}}\mathsf{ReLU}(1+f(\textbf{z}^{l,H}_{out},\phi_{t^{{}^{\prime}}}))$ ##### Loss for noisy mentions: In order to model the noisy entity mentions $D_{tr\text{-}noisy}$, we use a variant of above-mentioned loss function to embed the mention close to most relevant type label $t^{*}$, where $t^{*}=\operatorname*{argmax}_{t\in T_{y}}f(\textbf{z}^{l,H}_{out},\phi_{t})$, among the set of noisy type labels $(T_{y})$ and push it away from the irrelevant type labels ($T_{y^{{}^{\prime}}}$). The loss function is mentioned as follows: $\displaystyle L_{noisy}$ $\displaystyle=\mathsf{ReLU}(1-f(\textbf{z}^{l,H}_{out},\phi_{t^{*}}))+$ (3.13) $\displaystyle\sum_{t^{{}^{\prime}}\in T_{y^{{}^{\prime}}}}\mathsf{ReLU}(1+f(\textbf{z}^{l,H}_{out},\phi_{t^{{}^{\prime}}}))$ Finally, we minimize $L_{clean}+L_{noisy}$ as the final loss function of the FGNET-RH. ## 4 Experimentation ### 4.1 Dataset We evaluate our model using a set of publicly available datasets for FG-NET. We chose these datasets because they contain fairly large proportion of test instances and corresponding evaluation will be more concrete. Statistics of these dataset is shown in Table 1. These datasets are explained as follows: ##### BBN: Its training corpus is acquired from the Wall Street Journal annotated by [22] using DBpedia Spotlight. ##### OntoNotes: It is acquired from newswire documents contained in the OntoNotes corpus [23]. The training data is mapped to Freebase types via DBpedia Spotlight [5]. The testing data is manually annotated by Gillick et al., [8]. Dataset | BBN | OntoNotes ---|---|--- Training Mentions | 86078 | 220398 Testing Mentions | 13187 | 9603 % clean mentions (training) | 75.92 | 72.61 % clean mentions (testing) | 100 | 94.0 Entity Types | 47 | 89 Table 1: Fine-Grained Named Entity Typing data sets ### 4.2 Experimental Settings In order to set up a fair platform for comparative evaluation, we use the same data settings (training, dev and test splits) as used by all the models considered as baselines in Table 2. All the experiments are performed using Intel Gold 6240 CPU with 256 GB main memory. ##### Model Parameters: For stage-I, the hidden layer size of the context and the position encoders is set to 100d. The hidden layer size of the mention character encoder is 200d. Character, position and label embeddings are randomly initialized. We report the model performance using 300d Glove [16] and 1024d deep contextualized embeddings [17]. For stage-II, we construct graphs with 5.4M and 0.6M edges for BBN and OntoNotes respectively. Curvature constant of the hyperbolic space is set to $K=1$. All the models are trained using Adam optimizer [10] with learning rate = 0.001. ### 4.3 Model Comparison We evaluate FGNET-RH against the following baseline models: (i) Figer [13]; (ii) Hyena [29]; (iii) AFET, AFET-NoCo and AFET-NoPa [19]; (iv) Attentive [21]; (v) FNET [1]; (vi) NFGEC + LME [25]; and (vii) FGET-RR [2]. For performance comparison, we use the scores reported in the original papers, as they are computed using a similar data settings as that of ours. Note that we do not compare our model against [4, 14] because these models use crowd-sourced data in addition to the distantly supervised data for model training. Likewise, we exclude [26] from evaluation because Xu and Barbosa changed the fine-grained problem definition from multi-label to single-label classification problem. This makes their problem settings different from that of ours and the end results are no longer comparable. ### 4.4 Main Results The results of the proposed model are shown in Table 2. For each data, we boldface the best scores with the existing state-of-the art underlined. These results show that FGNET-RH outperforms the existing state-of-the-art models by a significant margin. For the BBN data, FGNET-RH achieves 3.5%, 1.2% and 1.5% improvement in strict accuracy, mac-F1 and mic-F1 respectively, compared to the FGET-RR. For OntoNotes, FGNET-RH improves the mac-F1 and mic-F1 scores by 1.2% and 1.6%. These results show that FGNET-RH offers multi-faceted benefits, i.e., using hyperbolic space in combination with the graphs to encode the hierarchy, while at the same time catering to noise in the best possible way. Especially, augmented context sharing along the hierarchy leads to considerable improvement in the performance compared to the baseline models. | OntoNotes | BBN ---|---|--- | strict | mac-F1 | mic-F1 | strict | mac-F1 | mic-F1 FIGER [13] | 0.369 | 0.578 | 0.516 | 0.467 | 0.672 | 0.612 HYENA [29] | 0.249 | 0.497 | 0.446 | 0.523 | 0.576 | 0.587 AFET-NoCo [19] | 0.486 | 0.652 | 0.594 | 0.655 | 0.711 | 0.716 AFET-NoPa [19] | 0.463 | 0.637 | 0.591 | 0.669 | 0.715 | 0.724 AFET-CoH [19] | 0.521 | 0.680 | 0.609 | 0.657 | 0.703 | 0.712 AFET [19] | 0.551 | 0.711 | 0.647 | 0.670 | 0.727 | 0.735 Attentive [21] | 0.473 | 0.655 | 0.586 | 0.484 | 0.732 | 0.724 FNET-AllC [1] | 0.514 | 0.672 | 0.626 | 0.655 | 0.736 | 0.752 FNET-NoM [1] | 0.521 | 0.683 | 0.626 | 0.615 | 0.742 | 0.755 FNET [1] | 0.522 | 0.685 | 0.633 | 0.604 | 0.741 | 0.757 NFGEC+LME [25] | 0.529 | 0.724 | 0.652 | 0.607 | 0.743 | 0.760 FGET-RR[2] (Glove) | 0.567 | 0.737 | 0.680 | 0.740 | 0.811 | 0.817 FGET-RR[2] (ELMO) | 0.577 | 0.743 | 0.685 | 0.703 | 0.819 | 0.823 FGNET-RH (Hyperboloid + Glove) | 0.580 | 0.738 | 0.685 | 0.766 | 0.828 | 0.835 FGNET-RH (Hyperboloid + ELMO) | 0.575 | 0.752 | 0.696 | 0.712 | 0.824 | 0.823 FGNET-RH (Poincaré-Ball + Glove) | 0.579 | 0.741 | 0.684 | 0.760 | 0.829 | 0.833 FGNET-RH (Poincaré-Ball + ELMO) | 0.573 | 0.740 | 0.685 | 0.698 | 0.828 | 0.830 Table 2: FG-NET performance comparison against baseline models ### 4.5 Ablation Study In this section, we evaluate the impact of different model components on label de-noising. Specifically, we analyze the performance of FGNET-RH using variants of the adjacency graph, including: (i) randomly generated adjacency graph of approximately the same size as $G$: $\text{FGNET-RH}{}\,(R)$, (ii) unweighted adjacency graph: $\text{FGNET-RH}{}\,(A)$, and (iii) pairwise contextual similarity as the attention weights $\text{FGNET- RH}{}\,(\widetilde{\eta\odot A})$. The results in Table 3 show that for the given model architecture, the performance improvement (correspondingly noise- reduction) can be attributed to using the appropriate adjacency graph. A drastic reduction in the model performance for $\text{FGNET-RH}{}\,(R)$ shows that once the contextual similarity structure of the graph is lost, the label- smoothing is no longer effective. Likewise, improvement in performance for the models: $\text{FGNET-RH}{}\,(A)$ and $\text{FGNET-RH}{}\,(\widetilde{\eta\odot A})$, implies that the adjacency graphs $(A)$ and especially $(\widetilde{\eta\odot A})$ indeed incorporate the required type-specific contextual clusters at the needed level of granularity to effectively smoothen the noisy labels prior to the entity typing. Model | OntoNotes | BBN ---|---|--- strict | mac-F1 | mic-F1 | strict | mac-F1 | mic-F1 $\text{FGNET-RH}{}\,(R)$ | 0.484 | 0.643 | 0.597 | 0.486 | 0.647 | 0.653 $\text{FGNET-RH}{}\,(A)$ | 0.531 | 0.699 | 0.632 | 0.735 | 0.808 | 0.815 $\text{FGNET-RH}{}\,(\widetilde{\eta\odot A})$ | 0.580 | 0.738 | 0.685 | 0.766 | 0.828 | 0.835 | Hyperboloid ($\mathbb{H}^{d}$) $\text{FGNET-RH}{}\,(R)$ | 0.490 | 0.665 | 0.608 | 0.633 | 0.704 | 0.724 $\text{FGNET-RH}{}\,(A)$ | 0.571 | 0.737 | 0.679 | 0.746 | 0.814 | 0.822 $\text{FGNET-RH}{}\,(\widetilde{\eta\odot A})$ | 0.579 | 0.741 | 0.684 | 0.760 | 0.829 | 0.833 | Poincaré-Ball ($\mathbb{D}^{d}$) Table 3: FGNET-RH performance comparison using different adjacency matrices and Glove Embeddings #### 4.5.1 Effectiveness of Hyperbolic Geometry In order to verify the effectiveness of refining the mention encodings in the hyperbolic space (stage-II), we perform label-wise performance analysis for the dominant labels in the BBN dataset. Corresponding results for the Hyperboloid and the Poincaré-Ball model (in Table 4) show that FGNET-RH outperforms the existing state-of-the-art, i.e., FGET-RR by Ali et al., [2], achieving higher F1-scores across all the labels. Note that FGNET-RH can achieve higher performance for the base type labels: {e.g., _“/Person”, “/Organization”, “/GPE”_ etc.,}, as well as other type labels down in the hierarchy, {e.g., _“/Organization/Corporation”, “/GPE/City”_ etc.,}. For {_“Organization”_ and _“Corporation”_} FGNET-RH achieves a higher F1=0.896 and F1=0.855 respectively, compared to the F1=0.881 and F1=0.844 by FGET-RR. This is made possible because embedding in the hyperbolic space enables type- specific context sharing at each level of the type hierarchy by appropriately adjusting the norm of the label vector. To further strengthen our claims regarding the effectiveness of using hyperbolic space for FG-NET, we analyzed the context of the entity types along the type-hierarchy. We observed, for the fine-grained type labels, the context is additive and may be arranged in a hierarchical structure with the generic terms lying at the root and the specific terms lying along the children nodes. For example, _“Government Organization”_ being a subtype of _“Organization”_ adds tokens similar to {bill, treasury, deficit, fiscal, senate etc., } to the context of _“Organization”_. Likewise, _“Hospital”_ adds tokens similar to {family, patient, kidney, stone, infection etc., } to the context of _“Organization”_. Labels | Support | FGET-RR [2] | FGNET-RH (Poincaré-Ball) | FGNET-RH (Hyperboloid) ---|---|---|---|--- Prec | Rec | F1 | Prec | Rec | F1 | Prec | Rec | F1 /Organization | 45.30% | 0.924 | 0.842 | 0.881 | 0.916 | 0.876 | 0.896 | 0.926 | 0.860 | 0.891 /Organization/Corporation | 35.70% | 0.921 | 0.779 | 0.844 | 0.903 | 0.812 | 0.855 | 0.908 | 0.801 | 0.851 /Person | 22.00% | 0.86 | 0.886 | 0.872 | 0.876 | 0.902 | 0.889 | 0.843 | 0.911 | 0.876 /GPE | 21.30% | 0.924 | 0.845 | 0.883 | 0.92 | 0.868 | 0.893 | 0.924 | 0.885 | 0.904 /GPE/City | 9.17% | 0.802 | 0.767 | 0.784 | 0.806 | 0.750 | 0.777 | 0.804 | 0.795 | 0.799 Table 4: Label-wise Precision, Recall and F1 scores for the BBN data compared with FGET-RR [2] This finding correlates with the norm of the label vectors, shown in Table 5 for the Poincaré-Ball model. The vector norm of the entity types deep in the hierarchy {e.g., _“/Facility/Building”, “/Facility/Bridge”, “/Facility/Highway”_ etc., } is greater than that of the base entity type { _“/Facility”_ }. A similar trend is observed for the fine-grained types: {_“/Organization/Government”, “/Organization/Political”_ etc.,} compared to the base type: {_“/Organization”_}. It justifies that FGNET-RH indeed adjusts the norm of the label vector according to the depth of the type-label in the label-hierarchy, which allows the model to consequently cluster the type- specific context along the hierarchy in an augmented fashion. In addition, we also analyzed the entity mentions corrected especially by the label-smoothing process, i.e., the stage-II of FGNET-RH. For this, we examined the model performance with and without the label-smoothing, i.e., we separately build a classification model by using the output of stage-I. For the BBN data, the stage-II is able to correct about 18% of the mis- classifications made by stage-I. For example in the sentence: _“CNW Corp. said the final step in the acquisition of the company has been completed with the merger of CNW with a subsidiary of Chicago & amp.”_, the bold-faced entity mention CNW is labeled {_“/GPE”_} by stage-I. However, after label-smoothing in stage-II, the label predicted by FGNET-RH is {_“/Organization/Corporation”_}, which indeed is the correct label. A similar trend was observed for the OntoNotes data set. This analysis concludes that the FGNET-RH using a blend of the contextual graphs and the hyperbolic space incorporates the right geometry to embed the noisy FG-NET data with lowest possible distortion. Compared to the euclidean space, the hyperbolic space being a non-euclidean space allows the graph volume (number of nodes within a fixed radius) to grow exponentially along the hierarchy. This enables the FGNET-RH to perform label-smoothing by forming type-specific contextual clusters across noisy mentions along the type hierarchy. #### 4.5.2 Error Cases We analyzed the prediction errors of FGNET-RH and attribute them to the following factors: ##### Inadequate Context: For these cases, type-labels are dictated entirely by the mention tokens, with very little information contained in the context. For example, in the sentence: _“The IRS recently won part of its long-running battle against John.”_, the entity mention _“ IRS”_ is labeled as {_“/Organization/Corporation”_} irrespective of any information contained in the mention’s context. Limited information contained in the mention’s context in turn limits the end-performance of FGNET-RH in predicting all possible fine-grained labels thus effecting the recall. For the BBN data set, more than 30% errors may be attributed to the inadequate mention’s context. ##### Correlated Context: FG-NET type hierarchy encompasses semantically correlated entity types, e.g., { _“Organization”_ vs _“Corporation”_}; {_“Actor”_ vs _“Artist”_}; {_“Actor”_ vs _“Director”_}; {_“Ship”_ vs _“Spacecraft”_}; {_“Coach”_ vs _“Athlete”_} etc., with highly convoluted context. For example, the context of the entity types {_“actor”_} and {_“artist”_} is extremely overlapping, it contains semantically-related tokens like: {direct, dialogue, dance, acting, etc.,}. This high contextual overlap makes it hard for the FGNET-RH to delineate the decision boundary across these correlated entity types. It leads to false predictions by the model thus effecting the precision. For the BBN data set, more than 35% errors may be attributed to the correlated context. ##### Label Bias: Label bias originating from the distant supervision may result in the label- smoothing to be in-effective. This occurs specifically if all the labels originating from the distant supervision are incorrect. For the BBN data approximately 5% errors may be attributed to the label bias. The rest of the errors may be attributed to the inability of FGNET-RH to explicitly deal with different word senses, in-depth syntactic analysis, in- adequacy of underlying embedding models to handle semantics, etc. We plan to accommodate these aspects in the future work. Label | Norm | Label | Norm ---|---|---|--- /Organization | 0.855 | /Facility | 0.643 /Organization/Religious | 0.860 | /Facility/Building | 0.725 /Organization/Government | 0.870 | /Facility/Bridge | 0.745 /Organization/Political | 0.875 | /Facility/Highway | 0.815 Table 5: FGNET-RH Label-norms for the Poincaré-Ball model, the norm for the base type-labels is lower than the type-labels deep in the hierarchy ## 5 Conclusions In this paper, we introduced FGNET-RH, a novel approach that combines the benefits of graph structures and hyperbolic geometry to perform entity typing in a robust fashion. FGNET-RH initially learns noisy mention encodings using LSTM networks and constructs a graph to cluster contextually similar mentions using embeddings in euclidean domain, later it performs label-smoothing in hyperbolic domain to refine the noisy encodings prior to the entity-typing. Performance evaluation using the benchmark datasets shows that the FGNET-RH offers a perfect geometry for context sharing across distantly supervised data, and in turn outperforms the existing research on FG-NET by a significant margin. ## References * [1] Abhishek, Ashish Anand, and Amit Awekar. Fine-grained entity type classification by jointly learning representations and label embeddings. In EACL (1), pages 797–807. Association for Computational Linguistics, 2017. * [2] Muhammad Asif Ali, Yifang Sun, Bing Li, and Wei Wang. Fine-grained named entity typing over distantly supervised data based on refined representations. In AAAI, pages 7391–7398. AAAI Press, 2020. * [3] Ines Chami, Zhitao Ying, Christopher Ré, and Jure Leskovec. Hyperbolic graph convolutional neural networks. In NeurIPS, pages 4869–4880, 2019. * [4] Eunsol Choi, Omer Levy, Yejin Choi, and Luke Zettlemoyer. Ultra-fine entity typing. In ACL (1), pages 87–96. Association for Computational Linguistics, 2018. * [5] Joachim Daiber, Max Jakob, Chris Hokamp, and Pablo N. Mendes. Improving efficiency and accuracy in multilingual entity extraction. In I-SEMANTICS, pages 121–124. ACM, 2013. * [6] Bhuwan Dhingra, Christopher J. Shallue, Mohammad Norouzi, Andrew M. Dai, and George E. Dahl. Embedding text in hyperbolic spaces. In TextGraphs@NAACL-HLT, pages 59–69. Association for Computational Linguistics, 2018. * [7] Xin Dong, Evgeniy Gabrilovich, Geremy Heitz, Wilko Horn, Ni Lao, Kevin Murphy, Thomas Strohmann, Shaohua Sun, and Wei Zhang. Knowledge vault: a web-scale approach to probabilistic knowledge fusion. In KDD, pages 601–610. ACM, 2014. * [8] Dan Gillick, Nevena Lazic, Kuzman Ganchev, Jesse Kirchner, and David Huynh. Context-dependent fine-grained entity type tagging. CoRR, abs/1412.1820, 2014. * [9] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation, 9(8):1735–1780, 1997. * [10] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR (Poster), 2015. * [11] Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. In ICLR (Poster). OpenReview.net, 2017. * [12] Ni Lao and William W Cohen. Relational retrieval using a combination of path-constrained random walks. Machine learning, 81(1):53–67, 2010. * [13] Xiao Ling and Daniel S. Weld. Fine-grained entity recognition. In AAAI. AAAI Press, 2012. * [14] Federico López, Benjamin Heinzerling, and Michael Strube. Fine-grained entity typing in hyperbolic space. In RepL4NLP@ACL, pages 169–180. Association for Computational Linguistics, 2019. * [15] Maximilian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. In NIPS, pages 6338–6347, 2017. * [16] Jeffrey Pennington, Richard Socher, and Christopher D. Manning. Glove: Global vectors for word representation. In EMNLP, pages 1532–1543. ACL, 2014. * [17] Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. Deep contextualized word representations. In NAACL-HLT, pages 2227–2237. Association for Computational Linguistics, 2018. * [18] Deepak Ravichandran and Eduard Hovy. Learning surface text patterns for a question answering system. In Proceedings of the 40th annual meeting on association for computational linguistics, pages 41–47. Association for Computational Linguistics, 2002. * [19] Xiang Ren, Wenqi He, Meng Qu, Lifu Huang, Heng Ji, and Jiawei Han. AFET: automatic fine-grained entity typing by hierarchical partial-label embedding. In EMNLP, pages 1369–1378. The Association for Computational Linguistics, 2016. * [20] Xiang Ren, Wenqi He, Meng Qu, Clare R. Voss, Heng Ji, and Jiawei Han. Label noise reduction in entity typing by heterogeneous partial-label embedding. In KDD, pages 1825–1834. ACM, 2016. * [21] Sonse Shimaoka, Pontus Stenetorp, Kentaro Inui, and Sebastian Riedel. An attentive neural architecture for fine-grained entity type classification. In AKBC@NAACL-HLT, pages 69–74. The Association for Computer Linguistics, 2016. * [22] Ralph Weischedel and Ada Brunstein. Bbn pronoun coreference and entity type corpus. Linguistic Data Consortium, Philadelphia, 112, 2005. * [23] Ralph Weischedel, Sameer Pradhan, Lance Ramshaw, Martha Palmer, Nianwen Xue, Mitchell Marcus, Ann Taylor, Craig Greenberg, Eduard Hovy, Robert Belvin, et al. Ontonotes release 4.0. LDC2011T03, Philadelphia, Penn.: Linguistic Data Consortium, 2011\. * [24] Wentao Wu, Hongsong Li, Haixun Wang, and Kenny Qili Zhu. Probase: a probabilistic taxonomy for text understanding. In SIGMOD Conference, pages 481–492. ACM, 2012. * [25] Ji Xin, Hao Zhu, Xu Han, Zhiyuan Liu, and Maosong Sun. Put it back: Entity typing with language model enhancement. In EMNLP, pages 993–998. Association for Computational Linguistics, 2018. * [26] Peng Xu and Denilson Barbosa. Neural fine-grained entity type classification with hierarchy-aware loss. In NAACL-HLT, pages 16–25. Association for Computational Linguistics, 2018. * [27] Yadollah Yaghoobzadeh, Heike Adel, and Hinrich Schütze. Noise mitigation for neural entity typing and relation extraction. arXiv preprint arXiv:1612.07495, 2016. * [28] Dani Yogatama, Daniel Gillick, and Nevena Lazic. Embedding methods for fine grained entity type classification. In ACL (2), pages 291–296. The Association for Computer Linguistics, 2015. * [29] Mohamed Amir Yosef, Sandro Bauer, Johannes Hoffart, Marc Spaniol, and Gerhard Weikum. Hyena-live: Fine-grained online entity type classification from natural-language text. In ACL (Conference System Demonstrations), pages 133–138. The Association for Computer Linguistics, 2013.
# Towards Robustness to Label Noise in Text Classification via Noise Modeling Siddhant Garg Amazon Alexa AI<EMAIL_ADDRESS>, Goutham Ramakrishnan Health at Scale Corporation<EMAIL_ADDRESS>and Varun Thumbe KLA Corporation<EMAIL_ADDRESS> (2021) ###### Abstract. Large datasets in NLP tend to suffer from noisy labels due to erroneous automatic and human annotation procedures. We study the problem of text classification with label noise, and aim to capture this noise through an auxiliary noise model over the classifier. We first assign a probability score to each training sample of having a clean or noisy label, using a two- component beta mixture model fitted on the training losses at an early epoch. Using this, we jointly train the classifier and the noise model through a novel de-noising loss having two components: (i) cross-entropy of the noise model prediction with the input label, and (ii) cross-entropy of the classifier prediction with the input label, weighted by the probability of the sample having a clean label. Our empirical evaluation on two text classification tasks and two types of label noise: random and input- conditional, shows that our approach can improve classification accuracy, and prevent over-fitting to the noise. Label Noise; Noise Model; Robustness; Text Classification; NLP ††journalyear: 2021††copyright: acmcopyright††conference: Proceedings of the 30th ACM International Conference on Information and Knowledge Management; November 1–5, 2021; Virtual Event, QLD, Australia††booktitle: Proceedings of the 30th ACM International Conference on Information and Knowledge Management (CIKM ’21), November 1–5, 2021, Virtual Event, QLD, Australia††price: 15.00††doi: 10.1145/3459637.3482204††isbn: 978-1-4503-8446-9/21/11††ccs: Computing methodologies Natural language processing ## 1\. Introduction Training modern ML models requires access to large accurately labeled datasets, which are difficult to obtain due to errors in automatic or human annotation techniques (Wang et al., 2018; Zlateski et al., 2018). Recent studies (Zhang et al., 2016) have shown that neural models can over-fit on noisy labels and thereby not generalize well. Human annotations for language tasks have been popularly obtained from platforms like Amazon Mechanical Turk (Ipeirotis et al., 2010), resulting in noisy labels due to ambiguity of the correct label (Zhan et al., 2019), annotation speed, human error, inexperience of annotator, etc. While learning with noisy labels has been extensively studied in computer vision (Reed et al., 2015; Zhang et al., 2018; Thulasidasan et al., 2019), the corresponding progress in NLP has been limited. With the increasing size of NLP datasets, noisy labels are likely to affect several practical applications (Agarwal et al., 2007). Figure 1. Illustration of our approach, with an auxiliary noise model $N_{M}$ on top of the classifier $M$. We jointly train the models using a de-noising loss $\mathcal{L}_{DN}$, and use the clean label prediction $\hat{y}^{(c)}$ during inference. In this paper, we consider the problem of text classification, and capture the label noise through an auxiliary noise model (See Fig. 1). We leverage the finding of learning on clean labels being easier than on noisy labels (Arazo et al., 2019), and first fit a two-component beta-mixture model (BMM) on the training losses from the classifier at an early epoch. Using this, we assign a probability score to every training sample of having a clean or noisy label. We then jointly train the classifier and the noise model by selectively guiding the former’s prediction for samples with high probability scores of having clean labels. More specifically, we propose a novel de-noising loss having two components: (i) cross-entropy of the noise model prediction with the input label and (ii) cross-entropy of the classifier prediction with the input label, weighted by the probability of the sample having a clean label. Our formulation constrains the noise model to learn the label noise, and the classifier to learn a good representation for the prediction task from the clean samples. At inference time, we remove the noise model and use the predictions from the classifier. (a) Epoch 1 (b) Epoch 9 (c) Epoch 30 (d) Fitting a BMM at Epoch 9 Figure 2. (a), (b) and (c) show the histogram of the training loss from the classifier for the train split (with Clean/ Noisy label) at different epochs (word-LSTM on TREC with 40% random noise). Initially (see (a)), the losses are high for all data points (both clean and noisy labels). A fully trained model achieves low losses on both clean and noisy data points, indicating over- fitting to the noise, as seen in (c). However, at an early epoch of training, we observe that samples with clean labels have lower losses while those with noisy labels have high losses, leading to the formation of two clusters as seen in (b). We fit a beta-mixture model at this intermediate epoch to estimate the probability of a sample having a clean or noisy label, as shown in (d). Most existing works on learning with noisy labels assume that the label noise is independent of the input and only conditional on the true label. Text annotation complexity has been shown to depend on the lexical, syntactic and semantic input features (Joshi et al., 2014) and not solely on the true label. The noise model in our formulation can capture an arbitrary noise function, which may depend on both the input and the original label, taking as input a contextualized input representation from the classifier. While de-noising the classifier for sophisticated noise functions is a challenging problem, we take the first step towards capturing a real world setting. We evaluate our approach on two popular datasets, for two different types of label noise: random and input-conditional; at different noise levels. Across two model architectures, our approach results in improved model accuracies over the baseline, while preventing over-fitting to the label noise. ## 2\. Related Work There have been several research works that have studied the problem of combating label noise in computer vision (Frénay and Verleysen, 2014; Jiang et al., 2018, 2019) through techniques like bootstrapping (Reed et al., 2015), mixup (Zhang et al., 2018), etc. Applying techniques like mixup (convex combinations of pairs of samples) for textual inputs is challenging due to the discrete nature of the input space and retaining overall semantics. In natural language processing, Agarwal et al. (2007) study the effect of different kinds of noise on text classification, Ardehaly and Culotta (2018) study social media text classification using label proportion (LLP) models, and Malik and Bhardwaj (2011) automatically validate noisy labels using high-quality class labels. Jindal et al. (2019) capture random label noise via a $\ell_{2}$-regularized matrix learned on the classifier logits. Our work differs from this as we i) use a neural network noise model over contextualized embeddings from the classifier, with (ii) a new de-noising loss to explicitly guide learning. It is difficult to draw a distinction between noisy labels, and outliers which are hard to learn from. While several works perform outlier detection (Goodman et al., 2016; Larson et al., 2019) to discard these samples while learning the classifier, we utilise the noisy data in addition to the clean data for improving performance. ## 3\. Methodology Problem Setting Let $(X,Y^{(c)}){=}\\{(x_{1},y_{1}^{(c)}),\dots,(x_{N},y_{N}^{(c)})\\}$ denote clean training samples from a distribution $\mathcal{D}{=}{\mathcal{X}}{\times}{\mathcal{Y}}$. We assume a function $\mathcal{F}:{\mathcal{X}}{\times}{\mathcal{Y}\rightarrow\mathcal{Y}}$ that introduces noise in labels $Y^{(c)}$. We apply $\mathcal{F}$ on $(X,Y^{(c)})$ to obtain the noisy training data $(X,Y^{(n)})=\\{(x_{1},y_{1}^{(n)}),\dots,(x_{N},y_{N}^{(n)})\\}$. $(X,Y^{(n)})$ contains a combination of clean samples (whose original label is retained $y^{(n)}{=}y^{(c)}$) and noisy samples (whose original label is corrupted $y^{(n)}{\neq}y^{(c)}$). Let $(X_{T},Y_{T})$ be a test set sampled from the clean distribution $\mathcal{D}$. Our goal is to learn a classifier model $\mathcal{M}:\mathcal{X}{\rightarrow}\mathcal{Y}$ trained on the noisy data $(X,Y^{(n)})$, which generalizes well on $(X_{T},Y_{T})$. Note that we do not have access to the clean labels $Y^{(c)}$ at any point during training. Modeling Noise Function $\mathcal{F}$ We propose to capture $\mathcal{F}$ using an auxiliary noise model $N_{M}$ on top of the classifier model $M$, as shown in Fig. 1. For an input $x$, a representation $R_{M}(x)$, derived from $M$, is fed to $N_{M}$. $R_{M}(x)$ can typically be the contextualized input embedding from the penultimate layer of $M$. We denote the predictions from $M$ and $N_{M}$ to be $\hat{y}^{(c)}$(clean prediction) and $\hat{y}^{(n)}$(noisy prediction) respectively. The clean prediction $\hat{y}^{(c)}$ is used for inference. Model | | | TREC (word-LSTM: 93.8, word-CNN: 92.6) | | AG-News (word-LSTM: 92.5, word-CNN: 91.5) ---|---|---|---|---|--- Noise % | | 10 | 20 | 30 | 40 | 50 | | 10 | 20 | 30 | 40 | 50 word LSTM | Baseline | | 88.0 (-0.6) | 89.4 (-9.6) | 83.4 (-19.0) | 79.6 (-24.8) | 77.6 (-27.2) | | 91.9 (-1.7) | 91.3 (-1.5) | 90.5 (-2.5) | 89.3 (-3.7) | 88.6 (-10.5) $\mathcal{L}_{DN{-}H}$ | | 92.2 (-0.6) | 90.2 (-0.2) | 88.8 (-0.4) | 83.0 (-3.6) | 82.4 (0.0) | | 91.5 (-0.1) | 90.6 (-0.1) | 90.8 (-0.1) | 90.3 (0.0) | 89.0 (-0.1) $\mathcal{L}_{DN{-}S}$ | | 92.4 (-1.0) | 90.0 (-0.2) | 87.4 (-2) | 83.4 (-1.0) | 82.6 (-8.4) | | 91.8 (-0.3) | 90.8 (-0.2) | 91.0 (-0.1) | 90.3 (-0.1) | 88.6 (-0.1) word CNN | Baseline | | 88.8 (-1.4) | 89.2 (-1.8) | 84.8 (-8.0) | 82.2 (-15.0) | 77.6 (-16.0) | | 90.9 (-2.7) | 90.6 (-6.2) | 89.3 (-10.2) | 89.2 (-17.9) | 87.4 (-25.2) $\mathcal{L}_{DN{-}H}$ | | 91 (-0.2) | 90.8 (-0.2) | 89.4 (-1.0) | 81.4 (0.0) | 81.4 (-4.8) | | 91.3 (-0.2) | 91.0 (-0.4) | 90.3 (-0.3) | 88.3 (-3.2) | 86.6 (-3.5) $\mathcal{L}_{DN{-}S}$ | | 92.2 (-1.4) | 91.8 (-2.0) | 88.8 (-2.8) | 77.0 (-2.4) | 77.2 (-7.0) | | 90.9 (0.0) | 90.4 (-0.1) | 88.7 (-1.1) | 86.6 (-3.5) | 84.5 (-10.2) Table 1. Results from experiments using random noise. Here for A(B): A refers to the Best model accuracy while B refers to (Last-Best) accuracy. The models with highest Best accuracies are in bold. For each noise $\%$, the least and most reductions in Last accuracy are highlighted in green and red. Baseline ($0\%$ noise) reported beside dataset. ### 3.1. Estimating clean/noisy label using BMM It has been empirically observed that classifiers that capture input semantics do not fit the noise before significantly learning from the clean samples (Arazo et al., 2019). For a classifier trained using a cross entropy loss($\mathcal{L}_{CE}$) on the noisy dataset, this can be exploited to cluster the input samples as being clean/noisy in an unsupervised manner. Initially the training loss on both clean and noisy samples is large, and after a few training epochs, the loss of majority of the clean samples reduces. Since the loss of the noisy samples is still large, this segregates the samples into two clusters with different loss values. On further training, the model over-fits on the noisy samples and the training loss on both samples reduces. We illustrate this in Fig. 2(a)$-$(c). We fit a 2-component Beta mixture model (BMM) over the normalized training losses ($\mathcal{L}_{CE}(\hat{y}^{(c)},\cdot)\in[0,1]$) obtained after training the model for some warmup epochs $T_{0}$. Using a Beta mixture model works better than using a Gaussian mixture model as it allows for asymmetric distributions and can capture the short left-tails of the clean sample losses. For a sample $(x,y)$ with normalized loss $\mathcal{L}_{CE}(\hat{y}^{(c)},y)=\ell$, the BMM is given by: $\displaystyle p(\ell)=\lambda_{c}\cdot p(\ell|\text{clean})+\lambda_{n}\cdot p(\ell|\text{noisy})$ $\displaystyle p(\ell|\text{clean})=\frac{\Gamma(\alpha_{c}+\beta_{c})}{\Gamma(\alpha_{c})\Gamma(\beta_{c})}\ell^{\alpha_{c}-1}{(1-\ell)}^{\beta_{c}-1}$ $\displaystyle p(\ell|\text{noisy})=\frac{\Gamma(\alpha_{n}+\beta_{n})}{\Gamma(\alpha_{n})\Gamma(\beta_{n})}\ell^{\alpha_{n}-1}{(1-\ell)}^{\beta_{n}-1}$ where $\Gamma$ denotes the gamma distribution and $\alpha_{c/n},\beta_{c/n}$ are the parameters corresponding to the individual clean/noisy Beta distributions. The mixture coefficients $\lambda_{c}$ and $\lambda_{n}$, and parameters ($\alpha_{c/n},\beta_{c/n}$) are learnt using the EM algorithm. On fitting the BMM $\mathcal{B}$, for a given input $x$ with a normalized loss $\mathcal{L}_{CE}(\hat{y}^{(c)},y)=\ell$, we denote the posterior probability of $x$ having a clean label by: $\mathcal{B}(x)=\frac{\lambda_{c}\cdot p(\ell|\text{clean})}{\lambda_{c}\cdot p(\ell|\text{clean})+\lambda_{n}\cdot p(\ell|\text{noisy})}$ The BMM $\mathcal{B}$ learnt from Fig. 2b is shown in Fig. 2d. Algorithm 1 Training using $\mathcal{L}_{DN-H}$ Input: Train data $(x_{i},y_{i}^{(n)})_{i=1}^{N}$, warmup epochs $T_{0}$, total epochs $T$, parameter $\beta$, classifier $M$, noise model $N_{M}$ for epoch in $\\{1,\dots,T_{0}\\}$ do $\hat{y_{i}}^{(c)}\leftarrow M(x_{i})\ \forall\ i\in[N]$ Train $M$ with $\sum_{i}\mathcal{L}_{CE}(\hat{y_{i}}^{(c)},y_{i}^{(n)})$ end for Fit a 2-mixture BMM $\mathcal{B}$ on $\\{\mathcal{L}_{CE}(\hat{y_{i}}^{(c)},y_{i}^{(n)})\\}_{i=1}^{N}$ for epoch in $\\{T_{0}+1,\dots,T\\}$ do $\hat{y_{i}}^{(c)}\leftarrow M(x_{i})$, $\hat{y_{i}}^{(n)}\leftarrow N_{M}(R_{M}(x_{i}))\ \ \forall\ i\in[N]$ Train $M,N_{M}$ with $\mathcal{L}_{DN{-}H}=$ $\qquad\sum_{i}\big{(}\mathcal{L}_{CE}(\hat{y_{i}}^{(n)},y_{i}^{(n)}){+}$ $\beta\cdot\mathbbm{1}[\mathcal{B}(x){>}0.5]\cdot\mathcal{L}_{CE}(\hat{y_{i}}^{(c)},y_{i}^{(n)})\big{)}$ end for Return: Trained classifier model $M$ ### 3.2. Learning $M$ and $N_{M}$ We aim to train $M,N_{M}$ such that when given an input, $M$ predicts the clean label and $N_{M}$ predicts the noisy label for this input (if $\mathcal{F}$ retains the original clean label for this input, then both $M,N_{M}$ predict the clean label). Thus for an input $(x,y)$ having a clean label, we want $\hat{y}^{(c)}{=}\hat{y}^{(n)}{=}y$; and for an input $(x,y)$ having a noisy label, we want $\hat{y}^{(n)}{=}y$ and $\hat{y}^{(c)}$ to be the clean label for $x$. We jointly train $M,N_{M}$ using the de-noising loss proposed below: (1) $\displaystyle\mathcal{L}_{DN}{=}\ \mathcal{L}_{CE}(\hat{y}^{(n)},y)\ {+}\beta{\cdot}\mathcal{B}(x){\cdot}\mathcal{L}_{CE}({\hat{y}^{(c)}},y)$ The first term trains the $M{-}N_{M}$ cascade jointly using cross entropy between $\hat{y}^{(n)}$ and $y$. The second term trains $M$ to predict $\hat{y}^{(c)}$ correctly for samples believed to be clean, weighted by $\mathcal{B}(x)$. Here $\beta$ is a parameter that controls the trade-off between the two terms. By jointly training $M,N_{M}$ with $\mathcal{L}_{DN}$, we implicitly constrain the label noise in $N_{M}$. We use an alternative loss formulation by replacing the Bernoulli $\mathcal{B}(x)$ with the indicator $\mathbbm{1}[\mathcal{B}(x){>}0.5]$. For ease of notation, we refer the former (using $\mathcal{B}(x)$) as the soft de-noising loss $\mathcal{L}_{DN{-}S}$ and the latter as the hard de-noising loss $\mathcal{L}_{DN{-}H}$. Thus we use the following 3-step approach to learn $M$ and $N_{M}$: 1. (1) Warmup: Train $M$ using $\mathcal{L}_{CE}(\hat{y}^{(c)},y)$. 2. (2) Fitting BMM: Fit a 2-component BMM $\mathcal{B}$ on the $\mathcal{L}_{CE}(\hat{y}^{(c)},y)$ for all $(x,y)\in(X,Y^{(n)})$. 3. (3) Training with $\mathcal{L}_{DN}$: Jointly train $M$ and $N_{M}$ end-to-end using $\mathcal{L}_{DN{-}S/H}$. We summarize our methodology in Algorithm 1, when using $\mathcal{L}_{DN-H}$. Dataset | Num. Classes | Train | Validation | Test ---|---|---|---|--- TREC (Li and Roth (2002)) | 6 | 4949 | 503 | 500 AG-News (Gulli (2005)) | 4 | 112000 | 8000 | 7600 Table 2. Summary statistics of the datasets Model | | | Token (How/What) based | | Length based ---|---|---|---|---|--- Noise % | | 10 | 20 | 30 | 40 | 50 | | 10 | 20 | 30 | 40 | 50 word LSTM | Baseline | | 89.2 (-0.4) | 84.4 (-8.2) | 77.8 (-10.6) | 76 (-17) | 71.8 (-15.8) | | 91.4 (-1.0) | 87 (0.6) | 82.2 (1.8) | 82.4 (-2.6) | 74.2 (-3.0) $\mathcal{L}_{DN{-}H}$ | | 91.8 (0) | 87.4 (-2.2) | 84.2 (0.4) | 79 (1) | 67.8 (1.4) | | 91.6 (-0.6) | 90.2 (-0.8) | 87.4 (-0.2) | 87.4 (-0.8) | 79 (0) $\mathcal{L}_{DN{-}S}$ | | 91.8 (0.2) | 90.6 (-1.2) | 83.8 (-6.8) | 79.2 (-19.2) | 75.6 (-15.8) | | 92 (0.4) | 90.6 (1) | 85.4 (0.2) | 84 (-2.8) | 75 (-3) word CNN | Baseline | | 90.4 (-3.6) | 83.8 (-1.8) | 82.4 (-7.4) | 78.8 (-17.2) | 52 (1.4) | | 91 (0) | 88 (1.2) | 85.2 (-1.2) | 82 (-2.6) | 73.6 (-1.4) $\mathcal{L}_{DN{-}H}$ | | 90 (0.8) | 86.6 (-3) | 84.4 (-0.6) | 80.6 (-4.2) | 74 (-7.6) | | 90.6 (-0.8) | 89.6 (-1) | 87.2 (-0.2) | 82.6 (-0.4) | 77 (-6.2) $\mathcal{L}_{DN{-}S}$ | | 91.2 (-0.4) | 86.8 (-1.8) | 84.2 (-4.2) | 81.8 (-12) | 65.2 (-12.4) | | 92.8 (-3.6) | 91 (-2.2) | 86.8 (-1.4) | 86 (-4) | 75.4 (1.8) Table 3. Results from experiments using input-conditional noise on the TREC dataset. ## 4\. Evaluation Datasets We experiment with two popular text classification datasets: (i) TREC question-type dataset (Li and Roth, 2002), and (ii) AG-News dataset (Gulli, 2005) (Table 2). We inject noise in the training and validation sets, while retaining the original clean test set for evaluation. Note that collecting real datasets with known patterns of label noise is a challenging task, and out of the scope of this work. We artificially inject noise in clean datasets, which enables easy and extensive experimentation. Models We conduct experiments on two popular model architectures: word-LSTM (Hochreiter and Schmidhuber, 1997) and word-CNN (Kim, 2014). For word-LSTM, we use a 2-layer BiLSTM with hidden dimension of 150. For word-CNN, we use 300 kernel filters each of size 3, 4 and 5. We use the pre-trained GloVe embeddings (Pennington et al., 2014) for initializing the word embeddings for both models. We train models on TREC and AG-News for 100 and 30 epochs respectively. We use an Adam optimizer with a learning rate of $10^{-5}$ and a dropout of $0.3$ during training. For the noise model $N_{M}$, we use a simple 2-layer feedforward neural network, with the number of hidden units $n_{hidden}=4{\cdot}n_{input}$. We choose the inputs to the noise model $R_{M}(x)$ as per the class of label noise, which we describe in Section 4.1 and 4.2. We conduct hyper-parameter tuning for the number of warmup epochs $T_{0}$ and $\beta$ using grid search over the ranges of {6,10,20} and {2,4,6,8,10} respectively. Metrics and Baseline We evaluate the robustness of the model to label noise on two fronts: (i) how well it performs on clean data, and (ii) how much it over- fits the noisy data. For the former, we report the test set accuracy (denoted by Best) corresponding to the model with best validation accuracy . For the latter, we examine the gap in test accuracies between the Best, and the Last model (after last training epoch). We evaluate our approach against only training $M$ (as the baseline), for two types of noise: random and input- conditional, at different noise levels. ### 4.1. Results: Random Noise For a specific Noise %, we randomly change the original labels of this percentage of samples. Since the noise function is independent of the input, we use logits from $M$ as the input $R(x)$ to $N_{M}$. We report the Best and (Last \- Best) test accuracies in Table 1. From the experiments, we observe that: (i) $L_{DN-S}$ and $L_{DN-H}$ almost always outperforms the baseline across different noise levels. The performance of $L_{DN-S}$ and $L_{DN-H}$ are similar. We observe that training with $L_{DN{-}S}$ tends to be better at low noise %, whereas $L_{DN{-}H}$ tends to be better at higher noise %. Our method is more effective for TREC than AG-News, since even the baseline can learn robustly on AG-News. (ii) Our approach using $L_{DN{-}S}$ and $L_{DN{-}H}$ drastically reduces over-fitting on noisy samples (visible from small gaps between Best and Last accuracies). For the baseline, this gap is significantly larger, especially at high noise levels, indicating over-fitting to the label noise. For example, consider word-LSTM on TREC at 30% noise: while the baseline suffers a sharp drop of 24.8 points from 79.6%, the accuracy of the $L_{DN{-}S}$ model drops just 1.0% from 83.4%. Figure 3. Test accuracy across training epochs of word-LSTM model on the TREC dataset with two levels of random noise: $30\%$ and $40\%$. The baseline heavily over-fits on the noise degrading performance, while $L_{DN{-}S}$ and $L_{DN{-}H}$ avoid this. We further demonstrate that our approach avoids over-fitting, thereby stabilizing the model training by plotting the test accuracies across training epochs in Fig. 3. We observe that the baseline model over-fits the label noise with more training epochs, thereby degrading test accuracy. The degree of over-fitting is greater at higher levels of noise (Fig. 3(b) vs Fig. 3(a)). In comparison, our de-noising approach using both $L_{DN-S}$ and $L_{DN-H}$ does not over-fit on the noisy labels as demonstrated by stable test accuracies across epochs. This is particularly beneficial when operating in a few-shot setting where one does not have access to a validation split that is representative of the test split for early stopping. | Noise | AP (7.8%) | Reuters (10.8%) | Either (18.6%) ---|---|---|---|--- word LSTM | Baseline | 82.8 (-0.5) | 85.6 (-0.8) | 75.7 (-0.4) $\mathcal{L}_{DN{-}H}$ | 82.7 (0) | 85.7 (-0.1) | 76.6 (-0.4) $\mathcal{L}_{DN{-}S}$ | 82.8 (0.3) | 85.5 (0.1) | 76 (-0.1) word CNN | Baseline | 83.1 (-0.2) | 85.7 (0) | 76.6 (-0.9) $\mathcal{L}_{DN{-}H}$ | 82.4 (0.8) | 86.2 (0) | 76.1 (0.1) $\mathcal{L}_{DN{-}S}$ | 82.5 (0.5) | 86.1 (0.1) | 76.4 (0) Table 4. Results from input-conditional noise on AG-News. ### 4.2. Results: Input-Conditional Noise We heuristically condition the noise function $\mathcal{F}$ on lexical and syntactic input features. We are the first to study such label noise for text inputs, to our knowledge. For both the TREC and AG-News, we condition $\mathcal{F}$ on syntactic features of the input: (i) The TREC dataset contains different types of questions. We selectively corrupt the labels of inputs that contain the question words ‘How’ or ‘What’ (chosen based on occurrence frequency). For texts starting with ‘How’ or ‘What’, we insert random label noise (at different levels). We also consider $\mathcal{F}$ conditional on the text length (a lexical feature). More specifically, we inject random label noise for the longest x% inputs in the dataset. (ii) The AG-News dataset contains news articles from different news agency sources. We insert random label noise for inputs containing the token ‘AP’, ‘Reuters’ or either one of them. We concatenate the contextualised input embedding from the penultimate layer of $M$ and the logits corresponding to $\hat{y}^{(c)}$ as the input $R_{M}(x)$ to $N_{M}$. We present the results in Tables 3 and 4. On TREC, our method outperforms the baseline for both the noise patterns we consider. For the question-length based noise, we observe the same trend of $L_{DN{-}H}$ outperforming $L_{DN{-}S}$ at high noise levels, and vice-versa. On AG-News, the noise % for inputs having the specific tokens ’AP’ and ’Reuters’ are relatively low, and our method performs at par or marginally improves over the baseline performance. Interestingly, the input–conditional noise we consider makes effective learning very challenging, as demonstrated by significantly lower _Best_ accuracies for the baseline model than for random noise. As the classifier appears to overfit to the noise very early during training, we observe relatively smaller gaps between Best and Last accuracies. Compared to random noise, our approach is less efficient at alleviating the _(Best-Last)_ accuracy gap for input-conditional noise. These experiments however reveal promising preliminary results on learning with input-conditional noise. ## 5\. Conclusion We have presented an approach to improve text classification when learning from noisy labels by jointly training a classifier and a noise model using a de-noising loss. We have evaluated our approach on two text classification tasks. demonstrate its effectiveness through an extensive evaluation. Future work includes studying more complex $\mathcal{F}$ for other NLP tasks like language inference and QA. ## References * (1) * Agarwal et al. (2007) Sumeet Agarwal, Shantanu Godbole, Shourya Roy, and Diwakar Punjani. 2007. How much noise is too much: A study in automatic text classification. In _In Proc. of ICDM_. * Arazo et al. (2019) Eric Arazo, Diego Ortego, Paul Albert, Noel O’Connor, and Kevin Mcguinness. 2019. Unsupervised Label Noise Modeling and Loss Correction. In _Proceedings of the 36th International Conference on Machine Learning_ _(Proceedings of Machine Learning Research, Vol. 97)_ , Kamalika Chaudhuri and Ruslan Salakhutdinov (Eds.). PMLR, Long Beach, California, USA, 312–321. http://proceedings.mlr.press/v97/arazo19a.html * Ardehaly and Culotta (2018) Ehsan Ardehaly and Aron Culotta. 2018. Learning from noisy label proportions for classifying online social data. _Social Network Analysis and Mining_ 8 (12 2018). https://doi.org/10.1007/s13278-017-0478-6 * Frénay and Verleysen (2014) Benoît Frénay and Michel Verleysen. 2014. Classification in the Presence of Label Noise: A Survey. _IEEE Transactions on Neural Networks and Learning Systems_ 25 (2014), 845–869. * Goodman et al. (2016) James Goodman, Andreas Vlachos, and Jason Naradowsky. 2016\. Noise reduction and targeted exploration in imitation learning for Abstract Meaning Representation parsing. In _Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)_. Association for Computational Linguistics, Berlin, Germany, 1–11. https://doi.org/10.18653/v1/P16-1001 * Gulli (2005) A. Gulli. 2005. The Anatomy of a News Search Engine. In _Special Interest Tracks and Posters of the 14th International Conference on World Wide Web_ (Chiba, Japan) _(WWW ’05)_. Association for Computing Machinery, New York, NY, USA, 880–881. * Hochreiter and Schmidhuber (1997) Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long Short-Term Memory. _Neural Comput._ 9, 8 (Nov. 1997), 1735–1780. https://doi.org/10.1162/neco.1997.9.8.1735 * Ipeirotis et al. (2010) Panagiotis G. Ipeirotis, Foster Provost, and Jing Wang. 2010\. Quality Management on Amazon Mechanical Turk. In _Proceedings of the ACM SIGKDD Workshop on Human Computation_ (Washington DC) _(HCOMP ’10)_. Association for Computing Machinery, New York, NY, USA, 64–67. https://doi.org/10.1145/1837885.1837906 * Jiang et al. (2019) Junjun Jiang, Jiayi Ma, Zheng Wang, Chen Chen, and Xianming Liu. 2019. Hyperspectral Image Classification in the Presence of Noisy Labels. _IEEE Transactions on Geoscience and Remote Sensing_ 57 (2019), 851–865. * Jiang et al. (2018) Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. 2018. MentorNet: Learning Data-Driven Curriculum for Very Deep Neural Networks on Corrupted Labels. In _Proceedings of the 35th International Conference on Machine Learning_ , Jennifer Dy and Andreas Krause (Eds.). PMLR. http://proceedings.mlr.press/v80/jiang18c.html * Jindal et al. (2019) Ishan Jindal, Daniel Pressel, Brian Lester, and Matthew Nokleby. 2019. An Effective Label Noise Model for DNN Text Classification. In _Proceedings of North American Chapter of the Association of Computational Linguistics 2019_. * Joshi et al. (2014) Aditya Joshi, Abhijit Mishra, Nivvedan Senthamilselvan, and Pushpak Bhattacharyya. 2014. Measuring Sentiment Annotation Complexity of Text. In _Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers)_. Association for Computational Linguistics, Baltimore, Maryland. * Kim (2014) Yoon Kim. 2014. Convolutional Neural Networks for Sentence Classification. In _Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing, EMNLP 2014_. 1746–1751. * Larson et al. (2019) Stefan Larson, Anish Mahendran, Andrew Lee, Jonathan K. Kummerfeld, Parker Hill, Michael A. Laurenzano, Johann Hauswald, Lingjia Tang, and Jason Mars. 2019. Outlier Detection for Improved Data Quality and Diversity in Dialog Systems. In _Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers)_. Association for Computational Linguistics, Minneapolis, Minnesota, 517–527. https://doi.org/10.18653/v1/N19-1051 * Li and Roth (2002) Xin Li and Dan Roth. 2002\. Learning Question Classifiers. In _Proceedings of the 19th International Conference on Computational Linguistics - Volume 1_ (Taipei, Taiwan) _(COLING ’02)_. Association for Computational Linguistics, Stroudsburg, PA, USA, 1–7. * Malik and Bhardwaj (2011) H. H. Malik and V. S. Bhardwaj. 2011. Automatic Training Data Cleaning for Text Classification. In _2011 IEEE 11th International Conference on Data Mining Workshops_. 442–449. * Pennington et al. (2014) Jeffrey Pennington, Richard Socher, and Christopher D. Manning. 2014. GloVe: Global Vectors for Word Representation. In _Empirical Methods in Natural Language Processing (EMNLP)_. 1532–1543. http://www.aclweb.org/anthology/D14-1162 * Reed et al. (2015) Scott E. Reed, Honglak Lee, Dragomir Anguelov, Christian Szegedy, Dumitru Erhan, and Andrew Rabinovich. 2015. Training Deep Neural Networks on Noisy Labels with Bootstrapping.. In _ICLR (Workshop)_ , Yoshua Bengio and Yann LeCun (Eds.). http://dblp.uni-trier.de/db/conf/iclr/iclr2015w.html#ReedLASER14 * Thulasidasan et al. (2019) Sunil Thulasidasan, Tanmoy Bhattacharya, Jeff Bilmes, Gopinath Chennupati, and Jamal Mohd-Yusof. 2019\. Combating Label Noise in Deep Learning using Abstention. In _Proceedings of the 36th International Conference on Machine Learning_ , Kamalika Chaudhuri and Ruslan Salakhutdinov (Eds.). PMLR. http://proceedings.mlr.press/v97/thulasidasan19a.html * Wang et al. (2018) Fei Wang, Liren Chen, Cheng Li, Shiyao Huang, Yanjie Chen, Chen Qian, and Chen Change Loy. 2018. The Devil of Face Recognition is in the Noise. _arXiv preprint arXiv:1807.11649_ (2018). * Zhan et al. (2019) Xueying Zhan, Yaowei Wang, Yanghui Rao, and Qing Li. 2019\. Learning from Multi-Annotator Data: A Noise-Aware Classification Framework. _ACM Trans. Inf. Syst._ 37, 2, Article 26 (Feb. 2019), 28 pages. https://doi.org/10.1145/3309543 * Zhang et al. (2016) Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. 2016. Understanding deep learning requires rethinking generalization. _ICLR_ (2016). http://arxiv.org/abs/1611.03530 cite arxiv:1611.03530Comment: Published in ICLR 2017. * Zhang et al. (2018) Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. 2018. mixup: Beyond Empirical Risk Minimization. In _International Conference on Learning Representations_. https://openreview.net/forum?id=r1Ddp1-Rb * Zlateski et al. (2018) Aleksandar Zlateski, Ronnachai Jaroensri, Prafull Sharma, and Fredo Durand. 2018. On the Importance of Label Quality for Semantic Segmentation. In _CVPR_. 1479–1487. https://doi.org/10.1109/CVPR.2018.00160
# How does an external electric field trigger the Cassie-Baxter-Wenzel wetting transition on a textured surface? Ke Xiao†, Xi Chen†, and Chen-Xu Wu<EMAIL_ADDRESS>Fujian Provincial Key Laboratory for Soft Functional Materials Research, Department of Physics, College of Physical Science and Technology, Xiamen University, Xiamen 361005, People’s Republic of China (August 27, 2024) ###### Abstract Understanding the critical condition and mechanism of the droplet wetting transition between Cassie-Baxter state and Wenzel state triggered by an external electric field is of considerable importance because of its numerous applications in industry and engineering. However, such a wetting transition on a patterned surface is still not fully understood, e.g., the effects of electro-wetting number, geometry of the patterned surfaces, and droplet volume on the transition have not been systematically investigated. In this paper, we propose a theoretical model for the Cassie-Baxter-Wenzel wetting transition triggered by applying an external voltage on a droplet placed on a mirco- pillared surface or a porous substrate. It is found that the transition is realized by lowering the energy barrier created by the intermediate composite state considerably, which enables the droplet to cross the energy barrier and complete the transition process. Our calculations also indicate that for fixed droplet volume, the critical electrowetting number (voltage) will increase (decrease) along with the surface roughness for a micro-pillar patterned (porous) surface, and if the surface roughness is fixed, a small droplet tends to ease the critical electrowetting condition for the transition. Besides, three dimensional phase diagrams in terms of electrowetting number, surface roughness, and droplet volume are constructed to illustrate the Cassie-Baxter- Wenzel wetting transition. Our theoretical model can be used to explain the previous experimental results about the Cassie-Baxter-Wenzel wetting transition reported in the literature. ## I INTRODUCTION The role of wettability is widely studied in order to gain highly water- repellent substrates referred to as superhydrophobic surfaces with an ultrahigh apparent contact angle, a much smaller contact angle hysteresis D.Quere2005 ; D.Quere2008 , and hydrodynamic slip C.Choi2006 ; P.Joseph2006 ; A.Steinberger2007 . These properties rely on the nano- or the microscale topological structures of the surfaces which exhibit a broad range of applications in engineering such as self-cleaning R.Blossey2003 ; K.M.Wisdom2013 , water proofing A.Lafuma2003 ; M.Nosonovsky2007 , drag reduction R.Truesdell2006 , anti-dew/reflection J.B.Boreyko2009 ; R.H.Siddique2015 , bactericidal activity E.P.Ivanova2012 ; E.P.Ivanova2013 ; X.L.Li2016 ; KeXiao2020 , and so on. Typically, when a liquid droplet rests on such a roughened surface, a classical description of the droplet is characterized by Wenzel (W) R.N.Wenzel1936 and Cassie-Baxter (CB) A.B.D.Cassie1944 wetting states. The former one corresponds to a homogenous wetting state in which liquid penetrates into the texture, i.e. a fully wetted state losing superhydrophobicity, whereas the droplet in Cassie-Baxter state merely suspends on the tips of the rough surface, featured with air being trapped in the cavities of the rough surface topography. Generally, the transition between these two states can be triggered via various approaches by tuning external control parameters, such as passive strategies that rely, for example, on the utilization of gravity force Z.Yoshimitsu2002 ; B.Majhy2020 , evaporation P.Tsai2010 ; X.M.Chen2012 ; H.C.M.Fernandes2015 , and Laplace pressure A.Giacomello2012 ; P.Lv2014 , and active strategies that employ surface acoustic wave A.Sudeepthi2020 , vibration E.Bormashenko2007 ; J.B.Boreyko2009October , and even electric field G.Manukyan2011 ; J.M.Oh2011 ; R.Roy2018 ; B.X.Zhang2019 , a technique called electrowetting (EW). The advantages of adaption to various geometries, little power consumption, and fast and precise fine-tuning of the wetting state make EW a prevalent technique receiving significant research interests in the past decade F.Mugele2005 ; W.C.Nelson2012 . It is well known that once an electric voltage is applied between a substrate and a droplet settling on it, the initial equilibrium contact angle of the droplet will be reduced to a new smaller value, leading to an alteration of its apparent wettability referred to as electrowetting-on-dielectric F.Mugele2005 ; L.Q.Chen2014 . Such an EW phenomenon has attracted significant attention due to its extensive applications in lab-on-chip systems F.Mugele2005 and microfluidic operations M.G.Pollack2002 ; V.Bahadur2007 , and the discovery of the ability to induce wetting states transition G.Manukyan2011 ; J.M.Oh2011 ; R.Roy2018 ; B.X.Zhang2019 and the droplet detachment A.Cavalli2016 ; Q.Vo2019 ; Q.G.Wang2020 ; K.Xiao . Recently, numerous efforts using experiment G.Manukyan2011 ; S.Berry2012 ; Y.Chen2019 , theoretical modeling V.Bahadur2007 ; R.Roy2018 , and computer simulation J.M.Oh2011 ; B.X.Zhang2019 ; A.M.Miqdad2016 have been devoted to getting a better understanding of the wetting transition triggered by electric field. By combining experiment and numerical simulation, it has been found that the stability of the CB state under EW is determined by the balance of the Maxwell stress and the Laplace stress G.Manukyan2011 . Meanwhile, the wetting transition from CB state to W state is controlled by the energy barrier stemming from the pinning of the contact lines at the edges of the hydrophobic pillars G.Manukyan2011 . Based on a surface energy model, Roy et al. estimated the energy barriers for the EW-induced CB-to-W transition of a droplet on a mushroom-shaped re-entrant microstructures and an array of cylindrical microposts. They experimentally demonstrated that the transition on a mushroom structure is more resilient than that on an array of microposts R.Roy2018 . Besides, computer simulation also provides a useful complement to revealing the underlying mechanism of droplet wetting transition on textured surfaces. By employing molecular dynamics simulations, Zhang et al. B.X.Zhang2019 studied the mechanism behind the CB-to-W transition of a nanoscale water droplet resting on a nanogrooved surface under an external electric field, and found that there exists an energy barrier separating the CB state and the W state. In addition, they also discussed the dependence of the energy barrier on the electric filed strength, the groove aspect ratio, and the intrinsic contact angle of the groove. Despite of the fact that the EW-induced transition have been extensively studied either via experimental or theoretical approaches, a systematic analytical understanding of the underlying mechanism, in particular, the dependence of critical electric voltage on surface roughness and droplet volume has not been explored. In this paper, we establish a theoretical model to study the EW transition on micro-patterned surfaces through analyzing the difference of interfacial free energy between CB state and intermediate composite state. The effects of surface roughness and droplet volume on the threshold voltage are discussed. To further explore the interrelation among threshold voltage, surface roughness and droplet volume, three dimensional (3D) phase diagrams in the corresponding parameter space are also constructed. We expect that our model can offer some guidance to the design and fabrication of the patterned surfaces and allow one to study EW transition on other types of patterned surfaces. ## II THEORETICAL MODELING We begin our investigation by considering an EW setup consisting of a millimeter-sized sessile water droplet deposited on two different types of superhydrophobic surface decorated with, respectively, a square lattice of cylindric mircropillars and a regular array of pores, as shown in Fig. 1. The three dimensional (3D) geometry of the micro-pillar patterned surface in this paper is schematically shown in Fig. 1(a). The top view and the side view of the squarely distributed micropore-patterned surface are sketched in Figs. 1(b) and 1(c) respectively. Here, the cylindrical pillars and the pores are characterized by their radius $R$, height $H$, and gap pitch $P$ (or center- to-center interspacing $S$) between neighboring pillars or pores, respectively. Traditionally, nondimensional parameters roughness factor $r$ and solid fraction $\phi_{s}$, which are defined as the ratio of the actual area of the solid surface to its projection area and the ratio of the contact solid surface (tip of the pillars or the pores) to the total horizontal surface respectively, are commonly used to represent the level of roughness for the textured surfaces. Geometrically, in this paper, the roughness factor is given by $r=1+2\pi RH/S^{2}$, and the solid fraction $\phi_{s}$ are written as $\pi R^{2}/S^{2}$ for the pillar patterned surface and $1-\pi R^{2}/S^{2}$ for the porous surface respectively. The two possible wetting states, i.e., CB state and W state, are illustrated by Figs. 1(d) and 1(f), between which there exists an energy barrier, i.e., an intermediate composite state [see Fig. 1(e)], which can be lowered to a level below CB state by applying an external electric voltage across the droplet. Figure 1: (Color online) (a) Schematic 3D picture of a microstructured surface with cylindrical pillars of radius $R$, height $H$ and gap pitch $P$ on a periodic square lattice with pillar-to-pillar spacing $S$. Schematic (b) top view and (c) side view of a regular array of pores of radius $R$, depth $H$, gap pitch $P$, and center-to-center pore spacing $S$. Different wetting states of a droplet on a microtextured surface in the presence of an external electric voltage: (d) CB state, (e) intermediate composite state, and (f) W state. This leads to a CB-W wetting transition. In order to probe the critical electric voltage that triggers the wetting transition, it is necessary to calculate the total interfacial free energy for CB state and intermediate composite state, which, in general, is given by the sum of all the surface energies between the droplet and the patterned surface, i.e., $\displaystyle G=\gamma_{\rm lv}A_{\rm lv}+\gamma_{\rm sv}A_{\rm sv}+\gamma_{\rm ls}^{\rm eff}A_{\rm ls},$ (1) where $A_{\rm lv}$, $A_{\rm sv}$, and $A_{\rm ls}$ are the areas of the liquid-vapor, solid-vapor, and liquid-solid interfaces, and $\gamma_{\rm lv}$, $\gamma_{\rm sv}$, and $\gamma_{\rm ls}^{\rm eff}$ are the liquid-vapor, solid-vapor, and effective liquid-solid interfacial energy density, respectively. Here $\gamma_{\rm ls}^{\rm eff}=\gamma_{\rm ls}-\eta\gamma_{\rm lv}$ with $\eta=\varepsilon_{0}\varepsilon U^{2}/2d\gamma_{\rm lv}$ represents the dimensionless electrowetting number, where $\varepsilon_{0}$, $\varepsilon$, and $d$ are the dielectric permittivity in vacuum, the relative dielectric constant, and the thickness of the insulating layer, respectively. Let $A_{\rm t}=A_{\rm sv}+A_{\rm ls}$, the total area of the solid surface including solid-vapor and liquid-solid interfaces, and with a consideration of the Young’s equation $\gamma_{\rm sv}-\gamma_{\rm ls}=\gamma_{\rm lv}\cos{\theta_{\rm Y}}$, Eq. (1) can be converted to $\displaystyle G=\gamma_{\rm lv}A_{\rm lv}-\gamma_{\rm lv}(\cos{\theta_{\rm Y}}+\eta)A_{\rm ls}+\gamma_{\rm sv}A_{\rm t},$ (2) where $\theta_{\rm Y}$ is the apparent contact angle at the equilibrium state. Here we assume that the volume of the droplet is small, corresponding to a characteristic size smaller than the capillary length $l_{\rm c}=\sqrt{\gamma_{\rm lv}/\rho g}\sim 2.7\leavevmode\nobreak\ {\rm mm}$, where $\rho$ and $g$ are the density of the water and the gravitational acceleration respectively. In this case the gravitational effect can be neglected, and the shape of the droplet associated to the transition can be treated as a sphere. In particular, the total interfacial free energy of a water droplet in CB state on a patterned surface, as schematically illustrated by Fig. 1(d), can be calculated as $\displaystyle G_{\rm CB}=$ $\displaystyle\gamma_{\rm lv}\pi R_{\rm CB}^{2}\Big{[}\nu(\theta_{\rm CB})+(1-\phi_{s})-(\cos{\theta_{\rm Y}}+\eta)\phi_{s}\Big{]}$ $\displaystyle+\gamma_{\rm sv}A_{\rm t},$ (3) where $\nu(\theta)=2/(1+\cos\theta)$ is a dimensionless function. In addition, as the evaporation effect of the water droplet is excluded as well, it is reasonable to assume that the droplet volume $V_{0}$ is conserved. Therefore the base radius $R_{\rm CB}$ and the contact angle $\theta_{\rm CB}$ of the droplet can be determined according to a minimization of the global energy under the constraint of fixed droplet volume. Similarly, the total interfacial free energy of the intermediate composite state reads $\displaystyle G_{\rm inter}=$ $\displaystyle\gamma_{\rm lv}\pi R_{\rm inter}^{2}\Bigg{\\{}\nu(\theta_{\rm inter})+(1-\phi_{s})\nu(\theta_{\rm Y}-\frac{\pi}{2})-$ $\displaystyle(\cos{\theta_{\rm Y}}+\eta)\bigg{[}\phi_{s}+(r-1)\frac{h}{H}\bigg{]}\Bigg{\\}}+\gamma_{\rm sv}A_{\rm t},$ (4) where $R_{\rm inter}$ and $h$ are the base radius of the droplet and the penetration height from the tip of the pillars or pores to the point at which the curved interstitial liquid-vapor meniscus touches the edges of the walls (see Appendix for the calculation of $h$), respectively. Figure 2: (Color online) Representative total interfacial free energies of the CB and intermediate states of a droplet on a micropillar-patterned surface: $G_{\rm CB}$ and $G_{\rm inter}$, respectively, as a function of EW number. The inset shows their difference, $\Delta G=G_{\rm inter}-G_{\rm CB}$, which gives the critical EW number (critical external electric voltage) for the wetting transition when $\Delta G=0$. ## III RESULTS AND DISCUSSION In our model, we take into account all the interfacial free energies for a CB state and an intermediate composite state of a three-dimensional droplet when placed on a micropillar- or a pore- patterned surface. Our calculations in this paper were carried out by using $\gamma_{\rm lv}=72.8\leavevmode\nobreak\ {\rm mN\cdot m^{-1}}$, $\varepsilon=3.2$, $d=3\leavevmode\nobreak\ \mu{\rm m}$ and $\theta_{\rm Y}=115^{\circ}$, respectively R.Roy2018 . Without the application of an external field, as valued by the intercept in Fig. 2, the energy of the intermediate composite state is higher than that of the CB state, denying the occurrence of a CB-W transition. However, once an external field is applied and increased, it is found that although the energy of the CB state decreases, the energy of the intermediate composite state decreases in a steady and more rapid way, and there exists an intersection point where the droplet in CB state crosses the energy barrier created by the intermediate composite state and changes to a Wenzel state, as shown in Fig. 2. Such a critical condition corresponds to a critical voltage (or a critical EW number $\eta_{\rm c}$) which can be estimated by equating the energies of these two states for a given textured surface. Figure 3: (Color online) The critical EW number $\eta_{c}$ for the CB-W transition vs (a) aspect ratio $R/H$, (b) surface roughness $r$, (c) solid fraction $\phi_{s}$, and (d) $P/H$, where the black square dot curve and the red circle dot curve stand for micropillar- and pore- patterned surfaces, respectively. For a fixed droplet volume, the EW number $\eta_{\rm c}$ is found to remarkably depend on the geometric features of the structured surface (surface roughness), such as aspect ratio $R/H$, relative pitch $P/H$ (density $1/S^{2}=1/(P+2R)^{2}$), surface roughness $r$ and solid fraction $\phi_{\rm s}$, as shown in Fig. 3. It can be seen that, as shown by the black square dot curve in Fig. 3(a), it becomes harder (easier) for a CB-W transition on a micropillar-patterned surface (porous surface) to occur with the increase of its aspect ratio $R/H$. An alternative illustration of Fig. 3(a) is to replace the dimensionless aspect ratio $R/H$ by surface roughness $r$, indicating that roughness suppresses (enhances) EW-induced CB-W transition on a micropillar-patterned surface (porous surface) (Fig. 3(b)). Besides, as the aspect ratio also correlates with solid fraction $\phi_{\rm s}$, it is possible to depict the critical EW number in terms of solid fraction $\phi_{\rm s}$ for both pillar- and pore- patterned surfaces, as shown in Fig. 3(c). Apart from the aspect ratio, the distribution density of pillars and pores also plays an important role in determining the critical EW number. A more deeper investigation, as shown in Fig. 3(d), exhibits that a reduction of critical EW number $\eta_{\rm c}$ can be achieved by increasing (decreasing) the pitch of pillar-patterned (pore-patterned) surfaces. Figure 4: (Color online) The critical EW number $\eta_{c}$ for the CB-W transition vs droplet volume $V_{0}$, where the black square dot curve and the red circle dot curve represent micropillar- and pore- patterned surfaces, respectively. It has been found experimentally that the onset of CB-W wetting transition occurs at a certain droplet size as the droplet gets smaller during evaporation process P.Tsai2010 ; X.M.Chen2012 ; H.C.M.Fernandes2015 . Meanwhile, the thermodynamic favorable wetting state also depends on droplet volume K.Xiao2017 . Such a phenomenon can be explained by our theoretical model. Figure 4 exhibits the effect of droplet volume $V_{0}$ on wetting for a set of fixed surface geometric parameters ($R=25\leavevmode\nobreak\ {\rm\mu m}$, $H=50\leavevmode\nobreak\ {\rm\mu m}$, and $S=100\leavevmode\nobreak\ {\rm\mu m}$), revealing that the larger the droplet size, the higher the critical voltage for the wetting transition to occur. Figure 5: (Color online) 3D phase diagrams in terms of the critical EW number $\eta_{c}$, surface roughness $r$ and droplet volume $V_{0}$ for (a) a micropillar-patterned surface and (b) a pore-patterned surface. Finally, in order to comprehensively understand how surface roughness and droplet size affect the critical value of $\eta_{c}$ as a whole, 3D phase diagrams for micropillar and pore-patterned surface are constructed in terms of EW number, surfce roughness, and droplet volume, as demonstrated in Fig. 5. It is found that all phase diagrams are divided into two regimes, namely CB state (under the curved surface) and W state (above the curved surface) separated by a coexisting curved surface representing the critical condition for the wetting transition. According to the 3D phase diagrams, we can deduce that a higher (lower) critical EW number $\eta_{c}$ is required to trigger the CB-W transition for rougher pillar- (pore-) patterned surfaces. While for large droplet, higher $\eta_{c}$ is needed for the wetting transition regardless of the geometric pattern of the surfaces. Therefore, it becomes possible that the CB-W transition triggered by EW effect can be effectively inhibited by engineering a surface with hierarchical roughness or by adopting large droplet. To examine the validity of our theoretical model, it is necessary to compare our theoretical predictions with experimental results. For example, Bahadur et al. V.Bahadur2008 showed via experiments that the observed transition voltage is 35 V for microstructured surface with roughness 2.87, solid fraction 0.23, and pillar height 43.1 ${\rm\mu m}$, while the transition voltage has to be increased to 58 V to observe the wetting transition for the microstructured surface with same pillar height but an increased surface roughness 3.71 and solid fraction 0.55, a result in good agreement with our present conclusions. In addition, it has been found by the experimental observations made by Manukyan et al. G.Manukyan2011 that the critical voltage causing the cavities filled with water and the lateral propagation decreases as the gap widths between the pillars were widened (equivalent to a reduction of surface roughness and solid fraction), a conclusion also in accordance with our results. What’s more, Roy et al. R.Roy2018 also reported that the more sparsely the cylindrical microposts are distributed, the higher the EW voltage is required to trigger the CB-W transition. These results all support the theoretical model proposed in this paper. ## IV CONCLUSION In this paper we developed a model to interpret the CB-W wetting transition triggered by an external electric field for a three-dimensional water droplet deposited on a micropillar- or a pore-patterned surface. It is found that the electric field lowers the energy barrier created by the intermediate composite state, which allows the droplet initially in CS state to cross and complete the EW-induced CB-W transition. The critical value of the electric field applied for the transition is influenced by the geometrical parameters for the thermodynamic wetting states, such as the base radius and apparent contact angle of a droplet, and the critical voltage for the CB-W wetting transition. 3D phase diagrams in terms of EW number, surface roughness, and droplet volume are constructed. It is shown that low (high) roughness, low (high) pitch, small (small) solid fraction, and small (small) droplet size encourages a CB-W wetting transition triggered by an external field, a conclusion in good agreement with previous investigations reported in the literature. ${{\dagger}}$ These authors contributed equally to this work. ###### Acknowledgements. This work was funded by the National Science Foundation of China under Grant No. 11974292 and No. 11947401. ## V Appendix: The calculation of the base radius and the penetration height of a droplet placed on a patterned surface For all the wetting states in the main text, the base radius of a droplet is determined under the constraint of fixed volume, which, in CB state, can be written as $\displaystyle V_{\rm CB}=\frac{\pi}{3}R_{\rm CB}^{3}\mu(\theta_{\rm CB})=V_{0},$ (5) where $\mu(\theta)=(2+\cos\theta)(1-\cos\theta)^{2}/\sin^{3}\theta$ is a dimensionless function, and $R_{\rm CB}$ and $\theta_{\rm CB}$ are the base radius and the apparent contact angle of the droplet respectively. Such an equation can also be rewritten as $\displaystyle R_{\rm CB}=\bigg{[}\frac{3V_{0}}{\pi\mu(\theta_{\rm CB})}\bigg{]}^{1/3}.$ (6) When a droplet on a microstructured surface reaches the intermediate state, its volume can be divided into two parts, i.e., the one on the top of the patterned surface and the other penetrating into the interspacing of the pillars. The volume of the spherical cap above the patterned surface is given by $\displaystyle V_{\rm top}^{\rm inter}=\frac{\pi}{3}R_{\rm inter}^{3}\mu(\theta_{\rm inter}).$ (7) Strictly speaking, the equilibrium configuration of the curved liquid-vapor interface in a unit cell between the pillars needs to be determined by the Young-Laplace equation. However, due to the fact that the droplet volume filling the space around the pillars is much smaller than that of the spherical cap above the patterned surface, it is reasonable to treat the shape of the liquid-vapor interface in a unit cell between the pillars as a spherical cap with effective capillary radius $R_{\rm cap}^{\rm eff}$ defined by R.Roy2018 $\displaystyle R_{\rm cap}^{\rm eff}=S\bigg{(}\frac{1-\phi_{s}}{\pi}\bigg{)}^{1/2}.$ (8) Thus, the height of the corresponding spherical cap is calculated as $\displaystyle h_{\rm cap}^{\rm eff}=R_{\rm cap}^{\rm eff}\frac{1-\cos\big{(}\theta_{\rm Y}-\frac{\pi}{2}\big{)}}{\sin\big{(}\theta_{\rm Y}-\frac{\pi}{2}\big{)}}.$ (9) The penetration height $h$ in the main text can be obtained as $h=H-h_{\rm cap}^{\rm eff}$. Then the droplet volume underneath the top spherical cap corresponding to the volume penetrated into the interspacing of the pillars is given by $\displaystyle V_{\rm bottom}^{\rm inter}=\frac{\pi}{3}R_{\rm inter}^{2}(1-\phi_{s})\bigg{[}3h+R_{\rm cap}^{\rm eff}\mu\big{(}\theta_{\rm Y}-\frac{\pi}{2}\big{)}\bigg{]}.$ (10) Given this, the base radius $R_{\rm mid}$ of a droplet in the intermediate state can be found by solving the following equation $\displaystyle V_{0}=\frac{\pi}{3}R_{\rm inter}^{3}\mu(\theta_{\rm inter})+\frac{\pi}{3}R_{\rm inter}^{2}(1-\phi_{s})\bigg{[}3h+R_{\rm cap}^{\rm eff}\mu\big{(}\theta_{\rm Y}-\frac{\pi}{2}\big{)}\bigg{]}.$ (11) Similarly, when a droplet on a pore-patterned surface gets into intermediate state, the calculation of its base radius can be done in the same manner as above so long as we replace the effective capillary radius of the spherical cap of the bottom part by the radius of the pore $R$. ## References * (1) D. Quéré, Non-sticking drops, Rep. Prog. Phys. 68, 2495 (2005). * (2) D. Quéré, Wetting and Roughness, Annu. Rev. Mater. Res. 38, 71 (2008). * (3) C. Choi and C. J. Kim, Large Slip of Aqueous Liquid Flow over a Nanoengineered Superhydrophobic Surface, Phys. Rev. Lett. 96, 066001 (2006). * (4) P. Joseph, C. Cottin-Bizonne, J.-M. Benoît, C. Ybert, C. Journet, P. Tabeling, and L. Bocquet, Slippage of Water Past Superhydrophobic Carbon Nanotube Forests in Microchannels, Phys. Rev. Lett. 97, 156104 (2006). * (5) A. Steinberger, C. Cottin-Bizonne, P. Kleimann, and E. Charlaix, High friction on a bubble mattress, Nature Mater. 6, 665 (2007). * (6) R. Blossey, Self-Cleaning Surfaces Virtual Realities, Nat. Mater. 2 301-306 (2003) * (7) K. M. Wisdom, J. A. Watson, X. Qu, F. Liu, G. S. Watson, and C. H. Chen, Selfcleaning of superhydrophobic surfaces by self-propelled jumping condensate, Proc. Natl. Acad. Sci. U. S. A. 20, 7992-7997 (2013). * (8) A. Lafuma and D. Quere, Superhydrophobic States, Nat. Mater., 2 457-460 (2003). * (9) M. Nosonovsky and B. Bhushan, Biomimetic Superhydrophobic Surfaces: Multiscale Approach, Nano Lett. 7, 2633 (2007). * (10) R. Truesdell, A. Mammoli, P. Vorobieff, F. van Swol, and C. J. Brinker, Drag reduction on a patterned superhydrophobic surface, Phys. Rev. Lett. 97, 044504 (2006). * (11) J. B. Boreyko and C. H. Chen, Self-propelled dropwise condensate on superhydrophobic surfaces, Phys. Rev. Lett. 103, 184501 (2009). * (12) R. H. Siddique, G. Gomard, and H. Holscher, The role of random nanostructures for the omnidirectional anti-reflection properties of the glasswing butterfly, Nat. Commun. 6, 301 (2015). * (13) E. P. Ivanova, J. Hasan, H. K. Webb, V. K. Truong, G. S. Watson, J. A. Watson, V. A. Baulin, S. Pogodin, J. Y. Wang, M. J. Tobin, C. Lobbe, and R. J. Crawford, Natural Bactericidal Surfaces: Mechanical Rupture of Pseudomonas aeruginosa Cells by Cicada Wings, Small 8, 2489 (2012). * (14) E. P. Ivanova, J. Hasan, H. K. Webb, G. Gervinskas, S. Juodkazis, V. K. Truong, A. H. F. Wu, R. N. Lamb, V. A. Baulin, G. S. Watson, J. A. Watson, D. E. Mainwaring, and R. J. Crawford, Bactericidal activity of black silicon, Nat. Commun. 4, 359 (2013). * (15) X. L. Li, Bactericidal mechanism of nanopatterned surfaces, Phys. Chem. Chem. Phys. 18, 1311-1316 (2016). * (16) K. Xiao, X. Z. Cao, X. Chen, H. Z. Hu, and C. X. Wu, Bactericidal efficacy of nanopatterned surface tuned by topography, J. Appl. Phys. 128, 064701 (2020). * (17) R. N. Wenzel, Resistance of Solid Surfaces to Wetting by Water, Ind. Eng. Chem. 28, 988-994 (1936). * (18) A. B. D. Cassie and S. Baxter, Wettability of Porous Surface, Trans. Faraday Soc. 40, 546-551 (1944). * (19) Z. Yoshimitsu, A. Nakajima, T. Watanabe, and K. Hashimoto, Effects of Surface Structure on the Hydrophobicity and Sliding Behavior of Water Droplets, Langmuir 18, 5818 (2002). * (20) B. Majhy, V.P. Singh, A. K. Sen, Understanding wetting dynamics and stability of aqueous droplet over superhydrophilic spot surrounded by superhydrophobic surface, J. Colloid Interface Sci. 565, 582-591 (2020). * (21) P. Tsai, R. G. H. Lammertink, M. Wessling, and D. Lohse, Evaporation-Triggered Wetting Transition for Water Droplets upon Hydrophobic Microstructures, Phys. Rev. Lett. 104, 116102 (2010). * (22) X. M. Chen, R. Y. Ma, J. T. Li, C. L. Hao, W. Guo, B. L. Luk, S. C. Li, S. H. Yao, and Z. K. Wang, Evaporation of Droplets on Superhydrophobic Surfaces: Surface Roughness and Small Droplet Size Effects, Phys. Rev. Lett. 109, 116101 (2012). * (23) H. C. M. Fernandes, M. H. Vainstein, and C. Brito, Modeling of Droplet Evaporation on Superhydrophobic Surfaces, Langmuir 31, 7652-7659 (2015). * (24) A. Giacomello, M. Chinappi, S. Meloni, and C. M. Casciola, Metastable Wetting on Superhydrophobic Surfaces: Continuum and Atomistic Views of the Cassie-Baxter-Wenzel Transition, Phys. Rev. Lett. 109, 226102 (2012). * (25) P. Lv, Y. Xue, Y. Shi, H. Lin, and H. Duan, Metastable States and Wetting Transition of Submerged Superhydrophobic Structures, Phys. Rev. Lett. 112, 196101 (2014). * (26) A. Sudeepthi, L. Yeo, and A. K. Sen, Cassie-Wenzel wetting transition on nanostructured superhydrophobic surfaces induced by surface acoustic waves, Appl. Phys. Lett. 116, 093704 (2020). * (27) E. Bormashenko, R. Pogreb, G. Whyman, Y. Bormashenko, and M. Erlich, Vibration-induced Cassie-Wenzel wetting transition on rough surfaces, Appl. Phys. Lett. 90, 201917 (2007). * (28) J. B. Boreyko and C. H. Chen, Restoring Superhydrophobicity of Lotus Leaves with Vibration-Induced Dewetting, Phys. Rev. Lett. 103, 174502 (2009). * (29) G. Manukyan, J. M. Oh, D. van den Ende, R. G. H. Lammertink, and F. Mugele, Electrical Switching of Wetting States on Superhydrophobic Surfaces: A Route Towards Reversible Cassie-to-Wenzel Transitions, Phys. Rev. Lett. 106, 014501 (2011). * (30) J. M. Oh, G. Manukyan, D. van den Ende, and F. Mugele, Electric-field-driven instabilities on superhydrophobic surfaces, EPL 93, 56001 (2011). * (31) R. Roy, J. A. Weibel, and S. V. Garimella, Re-entrant Cavities Enhance Resilience to the Cassie-to-Wenzel State Transition on Superhydrophobic Surfaces during Electrowetting, Langmuir 34, 12787-12793 (2018). * (32) B. X. Zhang, S. L. Wang, and X. D. Wang, Wetting Transition from the Cassie-Baxter State to the Wenzel State on Regularly Nanostructured Surfaces Induced by an Electric Field, Langmuir 35, 662-670 (2019). * (33) F. Mugele and J. C. Baret, Electrowetting: from basics to applications, J. Phys.: Condens Matter 17, R705-R774 (2005). * (34) W. C. Nelson and C. J. Kim, Droplet Actuation by Electrowetting-on-Dielectric (EWOD): A Review, J. Adhes. Sci. Technol. 26, 1747-1771 (2012). * (35) L. Q. Chen,E. Bonaccurso, Electrowetting - From statics to dynamics, Adv. Colloid Interface Sci. 210, 2-12 (2014). * (36) M. G. Pollack, A. D. Shenderov, and R. B. Fair, Electrowetting-based actuation of droplets for integrated microfluidics, Lab Chip 2, 96-101 (2002). * (37) V. Bahadur and S. V. Garimella, Electrowetting-Based Control of Static Droplet States on Rough Surfaces, Langmuir 23, 4918-4924 (2007). * (38) A. Cavalli, D. J. Preston, E. Tio, D. W. Martin, N. Miljkovic, E. N. Wang, F. Blanchette, and J. W. M. Bush, Electrically induced drop detachment and ejection, Phys. Fluids 28, 022101 (2016). * (39) Q. Vo and T. Tran, Critical Conditions for Jumping Droplets, Phys. Rev. Lett. 123, 024502 (2019). * (40) Q. G. Wang, M. Xu, C. Wang, J. P. Gu, N. Hu, J. F. Lyu, and W. Yao, Actuation of a Nonconductive Droplet in an Aqueous Fluid by Reversed Electrowetting Effect, Langmuir 36, 8152-8164 (2020). * (41) K. Xiao and C. X. Wu, Critical condition for electrowetting-induced detachment of a droplet from a curved surface, arXiv:2012.07255 (2020). * (42) S. Berry, T. Fedynyshyn, L. Parameswaran, and A. Cabral, Switchable electrowetting of droplets on dual-scale structured surfaces, J. Vac. Sci. Technol. B 30, 06F801 (2012). * (43) Y. Chen, Y. Suzuki, and K. Morimoto, Electrowetting-Dominated Instability of Cassie Droplets on Superlyophobic Pillared Surfaces, Langmuir 35, 2013-2022 (2019). * (44) A. M. Miqdad, S. Datta, A. K. Das, and P. K. Das, Effect of electrostatic incitation on the wetting mode of a nano-drop over a pillar-arrayed surface, RSC Adv. 6, 110127 (2016). * (45) K. Xiao, Y. P. Zhao, G. Ouyang, and X. L. Li, An analytical model of nanopatterned superhydrophobic surfaces, J. Coat. Technol. Res. 14, 1297-1306 (2017). * (46) V. Bahadur and S. V. Garimella, Electrowetting-Based Control of Droplet Transition and Morphology on Artificially Microstructured Surfaces, Langmuir 24, 8338-8345 (2008).
# PPT: Parsimonious Parser Transfer for Unsupervised Cross-Lingual Adaptation Kemal Kurniawan1 Lea Frermann1 Philip Schulz2 Trevor Cohn1 1School of Computing and Information Systems, University of Melbourne 2Amazon Research <EMAIL_ADDRESS> <EMAIL_ADDRESS> <EMAIL_ADDRESS> <EMAIL_ADDRESS> Work done outside Amazon. ###### Abstract Cross-lingual transfer is a leading technique for parsing low-resource languages in the absence of explicit supervision. Simple ‘direct transfer’ of a learned model based on a multilingual input encoding has provided a strong benchmark. This paper presents a method for unsupervised cross-lingual transfer that improves over direct transfer systems by using their output as implicit supervision as part of self-training on unlabelled text in the target language. The method assumes minimal resources and provides maximal flexibility by (a) accepting any pre-trained arc-factored dependency parser; (b) assuming no access to source language data; (c) supporting both projective and non-projective parsing; and (d) supporting multi-source transfer. With English as the source language, we show significant improvements over state- of-the-art transfer models on both distant and nearby languages, despite our conceptually simpler approach. We provide analyses of the choice of source languages for multi-source transfer, and the advantage of non-projective parsing. Our code is available online.111https://github.com/kmkurn/ppt- eacl2021 ## 1 Introduction Figure 1: Illustration of our technique. For a target language sentence ($x_{i}$), a source parser $P_{\theta_{0}}$ predicts a set of candidate arcs $\tilde{A}(x_{i})$ (subset shown in the figure), and parses $\tilde{Y}(x_{i})$. The highest scoring parse is shown on the bottom (green), and the true gold parse (unknown to the parser) on top (red). A target language parser $P_{\theta}$ is then fine-tuned on a data set of ambiguously labelled sentences $\\{x_{i},\tilde{Y}(x_{i})\\}$. Recent progress in natural language processing (NLP) has been largely driven by increasing amounts and size of labelled datasets. The majority of the world’s languages, however, are low-resource, with little to no labelled data available (Joshi et al., 2020). Predicting linguistic labels, such as syntactic dependencies, underlies many downstream NLP applications, and the most effective systems rely on labelled data. Their lack hinders the access to NLP technology in many languages. One solution is cross-lingual model transfer, which adapts models trained on high-resource languages to low- resource ones. This paper presents a flexible framework for cross-lingual transfer of syntactic dependency parsers which can leverage _any_ pre-trained arc-factored dependency parser, and assumes no access to labelled target language data. One straightforward method of cross-lingual parsing is direct transfer. It works by training a parser on the source language labelled data and subsequently using it to parse the target language directly. Direct transfer is attractive as it does not require labelled target language data, rendering the approach fully unsupervised.222Direct transfer is also called zero-shot transfer or model transfer in the literature. Recent work has shown that it is possible to outperform direct transfer if unlabelled data, either in the target language or a different auxiliary language, is available (He et al., 2019; Meng et al., 2019; Ahmad et al., 2019b). Here, we focus on the former setting and present flexible methods that can adapt a pre-trained parser given unlabelled target data. Despite their success in outperforming direct transfer by leveraging unlabelled data, current approaches have several drawbacks. First, they are limited to generative and projective parsers. However, discriminative parsers have proven more effective, and non-projectivity is a prevalent phenomenon across the world’s languages (de Lhoneux, 2019). Second, prior methods are restricted to single-source transfer, however, transfer from multiple source languages has been shown to lead to superior results (McDonald et al., 2011; Duong et al., 2015a; Rahimi et al., 2019). Third, they assume access to the source language data, which may not be possible because of privacy or legal reasons. In such source-free transfer, only a pre-trained source parser may be provided. We address the three shortcomings with an alternative method for unsupervised target language adaptation (Section 2). Our method uses high probability edge predictions of the source parser as a supervision signal in a self-training algorithm, thus enabling unsupervised training on the target language data. The method is feasible for discriminative and non-projective parsing, as well as multi-source and source-free transfer. Building on a framework introduced in Täckström et al. (2013), this paper for the first time demonstrates their effectiveness in the context of state-of-the-art neural dependency parsers, and their generalizability across parsing frameworks. Using English as the source language, we evaluate on eight distant and ten nearby languages (He et al., 2019). The single-source transfer variant (Section 2.1) outperforms previous methods by up to $11\text{\,}\mathrm{\char 37\relax}$ UAS, averaged over nearby languages. Extending the approach to multi-source transfer (Section 2.2) gives further gains of $2\text{\,}\mathrm{\char 37\relax}$ UAS and closes the performance gap against the state of the art on distant languages. In short, our contributions are: 1. 1. A conceptually simple and highly flexible framework for unsupervised target language adaptation, which supports multi-source and source-free transfer, and can be employed with any pre-trained state-of-the-art arc-factored parser(s); 2. 2. Generalisation of the method of Täckström et al. (2013) to state-of-the-art, non-projective dependency parsing with neural networks; 3. 3. Up to $13\text{\,}\mathrm{\char 37\relax}$ UAS improvement over state-of-the- art models, considering nearby languages, and roughly equal performance over distant languages; and 4. 4. Analysis of the impact of choice of source languages on multi-source transfer quality. ## 2 Supervision via Transfer In our scenario of unsupervised cross-lingual parsing, we assume the availability of a pre-trained source parser, and unlabelled text in the target language. Thus, we aim to leverage this data such that our cross-lingual transfer parsing method out-performs direct transfer. One straightforward method is self-training where we use the predictions from the source parser as supervision to train the target parser. This method may yield decent performance as direct transfer is fairly good to begin with. However, we may be able to do better if we also consider a set of parse trees that have high probability under the source parser (cf. Fig. 1 for illustration). If we assume that the source parser can produce a set of possible trees instead, then it is natural to use all of these trees as supervision signal for training. Inspired by Täckström et al. (2013), we formalise the method as follows. Given an unlabelled dataset $\\{x_{i}\\}_{i=1}^{n}$, the training loss can be expressed as $\displaystyle\mathcal{L}(\theta)$ $\displaystyle=-\frac{1}{n}\sum_{i=1}^{n}\log\sum_{y\in\tilde{Y}(x_{i})}P_{\theta}(y|x_{i})$ (1) where $\theta$ is the target parser parameters and $\tilde{Y}(x_{i})$ is the set of trees produced by the source parser. Note that $\tilde{Y}(x_{i})$ must be smaller than the set of all trees spanning $x$ (denoted as $\mathcal{Y}(x_{i})$ ) because $\mathcal{L}(\theta)=0$ otherwise. This training procedure is a form of self-training, and we expect that the target parser can learn the correct tree as it is likely to be included in $\tilde{Y}(x_{i})$. Even if this is not the case, as long as the correct arcs occur quite frequently in $\tilde{Y}(x_{i})$, we expect the parser to learn a useful signal. We consider an arc-factored neural dependency parser where the score of a tree is defined as the sum of the scores of its arcs, and the arc scoring function is parameterised by a neural network. The probability of a tree is then proportional to its score. Formally, this formulation can be expressed as $\displaystyle P_{\theta}(y|x)$ $\displaystyle=\frac{\exp s_{\theta}(x,y)}{Z(x)}$ (2) $\displaystyle s_{\theta}(x,y)$ $\displaystyle=\sum_{(h,m)\in A(y)}s_{\theta}(x,h,m)$ (3) where $Z(x)=\sum_{y\in\mathcal{Y}(x)}\exp s_{\theta}(x,y)$ is the partition function, $A(y)$ is the set of head-modifier arcs in $y$, and $s_{\theta}(x,y)$ and $s_{\theta}(x,h,m)$ are the tree and arc scoring function respectively. ### 2.1 Single-Source Transfer Here, we consider the case where a single pre-trained source parser is provided and describe how the set of trees is constructed. Concretely, for every sentence $x=w_{1},w_{2},\ldots,w_{t}$ in the target language data, using the source parser, the set of high probability trees $\tilde{Y}(x)$ is defined as the set of dependency trees that can be assembled from the high probability arcs set $\tilde{A}(x)=\bigcup_{m=1}^{t}\tilde{A}(x,m)$, where $\tilde{A}(x,m)$ is the set of high probability arcs whose dependent is $w_{m}$. Thus, $\tilde{Y}(x)$ can be expressed formally as $\displaystyle\tilde{Y}(x)$ $\displaystyle=\\{y|y\in\mathcal{Y}(x)\wedge A(y)\subseteq\tilde{A}(x)\\}.$ (4) $\tilde{A}(x,m)$ is constructed by adding arcs $(h,m)$ in order of decreasing arc marginal probability until their cumulative probability exceeds a threshold $\sigma$ (Täckström et al., 2013). The predicted tree from the source parser is also included in $\tilde{Y}(x)$ so the chart is never empty. This prediction is simply the highest scoring tree. This procedure is illustrated in Fig. 1. Since $\mathcal{Y}(x)$ contains an exponential number of trees, efficient algorithms are required to compute the partition function $Z(x)$, arc marginal probabilities, and the highest scoring tree. First, arc marginal probabilities can be computed efficiently with dynamic programming for projective trees (Paskin, 2001) and Matrix-Tree Theorem for the non-projective counterpart (Koo et al., 2007; McDonald and Satta, 2007; Smith and Smith, 2007). The same algorithms can also be employed to compute $Z(x)$. Next, the highest scoring tree can be obtained efficiently with Eisner’s algorithm (Eisner, 1996) or the maximum spanning tree algorithm (McDonald et al., 2005; Chu and Liu, 1965; Edmonds, 1967) for the projective and non-projective cases, respectively. The transfer is performed by initialising the target parser with the source parser’s parameters and then fine-tuning it with the training loss in Eq. 1 on the target language data. Following previous works (Duong et al., 2015b; He et al., 2019), we also regularise the parameters towards the initial parameters to prevent them from deviating too much since the source parser is already good to begin with. Thus, the final fine-tuning loss becomes $\displaystyle\mathcal{L}^{\prime}(\theta)$ $\displaystyle=\mathcal{L}(\theta)+\lambda||\theta-\theta_{0}||_{2}^{2}$ (5) where $\theta_{0}$ is the initial parameters and $\lambda$ is a hyperparameter regulating the strength of the $L_{2}$ regularisation. This single-source transfer strategy was introduced as ambiguity-aware self-training by Täckström et al. (2013). A difference here is that we regularise the target parser’s parameters against the source parser’s as the initialiser, and apply the technique to modern lexicalised state-of-the-art parsers. We refer to this transfer strategy as PPT hereinafter. Note that the whole procedure of PPT can be performed even when the source parser is trained with monolingual embeddings. Specifically, given a source parser trained _only on monolingual embeddings_ , one can align pre-trained target language word embeddings to the source embedding space using an offline cross-lingual alignment method (e.g., of Smith et al. (2017)), and use the aligned target embeddings with the source model to compute $\tilde{Y}(x)$. Thus, our method can be used with any pre-trained monolingual neural parser. ### 2.2 Multi-Source Transfer We now consider the case where multiple pre-trained source parsers are available. To extend PPT to this multi-source case, we employ the ensemble training method from Täckström et al. (2013), which we now summarise. We define $\tilde{A}(x,m)=\bigcup_{k}\tilde{A}_{k}(x,m)$ where $\tilde{A}_{k}(x,m)$ is the set of high probability arcs obtained with the $k$-th source parser. The rest of the procedure is exactly the same as PPT. Note that we need to select one source parser as the main source to initialise the target parser’s parameters with. Henceforth, we refer to this method as PPTX. Multiple source parsers may help transfer better because each parser will encode different syntactic biases from the languages they are trained on. Thus, it is more likely for one of those biases to match that of the target language instead of using just a single source parser. However, multi-source transfer may also hurt performance if the languages have very different syntax, or the source parsers are of poor quality, which can arise from poor quality cross-lingual word embeddings. ## 3 Experiments ### 3.1 Setup We run our experiments on Universal Dependency Treebanks v2.2 (Nivre et al., 2018). We reimplement the self-attention graph-based parser of Ahmad et al. (2019a) that has been used with success for cross-lingual dependency parsing. Averaged over 5 runs, our reimplementation achieves $88.8\text{\,}\mathrm{\char 37\relax}$ unlabelled attachment score (UAS) on English Web Treebank using the same hyperparameters,333Reported in Table 4. slightly below their reported $90.3\text{\,}\mathrm{\char 37\relax}$ result.444UAS and LAS are reported excluding punctuation tokens. We select the run with the highest labelled attachment score (LAS) as the source parser. We obtain cross-lingual word embeddings with the offline transformation of Smith et al. (2017) applied to fastText pre-trained word vectors (Bojanowski et al., 2017). We include the universal POS tags as inputs by concatenating the embeddings with the word embeddings in the input layer. We acknowledge that the inclusion of gold POS tags does not reflect a realistic low-resource setting where gold tags are not available, which we discuss more in Section 3.3. We evaluate on 18 target languages that are divided into two groups, distant and nearby languages, based on their distance from English as defined by He et al. (2019).555We exclude Japanese and Chinese based on Ahmad et al. (2019a), who reported atypically low performance on these two languages, which they attributed to the low quality of their cross-lingual word embeddings. In subsequent work they excluded these languages (Ahmad et al., 2019b). During the unsupervised fine-tuning, we compute the training loss over all trees regardless of projectivity (i.e. we use Matrix-Tree Theorem to compute Eq. 1) and discard sentences longer than 30 tokens to avoid out-of-memory error. Following He et al. (2019), we fine-tune on the target language data for 5 epochs, tune the hyperparameters (learning rate and $\lambda$) on Arabic and Spanish using LAS, and use these values666Reported in Table 5. for the distant and nearby languages, respectively. We set the threshold $\sigma=0.95$ for both PPT and PPTX following Täckström et al. (2013). We keep the rest of the hyperparameters (e.g., batch size) equal to those of Ahmad et al. (2019a). For PPTX, unless otherwise stated, we consider a leave-one-out scenario where we use all languages except the target as the source language. We use the same hyperparameters as the English parser to train these non-English source parsers and set the English parser as the main source. ### 3.2 Comparisons We compare PPT and PPTX against several recent unsupervised transfer systems. First, He is a neural lexicalised DMV parser with normalising flow that uses a language modelling objective when fine-tuning on the unlabelled target language data (He et al., 2019). Second, Ahmad is an adversarial training method that attempts to learn language-agnostic representations (Ahmad et al., 2019b). Lastly, Meng is a constrained inference method that derives constraints from the target corpus statistics to aid inference (Meng et al., 2019). We also compare against direct transfer (DT) and self-training (ST) as our baseline systems.777ST requires significantly less memory so we only discard sentences longer than 60 tokens. Complete hyperparameter values are shown in Table 5. ### 3.3 Results Target | UAS | LAS ---|---|--- DT | ST | PPT | PPTX | He | Ahmad | Meng | DT | ST | PPT | PPTX | Ahmad fa | $37.53$ | $38.0$ | $39.47$ | $53.58$ | $63.20$ | — | — | $29.24$ | $30.5$ | $31.60$ | $44.52$ | — ar† | $37.60$ | $39.2$ | $39.46$ | $48.29$ | $55.44$ | $38.98$ | $47.3$ | $27.34$ | $30.0$ | $29.94$ | $38.48$ | $27.89$ id | $51.63$ | $49.9$ | $50.28$ | $71.91$ | $64.20$ | $51.57$ | $53.1$ | $45.23$ | $44.4$ | $44.72$ | $59.02$ | $45.31$ ko | $35.07$ | $37.1$ | $37.47$ | $34.59$ | $37.03$ | $34.23$ | $37.1$ | $16.57$ | $18.2$ | $17.99$ | $16.11$ | $16.08$ tr | $36.93$ | $38.1$ | $39.16$ | $38.44$ | $36.05$ | — | $35.2$ | $18.49$ | $19.5$ | $19.04$ | $20.60$ | — hi | $33.70$ | $34.7$ | $33.96$ | $36.39$ | $33.17$ | $37.37$ | $52.4$ | $25.40$ | $26.6$ | $26.37$ | $28.27$ | $28.01$ hr | $61.98$ | $63.4$ | $63.79$ | $71.90$ | $65.31$ | $63.11$ | $63.7$ | $51.87$ | $54.2$ | $54.18$ | $61.17$ | $53.62$ he | $56.62$ | $59.2$ | $60.49$ | $64.17$ | $64.80$ | $57.15$ | $58.8$ | $47.61$ | $50.5$ | $51.11$ | $53.92$ | $49.36$ average | $43.88$ | $44.95$ | $45.51$ | $52.41$ | $52.40$ | — | — | $32.72$ | $34.24$ | $34.37$ | $40.26$ | — bg | $77.68$ | $80.0$ | $81.22$ | $81.92$ | $73.57$ | $79.72$ | $79.7$ | $66.22$ | $68.9$ | $70.02$ | $70.24$ | $68.39$ it | $77.89$ | $79.7$ | $81.36$ | $83.65$ | $70.68$ | $80.70$ | $82.0$ | $71.07$ | $74.0$ | $75.53$ | $77.73$ | $75.57$ pt | $74.07$ | $76.3$ | $77.07$ | $80.95$ | $66.61$ | $77.09$ | $77.5$ | $65.05$ | $67.6$ | $68.26$ | $70.64$ | $67.81$ fr | $74.80$ | $77.5$ | $78.64$ | $80.57$ | $67.66$ | $78.31$ | $79.1$ | $68.13$ | $71.7$ | $72.76$ | $74.46$ | $73.29$ es† | $72.45$ | $74.9$ | $75.21$ | $78.25$ | $64.28$ | $74.08$ | $75.8$ | $63.78$ | $66.5$ | $67.01$ | $69.15$ | $65.84$ no | $77.86$ | $80.4$ | $81.21$ | $80.01$ | $65.29$ | $80.98$ | $80.4$ | $69.06$ | $71.9$ | $72.66$ | $71.75$ | $73.10$ da | $75.33$ | $76.0$ | $77.29$ | $76.57$ | $61.08$ | $76.25$ | $76.6$ | $66.25$ | $67.4$ | $68.55$ | $67.86$ | $68.03$ sv | $78.89$ | $80.5$ | $82.09$ | $81.03$ | $64.43$ | $80.43$ | $80.5$ | $71.12$ | $72.7$ | $74.20$ | $72.72$ | $76.68$ nl | $68.00$ | $68.9$ | $69.85$ | $74.39$ | $61.72$ | $69.23$ | $67.6$ | $59.47$ | $60.7$ | $61.53$ | $65.39$ | $60.51$ de | $66.79$ | $69.9$ | $69.52$ | $74.05$ | $69.52$ | $71.05$ | $70.8$ | $56.40$ | $60.0$ | $59.69$ | $63.45$ | $61.84$ average | $74.38$ | $76.41$ | $77.35$ | $79.14$ | $66.48$ | $76.78$ | $77.0$ | $65.66$ | $68.14$ | $69.02$ | $70.34$ | $69.11$ Table 1: Test UAS and LAS (avg. 5 runs) on distant (top) and nearby (bottom) languages, sorted from most distant (fa) to closest (de) to English. PPTX is trained in a leave-one-out fashion. The numbers for He, Ahmad, and Meng are obtained from the corresponding papers, direct transfer (DT) and self-training (ST) are based on our own implementation. † indicates languages used for hyper-parameter tuning, and thus have additional supervision through the use of a labelled development set. Table 1 shows the main results. We observe that fine-tuning via self-training already helps DT, and by incorporating multiple high probability trees with PPT, we can push the performance slightly higher on most languages, especially the nearby ones. Although not shown in the table, we also find the PPT has up to 6x lower standard deviation than ST, which makes PPT preferrable to ST. Thus, we exclude ST as a baseline from our subsequent experiments. Our results seem to agree with that of Täckström et al. (2013) and suggest that PPT can also be employed for neural parsers. Therefore, it should be considered for target language adaptation if unlabelled target data is available. Comparing to He (He et al., 2019), PPT performs worse on distant languages, but better on nearby languages. This finding means that if the target language has a closely related high-resource language, it may be better to transfer from that language as the source and use PPT for adaptation. Against Ahmad (Ahmad et al., 2019b), PPT performs better on 4 out of 6 distant languages. On nearby languages, the average UAS of PPT is higher, and the average LAS is on par. This result shows that leveraging unlabelled data for cross-lingual parsing without access to the source data is feasible. PPT also performs better than Meng (Meng et al., 2019) on 4 out of 7 distant languages, and slightly better on average on nearby languages. This finding shows that PPT is competitive to their constrained inference method. Also reported in Table 1 are the ensemble results for PPTX, which are particularly strong. PPTX outperforms PPT, especially on distant languages with the average UAS and LAS absolute improvements of $7\text{\,}\mathrm{\char 37\relax}$ and $6\text{\,}\mathrm{\char 37\relax}$ respectively. This finding suggests that PPTX is indeed an effective method for multi-source transfer of neural dependency parsers. It also gives further evidence that multi-source transfer is better than the single-source counterpart. PPTX also closes the gap against the state-of-the-art adaptation of He et al. (2019) in terms of average UAS on distant languages. This result suggests that PPTX can be an option for languages that do not have a closely related high-resource language to transfer from. #### Treebank Leakage Figure 2: Relationship between treebank leakage and LAS for PPTX. Shaded area shows $95\text{\,}\mathrm{\char 37\relax}$ confidence interval. Korean and Turkish (in red) are excluded when computing the regression line. The success of our cross-lingual transfer can be attributed in part to treebank leakage, which measures the fraction of dependency trees in the test set that are isomorphic to a tree in the training set (with potentially different words); accordingly these trees are not entirely unseen. Such leakage has been found to be a particularly strong predictor for parsing performance in monolingual parsing (Søgaard, 2020). Fig. 2 shows the relationship between treebank leakage and parsing accuracy, where the leakage is computed between the English training set as source and the target language’s test set. Excluding outliers which are Korean and Turkish because of their low parsing accuracy despite the relatively high leakage, we find that there is a fairly strong positive correlation ($r=0.57$) between the amount of leakage and accuracy. The same trend occurs with DT, ST, and PPT. This finding suggests that cross-lingual parsing is also affected by treebank leakage just like monolingual parsing is, which may present an opportunity to find good sources for transfer. #### Use of Gold POS Tags As we explained in Section 3.1, we restrict our experiments to gold POS tags for comparison with prior work. However, the use of gold POS tags does not reflect a realistic low-resource setting where one may have to resort to automatically predicted POS tags. Tiedemann (2015) has shown that cross- lingual delexicalised parsing performance degrades when predicted POS tags are used. The degradation ranges from 2.9 to 8.4 LAS points depending on the target language. Thus, our reported numbers in Table 1 are likely to decrease as well if predicted tags are used, although we expect the decline is not as sharp because our parser is lexicalised. ### 3.4 Parsimonious Selection of Sources for PPTX Figure 3: Comparison of selection of source languages for PPTX on distant and nearby languages, sorted from most distant (fa) to closest (de) to English. PPTX-LOO is trained in a leave-one-out fashion. PPTX-REPR uses the representative source language set, while PPTX-PRAG is adapted from five high- resource languages. A source language is excluded from the source if it is also the target language. In our main experiment, we use all available languages as source for PPTX in a leave-one-out setting. Such a setting may be justified to cover as many syntactic biases as possible, however, training dozens of parses may be impractical. In this experiment, we consider the case where we can train only a handful of source parsers. We investigate two selections of source languages: (1) a representative selection (PPTX-REPR) which covers as many language families as possible and (2) a pragmatic selection (PPTX-PRAG) containing truly high-resource languages for which quality pre-trained parsers are likely to exist. We restrict the selections to 5 languages each. For PPTX- REPR, we use English, Spanish, Arabic, Indonesian, and Korean as source languages. This selection covers Indo-European (Germanic and Romance), Afro- Asiatic, Austronesian, and Koreanic language families respectively. We use English, Spanish, Arabic, French, and German as source languages for PPTX- PRAG. The five languages are classified as exemplary high-resource languages by Joshi et al. (2020). We exclude a language from the source if it is also the target language, in which case there will be only 4 source languages. Other than that, the setup is the same as that of our main experiment.888Hyperparameters are tuned; values are shown in Table 5. We present the result in Fig. 3 where we also include the results for PPT, and PPTX with the leave-one-out setting (PPTX-LOO). We report only LAS since UAS shows a similar trend. We observe that both PPTX-REPR and PPTX-PRAG outperform PPT overall. Furthermore, on nearby languages except Dutch and German, both PPTX-REPR and PPTX-PRAG outperform PPTX-LOO, and PPTX-PRAG does best overall. In contrast, no systematic difference between the three PPTX variants emerges on distant languages. This finding suggests that instead of training dozens of source parsers for PPTX, training just a handful of them is sufficient, and a “pragmatic” selection of a small number of high-resource source languages seems to be an efficient strategy. Since pre-trained parsers for these languages are most likely available, it comes with the additional advantage of alleviating the need to train parsers at all, which makes our method even more practical. #### Analysis on Dependency Labels Figure 4: Comparison of direct transfer (DT), PPT, and PPTX-PRAG on select dependency labels of Indonesian (top) and German (bottom). Next, we break down the performance of our methods based on the dependency labels to study their failure and success patterns. Fig. 4 shows the UAS of DT, PPT, and PPTX-PRAG on Indonesian and German for select dependency labels. Looking at Indonesian, PPT is slightly worse than DT in terms of overall accuracy scores (Table 1), and this is reflected across dependency labels. However, we see in Fig. 4 that PPT outperforms DT on amod. In Indonesian, adjectives follow the noun they modify, while in English the opposite is true in general. Thus, unsupervised target language adaptation seems able to address these kinds of discrepancy between the source and target language. We find that PPTX-PRAG outperforms both DT and PPT across dependency labels, especially on flat and compound labels as shown in Fig. 4. Both labels are related to multi-word expressions (MWEs), so PPTX appears to improve parsing MWEs in Indonesian significantly. For German we find that both PPT and PPTX-PRAG outperform DT on most dependency labels, with the most notable gain on nmod, which appear in diverse, and often non-local relations in both languages many of which do not structurally translate, and fine-tuning improves performance as expected. Also, we see PPTX-PRAG significantly underperforms on compound while PPT is better than DT. German compounds are often merged into a single token, and self-training appears to alleviate over-prediction of such relations. The multi-source case may contain too much diffuse signal on compound and thus the performance is worse than that of DT. We find that PPT and PPTX improves over DT on mark, likely because markers are often used in places where German deviates from English by becoming verb-final (e.g., subordinate clauses). Both PPT and PPTX-PRAG seem able to learn this characteristic as shown by their performance improvements. This analysis suggests that the benefits of self- training depend on the syntactic properties of the target language. ### 3.5 Effect of Projectivity Model | Target | AVG ---|---|--- id | hr | fr | nl Non-projective DT | $45.2$ | $51.9$ | $68.1$ | $59.5$ | $56.2$ PPT | $44.7$ | $54.2$ | $72.8$ | $61.5$ | $58.3$ PPTX-PRAG | $57.4$ | $62.2$ | $77.9$ | $66.4$ | $66.0$ Projective DT | $45.7$ | $52.1$ | $68.4$ | $59.6$ | $56.4$ PPT | $45.0$ | $54.0$ | $72.3$ | $61.7$ | $58.3$ PPTX-PRAG | $57.5$ | $61.1$ | $78.1$ | $67.7$ | $66.1$ Table 2: Comparison of projective and non-projective direct transfer (DT), PPT, and PPTX-PRAG. Scores are LAS, averaged over 5 runs. In this experiment, we study the effect of projectivity on the performance of our methods. We emulate a projective parser by restricting the trees in $\tilde{Y}(x)$ to be projective. In other words, the sum in Eq. 1 is performed only over projective trees. At test time, we search for the highest scoring projective tree. We compare DT, PPT, and PPTX-PRAG, and report LAS on Indonesian (id) and Croatian (hr) as distant languages, and on French (fr) and Dutch (nl) as nearby languages. The trend for UAS and on the other languages is similar. We use the dynamic programming implementation provided by torch- struct for the projective case (Rush, 2020). We find that it consumes more memory than our Matrix-Tree Theorem implementation, so we set the length cutoff to 20 tokens.999Hyperparameters are tuned; values are shown in Table 5. Table 2 shows result of our experiment, which suggests that there is no significant performance difference between the projective and non-projective variant of our methods. This result suggests that our methods generalise well to both projective and non-projective parsing. That said, we recommend the non-projective variant as it allows better parsing of languages that are predominantly non-projective. Also, we find that it runs roughly 2x faster than the projective variant in practice. ### 3.6 Disentangling the Effect of Ensembling and Larger Data Size Model | Target ---|--- ar | es DT | $28.09$ | $64.11$ PPT | $30.84$ | $67.27$ PPTXEN5 | $30.92$ | $66.25$ PPTX-PRAGS | $36.46$ | $70.32$ PPTX-PRAG | $36.45$ | $71.88$ Table 3: Comparison of LAS on Arabic and Spanish on the development set, averaged over 5 runs. PPTXEN5 is PPTX with 5 English parsers as source, each trained on 1/5 size of the English corpus. PPTX-PRAGS is PPTX with the pragmatic selection of source languages (PPTX-PRAG) but each source parser is trained on the same amount of data as PPTXEN5. The effectiveness of PPTX can be attributed to at least three factors: (1) the effect of ensembling source parsers (ensembling), (2) the effect of larger data size used for training the source parsers (data), and (3) the diversity of syntactic biases from multiple source languages (multilinguality). In this experiment, we investigate to what extent each of those factors contributes to the overall performance. To this end, we design two additional comparisons: PPTXEN5 and PPTX-PRAGS. PPTXEN5 is PPTX with only English source parsers, where each parser is trained on 1/5 of the English training set. That is, we randomly split the English training set into five equal-sized parts, and train a separate parser on each. These parsers then serve as the source parsers for PPTXEN5. Thus, PPTXEN5 has the benefit of ensembling but not data and multilinguality compared with PPT. PPTX-PRAGS is PPTX whose source language selection is the same as PPTX-PRAG, but each source parser is trained on the training data whose size is roughly the same as that of the training data of PPTXEN5 source parsers. In other words, the training data size is roughly equal to 1/5 of the English training set. To obtain this data, we randomly sub-sample the training data of each source language to the appropriate number of sentences. Therefore, PPTX-PRAGS has the benefit of ensembling and multilinguality but not data. Table 3 reports their LAS on the development set of Arabic and Spanish, averaged over five runs. We also include the results of PPTX-PRAG that enjoys all three benefits. We observe that PPT and PPTXEN5 perform similarly on Arabic, and PPTXEN5 has a slightly lower performance on Spanish. This result suggests a negligable effect of ensembling on performance. On the other hand, PPTX-PRAGS outperforms PPTXEN5 remarkably, with approximately $6\text{\,}\mathrm{\char 37\relax}$ and $4\text{\,}\mathrm{\char 37\relax}$ LAS improvement on Arabic and Spanish respectively, showing that multilinguality has a much larger effect on performance than ensembling. Lastly, we see that PPTX-PRAG performs similarly to PPTX-PRAGS on Arabic, and about $1.6\text{\,}\mathrm{\char 37\relax}$ better on Spanish. This result demonstrates that data size has an effect, albeit a smaller one compared to multilinguality. To conclude, the effectiveness of PPTX can be attributed to the diversity contributed through multiple languages, and not to ensembling or larger source data sets. ## 4 Related Work Cross-lingual dependency parsing has been extensively studied in NLP. The approaches can be grouped into two main categories. On the one hand, there are approaches that operate on the data level. Examples of this category include annotation projection, which aims to project dependency trees from a source language to a target language (Hwa et al., 2005; Li et al., 2014; Lacroix et al., 2016; Zhang et al., 2019); and source treebank reordering, which manipulates the source language treebank to obtain another treebank whose statistics approximately match those of the target language (Wang and Eisner, 2018; Rasooli and Collins, 2019). Both methods have no restriction on the type of parsers as they are only concerned with the data. Transferring from multiple source languages with annotation projection is also feasible (Agić et al., 2016). Despite their effectiveness, these data-level methods may require access to the source language data, hence are unusable when it is inaccessible due to privacy or legal reasons. In such source-free transfer, only a model pre- trained on the source language data is available. By leveraging parallel data, annotation projection is indeed feasible without access to the source language data. That said, parallel data is limited for low-resource languages or may have a poor domain match. Additionally, these methods involve training the parser from scratch for every new target language, which may be prohibitive. On the other hand, there are methods that operate on the model level. A typical approach is direct transfer (aka., zero-shot transfer) which trains a parser on source language data, and then directly uses it to parse a target language. This approach is enabled by the shared input representation between the source and target language such as POS tags (Zeman and Resnik, 2008) or cross-lingual embeddings (Guo et al., 2015; Ahmad et al., 2019a). Direct transfer supports source-free transfer and only requires training a parser once on the source language data. In other words, direct transfer is unsupervised as far as target language resources. Previous work has shown that unsupervised target language adaptation outperforms direct transfer. Recent work by He et al. (2019) used a neural lexicalised dependency model with valence (DMV) (Klein and Manning, 2004) as the source parser and fine-tuned it in an unsupervised manner on the unlabelled target language data. This adaptation method allows for source-free transfer and performs especially well on distant target languages. A different approach is proposed by Meng et al. (2019), who gathered target language corpus statistics to derive constraints to guide inference using the source parser. Thus, this technique also allows for source-free transfer. A different method is proposed by Ahmad et al. (2019b) who explored the use of unlabelled data from an auxiliary language, which can be different from the target language. They employed adversarial training to learn language-agnostic representations. Unlike the others, this method can be extended to support multi-source transfer. An older method is introduced by Täckström et al. (2013), who leveraged ambiguity-aware training to achieve unsupervised target language adaptation. Their method is usable for both source-free and multi- source transfer. However, to the best of our knowledge, its use for neural dependency parsing has not been investigated. Our work extends theirs by employing it for the said purpose. The methods of both He et al. (2019) and Ahmad et al. (2019b) have several limitations. The method of He et al. (2019) requires the parser to be generative and projective. Their generative parser is quite impoverished with an accuracy that is $21$ points lower than a state-of-the-art discriminative arc-factored parser on English. Thus, their choice of generative parser may constrain its potential performance. Furthermore, their method performs substantially worse than direct transfer on nearby target languages. Because of the availability of resources such as Universal Dependency Treebanks (Nivre et al., 2018), it is likely that a target language has a closely related high- resource language which can serve as the source language. Therefore, performing well on nearby languages is more desirable pragmatically. On top of that, it is unclear how to employ this method for multi-source transfer. The adversarial training method of Ahmad et al. (2019b) does not suffer from the aforementioned limitations but is unusable for source-free transfer. That is, it assumes access to the source language data, which may not always be feasible due to privacy or legal reasons. ## 5 Conclusions This paper presents a set of effective, flexible, and conceptually simple methods for unsupervised cross-lingual dependency parsing, which can leverage the power of state-of-the-art pre-trained neural network parsers. Our methods improve over direct transfer and strong recent unsupervised transfer models, by using source parser uncertainty for implicit supervision, leveraging only unlabelled data in the target language. Our experiments show that the methods are effective for both single-source and multi-source transfer, free from the limitations of recent transfer models, and perform well for non-projective parsing. Our analysis shows that the effectiveness of the multi-source transfer method is attributable to its ability to leverage diverse syntactic signals from source parsers from different languages. Our findings motivate future research into advanced methods for generating informative sets of candidate trees given one or more source parsers. ## Acknowledgments We thank the anonymous reviewers for the useful feedback. A graduate research scholarship is provided by Melbourne School of Engineering to Kemal Kurniawan. ## References * Agić et al. (2016) Željko Agić, Anders Johannsen, Barbara Plank, Héctor Martínez Alonso, Natalie Schluter, and Anders Søgaard. 2016. Multilingual projection for parsing truly low-resource languages. _Transactions of the Association for Computational Linguistics_ , 4:301–312. * Ahmad et al. (2019a) Wasi Ahmad, Zhisong Zhang, Xuezhe Ma, Eduard Hovy, Kai-Wei Chang, and Nanyun Peng. 2019a. On difficulties of cross-lingual transfer with order differences: A case study on dependency parsing. In _Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers)_ , pages 2440–2452. * Ahmad et al. (2019b) Wasi Uddin Ahmad, Zhisong Zhang, Xuezhe Ma, Kai-Wei Chang, and Nanyun Peng. 2019b. Cross-lingual dependency parsing with unlabeled auxiliary languages. In _Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL)_ , pages 372–382. * Bojanowski et al. (2017) Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2017. Enriching word vectors with subword information. _Transactions of the Association for Computational Linguistics_ , 5:135–146. * Chu and Liu (1965) Yoeng-Jin Chu and Tseng-Hong Liu. 1965. On the shortest arborescence of a directed graph. _Scientia Sinica_ , 14:1396–1400. * de Lhoneux (2019) Miryam de Lhoneux. 2019. _Linguistically Informed Neural Dependency Parsing for Typologically Diverse Languages_. Ph.d. thesis, Uppsala University. * Duong et al. (2015a) Long Duong, Trevor Cohn, Steven Bird, and Paul Cook. 2015a. Cross-lingual transfer for unsupervised dependency parsing without parallel data. In _Proceedings of the Nineteenth Conference on Computational Natural Language Learning_ , pages 113–122. * Duong et al. (2015b) Long Duong, Trevor Cohn, Steven Bird, and Paul Cook. 2015b. Low resource dependency parsing: Cross-lingual parameter sharing in a neural network parser. In _Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 2: Short Papers)_ , pages 845–850. * Edmonds (1967) Jack Edmonds. 1967. Optimum branchings. _Journal of Research of the national Bureau of Standards B_ , 71(4):233–240. * Eisner (1996) Jason M. Eisner. 1996. Three new probabilistic models for dependency parsing: An exploration. In _COLING 1996 Volume 1: The 16th International Conference on Computational Linguistics_. * Guo et al. (2015) Jiang Guo, Wanxiang Che, David Yarowsky, Haifeng Wang, and Ting Liu. 2015. Cross-lingual dependency parsing based on distributed representations. In _Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)_ , pages 1234–1244. * He et al. (2019) Junxian He, Zhisong Zhang, Taylor Berg-Kirkpatrick, and Graham Neubig. 2019. Cross-lingual syntactic transfer through unsupervised adaptation of invertible projections. In _Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics_ , pages 3211–3223. * Hwa et al. (2005) Rebecca Hwa, Philip Resnik, Amy Weinberg, Clara Cabezas, and Okan Kolak. 2005. Bootstrapping parsers via syntactic projection across parallel texts. _Natural Language Engineering_ , 11(3):311–325. * Joshi et al. (2020) Pratik Joshi, Sebastin Santy, Amar Budhiraja, Kalika Bali, and Monojit Choudhury. 2020. The state and fate of linguistic diversity and inclusion in the NLP world. In _Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics_. * Klein and Manning (2004) Dan Klein and Christopher Manning. 2004. Corpus-based induction of syntactic structure: Models of dependency and constituency. In _Proceedings of the 42nd Annual Meeting of the Association for Computational Linguistics (ACL-04)_ , pages 478–485. * Koo et al. (2007) Terry Koo, Amir Globerson, Xavier Carreras, and Michael Collins. 2007. Structured prediction models via the matrix-tree theorem. In _Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL)_ , pages 141–150. * Lacroix et al. (2016) Ophélie Lacroix, Lauriane Aufrant, Guillaume Wisniewski, and François Yvon. 2016. Frustratingly easy cross-lingual transfer for transition-based dependency parsing. In _Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies_ , pages 1058–1063. * Li et al. (2014) Zhenghua Li, Min Zhang, and Wenliang Chen. 2014. Soft cross-lingual syntax projection for dependency parsing. In _Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers_ , pages 783–793. * McDonald et al. (2005) Ryan McDonald, Fernando Pereira, Kiril Ribarov, and Jan Hajič. 2005. Non-projective dependency parsing using spanning tree algorithms. In _Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing_ , pages 523–530. * McDonald et al. (2011) Ryan McDonald, Slav Petrov, and Keith Hall. 2011. Multi-source transfer of delexicalized dependency parsers. In _Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing_ , pages 62–72. * McDonald and Satta (2007) Ryan McDonald and Giorgio Satta. 2007. On the complexity of non-projective data-driven dependency parsing. In _Proceedings of the Tenth International Conference on Parsing Technologies_ , pages 121–132. * Meng et al. (2019) Tao Meng, Nanyun Peng, and Kai-Wei Chang. 2019. Target language-aware constrained inference for cross-lingual dependency parsing. In _Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)_ , pages 1117–1128. * Nivre et al. (2018) Joakim Nivre, Mitchell Abrams, Željko Agić, and et al. 2018. Universal dependencies 2.2. LINDAT/CLARIAH-CZ digital library at the Institute of Formal and Applied Linguistics (ÚFAL), Faculty of Mathematics and Physics, Charles University. * Paskin (2001) Mark A Paskin. 2001. _Cubic-time parsing and learning algorithms for grammatical bigram models_. * Rahimi et al. (2019) Afshin Rahimi, Yuan Li, and Trevor Cohn. 2019. Massively multilingual transfer for NER. In _Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics_ , pages 151–164. * Rasooli and Collins (2019) Mohammad Sadegh Rasooli and Michael Collins. 2019. Low-resource syntactic transfer with unsupervised source reordering. In _Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers)_ , pages 3845–3856. * Rush (2020) Alexander Rush. 2020. Torch-struct: Deep structured prediction library. In _Proceedings of the 58th annual meeting of the association for computational linguistics: System demonstrations_ , pages 335–342. * Smith and Smith (2007) David A. Smith and Noah A. Smith. 2007. Probabilistic models of nonprojective dependency trees. In _Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL)_ , pages 132–140. * Smith et al. (2017) Samuel L. Smith, David H. P. Turban, Steven Hamblin, and Nils Y. Hammerla. 2017\. Offline bilingual word vectors, orthogonal transformations and the inverted softmax. In _International Conference on Learning Representations_. * Søgaard (2020) Anders Søgaard. 2020. Some languages seem easier to parse because their treebanks leak. In _Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)_ , pages 2765–2770. * Täckström et al. (2013) Oscar Täckström, Ryan McDonald, and Joakim Nivre. 2013. Target language adaptation of discriminative transfer parsers. In _Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies_ , pages 1061–1071. * Tiedemann (2015) Jörg Tiedemann. 2015. Cross-lingual dependency parsing with universal dependencies and predicted PoS labels. In _Proceedings of the Third International Conference on Dependency Linguistics (Depling 2015)_ , pages 340–349. * Wang and Eisner (2018) Dingquan Wang and Jason Eisner. 2018. Synthetic data made to order: The case of parsing. In _Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing_ , pages 1325–1337. * Zeman and Resnik (2008) Daniel Zeman and Philip Resnik. 2008. Cross-language parser adaptation between related languages. In _Proceedings of the IJCNLP-08 Workshop on NLP for Less Privileged Languages_. * Zhang et al. (2019) Meishan Zhang, Yue Zhang, and Guohong Fu. 2019. Cross-lingual dependency parsing using code-mixed treebank. In _Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)_ , pages 996–1005. ## Appendix A Hyperparameter values Here we report the hyperparameter values for experiments presented in the paper. Table 4 shows the hyperparameter values of our English source parser explained in Section 3.1. Table 5 reports the tuned hyperparameter values for our experiments shown in Table 1, Fig. 3, and Table 2. Hyperparameter | Value ---|--- Sentence length cutoff | 100 Word embedding size | 300 POS tag embedding size | 50 Number of attention heads | 10 Number of Transformer layers | 6 Feedforward layer hidden size | 512 Attention key vector size | 64 Attention value vector size | 64 Dropout | 0.2 Dependency arc vector size | 512 Dependency label vector size | 128 Batch size | 80 Learning rate | ${10}^{-4}$ Early stopping patience | 50 Table 4: Hyperparameter values of the source parser. Hyperparameter | Value ---|--- Nearby | Distant ST Sentence length cutoff | 60 | 60 Learning rate | $5.6\text{\times}{10}^{-4}$ | $3.7\text{\times}{10}^{-4}$ L2 coefficient ($\lambda$) | $3\text{\times}{10}^{-4}$ | $2.8\text{\times}{10}^{-4}$ PPT Learning rate | $3.8\text{\times}{10}^{-5}$ | $2\text{\times}{10}^{-5}$ L2 coefficient ($\lambda$) | $0.01$ | $0.39$ PPTX/PPTX-LOO Learning rate | $2.1\text{\times}{10}^{-5}$ | $5.9\text{\times}{10}^{-5}$ L2 coefficient ($\lambda$) | $0.079$ | $1.2\text{\times}{10}^{-4}$ PPTX-REPR Learning rate | $1.7\text{\times}{10}^{-5}$ | $9.7\text{\times}{10}^{-5}$ L2 coefficient ($\lambda$) | $4\text{\times}{10}^{-4}$ | $0.084$ PPTX-PRAG Learning rate | $4.4\text{\times}{10}^{-5}$ | $8.5\text{\times}{10}^{-5}$ L2 coefficient ($\lambda$) | $2.7\text{\times}{10}^{-4}$ | $2.8\text{\times}{10}^{-5}$ Projective PPT Sentence length cutoff | 20 | 20 Learning rate | ${10}^{-4}$ | ${10}^{-4}$ L2 coefficient ($\lambda$) | $7.9\text{\times}{10}^{-4}$ | $7.9\text{\times}{10}^{-4}$ Projective PPTX-PRAG Sentence length cutoff | 20 | 20 Learning rate | $9.4\text{\times}{10}^{-5}$ | $9.4\text{\times}{10}^{-5}$ L2 coefficient ($\lambda$) | $2.4\text{\times}{10}^{-4}$ | $2.4\text{\times}{10}^{-4}$ Table 5: Hyperparameter values of ST, PPT, PPTX, PPTX-REPR, PPTX-PRAG, projective PPT, and projective PPTX-PRAG. Sentence length cutoff for PPT, PPTX, PPTX-REPR, and PPTX-PRAG is 30, as explained in Section 3.1.
# What We Can Learn From Visual Artists About Software Development Jingyi Li Stanford University , Sonia Hashim University of California, Santa Barbara and Jennifer Jacobs University of California, Santa Barbara (2021) ###### Abstract. This paper explores software’s role in visual art production by examining how artists use and develop software. We conducted interviews with professional artists who were collaborating with software developers, learning software development, and building and maintaining software. We found artists were motivated to learn software development for intellectual growth and access to technical communities. Artists valued efficient workflows through skilled manual execution and personal software development, but avoided high-level forms of software automation. Artists identified conflicts between their priorities and those of professional developers and computational art communities, which influenced how they used computational aesthetics in their work. These findings contribute to efforts in systems engineering research to integrate end-user programming and creativity support across software and physical media, suggesting opportunities for artists as collaborators. Artists’ experiences writing software can guide technical implementations of domain-specific representations, and their experiences in interdisciplinary production can aid inclusive community building around computational tools. visual art, software development, creativity support tools ††copyright: iw3c2w3††journalyear: 2021††conference: CHI Conference on Human Factors in Computing Systems; May 8–13, 2021; Yokohama, Japan††booktitle: CHI Conference on Human Factors in Computing Systems (CHI ’21), May 8–13, 2021, Yokohama, Japan††doi: 10.1145/3411764.3445682††isbn: 978-1-4503-8096-6/21/05††ccs: Human-centered computing User interface programming††ccs: Applied computing Media arts ## 1\. Introduction Software has become closely integrated with visual art production. Illustration, painting, animation, and photography are just a small sample of once non-digital fields where many practitioners now incorporate software. Like all creative tools, software offers artists new opportunities while simultaneously imposing constraints. For example, digital drawing software allows artists to continually edit their pieces through layers and undo, but limits them to the functionality provided on a toolbar and can make it difficult to integrate physical media. Developing software for visual art brings many specific challenges. Artists are distinguished from many other software users by their nonconformity and unconventional approaches to making artifacts (Feist, 1999). Artists also work extensively with physical media and manual forms of expression—qualities that are challenging to integrate in digital representations (Schachman, 2012; Devendorf and Ryokai, 2015). Designing software abstractions and constraints for different domains of visual art poses additional challenges. The ways software designers expect artists to interact with the system may be different from the ways artists are used to engaging with materials to make art in their domain (Victor, 2011). Furthermore, artists who use domain-specific creative coding languages (Reas and Fry, 2004; Lieberman, 2014) face additional challenges in having to understand abstract representations and work in a highly structured manner (Victor, 2011). These forms of working can be incompatible with their existing practices of manual manipulation and non- objective exploration (Berger, 2014; Goodman, 1968). These challenges, as well the opportunities they surface, affect a variety of stakeholders. They include systems engineering researchers who research creativity support technologies. They also include artists who collaborate with software developers in their own practice. Finally, they include artists who are software developers. Such individuals build personal tools, but also can support communities of creative practitioners by leveraging their domain knowledge to develop and distribute new, artist-specific platforms (Reas and Fry, 2004; Levin, 2003). To support multiple stakeholders in software production for visual art and address these challenges, we sought to develop a more detailed understanding of the ways visual artists work with software. By studying how artists develop and use their own forms of software, we synthesized design implications across the joint interests of artists and professional software developers. In this paper, we ask two research questions: (1) What factors lead artists to embrace or reject software tools created by others? (2) What factors motivate artists to develop their own software, and how do their approaches diverge or align with other domains of software development? Our research questions are motivated by our efforts to develop programming systems for visual artists over the past seven years and oriented towards informing future efforts in systems engineering research for creativity support. Like researchers, artists are motivated, in-part, by creating novel outcomes. We argue that artists’ motivations for building their own software are well aligned with the kinds of contributions valued by systems engineering for visual art production HCI research. Artists’ joint concerns around expressiveness, audience, and commercial viability, as well as their unique workflows, can inform approaches in end-user software development. But at the same time, artist communities have different values, norms, and forms of dissemination than those of professional software developers (Levin, 2015). Understanding and integrating these values is the first step in informing research partnerships between these communities. To investigate these questions and opportunities, we formalized our initial observations into 13 in-depth interviews conducted with artists across a spectrum of software development expertise, deliberately selected from a larger set of conversations over the past four years. Our interviews sought to understand how artists’ tools and materials–-digital, physical, and programmatic–impacted their processes and outcomes. This paper makes two contributions. First, through a thematic analysis of our interviews, we surface themes on how software intersects with visual art practice: how artists were motivated to build software as a mechanism for intellectual and community engagement, how existing software representations worked with or against artists’ complex workflows, how artists valued efficiency but resisted forms of software automation, how artists used software beyond making art to organize, motivate and reflect, and how interaction with technical communities impacted artists’ aesthetic choices. Second, our findings suggest opportunities for how efforts in end-user programming can map onto creativity support tools for visual artists. Current forms of software automation misalign with how artists work, while forms of higher level abstraction hinder their ability to manipulate data representations. We argue that artists, who innovate in adapting and writing software to fit their idiosyncratic working styles, can guide technical implementations of domain-specific program representations and that their experiences in interdisciplinary production can aid inclusive community-building around computational tools. ## 2\. Background Our research contributes to HCI efforts to inform technical systems through the study of creative communities. Because we studied how visual artists develop software, our work builds on and informs end-user programming research. To guide our analysis, we also examined artist-built programming tools and communities, and research-driven creativity support tools. ### 2.1. Informing HCI by Studying Creative Practice Understanding creative practice requires different forms of inquiry beyond controlled study (Shneiderman, 2007). By investigating art and craft practice, researchers have demonstrated alternative strategies for established domains of HCI. This includes studying ceramics to inform interaction design (Bardzell et al., 2012), furniture production to inform CAD and digital fabrication research (Cheatle and Jackson, 2015), and house construction to inform design for living materials (Dew and Rosner, 2018). Researchers have also used collaborative models working with artists or craftspeople to engage in joint production (Jacobs and Zoran, 2015) or co-author HCI publications (Devendorf et al., 2020; Fiebrink and Sonami, [n.d.]). These investigations challenge notions that artists or craftspeople have limited technical proficiency, demonstrate the technical sophistication of their work, and describe concrete ways in which they can inform technical systems. Inspired by prior efforts, we focus on how software use and development by visual artists can inform end- user programming and systems engineering research. Studies of art practices have also examined the role of manual production in creative workflows. HCI researchers have theorized that ideas emerge from interacting with materials (Ingold, 2010; Schon, [n.d.]), tangible making builds mental models (Papert, 1980; Tversky, 2019), and physical manipulation facilitates concrete cognitive tasks (Gross, 2009; Klemmer et al., 2006). Manipulating physical media helps artists develop manual skill and knowledge (Needleman, 1979), react and plan in step (Berger, 2014; Do and Gross, 2007), and reason about building tools (McCullough, 1996). Manual tools also afford expressiveness, by preserving gesture (Mumford, 1952), and speed, which supports open-ended exploration (Berger, 2014). In contrast to prior work on manual production, we investigate how efficiency and automation in software align or conflict with artists’ manual processes and aesthetic preferences. ### 2.2. Creative Coding Systems: Programming Tools & Communities for Visual Art The expressive power of programming has led visual artists to develop and disseminate their own programming tools. Lieberman, Watson, and Castro developed OpenFrameworks, a textual programming toolkit, to create interactive audiovisual artwork (Lieberman, 2014). The textual languages Processing and p5.js (Reas and Fry, 2016; McCarthy and Turner, [n.d.]a) emerged from exploring how to teach programming through the lens of art (Reas and Fry, 2016). Node-based visual programming frameworks like Max and vvvv apply ideas from signal processing for computational music towards producing computational artwork (Cycling74, 2017; vvvv group, 2017). Researchers in music technology and live-coding have developed domain-specific representations and tools (Eaglestone et al., 2001; Barbosa et al., 2018; Malloch et al., 2019). The goals of these domains differ from visual art. For instance, latency and scrubbing metaphors are less relevant to visual art. Research on high ceilings in computational music achieved via design metaphors of control intimacy of musical instruments (Wessel and Wright, 2001) is different from our work that also focuses on expert practitioners. Our focus is to study the ways visual artists build software as a means to inform creativity support systems research. Some artists who create software tools also engage in community building. From the OpenFrameworks design philosophy that prioritizes a “do it with others” approach to making art (Lieberman, 2014), to the Processing community that builds and maintains the platform’s many extensions (Reas and Fry, 2016), to the p5.js community statement that recognizes diverse contributions from new programmers (McCarthy and Turner, [n.d.]b), all of these frameworks rely on collaborative development from their communities. The School for Poetic Computation (SFPC), a school run by artists for artists who often build computational tools, is another example of an artist-led creative coding community that extends software use (Jacobs, 2018). Artist-developed software tools align with art practice and are shaped by community engagement. Our work provides greater detail on this process by examining how artists move from creating artifacts to authoring software and how artists’ software use is shaped by interactions with technical communities. ### 2.3. End-User Programming in Art and Design Research in end-user programming (EUP) supports non-professional programmers as they develop software to further their work (Lieberman et al., 2006) and has focused largely on interaction designers. Research has shown that visual designers seek programming tools that directly integrate with visual drawing tools (Myers et al., 2008) and use high-level tools mapped to specific tasks or glued with general purpose languages rather than learn new programming frameworks (Brandt et al., 2008). Systems like Juxtapose (Hartmann et al., 2008) and Interstate (Oney et al., 2014) improve programming for interaction designers through better version management and visualizations. Re-envisioning software as information substrates (Beaudouin-Lafon, 2017) that integrate data and application functionality supports greater software malleability and more varied forms of collaboration in web (Klokmose et al., 2015) and video editing (Klokmose et al., 2019). While there has been extensive EUP research targeting designers, less research examines EUP for visual art. Researchers have developed a variety of graphic art tools that enable programming through direct manipulation. Some systems support two pane interfaces that place visual output side-by-side with code (Chugh et al., 2016; McDirmid, 2016). Recent work demonstrated that allowing users to directly manipulate intermediate execution values, in addition to output, minimized textual editing (Hempel et al., 2019). Other work, like Dynamic Brushes, aims to support expressiveness through a direct manipulation environment coupled with a programming language (Jacobs et al., 2018). Results from a study of debugging tools developed for Dynamic Brushes suggested that artists inspect program functionality while making artwork (Li et al., 2020). Our research is aimed at informing future efforts in EUP for visual art by investigating how artists approach software development and work with software representations. We provide insights into the ways that visual artists’ objectives differ from other end-user programmers and highlight opportunities in building domain-specific programming tools for visual art. ### 2.4. Creativity Support Tools for Visual Art Creativity support tools researchers have extensively studied how software can support visual art workflows. While Shneiderman outlined several opportunities for creativity support including exploratory discovery and record keeping (Shneiderman, 2002), many HCI systems for visual art emphasize producing artifacts. Systems such as large-scale generative design visualizations (Matejka et al., 2018; Zaman et al., 2015), cross-modal generative sketching for 3D design (Kazi et al., 2017), or text-based icon design (Zhao et al., 2020) aid practitioners in exploring ideas and selecting artifacts. Researchers have also explored supporting specific affordances in tools such as speed-aware line smoothing algorithms (Thiel et al., 2011), manipulations of negative space to edit vector graphics (Bernstein and Li, 2015), or selective undo in digital painting (Myers et al., 2015). Another category of systems examine ways to digitally emulate physical forms of production (Barnes et al., 2008; Kazi et al., 2011; Leung and Lara, 2015). A large body of creativity support research focuses on broadening participation. Shneiderman, as well as Silverman and Resnick (Resnick and Silverman, 2005), advocated for creative systems with “low-floors” that reduce the barriers to entry and “wide walls” that support a diverse range of outcomes. Researchers have examined reducing barriers to making computational art through direct manipulation interfaces for creating procedural constraints (Jacobs et al., 2017; Kazi et al., 2014). Other systems guide novices in tasks like photographic alignment (E et al., 2020) and narrative video composition (Kim et al., 2015). Machine learning (ML) based systems including applications of neural style transfer (NST) (Gatys et al., 2015; Champandard, 2016; Iizuka et al., 2016), user-in-the-loop tools (Runway AI, 2020), or support for specific automated tasks like auto-complete for hand-drawn animation (Xing et al., 2015), sketch simplification (Simo-Serra et al., 2018), and layout generation (Batra et al., 2019), are increasingly used to facilitate easy entry to visual art production through high-level automation. ML-based tools have raised new questions about the relationships between artists and software. Semmo et al. challenged the use of NST arguing that NST must support interactivity and customization for use in creative production (Semmo et al., 2017). Hertzmann critiqued the notion that ML-based automation creates AI- generated art, arguing that computational systems are not social agents and therefore cannot make art (Hertzmann, 2018). We seek to understand and critique high-level forms of software automation for visual art by examining the ways artists use or reject these systems in practice. ## 3\. Methods This work is structured around a thematic analysis of interviews with professional visual artists. These interviews were motivated and informed by the authors’ personal experiences working between systems engineering and fine art. ### 3.1. Author Backgrounds To understand the perspectives that shaped our work, we provide background on the research team’s expertise and focus. The authors represent a spectrum of art experience: Jingyi maintains a hobbyist painting and illustration practice, Sonia studied art history, and Jennifer has formal art training and worked as a professional artist. Presently, Jingyi and Sonia are graduate students in computer science and Jennifer is an HCI professor in an interdisciplinary art and engineering department. As we transitioned from creating artwork to researching and building software tools, we had the opportunity to test our tools with practicing artists. Our discussions about software use, programming, and creative production with artists indicated differences in how practicing artists and professional software developers viewed the opportunities of software. Name | Years Active | Description | Role ---|---|---|--- Molly Morin | 10+ | Digital fabrication artist | Studio artist, fine arts professor Lynne Yun | 15+ | Letterform designer | Designer at type firm Eran Hilleli | 10+ | Animator | Animator, software developer Miwa Matreyek | 10+ | Animator & dancer | Independent artist Emily Gobeille | 15+ | Interaction designer & illustrator | Studio founder Shantell Martin | 15+ | Large format & visual artist | Independent artist Michael Salter | 25+ | Studio artist | Studio artist, fine arts professor Chris Coleman | 15+ | Emergent media practitioner | Studio artist, fine arts professor Fish McGill | 8 | Pen & ink illustrator | Studio artist, fine arts professor Ben Tritt | 20+ | Painter | Studio founder Nina Wishnok | 10+ | Printmaker | Studio founder Kim Smith | 10+ | Painter & illustrator | Studio artist, art education company founder Mackenzie Schubert | 10+ | Illustrator | Studio artist, art technology company founder Table 1. Artist demographics. ### 3.2. Interview Methodology and Participants Building from this preliminary observation, we conducted a formal set of semi- structured interviews with 20 professional visual artists over a period of four years. We recruited subjects through our established networks drawn from the artist residencies, exhibitions, and educational networks that Jennifer participated in. We aimed to interview artists who worked across a diverse set of materials (e.g., software, paints, code), processes (e.g., coding, manual drawing, performance), and domains (e.g., illustration, animation, interactive art). Figure 1 shows sample artworks from each artist and Table 1 shows basic demographics information, with full artist names released with consent. While these interviews were interspersed with and helped guide research in building software tools, in this work, we foreground the insights from these interviews, rather than presenting a limited subset to motivate a specific system. For this paper, we included data from 13 out of the 20 interviews. We omitted interviews from artists who did not work with software or visual art as well as those in which the artist focused on their conceptual stances over their process. Our interviews were primarily in person, with five conducted through video conferencing software, and on average lasted 1.5 hours. Our interview objectives were to understand what artists perceive to be the primary opportunities and limitations of digital, manual, and physical media; how different media shaped artists’ learning, process, and outcomes; and what factors encouraged or prevented artists from engaging in software development in their work. Prior to each interview, we reviewed each artist’s work to direct process-based questions towards specific pieces. ### 3.3. Data Collection and Analysis We audio recorded and transcribed each interview. For analysis, we conducted a reflexive thematic analysis (Braun and Clarke, 2006, 2019), focusing on an inductive approach open to latent themes. Each author reviewed each transcript, and, following a discussion of initial patterns, each author coded a subset of transcripts to initially identify as many interesting data extracts as possible. The research team refined the codes and conceptualized them into preliminary themes through weekly meetings and discussions. After all authors collaboratively drafted a written description of each theme, Sonia and Jennifer reviewed them with respect to the coded transcripts to confirm they were representative of the original codes. ### 3.4. Limitations Our work focuses on a deep qualitative examination of 13 artists. This approach was necessary to gain insight into the specific factors that shaped the workflows of our interviewees. Artists represented here had a range of experiences with software; future research engaging exclusively with artists who build software will likely uncover additional details such as a quantitative breakdown of time spent across different software development tasks. We chose our methodology based on our personal experiences transitioning between art and software development and research, which we disclose in our author background statement to contextualize our analysis and discussion. Any interview risks social desirability bias. Because we had personal connections with our interviewees, trust contributed to them being comfortable giving honest answers, increasing the reliability of the responses (Fujii, 2017). ## 4\. Intersections of Software and Visual Art We conceptualized themes across six dimensions presented below. Revisiting our research questions, we describe how software constraints and representations, use cases beyond producing art, and cultural associations of computational aesthetics impacted how artists used software tools made by others. Artists were motivated to develop their own software to create new functionality in their works, grow intellectually, and gain technical legitimacy in their communities. Furthermore, their idiosyncratic workflows spanning digital and physical mediums were often misaligned with both the constraints and forms of automation provided by existing software. ### 4.1. Artists’ Motivations for Software Development Figure 1. Artwork created by interviewed artists: A) Interactive installation by Emily Gobeille. B) Oil painting by Ben Tritt. C) New media sculpture by Chris Coleman. D) Pencil illustration by Nina Wishnok. E) Line drawing mural by Shantell Martin. F) Typeface by Lynne Yun. G) Poster by Fish McGill. H) Comic page by Mackenzie Schubert. I) Digital illustration by Michael Salter in collaboration with Chris Coleman. J) Vinyl cut sculptures by Molly Morin. K) Synthesizer controlled animated character by Eran Hilleli. L) Painting by Kim Smith. M) 3D printed vector graphic by Michael Salter. N) Shadow installation by Miwa Matreyek. 14 pieces of artwork arranged in a grid. A: People interacting with a wall with a projected green background and plant-like structures. B: A textured monochromatic brown oil painting of a human figure. C: A digital sculpture of a person wearing a hat with web-like textures. D : An abstract pencil line drawing on light paper, with many geometric forms and an abstract house. E: An artist standing on a lift drawing a large line drawing mural on a wall. The mural has a white background and is of people’s faces. F: A poster demonstrating a font called “Trade Gothic Display Family” in all uppercase, sans-serif, bold lettering. G: A poster advertising Boston National Portfolio Day, using primarily purple, red, and blue, with many stylized people walking into a building. H: A three panel digital comic book page done entirely in shades of blue on a yellow background. A man is outside of a cluttered house pulling up something from a cable. I: A digital piece that morphs hand drawn illustrations with procedurally generated forms. The top half of the piece is various yellow lines with an indistinguishable vector graphic and the bottom half is white with some flowers. J: Vinyl cut sculptures, thin white floral shapes, hanging on the wall with shadows accentuating their form. K: On the top, a 3D modeled man with a beard, blue skin, and a yellow nose. On the bottom, the artist interacting with a synthesizer to control the man. L: An abstract painting, mainly greys, blues, and oranges, on a large canvas that has been deliberately crumpled and reshaped. M: A black 3D printed gun, except the barrel of the gun loops in a circle to the mouth of the gun. N: A wall showing a projection and the shadow of a human. The background is teal and there is a black and white anemone form obscuring part of the human shadow. Out of the 13 artists from our interviews, eight either were software developers or were in the process of learning software development. We distinguish between software development and programming in that programming involves writing code, while software development also encompasses testing, maintaining, and sharing software (Ko et al., 2011). Our software developers described a variety of motivating factors. Initially, artists developed software to add interactivity to their artwork. For example, Mackenzie learned Unity development to facilitate dynamic 3D transitions between panels in an interactive digital comic, and Emily learned C++ and Macromedia Director’s Lingo programming language because it enabled her to create interactive animations. When adding software as a component of their output, artists valued robustness. Artists also developed software to automate tasks. While using software as a tool for art-making, artists placed a higher emphasis on reusability. Molly learned to procedurally generate vector graphics to reduce manual labor when repeating forms in Adobe Illustrator, and Lynne learned Python to reduce the effort required to test different combinations of type when creating new fonts. These objectives align with the established framing of end-user programmers as individuals who write code in service of another practice. In addition to functional applications, artists developed software because of the opportunities for intellectual and creative growth. For instance, Ben described an interest diving into programming in order to “look under the hood” of software. Redesigning or building new software systems enabled artists to identify constraints in their tools to imagine alternatives. Lynne’s initial experience with Python acted as the catalyst for her to enroll in a creative programming course. She described how learning to use a C++ based framework to make her own graphical user interfaces (GUIs) enabled her to recognize third party software constraints and envision other options: > If I could make my own GUI for things, maybe I could be using things in a > different manner. This is a recent thought…I don’t think I ever realized how > much it was impacting my work. The notion that authoring software could expand one’s awareness of creative possibilities was remarkably consistent among the artists we spoke with, though artists varied how they used this idea in practice. Emily, Chris, and Eran all developed their own software interfaces that exposed specific parameters to explore and fine-tune visual properties of their work. Molly described the intellectual satisfaction she derived from translating her manual process generating vector geometry for fabrication to an algorithmic description, enjoying solving complex geometric problems while creating a reusable tool that reflected her manual practice. Similarly, Eran created experimental software tools primarily to investigate new concepts, and he was less concerned if the resulting tool would lead to a finished piece. Artists instead valued speed in designing software sketches to quickly test ideas. Finally, four artists described being motivated to develop software to influence future forms of software design, different from what they currently observed. Eran and Chris supported students and newcomers to animation and embedded programming, respectively, by designing tools that addressed obstacles they experienced in their own work. Lynne described how developing software would allow her to “have a seat at the table” around media software production. Likewise, Chris recognized that, as an independent artist and professor, he couldn’t compete with the speed and resources of professional software companies, but he could release different kinds of tools that influenced the direction those companies might take, saying, “[a]ll I can do is shape the conversation for the professional tools that get made afterwards.” Kim went a step further, describing her desire to improve her software development skills as a means to bypass negative experiences and communication breakdowns she had experienced when working directly with professional developers: > It’s not always clear to the developers that I’ve worked with why things are > important from a designer’s point of view …I think if I could really get > that skill down and design as well …the end results would be better. These experiences demonstrate how learning and participating in software development enabled multiple forms of power in visual art production. First, artists developed software to create new functionality in artworks. Second, artists developed software to grow intellectually and they built their own interfaces to explore or refine work. Finally, by demonstrating knowledge in software development, artists gained technical legitimacy and could engage in dialogue with professional software developers or circumvent them altogether. ### 4.2. Selecting Software Constraints and Representations Our interviews revealed two ways software made by others impacted artists’ processes and outputs: they negotiated different forms of software constraints, and they carefully selected specific graphic representations. Artists viewed software constraints as constructive when they could define the constraints’ parameters. Sometimes, this involved choosing to not use a tool’s features. Shantell, Eran, Fish, Ben, and Mackenzie all described points when they constrained themselves from using software-based undo. This constraint replicated the quality of physical ink, forcing them to work with their mistakes or avoided breaking their flow of drawing with editing. Artists also imposed constraints on their practices through their formal knowledge of design and composition. Michael described how, in Illustrator, he manually laid out his compositions to follow grid structures but broke those structures at arbitrary points—a process that would have been more difficult if Illustrator enforced the grid constraint. In other instances, artists developed software to enact constraints. Emily, Eran, and Fish authored software tools that restricted a user’s ability to erase, define geometry, or select colors. Fish described programming a drawing tool that automatically faded past strokes over time: > I worked in Processing on creating drawing tools that would fade over time…I > was storing screenshots of every stroke, so then I could watch how someone’s > drawing came together on a loop…and that came out of experiences, just > looking for ways to get other people that are afraid of drawing, just to > jump in and try something. So, it was a combination of creating the software > that would capture each line and mark and letting it fade, so they can get a > sense of depth as they are working. While artists had positive experiences imposing their own constraints, they struggled with the stylistic constraints imposed by feature-rich commercial software tools. Despite their respect for and reliance on commercial tools like Adobe After Effects and Unity, Miwa, Molly, Nina, Eran, and Chris struggled with feeling like stylistic aspects of their work were, in Molly’s words, “predetermined by the program.” In part, this reaction was tied to the expectation for fine artists to create novel imagery. Miwa and Chris both described their need to obscure the means of development when using After Effects. The constraints of high-level software tools were also at odds with artists’ desires to enact custom workflows. Artists avoided “defaults” and “presets” for this reason. For example, Nina avoided Photoshop filters because they were “straight out of the box” and incompatible with her personal workflow, and Fish described feeling “stifled” by the aesthetic constraints of Illustrator defaults until he learned how to author custom brushes. Similar to constraints, artists embraced or rejected lower-level graphic representations—e.g., 3D meshes, Bezier curves, bitmaps—based on the extent that a representation supported their workflow. Both Michael and Molly worked primarily with vector graphics, despite being adept in other representations, because vectors were best suited to the curvilinear geometry and “clean” aesthetics of their work. Similarly, other artists deliberately rejected some digital graphic representations or were frustrated with the inconsistencies that emerged when they tried to blend two different representations. Emily described an extreme dislike for the “smoothness” of 3D graphics, and Mackenzie described how integrating work from Photoshop and Illustrator created a stylistic “gap.” In most cases, these tensions with software representations were not the result of nostalgia for physical media. Emily wasn’t interested in recreating traditional techniques in a digital format, stating her goal was to “push both traditional artistic practice and software tools out of their individual comfort zones, to create something that is unique and blurs the lines between the two.” Alternatively, Shantell, Michael, and Mackenzie felt digital representations actually had comparable aesthetic qualities to physical media. Instead, the degree to which artists were comfortable working with a digital representation was determined by how it supported their workflows. At a low level of individual mark making, Mackenzie described how Bezier curves afforded editing existing work (a process he described as “finicky”) whereas bitmap brushes pushed him to “[sketch] one thing once and move through it,” because the bitmap representation didn’t support the same level of editing after drawing. At a higher level, Eran described how the timeline representation in animation software required animators to painstakingly edit individual frames. As an animation instructor, he observed students transitioning between drawing the animation and adjusting the timeline to the point of fatigue: > The task the person has to do is finish his drawing, move the timeline, > [and] change something. I see students that, every time they do that, their > mind hurts…It’s a break in their flow. These observations led Eran to develop animation tools with continuous, rather than frame-based, timelines. He recognized that these different representations lead to trade-offs in workflows—a continuous timeline would afford speed and shift the focus to drawing, but a key-frame timeline enabled low-level but laborious control. Artists also considered aesthetics clearly shaped by a specific representation to be an indicator of novice work. Michael and Molly described how, as teachers, they worked with students to master software so that their own drawing style was preserved, rather than obscured by the qualities of vector graphics. Mackenzie described how representing applications as standalone packages restricted iteration and workflows across different tools: > In some ways those programs are really behind kind of walls and are not very > modular. The program is this just inside of this [indicating to software > window] and it has many little tools that exist inside of that and some of > those have like smaller tools that exist inside of that. So you’re kind of > using like all of these tools in Photoshop and all of these tools in > Illustrator…but when I think about the way I work…passing things back and > forth and iterating in different ways and quickly. I feel like you’ve got > these big walls between programs…I’m interested in…how those things can be > broken out or built out separately. Overall, the degree to which artists embraced software constraints and representations corresponded to the degree that they could be adapted to an established workflow. When faced with complex tools with powerful black-box functionality or high-level representations, artists often tried to use these tools for unintended purposes. When this was not feasible, they opted for software tools with limited functionality or built their own. ### 4.3. Non-Linear Physical-Digital Workflows Despite working across a wide range of visual art domains, each artist described workflows that integrated digital and physical processes, working non-linearly between digital and physical production using a diverse set of tools and approaches. Foremost, artists heavily relied on the ability to provide manual inputs to digital tools. Miwa, Ben, and Nina all mentioned they liked the “organic” and “warm” quality of hand drawn art—deviations and irregularities in their artistic style that were built up over learning to draw and physically engaging with their body. Artists also improvised physical materials as digital inputs. Emily photographed and scanned objects to turn into textures to “incorporate as many non-computer elements into the digital artwork,” while Nina traced over copies of architectural plans as a starting point for her prints. Similarly, artists produced physical outputs with digital fabrication. Shantell used a CNC mill to fabricate a previously ink-based drawing as functional printed circuit board traces. Michael described how he would arbitrarily decide to convert flat vector graphics to 3D physical objects without advanced planning (Figure 1M): > I can extrude this. I can laser cut it 15 times and laminate it. …There’s > something to me, the formal experience of taking something that changes > dimension, which is exciting. I’m gonna find something that I normally > couldn’t have…Those opportunities past the computer. Artists also used physical production as a means to think through problems, either individually or with collaborators. For example, Emily worked out the computational rules to define an interactive generative puzzle game while working on a hand-sketched maze. Likewise, Molly shared how she would “solve code problems” while felting, constantly “switching back and forth between doing a little coding and doing a little felting or folding.” Emily and Michael, two artists who used software tools developed by their collaborators, shared that they would evaluate the tools by iteratively building physical artifacts with them. Both sets of collaborators moved back and forth between digital and physical production to create artwork together. For example, in his collaboration with Chris, Michael would “manually rip” outputs enabled by Chris’s tools and “start to draw with” them—he then sent these drawings and feedback to Chris, who in turn would modify his code. Moving between digital and physical spaces enabled artists to leverage the affordances that emerged from using both spaces when producing work, such as for painters Ben and Kim. Ben described working out ideas by alternating between digital painting on a tablet, where he could use color pickers to explore color choices instead of manually mixing paints, and explore material considerations through physical painting. He said the materiality of these physical paints added an abstraction to the way a piece communicated an idea beyond its literal representation, something that wasn’t available in digital software. Likewise, to inform her physical paintings on canvas (Figure 1L), Kim relied on “huge [digital] libraries of images and washes” to develop her ideas. For drawings to “look markedly better and more human,” Fish encouraged his students to cycle through physical and digital drawing: > Why don’t you put a piece of tracing paper over the screen right now, and > just physically draw it? And then let’s look at the drawing. And then let’s > open up Illustrator, and then create a version of it again. While appreciative of the benefits software tools provided, artists also encountered challenges of scale when translating physical artifacts to digital tools. Emily, who always began her process with physical sketches, described digitizing, segmenting, and sharing her sketches as unnecessarily laborious and repetitive. Artists were also unable to preserve the ways they manipulated physical elements of an artwork to explore scale and composition. Both Miwa and Emily noted the difficulty of using a screen to design animations that would be projected at large scale (Figure 1 A & N). Likewise, because she determined the scale of her paintings relative to her body, Kim was unable to make the same judgments while working with software interfaces. Nina went “back and forth” from working in software to printing out and “literally” putting down work on the floor to look at it when making judgments about layout and composition. Both Lynne and Nina elaborated how “proportions and visual relationships” needed to be assessed physically because balance and weight were perceived differently on a screen. In summary, artists flexibly and non-linearly moved back and forth between physical and digital spaces when creating work. Artists relied on physical manipulation as a means to refine producing artwork and as a form of reflection or problem solving. Finally, they encountered barriers when they were unable to use physical manipulation or embodied notions of proportion and scale in software tools. ### 4.4. Valuing Efficiency and Resisting Software Automation We observed that artists cared deeply about efficiency in their practice; some even developed their own forms of automation. While quickly working manually was important for aesthetic outcomes, existing forms of software-enabled automation imposed undesirable aesthetics that artists had to go back and manually refine. Artists valued speed and efficiency particularly in contexts of idea exploration, iteration, and turn-around time when working with collaborators. For example, Emily wrote software to procedurally generate and explore many different compositions, as this was faster and less effort than manually creating each one. Eran built tools for new ways of artistic expression with the goal of “getting to the quickest way [he] could test” them. In his collaboration with Michael, Chris described the importance of speed: > I love the fact that I have just enough proficiency with Processing that in > a day we could produce 20 different interesting iterations and then have a > longer dialogue about successes and failures and ways to change and ways to > improve. In forms of manual art production, working efficiently also resulted in desired aesthetic outcomes. For example, for Shantell, Mackenzie, and Michael, speed was synonymous with confidence in drawing and crucial to the aesthetic development of their line. In contrast, we saw artists reject aspects of automated efficiency that led to undesirable aesthetics, especially when they already possessed the manual skills to do something that looked better than what the software could. For example, because she disliked how the default algorithm on the vinyl cutter produced intersections, Molly wrote her own to outline Bezier curves with some thickness in order to vinyl cut her line drawings (Figure 1J). Similarly, Lynne hesitated to ever arrange typography along a path because she disliked how the automated result looked: > In Illustrator…if you try to set text on a path in that circle, it looks > really crappy, so I’ll never do it. But in an analog format, where I can cut > and paste the letters or draw them to be there, it looks fine…Maybe it’s > because of the program that it was almost taboo for me to put type on a > shape, because it looked terrible in the interface. In fact, artists often chose to create works by hand even when they recognized code could have achieved a similar aesthetic outcome. For instance, Michael preferred to execute a painting that had generative art aesthetics by hand because, to him, drawing was more efficient than the overhead of programming a similar result. Many forms of software-enabled automation that artists relied on were established features, such as undo, redo, layers, saving multiple versions of a file, and digital editability. Artists described liking these forms of automation since they remained “in the loop” and still had aesthetic control over their pieces. Taken together, the experiences of artists using software for automation were at odds with the notion that automation would remove tedious manual labor. On the contrary, because artists lacked control over nuanced outcomes in automated systems, they often spent time fighting to achieve their desired aesthetics. Manually executing their pieces, on the other hand, was both expressive and efficient. ### 4.5. Using Software Tools Beyond Producing Artifacts Software tools served artists in aspects of their practice beyond working on a visual artifact. In this section, we report on visual artists’ experiences of using software tools across tasks of documenting, tracking, generating, and sharing ideas, as well as reflecting on the process of drawing. Artists described using digital software to collect, organize, and refer to artifacts, including sketches from their processes, while making artwork. Ben described storing and relying on digital recipes of how to mix precise colors of paint. Fish would often have his sketches in an art board to the side while working on a main piece, saying it was like having a “life raft” to have “some kind of composition to play off of.” Miwa shared a similar process, using Evernote to organize and reference inspiration she came across while working and her own drawn notes. Mackenzie depended on both Evernote and rigorous commenting in his code as a means to quickly resume personal projects after working on client work. He described his C# code in Unity as being “half comments, at least.” Artists also identified points where their software tools for organization fell flat. Emily, who frequently blended drawing and note taking, felt like compared to a paper notebook, using a computer was “less freeform.” Nina felt like duplicating her Illustrator artboards as to not lose old iterations while managing version history “wasn’t streamlined” and her ideas were “getting muddled,” as the cumbersome versioning obscured her creative decision making. To aid in reflecting on and analyzing their own processes, artists used forms of digital software, such as video recording. In creating a projection-based animation piece (Figure 1N), Miwa recorded footage of herself moving around the space to analyze how her shadow affected the piece and refine its composition. Likewise, Fish recorded himself drawing with a webcam in order to review and reflect upon his process “like a sports commentator.” Shantell also described recording herself, to rather simulate the pressures of having a live audience, which mentally forced her to draw. Beyond video, Shantell also cared about analyzing the metrics of her artwork and speculated on using computational tools to aid in understanding things she could not see: > I drew [Figure 1E] at an average of five inches per second and I had someone > work out the combined amount of line—it’s roughly 1668 feet long. Oh, and > the average coverage is between 10% and 12%. So, now what can I do with that > information? One thing I’m interested in as an artist is, can I break down > all the analytics of my drawing? Can I break down my speeds, my angles, my > distances? A few artists built their own software tools to aid in reflection. The Processing extension Fish created to observe how other artists built up their drawings (as quoted in section 4.2) was originally for his own practice, but he also found it valuable as a teaching tool for his students. Nina, on the other hand, was not interested in using tools for analyzing her own workflows because she felt like her lack of experience with coding prevented her from conceptualizing how those tools would work and how she would apply them to her own practice. In summary, artists engaged with software not only to create visual artifacts but also to organize and share materials in support of their pieces, as well as to introspect on their own and others’ artwork. The fact that artists prioritized using computational tools for reflection, analysis, organization, and management showed how they leveraged computational affordances to assist creative labor, as opposed to performing it. In our interviews, artists discussed at length computational tools that helped them in their process of making artwork, as opposed to tools that made the artwork itself. ### 4.6. Relationships between Aesthetics and Audiences As previously discussed, software shaped the visual characteristics of artists’ work. Artists’ decisions to obscure or embrace computational aesthetics were impacted by social and cultural perceptions of technology. Moreover, mismatches between their own values and those of established technical communities initially led artists to hesitate in identifying as technical creators. Many artists felt artwork produced with generative algorithms trended towards a specific aesthetic, referring to work that was “generative” or “glitch”-based in style, and work that deliberately suggested a technical sophistication by emphasising “shiny,” “sexy,” or “hardcore” elements in its construction. This idea is consistent with discussions in the computational art community around a digital or generative art “vernacular” of established and sometimes cliché aesthetics from a narrow set of algorithms (Watz, 2012). Chris, Michael, Miwa, Nina and Molly described how their audiences’ expectations surrounding computationally-produced artwork determined the ways in which they obscured or emphasized these aesthetics. In designing computationally generated or interactive works, Chris, Nina, and Miwa all tried to highlight the concepts in their pieces that were not about technology. When artists chose to incorporate a recently developed computational technique into their work, the incentive for originality and novelty created a contest to quickly map out all possible variations or unconventional applications. Citing the example of Google Deep Dream (Mordvintsev et al., 2015), Chris described how: > There’s this weird race to find the new edges of the new box every time an > update…or a new platform is pushed out because you know all the easy stuff > is going to be consumed into a more easy popular culture. Constantly finding > new aesthetics or what are the aesthetics that everybody is sort of working > in, but that you need to push just beyond. Michael, Chris’s frequent collaborator, described his appreciation for Chris’s ability to “finesse” the high-tech components of the work so that they were “embedded, subdued, and poetic,” recognizing that investing in a high-tech digital aesthetic required artists to continually adapt to rapidly changing trends and avoid clichés. He personally chose to avoid this “baggage” in his own practice. When artists brought computational work into more traditional art communities, they had to decide between obscuring the computational qualities of their work or devoting significant effort to explain and contextualize its technical properties to their audience. As Molly put it: > If you’re making work that looks like sculpture, then who’s your audience? > Is it an audience who understands sculpture, but not what an algorithm is? > Because I get real tired of explaining what an algorithm is. Similarly, several artists initially resisted using computation because they did not share the values of existing technical communities, despite finding later success. Kim described a “disconnect” between the decision making processes of designers and developers. When Shantell worked in engineering communities, she struggled to reconcile her desires for transparent and aesthetically varied works with engineering norms that emphasized efficiency through technological opacity and uniformity. Lynne delayed pursuing coding because she was “burned by the tech culture” of a major Silicon Valley tech company when she worked there as a designer. Likewise, Molly described how she initially felt pushed to exhibit mastery in computer programming because of the “power dynamics” that exist between programming and domains like knitting or drawing: > There’s a certain part of the population that’s going think you’re way > cooler if you can code that thing than if you can draw it out, which is > crazy. It is worthwhile to point out that the artists who experienced these conflicts were all women. The challenges they experienced are consistent with larger patterns of marginalization of certain groups—including women—in computer science and engineering (Margolis and Fisher, 2002). Despite these conflicts, Kim, Shantell, Lynne and Molly persisted—often flourished—in computational production because they were able to find computational communities or engineering collaborators with similar values who prioritized what they had to say. Molly’s immersion in traditional art communities strengthened her belief in the importance of manual and craft skills. She reached a point where she no longer felt like she needed to “prove” she could code well for her “art to matter.” ## 5\. Discussion Here we discuss how artists currently engage with software in relationship to the current state of end-user programming (EUP) and creativity support tool (CST) research. From our analysis, we present critiques of existing approaches, design implications in response to these approaches, and strategies for new research opportunities in four categories. (1) Current tools that seek to lower the barrier of entry for creating art may rely on forms of automation and abstraction that hinder how artists traditionally learn through manual engagement with materials. (2) Artists have complex and non-linear workflows that span physical and digital media; research can work towards unifying representations across both physical and digital objects. (3) Artists are uniquely suited as technical collaborators in defining domain- specific programming representations and their contributions can expand the dimensions of what counts as systems research. (4) If systems engineering researchers wish to engage with artists, they have to consider not only the tools they build, but also the communities that surround them. ### 5.1. Automation Obscures Processes, Abstraction Obscures Data Based on our interview findings, we challenge the notion that computational automation reduces tedious manual labor. In section 4.4, we described how forms of automation and abstraction that forgo manual control presented barriers to artists becoming self-reliant and producing aesthetically sophisticated outcomes. Artists instead relied on skilled manual execution and custom software to be efficient while preserving manual style. We argue that CST research can focus on not only making tasks faster or easier to accomplish, but also helping artists develop self-sufficiency through preserving manual control. While simpler controls make tasks more accessible to novices, who are the second most targeted user group in CST research (Frich et al., 2019), forms of black-box automation can prevent artists from both using these tools in their existing workflows and in flexibly extending them across multiple workflows. For example, Molly had to write her own vinyl cutting algorithm from scratch because she could not edit the existing one provided with the software. As these forms of automation do not provide access to transparent algorithms, artists cannot adapt them in unique ways to develop idiosyncratic approaches to working. They instead may fall into aesthetics pre-determined by the tools, as described in section 4.2. Furthermore, higher-level abstractions sometimes prevented artists from working at multiple granularities. For instance, Mackenzie deliberately distinguished between the low-level representations of Bézier curves versus bitmaps when starting digital work as Bézier curves better afforded editing. In contrast, a higher-level representation, such as an automated effect or filter, would restrict this kind of meaningful decision making. By depending on computational scaffolds that do not allow them to manually manipulate data representations, artists may not develop the skills to produce sophisticated artifacts. For instance, tools that use generative adversarial networks (GANs) such as ArtBreeder (Simon, 2019) or various projects focusing on style transfer (Semmo et al., 2017) forego manual control as artists can only specify input images; the engineers of such systems are the ones who determine the visual aesthetics of the final artifact. Inspired by the processes artists described in section 4.5, forms of automation that do not remove manual control over processes can be applied to areas like exploration (Hartmann et al., 2008), project management, and new ways of “seeing” and reflecting upon artwork (Fraser et al., 2020). CST research has also investigated data abstractions artists are familiar with, such as better ways of selecting layers (Shimizu et al., 2019) and undoing/redoing actions (Chen et al., 2016; Myers et al., 2015). By allowing artists to have control over forms of automation and manipulate data representations, CSTs can focus on not only helping artists accomplish tasks, but also developing forms of self-reliance for unique outcomes. ### 5.2. Adapting Digital Tools to the Unique Workflows of Artists In section 4.3, we described how artists have complex and non-linear workflows that move across a variety of mediums, both physical and digital. We draw from these workflows to advocate for programming systems that integrate manual input and physical materials with digital output and computational control. We also see design opportunities that take into account the aesthetic experience of using digital tools that capture the user’s tacit knowledge. In creating work across physical and digital mediums, artists particularly highlighted the struggles of moving between separate tools and adapting their pieces to fit the constraints of the medium. In these transitions, artists experience what Winograd and Flores, in interpreting Heidegger, call “breakdowns” (Winograd and Flores, 1986)—instances when artists engage with low-level properties of objects because they fail to accommodate fluid and invisible interactions. Our interviews showed examples of productive breakdowns when working with physical materials, such as Michael exploring many fabrication techniques and materials to transform his vector drawings, which ultimately inspired reflections to shape the final artwork. However, artists also experienced frustrating breakdowns that were simply a result of separate applications and tools not being able to interface with each other, such as Emily having to laboriously transfer her sketches across different digital mediums. We argue that one opportunity space for CST researchers—particularly in end- user programming—is to design program representations that support non- linearly moving between work spaces of physical and digital media, such that the output of any single stage can be the input of any other stage. This is in contrast to current projects, such as those in digital fabrication (Chen et al., 2018; Savage et al., 2015; Yamaoka et al., 2019), that assume more linear workflows: artists might start with a digital design tool, then use an existing program to compile it into machine code, and then fabricate it. For example, representations that integrate digital graphics, manual drawing gesture data, and CNC toolpaths could enable artists to program custom workflows across physical and digital forms of creation, and software environments that unify interfaces for authoring graphics, modifying manual input parameters, and programming CNC machine behavior could support rapid digital-physical transitions. Projects that use interactive paper substrates (Garcia et al., 2012; Tsandilas et al., 2015) support domain experts in flexibly transitioning between the digital and physical. Devendorf and Rosner state that “hybridity,” a melding of exactly two dichotomous categories, narrows the scope of what designers work with and privileges some interactions over others (Devendorf and Rosner, 2017)—we imagine designing data representations so artists can smoothly transition between all forms of making during any stage of their process. Work has already investigated how to share workflows between different fabrication machines through integrated environments (Peek and Gershenfeld, 2018) or domain-specific languages (Tran O’Leary and Peek, 2019). Additionally, efforts in end-user programming like Enact (Leiva et al., 2019) or Webstrates (Klokmose et al., 2015) have made progress on accommodating workflows that span multiple application programs. Webstrates aims to unify software representations at the level of the operating system such that objects can be shared across different applications that traditionally have their own internal representations. However, many forms of art production rely on specific affordances of physical materials that cannot be digitally replicated (Ingold, 2010). In extending this concept to support the work of artists, we argue the “operating system” now has to include the physical spaces they also inhabit: for instance, from clay to 3D models to CAD software to G-Code to a 3D printer. Research efforts like Dynamicland (Victor, 2018) have investigated these concepts by building a programming representation and operating system to unify actions in software, and those over space and time in the physical world. When working with physical materials, artists also based their tool choices not only on how they helped them execute visual artifacts but also on how they integrated into their varied intentions while making art. The same tool could be used by different artists for sketching, for refinement, or simply because it brought both tactile and emotional pleasure. The focus artists had on their feelings and senses when using tools suggests a space for systems researchers to pay attention to the emotional and aesthetic experiences of their software, in addition to targeting contributions that make accomplishing tasks faster or easier. Finally, how artists interacted with tools was shaped by individual and embodied forms tacit knowledge, like knowing how much pressure to apply to a pen stroke while sketching versus lining. Capturing, formalizing, and evaluating these kinds of experiences—as well as emotional and aesthetic ones—is a challenge. Some CST researchers with backgrounds in art practice may draw from their own experiences (Torres et al., 2019), but when researchers lack familiarity with the practices they want to support, artists can be powerful technical collaborators. ### 5.3. Visual Artists as Technical Collaborators As detailed in section 4.2, artists have a deep domain knowledge of their medium and appropriately choose to use or not use tools based on how the constraints and representations of the tool fit into their workflows. Devendorf et al. encountered similar knowledge in their residency with weavers, and argue that craftspeople should be considered technical collaborators with HCI researchers—while craftspeople’s knowledge may be through a different medium than software engineering, that knowledge is still compatible with researchers’ practices of design iteration and innovation (Devendorf et al., 2020). We expand this notion of craftspeople as technical collaborators to visual artists; specifically, we argue visual artists are particularly well situated in defining domain-specific programming representations that integrate manual expression with computational automation and digital manipulation. Artists who code understand the constraints of writing software and applying code towards artifacts they have previously manually created, such as Lynne writing Python scripts to speed up typographic labor or Chris deriving insights about which tools to build immediately after he manually finished his pieces. Because certain representations may be better aligned with painting, rather than drawing, or felting, or may even misalign with working by hand, we suggest that the domain expertise of artists, who are well versed in their craft, can help define such representations. This is in line with ideas from participatory and co-design (Muller and Kuhn, 1993) methodologies; artists can make strong contributions in designing tools beyond the initial need-finding and final evaluation stages. However, beyond participatory design methods where artists help define high level features, we suggest that they also be involved in lower-level engineering discussions about data representations and implementation. Artists are specifically interested in forging new and different paths because they open up new avenues for artistic exploration, instead of basing their success on efficiency. Research collaborations are a two-way street that should benefit all parties in tangible ways. Before starting collaborations, researchers should consider how HCI can also bring value to the types of achievements valued in an art career. For instance, artists and researchers have different incentives in disseminating tools they make—artists value releasing tools to their communities, while less than a quarter of CSTs are publicly available (Frich et al., 2019), potentially due to the high value placed on the academic paper. We argue that HCI should broaden the scope of what is considered a systems engineering contribution to include the forms of output and inquiry artists value—such as making polished artifacts with systems over time, releasing useful technologies to support creative communities as opposed to constantly prioritizing novel systems, and investigating the art practices of individuals versus reporting on generalizable trends among large groups of artists. ### 5.4. Building Tool Communities In section 4.1, we described how artists were motivated to learn software development for various forms of power: in creating new functionality, intellectual growth, and technical legitimacy. In section 4.6, we described the experiences of four artists who overcame cultural barriers, differences in values, and feelings of exclusion to incorporate computation in their practices. Throughout this paper, we have argued for the value artists bring to our systems engineering research community. At the same time, these findings, as well as past research (Roque, 2016), show that the ways artists incorporate software and programming in their work is heavily influenced by the perceptions and values of their communities. If systems engineering researchers, as another technical community, seek to engage artists more broadly in developing tools, we also need to consider the communities surrounding the use of such tools. For inspiration, we can look to two artist-led organizations that have been successful in teaching programming to artists: p5.js, a JavaScript port of the Processing creative coding language, and the School for Poetic Computation (SFPC), an artist-run school with the motto “more poetry, less demo.” Both these organizations have been successful in devoting many resources towards building communities where artists feel a sense of belonging, in addition to providing the tools to make novel work. However, these groups—as well as the collaborations we reported on—represent a very narrow space within the larger software community. To support more artists, we need to find ways to build other communities elsewhere, as well as broaden our existing ones. The power of art comes from existing in intimate conversation with other humans, so leaving out such considerations of communities stifles the impact system builders can have on supporting artists working with computers. ## 6\. Conclusion In this paper, we examined the intersections of software and visual art to inform end-user programming and creativity support tools research. Our recommendations for software for creative production come from our thematic analysis of 13 artist interviews. By validating artists as core technical contributors in systems research and highlighting the need for inclusive community building around computational tools, we recognize and legitimize the value of distinct approaches to software use. Because systems that use high- level automation to make artwork conflict with artists’ desires for fine- grained manipulations of their artifacts and tools, we suggest using automation to instead support analysis, organization, and reflection. We argue that diverse, non-linear physical-digital workflows can inform building flexible data representations for artwork and digital fabrication. Finally, we believe artists can contribute to shaping these representations because of their distinct approaches to how they learn, use, and build software that are rooted in humanistic inquiry and art practice. Based on these findings, our vision is that software for artists will be written in close collaboration with artists. Through these collaborations, we see great potential to extend software use in creative practice and to grow inclusive communities around software development. ###### Acknowledgements. The authors extend a huge gratitude to all the artists—Chris Coleman, Emily Gobeille, Eran Hilleli, Shantell Martin, Miwa Matreyek, Fish McGill, Molly Morin, Michael Salter, Mackenzie Schubert, Kim Smith, Ben Tritt, Nina Wishnok, and Lynne Yun—without whom this work would not be possible. The authors would also like to thank colleagues Eric Rawn, Will Crichton, Evan Strasnick, Daniela Rosner, Laura Devendorf, Kristin Dew, Enric Boix-Adserà, and Kai Thaler for their valuable insights and conversations about this work. ## References * (1) * Barbosa et al. (2018) J. Barbosa, M. M. Wanderley, and S. Huot. 2018\. ZenStates: Easy-to-Understand Yet Expressive Specifications for Creative Interactive Environments. In _2018 IEEE Symposium on Visual Languages and Human-Centric Computing (VL/HCC)_. 167–175. https://doi.org/10.1109/VLHCC.2018.8506491 * Bardzell et al. (2012) Shaowen Bardzell, Daniela K. Rosner, and Jeffrey Bardzell. 2012. Crafting quality in design: integrity, creativity, and public sensibility. In _Proceedings of the Designing Interactive Systems Conference on - DIS ’12_. ACM Press, 11. https://doi.org/10.1145/2317956.2317959 * Barnes et al. (2008) Connelly Barnes, David E. Jacobs, Jason Sanders, Dan B Goldman, Szymon Rusinkiewicz, Adam Finkelstein, and Maneesh Agrawala. 2008. Video Puppetry: A Performative Interface for Cutout Animation. _ACM Trans. Graph._ 27, 5, Article 124 (Dec. 2008), 9 pages. https://doi.org/10.1145/1409060.1409077 * Batra et al. (2019) Vineet Batra, Ankit Phogat, and Tarun Beri. 2019\. Massively Parallel Layout Generation in Real Time. In _ACM SIGGRAPH 2019 Posters_ (Los Angeles, California) _(SIGGRAPH ’19)_. Association for Computing Machinery, New York, NY, USA, Article 3, 2 pages. https://doi.org/10.1145/3306214.3338596 * Beaudouin-Lafon (2017) Michel Beaudouin-Lafon. 2017\. Towards Unified Principles of Interaction. In _Proceedings of the 12th Biannual Conference on Italian SIGCHI Chapter_ _(CHItaly ’17)_. Association for Computing Machinery, 1–2. https://doi.org/10.1145/3125571.3125602 * Berger (2014) John Berger. 2014\. _Selected Essays of John Berger_. Bloomsbury Publishing Plc. * Bernstein and Li (2015) Gilbert Louis Bernstein and Wilmot Li. 2015. Lillicon: Using Transient Widgets to Create Scale Variations of Icons. _ACM Trans. Graph._ 34, 4, Article 144 (July 2015), 11 pages. https://doi.org/10.1145/2766980 * Brandt et al. (2008) Joel Brandt, Philip J. Guo, Joel Lewenstein, and Scott R. Klemmer. 2008. Opportunistic programming: how rapid ideation and prototyping occur in practice. In _Proceedings of the 4th international workshop on End-user software engineering - WEUSE ’08_. ACM Press, 1–5. https://doi.org/10.1145/1370847.1370848 * Braun and Clarke (2006) Virginia Braun and Victoria Clarke. 2006. Using thematic analysis in psychology. _Qualitative Research in Psychology_ 3, 2 (2006), 77–101. https://doi.org/10.1191/1478088706qp063oa arXiv:https://www.tandfonline.com/doi/pdf/10.1191/1478088706qp063oa * Braun and Clarke (2019) Virginia Braun and Victoria Clarke. 2019. Reflecting on reflexive thematic analysis. _Qualitative Research in Sport, Exercise and Health_ 11, 4 (2019), 589–597. https://doi.org/10.1080/2159676X.2019.1628806 arXiv:https://doi.org/10.1080/2159676X.2019.1628806 * Champandard (2016) Alex J. Champandard. 2016\. Semantic Style Transfer and Turning Two-Bit Doodles into Fine Artworks. _CoRR_ abs/1603.01768 (2016). arXiv:1603.01768 http://arxiv.org/abs/1603.01768 * Cheatle and Jackson (2015) Amy Cheatle and Steven J. Jackson. 2015. Digital Entanglements: Craft, Computation and Collaboration in Fine Art Furniture Production. In _Proceedings of the 18th ACM Conference on Computer Supported Cooperative Work & Social Computing_ _(CSCW ’15)_. Association for Computing Machinery, 958–968. https://doi.org/10.1145/2675133.2675291 * Chen et al. (2016) Hsiang-Ting Chen, Li-Yi Wei, Björn Hartmann, and Maneesh Agrawala. 2016. Data-Driven Adaptive History for Image Editing. In _Proceedings of the 20th ACM SIGGRAPH Symposium on Interactive 3D Graphics and Games_ (Redmond, Washington) _(I3D ’16)_. Association for Computing Machinery, New York, NY, USA, 103–111. https://doi.org/10.1145/2856400.2856417 * Chen et al. (2018) Xiang ’Anthony’ Chen, Ye Tao, Guanyun Wang, Runchang Kang, Tovi Grossman, Stelian Coros, and Scott E. Hudson. 2018. Forte: User-Driven Generative Design. In _Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems_ (Montreal QC, Canada) _(CHI ’18)_. Association for Computing Machinery, New York, NY, USA, 1–12. https://doi.org/10.1145/3173574.3174070 * Chugh et al. (2016) Ravi Chugh, Brian Hempel, Mitchell Spradlin, and Jacob Albers. 2016\. Programmatic and Direct Manipulation, Together at Last. _SIGPLAN Not._ 51, 6 (June 2016), 341–354. https://doi.org/10.1145/2980983.2908103 * Cycling74 (2017) Cycling74. 2017\. _Max_. https://cycling74.com/products/max/. * Devendorf et al. (2020) Laura Devendorf, Katya Arquilla, Sandra Wirtanen, Allison Anderson, and Steven Frost. 2020\. Craftspeople as Technical Collaborators: Lessons Learned through an Experimental Weaving Residency. In _Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems_ _(CHI ’20)_. Association for Computing Machinery, 1–13. https://doi.org/10.1145/3313831.3376820 * Devendorf and Rosner (2017) Laura Devendorf and Daniela K. Rosner. 2017. Beyond Hybrids: Metaphors and Margins in Design. In _Proceedings of the 2017 Conference on Designing Interactive Systems - DIS ’17_. ACM Press, 995–1000. https://doi.org/10.1145/3064663.3064705 * Devendorf and Ryokai (2015) Laura Devendorf and Kimiko Ryokai. 2015. Being the Machine: Reconfiguring Agency and Control in Hybrid Fabrication. In _Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems_ (Seoul, Republic of Korea) _(CHI ’15)_. Association for Computing Machinery, New York, NY, USA, 2477–2486. https://doi.org/10.1145/2702123.2702547 * Dew and Rosner (2018) Kristin N. Dew and Daniela K. Rosner. 2018. Lessons from the Woodshop: Cultivating Design with Living Materials. In _Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems - CHI ’18_. ACM Press, 1–12. https://doi.org/10.1145/3173574.3174159 * Do and Gross (2007) Ellen Yi-Luen Do and Mark D. Gross. 2007. Environments for Creativity: A Lab for Making Things. In _Proceedings of the 6th ACM SIGCHI Conference on Creativity & Cognition_ (Washington, DC, USA) _(C &C ’07)_. Association for Computing Machinery, New York, NY, USA, 27–36. https://doi.org/10.1145/1254960.1254965 * E et al. (2020) Jane L. E, Ohad Fried, Jingwan Lu, Jianming Zhang, Radomír Mech, Jose Echevarria, Pat Hanrahan, and James A. Landay. 2020\. Adaptive Photographic Composition Guidance. In _Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems_ (Honolulu, HI, USA) _(CHI ’20)_. Association for Computing Machinery, New York, NY, USA, 1–13. https://doi.org/10.1145/3313831.3376635 * Eaglestone et al. (2001) Barry Eaglestone, Nigel Ford, Ralf Nuhn, Adrian Moore, and Guy J Brown. 2001. Composition systems requirements for creativity: what research methodology. In _In Proc. MOSART Workshop_. 7–16. * Feist (1999) Gregory J Feist. 1999\. The influence of personality on artistic and scientific creativity. _Handbook of creativity_ (1999), 273. * Fiebrink and Sonami ([n.d.]) Rebecca Fiebrink and Laetitia Sonami. [n.d.]. Reflections on Eight Years of Instrument Creation with Machine Learning. ([n. d.]), 6. * Fraser et al. (2020) C. Ailie Fraser, Joy O. Kim, Hijung Valentina Shin, Joel Brandt, and Mira Dontcheva. 2020. Temporal Segmentation of Creative Live Streams. In _Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems_ (Honolulu, HI, USA) _(CHI ’20)_. Association for Computing Machinery, New York, NY, USA, 1–12. https://doi.org/10.1145/3313831.3376437 * Frich et al. (2019) Jonas Frich, Lindsay MacDonald Vermeulen, Christian Remy, Michael Mose Biskjaer, and Peter Dalsgaard. 2019. Mapping the Landscape of Creativity Support Tools in HCI. In _Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems_ (Glasgow, Scotland Uk) _(CHI ’19)_. Association for Computing Machinery, New York, NY, USA, 1–18. https://doi.org/10.1145/3290605.3300619 * Fujii (2017) Lee Ann Fujii. 2017\. _Interviewing in social science research: A relational approach_. Routledge. * Garcia et al. (2012) Jérémie Garcia, Theophanis Tsandilas, Carlos Agon, and Wendy Mackay. 2012. Interactive Paper Substrates to Support Musical Creation. In _Proceedings of the SIGCHI Conference on Human Factors in Computing Systems_ (Austin, Texas, USA) _(CHI ’12)_. Association for Computing Machinery, New York, NY, USA, 1825–1828. https://doi.org/10.1145/2207676.2208316 * Gatys et al. (2015) Leon A. Gatys, Alexander S. Ecker, and M. Bethge. 2015\. A Neural Algorithm of Artistic Style. _ArXiv_ abs/1508.06576 (2015). * Goodman (1968) Nelson Goodman. 1968\. Languages of Art: An Approach to a Theory of Symbols. The Bobbs-Merrill Company. _Inc. New York, Indianapolis_ (1968). * Gross (2009) M. D. Gross. 2009\. Visual Languages and Visual Thinking: Sketch Based Interaction and Modeling. In _Proceedings of the 6th Eurographics Symposium on Sketch-Based Interfaces and Modeling_ (New Orleans, Louisiana) _(SBIM ’09)_. Association for Computing Machinery, New York, NY, USA, 7–11. https://doi.org/10.1145/1572741.1572743 * Hartmann et al. (2008) Björn Hartmann, Loren Yu, Abel Allison, Yeonsoo Yang, and Scott R. Klemmer. 2008. Design as exploration: creating interface alternatives through parallel authoring and runtime tuning. In _Proceedings of the 21st annual ACM symposium on User interface software and technology - UIST ’08_. ACM Press, 91. https://doi.org/10.1145/1449715.1449732 * Hempel et al. (2019) Brian Hempel, Justin Lubin, and Ravi Chugh. 2019. Sketch-n-Sketch: Output-Directed Programming for SVG. In _Proceedings of the 32nd Annual ACM Symposium on User Interface Software and Technology_ (New Orleans, LA, USA) _(UIST ’19)_. Association for Computing Machinery, New York, NY, USA, 281–292. https://doi.org/10.1145/3332165.3347925 * Hertzmann (2018) Aaron Hertzmann. 2018\. Can Computers Create Art? _CoRR_ abs/1801.04486 (2018). arXiv:1801.04486 http://arxiv.org/abs/1801.04486 * Iizuka et al. (2016) Satoshi Iizuka, Edgar Simo-Serra, and Hiroshi Ishikawa. 2016\. Let There Be Color! Joint End-to-End Learning of Global and Local Image Priors for Automatic Image Colorization with Simultaneous Classification. _ACM Trans. Graph._ 35, 4, Article 110 (July 2016), 11 pages. https://doi.org/10.1145/2897824.2925974 * Ingold (2010) T. Ingold. 2010\. The textility of making. _Cambridge Journal of Economics_ 34, 1 (Jan 2010), 91–102. https://doi.org/10.1093/cje/bep042 * Jacobs (2018) J. Jacobs. 2018\. _SFPC Residency Reflections_. https://medium.com/sfpc/sfpc-residency-reflections-bf32204c92aa. * Jacobs et al. (2018) Jennifer Jacobs, Joel Brandt, Radomír Mech, and Mitchel Resnick. 2018. Extending Manual Drawing Practices with Artist-Centric Programming Tools. In _Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems_ _(CHI ’18)_. Association for Computing Machinery, 1–13. https://doi.org/10.1145/3173574.3174164 * Jacobs et al. (2017) Jennifer Jacobs, Sumit Gogia, Radomír Mundefinedch, and Joel R. Brandt. 2017. Supporting Expressive Procedural Art Creation through Direct Manipulation. In _Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems_ (Denver, Colorado, USA) _(CHI ’17)_. Association for Computing Machinery, New York, NY, USA, 6330–6341. https://doi.org/10.1145/3025453.3025927 * Jacobs and Zoran (2015) Jennifer Jacobs and Amit Zoran. 2015. Hybrid Practice in the Kalahari: Design Collaboration through Digital Tools and Hunter-Gatherer Craft. In _Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems_ _(CHI ’15)_. Association for Computing Machinery, 619–628. https://doi.org/10.1145/2702123.2702362 * Kazi et al. (2014) Rubaiat Habib Kazi, Fanny Chevalier, Tovi Grossman, and George Fitzmaurice. 2014. Kitty: Sketching Dynamic and Interactive Illustrations. In _Proceedings of the 27th Annual ACM Symposium on User Interface Software and Technology_ (Honolulu, Hawaii, USA) _(UIST ’14)_. Association for Computing Machinery, New York, NY, USA, 395–405. https://doi.org/10.1145/2642918.2647375 * Kazi et al. (2011) Rubaiat Habib Kazi, Kien Chuan Chua, Shengdong Zhao, Richard Davis, and Kok-Lim Low. 2011\. SandCanvas: New Possibilities in Sand Animation. In _CHI ’11 Extended Abstracts on Human Factors in Computing Systems_ (Vancouver, BC, Canada) _(CHI EA ’11)_. Association for Computing Machinery, New York, NY, USA, 483. https://doi.org/10.1145/1979742.1979562 * Kazi et al. (2017) Rubaiat Habib Kazi, Tovi Grossman, Hyunmin Cheong, Ali Hashemi, and George Fitzmaurice. 2017. DreamSketch: Early Stage 3D Design Explorations with Sketching and Generative Design. In _Proceedings of the 30th Annual ACM Symposium on User Interface Software and Technology_ (Québec City, QC, Canada) _(UIST ’17)_. Association for Computing Machinery, New York, NY, USA, 401–414. https://doi.org/10.1145/3126594.3126662 * Kim et al. (2015) Joy Kim, Mira Dontcheva, Wilmot Li, Michael S. Bernstein, and Daniela Steinsapir. 2015. Motif: Supporting Novice Creativity through Expert Patterns. In _Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems_ (Seoul, Republic of Korea) _(CHI ’15)_. Association for Computing Machinery, New York, NY, USA, 1211–1220. https://doi.org/10.1145/2702123.2702507 * Klemmer et al. (2006) Scott R. Klemmer, Björn Hartmann, and Leila Takayama. 2006\. How bodies matter: five themes for interaction design. In _Proceedings of the 6th conference on Designing Interactive systems_ _(DIS ’06)_. Association for Computing Machinery, 140–149. https://doi.org/10.1145/1142405.1142429 * Klokmose et al. (2015) Clemens N. Klokmose, James R. Eagan, Siemen Baader, Wendy Mackay, and Michel Beaudouin-Lafon. 2015. Webstrates: Shareable Dynamic Media. In _Proceedings of the 28th Annual ACM Symposium on User Interface Software & Technology_ _(UIST ’15)_. Association for Computing Machinery, 280–290. https://doi.org/10.1145/2807442.2807446 * Klokmose et al. (2019) Clemens N. Klokmose, Christian Remy, Janus Bager Kristensen, Rolf Bagge, Michel Beaudouin-Lafon, and Wendy Mackay. 2019. Videostrates: Collaborative, Distributed and Programmable Video Manipulation. In _Proceedings of the 32nd Annual ACM Symposium on User Interface Software and Technology_ _(UIST ’19)_. Association for Computing Machinery, 233–247. https://doi.org/10.1145/3332165.3347912 * Ko et al. (2011) Amy J. Ko, Robin Abraham, Laura Beckwith, Alan Blackwell, Margaret Burnett, Martin Erwig, Chris Scaffidi, Joseph Lawrance, Henry Lieberman, Brad Myers, and et al. 2011. The state of the art in end-user software engineering. _Comput. Surveys_ 43, 3 (Apr 2011), 21:1–21:44. https://doi.org/10.1145/1922649.1922658 * Leiva et al. (2019) Germán Leiva, Nolwenn Maudet, Wendy Mackay, and Michel Beaudouin-Lafon. 2019. Enact: Reducing Designer–Developer Breakdowns When Prototyping Custom Interactions. _ACM Trans. Comput.-Hum. Interact._ 26, 3, Article 19 (May 2019), 48 pages. https://doi.org/10.1145/3310276 * Leung and Lara (2015) Joshua Leung and Daniel M. Lara. 2015. Grease Pencil: Integrating Animated Freehand Drawings into 3D Production Environments. In _SIGGRAPH Asia 2015 Technical Briefs_ (Kobe, Japan) _(SA ’15)_. Association for Computing Machinery, New York, NY, USA, Article 16, 4 pages. https://doi.org/10.1145/2820903.2820924 * Levin (2003) Golan Levin. 2003\. _Essay for creative code_. http://www.flong.com/texts/essays/essay_creative_code * Levin (2015) Golan Levin. 2015\. _For Us, By Us_. http://www.flong.com/texts/essays/for-us-by-us/ * Li et al. (2020) Jingyi Li, Joel Brandt, Radomír Mech, Maneesh Agrawala, and Jennifer Jacobs. 2020. Supporting Visual Artists in Programming through Direct Inspection and Control of Program Execution. In _Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems_ (Honolulu, HI, USA) _(CHI ’20)_. Association for Computing Machinery, New York, NY, USA, 1–12. https://doi.org/10.1145/3313831.3376765 * Lieberman et al. (2006) Henry Lieberman, Fabio Paternò, Markus Klann, and Volker Wulf. 2006. End-user development: An emerging paradigm. In _End user development_. Springer, 1–8. * Lieberman (2014) Zach Lieberman. 2014\. _ofBook, a collaboratively written book about openFrameworks_. http://openframeworks.cc/ofBook/chapters/foreword.html. * Malloch et al. (2019) Joseph Malloch, Jérémie Garcia, Marcelo M. Wanderley, Wendy E. Mackay, Michel Beaudouin-Lafon, and Stéphane Huot. 2019\. _A Design Workbench for Interactive Music Systems_. Springer International Publishing, Cham, 23–40. https://doi.org/10.1007/978-3-319-92069-6_2 * Margolis and Fisher (2002) J. Margolis and A. Fisher. 2002. _Unlocking the Clubhouse: Women in Computing_. MIT Press. https://books.google.com/books?id=StwGQw45YoEC * Matejka et al. (2018) Justin Matejka, Michael Glueck, Erin Bradner, Ali Hashemi, Tovi Grossman, and George Fitzmaurice. 2018\. Dream Lens: Exploration and Visualization of Large-Scale Generative Design Datasets. In _Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems - CHI ’18_. ACM Press, 1–12. https://doi.org/10.1145/3173574.3173943 * McCarthy and Turner ([n.d.]a) L. McCarthy and M. Turner. [n.d.]a. _p5.js_. https://p5js.org/. * McCarthy and Turner ([n.d.]b) L. McCarthy and M. Turner. [n.d.]b. _p5.js Community Statement_. https://p5js.org/community/. * McCullough (1996) Malcom McCullough. 1996\. _Abstracting craft: the practiced digital hand_. MIT Press. * McDirmid (2016) Sean McDirmid. 2016\. A Live Programming Experience. https://www.youtube.com/watch?v=bnqkglrSqrg. * Mordvintsev et al. (2015) Alexander Mordvintsev, Christopher Olah, and Mike Tyka. 2015\. Deepdream-a code example for visualizing neural networks. _Google Research_ 2, 5 (2015). * Muller and Kuhn (1993) Michael J Muller and Sarah Kuhn. 1993. Participatory design. _Commun. ACM_ 36, 6 (1993), 24–28. * Mumford (1952) Lewsi Mumford. 1952\. _Art and Technics_. Columbia University Press. * Myers et al. (2008) Brad Myers, Sun Young Park, Yoko Nakano, Greg Mueller, and Andrew Ko. 2008. How designers design and program interactive behaviors. In _2008 IEEE Symposium on Visual Languages and Human-Centric Computing_. IEEE, 177–184. * Myers et al. (2015) Brad A. Myers, Ashley Lai, Tam Minh Le, YoungSeok Yoon, Andrew Faulring, and Joel Brandt. 2015\. Selective Undo Support for Painting Applications. In _Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems_ (Seoul, Republic of Korea) _(CHI ’15)_. Association for Computing Machinery, New York, NY, USA, 4227–4236. https://doi.org/10.1145/2702123.2702543 * Needleman (1979) Carla Needleman. 1979\. _The work of craft: an inquiry into the nature of crafts and craftsmanship_. Arkana. * Oney et al. (2014) Stephen Oney, Brad Myers, and Joel Brandt. 2014. InterState: a language and environment for expressing interface behavior. In _Proceedings of the 27th annual ACM symposium on User interface software and technology_ _(UIST ’14)_. Association for Computing Machinery, 263–272. https://doi.org/10.1145/2642918.2647358 * Papert (1980) Seymour Papert. 1980\. _Mindstorms: Children, Computers, and Powerful Ideas_. Basic Books, Inc., USA. * Peek and Gershenfeld (2018) Nadya Peek and Neil Gershenfeld. 2018. Mods: Browser-based rapid prototyping workflow composition. (2018). * Reas and Fry (2004) C. Reas and B. Fry. 2004\. _Processing_. http://processing.org. * Reas and Fry (2016) C. Reas and B. Fry. 2016\. _Processing Overview_. http://processing.org/overview. * Resnick and Silverman (2005) Mitchel Resnick and Brian Silverman. 2005. Some reflections on designing construction kits for kids. In _Proceedings of the 2005 conference on Interaction design and children_ _(IDC ’05)_. Association for Computing Machinery, 117–122. https://doi.org/10.1145/1109540.1109556 * Roque (2016) Ricarose Vallarta Roque. 2016\. _Family creative learning: designing structures to engage kids and parents as computational creators_. Ph.D. Dissertation. Massachusetts Institute of Technology. * Runway AI (2020) Inc. Runway AI. 2020\. _RunwayML_. https://runwayml.com/. * Savage et al. (2015) Valkyrie Savage, Sean Follmer, Jingyi Li, and Björn Hartmann. 2015. Makers’ Marks: Physical Markup for Designing and Fabricating Functional Objects. In _Proceedings of the 28th Annual ACM Symposium on User Interface Software & Technology_ (Charlotte, NC, USA) _(UIST ’15)_. Association for Computing Machinery, New York, NY, USA, 103–108. https://doi.org/10.1145/2807442.2807508 * Schachman (2012) Toby Schachman. 2012\. Alternative Programming Interfaces for Alternative Programmers. In _Proceedings of the ACM International Symposium on New Ideas, New Paradigms, and Reflections on Programming and Software_ (Tucson, Arizona, USA) _(Onward! 2012)_. Association for Computing Machinery, New York, NY, USA, 1–10. https://doi.org/10.1145/2384592.2384594 * Schon ([n.d.]) Donald A Schon. [n.d.]. Designing as reflective conversation with the materials of a design situation. ([n. d.]), 17. * Semmo et al. (2017) Amir Semmo, Tobias Isenberg, and Jürgen Döllner. 2017\. Neural Style Transfer: A Paradigm Shift for Image-Based Artistic Rendering?. In _Proceedings of the Symposium on Non-Photorealistic Animation and Rendering_ (Los Angeles, California) _(NPAR ’17)_. Association for Computing Machinery, New York, NY, USA, Article 5, 13 pages. https://doi.org/10.1145/3092919.3092920 * Shimizu et al. (2019) Evan Shimizu, Matt Fisher, Sylvain Paris, and Kayvon Fatahalian. 2019. Finding Layers Using Hover Visualizations. In _Proceedings of the 45th Graphics Interface Conference on Proceedings of Graphics Interface 2019_ (Kingston, Canada) _(GI’19)_. Canadian Human-Computer Communications Society, Waterloo, CAN, Article 16, 9 pages. https://doi.org/10.20380/GI2019.16 * Shneiderman (2002) Ben Shneiderman. 2002\. Creativity Support Tools. _Commun. ACM_ 45, 10 (Oct. 2002), 116–120. https://doi.org/10.1145/570907.570945 * Shneiderman (2007) Ben Shneiderman. 2007\. Creativity support tools: accelerating discovery and innovation. _Commun. ACM_ 50, 12 (Dec 2007), 20–32. https://doi.org/10.1145/1323688.1323689 * Simo-Serra et al. (2018) Edgar Simo-Serra, Satoshi Iizuka, and Hiroshi Ishikawa. 2018\. Mastering Sketching: Adversarial Augmentation for Structured Prediction. _Transactions on Graphics (Presented at SIGGRAPH)_ 37, 1 (2018). * Simon (2019) Joel Simon. 2019\. _ArtBreeder_. https://www.artbreeder.com/ * Thiel et al. (2011) Yannick Thiel, Karan Singh, and Ravin Balakrishnan. 2011\. Elasticurves: Exploiting Stroke Dynamics and Inertia for the Real-Time Neatening of Sketched 2D Curves. In _Proceedings of the 24th Annual ACM Symposium on User Interface Software and Technology_ (Santa Barbara, California, USA) _(UIST ’11)_. Association for Computing Machinery, New York, NY, USA, 383–392. https://doi.org/10.1145/2047196.2047246 * Torres et al. (2019) Cesar Torres, Matthew Jörke, Emily Hill, and Eric Paulos. 2019. Hybrid Microgenetic Analysis: Using Activity Codebooks to Identify and Characterize Creative Process. In _Proceedings of the 2019 on Creativity and Cognition_ (San Diego, CA, USA) _(C &C ’19)_. Association for Computing Machinery, New York, NY, USA, 2–14. https://doi.org/10.1145/3325480.3325498 * Tran O’Leary and Peek (2019) Jasper Tran O’Leary and Nadya Peek. 2019. Machine-o-Matic: A Programming Environment for Prototyping Digital Fabrication Workflows. In _The Adjunct Publication of the 32nd Annual ACM Symposium on User Interface Software and Technology_ (New Orleans, LA, USA) _(UIST ’19)_. Association for Computing Machinery, New York, NY, USA, 134–136. https://doi.org/10.1145/3332167.3356897 * Tsandilas et al. (2015) Theophanis Tsandilas, Magdalini Grammatikou, and Stéphane Huot. 2015. BricoSketch: Mixing Paper and Computer Drawing Tools in Professional Illustration. In _Proceedings of the 2015 International Conference on Interactive Tabletops & Surfaces_ (Madeira, Portugal) _(ITS ’15)_. Association for Computing Machinery, New York, NY, USA, 127–136. https://doi.org/10.1145/2817721.2817729 * Tversky (2019) Barbara Tversky. 2019\. _Mind in motion_. Basic Books. * Victor (2011) Bret Victor. 2011\. _Dynamic Pictures_. http://worrydream.com/DynamicPicturesMotivation * Victor (2018) Bret Victor. 2018\. Dynamicland. https://dynamicland.org * vvvv group (2017) vvvv group. 2017\. _vvvv_. https://vvvv.org/. * Watz (2012) Marius Watz. 2012\. _Algorithm Critique and Computational Aesthetics_. Vimeo. https://vimeo.com/46903693 * Wessel and Wright (2001) David Wessel and Matthew Wright. 2001. Problems and Prospects for Intimate Musical Control of Computers. In _Proceedings of the 2001 Conference on New Interfaces for Musical Expression_ (Seattle, Washington) _(NIME ’01)_. National University of Singapore, SGP, 1–4. * Winograd and Flores (1986) Terry Winograd and Fernando Flores. 1986. _Understanding computers and cognition: A new foundation for design_. Intellect Books. * Xing et al. (2015) Jun Xing, Li-Yi Wei, Takaaki Shiratori, and Koji Yatani. 2015\. Autocomplete Hand-Drawn Animations. _ACM Trans. Graph._ 34, 6, Article 169 (Oct. 2015), 11 pages. https://doi.org/10.1145/2816795.2818079 * Yamaoka et al. (2019) Junichi Yamaoka, Mustafa Doga Dogan, Katarina Bulovic, Kazuya Saito, Yoshihiro Kawahara, Yasuaki Kakehi, and Stefanie Mueller. 2019\. FoldTronics: Creating 3D Objects with Integrated Electronics Using Foldable Honeycomb Structures. In _Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems_ (Glasgow, Scotland Uk) _(CHI ’19)_. Association for Computing Machinery, New York, NY, USA, 1–14. https://doi.org/10.1145/3290605.3300858 * Zaman et al. (2015) Loutfouz Zaman, Wolfgang Stuerzlinger, Christian Neugebauer, Rob Woodbury, Maher Elkhaldi, Naghmi Shireen, and Michael Terry. 2015\. GEM-NI: A System for Creating and Managing Alternatives In Generative Design. In _Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems - CHI ’15_. ACM Press, 1201–1210. https://doi.org/10.1145/2702123.2702398 * Zhao et al. (2020) Nanxuan Zhao, Nam Wook Kim, Laura Mariah Herman, Hanspeter Pfister, Rynson W.H. Lau, Jose Echevarria, and Zoya Bylinskii. 2020\. ICONATE: Automatic Compound Icon Generation and Ideation. In _Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems_ (Honolulu, HI, USA) _(CHI ’20)_. Association for Computing Machinery, New York, NY, USA, 1–13. https://doi.org/10.1145/3313831.3376618
# Rabi oscillation of V${}_{\text{B}}^{-}$ spin in hexagonal boron nitride Wei Liu Zhi-Peng Li Yuan-Ze Yang Shang Yu Yu Meng Zhao-An Wang Nai-Jie Guo Fei-Fei Yan Qiang Li Jun-Feng Wang Jin-Shi Xu Yang Dong Xiang-Dong Chen Fang-Wen Sun Yi-Tao Wang<EMAIL_ADDRESS>Jian-Shun Tang <EMAIL_ADDRESS>Chuan-Feng Li<EMAIL_ADDRESS>Guang-Can Guo CAS Key Laboratory of Quantum Information, University of Science and Technology of China, Hefei, P.R.China CAS Center For Excellence in Quantum Information and Quantum Physics, University of Science and Technology of China, Hefei, P.R.China ###### Abstract VdW materials are a family of materials ranging from semimetal, semiconductor to insulator, and their common characteristic is the layered structure. These features make them widely applied in the fabrication of nano-photonic and electronic devices, particularly, vdW heterojunctions. HBN is the only layered material to date that is demonstrated to contain optically-detected electronic spins, which can benefit the construction of solid qubit and quantum sensor, etc., especially embedded in the nano-layered-devices. To realize this, Rabi oscillation is a crucial step. Here, we demonstrate the Rabi oscillation of V${}_{\text{B}}^{-}$ spins in hBN. Interestingly, we find the behaviors of the spins are completely different under the conditions of weak and relatively higher magnetic field. The former behaves like a single wide peak, but the latter behaves like multiple narrower peaks (e.g., a clear beat in Ramsey fringes). We conjecture a strong coupling of the spin with the surrounding nuclear spins, and the magnetic field can control the nuclear spin bath. Van der Waals (vdW) materials include a family of materials with various bandgap, and exhibit diverse electronic properties from semimetal (e.g., graphene) to semiconductor (e.g., transition metal dichalcogenides, TMDCs for short), and to insulator (e.g., hexagonal boron nitride, or hBN) XiaF2014 . Their common feature is the layered structure, namely, the atoms in the same layer are combined by the strong chemical bond, and the layers are connected by the relatively weak vdW force. This feature makes the layers from different vdW materials easy to be stacked together to form heterojunctions Geim2013 , which have the advantage of no lattice mismatch compared to their three- dimensional counterparts including GaAs, silicon, or diamond, etc. Besides, vdW materials have strong interaction with light since the two-dimensional confinement Trovatello2021 . These characteristics lead to a great of applications of vdW materials, such as photocurrent generation YuWJ2013 , light-emitting diode Ross2014 , field effect transistor LiuW2013 , single photon Srivastava2015 ; HeYM2015 ; Koperski2015 ; Chakraborty2015 ; Tonndorf2015 ; Palacios2017 ; Branny2017 ; Errando2020 ; TranTT2016n ; TranTT2016p ; TranTT2016a ; Martinez2016 ; Chejanovsky2016 ; Choi2016 ; Grosso2017 ; XueY2018 ; Proscia2018 ; LiuW2020 ; Fournier2020 ; Barthelmi2020 , and optical parametric amplification Trovatello2021 . Moreover, the light- valley interaction in TMDCs leads to the field of valleytronics XuX2014 ; Manzeli2017 . All these applications will contribute to the design and construction of photonic and electronic devices in very small scale, benefited from the atomic thickness of vdW materials. Among this family of layered materials, hBN has a large bandgap of $\sim$6 eV, which makes it have the ability to host plenty of kinds of defects, just similar to diamond Barry2020 ; Hanson2008 ; ChenX2015 and silicon carbide WangJF2020 ; YanFF2020 ; LiQ2020 , etc. Single-layer hBN was first found to emit single photons at room temperature in 2016 by Tran _et al._ TranTT2016n , and after that, a great of interest was stimulated to treat hBN defects as the promising single-photon emitters TranTT2016p ; TranTT2016a ; Martinez2016 ; Chejanovsky2016 ; Choi2016 ; Grosso2017 ; XueY2018 ; Proscia2018 ; LiuW2020 ; Fournier2020 , and furthermore, as the potential solid spin qubit Exarhos2019 ; Toledo2018 ; Gottscholl2020 ; Chejanovsky2019 ; Mendelson2020 ; Kianinia2020 ; GaoX2020 ; LiuW2021 ; Gottscholl2020a . As the single-photon source, hBN defect (in monolayer, flake or bulk) has the advantages of high brightness Grosso2017 ; LiuW2020 , broad spectral range TranTT2016a , easy tunability Grosso2017 ; XueY2018 , and easy fabrication, etc. The fabrication methods are diverse, including chemical etch Chejanovsky2016 , electron or ion irradiation Chejanovsky2016 ; Choi2016 ; Kianinia2020 ; Fournier2020 , laser ablation Choi2016 ; GaoX2020 , strain Proscia2018 , and so on. Especially, defects in hBN attract a lot of attentions to be a good candidate for solid spin qubit (particularly, in the vdW-nano-devices). Actually, electron paramagnetic resonance (EPR) signals in hBN have been found in very early decades Katzir1975 ; Moore1972 ; Fanciulli1992 . These signals are recognized by numerical calculations recently Sajid2018 ; Abdi2018 , and these theoretical works also predict many possible defects in hBN who have the potential to give optically detected magnetic resonance (ODMR) signals. In experiment, Exarhos _et al._ Exarhos2019 find the magnetic-field-dependent intensity of fluorescence emitted from a hBN defect in 2019. Later, the ODMR signals are revealed by Gottscholl _et al._ Toledo2018 ; Gottscholl2020 , Chejanovsky _et al._ Chejanovsky2019 and Mendelson _et al._ Mendelson2020 , respectively. The first kind of defects are assumed to be V${}_{\text{B}}^{-}$, and the third kind of defects are conjectured to be related to carbon. Based on the experimental results, several theoretical analyzations are carried out, especially for V${}_{\text{B}}^{-}$ Ivady2020 ; Sajid2020 , and the temperature-dependent features of this kind of spin defect are also detailedly investigated recently in experiment LiuW2021 . Figure 1: Simplified energy levels and ODMR results. (a) Simplified energy levels of V${}_{\text{B}}^{-}$ center and the related optical transitions among ground states (GS), excited states (ES) and metastable states (MS). The 532-nm laser (green) is used for the spin polarization and readout, and the microwave (pink) is used for coherent control of the spin state. (b) ODMR spectrum measured at room temperature with no external magnetic field. The bimodal signal due to the local strain has obvious hyperfine structures, which indicate the nucleus-electron interaction between V${}_{\text{B}}^{-}$ and the three neighboring nitrogen nuclei (14N). The experimental data are fitted using two-Lorentzian function to obtain $\nu_{1}\sim$3.424 GHz and $\nu_{2}\sim$3.533 GHz. (c) Dependence of the frequency shift of the $m_{s}=-1\leftrightarrow m_{s}=0$ transition on the magnetic field parallel to the hexagonal $c$ axis, from which we can fit the $g$ factor of V${}_{\text{B}}^{-}$ to be $1.992\pm 0.010$. Figure 2: Rabi oscillations. (a) Pulse sequence of Rabi measurement comprising the first laser pulse for spin polarization, then the microwave pulse with length $\tau$ for spin manipulation and the second laser pulse for state readout. (b) Rabi oscillations on the $m_{s}=-1\leftrightarrow m_{s}=0$ transition observed at room temperature and different magnetic field. The data are fitted using $\Sigma_{i=1}^{n}A_{i}\exp(-\tau/T_{i})\cos(2\pi f_{i}\tau+\phi_{i})+b\exp(-\tau/T_{b})+c$, with $n=1,2$ and $3$ (red lines) from top to down, respectively. $A_{i}$, $T_{i}$, $f_{i}$, $\phi_{i}$, $b$, $T_{b}$ and $c$ are the amplitude, oscillation decay time, frequency, phase, decayed background and its decay time, constant background, respectively. (c) Rabi frequency measured with a linear dependence on the square root of microwave power $\sqrt{P}$. Figure 3: $T_{1}$ measurement and spin echo detections. (a) Pulse sequence for characterizing the spin-lattice relaxation dynamics, including the spin polarization and readout, the microwave $\pi$-pulse obtained from Rabi-measurement and the free evolution time $\tau$ for changes. (b) $T_{1}$ measurement at 0 mT revealing the spin-lattice relaxation time of $T_{1}=16.377\pm 0.416$ $\mu$s. (c) $T_{1}$ time versus magnetic field, suggesting that there is roughly no $T_{1}$ dependence on magnetic field. (d) Pulse scheme for spin echo measurement with $\frac{\pi}{2}-\frac{\tau}{2}-\pi-\frac{\tau}{2}-\frac{\pi}{2}$ sequence, where $\tau$ is the free evolution time. (e) Optically-detected spin-echo measurement at 0 mT. (f) Spin echo at 36 mT which cannot be fitted well, showing complicated oscillations induced by the nuclear spin bath and the red line is only a guide for eyes. Figure 4: Ramsey interference. (a) Ramsey pulse sequence with $\frac{\pi}{2}-\tau-\frac{\pi}{2}$. (b) Ramsey result with $B=0$ mT and $f_{\text{MW}}=3428$ MHz. No oscillation is observed but a fast decay with $T_{2}^{*}=60.198\pm 2.747$ ns. A slow background decay is also observed as that in Rabi results. (c) Ramsey result at 44 mT and 2200-MHz microwave frequency. Three frequencies are observed and two of them form a clear beat. The red line is the fitting. The distances between the adjacent frequencies are both around $47$ MHz (the hyperfine splitting due to the nucleus-electron interaction observed and calculated previously). $T_{2}^{*}$ corresponding to these three frequencies are $0.665\pm 0.108$ $\mu$s, $2.500\pm 2.160$ $\mu$s and $1.448\pm 0.841$ $\mu$s, respectively. The energy levels of V${}_{\text{B}}^{-}$ is gradually clear, and a simplified diagram is sketched in Fig. 1(a). As discussed in Refs. Gottscholl2020 ; Ivady2020 , this defect contains a triplet ground state (GS), which is primarily constituted of three energy levels with $m_{s}=0$ and $m_{s}=\pm 1$, and $D$ is the zero-field splitting (ZFS) between them. ES and MS represent the excited states and metastable states, respectively. The green arrows stand for the excitation laser which pump the population to above ES, and the red arrow represents the fluorescence to be detected. The gray wavy arrows represent the inter-system crossings (ISC) between $S=3$ and $S=1$. The pink circled arrow is the applied microwave (MW) between $m_{s}=0$ and $m_{s}=\pm 1$, which will change populations of these states, and hence change the intensity of fluorescence. By recording the difference of the fluorescence intensities, we can read the state of spin qubit. This method is called ODMR. A typical ODMR signal of V${}_{\text{B}}^{-}$ at zero-magnetic-field is shown in Fig. 1(b). The excitation laser is always on, and the MW works at the on/off mode, then the contrast is calculated from the difference between the on and off fluorescences. $\nu_{0}=(\nu_{1}+\nu_{2})/2=3.479$ GHz is the frequency corresponding to ZFS $D$, and $\nu_{2,1}$ correspond to the transitions between $m_{s}=\pm 1$ states to $m_{s}=0$, respectively. Remarkably, for each transition peak, we can clearly see several hyperfine splittings, which can be identified as the nucleus-electron interaction. For V${}_{\text{B}}^{-}$ defect, there are three nitrogen nuclei (14N) around, each of which has a nuclear spin of $I=1$. Therefore, totally $2(3I)+1=7$ hyperfine transitions should be observed, and the interval between them is detected Gottscholl2020 and calculated Ivady2020 to be approximately $47$ MHz. Fig. 1(c) shows the frequency shift of the $m_{s}=-1\leftrightarrow m_{s}=0$ transition as the magnetic field varies, from which we can fit the $g$ factor of this spin to be $1.992\pm 0.010$. The next step is naturally to coherently operate the solid spin, of which the Rabi oscillation is a key tool. Here we utilize the two-level states $m_{s}=-1,0$ as the spin qubit to perform the coherent control. Fig. 2(a) is the time diagram of Rabi oscillation. After a long excitation-laser pulse, the spin is polarized to $m_{s}=0$ state, then a MW pulse with length of $\tau$ is applied to rotate the spin, followed by a readout pulse. The results are shown in Fig. 2(b). At magnetic field of $B=0$ mT, we see a standard decayed Rabi oscillation, but we also observe a tiny decay of background, which may be caused by the overlarge density of V${}_{\text{B}}^{-}$ defects, since our integrated dose of neutron irradiation is quite large ($\sim 2.0\times 10^{17}$ $n$ cm-2). For results of non-zero magnetic field, we observe an oscillation of multiple Rabi frequencies. We conjecture that a more orderly ensemble of defects is formed because of the magnetic field, which makes the hyperfine peaks narrower and can to some extent be separated from each other; or maybe some other reasons such as the influences of the environmental nuclear or electronic spins Hanson2008 . By varying the MW power, we derive the linear dependence of the fitted Rabi frequency versus the square root of power $\sqrt{P}$ (see Fig. 2(c)), which indicates the validity of our Rabi results. With the Rabi frequency, we can define $\frac{\pi}{2}$-pulse and $\pi$-pulse. Utilizing $\pi$-pulse, we can measure the spin-lattice relaxation time $T_{1}$. The pulse sequence is shown in Fig. 3(a) and Fig. 3(b) exhibits the relaxation result at $B=0$ mT. By fitting this result, we derive $T_{1}=16.377\pm 0.416$ $\mu$s. Then we repeat this sequence for various magnetic fields and obtain the results in Fig. 3(c). We find $T_{1}$ is approximately independent of the magnetic field. Next, we perform the sequence $\frac{\pi}{2}-\frac{\tau}{2}-\pi-\frac{\tau}{2}-\frac{\pi}{2}$ (spin echo, see Fig. 3(d)) to measure $T_{2}$. At $B=0$ mT, we observe a monotonic decay of contrast shown in Fig. 3(e), and $T_{2}$ is fitted to be $82.121\pm 2.462$ ns, which is quite short. We conjecture that it may also be caused by the overlarge density of V${}_{\text{B}}^{-}$ defects. At $B=36$ mT, we find the decayed-contrast curve is complicatedly modulated (see Fig. 3(f)). We cannot fit it well, and the red line is only a guide for eyes. Here we want to note that, since $T_{2}$ is quite short, the impact of time durations of the MW pulses, especially the $\pi$-pulse, can not be ignored, therefore, the values of the fitted $T_{2}$ will have inaccuracy, but the order of magnitude can be determined. For $T_{1}$ which is far longer than MW-pulse durations, this problem is not met. We also perform the Ramsey interference experiment on the V${}_{\text{B}}^{-}$ spins. The pulse sequence is presented in Fig. 4(a), and Fig. 4(b) shows the result at $B=0$ mT. We have not seen the oscillations, which may be caused by the fast decay corresponding to a short $T_{2}^{*}$. On one hand, the nucleus- electron interaction splits the $m_{s}=-1\leftrightarrow m_{s}=0$ transition into 7 peaks; on the other hand, every peak is broad at zero magnetic field due to the inhomogeneous ensemble, and the peaks adhere to each other to form a single wide peak and thus induce a short $T_{2}^{*}$. Similar to the result in Fig. 2(b), we also see a slow decay of background in this figure (this data cannot be fitted well using single-decay curve, but a double-decay curve). The reason may also be attributed to the overlarge density of V${}_{\text{B}}^{-}$ defects in the sample. In contrast, when we apply a magnetic field of $B=44$ mT and set the MW frequency at $f_{\text{MW}}=2200$ MHz, we see a multiple- frequency oscillation, in which a beat is clearly recognized, and it is superposed on another slow oscillation. These three frequencies are fitted as $f_{-1}=-44.171\pm 0.039$ MHz, $f_{0}=0.934\pm 0.131$ MHz, $f_{1}=45.872\pm 0.063$ MHz, respectively. We therefore conjecture that the three peaks that contribute to the Ramsey oscillation are located at $h(f_{\text{MW}}+(-)f_{i})$ ($i=-1,0,1$, $h$ is Planck constant, and the sign “-” represents the possibility of reverse direction), and other peaks contribute to this oscillation little since they are much farther from the MW frequency. The distance between the adjacent peaks is $f_{0}-f_{-1}=45.105\pm 0.136$ MHz$\approx f_{1}-f_{0}=44.938\pm 0.145$ MHz (the Planck constant is omitted here), which is coincident with the calculated Ivady2020 and observed Gottscholl2020 energy separation of the hyperfine splittings induced by nucleus-electron interaction (approximately $47$ MHz). The fitted $T_{2}^{*}$ of these three peaks are $0.665\pm 0.108$ $\mu$s, $2.500\pm 2.160$ $\mu$s and $1.448\pm 0.841$ $\mu$s, respectively. Although we have not extracted $T_{2}$ in the case that magnetic field is nonzero due to the quite complex modulations, we can conjecture that $T_{2}$ will be in the order of microseconds from the results of $T_{2}^{*}$ of the individual peaks. It seems the magnetic field help elongate the spin coherence time. It is interesting to find that in both the Rabi- and Ramsey-oscillation results, the observed phenomena are very different when the magnetic field is weak or relatively strong. At zero magnetic field, the defects behave as a large ensemble with a single broad peak; but when the magnetic field is strong enough, the defects behave like a more orderly ensemble with multiple narrow peaks, i.e., the multiple-frequency fittings of each data with higher magnetic field. This phenomena suggest that the V${}_{\text{B}}^{-}$ spin is highly correlated to the neighboring nuclear spins, which can be utilized to study the nuclear spins or the correlations between them. As the only candidate for spin qubit in vdW material (to date), the coherent operations of defects in hBN based on Rabi oscillation play the crucial role, and provide a powerful tool for the design and construction of spin-based vdW- nano-devices, especially when the techniques of vdW heterojunction are combined. Although $T_{2}$ is still quite short, which may be primarily due to the overlarge density of V${}_{\text{B}}^{-}$ defects in the sample caused by the high-dose neutron irradiation, there will be several methods to improve it. For example, we can decrease the integrated dose of neutron irradiation; or perform a suitable annealing on the sample since high-temperature condition can reduce the V${}_{\text{B}}^{-}$ defect number; or put the sample into a low-temperature cryostat; or apply higher magnetic field; etc. In summary, we have realized the Rabi oscillation of the V${}_{\text{B}}^{-}$ spins in hBN, based on which we also detect $T_{1}$ and perform the spin-echo and Ramsey-interference experiments. We find $T_{1}$ is almost not affected by magnetic field, which is roughly around $16.377\pm 0.416$ $\mu$s; however, the results of Rabi oscillation, spin echo and Ramsey oscillation are very different under the conditions of weak and relatively strong magnetic field. At zero magnetic field, the defects behave as a large ensemble with a single broad peak, i.e., the Rabi result is well fitted by a single-frequency oscillation, the echo result shows a fast decay of $82.121\pm 2.462$ ns, and the Ramsey result also decays fast without any oscillations. When the magnetic field goes higher, the defects behave like a more orderly ensemble with multiple narrow peaks, and we see multiple-frequency oscillations in both Rabi and Ramsey results. Especially, we see a clear beat and an additional slow oscillation in the Ramsey result, and by fitting this result, we find the distances among these three frequencies ($45.105\pm 0.136$ MHz$\approx 44.938\pm 0.145$ MHz) are roughly equal to the energy separation of the nucleus-electron-interaction-induced hyperfine splitting (approximately $47$ MHz, $h$ is omitted). The decay times ($T_{2}^{*}$) of these three oscillations are $0.665\pm 0.108$ $\mu$s, $2.500\pm 2.160$ $\mu$s and $1.448\pm 0.841$ $\mu$s, respectively, from which we conjecture that $T_{2}$ for each separated peak will be in the order of microsecond. It seems the magnetic field freezes the environmental spins to some extend and elongates the $T_{2}$. Our results suggest that the V${}_{\text{B}}^{-}$ spin is highly correlated to the neighboring nuclear spins, which provides a potential tool to study them. ## Acknowledgments This work is supported by the National Key Research and Development Program of China (No. 2017YFA0304100), the National Natural Science Foundation of China (Grants Nos. 11822408, 11674304, 11774335, 11821404, and 11904356), the Key Research Program of Frontier Sciences of the Chinese Academy of Sciences (Grant No. QYZDY-SSW-SLH003), the Fok Ying-Tong Education Foundation (No. 171007), the Youth Innovation Promotion Association of Chinese Academy of Sciences (Grants No. 2017492), Science Foundation of the CAS (No. ZDRW- XH-2019-1), Anhui Initiative in Quantum Information Technologies (AHY020100, AHY060300), the Fundamental Research Funds for the Central Universities (Nos. WK2470000026, WK2030000008 and WK2470000028). ## References * (1) Xia, F., Wang, H., Xiao, D., Dubey, M. & Ramasubramaniam, A. Two-dimensional material nanophotonics. _Nature Photon._ 8, 899-907 (2014). * (2) Geim, A. K. & Grigorieva, I. V. Van der Waals heterostructures. _Nature_ 499, 419-425 (2013). * (3) Trovatello, C. _et al._ Optical parametric amplification by monolayer transition metal dichalcogenides. _Nature Photon._ 15, 6-10 (2021). * (4) Yu, W. J. _et al._ Highly efficient gate-tunable photocurrent generation in vertical heterostructures of layered materials. _Nature Nanotechnol._ 8, 952-958 (2013). * (5) Ross, J. S. _et al._ Electrically tunable excitonic light-emitting diodes based on monolayer WSe2 p-n junctions. _Nature Nanotechnol._ 9, 268-272 (2014). * (6) Liu, W. _et al._ Role of metal contacts in designing high-performance monolayer n-type WSe2 field effect transistors. _Nano Lett._ 13, 1983-1990 (2013). * (7) Srivastava, A. _et al._ Optically active quantum dots in monolayer WSe2. _Nature Nanotechnol._ 10, 491-496 (2015). * (8) He, Y. M. _et al._ Single quantum emitters in monolayer semiconductors. _Nature Nanotechnol._ 10, 497-502 (2015). * (9) Koperski, M. _et al._ Single photon emitters in exfoliated WSe2 structures. _Nature Nanotechnol._ 10, 503-506 (2015). * (10) Chakraborty, C., Kinnischtzke, L., Goodfellow, K. M., Beams, R. & Vamivakas, A. N. Voltage-controlled quantum light from an atomically thin semiconductor. _Nature Nanotechnol._ 10, 507-511 (2015). * (11) Tonndorf, P. _et al._ Single-photon emission from localized excitons in an atomically thin semiconductor. _Optica_ 2, 347-352 (2015). * (12) Palacios-Berraquero, C. _et al._ Large-scale quantum-emitter arrays in atomically thin semiconductors. _Nature Commun._ 8, 1-6 (2017). * (13) Branny, A., Kumar, S., Proux, R. & Gerardot, B. D. Deterministic strain-induced arrays of quantum emitters in a two-dimensional semiconductor. _Nature Commun._ 8, 1-7. (2017). * (14) Errando-Herranz, C. _et al._ On-chip single photon emission from a waveguide-coupled two-dimensional semiconductor. Preprint at https://arxiv.org/abs/2002.07657 (2020). * (15) Tran, T. T., Bray, K., Ford, M. J., Toth, M. & Aharonovich, I. Quantum emission from hexagonal boron nitride monolayers. _Nature Nanotechnol._ 11, 37-41 (2016). * (16) Tran, T. T. _et al._ Quantum emission from defects in single-crystalline hexagonal boron nitride. _Phys. Rev. Appl._ 5, 034005 (2016). * (17) Martínez, L. J. _et al._ Efficient single photon emission from a high-purity hexagonal boron nitride crystal. _Phys. Rev. B_ 94, 121405(R) (2016). * (18) Chejanovsky, N. _et al._ Structural attributes and photodynamics of visible spectrum quantum emitters in hexagonal boron nitride. _Nano Lett._ 16, 7037-7045 (2016). * (19) Choi, S. _et al._ Engineering and localization of quantum emitters in large hexagonal boron nitride layers. _ACS Appl. Mater. Interfaces_ 8, 29642-29648 (2016). * (20) Tran, T. T. _et al._ Robust multicolor single photon emission from point defects in hexagonal boron nitride. _ACS Nano_ 10, 7331-7338 (2016). * (21) Grosso, G. _et al._ Tunable and high-purity room temperature single-photon emission from atomic defects in hexagonal boron nitride. _Nature Commun._ 8, 1-8 (2017). * (22) Xue, Y. _et al._ Anomalous pressure characteristics of defects in hexagonal boron nitride flakes. _ACS Nano_ 12, 7127-7133 (2018). * (23) Proscia, N. V. _et al._ Near-deterministic activation of room-temperature quantum emitters in hexagonal boron nitride. _Optica_ 5, 1128-1134 (2018). * (24) Liu, W., Wang, Y.-T., Li, Z.-P., Yu, S., Ke, Z.-J., Meng, Y., Tang, J.-S., Li, C.-F. & Guo, G.-C. An ultrastable and robust single-photon emitter in hexagonal boron nitride. _Physica E_ 124, 114251 (2020). * (25) Fournier, C. _et al._ Position-controlled quantum emitters with reproducible emission wavelength in hexagonal boron nitride. Preprint at https://arxiv.org/abs/2011.12224 (2020). * (26) Barthelmi, K. _et al._ Atomistic defects as single-photon emitters in atomically thin MoS2. _Appl. Phys. Lett._ 117, 070501 (2020). * (27) Xu, X., Yao, W., Xiao, D. & Heinz, T. F. Spin and pseudospins in layered transition metal dichalcogenides. _Nature Phys._ 10, 343-350 (2014). * (28) Manzeli, S., Ovchinnikov, D., Pasquier, D., Yazyev, O. V. & Kis, A. 2D transition metal dichalcogenides. _Nature Rev. Mater._ 2, 17033 (2017). * (29) Barry, J. F. _et al._ Sensitivity optimization for NV-diamond magnetometry. _Rev. Mod. Phys._ 92, 015004 (2020). * (30) Hanson, R., Dobrovitski, V. V., Feiguin, A. E., Gywat, O. & Awschalom, D. D. Coherent dynamics of a single spin interacting with an adjustable spin bath. _Science_ 320, 352-355 (2008). * (31) Chen, X.-D., Zou, C.-L., Gong, Z.-J., Dong, C.-H., Guo, G.-C. & Sun, F.-W. Subdiffraction optical manipulation of the charge state of nitrogen vacancy center in diamond. _Light Sci. Appl._ 4, e230-e230 (2015). * (32) Wang, J.-F., Yan, F.-F., Li, Q., Liu, Z.-H., Liu, H., Guo, G.-P., Guo, L.-P., Zhou, X., Cui, J.-M., Wang, J., Zhou, Z.-Q., Xu, X.-Y., Xu, J.-S., Li, C.-F. & Guo, G.-C. Coherent control of nitrogen-vacancy center spins in silicon carbide at room temperature. _Phys. Rev. Lett._ 124, 223601 (2020). * (33) Yan, F.-F., Yi, A.-L., Wang, J.-F., Li, Q., Yu, P., Zhang, J.-X., Gali, A., Wang, Y., Xu, J.-S., Ou, X., Li, C.-F. & Guo, G.-C. Room-temperature coherent control of implanted defect spins in silicon carbide. _npj Quantum Inf._ 6, 1-6 (2020). * (34) Li, Q., Wang, J.-F., Yan, F.-F., Zhou, J.-Y., Wang, H.-F., Liu, H., Guo, L.-P., Zhou, X., Gali, A., Liu, Z.-H., Wang, Z.-Q., Sun, K., Guo, G.-P., Tang, J.-S., Xu, J.-S., Li, C.-F. & Guo, G.-C. Room temperature coherent manipulation of single-spin qubits in silicon carbide with high readout contrast. Preprint at https://arxiv.org/abs/2005.07876 (2020). * (35) Exarhos, A .L., Hopper, D. A., Patel, R. N., Doherty, M. W. & Bassett, L. C. Magnetic-field-dependent quantum emission in hexagonal boron nitride at room temperature. _Nature Commun._ 10, 222 (2019). * (36) Toledo, J. R. _et al._ Electron paramagnetic resonance signature of point defects in neutron-irradiated hexagonal boron nitride. _Phys. Rev. B_ 98, 155203 (2018). * (37) Gottscholl, A. _et al._ Initialization and read-out of intrinsic spin defects in a van der Waals crystal at room temperature. _Nature Mater._ 19, 540-545 (2020). * (38) Chejanovsky, N. _et al._ Single spin resonance in a van der Waals embedded paramagnetic defect. Preprint at https://arxiv.org/abs/1906.05903 (2019). * (39) Mendelson, N. _et al._ Identifying carbon as the source of visible single-photon emission from hexagonal boron nitride. _Nature Mater._ (2020). https://doi.org/10.1038/s41563-020-00850-y. * (40) Kianinia, M., White, S., Froch, J. E., Bradac, C. & Aharonovich, I. Generation of spin defects in hexagonal boron nitride. _ACS Photon._ 7, 2147-2152 (2020). * (41) Gao, X. _et al._ Femtosecond Laser Writing of Spin Defects in Hexagonal Boron Nitride. Preprint at https://arxiv.org/abs/2012.03207 (2020). * (42) Liu, W., Li, Z.-P., Yang, Y.-Z., Yu, S., Meng, Y., Wang, Z.-A., Li, Z.-C., Guo, N.-J., Yan, F.-F, Li, Q., Wang, J.-F., Xu, J.-S., Wang, Y.-T., Tang, J.-S., Li, C.-F. & Guo, G.-C. Temperature-dependent energy-level shifts of Spin Defects in hexagonal Boron Nitride. Preprint at https://arxiv.org/abs/2101.09920 (2021). * (43) Gottscholl, A. _et al._ Room Temperature Coherent Control of Spin Defects in hexagonal Boron Nitride. Preprint at https://arxiv.org/abs/2010.12513 (2020). * (44) Moore, A. W. & Singer, L. S. Electron spin resonance in carbon-doped boron nitride. _J. Phys. Chem. Solids_ 33, 343-356 (1972). * (45) Katzir, A., Suss, J. T., Zunger, A. & Halperin, A. Point defects in hexagonal boron nitride. I. EPR, thermoluminescence, and thermally-stimulated-current measurements. _Phys. Rev. B_ 11, 2370 (1975). * (46) Fanciulli, M. & Moustakas, T. D. in _Wide Band Gap Semiconductors, Proceedings of the Annual Fall Meeting of the Materials Research Society_ (Materials Research Society, Pittsburgh, PA, 1992). * (47) Sajid, A., Reimers, J. R. & Ford, M. J. Defect states in hexagonal boron nitride: Assignments of observed properties and prediction of properties relevant to quantum computation. _Phys. Rev. B_ 97, 064101 (2018). * (48) Abdi, M., Chou, J. P., Gali, A. & Plenio, M. B. Color centers in hexagonal boron nitride monolayers: a group theory and ab initio analysis. _ACS Photon._ 5, 1967-1976 (2018). * (49) Ivády, V. _et al._ Ab initio theory of the negatively charged boron vacancy qubit in hexagonal boron nitride. _npj Comput. Mater._ 6, 1-6 (2020). * (50) Sajid, A., Thygesen, K. S., Reimers, J. R. & Ford, M. J. Edge effects on optically detected magnetic resonance of vacancy defects in hexagonal boron nitride. _Commun. Phys._ 3, 1-8 (2020).
# Multi-Instance Pose Networks: Rethinking Top-Down Pose Estimation ## 1 Multi-Instance Modulation Block (MIMB) Code ## 2 Implementation Details ## 3 Diminishing Returns with $\mathbf{N=3,4}$ ## 4 Additional Results on COCO, CrowdPose and OCHuman ### 4.1 Additional results on COCO ### 4.2 Additional results on CrowdPose ### 4.3 Additional results on OCHuman ### 4.4 Robustness to Bounding Box Confidence ## 5 Individual Instance Performance ## 6 Ablation: MIMB ## 7 OCPose Dataset ## 8 Qualitative Results
# Reciprocal Landmark Detection and Tracking with Extremely Few Annotations Jianzhe Lin Student Member, IEEE Ghazal Sahebzamani Student Member, IEEE Christina Luong Fatemeh Taheri Dezaki Student Member, IEEE Mohammad Jafari Student Member, IEEE Purang Abolmaesumi Fellow, IEEE and Teresa Tsang Jianzhe Lin, Ghazal Sahebzamani, Fatemeh Taheri Dezaki, Mohammad Jafari, and Purang Abolmaesumi are with the Electrical and Computer Engineering Department, University of British Columbia, Vancouver, BC, V6T 1Z2, Canada.(e-mail: jianzhelin, ghazal, fatemeht, mohammadj, purang@ece.ubc.ca).Teresa Tsang is the Associate Head Research and Co-Acting Head, Department of Medicine,University of British Columbia. She is Director of Echo Laboratory at Vancouver General Hospital, Vancouver, BC, <EMAIL_ADDRESS>Luong isa Clinical Assistant Professor within the Division of Cardiology at the University of British Columbia. She is the Head of Stress Echocardiography at Vancouver General Hospital, Vancouver, BC, Canada..(e-mail:t.tsang@ubc.ca). ###### Abstract Localization of anatomical landmarks to perform two-dimensional measurements in echocardiography is part of routine clinical workflow in cardiac disease diagnosis. Automatic localization of those landmarks is highly desirable to improve workflow and reduce interobserver variability. Training a machine learning framework to perform such localization is hindered given the sparse nature of gold standard labels; only few percent of cardiac cine series frames are normally manually labeled for clinical use. In this paper, we propose a new end-to-end reciprocal detection and tracking model that is specifically designed to handle the sparse nature of echocardiography labels. The model is trained using few annotated frames across the entire cardiac cine sequence to generate consistent detection and tracking of landmarks, and an adversarial training for the model is proposed to take advantage of these annotated frames. The superiority of the proposed reciprocal model is demonstrated using a series of experiments. localization, gold standard labels, adversarial training, reciprocal model. ## 1 Introduction Data scarcity and lack of annotation is a general problem for developing machine learning models in medical imaging. Among various medical imaging modalities, ultrasound (US) is the most frequently used modality given its widespread availability, lower cost, and safety since it does not involve ionizing radiation. Specifically, US imaging, in the form of echocardiography (echo), is the standard-of-care in cardiac imaging for the detection of heart disease. Echo examinations are performed across up to 14 standard views from several acoustic windows on the chest. In this paper, we specifically focus on the parasternal long axis (PLAX), which is one of the most common view acquired in the point-of-care US for rapid examination of cardiac function (Fig. 1). Several measurements from PLAX require the localization of anatomical landmarks across discrete points in the cardiac cycle. Our work specifically investigates automatic localization of the left ventricle (LV) internal dimension (LVID), which is routinely used to estimate the ejection fraction (EF), a strong indicator of cardiac function abnormality. In clinics, LVID landmarks are determined in two frames of the cardiac cycle, i.e. end- diastolic and end-systolic. However, such annotation is challenging, specially for general physicians at point-of-care who do not have the experience of cardiologists. As such, the automation of landmark localization is highly desirable. However, developing a machine learning model for such automation has been hindered by the availability of only sparse set of labeled frames in cardiac cines. Manually labeling all cardiac frames for a large set of cardiac cines is virtually impractical, given limited expert time. Figure 1: Example of PLAX images, as one of the most common standard views acquired in point-of-care echocardiography. Landmarks identified on the left ventricle are used to measure the EF, a strong indicator of cardiac disease. Two landmarks on inferolateral and anteroseptal walls (IW, AW) are yellow color while the LVID is the red line. LVID can be localized with IW and AW. Instead of manually labeling, we propose a new Reciprocal landmark Detection and Tracking (RDT) model that enables automation in measurements across the entire cardiac cycle. The model only uses prior knowledge from sparsely labeled key frames that are temporally distant in a cardiac cycle. Meanwhile, we take advantage of temporal coherence of cardiac cine series to impose cycle consistency in tracking landmarks across unannotated frames that are between these two annotated frames. To impose consistent detection and tracking of the landmarks, we propose a reciprocal training as a self-supervision process. Figure 2: The general flowchart of the proposed detection and tracking model. Gold standard labels are only available for end-diastolic and end-systolic frames. The propagation starts from the end-diastolic frame and ends at the end-systolic frame. The tracking is completed in a cycle way. The two annotated frames serve as a weak supervision for the model. The detection and tracking results from the unannotated frames jointly reciprocally provide another self-supervision. In summary, we propose a RDT model, which is weakly supervised by only two annotated keyframes in an image sequence for model training. For testing, the model is an end-to-end model that detects the landmark in the first frame, followed by a tracking process. Our contributions are: * • A novel Reciprocal landmark Detection and Tracking (RDT) model is proposed. In the model, the spatial constraint for detection and temporal coherence for tracking of cardiac cine series work reciprocally, to generate accurate localization of landmarks; * • The sparse nature of echocardiography labels is handled by the proposed model. The model is only weakly supervised by two annotated image frames that are temporally distant from each other. The annotation sparsity is also analyzed in the experimental part; * • A novel adversarial training approach (Ad-T) for optimization of the proposed RDT. Such training is made possible by introducing four complementary losses as in Fig. 2, i.e. reciprocal loss, motion loss, focal loss, and cycle loss. Compared with conventional training approaches, Ad-T indirectly achieves feature augmentation, which is extremely important for model training given the extremely few annotations. the advantage of such Ad-T is highlighted in our ablation study. ## 2 Related Work As a low cost, low risk, and easily accessible modality, the cardiac US is widely used as an assessment tool in point-of-care. With the utilization of US technology in various form factors from cart-based to hand-held devices, measurement of cardiac structures can be typically conducted by users with diverse levels of expertise. However, due to US images’ noisy nature, studies indicate large amounts of inter- and intra-observer variability even among experts [1]. This amount of observer variability may easily lead to errors in reporting an abnormal patient as normal, or vice versa for borderline cases. This fact has raised the significance of automated measurement systems by reducing the variability and increasing the reliability of cardiac reports among US operators. Furthermore, automation saves a considerable amount of time by improving clinical workflow. The problem of automated prediction of clinical measurements, such as segmentation and keypoint localization of anatomical structures, has been approached from different angles, especially within the deep learning literature, where leveraging large size training datasets has led to significant improvements in the accuracy of predicted measurements. Most of the recent methods have used fully convolutional neural networks (FCN) as their main building block to predict pixel-level labels [2, 3, 4, 5, 6, 7, 8, 9]. Similar to numerous works in pose detection literature [10, 11], in many FCN-based methods, the structure localization problem has been approached by predicting heatmaps corresponding to the regions of interest at some point in the network [12]. In [12], a convolutional neural network (CNN) architecture was explored to combine the local appearance of one landmark with the spatial configuration of other landmarks for multiple landmark localization. However, these methods are introduced for problems where data consists of individual frames, rather than temporal sequences. On the contrary, time plays an important role in the calculation of measurements such as EF in cardiac cycles. Therefore, the sole use of these methods may not be sufficient for our problem of interest and other temporally constrained or real-time applications. Recent studies have made use of spatio-temporal models to overcome limitations of previous models in problems dealing with sequential data, and particularly, echo cine loops [13, 14]. In [15], while a center of the mass layer was introduced and placed on top of an FCN architecture to regress keypoints out of the predicted heatmaps directly, a convolutional long short-term memory (CLSTM) network was also utilized for improving temporal consistency. In the cardiac segmentation domain, many works such as [16] have applied recurrent neural networks to their pipeline. In [17], multi-scale features are first extracted with pyramid ConvBlocks, and these features are aggregated using hierarchical ConvLSTMs. Other types of studies have fed motion information to their network based on estimating the motion vector between consecutive frames [18, 19, 20]. Another case of this method is presented by [21], in which similar to our weakly-supervised problem, motion estimation is obtained from an optical flow branch to enforce spatio-temporal smoothness over a weakly supervised segmentation task with sparse labels in the temporal dimension. However, optical flow estimation might contain drastic errors in consecutive frames with large variability, especially in US images where the boundaries are fuzzy, and considerable amounts of noise and artifacts may be present. Therefore, they may not be suitable for a weakly supervised task where the labels are distant in the time domain. Moreover, although most of the mentioned methods take temporal coherence into account, these constraints may not be directly enforced on the model in a desired way [18, 19, 13, 21, 17, 20, 16]. In order to overcome these shortcomings, [22] proposed a method for consistent segmentation of echocardiograms in the time dimension, where only end-diastolic and end-systolic frames have segmentation labels per cycle. This method consists of two co-learning strategies for segmentation and tracking, in which the first strategy estimates shape and motion fields in appearance level, and the second one imposes further temporal consistency in shape level for the previous segmentation predictions. In our method, however, instead of a segmentation task, we perform detection and tracking with reciprocal learning in a landmark detection paradigm in the presence of sparse temporal labels. ## 3 Approach Our general RDT framework can be found in Figure 2. The model can be divided into three parts, the _feature encoder_ (blue color), _detection head_ (orange color), and _tracking head_ (green color). The feature encoder and detection head combined can be viewed as a Unet-like model, for which the general structure is similar to Unet. In the model training phase, the input of the RDT model is an echo sequence starting from the end-diastolic frame and ending at the end-systolic frame. For the detection branch, the input is the whole frame, while for the tracking branch, the inputs are patches from two neighboring frames. The output of the network is two predicted landmark pair locations for LVID. ### 3.1 Problem Formulation Suppose the frames in the cardiac cine series are represented by $\\{I_{1},I_{2},I_{3},...,I_{k}\\}$. For model training, we suppose the end- diastolic frame to be the $1^{st}$ frame, and the end-systolic frame to be the $k^{th}$ frame. The $1^{st}$ and $k^{th}$ frames are with annotation, while the in-between frames are unannotated. The landmark pairs are represented by ${i_{t},a_{t}}$ ($i_{t}=\\{x^{i}_{t},y^{i}_{t}\\},a_{t}=\\{x^{a}_{t},y^{a}_{t}\\}$) corresponding to the landmarks on the inferolateral and anteroseptal walls of LV in the $t^{th}$ frame, respectively. We use $\phi$ to represent the _feature encoder_ , and the feature generated for $I_{t}$ is represented by $\phi_{I_{t}}$. The $\phi_{I_{t}}$ is solely input to the _detection head_ $D$ to get the predicted landmark locations ${i_{t}^{D},a_{t}^{D}}$. For _tracking head_ , the input is the cropped features of two consecutive frames. One serves as the template frame while the other serves as the search frame. For landmark tracking, the predicted locations start from the $2^{nd}$ frame. After a cycle forward and backward propagation, the predicted location will end at the $1_{st}$ frame. ### 3.2 Network Architecture and Losses #### 3.2.1 Shared Feature Encoder The feature encoder consists of six 3$\times$3 convolution layers, each followed by a rectified linear unit (ReLU). The third convolution layer is with a stride equal to 2. Since a single feature encoder is sufficient for the tracking head, we share this part of the encoder with both tracking and detection branches. Since the shared encoder is optimized by losses generated from different heads, the encoded feature should be robust since its optimization considers both the spatial information exploited by the detection branch and temporal information explored by the tracking branch. #### 3.2.2 Detection Head and Focal Loss The detection head combined with the feature encoder together can be viewed as an Unet-like structure, which consists of a contracting path and an expansive path. The contracting path follows the typical architecture of a convolutional network. The beginning of the detection head is another six layers for feature generation. There are two similar downsampling steps to the shared feature encoder. However, we also double the number of feature channels in these two steps. Every step in the expansive path consists of an upsampling of the feature map followed by a 2$\times$2 convolution (“up-convolution”). The first two upsampling layers halve the number of feature channels. We also concatenate the output of each upsampling layer with a correspondingly cropped feature map from the contracting path. Each 3$\times$3 convolutions is followed by a ReLU. As padding is applied, there is no cropping in the whole neural network. For the final two layers used for classification, the first one is a 3$\times$3 convolution layer, and the second is a 1$\times$1 layer, which is used to map each 48-component feature vector to the desired number of landmarks (Here, the number of landmarks is 2). The last layer’s output is a two-dimension heatmap, and each location of the heatmap represents the probability of a target landmark. Focal loss is generated on annotated frames. For each landmark, there is one ground-truth positive location in each dimension of the heatmap (Two landmarks correspond to two dimensions), and all the other locations are negative. For such ground truth, penalizing negative location equally with the positive ones is not appropriate, therefore we apply the focal loss. During training, we reduce the penalty given to negative locations within a radius of the positive location. We empirically set the radius to be 10 pixels. The amount of penalty reduction is given by an unnormalized 2D Gaussian $e^{-(x^{2}+y^{2})/2\delta^{2}}$, whose center is at the positive location and whose $\sigma$ is 1/3 of the radius. Let $p_{c_{i,j}}$ be the score at location (i, j) for landmark c in the predicted heatmap, and let $y_{c_{i,j}}$ be the ground-truth heatmap augmented with the unnormalized Gaussians. We create a variant of focal loss [23]: $\scriptsize{\mathcal{L}_{\det}}=\sum\limits_{c=1}^{2}{\sum\limits_{i=1}^{H}{\sum\limits_{j=1}^{W}{\left\\{{\begin{array}[]{*{20}{c}}{{{(1-{p_{c_{i,j}}})}^{\alpha}}\log({p_{c_{i,j}}})\quad if\quad{y_{c_{i,j}}}=1}\\\ {{{(1-y{}_{c_{i,j}})}^{\beta}}{{({p_{c_{i,j}}})}^{\alpha}}\log(1-{p_{c_{i,j}}})\quad else,}\end{array}}\right.}}}$ (1) where $\alpha$ and $\beta$ are the hyperparameters that control the contribution of each point (we empirically set $\alpha$ to 2 and $\beta$ to 4 in all experiments). With the Gaussian distribution encoded in the $y_{c_{i,j}}$, the term $1-{y_{c_{i,j}}}$ is used for reducing the penalty around the ground truth locations. #### 3.2.3 Tracking Head and Cycle Loss: For the tracking head, when we get $\phi_{I_{t}}$ and $\phi_{I_{t-1}}$, we first crop the search patches and the template patches both centering at the landmark pairs in the two consecutive frames, respectively. The two template patches for inferolateral/anteroseptal landmarks get concatenated and are represented by $P_{t-1}$, while the two search patches for inferolateral/anteroseptal landmarks get concatenated and are represented by $N_{t}$. The input for the tracking branch is the template patch $P_{t-1}$ with size $25\times 25$ and the search patch $N_{t}$ with size $29\times 29$, both centering at the landmark patch $i_{t-1},a_{t-1}$. The size of $P_{t-1}$ and $N_{t}$ are labeled in Fig. 2, which are set empirically. We formulate the _tracking head_ $T$ as $\delta_{i_{t}},\delta_{a_{t}}=T(\phi_{P_{t}},\phi_{N_{t+1}})$. For the tracking head, we first define a convolutional operation between $\phi_{P_{t-1}}$ and $\phi_{N_{t}}$ in order to compute the affinity (similarity) between each sub-patch of $\phi_{N_{t}}$ and $\phi_{P_{t-1}}$. To be more specific, $\phi_{P_{t-1}}$ and $\phi_{N_{t}}$ are combined by using a cross-correlation layer $f(\phi_{N_{t}},\phi_{P_{t-1}})=\phi_{P_{t-1}}*\phi_{N_{t}}.$ (2) Note that the output of this function is a feature map indicating the _affinity score_. For hands-on implementation, it is simple to take $\phi_{P_{t-1}}$ as a kernel matrix to compute dense convolution on $\phi_{N_{t}}$ within the framework of existing conv-net libraries. The output feature map is followed by another three fully connected layers (represented by m in Eq. 3) to predict the landmark motion. Such regression operation is further formulated as $T(\phi_{P_{t}},\phi_{N_{t+1}})=\delta_{i_{t}},\delta_{a_{t}}=m(f(\phi_{N_{t}},\phi_{P_{t-1}});\theta_{f}).$ (3) where $\theta_{f}$ represents the parameters for the fully connected network. $\delta_{i_{t}}$ and $\delta_{a_{t}}$ are both two-dimensional moves (along x-axis and y-axis, respectively). The new landmark location is calculated by adding its previous location to the predicted motion. Such motion prediction is generally similar with optical flow, in which a new three-layer regression is also incorporated. This regression makes the learning process adaptive. Figure 3: Optimization of the proposed reciprocal training. As the tracking process is only supervised by end-diastolic and end-systolic frames, we introduce the cycle loss and motion loss to supervise the tracking branch. To model the cycle process, we iteratively apply the tracking head $T$ in a forward manner: $\begin{array}[]{l}{L_{t}}^{*}=T({\phi_{{P_{t-1}}}},{\phi_{{N_{t}}}})+{L_{t-1}}^{*}\\\ =T({\phi_{{P_{t-1}}}},{\phi_{{N_{t}}}})+T({\phi_{{P_{t-2}}}},{\phi_{{N_{t-1}}}})+{L_{t-2}}^{*}\\\ =T({\phi_{{P_{t-1}}}},{\phi_{{N_{t}}}})+...T({\phi_{{P_{1}}}},{\phi_{{N_{2}}}})+{L_{1}}^{*},\end{array}$ (4) in which $L_{t}^{*}=\\{i_{t},a_{t}\\}$ represents the predicted location of landmark pairs in $t^{th}$ frame, while $L_{1}^{*}=\\{i_{1},a_{1}\\}$ represents the ground truth location of landmark pairs in the first annotated frame. Here ”+” represents the element-wise addition between the location of landmarks in the current frame and motion calculated in Eq.3. Also, we use the same formulation in backward manner as: $\begin{array}[]{l}{L_{1}}^{*}=T({\phi_{{P_{2}}}},{\phi_{{N_{1}}}})+{L_{2}}^{*}\\\ =T({\phi_{{P_{2}}}},{\phi_{{N_{1}}}})+T({\phi_{{P_{3}}}},{\phi_{{N_{2}}}})+{L_{3}}^{*}\\\ =T({\phi_{{P_{t}}}},{\phi_{{N_{t-1}}}})+...T({\phi_{{P_{2}}}},{\phi_{{N_{1}}}})+{L_{t}}^{*}.\end{array}$ (5) We use the labeled end-diastolic frame as the beginning frame of the echo cine series, and the end-systolic frame as the end frame. The motion loss is defined by the deviation between the predicted landmark pair locations in the end-systolic frame and their ground truth locations. Suppose the labeled end- systolic frame is the $k^{th}$ frame; after forward propagation, the motion loss $\mathcal{L}$ is defined as $\begin{array}[]{l}\mathcal{L}_{motion}^{k}=\mathcal{L}_{1\rightarrow k}=\|{L_{k}}-{L_{k}}^{*}\|^{2}\\\ =\|{L_{k}}-(T({\phi_{{P_{k-1}}}},{\phi_{{N_{k}}}})+...T({\phi_{{P_{1}}}},{\phi_{{N_{2}}}})+{L_{1}})\|^{2}.\end{array}$ (6) The forward propagation is followed by backward propagation that ends at the end-diastolic frame. By combining Eq. 4 and Eq. 5, the current predicted landmark pair location in the diastolic frame $L_{1}^{*}$ can actually be represented by its ground truth location $L_{1}$, and we use the deviation between these two terms to represent the cycle loss as follow: $\begin{array}[]{l}\mathcal{L}_{cycle}^{k}=\mathcal{L}_{1\rightarrow k\rightarrow 1}=\|{L_{1}}-{L_{1}}^{*}\|^{2}\\\ =\|{L_{1}}-{L_{k}}^{*}+{L_{k}}^{*}-{L_{1}}^{*}\|^{2}\\\ =\|(T({\phi_{{P_{k-1}}}},{\phi_{{N_{k}}}})+...T({\phi_{{P_{1}}}},{\phi_{{N_{2}}}}))+\\\ (T({\phi_{{P_{k}}}},{\phi_{{N_{k-1}}}})+...T({\phi_{{P_{2}}}},{\phi_{{N_{1}}}}))\|^{2}.\end{array}$ (7) Finally, the cycle loss can be simplified as $\mathcal{L}_{cycle}^{k}=-(\mathcal{L}_{motion}^{k}+\mathcal{L}_{motion}^{1}).$ (8) #### 3.2.4 Reciprocal Loss for Unannotated Frames: The former motion loss, cycle loss, and focal loss are applied for the annotated frames, whereas the reciprocal loss is proposed only for the unannotated frames, which can be viewed as a self-supervision. In the training phase, only the end-diastolic and end-systolic frames are annotated while the in-between frames are unannotated. For these unannotated frames, we can generate both the $i_{t}^{D},a_{t}^{D}=max(D(\phi_{I_{t}}))$ and the $i_{t}^{T},a_{t}^{T}=T(\phi_{P_{t-1}},\phi_{N_{t}})+i_{t-1}^{T},a_{t-1}^{T}$. Although no annotation was assigned, the two predicted landmark pair locations are assumed to be the same. The discrepancy between these two formulates the reciprocal loss. The frame rate for reciprocal loss is set as 3, which means such loss is generated in every three frames. As $D(\phi_{I_{t}})$ is a heatmap with each location indicating the probability of target location, we define the reciprocal loss similar to the focal loss. We assume $i_{t}^{T}$ and $a_{t}^{T}$ to be the only positive locations of frame $t$, which is augmented as a 2D Gaussian distribution centering at each positive location. The predicted heatmap from the detection branch is viewed as predicted locations. The formulated reciprocal loss ${\mathcal{L}_{rec}(D,T)}$ is the same as defined in Eq. 1. ## 4 Optimization The basic idea for the proposed RDT model is to create a reciprocal learning between the detection task and the tracking task, as the detection task mainly focuses on the spatial information of a single frame, while the tracking task considers the temporal correlation between consecutive frames. However, the detected landmark pair locations and the tracked landmark pair locations are assumed to be the same. Therefore, we want the two branches to generate a discrepancy to optimize both the feature encoder $\phi$ and the detection/tracking head. We propose a novel adversarial optimization mechanism. The motivation is for feature augmentation as the number of data is really limited. Trained by the augmented feature, both the detection head D and the tracking head T in Fig. 3 can be more robust. In Fig. 3, we use blue color to represent the feature distribution of the target landmark pair, and orange color to represent the background. In order to generate a more different distribution of features from unannotated frames, we propose to utilize the disagreement between D and T on the prediction of unannotated frames. We assume D and T can predict the location of annotated frames correctly. Here, we use a key intuition that the feature distribution of unannotated data outside the support of the annotated ones is likely to be predicted differently by D and T. Black lines denote this region as in Fig. 3 (Discrepancy Region). Therefore, if we can measure the disagreement between D and T and train $\phi$ to maximize the disagreement, the encoder will generate more unknown feature distributions outside the support of the annotated ones. The disagreement here is our formerly formulated reciprocal loss ${\mathcal{L}_{rec}(D,T)}$. This goal can be achieved by iterative steps as in Fig. 4. We first update the feature encoder to maximize the ${\mathcal{L}_{rec}(D,T)}$. Then we freeze this encoder part, and update the D and T to minimize the ${\mathcal{L}_{rec}(D,T)}$, in order to get the uniformed predicted results for the newly generated unknown feature from the feature encoder. Detailed optimization steps are described as follows. ### 4.1 Training Steps: We need to train D and T, which take inputs from $\phi$. Both D and T must predict the annotated landmark pair locations correctly. We solve this problem in three steps, as can be found in Fig. 4. Step A. First, we train D, T, and $\phi$ to predict the landmark pairs of annotated frames correctly. We train the networks to minimize three losses applied to annotated frames. The objective is as follows: $\mathop{\min}\limits_{\phi,D,T}({\mathcal{L}_{\det}}+\mathcal{L}_{motion}^{k}+\mathcal{L}_{cycle}^{k});$ (9) Step B. In this step, we train the feature encoder $\phi$ for fixed D and T. By training the encoder to increase the discrepancy, more unknown feature distributions different from the annotated data can be generated. Note that this step only uses the unannotated data. The objective can be formulated as: $\mathop{\max}\limits_{\phi}({\mathcal{L}_{rec}}({\rm{D}},T));$ (10) Step C. We train D and T to minimize the discrepancy with a fixed $\phi$. As this step is to get the uniformed and correct detection/tracking results, the step is repeated for three times for the same mini-batch empirically. This setting achieves a trade-off between the encoder and the heads (detection, tracking). This step is applied on both annotated and unannotated frames, to get the best model weights of detection/tracking heads for all the existing features. The objective is as follows: $\mathop{\min}\limits_{D,T}({\mathcal{L}_{\det}}+\mathcal{L}_{motion}^{k}+\mathcal{L}_{cycle}^{k}+{\mathcal{L}_{rec}}({\rm{D}},T)).$ (11) These three steps are repeated until convergence. Weights for different losses are emprically set as 1, in both Step A and Step C. Based on our experience, the order of the three steps is not essential. However, our primary concern is to train D, T, and $\phi$ in an adversarial manner. Figure 4: Stepwise model training process. ## 5 Experiments ### 5.1 Dataset and Setup Our echocardiography dataset is collected from our local hospital, following approvals from the Medical Research Ethics Board in coordination with the privacy office. Data were randomly split into mutually exclusive training and testing datasets, where no patient data were shared across the two datasets. The training dataset includes 995 echo cine series with 1990 annotated frames, while the testing dataset includes 224 sequences with 448 annotated frames. Different sequences have a different number of frames ranging from 10s to 100s. The number of frames between end-diastolic and end-systolic phases is different for each cine sample, ranging from 5 to 20 frames. We run the experiments on our 8x Tesla V100 Server. For the hardware, the CPU is Intel(R) Xeon(R) CPU E5-2698 v4. All comparison methods are trained until convergence. For the proposed method trained by a single GPU, the model converges at 30 epochs, and the running time is 31min/epoch. ### 5.2 Quantitative Results EF in PLAX view is estimated based on the distance between inferolateral and anteroseptal landmarks, i.e. LVID. We use the length error (LE) of LVID as well as the location deviation error (LDE) of inferolateral/anteroseptal landmarks (abbreviated as IL/AL) as key errors. LDE is also the most widely used criterion for detection/tracking methods. The comparison is mainly made among the proposed method, the most recently proposed frame by frame detection-based method (Modified U-Net[24], CenterNet[25]), and the regular detection+tracking method (Unet+C-Ynet[26]). Unet here is with the same structure as the proposed method. Unet and C-Ynet are trained separately. A general comparison can be found in Table 6. Table 1: Statistical comparison with the state-of-the-art methods. Errors (’cm’) for different sequences are sorted in ascending order. Evaluation criteria are the Length Error (LE) and the Location Deviation Error (LDE) of Inferolateral/Anteroseptal Landmarks (IL/AL) Method | Frame | Criterion(cm) | Mean$\pm$ std | min | $25\%$ | Median | $75\%$ | $90\%$ | max ---|---|---|---|---|---|---|---|---|--- | | LDE of AL | 1.28 $\pm$ 1.43 | 0.01 | 0.51 | 0.96 | 1.60 | 2.26 | 11.71 Proposed RDT | end-diastolic | LDE of IL | 1.16$\pm$1.27 | 0.06 | 0.40 | 0.88 | 1.48 | 2.15 | 10.67 | | LE of LVID | 0.81$\pm$1.04 | 0.00 | 0.25 | 0.51 | 1.00 | 1.63 | 8.07 CenterNet | | LDE of AL | 1.79 $\pm$ 1.96 | 0.07 | 0.72 | 1.21 | 2.04 | 3.81 | 13.62 [25] | end-diastolic | LDE of IL | 1.71$\pm$1.82 | 0.09 | 0.61 | 1.34 | 2.49 | 3.13 | 12.60 | | LE of LVID | 1.22$\pm$1.94 | 0.03 | 0.44 | 1.15 | 1.81 | 2.24 | 10.33 Unet+C-Ynet | | LDE of AL | 2.29$\pm$3.02 | 0.05 | 0.68 | 1.41 | 2.35 | 5.21 | 19.01 [26] | end-diastolic | LDE of IL | 3.72$\pm$4.05 | 0.07 | 0.78 | 1.91 | 5.26 | 10.89 | 18.81 | | LE of LVID | 2.39$\pm$2.61 | 0.00 | 0.64 | 1.38 | 3.28 | 5.77 | 12.16 Modified U-Net | | LDE of AL | 5.15$\pm$4.86 | 0.10 | 1.27 | 2.99 | 8.18 | 12.72 | 19.76 [24] | end-diastolic | LDE of IL | 5.36$\pm$4.74 | 0.03 | 1.01 | 4.13 | 8.86 | 12.31 | 17.22 | | LE of LVID | 3.40 $\pm$ 3.02 | 0.02 | 0.97 | 2.49 | 5.07 | 7.63 | 15.17 | | LDE of AL | 1.44 $\pm$ 1.30 | 0.06 | 0.66 | 1.16 | 1.75 | 2.67 | 10.37 Proposed RDT | end-systolic | LDE of IL | 1.13$\pm$1.22 | 0.06 | 0.51 | 0.90 | 1.25 | 1.89 | 10.10 | | LE of LVID | 1.09$\pm$0.95 | 0.00 | 0.37 | 0.90 | 1.51 | 2.43 | 5.81 CenterNet | | LDE of AL | 1.90 $\pm$ 1.64 | 0.09 | 0.98 | 1.73 | 2.98 | 3.75 | 13.57 [25] | end-systolic | LDE of IL | 2.03$\pm$2.21 | 0.12 | 0.92 | 1.98 | 3.68 | 4.42 | 14.54 | | LE of LVID | 1.83$\pm$1.48 | 0.06 | 0.95 | 1.78 | 2.93 | 4.31 | 9.25 Unet+C-Ynet | | LDE of AL | 2.78$\pm$2.87 | 0.14 | 0.98 | 1.82 | 3.29 | 5.85 | 19.8 [26] | end-systolic | LDE of IL | 3.42$\pm$3.80 | 0.06 | 0.78 | 1.74 | 4.71 | 9.73 | 17.33 | | LE of LVID | 2.45$\pm$2.61 | 0.00 | 0.73 | 1.41 | 3.02 | 5.14 | 11.51 Modified U-Net | | LDE of AL | 5.05$\pm$4.34 | 0.16 | 1.42 | 2.90 | 8.47 | 12.04 | 16.79 [24] | end-systolic | LDE of IL | 5.72$\pm$4.59 | 0.03 | 1.70 | 4.65 | 9.26 | 12.24 | 17.91 | | LE of LVID | 3.87$\pm$3.21 | 0.03 | 1.64 | 3.02 | 5.18 | 8.39 | 19.38 Comparison with state-of-the-art methods is reported in Table 6. Our results verify that our detection on the end-diastolic frame performs best over compared methods. Results also demonstrate that errors in end-systolic and end-diastolic frames are of the same range, suggesting that the tracking error is not accumulative over in-between unannotated frames. Figure 5: Four examples of frames with median LDE. The predicted LVID is the orange color line with landmarks in yellow color, while the ground truth LVID is the green color line with landmarks in red color. ### 5.3 Qualitative Results with Visualized Examples Fig. 5 shows four examples with the location error around the median. Here the Location Deviation Error (LDE) is the average location error of Inferolateral/Anteroseptal Landmarks (AL and IL), as there are no cases in our test data for which the AL and IL are both at the median. For the end-systolic frame, the average LDE is 0.95$\pm$0.68 cm for mean $\pm$ std, and 0.85 cm for the median. For the end-diastolic frame, the average LDE is with 0.91$\pm$0.66 cm for mean $\pm$ std, and 0.82 cm for the median. ### 5.4 Ablation Study In our ablation study, we verify the effectiveness of the adversarial training (Ad-T), as well as the reciprocal loss (Rec-L). Without the reciprocal loss, the structural information of in-between unannotated frames is ignored. As Ad-T is based on Rec-L, without Rec-L the Ad-T cannot be achieved. A detailed comparison can be found in Table 2. Table 2: Ablation study for Ad-T and Rec-L. Frame | Criterion(cm) | Ad-T | Rec-L | mean | median ---|---|---|---|---|--- | LDE-AL | $\color[rgb]{0,0,1}{\times}$ | $\color[rgb]{0,0,1}{\times}$ | 3.22 | 4.50 | LDE-IL | $\color[rgb]{0,0,1}{\times}$ | $\color[rgb]{0,0,1}{\times}$ | 5.02 | 6.74 | LE | $\color[rgb]{0,0,1}{\times}$ | $\color[rgb]{0,0,1}{\times}$ | 2.65 | 3.53 | LDE-AL | $\color[rgb]{0,0,1}{\times}$ | $\color[rgb]{1,0,0}{\checkmark}$ | 1.70 | 1.95 ED | LDE-IL | $\color[rgb]{0,0,1}{\times}$ | $\color[rgb]{1,0,0}{\checkmark}$ | 1.76 | 2.02 | LE | $\color[rgb]{0,0,1}{\times}$ | $\color[rgb]{1,0,0}{\checkmark}$ | 1.00 | 1.04 | LDE-AL | $\color[rgb]{1,0,0}{\checkmark}$ | $\color[rgb]{1,0,0}{\checkmark}$ | 1.28 | 0.96 | LDE-IL | $\color[rgb]{1,0,0}{\checkmark}$ | $\color[rgb]{1,0,0}{\checkmark}$ | 1.16 | 0.88 | LE | $\color[rgb]{1,0,0}{\checkmark}$ | $\color[rgb]{1,0,0}{\checkmark}$ | 0.81 | 0.51 | LDE-AL | $\color[rgb]{0,0,1}{\times}$ | $\color[rgb]{0,0,1}{\times}$ | 3.17 | 3.85 | LDE-IL | $\color[rgb]{0,0,1}{\times}$ | $\color[rgb]{0,0,1}{\times}$ | 4.79 | 6.94 | LE | $\color[rgb]{0,0,1}{\times}$ | $\color[rgb]{0,0,1}{\times}$ | 2.36 | 3.47 | LDE-AL | $\color[rgb]{0,0,1}{\times}$ | $\color[rgb]{1,0,0}{\checkmark}$ | 1.76 | 1.92 ES | LDE-IL | $\color[rgb]{0,0,1}{\times}$ | $\color[rgb]{1,0,0}{\checkmark}$ | 1.88 | 1.94 | LE | $\color[rgb]{0,0,1}{\times}$ | $\color[rgb]{1,0,0}{\checkmark}$ | 1.41 | 1.54 | LDE-AL | $\color[rgb]{1,0,0}{\checkmark}$ | $\color[rgb]{1,0,0}{\checkmark}$ | 1.44 | 1.75 | LDE-IL | $\color[rgb]{1,0,0}{\checkmark}$ | $\color[rgb]{1,0,0}{\checkmark}$ | 1.13 | 1.25 | LE | $\color[rgb]{1,0,0}{\checkmark}$ | $\color[rgb]{1,0,0}{\checkmark}$ | 1.09 | 0.90 Table 2 shows that the reciprocal loss substantially improves the framework. By adding the reciprocal loss, the errors decrease around 2 cm for all different criteria. The results again improve a lot when the model is trained with our proposed Ad-T method. ### 5.5 Model Extension We also test our proposed model’s extension ability, in which we try only to use one frame (end-diastolic) in each sequence for model training. Such training would start from the first annotated frame, and then track in a cycle way. The motion loss and the focal loss in the last frame are not available in such a model. The model is mainly trained by the reciprocal loss from the unannotated frames and the focal loss as well as the cycle loss in the annotated frame (i.e., the end-diastolic frame). Detailed results are reported in Table 3. We simply use the medium value of two LDEs (AL and IL) to represent the LDE. Table 3: Statistics analysis for model trained by one frame only. Frame | Criterion(cm) | Mean$\pm$ std | min | Medium ---|---|---|---|--- ED | LDE | 1.59$\pm$1.85 | 0.04 | 0.95 | LE | 1.04$\pm$1.20 | 0.02 | 0.68 ES | LDE | 1.76$\pm$1.49 | 0.10 | 1.34 | LE | 1.77$\pm$1.39 | 0.01 | 1.49 We can find even with only one frame annotated, the proposed model can get satisfying results, when compared with the state-of-the-art. However, results on end-systolic are much worse than on end-diastolic, which means the second annotated frame affects the tracking branch a lot. Table 4: Analysis for the annotation sparsity. Annotation rate | 5-8 | 8-12 | 12-16 | 16-20 ---|---|---|---|--- Average LDE (cm)/ sequence | 0.23 | 0.26 | 0.24 | 0.27 Average LDE (cm) /frame | 0.031 | 0.025 | 0.021 | 0.018 Table 5: Analysis for the sparsity of reciprocal loss. Loss rate | 2 | 3 | 4 | 5 ---|---|---|---|--- Average LDE (cm)/ sequence | 0.46 | 0.25 | 0.29 | 0.38 ### 5.6 Annotation Sparsity Analysis Sparsity of annotation: As the number of in-between unannotated frames is random ranging from 5 to 20, such number may influence the tracking branch, while the detection branch may not be affected. Therefore, to analyze the influence of annotation sparsity on tracking, we just start from the ground truth location of the first frame (end-diastolic) to do the tracking. The whole model does not change. We get the predicted location in the second annotated frame (end-systolic), and use the location deviation errors (LE) on this frame for different sequences with different annotation sparsity for evaluation. Results can be found in Table 4. We observe that the proposed method is not affected by the annotation sparsity. The Average LDEs for different sequences are generally the same around 0.25 cm. The average LDE/frame is the Average LDE divided by the in- between frame number. As the reciprocal loss is generated every three frames, whenever a large error is generated from the tracking branch, the discrepancy between detected and tracked location will also be large, which brings a significant reciprocal loss. Such loss overcomes the problem brought by large annotation sparsity. Sparsity of reciprocal loss: The frequency for applying reciprocal loss for in-between unannotated frames is also important. A comparison can be found in Table 5. We can find that only when the reciprocal loss is applied every three frames, the results are best. Therefore, we empirically set such rate as 3. The reciprocal loss should not be applied too densely or with a large sparsity. ### 5.7 Failed Cases Analysis There are still a few failure cases in our current result. An example is shown in Fig. 6. The result is with the maximum LDE (LDE of AL: 5.02 cm, LDE of IL: 8.07 cm, LD: 0.48cm). We hypothesize that the reason for such failure is as follows: During the image acquisition, the operator appears to have zoomed the ultrasound image on the LV. Hence, no other cardiac chamber is clearly visible and the appearance of the image is substantially different from a typical PLAX image. A much larger training data set will be required to avoid failure in such cases. If we set 2 cm error for the average LDE (average of IL and AL) as the critical point for failure, for the end-systolic frame the failure percentage is $6.1\%$, while for the end-diastolic frame the percentage is $3.7\%$. These are promising results, compared with the results in [24] whose failure is $6.7\%$. We note that the model in [24] is trained by densely annotated sequences, instead of sparsely annotated sequences as in our method. Figure 6: An example of a discrepant case. This PLAX view is suboptimal and has been imaged at a low imaging window on the chest resulting altered axis of the LV. The ground truth LVID label (shown in green color), used clinically, has been placed in an atypical position based on operator judgment (closer to the apex) to account for the altered geometry. The predicted LVID is the orange color line with landmarks in yellow color. It should be noted that despite relatively large LDE error, both measurements are likely clinical acceptable, as the distance between AL and IL, rather than their absolute image coordinates, is the main metric used to measure EF. ### 5.8 Ejection Fraction Error Analysis We also analyze the proposed method from medical perspective. We calculate the Ejection Fraction error on the testing dataset. Ejection fraction (EF) is a measurement, expressed as a percentage, of how much blood the left ventricle pumps out with each contraction. The ejection fraction represents the percent of the total amount of blood being pushed out with each heartbeat in the left ventricle. The Ejection Fraction (EF) is formulated as $EF=100\times(ED_{vol}-ES_{vol})/ED_{vol},$ (12) in which $ED_{vol}$ and $ES_{vol}$ are End-diastolic volume and End-systolic volume respectively, which are formulated by Teichholz formula as below $ED_{vol}=7\times EDD/(2\textperiodcentered 4+EDD),$ (13) $ES_{vol}=7\times ESD/(2\textperiodcentered 4+ESD).$ (14) Here EDD means the length of LV in the end-diastolic frame and the end- systolic frame respectively. The EF error is the difference between the predicted EF and ground truth EF. The calculated result can be found in Table 6. Also, we draw the EF scatter as can be found in Fig. 7. Table 6: Statistical result of EF error for the proposed method. Results | Mean$\pm$ std | min | Median | 90% ---|---|---|---|--- EF prediction | 37.08 $\pm$17.39 | 0.59 | 37.25 | 63.75 EF error | 19.00 $\pm$26.25 | 0.02 | 12.28 | 39.21 Figure 7: The EF scatter plot for the proposed method. ## 6 Conclusion In this paper, we proposed a novel reciprocal landmark detection and tracking model. The model is designed to tackle the data and annotation scarcity problem for ultrasound sequences. The model achieves reliable landmark detection and tracking with only around 2,000 annotated frames (995 sequences) for training. For each sequence, only two key frames are annotated. The model is optimized by a novel adversarial training way, which can better exploit the training data’s limited information. The comparison with state-of-the-art and analysis of results verify the effectiveness of our proposed method. ## References * [1] A. Thorstensen, H. Dalen, B. H. Amundsen, S. A. Aase, and A. Stoylen, “Reproducibility in echocardiographic assessment of the left ventricular global and regional function, the hunt study,” _European Journal of Echocardiography_ , vol. 11, no. 2, pp. 149–156, 2010. * [2] O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in _International Conference on Medical Image Computing and Computer-Assisted Intervention_. Springer, 2015, pp. 234–241. * [3] F. Milletari, N. Navab, and S.-A. Ahmadi, “V-net: Fully convolutional neural networks for volumetric medical image segmentation,” in _2016 Fourth International Conference on 3D Vision (3DV)_. IEEE, 2016, pp. 565–571. * [4] M. Avendi, A. Kheradvar, and H. Jafarkhani, “A combined deep-learning and deformable-model approach to fully automatic segmentation of the left ventricle in cardiac mri,” _Medical Image Analysis_ , vol. 30, pp. 108–119, 2016. * [5] O. Oktay, E. Ferrante, K. Kamnitsas, M. Heinrich, W. Bai, J. Caballero, S. A. Cook, A. De Marvao, T. Dawes, D. P. O‘Regan _et al._ , “Anatomically constrained neural networks (acnns): application to cardiac image enhancement and segmentation,” _IEEE Transactions on Medical Imaging_ , vol. 37, no. 2, pp. 384–395, 2017. * [6] T. A. Ngo, Z. Lu, and G. Carneiro, “Combining deep learning and level set for the automated segmentation of the left ventricle of the heart from cardiac cine magnetic resonance,” _Medical Image Analysis_ , vol. 35, pp. 159–171, 2017. * [7] W. Bai, O. Oktay, M. Sinclair, H. Suzuki, M. Rajchl, G. Tarroni, B. Glocker, A. King, P. M. Matthews, and D. Rueckert, “Semi-supervised learning for network-based cardiac mr image segmentation,” in _International Conference on Medical Image Computing and Computer-Assisted Intervention_. Springer, 2017, pp. 253–260. * [8] W. Bai, M. Sinclair, G. Tarroni, O. Oktay, M. Rajchl, G. Vaillant, A. M. Lee, N. Aung, E. Lukaschuk, M. M. Sanghvi _et al._ , “Automated cardiovascular magnetic resonance image analysis with fully convolutional networks,” _Journal of Cardiovascular Magnetic Resonance_ , vol. 20, no. 1, p. 65, 2018. * [9] L. Yao, J. Prosky, E. Poblenz, B. Covington, and K. Lyman, “Weakly supervised medical diagnosis and localization from multiple resolutions,” _Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition_ , 2018. * [10] T. Pfister, J. Charles, and A. Zisserman, “Flowing convnets for human pose estimation in videos,” in _Proceedings of the IEEE International Conference on Computer Vision_ , 2015, pp. 1913–1921. * [11] J. J. Tompson, A. Jain, Y. LeCun, and C. Bregler, “Joint training of a convolutional network and a graphical model for human pose estimation,” in _Advances in Neural Information Processing Systems_ , 2014, pp. 1799–1807. * [12] C. Payer, D. Štern, H. Bischof, and M. Urschler, “Regressing heatmaps for multiple landmark localization using cnns,” in _International Conference on Medical Image Computing and Computer-Assisted Intervention_. Springer, 2016, pp. 230–238. * [13] N. Savioli, M. S. Vieira, P. Lamata, and G. Montana, “Automated segmentation on the entire cardiac cycle using a deep learning work-flow,” in _2018 Fifth International Conference on Social Networks Analysis, Management and Security (SNAMS)_. IEEE, 2018, pp. 153–158. * [14] F. T. Dezaki, Z. Liao, C. Luong, H. Girgis, N. Dhungel, A. H. Abdi, D. Behnami, K. Gin, R. Rohling, and P. Abolmaesumi, “Cardiac phase detection in echocardiograms with densely gated recurrent neural networks and global extrema loss,” _IEEE Transactions on Medical Imaging_ , vol. 38, no. 8, pp. 1821–1832, 2018. * [15] M. Sofka, F. Milletari, J. Jia, and A. Rothberg, “Fully convolutional regression network for accurate detection of measurement points,” in _Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support_. Springer, 2017, pp. 258–266. * [16] X. Du, S. Yin, R. Tang, Y. Zhang, and S. Li, “Cardiac-deepied: Automatic pixel-level deep segmentation for cardiac bi-ventricle using improved end-to-end encoder-decoder network,” _IEEE Journal of Translational Engineering in Health and Medicine_ , vol. 7, pp. 1–10, 2019. * [17] M. Li, W. Zhang, G. Yang, C. Wang, H. Zhang, H. Liu, W. Zheng, and S. Li, “Recurrent aggregation learning for multi-view echocardiographic sequences segmentation,” in _International Conference on Medical Image Computing and Computer-Assisted Intervention_. Springer, 2019, pp. 678–686. * [18] W. Yan, Y. Wang, Z. Li, R. J. Van Der Geest, and Q. Tao, “Left ventricle segmentation via optical-flow-net from short-axis cine mri: preserving the temporal coherence of cardiac motion,” in _International Conference on Medical Image Computing and Computer-Assisted Intervention_. Springer, 2018, pp. 613–621. * [19] M. H. Jafari, H. Girgis, Z. Liao, D. Behnami, A. Abdi, H. Vaseli, C. Luong, R. Rohling, K. Gin, T. Tsang _et al._ , “A unified framework integrating recurrent fully-convolutional networks and optical flow for segmentation of the left ventricle in echocardiography data,” in _Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support_. Springer, 2018, pp. 29–37. * [20] S. Chen, K. Ma, and Y. Zheng, “Tan: Temporal affine network for real-rime left ventricle anatomical structure analysis based on 2d ultrasound videos,” _arXiv preprint arXiv:1904.00631_ , 2019. * [21] C. Qin, W. Bai, J. Schlemper, S. E. Petersen, S. K. Piechnik, S. Neubauer, and D. Rueckert, “Joint learning of motion estimation and segmentation for cardiac mr image sequences,” in _International Conference on Medical Image Computing and Computer-Assisted Intervention_. Springer, 2018, pp. 472–480. * [22] H. Wei, H. Cao, Y. Cao, Y. Zhou, W. Xue, D. Ni, and S. Li, “Temporal-consistent segmentation of echocardiography with co-learning from appearance and shape,” in _International Conference on Medical Image Computing and Computer-Assisted Intervention_. Springer, 2020, pp. 1–8. * [23] T.-Y. Lin, P. Goyal, R. Girshick, K. He, and P. Dollár, “Focal loss for dense object detection,” in _Proceedings of the IEEE International Conference on Computer Vision_ , 2017, pp. 2980–2988. * [24] A. Gilbert, M. Holden, L. Eikvil, S. A. Aase, E. Samset, and K. McLeod, “Automated left ventricle dimension measurement in 2d cardiac ultrasound via an anatomically meaningful cnn approach,” in _Smart Ultrasound Imaging and Perinatal, Preterm and Paediatric Image Analysis_. Springer, 2019, pp. 29–37. * [25] X. Zhou, D. Wang, and P. Krähenbühl, “Objects as points,” _arXiv preprint arXiv:1904.07850_ , 2019. * [26] J. Lin, Y. Zhang, A. Amadou, I. Voigt, T. Mansi, and R. Liao, “Cycle ynet: Semi-supervised tracking of 3d anatomical landmarks,” in _International MICCAI Workshop on Machine Learning in Medical Imaging_. Springer, 2020, pp. 1–8.
# Controlling core-hole lifetime through an x-ray planar cavity Xin-Chao Huang,1 Xiang-Jin Kong,2 Tian-Jun Li,1 Zi-Ru Ma,1 Hong-Chang Wang,3 Gen-Chang Liu,4 Zhan-Shan Wang,4 Wen-Bin Li,4<EMAIL_ADDRESS>Lin-Fan Zhu,1 <EMAIL_ADDRESS>1Hefei National Laboratory for Physical Sciences at Microscale and Department of Modern Physics, University of Science and Technology of China, Hefei, Anhui 230026, People’s Republic of China 2Department of Physics, National University of Defense Technology, Changsha, Hunan 410073, People’s Republic of China 3Diamond Light Source, Harwell Science and Innovation Campus, Didcot, Oxfordshire, OX11 0DE, UK 4MOE Key Laboratory of Advanced Micro-Structured Materials, Institute of Precision Optical Engineering (IPOE), School of Physics science and Engineering, Tongji University, Shanghai 200092, People’s Republic of China ###### Abstract It has long been believed that core-hole lifetime (CHL) of an atom is an intrinsic physical property, and controlling it is significant yet is very hard. Here, CHL of the 2$p$ state of W atom is manipulated experimentally through adjusting the emission rate of a resonant fluorescence channel with the assistance of an x-ray thin-film planar cavity. The emission rate is accelerated by a factor linearly proportional to the cavity field amplitude, that can be directly controlled by choosing different cavity modes or changing the angle offset in experiment. This experimental observation is in good agreement with theoretical predictions. It is found that the manipulated resonant fluorescence channel even can dominate the CHL. The controllable CHL realized here will facilitate the nonlinear investigations and modern x-ray scattering techniques in hard x-ray region. PACS: 32.80.-t, 32.80.Qk, 42.50.Ct, 32.30.Rj, 78.70.Ck. The particularity of an inner-shell excitation or ionization is to produce a core vacancy, which has a finite lifetime, i.e., the so-called core-hole lifetime (CHL), and then it decays into lower-lying states. There are two main relaxation pathways, i.e., radiative (fluorescence) and non-radiative (Auger decay or autoionization) channels, and the CHL is determined by the total decay rate of all relaxation channels. Normally, Auger effect dominates the decay routes of K shell for low-Z atoms Auger (1925) and L and M shells for higher-Z atoms Krause (1979), so the CHL is sometimes called Auger lifetime. The CHL has long been considered as an intrinsic factor and controlling it is very difficult because the relaxation channels are hard to be manipulated with common methods. Nevertheless, an adjustable CHL is strongly desired, since CHL changes are useful to detect ultrafast dynamics. An adjustable CHL is needed to give a deep insight to nonlinear light-matter interaction with the advent of x-ray free electron laser (XFEL), since the ratio of CHL to XFEL pulse width does matter for multiphoton ionization Fukuzawa _et al._ (2013), two-photon absorption Tamasaku _et al._ (2018), population inversion Yoneda _et al._ (2015) and stimulated emission Wu _et al._ (2016); Chen _et al._ (2018). CHL is also a key factor in resonant x-ray scattering (RXS) process Gel’mukhanov and Ågren (1999), where the dynamics of the core-excited state is controlled by the duration time which is determined by both energy detuning and CHL Gel’mukhanov _et al._ (1999). Because of a lack of an efficient method to manipulate CHL experimentally, the controlling schemes for duration time were based on the energy detuning up to now Skytt _et al._ (1996); Feifel _et al._ (2004); Kimberg _et al._ (2013); Feifel and Piancastelli (2011); Morin and Miron (2012); Miron and Morin (2011). The dynamics of the core-excited state determines the application range for RXS techniques, e.g., resonant inelastic x-ray scattering (RIXS) Ament _et al._ (2011). Since Coulomb interaction between core-hole and valence electrons only exists during the existence of core-excited state, the relative timescale between CHL and elementary excitations governs the effectiveness of indirect RXIS Ament _et al._ (2007, 2009); Dean _et al._ (2012), especially for charge and magnon excitations van den Brink (2007); Ament _et al._ (2009); Haverkort (2010); Jia _et al._ (2016); Tohyama and Tsutsui (2018). In time-resolved RIXS (tr- RIXS), CHL also needs to be flexibly adjusted for pursuing higher time resolution Dean _et al._ (2016); Wang _et al._ (2018); Chen _et al._ (2019); Buzzi _et al._ (2018). Therefore, a controllable CHL will be very useful thus is strongly wished for, from both fundamental and application perspectives. Because CHL is determined by the total decay rate of all relaxation channels, controlling CHL means manipulatable decay channels, at least one of them, which is a challenging task. Stimulated emission channel could be opened by intense and short x-ray pulses to accelerate CHL Wu _et al._ (2016); Chen _et al._ (2018), while such scheme can only be implemented in XFEL. The present work proposes another scheme that controls the spontaneous emission channel. R. Feynman once said, the theory behind chemistry is quantum electrodynamics (QED) Richard (1985), indicating that the spontaneous emission rate of atom depends on the environment (photonic density of states). A cavity is such an outstanding system to robustly structure environment and modify the spontaneous emission rate in visible wavelength regime Tomaš (1995); Raimond _et al._ (2001), as known cavity-QED. With the dramatic progress of new generation x-ray source and thin-film technology, cavity-QED effect in hard x-ray range was demonstrated by the laboratory of thin film planar cavity with nuclear ensembles Röhlsberger _et al._ (2010, 2012); Heeg _et al._ (2013, 2015a, 2015b); Haber _et al._ (2017) or electronic resonance Haber _et al._ (2019), which breeds the new field of x-ray quantum optics Adams _et al._ (2013). In this work, a controllable CHL for 2$p$ state of W atom is realized through adjusting the emission rate of a resonant fluorescence channel with the assistance of an x-ray thin-film planar cavity. WSi2 has a remarkable white line around the L${}_{\textrm{III}}$ edge of W, which is a resonant channel and generally known to be associated with an atomic-like electric dipole allowed transition, from an inner shell 2$p$ to an unoccupied level 5$d$ Haber _et al._ (2019); Brown _et al._ (1977); Wei and Lytle (1979). Inside the cavity, the emission rate of the resonant channel depends on the photonic density of states where the atom locates, which can be modified by the cavity field amplitude in experiment. Because the thin-film planar cavity can only enhance the photonic density of states, but not suppress it, only CHL shortening is realized in the present experiment. As long as the cavity effect is strong enough, the total decay rate will have measurable changes and lead to an controllable CHL. Fig. 1: The schematic for controlling core-hole lifetime. (a) Cavity sample and measurement setup. The cavity has a structure of Pt (2.1 nm)/C (18.4 nm)/WSi2 (2.8 nm)/C (18.0 nm)/Pt (16.0 nm)/Si100, and the middle-right inset shows the energy-level of L${}_{\textrm{III}}$ edge of atom W. The sample is probed by a monochromatic x-ray, and the resonant fluorescence is measured in the reflection direction by a CCD and the inelastic fluorescence signals are collected by an energy-resolved fluorescence detector. The distance between collimator and sample surface is 31.0 mm, and the hole diameter and the length of the collimator are 2.8 mm and 20.1 mm. An example of full range fluorescence spectrum is shown in inset at top-left side, and the grey region corresponds to the fluorescence photon energy of Lα line. The inset at top- right side is the reflectivity curve with an incident energy detuning 30 eV from $E_{0}$, and the pink solid bar indicates the critical angle of Pt (0.46 degree). (b) The values of Re($\eta$) and Im($\eta$) as a function of incident angle, which is calculated by a transfer matrix formulism. (c) The simplified energy levels of W. The driving is labeled by the blue arrow, and the cavity enhanced emission is labeled by the red thick arrow. The inelastic fluorescence decay is labeled by the red thin arrow. Fig. 1(a) depicts the cavity structure used in the present work. The thin-film cavity is made of a multilayer of Pt and C. The top and bottom layers of Pt with a high electron density are used as mirrors. The layers of C in the middle with a low electron density are used to guide the x-ray and to stack the cavity space. In this design, at certain incident angles $\theta_{\textrm{th}}$ below the critical angle of Pt, x-ray can resonantly excite specific cavity guided modes where dips in the reflectivity curve appear as shown in the top-right inset of Fig. 1(a). In the present work, $\theta_{\textrm{th}}$ are $\theta_{\textrm{1st}}$=0.218∘, $\theta_{\textrm{3rd}}$=0.312∘ and $\theta_{\textrm{5th}}$=0.440∘ for the 1${}^{\textrm{st}}$, 3${}^{\textrm{rd}}$ and 5${}^{\textrm{th}}$ odd orders of cavity mode. Then the coupling between the cavity and atom is built by embedding a thin layer of WSi2 at the middle of the cavity where the cavity field amplitudes are the strongest. The field distributions of the 1${}^{\textrm{st}}$, 3${}^{\textrm{rd}}$ and 5${}^{\textrm{th}}$ orders of cavity mode are sketched in Fig. 1(a). As shown in the middle-right inset of Fig. 1(a), the inner shell energy-level system is different from the simple two-level one, and both resonant channel and incoherent processes such like inelastic radiative channels (Auger decay channels is not exhibited here) can annihilate the core vacancy state, so the decay width is determined by the total decay rates of all relaxation channels. Excited by the incoming x-ray field, the atomic dipole emits the resonant fluorescence through the resonant channel, and the resonant response could be written as a simple form of Lorentz function, $f=-f_{0}\frac{i\gamma_{\textrm{re}}/2}{\delta+i(\gamma_{\textrm{re}}/2+\gamma_{\textrm{in}}/2)}$ (1) The electronic continuum in higher energy range is not considered here. $f_{0}$ is a constant, and $\delta$ is the energy detuning between the incident x-ray energy $E$ and the white line transition energy $E_{0}$. $\gamma_{\textrm{re}}$ is the natural spontaneous emission rate of the resonant channel, while $\gamma_{\textrm{in}}$ is the incoherent decay rate which sums two branches: the radiative decay rate of inelastic channels $\gamma_{\textrm{ie}}$ and the non-radiative decay rate of Auger process $\gamma_{\textrm{A}}$, i.e., $\gamma_{\textrm{in}}=\gamma_{\textrm{ie}}+\gamma_{\textrm{A}}$. It is clear that the inverse core-hole lifetime is expressed by the natural width as $\gamma=\gamma_{\textrm{re}}+\gamma_{\textrm{in}}$. Cavity strengthens the photonic density of states Tomaš (1995); Röhlsberger _et al._ (2005) at the position of the radiating atom, so the resonant fluorescence will be enhanced. Applying the transfer matrix combined with a perturbation expansion method (SM Sec. I), the resonant fluorescence in the reflection direction is solved as, ${{r}_{a}}=-\frac{id{{f}_{0}}\times{{\left|{{a}^{z_{a}}}\right|}^{2}}{{\gamma}_{\text{re}}}/2}{\delta+{{\delta}_{c}}+i\left({{\gamma}_{c}}+\gamma\right)/2}$ (2) $d$ is the thickness of the atomic layer, and ${{\left|{{a}^{z_{a}}}\right|}^{2}}$ is the field intensity where the atom locates. It can be seen that Eq. (2) still has a Lorentzian resonant response, while contains additional cavity effects: the cavity enhanced emission rate $\gamma_{c}$ and the cavity induced energy shift $\delta_{c}$, $\begin{array}[]{lll}{{\gamma}_{c}}&=d{{f}_{0}}\gamma_{\textrm{re}}\times\operatorname{Re}\left(\eta\right)\\\ {{\delta}_{c}}&=d{{f}_{0}}\gamma_{\textrm{re}}\times\operatorname{Im}\left(\eta\right)\\\ \eta&=pq\\\ \end{array}$ (3) Thus the emission rate is enhanced by a factor of Re($\eta$), where $p$ and $q$ are the field amplitudes corresponding to the wave scattered from up (down) direction into both up and down directions at the position of atomic layer (Sec. I of SM). Note here that the photonic density of states is directly related to the cavity field amplitudes Röhlsberger _et al._ (2005, 2012), so Eq. (3) conforms to the typical cavity Purcell effect Raimond _et al._ (2001) which describes the well-known linear relation between lifetime shortening and photonic density of states strengthening. It is clear that the real part of $\eta$ is an essential factor to control the enhanced emission rate, and the energy shift is modified by the image part of $\eta$. The real and image parts of $\eta$ as a function of incident angle are depicted in Fig. 1(b), and $\gamma_{c}$ and $\delta_{c}$ are simultaneously modified by the incident angle around the mode angles, which has been observed by Haber _et al_ recently Haber _et al._ (2019). On the other hand, Fig. 1(b) suggests that the strongest enhanced emission rate can be achieved without introducing additional energy shift by exactly choosing the angles of odd orders of cavity mode, which will be more convenient to study the individual influence of the CHL on core-hole dynamics (SM Sec. IV). The fully controllable resonant channel makes an adjustable total inverse core-hole lifetime, ${{\Gamma}_{n}}=\gamma_{c}+\gamma_{\textrm{re}}+\gamma_{\textrm{ie}}+\gamma_{\textrm{A}}$ (4) where all four contributions are included, herein $\gamma_{c}$ is the cavity enhanced emission rate, and $\gamma=\gamma_{\textrm{re}}+\gamma_{\textrm{ie}}+\gamma_{\textrm{A}}$ is the natural inverse CHL as the sum of three branches: the natural spontaneous emission rate of the resonant fluorescence channel, the radiative decay rate of inelastic fluorescence channels and the Auger decay rate. $\gamma$ is a fixed value which can be obtained from the experimental spectrum at a large incident angle (Fig. S3 of SM), i.e., $\gamma/2$=3.6 eV. As long as $\gamma_{c}$ is large enough, this controllable part will dominate the CHL. As shown in the simplified energy levels in Fig. 1(c), the core-hole lifetime determines the linewidth of inelastic scattering, i.e, the fluorescence spectrum. We employ a RXS formalism known as Kramers-Heisenberg equation to character the inelastic scattering Gel’mukhanov and Ågren (1999); Ament _et al._ (2011) as, ${{F}_{if}}\left(\overset{\scriptscriptstyle\rightharpoonup}{k},{\overset{\scriptscriptstyle\rightharpoonup}{k}}^{\prime},\omega,{\omega}^{\prime}\right)=\frac{\left\langle f\right|{\hat{D}}^{\prime}\left|n\right\rangle\left\langle n\right|\hat{D}\left|i\right\rangle}{\delta+i{{\Gamma}_{n}}/2}$ (5) Herein the initial state $\left|i\right\rangle=\left|g,\overset{\scriptscriptstyle\rightharpoonup}{k}\right\rangle$, the final state $\left|f\right\rangle=\left|f,{\overset{\scriptscriptstyle\rightharpoonup}{k}}^{\prime}\right\rangle$, and the intermediate state $\left|n\right\rangle=\left|e,0\right\rangle$. $\overset{\scriptscriptstyle\rightharpoonup}{k}$ is the wave vector and $\hat{D}$ is the transition operator. For the present system, an intermediate state and a final state are considered, since we choose the L${}_{\textrm{III}}$ white line transition and measure Lα with energy $E^{\prime}$. Eq. (5) indicates that CHL changes can be monitored by the inelastic fluorescence spectrum, and in the present work Lα line (Lα1 is much stronger than Lα2) is chosen to obtain the inelastic fluorescence spectra. The measurement was performed on the B16 Test beamline in Diamond Light Source. Monochromatic x-ray from a double crystal monochromator was used to scan the incident x-ray energy, and two small apart slits were used to obtain a good collimation beam with a small vertical beam size of about 50 $\mu$m. The multilayer was deposited onto a polished silicon wafer (100) using DC magnetron sputtering method which is popular to fabricate diverse cavity structures Röhlsberger _et al._ (2010, 2012); Heeg _et al._ (2013, 2015a, 2015b); Haber _et al._ (2017, 2019). Before sample fabrication, the deposition rate was calibrated carefully to guarantee the layer thickness with a good accuracy of better than 1 Å. The size of the wafer is 30$\times$30 mm2 which is larger than the footprint to avoid the beam overpassing the sample length. As shown in the top-right inset of Fig. 1(a), the $\theta-2\theta$ rocking curve with an incident energy detuning 30 eV from the white line position was measured firstly to find the desired specific incident angles corresponding to the 1${}^{\textrm{st}}$, 3${}^{\textrm{rd}}$ and 5${}^{\textrm{th}}$ orders of the guided modes, i.e., the corresponding reflection dips. For a given incident angle, the incident energy $E$ was scanned from 10161 eV to 10261 eV across the transition energy $E_{0}$=10208 eV. Then the reflectivity corresponding to the resonant channel was measured by a CCD detector, and the inelastic fluorescence lines were measured simultaneously by a silicon drift detector (_Vortex_) with a resolution of about 180 eV at a perpendicular direction. In front of the fluorescence detector, a collimator guarantees a constant detected area of the sample, and the footprint $bw/\textrm{sin}\theta$ on the sample surface is determined by the beam width $bw$ and the incident angle, so the inelastic fluorescence intensities need to be normalized by taking into account a geometry factor Li _et al._ (2012). The strongest Lα fluorescence lines (Lα1 at 8398 eV and Lα2 at 8335 eV) of W are far from the white line (10208 eV) and other weak fluorescence lines (9951 eV of Lβ2, 7387 eV of Ll and other negligible lines), so Lα can be extracted separately from the energy-resolved fluorescence spectrum. Firstly, the 1${}^{\textrm{st}}$, 3${}^{\textrm{rd}}$ and 5${}^{\textrm{th}}$ orders are exactly chosen to control the CHL without introducing additional energy shift, and the results are depicted in Fig. (2). Fig. 2(a) shows the experimental and theoretically fitted reflectivity curves. The present theoretical model for resonant fluorescence does not take into account the influence of the absorption edge due to its nature of the electronic continuum, and the continuum overlaps with the right side of the white line. The sudden increase of the absorption coefficient changes the refractive index and dramatically alters the cavity properties Haber _et al._ (2019). So the data below 10220 eV are selected for fitting (labeled by the green region). Above 10220 eV, the theoretical results diverge from the experimental datum. The reflection coefficient includes the contributions from two pathways (SM Sec. II): the first one of $r_{0}$ is from the multilayer cavity itself that the photon does not interact with the resonant atom, and the second one of $r_{a}$ is from the resonant atom inside the cavity, i.e., the resonant fluorescence. The linewidth of the cavity is much larger than the one of atom, which means that $r_{0}$ is more like a flat continuum state and $r_{a}$ is more like a sharp discrete state Liu _et al._ (2003); Heeg _et al._ (2015a). Therefore, the reflectivity spectrum is a result of Fano interference. It can be seen from Fig. 2(a) that the profile of the reflectivity spectra shows Fano line-shape. The reflectivity spectra give the values of $(\gamma_{c}+\gamma)/2$ as 7.9 eV, 6.9 eV and 5.2 eV for the 1${}^{\textrm{st}}$, 3${}^{\textrm{rd}}$ and 5${}^{\textrm{th}}$ orders of the cavity mode. Fig. 2: (a) The measured and theoretical reflectivity spectra for the 1${}^{\textrm{st}}$, 3${}^{\textrm{rd}}$ and 5${}^{\textrm{th}}$ orders as a function of incident photon energy. The red solid line is the theoretically fitted result. (b) The measured and fitted inelastic fluorescence spectra of Lα as a function of incident photon energy. The solid lines in pink, red, green and blue are the fitted result, Lorentzian resonance line, electronic continuum line and the flat background respectively. Fig. 2(b) shows the experimental and fitted inelastic fluorescence spectra of Lα as a function of incident x-ray energy for the 1${}^{\textrm{st}}$, 3${}^{\textrm{rd}}$ and 5${}^{\textrm{th}}$ orders. The inelastic fluorescence spectrum is fitted by a custom function combining a simple Lorentz function $L(E)$ with a Heaviside step function $H(E)$ (SM Sec.III), herein $L(E)$ with a linewidth $\Gamma_{n}/2$ is used to describe Eq. (5) and $H(E)$ is used to describe the absorption edge. The fitted values of $\Gamma_{n}/2$ are 8.6 eV, 6.2 eV and 5.1 eV which match well with the derived values from the resonant fluorescence spectra of Fig. 2(a), demonstrating that the shortening of CHL indeed comes from the regulation of resonant fluorescence channel. Moreover, the value of $\gamma_{c}=\Gamma_{n}-\gamma$ is even larger than $\gamma$ in the 1${}^{\textrm{st}}$ order, indicating that the adjustable resonant channel breaks the limitation of Auger processes and unchangeable radiative decay channels and dominantly determines CHL. A behavior of widening linewidth is cross-checked by Ll and Lβ2 lines (SM Fig. S5). Fig. 3: The inelastic fluorescence spectra of Lα as functions of incident photon energy and angle offset. Angle offset is the deviation between the incident angle and the $\theta_{\textrm{1st}}$ ($\theta_{\textrm{3rd}}$, $\theta_{\textrm{5th}}$). The measured inelastic fluorescence 2D spectra are shown in Fig. 3 for selected incident angles around the mode angles of the 1${}^{\textrm{st}}$, 3${}^{\textrm{rd}}$ and 5${}^{\textrm{th}}$ orders ($\theta_{\textrm{1st}}$=0.218∘, $\theta_{\textrm{3rd}}$=0.312∘ and $\theta_{\textrm{5th}}$=0.440∘ respectively). As discussed in Eq. (3), CHL and the cavity induced energy shift are simultaneously controlled by the incident angle. When the incident angle scans across the mode angle, a phenomenon of firstly increasing to maximum at the mode angle then decreasing of the inverse CHL will be observed along with an additional energy shift, which is demonstrated by Fig. 3. For the 3${}^{\textrm{rd}}$ order, the maximum linewidth does not seem to be where the angle offset is 0, this may due to the occasionally angle shift from the instabilities of the goniometer or sample holder. Note here that Fig. 3 suggests a way to continuously modify CHL but introduce additional energy shift. Fig. 4: (a) The measured and fitted inelastic fluorescence spectra of Lα for selected angle offsets. (b) Enhanced emission rate $\gamma_{c}$ as a function of Re$(\eta)$. The values of $\gamma_{c}$ are derived from the fitted linewidth $\Gamma_{n}$ as $\gamma_{c}=\Gamma_{n}-\gamma$, and the values of Re$(\eta)$ are obtained by transfer matrix calculation. The dashed blue line is a linear fitting of the experimental dots to guide the eyes. As predicted in Eq. (3), the enhanced emission rate $\gamma_{c}$ is linearly connected with the real part of the cavity filed amplitude $\eta$, and this is the essential to discuss the magnitude of inverse CHL. It should be noted here that the present method to control CHL is different from the scenario of stimulated emission Wu _et al._ (2016); Chen _et al._ (2018) where a non- linear relationship between the stimulated emission rate and the x-ray field intensity is expected. The present scheme actually employs a cavity to manipulate the enhanced spontaneous emission whose decay rate is linearly determined by the photonic density of states. The inelastic fluorescence spectra in Fig. 3 are fitted to get the values of $\Gamma_{n}$, and some selected spectra are shown in Fig. 4(a). Then the values of $\gamma_{c}$ are obtained based on Eq. (4), and the values of Re$(\eta)$ are calculated by the transfer matrix formulism. A good linear relationship between $\gamma_{c}$ and Re$(\eta)$ is depicted in Fig. 4(b) which is consistent with the prediction of Eq. (3). From the general viewpoint of cavity-QED in optical regime, the inelastic channel is an incoherent process which accelerates the decoherence, so it is regarded as a defect for the system Van Loo _et al._ (2013). However, the inelastic channel is a natural character and widely exists in atomic inner- shell systems, herein we demonstrate it can be very useful to monitor CHL changes, enriching the picture of cavity effect. In conclusion, the core-hole lifetime for 2$p$ state of W is manipulated experimentally through constructing an x-ray thin-film planar cavity system. The core-hole lifetime directly depends on the cavity field amplitude at the position of W atom (SM Sec. I and Huang _et al._ (2020)), which can be adjusted by choosing the different orders of cavity mode or varying the incident angle offset. With a high quality cavity sample, the core-hole lifetime is conveniently manipulated in experiment. Notably for the case of the 1${}^{\textrm{st}}$ order, the decay rate of the resonant channel is even stronger than the natural inverse core-hole lifetime which is dominated by the Auger process for L${}_{\textrm{III}}$ shell of atom W in common scenarios. Moreover, the inelastic fluorescence spectra are utilized as a good monitor to reflect the core-hole lifetime changes. The cavity structure is suitable for a wide range of x-ray energy from few to tens of keV, so the present scheme could be extended to a lot of elements which have resonant fluorescence channel. Utilizing the present cavity technique, the duration time of RXS process can be controlled not only by the energy detuning, but also by the core-hole lifetime, which will enrich the physical studies for RXS (SM Sec. VI) in future. Combing with the high-resolution $\sim$100 meV analyzer Hill _et al._ (2007), a cavity-manipulating RXS is expected to be achievable. This work is supported by National Natural Science Foundation of China (Grants No. U1932207), and the National Key Research and Development Program of China (Grants No. 2017YFA0303500 and 2017YFA0402300). The experiment was carried out in instrument B16 of Diamond Light Source Ltd (No. MM21446-1), United Kingdom. Authors thank Xiao-Jing Liu for fruitful discussion. ## References * Auger (1925) P. Auger, Comptes Rendus 180, 65 (1925). * Krause (1979) M. O. Krause, J. Phys. Chem. Ref. Data 8, 307 (1979). * Fukuzawa _et al._ (2013) H. Fukuzawa, S.-K. Son, K. Motomura, _et al._ , Phys. Rev. Lett. 110, 173005 (2013). * Tamasaku _et al._ (2018) K. Tamasaku, E. Shigemasa, Y. Inubushi, _et al._ , Phys. Rev. Lett. 121, 083901 (2018). * Yoneda _et al._ (2015) H. Yoneda, Y. Inubushi, K. Nagamine, _et al._ , Nature 524, 446 (2015). * Wu _et al._ (2016) B. Wu, T. Wang, C. E. Graves, _et al._ , Phys. Rev. Lett. 117, 027401 (2016). * Chen _et al._ (2018) Z. Chen, D. J. Higley, M. Beye, _et al._ , Phys. Rev. Lett. 121, 137403 (2018). * Gel’mukhanov and Ågren (1999) F. Gel’mukhanov and H. Ågren, Phys. Rep. 312, 87 (1999). * Ament _et al._ (2011) L. J. P. Ament, M. van Veenendaal, T. P. Devereaux, J. P. Hill, and J. van den Brink, Rev. Mod. Phys. 83, 705 (2011). * Gel’mukhanov _et al._ (1999) F. Gel’mukhanov, P. Sałek, T. Privalov, and H. Ågren, Phys. Rev. A 59, 380 (1999). * Skytt _et al._ (1996) P. Skytt, P. Glans, J.-H. Guo, K. Gunnelin, C. Såthe, J. Nordgren, F. K. Gel’mukhanov, A. Cesar, and H. Ågren, Phys. Rev. Lett. 77, 5035 (1996). * Feifel _et al._ (2004) R. Feifel, A. Baev, F. Gelmukhanov, _et al._ , Phys. Rev. A 69, 022707 (2004). * Kimberg _et al._ (2013) V. Kimberg, A. Lindblad, J. Söderström, O. Travnikova, C. Nicolas, Y. P. Sun, F. Gel’mukhanov, N. Kosugi, and C. Miron, Phys. Rev. X 3, 011017 (2013). * Feifel and Piancastelli (2011) R. Feifel and M. N. Piancastelli, J. Electron Spectrosc. Relat. Phenom. 183, 10 (2011). * Morin and Miron (2012) P. Morin and C. Miron, J. Electron Spectrosc. Relat. Phenom. 185, 259 (2012). * Miron and Morin (2011) C. Miron and P. Morin, _Handbook of High-Resolution Spectroscopy_ (Wiley, Chichester, UK, 2011) pp. 1655–1690. * Ament _et al._ (2007) L. J. P. Ament, F. Forte, and J. van den Brink, Phys. Rev. B 75, 115118 (2007). * Ament _et al._ (2009) L. J. P. Ament, G. Ghiringhelli, M. M. Sala, L. Braicovich, and J. van den Brink, Phys. Rev. Lett. 103, 117003 (2009). * Dean _et al._ (2012) M. Dean, R. Springell, C. Monney, _et al._ , Nat. Mater. 11, 850 C854 (2012) . * van den Brink (2007) J. van den Brink, Euro. Phys. Lett. 80, 47003 (2007). * Haverkort (2010) M. W. Haverkort, Phys. Rev. Lett. 105, 167404 (2010). * Jia _et al._ (2016) C. Jia, K. Wohlfeld, Y. Wang, B. Moritz, and T. P. Devereaux, Phys. Rev. X 6, 021020 (2016). * Tohyama and Tsutsui (2018) T. Tohyama and K. Tsutsui, Inter. J. Mod. Phys. B 32, 1840017 (2018). * Dean _et al._ (2016) M. P. Dean, Y. Cao, X. Liu, _et al._ , Nat. Mat. 15, 601 (2016). * Wang _et al._ (2018) Y. Wang, M. Claassen, C. D. Pemmaraju, C. Jia, B. Moritz, and T. P. Devereaux, Nat. Rev. Mat. 3, 312 (2018). * Chen _et al._ (2019) Y. Chen, Y. Wang, C. Jia, B. Moritz, A. M. Shvaika, J. K. Freericks, and T. P. Devereaux, Phys. Rev. B 99, 104306 (2019). * Buzzi _et al._ (2018) M. Buzzi, M. Först, R. Mankowsky, and A. Cavalleri, Nat. Rev. Mat. 3, 299 (2018). * Richard (1985) F. Richard, _QED: The strange theory of light and matter_ (Princeton University Press, USA, 1985). * Tomaš (1995) M. S. Tomaš, Phys. Rev. A 51, 2545 (1995). * Raimond _et al._ (2001) J. M. Raimond, M. Brune, and S. Haroche, Rev. Mod. Phys. 73, 565 (2001). * Röhlsberger _et al._ (2010) R. Röhlsberger, K. Schlage, B. Sahoo, S. Couet, and R. Rüffer, Science 328, 1248 (2010). * Röhlsberger _et al._ (2012) R. Röhlsberger, H.-C. Wille, K. Schlage, and B. Sahoo, Nature 482, 199 (2012). * Heeg _et al._ (2013) K. P. Heeg, H.-C. Wille, K. Schlage, _et al._ , Phys. Rev. Lett. 111, 073601 (2013). * Heeg _et al._ (2015a) K. P. Heeg, C. Ott, D. Schumacher, H.-C. Wille, R. Röhlsberger, T. Pfeifer, and J. Evers, Phys. Rev. Lett. 114, 207401 (2015a). * Heeg _et al._ (2015b) K. P. Heeg, J. Haber, D. Schumacher, _et al._ , Phys. Rev. Lett. 114, 203601 (2015b). * Haber _et al._ (2017) J. Haber, X. Kong, C. Strohm, _et al._ , Nat. Photon. 11, 720 (2017). * Haber _et al._ (2019) J. Haber, J. Gollwitzer, S. Francoual, M. Tolkiehn, J. Strempfer, and R. Röhlsberger, Phys. Rev. Lett. 122, 123608 (2019). * Adams _et al._ (2013) B. W. Adams, C. Buth, S. M. Cavaletto, _et al._ , J. Mod. Optic. 60, 2 (2013). * Brown _et al._ (1977) M. Brown, R. E. Peierls, and E. A. Stern, Phys. Rev. B 15, 738 (1977). * Wei and Lytle (1979) P. S. P. Wei and F. W. Lytle, Phys. Rev. B 19, 679 (1979). * Röhlsberger _et al._ (2005) R. Röhlsberger, K. Schlage, T. Klein, and O. Leupold, Phys. Rev. Lett. 95, 097601 (2005). * Huang _et al._ (2020) X.-C. Huang, Z.-R. Ma, X.-J. Kong, W.-B. Li, and L.-F. Zhu, J. Opt. Soc. Am. B 37, 745 (2020). * Li _et al._ (2012) W. Li, J. Zhu, X. Ma, H. Li, H. Wang, K. J. Sawhney, and Z. Wang, Rev. Sci. Instrum. 83, 053114 (2012). * Liu _et al._ (2003) X.-J. Liu, L.-F. Zhu, Z.-S. Yuan, _et al._ , Phys. Rev. Lett. 91, 193203 (2003). * Van Loo _et al._ (2013) A. F. Van Loo, A. Fedorov, K. Lalumière, B. C. Sanders, A. Blais, and A. Wallraff, Science 342, 1494 (2013). * Hill _et al._ (2007) J. Hill, D. Coburn, Y.-J. Kim, T. Gog, D. Casa, C. Kodituwakku, and H. Sinn, J. Synchrotron Radiat. 14, 361 (2007).
††thanks: They contribute equally to this work. ††thanks: They contribute equally to this work. # Rényi Entropy Dynamics and Lindblad Spectrum for Open Quantum System Yi-Neng Zhou Institute for Advanced Study, Tsinghua University, Beijing 100084, China Liang Mao Institute for Advanced Study, Tsinghua University, Beijing 100084, China Department of Physics, Tsinghua University, Beijing 100084, China Hui Zhai<EMAIL_ADDRESS>Institute for Advanced Study, Tsinghua University, Beijing 100084, China ###### Abstract In this letter we point out that the Lindblad spectrum of a quantum many-body system displays a segment structure and exhibits two different energy scales in the strong dissipation regime. One energy scale determines the separation between different segments, being proportional to the dissipation strength, and the other energy scale determines the broadening of each segment, being inversely proportional to the dissipation strength. Ultilizing a relation between the dynamics of the second Rényi entropy and the Lindblad spectrum, we show that these two energy scales respectively determine the short- and the long-time dynamics of the second Rényi entropy starting from a generic initial state. This gives rise to opposite behaviors, that is, as the dissipation strength increases, the short-time dynamics becomes faster and the long-time dynamics becomes slower. We also interpret the quantum Zeno effect as specific initial states that only occupy the Lindblad spectrum around zero, for which only the broadening energy scale of the Lindblad spectrum matters and gives rise to suppressed dynamics with stronger dissipation. We illustrate our theory with two concrete models that can be experimentally verified. For a closed quantum system, the energy spectrums of Hamiltonian fully determine the time scales of its dynamics. For an open quantum system, when the environment is treated by the Markovian approximation, the couplings between system and environment are controlled by a set of dissipation operators. In this case, the dynamics of the system is governed by the Lindblad equation which contains the contributions from both the Hamiltonian and the dissipation operators open . Obviously, the spectrum of the Hamiltonian alone can no longer determine the time scales of the entire dynamics, and a natural question is then what energy scales set the time scales of dynamics of an open quantum system. There are various directions to approach this issue, and the answer also relies on what type of dynamics that we are concerned with. Here let us focus on the dissipation driven dynamics. There are still different physical intuitions from different perspectives. One intuition is from the perturbation theory when the dissipation strength is weaker compared with the typical energy scales of the Hamiltonian Pan . In this regime, by treating the dissipation perturbatively, it leads to a scenario that the dissipation dynamics becomes faster when the dissipation strength is stronger. Another intuition is from the studies of the quantum Zeno effect Zeno_paradox ; Zeno1990 ; quantum_Zeno ; Zeno_review , which states that frequent measurements can slow down the dynamics, provided that the typical time interval between two successive measurements are shorter than the intrinsic time scale of the system. Since the measurement can also be understood in term of dissipations in the Lindblad master equation, it provides another scenario that the dissipation dynamics is suppressed when the dissipation becomes stronger, in the regime that the dissipation strength is stronger compared with the typical energy scales of the Hamiltonian. It seems that these two scenarios respectively work on different parameter regimes and the results are also opposite to each other. It will be interesting to see that there actually exists a framework that can unify these two scenarios. When a system is coupled to a Markovian environment, the entropy of the system will increase in time. The entropy dynamics of an open quantum many-body system is a subject that attracts lots of interests recently entang_review ; Qi_Cricuit ; Chen ; Kitaev ; YuChen ; Zhai . In this letter, we address the issue of typical time scales of the entropy increasing dynamics of a quantum many-body system coupled to a Markovian invironment, and especially, we should focus on the second Rényi entropy, for the reason that will be clear below, and answer the question whether the entropy dynamics is faster or slower when the dissipation strength increases. Figure 1: Schematic of the mapping between the Lindblad equation (left) and the Schödinger like equation in a doubled system (right). Here $\hat{L}\hat{\rho}$ denotes the r.h.s. of Eq. 1. Our studies are based on a mapping between the Lindblad master equation and a non-unitary evolution of wave function in a doubled space, as shown in Fig. 1. Let us first review this mapping Operator_Schmidt ; Mixed_state . Considering a density matrix $\hat{\rho}$, and given a set of complete bases $\\{|n\rangle\\},(n=1,\dots,\mathcal{D}_{\text{H}})$ of the Hilbert space with dimension $\mathcal{D}_{\text{H}}$ (say, the eigenstates of the Hamiltonian $\hat{H}$ with eigenenergies $E_{n}$), the density matrix $\hat{\rho}$ can be expressed as $\hat{\rho}=\sum_{mn}\rho_{mn}|m\rangle\langle n|$. By the operator-to-state mapping, we can construct a wave function $\Psi_{\rho}=\sum_{mn}\rho_{mn}|m\rangle\otimes|n\rangle$, which contains exact the same amount information as $\hat{\rho}$. Here $\Psi_{\rho}$ is a wave function on a system whose size is doubled compared to the original system, and we will refer these two copies of original system as the “left” (L) and the “right” (R) systems. Under this mapping, for instance, a density matrix of a pure state $\hat{\rho}=|\psi\rangle\langle\psi|$ is mapped to a product state $\Psi_{\rho}=|\psi\rangle\otimes|\psi\rangle$ in the double system, and a thermal density matrix at temperature $T$ as $\hat{\rho}=\sum_{n}e^{-E_{n}/(k_{\text{b}}T)}|n\rangle\langle n|$ is mapped to a thermofield double state at temperature $T/2$ as $\Psi_{\rho}=\sum e^{-E_{n}/(k_{\text{b}}T)}|n\rangle\otimes|n\rangle$ in the double system. For an open system coupled to a Markovian environment, the density matrix obeys the Lindblad master equation given by $\hbar\frac{d\hat{\rho}}{dt}=-i[\hat{H},\hat{\rho}]+\sum\limits_{\mu}\gamma_{\mu}\left(2\hat{L}_{\mu}\hat{\rho}\hat{L}_{\mu}^{\dagger}-\\{\hat{L}^{\dagger}_{\mu}\hat{L}_{\mu},\hat{\rho}\\}\right),$ (1) where $\hat{L}_{\mu}$ stand for a set of dissipation operators, and $\gamma_{\mu}$ are their corresponding dissipation strengths. After the mapping, the wave function $\Psi_{\rho}$ in the double system satisfies a Schrödinger-like equation $i\hbar\frac{d\Psi_{\rho}}{dt}=\left(\hat{H}_{\text{s}}-i\hat{H}_{\text{d}}\right)\Psi_{\rho}.$ (2) Here $\hat{H}_{\text{s}}$ is the Hermitian part of the Hamiltonian determined by system itself, and it is given by $\hat{H}_{\text{s}}=\hat{H}_{\text{L}}\otimes\hat{I}_{\text{R}}-\hat{I}_{\text{L}}\otimes\hat{H}^{\text{T}}_{\text{R}},$ (3) where operators with subscript “L” and “R” respectively stand for operators acting on the left and the right systems, and “T” stands for the transpose, and $\hat{I}$ represents the identity operator. $-i\hat{H}_{\text{d}}$ is the non-Hermitian part of the Hamiltonian determined by the dissipation operators, which is given by $\displaystyle\hat{H}_{\text{d}}=\sum_{\mu}\gamma_{\mu}$ $\displaystyle\left[-2\hat{L}_{\mu,\text{L}}\otimes\hat{L}^{\text{*}}_{\mu,\text{R}}\right.$ $\displaystyle\left.+(\hat{L}^{\dagger}_{\mu}\hat{L}_{\mu})_{\text{L}}\otimes\hat{I}_{\text{R}}+\hat{I}_{\text{L}}\otimes(\hat{L}^{\dagger}_{\mu}\hat{L}_{\mu})^{\text{*}}_{\text{R}}\right],$ (4) where the superscript * stands for taking complex conjugation. We can diagnolize this non-Hermitian Hamiltonian $\hat{H}_{\text{s}}-i\hat{H}_{\text{d}}$, which leads to a set of eigenstates as $(\hat{H}_{\text{s}}-i\hat{H}_{\text{d}})|\Psi^{l}_{\rho}\rangle=\epsilon_{l}|\Psi^{l}_{\rho}\rangle,$ (5) where $\epsilon_{l}$ is in general a complex number, and we denote them as $\epsilon_{l}=\alpha_{l}-i\beta_{l}$. This spectrum, originated from the Lindblad equation, is referred to as the Lindblad spectrum. The full Lindblad spectrum has been studied for a number of models before Prosen1 ; Prosen2 ; Universal_spectra ; tenfold ; local_random_Liouvillians ; Wang . Here we would like to make several useful comments on the Lindblad spectrum. i) $\alpha_{l}$ and $-\alpha_{l}$ always appear in pairs in the spectrum; ii) $\beta_{l}$ is always non-negative; iii) If $\hat{L}_{\mu}$ are all hermitian, there always exists a zero-energy eigenstate with $\epsilon_{l}=0$, and this eigenstate is labelled as $l=0$ and is given by $|\Psi^{l=0}_{\rho}\rangle=\frac{1}{\sqrt{\mathcal{D}_{\text{H}}}}\sum_{n}|n\rangle\otimes|n\rangle$. Figure 2: The dynamics of the second Rényi entropy $S^{(2)}$ as a function of $t\gamma$. $\gamma$ is the dissipation strength. Different curves have different $\gamma$ in unit of $J$. The inset show the long-time behavior of $S^{(2)}$ as functions of $tJ$ and $tJ^{2}/\gamma$. The dashed line is a fitting of initial slop based on Eq. 13. (a) is for the Bose-Hubbard model with $U=J$ and the number of sites $L=6$, and the number of bosons $N=3$. (b) is for hard core bosons model with $V=J$, $L=8$ and $N=4$. The initial state is taken as the ground state of $\hat{H}$. Rényi Entropy and Lindblad Spectrum. Here we bring out a close relation between the dynamics of the second Rényi entropy and the Lindblad spectrum. For any density matrix $\hat{\rho}(t)$, the second Rényi entropy $S^{(2)}(t)$ is given by $e^{-S^{(2)}}=\text{Tr}(\hat{\rho}^{2})=\sum\limits_{mn}\rho_{mn}(t)\rho_{nm}(t).$ (6) On the other hand, in the double system, the total amplitude of the wave function is given by $|\Psi_{\rho}|^{2}=\sum\limits_{mn}\rho_{mn}(t)\rho^{*}_{mn}(t).$ (7) Since the density matrix is always Hermitian, it gives $\rho_{nm}(t)=\rho^{*}_{mn}(t)$, and therefore, we have $e^{-S^{(2)}}=|\Psi_{\rho}|^{2}.$ (8) An initial state $\Psi_{\rho}(0)$ in the double space can be expanded as $\Psi_{\rho}(0)=\sum_{l}c_{l}|\Psi^{l}_{\rho}\rangle$, the subsequent evolution is given by $\Psi(t)=e^{-i\hat{H}_{\text{s}}t-\hat{H}_{\text{d}}t}|\Psi_{\rho}(0)\rangle=\sum_{l}c_{l}e^{-i\alpha_{l}t-\beta_{l}t}|\Psi^{n}_{\rho}\rangle$ (9) and therefore $e^{-S^{(2)}}=|\Psi_{\rho}|^{2}=\sum\limits_{n}|c_{l}|^{2}e^{-2\beta_{l}t}.$ (10) Since the evolution in double system is non-unitary and all $\beta_{l}$ are non-negative, the total amplitude of the wave function always decays in time. Hence, by this entropy-amplitude relation Eq. 8, the decaying of $|\Psi_{\rho}|^{2}$ gives rise to the increasing of $S^{(2)}$. Note that for any initial density matrix with trace unity and for hermitian $\hat{L}_{\mu}$, $c_{l=0}$ always equals $1/\sqrt{\mathcal{D}_{\text{H}}}$. This mode always does not decay in time because $\beta_{l=0}=0$. If there is no other eigenmodes with $\beta_{l}=0$, $l=0$ mode is the only remaining mode at infinite long time, which gives a maximum second Rényi entropy $\log\mathcal{D}_{\text{H}}$. Before reaching that limit, the imaginary parts of the Lindblad spectrum of occupied states determine the time scales of the Rényi entropy dynamics. Our discussion below will be based on this connection. Models. Although our discussion below is quite general for quantum many-body systems, we illustrate the results with two concrete models. The first model is the Bose-Hubbard model, which reads $\hat{H}=-J\sum\limits_{\langle ij\rangle}(\hat{b}^{\dagger}_{i}\hat{b}_{j}+\text{h.c.})+\frac{U}{2}\sum\limits_{i}\hat{n}_{i}(\hat{n}_{i}-1),$ (11) where $\hat{b}_{i}$ is the boson annihilation operator at site-$i$, and $\hat{n}_{i}=\hat{b}^{\dagger}_{i}\hat{b}_{i}$ is the boson number operator at site-$i$. $\langle ij\rangle$ denotes nearest neighbor sites. $J$ and $U$ are respectively the hopping and the on-site interaction strengths. For the second model, we consider hard-core bosons, which prevent two bosons to occupy the same site. In addition, we introduce the nearest-neighbor repulsion, and the model reads $\hat{H}=-J\sum\limits_{\langle ij\rangle}(\hat{b}^{\dagger}_{i}\hat{b}_{j}+\text{h.c.})+V\sum\limits_{\langle ij\rangle}\hat{n}_{i}\hat{n}_{j}.$ (12) In one-dimension, these two models are quite different, because the second model can be mapped to a spinless fermion model with nearest neighbor repulsion, and can also be mapped to a spin model with nearest neighbor couplings, but the first model cannot. In both cases, we take all $\hat{n}_{i}$ as the dissipation operators and we set the dissipation strengthes uniformly as $\gamma$. In the numerical results shown below, we have choose $J\sim U$ or $J\sim V$ such that $J$ sets the typical energy scale of the Hamiltonian part, and therefore, strong and weak dissipations respectively mean $\gamma/J>1$ or $\gamma/J<1$. Below we will show that both models exhibit similar features, which supports that our results are quite universal. Figure 3: The Lindblad spectrum for strong dissipation case (a1,a2,b1,b2) with $\gamma=5J$ and for weak dissipation case (c1,c2) with $\gamma=0.2J$. The red points mark the eigenstates with significant occupation ($|c_{l}|^{2}\geqslant 1/\mathcal{D}_{\text{H}}$) by the initial state. For (a1) and (a2) in the first raw, the initial state is taken as $\Psi_{\rho}=|\psi_{\text{g}}\rangle\otimes|\psi_{\text{g}}\rangle$, where $|\psi_{\text{g}}\rangle$ is the ground state of $\hat{H}$. For (b1) and (b2) in the second raw, the initial states are taken as the zero-energy eigenstate of $\hat{H}_{\text{d}}$, that are $|111000\rangle$ for (b1) and $|11110000\rangle$ for (b2) in Fock bases. The left column (a1,b1,c1) are for the Bose-Hubbard model with $U=J$ and the number of sites $L=6$, and the number of bosons $N=3$. The right column (a2,b2,c2) are for hard core bosons model with $V=J$, $L=8$ and $N=4$. Dynamics of the Rényi Entropy. We first consider the short-time behavior of the Rényi entropy dynamics. We apply the short-time expansion to Eq. 9 and ultilize the relation Eq. 8, and to the leading order of entropy change, we obtain $\lim\limits_{t\rightarrow 0}\frac{dS^{(2)}}{dt}=2\frac{\langle\Psi_{\rho}(0)|\hat{H}_{\text{d}}|\Psi_{\rho}(0)\rangle}{\langle\Psi_{\rho}(0)|\Psi_{\rho}(0)\rangle}.$ (13) The physical meaning of the r.h.s. of Eq. 13 in original system is the fluctuation of the dissipation operators. For instance, if the initial state is a pure state and $\hat{\rho}(0)=|\psi(0)\rangle\langle\psi(0)|$, then $|\Psi_{\rho}(0)\rangle=|\psi(0)\rangle\otimes|\psi(0)\rangle$, and Eq. 13 can be rewritten as $\displaystyle\lim\limits_{t\rightarrow 0}\frac{dS^{(2)}}{dt}=$ $\displaystyle 4\sum\limits_{\mu}\gamma_{\mu}\left(\langle\psi(0)|\hat{L}^{\dagger}_{\mu}\hat{L}_{\mu}|\psi(0)\rangle-|\langle\psi(0)|\hat{L}_{\mu}|\psi(0)\rangle|^{2}\right).$ (14) Suppose all $\gamma_{\mu}$ are taken as the same $\gamma$, this result shows that the time-dependence of $S^{(2)}$ is governed by a dimensionless time $\gamma t$. In other word, the larger $\gamma$ is, the faster the Rényi entropy dynamics increases. This $\gamma t$ scaling is shown in Fig. 2 for two different models, where one can see that the short-time parts of $S^{(2)}$ curves with different $\gamma$ collapse into a single line when plotted in term of $\gamma t$. The dashed lines compare the short-time behavior with the slope given by Eq. 13 and Eq. 14. In Fig. 2, one also finds that $S^{(2)}$ no longer obeys the $\gamma t$ scaling when $\gamma t>1$. Moreover, in the strong dissipation regime, the insets plotted in term of $tJ$ show an opposite trend at long-time, that is, the larger $\gamma$ is, the slower the Rényi entropy increases. In fact, the long-time behavior of $S^{(2)}$ exhibits a $t/\gamma$ scaling. As shown in the insets of Fig. 2, when the long-time part of $S^{(2)}$ curves with different $\gamma$ are ploted in term of $tJ^{2}/\gamma$, they all collapse into a single curve. Lindblad Spectrum with Strong Dissipation. This opposite behavior between short- and long-time can be understood very well in term of the Lindblad spectrum. As one can see from Fig. 3(a,b), for strong dissipation, the main feature of the Lindblad spectrum is that it separates into segments along the imaginary axes of the spectrum, and the separation between segments are approximately $2\gamma$. For each segment, the width along the imaginary axes is approximately given by $J^{2}/\gamma$. This feature can be understand by perturbation treatment of $\hat{H}_{\text{s}}-i\hat{H}_{\text{d}}$. Since the dissipation strength is stronger than the typical energy scales of the Hamiltonian, we can treat $\hat{H}_{\text{s}}$ as a perturbation to $\hat{H}_{\text{d}}$. To the zeroth order of $\hat{H}_{\text{d}}$, the spectrum is purely imaginary and different segments are separated by $2\gamma$. More importantly, it worth emphasizing that the eigenstates of $\hat{H}_{\text{d}}$ are usually highly degenerate, for instance, when different $\hat{L}_{\mu}$ commute with each other and are related by a symmetry, such as $\hat{L}_{\mu}$ being $\hat{n}_{i}$ in our examples. Usually, $\hat{H}_{\text{s}}$ and $\hat{H}_{\text{d}}$ do not commute with each other, and the perturbation in $\hat{H}_{\text{s}}$ lifts the degeneracy of the imaginary parts and gives rise to a broadening of the order of $J^{2}/\gamma$, due to the nature of the second order perturbation. We call these eigenstates with imaginary energies of the order of a few times of $\gamma$ as “high imaginary energy states”, and these eigenstates with imaginary energies of the order of a few times of $J^{2}/\gamma$ as “low-lying imaginary energy states”. For a generic initial state, both two types of eigenstates are occupied. Quite generally, the occupations of the “high imaginary energy states” are significant, for instance, when the initial state is taken as the eigenstates of $\hat{H}_{\text{s}}$. With the relation between the Rényi entropy dynamics and the Lindblad spectrum discussed above, it is clear that the short-time dynamics is dominated by these “high imaginary energy states” that gives a dynamics scaled by $t\gamma$. Nevertheless, when $\gamma t>1$, the weights on these “high imaginary energy states” mostly decay out and the long-time dynamics is therefore dominated by the “low-lying imaginary energy states” that gives a dynamics scaled by $tJ^{2}/\gamma$. Figure 4: The dynamics of the second Rényi entropy $S^{(2)}$ as a function of $tJ^{2}/\gamma$ for specific initial state. $\gamma$ is the dissipation strength. Different curves have different $\gamma$ in unit of $J$. The inset show the short-time behavior of $S^{(2)}$ as functions of $tJ$ and $t\gamma$. (a) is for the Bose-Hubbard model with $U=J$ and the number of sites $L=6$, and the number of bosons $N=3$. (b) is for hard core bosons model with $V=J$, $L=8$ and $N=4$. The initial states are taken as the zero-energy eigenstate of $\hat{H}_{\text{d}}$, that are $|111000\rangle$ for (a) and $|11110000\rangle$ for (b) in Fock bases. Quantum Zeno Effect Revisited. Here we consider a specific initial state that satisfies $\hat{H}_{\text{d}}|\Psi(0)\rangle=0$. In other word, such initial states do not exhibit fluctuation of dissipation operators. Thus, according to Eq. 13 and Eq. 14, the initial slop of $S^{(2)}$ is zero. Moreover, in the strong dissipation regime, the populations of the “high imaginary energy states” are strongly suppressed by the “gap” between different segments and their contribution becomes negligible, and such initial states mainly populate the “low-lying imaginary energy states”, as we shown in Fig. 3(b). Therefore, the entire dynamics of the second Rényi entropy is set by the energy scale $J^{2}/\gamma$ and it obeys the $t/\gamma$ scaling. This is shown in Fig. 4 for two models. To contrast such specific initial states with generic states discussed above, we plot in the inset of Fig. 4 the short-time behavior of $S^{(2)}$ as a function of $t\gamma$ and $tJ$. Unlike the results shown in Fig. 2, the short-time dynamics with $t\gamma<1$ are quite different, because it does not exhibit linear behavior and different curves do not collapse into a single line in term of $t\gamma$. For these initial states, that the dynamics is slower with stronger dissipation is reminiscent of the quantum Zeno effect. In fact, the quantum Zeno effect can indeed be understood in this way. Introducing $\\{|M\rangle\\},(M=1,\dots,\mathcal{D}_{\text{H}})$ as a set of complete and orthogonal measurement bases, we define the projection operators as $\hat{P}_{M}=|M\rangle\langle M|$, and the frequent measurement process can also be described by the Lindblad equation Eq. 1 with dissipation operator $\hat{L}_{\mu}$ given by all $\hat{P}_{M}$. With such dissipation operators, the Lindblad spectrum exhibits a set of “low-lying imaginary energy states” with energy scale given by $J^{2}/\gamma$. It can be shown that, as long as the initial state density matrix is diagonal in the measurement bases, the initial states satisfy $\hat{H}_{\text{d}}|\Psi(0)\rangle=0$. From Strong to Weak Dissipation. Finally we show that when $\gamma$ decreases and eventually becomes weaker compared with the typical energy scales in the Hamiltonian, the segments structure in the Lindblad spectrum disappears, as we shown in Fig. 3(c). Thus, the entropy dynamics for generic states no longer display the feature of two time scales. The quantum Zeno effect also disappears even for the specific initial states, and this is understandable because in this regime, the typical time interval between two measurements is already longer than the intrinsic evolution time of the system. Summary. In this work, we establish a relation between the Rényi entropy dynamics and the Lindblad spectrum in double space. At the strong dissipation regime, the Lindblad spectrum exhibits a segment structure, in which we can introduce the “high imaginary energy eigenstates” and the “low-lying imaginary energy eigenstates”. For a generic initial state with significantly occupied “high imaginary energy eigenstates”, the former dominates the short-time dynamics and the latter dominates the long-time dynamics, which respectively give rise to $t\gamma$ scaling and $t/\gamma$ scaling. For a specific initial state with only “low-lying imaginary energy eigenstates” significantly occupied, the dynamics is dominated by $t/\gamma$ scaling, and we show the quantum Zeno effect belongs to this class. We illustrate our results with two concrete models. The second Rényi entropy can now been measured in ultracold atomic gases in optical lattices, and in fact, it has been measured in the Bose-Hubbard model with or without disorder Greiner1 ; Greiner2 ; Greiner3 . The dissipation operators and their strenghes can also now be controlled in ultracold atomic gases BHMExp , our predictions can therefore be verified directly in the experimental setup. Acknowledgment. We thank Lei Pan, Tian-Shu Deng, Tian-Gang Zhou and Pengfei Zhang for helpful discussions. This work is supported by Beijing Outstanding Young Scientist Program, NSFC Grant No. 11734010, MOST under Grant No. 2016YFA0301600. Note Added. When finishing this work, we become aware of a work in which similar behaviors of the Lindblad spectrum in strong dissipation regime are also discussed full_spectrum . ## References * (1) H. P. Breuer and F. Petruccione, The Theory of Open Quantum Systems (Oxford University Press, Oxford, 2007). * (2) L. Pan, X. Chen, Y. Chen, and H. Zhai, Nat. Phys. 16, 767(2020). * (3) B. Misra and E. C. G. Sudarshan, Journal of Mathematical Physics 18, 756 (1977). * (4) W. M. Itano, D. J. Heinzen, J. J. Bollinger, and D. J. Wineland, Phys. Rev. A 41, 2295 (1990). * (5) A. G. Kofman and G. Kurizki, Nature 405, 546 (2000). * (6) K. Koshino and A. Shimizu, Phys. Rep. 412, 191 (2005). * (7) L. Aolita, F. de Melo and L. Davidovich, Rep. Prog. Phys. 78, 042001 (2015) * (8) P. Lorenzo, S. Christoph, and X.-L., Qi, J. High Energ. Phys. 2020, 63 (2020). * (9) Y. Chen, X.-L. Qi and P. Zhang, J. High Energ. Phys. 2020, 121 (2020). * (10) P. Dadras, A. Kitaev, arXiv: 2011.09622 * (11) Y. Chen, arXiv: 2012.00223 * (12) K. Su, P. Zhang and H. Zhai, arXiv: 2101.***** * (13) J. E. Tyson, J. Phys. A: Math. Gen. 36, 10101 (2003). * (14) M. Zwolak and G. Vidal, Phys. Rev. Lett. 93, 207205 (2004). * (15) T. Prosen, Phys. Rev. Lett. 109, 090404 (2012) * (16) M. V. Medvedyeva, F. H. L. Essler, T. Prosen, Phys. Rev. Lett. 117, 137202 (2016) * (17) S. Denisov, T. Laptyeva, W. Tarnowski, D. Chru?ci?ski, and K. Zyczkowski, Phys. Rev. Lett. 123, 140403 (2019) * (18) S. Lieu, M. McGinley, and N. R. Cooper, Phys. Rev. Lett. 124, 040401 (2020). * (19) K. Wang, F. Piazza, and D. J. Luitz, Phys. Rev. Lett. 124, 100604 (2020). * (20) D. Yuan, H. Wang, Z. Wang, D. L. Deng, arXiv: 2009.00019 * (21) R. Islam, R. Ma, P. M. Preiss, M. Eric Tai, A. Lukin, M. Rispoli, M. Greiner, Nature 528, 77 (2015). * (22) A. M. Kaufman, M. Eric Tai, A. Lukin, M. Rispoli, R. Schittko, P. M. Preiss, M. Greiner, Science 353, 794 (2016) * (23) A. Lukin, M. Rispoli, R. Schittko, M. Eric Tai, A. M. Kaufman, S. Choi, V. Khemani, J. Léonard, M. Greiner, Science 364, 256 (2019) * (24) R. Bouganne, M. B. Aguilera, A. Ghermaoui, J. Beugnon, F. Gerbier, Nat. Phys. 16, 2125 (2020). * (25) V. Popkov and C. Presilla, arXiv:2101.05708
# Page Curve from Non-Markovianity Kaixiang Su Institute for Advanced Study, Tsinghua University, Beijing 100084, China Department of Physics, University of California Santa Barbara, Santa Barbara, California, 93106, USA Pengfei Zhang <EMAIL_ADDRESS>Institute for Quantum Information and Matter, California Institute of Technology, Pasadena, California 91125, USA Walter Burke Institute for Theoretical Physics, California Institute of Technology, Pasadena, California 91125, USA Hui Zhai<EMAIL_ADDRESS>Institute for Advanced Study, Tsinghua University, Beijing 100084, China ###### Abstract In this letter, we use the exactly solvable Sachdev-Ye-Kitaev model to address the issue of entropy dynamics when an interacting quantum system is coupled to a non-Markovian environment. We find that at the initial stage, the entropy always increases linearly matching the Markovian result. When the system thermalizes with the environment at a sufficiently long time, if the environment temperature is low and the coupling between system and environment is weak, then the total thermal entropy is low and the entanglement between system and environment is also weak, which yields a small system entropy in the long-time steady state. This manifestation of non-Markovian effects of the environment forces the entropy to decrease in the later stage, which yields the Page curve for the entropy dynamics. We argue that this physical scenario revealed by the exact solution of the Sachdev-Ye-Kitaev model is universally applicable for general chaotic quantum many-body systems and can be verified experimentally in near future. Studying open quantum many-body systems is of fundamental importance for understanding quantum matters and for future applications of quantum technology because all systems are inevitably in contact with environments, and decoherence due to coupling with environments is a major obstacle for future applications of quantum devices Preskill2018 . So far, most studies of open quantum systems are limited to either situation in which the systems are small or weakly correlated quantum many-body systems, or situations that the environment is treated by the Born-Markovian approximation scully1999quantum ; breuer2002theory . Little effort has been made on strongly correlated quantum many-body systems coupled to a non-Markovian environment. This is simply because both strong correlation and non-Markovianity are difficult to handle theoretically. Open systems are also of interest to gravity studies, and the best-known problem is the black hole information paradox hawking1975 . The central issue of the black hole information paradox is whether the black hole evaporation can be considered as undergoing unitary dynamics. If so, the entropy should first increase and then decrease as the black hole evaporates. Such an entropy curve as shown in Fig. 1(a) is known as the Page curve page1993 . Here the entanglement entropy is the entropy of the reduced density matrix of the radiation part $\mathcal{A}$ after tracing out the remaining black hole part $\mathcal{B}$ . As shown in Fig. 1(a), this entanglement entropy reaches the maximum when half of the black hole is evaporated, giving rise to the Page curve. Reproducing the Page curve from gravity theory is a challenging part of the black hole information problem, and progresses have been made recently using the semi-classical gravity calculations penington2019entanglement ; almheiri2019entropy ; almheiri2020page ; almheiri2019replica ; penington2019replica . Figure 1: (a): Illustration of the Page curve: the entanglement entropy between two sub-systems with length $xL$ and $(1-x)L$. The total length is $L$ and $0\leqslant x\leqslant 1$. (b): Illustration of the setup: an SYK4 system with $N$ Majorana fermions serves as the system and an SYK2 system with $M$ Majorana fermions serves as the environment. Here $M\gg N$. In this letter, we explore the Sachdev-Ye-Kitaev model maldacena2016remarks ; Kitaev2018soft with random four-Majorana fermions interactions (SYK4), and this SYK4 model is coupled to a system with random quadratic Majorana fermions couplings (SYK2). The SYK2 part contains a lot more degrees-of-freedom compared with the SYK4 part such that the SYK2 part can be viewed as the environment. The motivations for considering such a model are two folds. First, the SYK4 model is exactly solvable in the large-N limit and its solution gives rise to a strongly correlated non-Fermi liquid state maldacena2016remarks . Recently, techniques related to SYK4 model has been widely used to construct exact solutions to address open issues of strongly correlated quantum matters gu2017local ; davison2017thermoelectric ; chen2017competition ; song2017strongly ; zhang2017dispersive ; jian2017model ; chen2017tunable ; eberlein2017quantum ; zhang2019evaporation ; almheiri2019universal ; gu2017spread ; chen2020Replica ; zhang2020 ; sk ; Altman ; sk jian ; zhang2020entanglement ; Liu2020non ; Kitaev . Here the situation we explored is also exactly solvable and we can use the solution to understand how a non-Markovian environment affects strongly correlated phases zhang2019evaporation ; almheiri2019universal ; chen2017tunable . Secondly, the SYK4 model is holographically dual to the Jackiw-Teitelboim gravity theory in the AdS2 geometry with a black hole bulk spectrum Polchinski ; bulk Yang ; bulk2 ; bulk3 ; bulk4 ; bulk5 . Thus, the entropy dynamics of the SYK4 system coupled to an environment gu2017spread ; chen2020Replica ; Kitaev resembles the black hole evaporation process penington2019entanglement ; almheiri2019entropy ; almheiri2020page ; almheiri2019replica ; penington2019replica and it will be interesting to study when a Page-like curve can emerge after turning on the coupling between the system and the environment. Remarkably, the main findings of this work bring these two aspects together, that is, we show that the Page curve emerges because of the non-Markovian effect of the environment. Model. The system under consideration is illustrated in Fig. 1(b) and the total Hamiltonian is given by $\displaystyle\hat{H}=\hat{H}_{\text{S}}+\hat{H}_{\text{E}}+\hat{H}_{\text{SE}}$ (1) $\displaystyle=\sum_{i<j<k<l}J^{S}_{ijkl}\psi_{i}\psi_{j}\psi_{k}\psi_{l}+i\sum_{a<b}J^{\text{E}}_{ab}\chi_{a}\chi_{b}+i\sum_{i,a}V_{ia}\psi_{i}\chi_{a}.$ Here $\psi_{i}$ ($i=1,\dots,N$) denotes $N$ Majorana fermions in the system and $\chi_{a}$ ($a=1,\dots,M$) denotes $M$ Majorana fermions in the environment, with $\\{\psi_{i},\psi_{j}\\}=\delta_{ij}$ and $\\{\chi_{a},\chi_{b}\\}=\delta_{ab}$. Throughout the letter we will use the subscript “S” and “E” to denote the system part and the environment part respectively. $\hat{H}_{\text{S}}$ and $\hat{H}_{\text{E}}$ are then SYK4 and SYK2 Hamiltonians. $J^{S}_{ijkl}$, $J^{\text{E}}_{ab}$ and $V_{ia}$ are independent random Gaussian variables with variances given by $\overline{(J^{\text{S}}_{ijkl})^{2}}=\frac{3!J_{S}^{2}}{N^{3}},\qquad\overline{(J^{\text{E}}_{ab})^{2}}=\frac{2!J_{\text{E}}^{2}}{M},\qquad\overline{(V_{ia})^{2}}=\frac{V^{2}}{M}.$ (2) Throughout the paper, we take $J_{S}=1$ as the energy unit. We focus on the limit $M\gg N\gg 1$ in which the Schwinger-Dyson equation for $\chi$ contains no contribution from $\psi$, justifying that the $\chi$ part can be viewed as the environment. Therefore, the Green’s function $G_{\chi}(\tau)=\left<T_{\tau}\chi_{i}(\tau)\chi_{i}(0)\right>$ of the environment takes the standard form of the SYK2 model as maldacena2016remarks $G_{\chi}(\omega)=-\frac{2}{i\omega+i\text{sgn}(\omega)\sqrt{4J_{\text{E}}^{2}+\omega^{2}}}.$ (3) The fact that this Green’s function has frequency dependence means that the environment is treated beyond the Markovian approximation. We consider the situation in which the system and the environment is initially decoupled, and both are in thermal equilibrium with inverse temperatures $\beta_{\text{S}}$ and $\beta_{\text{E}}$ respectively. The initial density matrix is therefore given by $\rho(0)=\frac{1}{Z_{\text{S}}Z_{\text{E}}}e^{-\beta_{S}\hat{H}_{\text{S}}}\otimes e^{-\beta_{\text{E}}\hat{H}_{\text{E}}}$ with $Z_{\text{S}}$ and $Z_{\text{E}}$ being the corresponding partition functions. Evolving the system with the Hamiltonian Eq. (1) and tracing out the environment, one obtains the reduced density matrix of the system $\rho_{\text{S}}(t)$ at time $t$ as $\rho_{S}(t)=\text{tr}_{\text{E}}\left[e^{-i\hat{H}t}\rho(0)e^{i\hat{H}t}\right].$ (4) The corresponding second Rényi entropy $S^{(2)}$ of the system is then given by $e^{-S^{(2)}_{\text{S}}(t)}=\text{tr}_{\text{S}}\left[\rho_{S}(t)^{2}\right].$ (5) Under the disorder replica diagonal assumption, $S^{(2)}_{\text{S}}(t)$ can be expressed in terms of path-integral over bilocal fields. In the large-$N$ limit, the integral is dominated by the saddle point solution, and the entropy can be obtained by evaluating the on-shell action. Figure 2: (a): Entropy curves for different $\kappa$ and $\beta_{\text{E}}$. Here $\kappa=0.1$ for solid lines and $\kappa=0.2$ for dashed lines. Three different $\beta_{\text{E}}=(0,3,6)$ are plotted. The red dashed straight line represents the same initial slope for all curves. Two curves with $\beta_{\text{E}}=0$ in (a) coincide with the Markovian results. (b): Entropy curves with different $\kappa$ and a fixed $\beta_{\text{E}}=6$. The curve with large $\kappa$ in (b) coincides with $\beta_{\text{E}}=0$ curves in (a) and the Markovian results. Recovering the Markovian Results. Below we will first discuss situations where the entropy dynamics of our model can recover the Markovian result. Here, by the Markovian result we mean dynamics obtained by solving the following Lindblad master equation scully1999quantum ; breuer2002theory $\partial_{t}\hat{\rho}=-i[\hat{H}_{\text{S}},\hat{\rho}]+\sum_{i}\kappa_{i}\left(\hat{L}_{i}\rho\hat{L}_{i}^{\dagger}-\frac{1}{2}\\{\hat{L}_{i}^{\dagger}\hat{L}_{i},\hat{\rho}\\}\right).$ (6) By treating the environment with the Markovian approximation, we only need to consider the Hamiltonian of the system and the dissipation operators, without having to explicitly include the environment. To make comparison with our model, we take $\kappa_{i}={\kappa}$ and $\hat{L}_{i}=\psi_{i}$. Similar to previous procedures, we consider the initial thermal density matrix $\rho_{\text{S}}(0)=\frac{1}{Z_{\text{S}}}e^{-\beta_{\text{S}}\hat{H}_{\text{S}}}$, and we then evolve the system with Eq. (6). The second Rényi entropy can also be represented as a path-integral where the saddle points approximation is applicable. In the Markovian case, the Rényi entropy first grows linearly in time and then saturates to its maximum value $(N/2)\log 2$, forbidding the possibility of any page-like behaviors Zhai . For reasons that will become clear below, we consider the large $J_{\text{E}}$ limit and fix $V^{2}/J_{\text{E}}$ as $\kappa$ in our model. Our discussions below will then identify the following conditions as sufficient for our model to recover the Markovian results. i) Infinite Environment Temperature. The SYK2 environment becomes a Markovian one when $\beta_{\text{E}}=0$. This is because only the two-point function enters the effective action for entropy dynamics, and in the large $J_{\text{E}}$ limit, the Fourier transformation of the real-time Green’s function of the environment $G_{\text{E}}^{>}(t,\beta)=\left<\chi_{i}(t)\chi_{i}(0)\right>_{\beta}/Z_{B}$ gives $G^{>}_{\text{E}}(\omega,\beta_{\text{E}})=\frac{1}{J_{\text{E}}}\frac{1}{1+e^{-\beta_{E}\omega}}.$ (7) When $\beta_{\text{E}}=0$, the second term vanishes and the Green’s function of the environment becomes frequency independent, which is equivalent to the Markovian approximation. Under this situation, the standard derivation of the master equation Eq. 6 through a second-order perturbation theory yields the dissipation strength $\kappa=V^{2}/J_{\text{E}}$. As one can see in Fig. 2 (a), when $\beta_{\text{E}}$ is set to zero, the entropy curve becomes independent of $\kappa$. This universal curve also coincides with the results from the Markovian approximation. ii) Short Time. In the Markovian approach, it can be shown by perturbation that the entropy grows linearly at the initial stage, and when $\kappa t\ll 1$ npZhai ; Zhai , the growth rate can be derived analytically as $\frac{dS^{(2)}_{\text{S}}(t)}{dt}=\kappa N(1-2G_{\text{S}}^{W}(0,2\beta_{S})).$ (8) Here we have defined the Wightman Green’s function for the single SYK4 model as $G_{\text{S}}^{W}(t,\beta)=\frac{1}{Z_{\text{S}}}\left<\psi_{i}(t-i\frac{\beta}{2})\psi_{i}(0)\right>_{\beta}.$ (9) For our model, similar perturbative calculation Kitaev for short-time yields $\displaystyle\frac{dS_{\text{S}}^{(2)}}{dt}=2V^{2}N\int_{-t}^{t}dt^{\prime}$ $\displaystyle\left[\left(G^{>}_{\text{S}}(t^{\prime},2\beta_{S})-G^{W}_{\text{S}}(t^{\prime},2\beta_{S})\right)\right.$ $\displaystyle\left.\times G^{>}_{\text{E}}(t^{\prime},\beta_{B})\right].$ (10) By approximating $t^{\prime}=0$ in the integrand, Eq. 10 becomes $\displaystyle\frac{dS_{\text{S}}^{(2)}}{dt}=\frac{V^{2}}{J_{\text{E}}}N(1-2G_{\text{S}}^{W}(0,2\beta_{S})).$ (11) By equalling $V^{2}/J_{E}=\kappa$, Eq. 11 is the same as Eq. 8. This can also be seen in Fig. 2(a) that the initial growth is linearly in $\kappa t$ and the slop is a constant for varying $\beta_{\text{E}}$ with fixed $\beta_{\text{S}}$. In other words, although the Green’s function of the environment Eq. 7 contains the frequency dependent part, it is not important for initial time and the short-time behavior is always dominated by the frequency independent part. iii) Large System-Environment Coupling. The entropy curve of our model also matches the Markovian result when $\kappa$ is sufficiently large compared with $J_{\text{S}}$. Since the short time limit is always Markovian as discussed in ii), here we focus on the long-time limit. Physically, when the coupling between system and environment is strong enough compared with the internal energy scales of the system, all Majorana fermions in the system tend to be maximally entangled with the environment, because the environment contains more degrees of freedom. Consequently, the entropy is expected to saturate to the maximum value $(N/2)\log 2$ in the long-time limit, which is the same as the Markovian case. This can also be shown more rigorously using the path- integral formalism by relating the Rényi entropy to the inner product of Kourkoulou-Maldacena pure states zhang2020entanglement ; Liu2020non . In Fig. 2, one can see that the entropy curve gradually approaches the Markovian result as $\kappa$ increases. Figure 3: The entropy curve for varying $\beta_{\text{S}}$ with fixed $\kappa=0.1$ and $\beta_{\text{E}}=3$. Three different $\beta_{\text{S}}=(2,3,8)$ are plotted. Page-like behaviours is guaranteed when $\beta_{\text{S}}=2$ because the initial entropy is higher than the saturated entropy. The Page Curve. Above we have shown that, under three situations, our model recovers the Markovian results, and the Markovian results do not display the Page curve for entropy dynamics. Hence, to reveal effects beyond the Markovian approximation, the following three conditions should be satisfied simultaneously, which are: i) the bath temperature should not be too high; ii) the evolution time should not be too short; iii) the coupling between system and environment should not be too large. Under these conditions, we find that Page curve for entropy dynamics is often observed, as was shown in Fig. 3. Thus, this attributes the emergence of the Page curve to the beyond Markovian effect. Since we have discussed that the entropy always increases at the initial time, it is essential to understand the decreasing behavior of the entropy curve at long time to understand the Page curve. Below we offer two physical understandings. The first understanding again relies on perturbation theory. When $\kappa$ is small, the entropy dynamics can be obtained by doing perturbation in $\kappa$, which yields the same perturbative results as Eq. 10. Here, since we focus on the long-time behavior, we can simply replace $t$ by infinity and the range of integration in Eq. 10 is set to be $(-\infty,\infty)$. By expressing the Green’s functions $G^{>}$ and $G^{W}$ in terms of the spectral functions $\rho$ maldacena2016remarks $\displaystyle G^{>}(\omega,\beta)=\rho(\omega)\frac{1}{1+e^{-\beta\omega}},$ (12) $\displaystyle G^{W}(\omega,\beta)=\rho(\omega)\frac{1}{2\cosh{(\beta\omega/2)}},$ (13) and making use of the fact that $\rho(\omega)$ is even in $\omega$, we obtain the following expression: $\displaystyle\frac{dS_{\text{S}}^{(2)}}{dt}=2V^{2}N\int_{0}^{\infty}\left[\frac{d\omega}{2\pi}\frac{\rho_{\text{S}}(\omega,2\beta_{S})\rho_{\text{E}}(\omega,\beta_{\text{E}})}{2\cosh\beta_{S}\omega}\right.$ $\displaystyle\left.\times\frac{(e^{\beta_{S}\omega}-1)(1-e^{(\beta_{\text{E}}-\beta_{S})\omega})}{1+e^{\beta_{\text{E}}\omega}}\right].$ (14) Note that in Eq. 14, all terms are positive definite except for the $1-e^{(\beta_{\text{E}}-\beta_{S})\omega}$ term. When the temperature of the environment is lower than the temperature of the system, $\beta_{\text{E}}>\beta_{\text{S}}$ and $1-e^{(\beta_{\text{E}}-\beta_{S})\omega}<0$. Therefore, $dS_{\text{S}}^{(2)}/dt<0$ and the entropy decreases at long time, which yields the Page curve. This gives a sufficient condition for the emergence of the Page curve, that is, the small $\kappa$ and the lower environment temperature, which also agrees with the three aforementioned conditions. The second understanding replies on inspecting how the system entropy saturates at sufficiently long times. It is reasonable to assume that the system eventually reaches thermal equilibrium with the environment, and since the environment contains much more degrees-of-freedom, the saturation entropy is determined by $\kappa$ and $\beta_{\text{E}}$ and is independent of $\beta_{\text{S}}$. The saturation entropy is smaller when the environment temperature is lower, which corresponds to a decrease in thermal entropy. When the coupling $\kappa$ is smaller the saturation entropy is also smaller because this lowers the entanglement entropy. On the other hand, the initial entropy of the system is mainly determined by parameters $J_{\text{S}}$ and $\beta_{\text{S}}$ of the system itself. When this saturation entropy is smaller than the initial entropy, the entropy has to decrease at a later stage, which also leads to a sufficient condition for the emergence of the Page curve. Summary. In summary, we address the issue of when a Page curve can emerge in entropy dynamics of a system coupled to the environment. Although we use SYK model as an exactly solvable model to study this problem, the lesson we learn from our model reveals a general physical picture that should be applicable in generic chaotic quantum many-body systems. This physical picture contains two ingredients. First, at the initial stage, the entropy dynamics is always dominated by the Markovian process which leads to a linear increase of entropy in time. Secondly, a chaotic system thermalizes with the environment in the long-time limit. After thermalization, a low environment temperature and a weak system-environment coupling respectively suppress the thermal and the entanglement contributions to the system entropy, which ensures a lower system entropy at long time and forces the entropy to decrease at the later stage. The long-time decreasing behavior is essential for the emergence of the Page curve. This long-time behavior is distinct from the Markovian case where the system is often heated to infinite temperature and the long-time steady state is described by a density matrix given by the identity matrix. Therefore, the Page curve is a consequence of the non-Markovian environment. Since the entanglement entropy can now be measured experimentally Greiner1 ; Greiner2 ; Greiner3 and the coupling to the environment can be also highly controllable, for instance, in ultracold atomic systems, the physical picture revealed in this work can be experimentally verified in near future. Acknowledgment. This work is supported by Beijing Outstanding Young Scientist Program, NSFC Grant No. 11734010, MOST under Grant No. 2016YFA0301600. Note added. When finishing the manuscript, we become aware of a work by Chen in which the Rényi entropy dynamics has been stuided in a similar model by the perturbation theory Yu . ## References * (1) J. Preskill, Quantum 2, 79 (2018). * (2) M. O. Scully, and M. S. Zubairy, Quantum Optics, Cambridge University Press, Cambridge, 1997. * (3) H. P. Breuer and F. Petruccione, The Theory of Open Quantum Systems, Oxford University Press, Oxford, 2007. * (4) S. W. Hawking, Commun.Math. Phys. 43 (1975) 199-220. * (5) D. N. Page, Phys. Rev. Lett. 71 (1993)1291. * (6) G. Penington, J. High Energ. Phys. 2020, 2 (2020). * (7) A. Almheiri, N. Engelhardt, D. Marolf, and H. Maxfield, J. High Energ. Phys. 2019 63 (2019). * (8) A. Almheiri, R. Mahajan, J. Maldacena, and Y. Zhao, J. High Energ. Phys. 2020, 149 (2020). * (9) A. Almheiri, T. Hartman, J. Maldacena, E. Shaghoulian, and A. Tajdini, J. High Energ. Phys. 2020, 13 (2020). * (10) G. Penington, S. H. Shenker, D. Stanford, and Z. Yang, arXiv:1911.11977. * (11) J. Maldacena, and D. Stanford, Physical Review D 94 (2016) 106002. * (12) A. Kitaev, and S. Josephine Suh, J. High Energ. Phys. 2018, 183 (2018). * (13) Y. Gu, X.-L. Qi and D. Stanford, J. High Energ. Phys. 2017 125 (2017). * (14) R. A. Davison, W. Fu, A. Georges, Y. Gu, K. Jensen and S. Sachdev, Phys. Rev. B 95 155131 (2017). * (15) X. Chen, R. Fan, Y. Chen, H. Zhai and P. Zhang, Phys. Rev. Lett. 119 207603 (2017). * (16) S. Banerjee and E. Altman, Phys. Rev. B 95, 134302 (2017). * (17) S.-K. Jian and H. Yao, Phys. Rev. Lett. 119, 206602 (2017). * (18) X.-Y. Song, C.-M. Jian and L. Balents, Phys. Rev. Lett. 119 216601 (2017). * (19) P. Zhang, Phys. Rev. B 96 205138 (2017). * (20) C.-M. Jian, Z. Bi and C. Xu, Phys. Rev. B 96 115122 (2017). * (21) A. Eberlein, V. Kasper, S. Sachdev and J. Steinberg, Phys. Rev. B 96 205123 (2017). * (22) Y. Gu, A. Lucas and X.-L. Qi, J. High Energ. Phys. 2017, 120 (2017). * (23) Y. Chen, H. Zhai and P. Zhang, J. High Energ. Phys. 2017 150 (2017). * (24) P. Zhang, Phys. Rev. B 100 245104 (2019). * (25) A. Almheiri, A. Milekhin and B. Swingle, arXiv:1912.04912. * (26) Y. Chen, X.-L. Qi and P. Zhang, J. High Energ. Phys. 2020, 121 (2020). * (27) P. Zhang, J. High Energ. Phys. 2020, 2 (2020). * (28) P. Zhang, C. Liu, and X. Chen, SciPost Phys. 8, 094 (2020). * (29) C. Liu, P. Zhang, and X. Chen, arXiv:2008.11955. * (30) S.-K. Jian, and B. Swingle, arXiv:2011.08158. * (31) P. Dadras, A. Kitaev, arXiv: 2011.09622. * (32) J. Maldacena, D. Stanford and Z. Yang, Prog Theor Exp Phys 2016 (12): 12C104. * (33) J. Polchinski and V. Rosenhaus, J. High Energ. Phys. 2016, 1 (2016). * (34) K. Jensen, Phys. Rev. Lett. 117, 111601 (2016). * (35) A. Jevicki and K. Suzuki, J. High Energ. Phys. 2016, 46 (2016). * (36) G. Mandal, P. Nayak, and S. R. Wadia, arXiv:1702.04266. * (37) D. J. Gross and V. Rosenhaus, J. High Energ. Phys. 2017, 92 (2017). * (38) Y. N. Zhou, L. Mao and H. Zhai, arXiv: 2101.*****. * (39) L. Pan, X. Chen, Y. Chen, and H. Zhai, Nat. Phys. 16, 767 (2020). * (40) R. Islam, R. Ma, P. M. Preiss, M. Eric Tai, A. Lukin, M. Rispoli, M. Greiner, Nature 528, 77 (2015). * (41) A. M. Kaufman, M. Eric Tai, A. Lukin, M. Rispoli, R. Schittko, P. M. Preiss, M. Greiner, Science 353, 794 (2016). * (42) A. Lukin, M. Rispoli, R. Schittko, M. Eric Tai, A. M. Kaufman, S. Choi, V. Khemani, J. Léonard, M. Greiner, Science 364, 256 (2019). * (43) Y. Chen, arXiv: 2012.00223.
Gravitational waves from type II axion-like curvaton model and its implication for NANOGrav result Masahiro Kawasaki(a,b) and Hiromasa Nakatsuka(a) (a)ICRR, The University of Tokyo, Kashiwa, Chiba 277-8582, Japan (b)Kavli IPMU (WPI), UTIAS, The University of Tokyo, Kashiwa, Chiba 277-8583, Japan ## 1 Introduction North American Nanohertz Observatory for Gravitational Waves (NANOGrav) is one of the Pulsar-timing-array experiments, which has been trying to detect the gravitational wave (GW) signal through the long-term observation of pulsars. Recently the NANOGrav reported a hint of the stochastic GW signal in their 12.5–year observation [1]. This signal can be explained by the stochastic GW with $\Omega_{\rm GW}\sim 10^{-9}$ at $f\sim 10^{-8}\mathrm{Hz}$, which can be sourced by cosmic string [2, 3, 4, 5, 6, 7], phase transitions [8, 9, 10, 11, 12], large density fluctuations in primordial black hole (PBH) formation models [13, 14, 15, 16, 17, 18, 19, 20, 21, 22] and other origin [23, 24, 25]. The PBH formation scenario is an attractive candidate for the NANOGrav signal since it can simultaneously explain the observed $30{M_{\odot}}$ black holes in the binary merger events in LIGO-Virgo collaboration [26, 27, 28, 29]. The PBH formation requires the large curvature power spectrum ($\mathcal{P}_{\zeta}(k_{\rm pbh})\sim 0.02$) at a wavenumber $k_{\rm pbh}$ related to the mass of the formed PBHs. Since PBHs are formed when the high- density regions enter the horizon and collapse, the PBH mass is roughly given by the horizon mass at the collapse, which is related to the frequency of the density fluctuation $f$ ($=k_{\text{pbh}}/2\pi$) as [30] $\displaystyle M(f)$ $\displaystyle\simeq 30{M_{\odot}}\left(\frac{\gamma}{0.2}\right)\left(\frac{\text{g}_{*}}{10.75}\right)^{-1/6}\left(\frac{f}{5.3\times 10^{-10}\,\mathrm{Hz}}\right)^{-2},$ (1) where $\gamma\sim 0.2$ is a ratio of PBH mass to the horizon mass [31] and $\text{g}_{*}$ is the numbers of relativistic degrees of freedom at the PBH formation. The same density fluctuations that collapse to form PBHs also generate a stochastic background of GWs through the nonlinear coupling, $\Braket{\zeta\zeta h}$ when they enter the horizon [32, 33, 34], with a spectrum peaking at the frequency $f\sim 10^{-9}\mathrm{Hz}$, very close to the one of the NANOGrav signal. In this paper, we focus on the axion-like curvaton model of the PBH formation scenario [35, 36, 37, 15, 38]. There are two types of the axion-like curvaton models; in one type the complex field $\Phi$ (whose phase part is the axion) rolls down toward origin during inflation [35, 36, 37, 15], and in the other type $\Phi$ rolls down from the origin [38]. These two types lead to the different dynamics and the power spectrum of induced GWs. The former type (type I) has been already studied and successfully explains the NANOGrav signal [15]. The type II model was first proposed in [38] where the model parameters were chosen to achieve a narrow density power spectrum of the density perturbations. However, a broader power spectrum is preferable to account for the NANOGrav signal. The goal of this paper is to investigate if such a broad power spectrum can be produced also in the type II model. It is found that by choosing appropriate parameters of the potential term in Eq. (15), we obtain the desired broader spectral shape. Moreover, since our model produces large positive non-Gaussianity on the density power spectrum, we can achieve the required abundance of PBHs by the smaller amplitude of the density power spectrum than the Gaussian one. We improve the treatment of the non- Gaussianity of this model to accurately estimate the required amplitude of density power spectrum, and also the induced GW. This paper is organized as follows. We calculate the curvature power spectrum of the type II axion-like curvaton model in Sec. 2. Next, we calculate the PBH abundance in Sec.3 and induced GWs in Sec. 4. We can achieve the broad power spectrum of induced GWs and explain the NANOGrav signal. In Sec. 5, we conclude this paper. ## 2 Type II axion-like curvaton model We now briefly summarize the dynamics of background and fluctuations in the type II axion-like curvaton model. (See [38] for the detailed calculation.) In this model, large fluctuations are produced in the phase direction of a complex scalar field $\Phi$, which we call “axion-like curvaton”. The potential of $\Phi$ is given by [38] $\displaystyle V_{\Phi}=\frac{\lambda}{4}\left(|\Phi|^{2}-\frac{v^{2}}{2}\right)^{2}+gI^{2}|\Phi|^{2}-v^{3}\epsilon(\Phi+\Phi^{*}),$ (2) where $I$ is the inflaton field, $\lambda$ and $g$ are coupling constants, $v/\sqrt{2}$ is the vacuum expectation value after inflation. The bias term, $v^{3}\epsilon(\Phi+\Phi^{*})$, is introduced to avoid the cosmic string problem and the stochastic effect on dynamics of $\Phi$, which requires the complicated treatment of fluctuations. Note that $\epsilon$ is naturally small ($\epsilon\ll 1$) in the sense of ’t Hooft’s naturalness [39] since $U(1)$ symmetry is restored for $\epsilon=0$. The field value of $\Phi$ changes during inflation due to the first and second term of Eq. (2). In the early stage of inflation ($I\gtrsim(\lambda/g)^{1/2}v$), the interaction term with the inflaton fixes $\Phi$ near the origin. In the late stage, the inflaton field value becomes smaller and $\Phi$ starts to roll down the Higgs-like potential towards $|\Phi|=v$. In this dynamics the phase direction acquires large fluctuations $\sim H/|\Phi|$ ($H$: Hubble parameter during inflation). The PBH formation requires the large fluctuations with wavenumber $k=k_{\rm pbh}\simeq 10^{5}\,\mathrm{Mpc}^{-1}$, which corresponds to the scale of $30{M_{\odot}}$ PBHs. Thus, we determine the inflaton coupling $g$ so that $\Phi$ starts to roll down potential at $t_{\rm pbh}$ satisfying $H(t_{\text{pbh}})=k_{\text{pbh}}/a(t_{\text{pbh}})$, which leads to $\displaystyle g\simeq\frac{\lambda}{4}\left(\frac{v}{I(t_{\rm pbh})}\right)^{2}.$ (3) We also assume that the effective mass of $\Phi$ is much larger than the Hubble parameter ($\lambda v^{2}\gg H^{2}$) until $t_{\rm pbh}$ to suppress the fluctuation at the larger wavelengths. This condition also ensures independence on the initial condition since $\Phi$ settles near the origin of the potential independently of the initial field value of $\Phi$. Based on the above background field dynamics, we calculate the fluctuation of $\Phi$, which is decomposed as $\displaystyle\Phi=\frac{1}{\sqrt{2}}(\varphi_{0}+\varphi)e^{i\theta}$ (4) with the homogeneous solution $\varphi_{0}$, and the perturbations $\varphi$ and $\theta$. The phase direction $\theta$ works as the curvaton, and the canonical field of the phase direction is defined as $\tilde{\sigma}\equiv\varphi_{0}\theta$. The inflation induces the quantum fluctuations with amplitude $H/(2\pi)$ for $\tilde{\sigma}$ at the horizon crossing. Thus, the power spectrum of the $\theta$ fluctuations is given by $\displaystyle\mathcal{P}_{\theta}(k,t_{k})=\left(\frac{H(t_{k})}{2\pi\varphi_{0}(t_{k})}\right)^{2},$ (5) where $t_{k}$ is the time when the fluctuation with $k$ crosses the horizon. $\mathcal{P}_{\theta}(k,t_{k})$ is suppressed for $k>k_{\text{pbh}}$ since the $\varphi_{0}(t_{k})$ quickly grows after $t_{\text{pbh}}$. While the $\Phi$ is fixed near the origin before $t_{\text{pbh}}$, $\tilde{\sigma}$ has the large effective mass given by $\displaystyle\tilde{m}_{\sigma}^{2}$ $\displaystyle\equiv\frac{\partial^{2}\left(-v^{3}\epsilon(\Phi+\Phi^{*})\right)}{\partial\tilde{\sigma}^{2}}\bigg{|}_{\tilde{\sigma}=0}={\sqrt{2}\epsilon v^{2}\frac{v}{\varphi_{0}}},$ (6) where we choose the model parameter so as $m_{\tilde{\sigma}}^{2}>H^{2}$ for $t<t_{\text{pbh}}$. Such large effective mass damps the fluctuations as $\tilde{\sigma}\propto a^{-3/2}$, and we define the damping factor of the phase direction with momentum $k$ as $\displaystyle R_{k}$ $\displaystyle\equiv\left(\frac{\tilde{\sigma}_{k}(t_{\text{pbh}})}{\tilde{\sigma}_{k}(t_{k})}\right)\sim\left(\frac{k}{k_{\text{pbh}}}\right)^{3/2}\text{ for }k<k_{\text{pbh}}.$ (7) ($R_{k}=1$ for $k>k_{\text{pbh}}$.) Finally, the power spectrum of the $\theta$ fluctuations at the end of inflation $t_{e}$ is given by $P_{\theta}(k,t_{e})=R_{k}^{2}\mathcal{P}_{\theta}(k,t_{k}).$ (8) Let us evaluate the density perturbations induced by the fluctuations of the phase direction. We suppose that the curvaton obtains the axion-like potential after inflation through some nonperturbative effect. The potential minimum of the nonperturbative term does not coincide with that of the primordial one determined by the bias term in our model. Suppose that nonperturbative potential takes the minimum at $\theta=\theta_{i}$, then it is written as $\displaystyle V_{\sigma}=\Lambda^{4}\left[1-\cos\left(\Theta\right)\right]\simeq\frac{1}{2}m_{\sigma}^{2}\sigma^{2},$ (9) where $\Theta\equiv\theta-\theta_{i}$, the curvaton $\sigma$ is defined as $\sigma\equiv v\Theta$ and $m_{\sigma}=\Lambda^{2}/v$. We assume that $m_{\sigma}$ is small enough to neglect the axion-like potential during inflation, i.e. $H^{2}\gg m_{\sigma}^{2}$. The density fluctuation is given by ${\delta\rho_{\sigma}}/{\rho_{\sigma}}=2\delta\theta/\theta_{i}$. Neglecting a small contribution from the inflaton ($\mathcal{P}_{\zeta}\sim 10^{-10}$) compared to the curvaton, the power spectrum of curvature perturbations is given by $\displaystyle\mathcal{P}_{\zeta}(k)$ $\displaystyle=\left(\frac{r}{4+3r}\right)^{2}\left(\frac{2}{\theta_{i}}\right)^{2}\mathcal{P}_{\theta}(k,t_{k})R_{k}^{2},$ (10) where $r$ is the ratio of the energy density of the curvaton to that of the inflaton (or radiation after reheating), $r=\rho_{\sigma}/\rho_{I}$. Assuming the instant reheating at $t_{\rm reheat}$, the ratio is given by $r(t_{\rm reheat})=(v^{2}\theta_{i}^{2})/(6M_{pl}^{2})$, which is chosen to be small to ensure that $\Phi$ does not disturb the inflation. $r$ grows during radiation- dominated era as $r\propto a$ due to the matter-like behavior of the curvaton, and its growth ends at the curvaton decay, $t_{\rm decay}$. We require that the curvaton decays into radiation before it overcloses the universe, $r(t_{\rm decay})\sim 0.5$, at which the temperature of the universe is $T_{\rm decay}\sim(r(t_{\rm reheat})/r(t_{\rm decay}))T_{\rm reheat}$. The typical decay rate of $\Phi$ is related to the curvaton mass as $\Gamma_{\sigma}=\kappa^{2}m_{\sigma}^{3}/(16\pi v^{2})$ where $\kappa$ is a coupling constant. We confirm that $T_{\rm decay}\sim 10^{3}\mathrm{GeV}$ is achieved for $m_{\sigma}\sim 10^{8}\mathrm{GeV}$ for our parameters in Eq. (15). In the following, $r$ refers to the energy ratio after decay, that is $r\equiv r(t_{\rm decay})$. The PBH abundance highly depends on the eventual non-Gaussian distribution of $\mathcal{P}_{\zeta}(k)$. It is known that the axion-like curvaton models produce large non-Gaussianity with local type bispectrum characterized by the following parameter: [37] $\displaystyle f_{\rm NL}=\frac{5}{12}\left(-3+\frac{4}{r}+\frac{8}{4+3r}\right).$ (11) We discuss the enhancement of the PBH abundance by non-Gaussianity in Sec.3. Non-Gaussianity also affects the power spectrum of the curvature perturbations and induced gravitational wave through the higher-order correlations [40, 41], whose contribution is characterized by $\mathcal{P}_{\zeta}f_{\rm NL}^{2}$. We take non-Gaussianity into account only approximately since the Gaussian contribution is the dominant for our choice of parameters given by Eq. (15), $\mathcal{P}_{\zeta}f_{\rm NL}^{2}<0.04$. We estimate non-Gaussian amplification on the curvature power spectrum as $\displaystyle Q^{\rm(NL)}_{\mathcal{P}}\equiv\frac{\mathcal{P}_{\zeta}^{\rm(NL)}(k_{*})}{\mathcal{P}_{\zeta}(k_{*})}$ $\displaystyle=1+\left(\frac{3}{5}f_{\rm NL}\right)^{2}\frac{k_{*}^{3}}{2\pi\mathcal{P}_{\zeta}({\bm{k}_{*}})}\int\text{d}^{3}q\frac{\mathcal{P}_{\zeta}({\bm{q}})\mathcal{P}(|{\bm{k}_{*}-\bm{q}}|)}{q^{3}|\bm{k}_{*}-\bm{q}|^{3}},$ (12) where $k_{*}$ is the wavenumber at the peak of $\mathcal{P}_{\zeta}$. $Q^{\rm(NL)}_{\mathcal{P}}$ is about $1.07$ for $r=0.5$ and $1.01$ for $r=1.0$ in our parameter set. Finally, the power spectrum of the density perturbations is given by $\displaystyle\mathcal{P}_{\delta}(k,t)$ $\displaystyle=\left(\frac{2}{3}\frac{k}{a(t)H(t)}\right)^{4}T(k\eta(t))^{2}\mathcal{P}_{\zeta}(k),$ (13) where $\eta(t)$ is the conformal time and $T(x)$ is the transfer function during radiation dominated era, which includes the suppression of the density perturbation in sub-horizon as $\displaystyle T(x)\equiv$ $\displaystyle~{}3\frac{\sin(x/\sqrt{3})-(x/\sqrt{3})\cos(x/\sqrt{3})}{(x/\sqrt{3})^{3}}.$ (14) In this paper, we use the following set of model parameters: $\displaystyle\lambda$ $\displaystyle=7.5\times 10^{-6},\quad v=5.77\times 10^{-2}{M_{\text{pl}}},\quad$ $\displaystyle g$ $\displaystyle=9.56\times 10^{-11},\quad\epsilon=2.57\times 10^{-10},$ $\displaystyle\begin{cases}&r=0.5,\quad\theta_{i}=5.3\times 10^{-2}\\\ &r=1.0,\quad\theta_{i}=6.5\times 10^{-2}\end{cases},$ (15) where we take $r$ and $\theta_{i}$ to achieve the PBH abundance required to explains the LIGO-Virgo event rate. Here we remark some relations between the parameters and the shape of the power spectrum. The larger $r$ or smaller $\theta_{i}$ increases the abundance of ALP as Eq. (10), and smaller $\epsilon$ changes the dynamics of $\varphi_{0}$ at $t\sim t_{\rm pbh}$, both of which result in larger amplitude of the power spectrum. The peak wavenumber of the power spectrum is determined by the ratio of potential terms, $g/(\lambda v^{2})$, as discussed in Eq. (3). The width of the spectrum depends on how slowly $\varphi_{0}$ changes since the fluctuation of $\theta$ with mode $k$ depends on the field value $\varphi_{0}$ at the horizon crossing as Eq. (5). We flatten the potential by choosing the smaller $v$, $g$ and $\lambda$ compared to the previous paper [38] to achieve a broad power spectrum. We numerically evaluate the dynamics of $\Phi$ assuming the chaotic inflation whose potential is given by $V_{I}=m_{I}^{2}I^{2}/2$ with $m_{\phi}\simeq 10^{13}\mathrm{GeV}$. The similar dynamics also holds for other inflation models. We show the curvature power spectrum $\mathcal{P}_{\zeta}(k)$ [Eq. (10)] in Fig. 1. We also show the constraints on $\mu$–distortion [42] by COBE/FIRAS [43] and BBN [44], and our model safely avoids the current constraints. Figure 1: The curvature power spectrum $Q^{\rm(NL)}_{\mathcal{P}}\mathcal{P}_{\zeta}(k)$ based on Eqs.(10) and (12). The non-Gaussian contribution is included by $Q^{\rm(NL)}_{\mathcal{P}}$. The orange regions are constraints on the curvature power spectrum from $\mu$–distortion [42] by COBE/FIRAS [43] and BBN [44]. ## 3 PBH formation The PBHs are formed by the collapse of high-density regions when they re-enter the horizon. The criterion on the PBH formation is estimated by the numerical simulation [45], in which the threshold value of the averaged density fluctuation is obtained. Since the threshold value is too large to neglect the nonlinear contribution, we need to use the effective threshold value including the nonlinear relation between density and curvature perturbations. The detailed analysis show that the effective threshold value is $\delta_{\rm th(eff)}\simeq\sqrt{2}\times 0.53$ for the averaged linear density perturbation defined by [46] $\displaystyle\bar{\delta}_{R}(\bm{x})$ $\displaystyle\equiv\int\text{d}^{3}yW(|\bm{x}-\bm{y}|,R)\delta(\bm{y})$ $\displaystyle=\int\frac{\text{d}^{3}p}{(2\pi)^{3}}\tilde{W}(pR)e^{i\bm{p}\cdot\bm{x}}\delta_{p},$ (16) where $R=k^{-1}$ is the scale of the horizon corresponding to the PBH mass [Eq.(1)], $W(x,R)$ and $\tilde{W}(z)$ are the window functions in the real and Fourier spaces, and the $\delta_{p}$ is the density fluctuation. Although it is known that the choice of window function causes a large uncertainly [47], a natural choice, often used in the literature, is the real-space top-hat window function used in the threshold value in the numerical simulation, $\displaystyle\tilde{W}(z)$ $\displaystyle=3\frac{\sin(z)-z\cos(z)}{z^{3}}.$ (17) Thus, a PBH is formed when a region has the averaged density $\bar{\delta}_{R}$ larger than the threshold value $\delta_{\rm th(eff)}$. We estimate the PBH formation rate based on the Press–Schechter formalism, where the PBH formation rate is calculated by the probability distribution of the averaged density perturbations. In our model, $\bar{\delta}_{R}$ follows the non-Gaussian distribution due to $f_{\rm NL}$ given by Eq. (11), which drastically changes the PBH abundance. The probability distribution of $\bar{\delta}_{R}$ is characterized by the variance and skewness defined by $\displaystyle\sigma_{R}^{2}$ $\displaystyle=\Braket{\bar{\delta}_{R}^{2}}-\Braket{\bar{\delta}_{R}}^{2},$ (18) $\displaystyle\mu_{R}$ $\displaystyle=\sigma_{R}^{-3}\left(\Braket{\bar{\delta}_{R}^{3}}-3\Braket{\bar{\delta}_{R}^{2}}\Braket{\bar{\delta}_{R}}+2\Braket{\bar{\delta}_{R}}^{3}\right),$ (19) where $\Braket{...}$ describes the ensemble average of $\delta_{p}$. For simplicity, we neglect the scale dependence of $\mu_{R}$ and evaluate it at $R=R_{\rm pbh}=k_{\rm pbh}^{-1}$, which corresponds to $30{M_{\odot}}$ PBHs. Using the formula in [46], the skewness is numerically given by $\displaystyle\mu\equiv\mu_{R}|_{R=R_{\rm pbh}}\simeq 3.39f_{\rm NL}\sigma_{R}|_{R=R_{\rm pbh}}.$ (20) We construct the statistical variable which reproduces the probability distribution of $\bar{\delta}_{R}$. Using the Gaussian variable $\chi$ characterized by $\Braket{\chi^{2}}=\sigma_{R}^{2}$, we define the statistical variable as $\displaystyle\bar{\delta}[\chi]\equiv\chi+\frac{\mu}{6\sigma_{R}}(\chi^{2}-\sigma_{R}^{2}),$ (21) which has the same variance and skewness in Eqs. (18) and (19) up to $\mathcal{O}(f_{\rm NL}\sigma_{R})$. The probability distribution of $\bar{\delta}[\chi]$ is given by [48, 49, 50] $\displaystyle P_{R}^{\rm(NG)}(\bar{\delta})$ $\displaystyle=\sum_{i=\pm}\left|\frac{\text{d}\chi_{i}(\bar{\delta})}{\text{d}\bar{\delta}}\right|P_{R}^{\rm(G)}(\chi_{i}(\bar{\delta})),$ (22) where $P_{R}^{\rm(G)}(\chi)$ is the Gaussian distribution function, $\displaystyle P_{R}^{\rm(G)}(\chi)=\frac{1}{\sqrt{2\pi}\sigma_{R}}\exp\left(-\frac{1}{2}\frac{\chi^{2}}{\sigma_{R}^{2}}\right),$ (23) and $\chi_{\pm}(\bar{\delta})$ are two solutions of $\bar{\delta}=\bar{\delta}[\chi]$, $\displaystyle\chi_{\pm}(\bar{\delta})=\frac{3\sigma_{R}}{\mu}\left(-1\pm\sqrt{1+\frac{2\mu}{3}\left(\frac{\mu}{6}+\frac{\bar{\delta}}{\sigma_{R}}\right)}\right).$ (24) The PBH formation rate is given by the probability of $\bar{\delta}>\delta_{\rm th(eff)}$, which leads to $\displaystyle\beta(R)$ $\displaystyle=\int_{\bar{\delta}>\delta_{\rm th(eff)}}P_{R}^{\rm(NG)}(\bar{\delta})\text{d}\bar{\delta}=\int_{\bar{\delta}[\chi]>\delta_{\rm th(eff)}}P_{R}^{\rm(G)}(\chi)\text{d}\chi$ $\displaystyle\simeq\frac{\sigma_{R}}{\sqrt{2\pi}\chi_{+}(\delta_{\rm th(eff)})}\exp\left(-\frac{\chi_{+}(\delta_{\rm th(eff)})^{2}}{2\sigma_{R}^{2}}\right).$ (25) Here we have used $\mu>0$ and $\chi_{+}(\delta_{\rm th(eff)})/\sigma_{R}\gg 1$ in the last line. The present PBH abundance is given by $\displaystyle f(M)\equiv\frac{\text{d}\Omega_{\text{PBH}}}{\text{d}\ln M}\frac{1}{\Omega_{\text{DM}}}$ (26) $\displaystyle=\frac{\beta(R(M))}{1.8\times 10^{-8}}\left(\frac{\gamma}{0.2}\right)^{3/2}\left(\frac{10.75}{\text{g}_{*}}\right)^{1/4}\left(\frac{0.12}{\Omega_{\text{DM}}h^{2}}\right)\left(\frac{M}{{M_{\odot}}}\right)^{-1/2},$ (27) where $\text{g}_{*}$ is the number of relativistic degrees of freedom at $T\sim 30\mathrm{MeV}$. We show the calculated mass spectrum of PBH in Fig. 2. We also plot the relevant constrains by microlensing experiments “MACHO/EROS/OGLE” [51, 52, 53] and energy injection into CMB through accretion around PBHs [54]. Since our model predicts the broad mass distribution, the large mass part of the distribution could conflict with the accretion constraints. However, it is noticed that accretion constraint has a large uncertainty. In fact, two different constraints are obtained depending on assumptions as shown in Fig. 2. Our mass distribution is consistent if we adopt the weaker accretion constraint. Figure 2: The mass spectrum of PBHs (blue line) and the constraints (orange regions) by microlensing experiments “MACHO/EROS/OGLE”(solid line) [51, 52, 53], CMB through spherical accretion (dashed line) and disk accretion (dotted line) [54]. ## 4 Induced gravitational waves The large density fluctuations induce the gravitational waves through the nonlinear coupling $\Braket{\zeta\zeta h}$ when they re-enter the horizon. The current energy fraction of GWs is written as $\displaystyle\Omega_{\rm GW}(t_{0},k)$ $\displaystyle=\left(\frac{a_{c}^{2}H_{c}}{a_{0}^{2}H_{0}}\right)^{2}\Omega_{\rm GW}(\eta_{c},k)$ $\displaystyle\simeq 0.83\left(\frac{\text{g}_{c}}{10.75}\right)^{-1/3}\Omega_{r,0}\Omega_{\rm GW}(t_{c},k),$ (28) where $\Omega_{r,0}$ is the current energy fraction of radiation, the subscript “c” denotes values when GW production effectively finishes. The energy density of the induced GWs at $t_{c}$ is calculated by solving the equation of motion of GWs with the source term of scalar fluctuations, and it is given by [30] $\displaystyle\Omega_{\rm GW}(t_{c},k)=\frac{8}{243}\int^{\infty}_{0}\text{d}y\int^{1+y}_{\left|1-y\right|}\text{d}x\mathcal{P}_{\zeta}(kx)\mathcal{P}_{\zeta}(ky)\frac{y^{2}}{x^{2}}$ $\displaystyle\qquad\times\left(1-\frac{(1+y^{2}-x^{2})^{2}}{4y^{2}}\right)^{2}\overline{\mathcal{I}(x,y,\eta_{c})}^{2},$ (29) where overline means the time average over $\eta_{c}$. $\mathcal{I}(x,y,\eta_{c})$ is written as $\displaystyle\mathcal{I}(x,y,\eta_{c})=\frac{k^{2}}{a(\eta_{c})}\int^{\eta_{c}}\text{d}\bar{\eta}a(\bar{\eta})g_{k}(\eta_{c};\bar{\eta})f(ky,kx,\bar{\eta}).$ (30) Here $g_{k}$ is the Green function, $\displaystyle g_{k}(\eta,\tilde{\eta})$ $\displaystyle=\frac{\sin(k(\eta-\bar{\eta}))}{k}\theta(\eta-\bar{\eta}),$ (31) and $f(k_{1},k_{2},\eta)$ is given by $\displaystyle f(k_{1},k_{2},\eta)$ $\displaystyle=\bigg{[}2T(k_{1},\eta)T(k_{2},\eta)$ $\displaystyle+$ $\displaystyle\left(\frac{\dot{T}(k_{1},\eta)}{H(\eta)}+T(k_{1},\eta)\right)\left(\frac{\dot{T}(k_{2},\eta)}{H(\eta)}+T(k_{2},\eta)\right)\bigg{]},$ (32) where $T(x)$ is the transfer function of the scalar fluctuations given by Eq.(14). We comment on the non-Gaussian contribution on the induced gravitational waves [55, 40, 41]. It is pointed out in [41] that non-Gaussianity of scalar fluctuations amplifies the induced gravitational waves when $\mathcal{P}_{\zeta}f_{\rm NL}^{2}$ is large. Since $\mathcal{P}_{\zeta}f_{\rm NL}^{2}<0.04$ in our calculation, the effect of non-Gaussianity is expected to be sub-dominant. Thus, we approximately include the effect of the non- Gaussianity on $\Omega_{\rm GW}$ by multiplying the factor $(Q^{\rm(NL)}_{\mathcal{P}})^{2}$ given by Eq. (12). This approximation includes a part of non-Gaussian contributions, “Hybrid” and “Reducible”-type terms discussed in [40]. Hybrid type is a product of the Gaussian and non- Gaussian contribution of curvature perturbation, and Reducible type is that of non-Gaussian and non-Gaussian contribution. Although there are other types of sources of GWs, it is known that Hybrid-type is one of the largest contributions among them. Thus, our calculation can estimate most of the effects of non-Gaussianity. The estimated GW spectra for $r=0.5$ (blue line) and $r=1.0$ (orange line) are shown in Fig. 3. The GW spectrum observed by the NANOGrav experiment is fitted by the power-law spectrum around $f\sim 10^{-8}$, $\displaystyle\Omega_{\rm GW}(f)h^{2}=\frac{2\pi^{2}}{3}\frac{h^{2}f^{2}}{H_{0}^{2}}h_{c}^{2}(f)=A_{\Omega}f^{5-\gamma},$ (33) where $\gamma$ is the tilt of the spectrum. In Fig. 3, we show the observed GWs for $\gamma=5$ and 6 with 2-$\sigma$ uncertainty on $A_{\Omega}$. We also plot current constraints by other PTA experiments, EPTA (solid) [56] and PPTA (dotted)[57], and future sensitivity by SKA (dashed) [58]. It is found that the broad power spectrum of GWs in the present model can explain the reported NANOGrav signal. Figure 3: The induced GW spectrum and the constraints by PTA experiments. We include the contribution of non-Gaussianity by using the factor $Q^{\rm(NL)}_{\mathcal{P}}$. The orange lines are current constraints by EPTA (solid) [56] and PPTA (Dotted)[57], and future sensitivity by SKA (Dashed) [58]. We show the reported NANOGrav signal with $\gamma=5$ (blue region) and $\gamma=6$ (pink region) with 2-$\sigma$ uncertainty (see Eq. (33)). ## 5 Conclusion The reported signal by the NANOGrav experiment indicates the various cosmological phenomena like cosmic string, phase transition and PBH formation. The PBH formation scenario is attractive among them since the reported frequency $f\sim 10^{-8}\mathrm{Hz}$ is close to the scale of the density fluctuations to produce $30{M_{\odot}}$ PBH, which can explain the binary black holes observed by LIGO-Virgo collaboration. The typical PBH formation models predict the induced GWs with smaller frequency and larger amplitude compared to the NANOGrav signal. To avoid this difficulty, one needs to modify the induced GW spectrum by the broad power spectrum and the non-Gaussianity of density fluctuations, which can enhance the PBH formation rate and give a good fit to the NANOGrav signal. The axion- like curvaton model can achieve the required features. There are two types of the axion-like curvaton models; in type I the complex field rolls down toward origin during inflation [35, 36, 37, 15], and in type II it rolls down from the origin[38]. In this paper, we focused on the type II model and chose the appropriate parameters of the potential term, which leads to a broader power spectrum of the density perturbations than that in [38]. As a result it was found that the induced GW spectrum can explain the NANOGrav signal as shown in Fig. 3. Moreover, the broad power spectrum of the density fluctuation results in the broad mass spectrum of PBHs as shown in Fig. 2, which can be tested by the accumulation of the binary merger observations. Our model predicts large local-type non-Gauusianity which can be probed through observation of GWs. The spectrum of induced gravitational waves is a useful tool to distinguish our model from others. For example, cosmic string and type-I axion-like curvaton models gererally predict much broader GW spectra than our model. ## Acknowledgment This work was supported by JSPS KAKENHI Grant Nos. 17H01131 (M.K.), 17K05434 (M.K.), 20H05851 (M.K.), 21K03567(M.K.), JP19J21974 (H.N.), Advanced Leading Graduate Course for Photon Science (H.N.), and World Premier International Research Center Initiative (WPI Initiative), MEXT, Japan (M.K.). ## References * [1] NANOGrav collaboration, Z. Arzoumanian et al., _The NANOGrav 12.5-year Data Set: Search For An Isotropic Stochastic Gravitational-Wave Background_ , 2009.04496. * [2] J. Ellis and M. Lewicki, _Cosmic String Interpretation of NANOGrav Pulsar Timing Data_ , 2009.06555. * [3] S. Blasi, V. Brdar and K. Schmitz, _Has NANOGrav found first evidence for cosmic strings?_ , 2009.06607. * [4] W. Buchmuller, V. Domcke and K. Schmitz, _From NANOGrav to LIGO with metastable cosmic strings_ , _Phys. Lett. B_ 811 (2020) 135914, [2009.10649]. * [5] R. Samanta and S. Datta, _Gravitational wave complementarity and impact of NANOGrav data on gravitational leptogenesis: cosmic strings_ , 2009.13452. * [6] S. Datta, A. Ghosal, R. Samanta and R. Sinha, _Baryogenesis from ultralight primordial black holes and strong gravitational waves_ , 2012.14981. * [7] N. Ramberg and L. Visinelli, _The QCD Axion and Gravitational Waves in light of NANOGrav results_ , 12, 2020, 2012.06882. * [8] Y. Nakai, M. Suzuki, F. Takahashi and M. Yamada, _Gravitational Waves and Dark Radiation from Dark Phase Transition: Connecting NANOGrav Pulsar Timing Data and Hubble Tension_ , 2009.09754. * [9] A. Addazi, Y.-F. Cai, Q. Gan, A. Marciano and K. Zeng, _NANOGrav results and Dark First Order Phase Transitions_ , 2009.10327. * [10] A. Neronov, A. Roper Pol, C. Caprini and D. Semikoz, _NANOGrav signal from MHD turbulence at QCD phase transition in the early universe_ , 2009.14174. * [11] S.-L. Li, L. Shao, P. Wu and H. Yu, _NANOGrav Signal from First-Order Confinement/Deconfinement Phase Transition in Different QCD Matters_ , 2101.08012. * [12] B. Barman, A. Dutta Banik and A. Paul, _Implications of NANOGrav results and UV freeze-in in a fast-expanding Universe_ , 2012.11969. * [13] V. Vaskonen and H. Veermäe, _Did NANOGrav see a signal from primordial black hole formation?_ , 2009.07832. * [14] V. De Luca, G. Franciolini and A. Riotto, _NANOGrav Hints to Primordial Black Holes as Dark Matter_ , 2009.08268. * [15] K. Inomata, M. Kawasaki, K. Mukaida and T. T. Yanagida, _NANOGrav results and LIGO-Virgo primordial black holes in axion-like curvaton model_ , 2011.01270. * [16] K. Kohri and T. Terada, _Solar-Mass Primordial Black Holes Explain NANOGrav Hint of Gravitational Waves_ , _Phys. Lett. B_ 813 (2021) 136040, [2009.11853]. * [17] G. Domènech and S. Pi, _NANOGrav Hints on Planet-Mass Primordial Black Holes_ , 2010.03976. * [18] S. Sugiyama, V. Takhistov, E. Vitagliano, A. Kusenko, M. Sasaki and M. Takada, _Testing Stochastic Gravitational Wave Signals from Primordial Black Holes with Optical Telescopes_ , 2010.02189. * [19] M. Braglia, D. K. Hazra, F. Finelli, G. F. Smoot, L. Sriramkumar and A. A. Starobinsky, _Generating PBHs and small-scale GWs in two-field models of inflation_ , _JCAP_ 08 (2020) 001, [2005.02895]. * [20] N. Bhaumik and R. K. Jain, _Stochastic induced gravitational waves and lowest mass limit of primordial black holes with the effects of reheating_ , 2009.10424. * [21] M. Braglia, X. Chen and D. Kumar Hazra, _Probing Primordial Features with the Stochastic Gravitational Wave Background_ , _JCAP_ 03 (2021) 005, [2012.05821]. * [22] V. Atal, A. Sanglas and N. Triantafyllou, _NANOGrav signal as mergers of Stupendously Large Primordial Black Holes_ , 2012.14721. * [23] S. Vagnozzi, _Implications of the NANOGrav results for inflation_ , _Mon. Not. Roy. Astron. Soc._ 502 (2021) L11, [2009.13432]. * [24] S. Bhattacharya, S. Mohanty and P. Parashari, _Implications of the NANOGrav result on primordial gravitational waves in nonstandard cosmologies_ , 2010.05071. * [25] S. Kuroyanagi, T. Takahashi and S. Yokoyama, _Blue-tilted inflationary tensor spectrum and reheating in the light of NANOGrav results_ , _JCAP_ 01 (2021) 071, [2011.03323]. * [26] LIGO Scientific, Virgo collaboration, B. P. Abbott et al., _Binary Black Hole Population Properties Inferred from the First and Second Observing Runs of Advanced LIGO and Advanced Virgo_ , 1811.12940. * [27] S. Bird, I. Cholis, J. B. Muñoz, Y. Ali-Haïmoud, M. Kamionkowski, E. D. Kovetz et al., _Did LIGO detect dark matter?_ , _Phys. Rev. Lett._ 116 (2016) 201301, [1603.00464]. * [28] S. Clesse and J. García-Bellido, _The clustering of massive Primordial Black Holes as Dark Matter: measuring their mass distribution with Advanced LIGO_ , _Phys. Dark Univ._ 15 (2017) 142–147, [1603.05234]. * [29] M. Sasaki, T. Suyama, T. Tanaka and S. Yokoyama, _Primordial Black Hole Scenario for the Gravitational-Wave Event GW150914_ , _Phys. Rev. Lett._ 117 (2016) 061101, [1603.08338]. * [30] K. Inomata, M. Kawasaki, K. Mukaida and T. T. Yanagida, _Double inflation as a single origin of primordial black holes for all dark matter and LIGO observations_ , _Phys. Rev._ D97 (2018) 043514, [1711.06129]. * [31] B. J. Carr, _The Primordial black hole mass spectrum_ , _Astrophys. J._ 201 (1975) 1–19. * [32] R. Saito and J. Yokoyama, _Gravitational wave background as a probe of the primordial black hole abundance_ , _Phys. Rev. Lett._ 102 (2009) 161101, [0812.4339]. * [33] R. Saito and J. Yokoyama, _Gravitational-wave constraints on the abundance of primordial black holes_ , _Progress of theoretical physics_ 123 (2010) 867–886. * [34] C. Unal, E. D. Kovetz and S. P. Patil, _Multi-messenger Probes of Inflationary Fluctuations and Primordial Black Holes_ , 2008.11184. * [35] M. Kawasaki, N. Kitajima and T. T. Yanagida, _Primordial black hole formation from an axionlike curvaton model_ , _Phys. Rev._ D87 (2013) 063519, [1207.2550]. * [36] M. Kawasaki, N. Kitajima and S. Yokoyama, _Gravitational waves from a curvaton model with blue spectrum_ , _JCAP_ 1308 (2013) 042, [1305.4464]. * [37] K. Ando, K. Inomata, M. Kawasaki, K. Mukaida and T. T. Yanagida, _Primordial Black Holes for the LIGO Events in the Axion-like Curvaton Model_ , 1711.08956. * [38] K. Ando, M. Kawasaki and H. Nakatsuka, _Formation of primordial black holes in an axionlike curvaton model_ , _Phys. Rev. D_ 98 (2018) 083508, [1805.07757]. * [39] G. ’t Hooft, _Naturalness, chiral symmetry, and spontaneous chiral symmetry breaking_ , _NATO Sci. Ser. B_ 59 (1980) 135–157. * [40] C. Unal, _Imprints of Primordial Non-Gaussianity on Gravitational Wave Spectrum_ , _Phys. Rev. D_ 99 (2019) 041301, [1811.09151]. * [41] R.-g. Cai, S. Pi and M. Sasaki, _Gravitational Waves Induced by non-Gaussian Scalar Perturbations_ , _Phys. Rev. Lett._ 122 (2019) 201101, [1810.11000]. * [42] J. Chluba, A. L. Erickcek and I. Ben-Dayan, _Probing the inflaton: Small-scale power spectrum constraints from measurements of the cosmic microwave background energy spectrum_ , _The Astrophysical Journal_ 758 (2012) 76. * [43] D. J. Fixsen, E. S. Cheng, J. M. Gales, J. C. Mather, R. A. Shafer and E. L. Wright, _The Cosmic Microwave Background spectrum from the full COBE FIRAS data set_ , _Astrophys. J._ 473 (1996) 576, [astro-ph/9605054]. * [44] K. Inomata, M. Kawasaki and Y. Tada, _Revisiting constraints on small scale perturbations from big-bang nucleosynthesis_ , _Phys. Rev._ D94 (2016) 043527, [1605.04646]. * [45] M. Shibata and M. Sasaki, _Black hole formation in the Friedmann universe: Formulation and computation in numerical relativity_ , _Phys. Rev. D_ 60 (1999) 084002, [gr-qc/9905064]. * [46] M. Kawasaki and H. Nakatsuka, _Effect of nonlinearity between density and curvature perturbations on the primordial black hole formation_ , _Phys. Rev. D_ 99 (2019) 123501, [1903.02994]. * [47] K. Ando, K. Inomata and M. Kawasaki, _Primordial black holes and uncertainties on choice of window function_ , 1802.06393. * [48] S. Young and C. T. Byrnes, _Primordial black holes in non-Gaussian regimes_ , _JCAP_ 1308 (2013) 052, [1307.4995]. * [49] C. T. Byrnes, E. J. Copeland and A. M. Green, _Primordial black holes as a tool for constraining non-Gaussianity_ , _Phys. Rev._ D86 (2012) 043512, [1206.4188]. * [50] S. Young, D. Regan and C. T. Byrnes, _Influence of large local and non-local bispectra on primordial black hole abundance_ , _JCAP_ 1602 (2016) 029, [1512.07224]. * [51] Macho collaboration, R. A. Allsman et al., _MACHO project limits on black hole dark matter in the 1-30 solar mass range_ , _Astrophys. J._ 550 (2001) L169, [astro-ph/0011506]. * [52] EROS-2 collaboration, P. Tisserand et al., _Limits on the Macho Content of the Galactic Halo from the EROS-2 Survey of the Magellanic Clouds_ , _Astron. Astrophys._ 469 (2007) 387–404, [astro-ph/0607207]. * [53] L. Wyrzykowski et al., _The OGLE View of Microlensing towards the Magellanic Clouds. IV. OGLE-III SMC Data and Final Conclusions on MACHOs_ , _Mon. Not. Roy. Astron. Soc._ 416 (2011) 2949, [1106.2925]. * [54] P. D. Serpico, V. Poulin, D. Inman and K. Kohri, _Cosmic microwave background bounds on primordial black holes including dark matter halo accretion_ , _Phys. Rev. Res._ 2 (2020) 023204, [2002.10771]. * [55] J. Garcia-Bellido, M. Peloso and C. Unal, _Gravitational Wave signatures of inflationary models from Primordial Black Hole Dark Matter_ , _JCAP_ 09 (2017) 013, [1707.02441]. * [56] L. Lentati et al., _European Pulsar Timing Array Limits On An Isotropic Stochastic Gravitational-Wave Background_ , _Mon. Not. Roy. Astron. Soc._ 453 (2015) 2576–2598, [1504.03692]. * [57] R. M. Shannon et al., _Gravitational waves from binary supermassive black holes missing in pulsar observations_ , _Science_ 349 (2015) 1522–1525, [1509.07320]. * [58] C. J. Moore, R. H. Cole and C. P. L. Berry, _Gravitational-wave sensitivity curves_ , _Class. Quant. Grav._ 32 (2015) 015014, [1408.0740].
# Modeling opinion leader’s role in the diffusion of innovation _Internship report originally written in June 2018 by intern N. Vodopivec under the supervision of C. Adam and J.-P. Chanteau_ Nataša Vodopivec Univ Grenoble-Alpes Grenoble INP Carole Adam Univ Grenoble-Alpes Grenoble Informatics Lab Jean-Pierre Chanteau Univ Grenoble-Alpes Centre de Recherche en Economie de Grenoble ###### Abstract The diffusion of innovations is an important topic for the consumer markets. Early research focused on how innovations spread on the level of the whole society. To get closer to the real world scenarios agent based models (ABM) started focusing on individual-level agents. In our work we will translate an existing ABM that investigates the role of opinion leaders in the process of diffusion of innovations to a new, more expressive platform designed for agent based modeling. We will do it to show that taking advantage of new features of the chosen platform should be encouraged when making models in the field of social sciences in the future, because it can be beneficial for the explanatory power of simulation results. ## 1 Introduction Diffusion refers to the process by which an innovation is adopted over time by members of a social system. An innovation commonly refers to a new technology, but it can be understood more broadly as a spread of ideas and practices Kiesling et al. (2012). The question whether a certain innovation will diffuse in society successfully or not has always been of important nature at the market level and has gained interest of many researchers since a number of pioneering works appeared in the 1960s. #### From the marketing perspective, it is of great importance to understand how information starting from mass media and traveling through word-of-mouth WoM affects adoption decisions of customers and consequently the diffusion of a new product van Eck et al. (2011). Mass media takes the role of an external influence to a society and WoM the role of an internal influence within the society. Traditionally, models were based on macro level looking at the society as a whole. Most such aggregate models stem from the model introduced by Bass Bass (1969), which takes the structure of a basic epidemic model where diffusion of innovation is seen as a contagious process driven by external and internal influences. This model assumes that the market is homogeneous, which means that all customers have the same characteristics. It further assumes that each consumer is connected with all other consumers and can thus influence all others. From these two assumptions it follows that the probability of adopting is linearly related to the number of past adopters. These assumptions are limitations of aggregate models as they ignore that in the real world consumers are individuals and as such heterogeneous and a part of complex social structures. #### To try to overcome these limitations, agent-based modeling Macal and North (2005) has been increasingly adopted in the recent times. Agent based modeling takes a different approach to diffusion of innovations, because it looks at the society from the micro level. Here, an observed entity is not a society as a whole, but an individual, represented as an agent. Customers’ heterogeneity, their social interactions and their decision making process can be modeled explicitly Kiesling et al. (2012). When simulated, macro level observations of network changes emerge from the micro level interactions between the individuals. Most agent based models of innovation diffusion have a similar structure and comprise of the following elements Jensen and Chappin (2017): 1. 1. Consumer agents define the individual entities that can adopt an innovation. These can be individual persons, households, or groups of households. They are heterogeneous. 2. 2. Social structure is a description of connections between singular agents, dividing them in different consumer groups. 3. 3. Decision making processes are the key actions of consumer agents in any social model, by which agents decide to adopt or reject the innovation. 4. 4. Social influence between agents often affects decision making processes and is commonly modeled as a social network graph. Models vary in the range at which social influence is exceeded. This can be influence from direct peers, from the respective social group or the entire population of agents. All these ranges of influence can be modeled as a social network graph. #### We are interested in modeling and simulating how innovation, both in the sense of the ideas, behaviours and in the sense of the products, spreads in the population. We have chosen to implement the model on a GAMA platform, which is seen as a current state-of-the-art agent-based modeling language and as an improved successor of the NetLogo platform. Figure 1 presents a screenshot of the implemented simulator, showing the social network with opinion leaders (in pink) and adopters of the innovation (in green). Figure 1: Screenshot of the GAMA simulator The main challenge when modeling social models is the verification of the final model. The verification can be done either using strong theoretical support or using obtained empirical data. However, due to the scope of our project it was difficult to obtain either. This is the reason why we have chosen to use an existing NetLogo model to rewrite and improve in GAMA, because this way we would be able to validate our model against it. A study by van Eck van Eck et al. (2011) (further use: reference study) was picked as it not only models the diffusion of innovations, but additionally investigates the role of opinion leaders in the process, which is another interesting phenomenon. Aside from agents being heterogeneous, they are further divided into two groups, namely the influentials or opinion leaders (OL) and followers or non-leaders (NL). #### Goldenberg et al Goldenberg et al. (2009) determine influentials by three factors: connectivity, knowledge and personality characteristics. Opinion leaders are a type of influential customers that have all of the characteristics of the influentials represented as central positions in the network (which means high connectivity), market knowledge (not necessarily about a specific product but about markets in general) and innovative behaviour. The reference study uses four critical assumptions about opinion leaders, which are later successfully checked by an empirical study: (1) OL have more contacts, (2) OL possess different characteristics, (3) OL exert different types of influence and (4) OL are among earlier adopters. Two important characteristics of opinion leaders are their innovativeness and their interpersonal influence. Regarding the degree of innovativeness, it means that opinion leaders have more experience and expertise with the product category than the other consumers and that they have been exposed to more information Lyons and Henderson (2005). Two main types of interpersonal influence exist: * • Informational influence is the tendency to accept information from others and believe it. Opinion leaders influence other consumers by giving them advice about a product. * • Normative influence stems from the people’s tendency to follow a certain norm; to adopt a product in order to be approved by other consumers. Normative influence can also be referred to as a social pressure. Reference study assumes that opinion leaders play an important role in both diffusion of information about products (informational influence) and the diffusion of products themselves (i.e. more product adoptions result in normative influence) van Eck et al. (2011). Therefore the influence of the opinion leaders on the speed of diffusion of both information and product, and on maximum adoption percentage in the process of diffusion of innovation, is investigated. #### The focus of this study is to investigate the speed of information diffusion, the speed of product diffusion, and the maximum adoption percentage of the product. The article is structured as follows: in Section 2 we describe the hypotheses that the reference study has set up and verified; Section 3 introduces the model with its agents, parameters, and social network; in Section 4 we present the experiments settings and discuss our simulation results; and in Section 5 we address the conclusions and suggestions for further work. ## 2 Hypotheses While investigating the role of opinion leaders in the innovation diffusion process, the impact of each of its three characteristics (innovative behaviour, normative influence, market knowledge) is looked at and thus more hypotheses are set up. We have chosen to validate our model against the following hypotheses put forward and successfully proven in the reference study van Eck et al. (2011): ##### H1a: ”The more innovative behaviour of the opinion leader results in a higher adoption percentage.” ##### H1b: ”If the weight of normative influence becomes more important to followers, the increase in the adoption percentage caused by the more innovative behavior of opinion leaders increases.” ##### H2a: ”Opinion leaders are less sensitive to normative influence than are followers.” ##### H2b: ”If opinion leaders are less sensitive to normative influence, adoption percentages increase.” ##### H2c: The less sensitive opinion leaders are to normative influence, the more the adoption percentages increase. ##### H3a: ”Opinion leaders are better at judging product quality, which results in a higher speed of information diffusion.” ##### H3b: ”Opinion leaders are better at judging product quality, which results in a higher speed of product diffusion.” ## 3 Model In this chapter we describe in more detail how our model was built. ### 3.1 Network Bohlman et al. Bohlman et al. (2010) indicate that specific network topologies in agent based modeling strongly influence the process of innovation diffusion: they affect the likelihood that diffusion spreads and the speed of adoption. This is because the network topology specifies the location and the number of links of innovators. A scale-free network structure proposed by Barabasi and Albert (1999) is used, because it stems from many empirical researches and is confirmed to imitate real world societies where some agents serve as hubs, meaning their number of connections greatly exceeds the average and they have central positions in the network. ### 3.2 Agents Each agent is described by the following attributes: * • Opinion leader tells whether an agent is an opinion leader or a non-leader. * • Quality threshold is a value given randomly uniformly to each agent before the beginning of a simulation. Its values are uniformly distributed in the range U(0,1). * • Known quality describes what the agent currently thinks of the product quality. This value is set dynamically when the agent gets aware of the product or adopts it, as is further explained in Section 4.1. * • Utility threshold is a value given randomly uniformly to each agent before the beginning of a simulation. OL and NL have different ranges from where this value can be taken. For NL it is U(0,1) and for OL it is U(0, max), where the maximum value is defined by a parameter of the experiment. * • Awareness tells whether an agent is aware of the product or not. * • Adopted tells whether an agent has adopted the product or not. * • Weight of normative influence is different for OL and NL and values are set dynamically as a normal distribution where average value and standard deviation are set as parameters of the simulation. The uniform distribution of the values of utility thresholds and of quality thresholds for individual agents makes the population heterogeneous. #### An agent’s decision to adopt is based on its utility threshold. The agent’s utility is calculated at each iteration of the simulation, and once it passes the agent’s utility threshold the agent adopts the innovation. The utility function consists of a weighted sum of the individual preference and the social influence. First represents informational influence and describes the agent’s opinion on the product quality, and second represents normative influence and takes into account the number of neighbouring agents that have already adopted the product. When the weight of the social influence of a certain agent is low the agent is very individualistic and is consequently hardly influenced by neighbours. On the contrary, high weight value means that the agent is very socially susceptible van Eck et al. (2011). ### 3.3 Parameters The model contains several parameters, which describe the influence of opinion leaders in various market settings van Eck et al. (2011). Some parameters are fixed for all experiments and others vary experimentally. The group of fixed parameters and their values, derived from the model by Delre et al. Delre et al. (2007), are presented in Table 1. The product quality is set to 0.5, meaning that if the agents base their decisions to adopt a product purely on their individual preferences, approximately 50% will never adopt. The mass media coefficient was set from prior studies. It represents a strong mass media support because many of the agents in the empirical study were reached by mass media (i.e. one percent of population is reached in each step). Variable | Parameter | Value ---|---|--- Product quality | q | 0.5 Mass media coefficient | m_m | 0.01 Number of agents | nb_agents | 500 Table 1: Settings for global parameters, that were fixed in current experiments Variable | Parameter | Value ---|---|--- Max utility threshold of OL | max | 0.8 Average normative influence, OL | avg_ni_ol | 0.51 Standard deviation for normative influence, OL | dev_ni_ol | 0.2 Average normative influence, NL | avg_ni_nl | 0.6 Standard deviation for normative influence, NL | dev_ni_nl | 0.2 OL judges product better | NA | Yes Table 2: Settings for base model parameters The varied parameters are changed one at a time per experiment to test the separate hypotheses. The parameters and their values, derived from empirical study conducted by reference study are presented in Table 2. First a base experiment with these values was run so that later hypotheses could be tested realistically. The innovativeness of opinion leaders is implemented as smaller possible values of it’s utility threshold with regard to that of the followers (the utility threshold of the followers has a uniform distribution in the range U(0, 1.0), for OL it’s in the range U(0, 0.8)), which makes them approximately 20% more likely to adopt the product. The difference is not big as OL are trying to avoid being too innovative, because if they adopted a product that turned out to be unsuccessful, they would loose followers. As observed in the empirical study, the weight of normative influence of opinion leaders holds a lower value ($\beta_{OL}=$ 0.51) than that of the followers ($\beta_{NL}=$ 0.6) as they care less about the social pressure. The weights of normative and informative influences sum up to 1, so the weight of informative influence is 1 - $\beta$. The model can be run either with opinion leaders in the network or without them. This was important to be able to see whether the diffusion of the innovation indeed spreads faster in the networks where innovative opinion leaders are present. | Innovativeness of OL | Weight of normative influence OL | Weight of normative influence NL | ---|---|---|---|--- Model (hypothesis tested with model) | $U_{i,min}$ | $\beta_{i,OL}$ | $\beta_{i,NL}$ | Quality of the product judgment (OL) Base Model 1 | U(0, 0.8) | N(0.51, 0.2) | N(0.6, 0.2) | Yes Model 2 (H1a) | U(0, 1) | N(0.51, 0.2) | N(0.6, 0.2) | Yes Model 3 (H1b) | U(0, 0.8) | N(0.51, 0.2) | N(0.8, 0.2) | Yes NA (H2a) | NA | NA | NA | NA Model 4 (H2b) | U(0, 0.8) | N(0.57, 0.2) | N(0.57, 0.2) | Yes Model 5 (H2c) | U(0, 0.8) | N(0.2, 0.2) | N(0.6, 0.2) | Yes Model 6 (H3a, H3b) | U(0, 0.8) | N(0.51, 0.2) | N(0.6, 0.2) | No Table 3: Parameter settings for hypotheses (adapted from van Eck et al. (2011)) ## 4 Experiments and results In this chapter we first present the experiment settings, then discuss the results and finally do a comparison between NetLogo and GAMA platforms. ### 4.1 Experiment settings A model was created for each separate hypothesis. The values of the varied parameters used for each model are shown in Table 3. Each model was run in a separate experiment that consisted of 25 time steps, which was enough for the maximum adoption percentage to be reached. To collect results for statistics each experiment was run with the same settings 60 times. We realize that 60 is a low number of repetitions for completely adequate statistics, but we faced a problem of the GAMA platform freezing due to too big memory consumption while trying to run it in batch mode, where more than one experiment is run one after another automatically. We did not anticipate this to happen as the calculations were very fast when running 500 consecutive experiments in NetLogo and GAMA is seen as it’s improved successor. Thus, we were reduced to having to run each experiment manually which proved to be quite time consuming so we limited the number of runs to 60 and might do more tests to calibrate the results if needed in the future. Each time step further consisted of three phases: mass media, WoM and adoption. In the beginning of the experiment no agents are aware of the product or have adopted it. Then mass media informs a predefined percentage (in our case 1%) of the population about it. In this step, the better market knowledge of the opinion leaders is implemented as such: because opinion leaders are able to make good product judgment they will have learned of a real product quality from mass media and their quality judgment will become equal to it (q = 0.5, Table 1). On contrary, followers are not able to make this judgment so they become aware of the product but their perceived product quality gets a random value. The followers are able to learn about the real product quality only by WoM from trusted sources, that is from opinion leaders and agents who have already adopted the product. In the word of mouth stage, agents talk with their neighbors and may learn about the real product quality if their neighbors are certain about it. In the adoption stage, agents can decide to adopt the product if they are aware about it and if the current value of utility function exceeds their utility threshold. ### 4.2 Results and Validation | | Adoption percentage --- (standard deviation) | Speed of information diffusion --- Average number of steps (standard deviation) | Speed of product diffusion --- Average number of steps (standard deviation) | | Reference --- study | Our --- study | Reference --- study | Our --- study | Reference --- study | Our --- study Base Model 1 - no OL | 0.398 (0.05) | 0.401 (0.05) | 3.64 (1.5) | 4.43 (1.47) | 6.27 (2.0) | 5.78 (1.51) Base Model 1 | 0.491 (0.05) | 0.454 (0.05) | 1.75 (1.2) | 1.96 (0.43) | 4.94 (1.2) | 2.80 (0.78) Model 2 (H1a) | 0.405 (0.04) | 0.398 (0.05) | | | | Model 3 (H1b) | 0.458 (0.06) | 0.395 (0.06) | | | | NA (H2a) | | | | | | Model 4 (H2b) | 0.480 (0.05) | 0.455 (0.06) | | | | Model 5 (H2c) | 0.515 (0.04) | 0.488 (0.05) | | | | Model 6 (H3a, H3b) | | | 4.73 (2.22) | 4.11 (1.70) | 7.76 (2.21) | 5.64 (1.65) Table 4: Results of tests of each hypothesis Before looking at the models testing the hypotheses we had to make sure that our model confirms the base assumption. It claims that in networks that include opinion leaders, higher speed of both the information and product diffusion as well as greater adoption percentage are achieved, than in the networks without them. Thus, we ran the base model in two experiments, once with opinion leaders and once without them. The weight of normative influence ($\beta_{i}$) in the comparison model with no opinion leaders is 0.75 (obtained from reference study). The average values of the results and their standard deviations for these two tests as well as for the rest of the tests can be found in Table 4. We can see that in the model with opinion leaders the information diffuses faster than in the model without the opinion leaders, in the first it takes 1.96 steps compared with 4.43 in the other. The same case happens for the speed of product diffusion, in the model with OL it takes 2.80 steps compared to 5.78 steps in the model without the OL, which means that the product diffuses faster in the model with OL. Thirdly, the value of the average adoption percentage is higher in the model with OL (0.45 in the model with OL and 0.40 in the model without), which also confirms our assumptions. Therefore, the base model successfully proves that the opinion leaders product higher speeds of information and product diffusion and higher adoption percentage. Table 4 shows the obtained averaged results from the reference study and from our model for each of the hypotheses models. Each hypothesis was run in a different experiment on it’s own model, for which the values are presented in the Table 3, except for the hypothesis H2a which the reference study validated by empirical study. When comparing the results we can see that even though values are a bit different, their proportions stay the same, meaning that our model was successfully validated against the reference NetLogo model and as such that the same as in the NetLogo model the hypotheses Ha1, H2a (empirical study), H2b, H2c, H3a and H3b got supported while the hypothesis H1b did not get supported. ### 4.3 Comparing the platforms The differences might be partially attributed to the smaller sample sizes that we use to average the results, but we think they’re mostly the reason of a different execution flow in the GAMA platform. It is here that GAMA platform introduces a difference that we find important when making social models. The execution flow of NetLogo for the model of diffusion of innovation is sequential and iterative. For each of the 25 steps of the simulation the three stages (mass-media, word of mouth, adoption) are executed one after another, where first one has to complete for all of the agents before the next stage can commence. Inside each stage, the agents execute the actions linked to it iteratively in a loop, the agent 1 does it first and agent 500 the last. The agents inherently act as small blocks of non-connected code and the order of their execution can never be different as the loop over the agents that calls each of them determines it. On the other hand, on GAMA platform each agent acts as it’s own entity with it’s own behaviours. During the simulation of the 25 steps, the only role of the world agent that stands above all other agents (on GAMA platform the world agent acts similar than a main function in many programming languages) is to schedule them, i.e. gives them an opportunity to act. While the agents still do not all run at the same time in parallel, they are not connected with actions of the other agents. When each agent gets it’s turn it runs it’s behaviours, which are in turn mass-media, WoM and adoption. So the prime difference is that on GAMA platform the main program only calls the agents and after that is has no control over how they execute, they act as individual entities. However, in our model the world agent mostly still calls the consumer agents iteratively, starting with agent 1 and finishing with agent 500, which is still not very representative of the real world because the order of the agents is the same at each simulation step. There exists a solution to this problem which is discussed in section 5. ## 5 Conclusions and Further research We have successfully established and validated a model of diffusion of innovation nn the state-of-the-art GAMA platform that is designed for agent- based modeling. We think that it is an important step to take towards the more realistic modeling of social interactions. However, as mentioned before in Section 4 the agents still get executed in the same order in each step of the simulation. This could lead to unrealistic simulations. We would like to highlight one example of this problem, namely discuss the execution of the Word of Mouth stage. When in this stage, the agent talks to all of its neighbours, and if any of them know of the true product quality, then the agent becomes aware of it by WoM. Now imagine the first simulation step when after the mass-media stage at most 1% of the population has become aware of the product. During the WoM stage agents will be called upon iteratively and each of them will have larger probability that some of its previously non- aware neighbours have now become aware and can thus share their knowledge about the product. Consequently, in each of the 25 steps of the simulation, agent number 1 will always have lesser probability to become aware by WoM than agent number 500. To solve this issue GAMA platform provides an option of calling agents in a random (shuffled) order. In NetLogo such option could be implemented manually, but would be hard to achieve. As a future work we think that adding this property and observing the obtained results could be a good idea. The results might stay the same, but the micro structure of the model would become closer to the real world social models. Another promising option of further research would be the extension of the current agents to BDI agents. Agent based modeling is already a step forward from the old aggregate models where humans were modeled as equal homogeneous entities. However, when handling ABMs in the field of social sciences, human agents can be further improved to become more human-like by giving them personality traits. These agents are called belief, desire and intention (BDI) agents. The model allows to use more complex and descriptive agent models to represent humans. It attempts to capture common understanding of how humans reason through: beliefs which represent the individual’s knowledge about the environment and about their own internal state; desires or more specifically goals (non-conflicting desires which the individual has decided they want to achieve); and intentions which are the set of plans or sequence of actions which the individual intends to follow in order to achieve their goals Adam and Gaudou (2016). Two other important functionalities a BDI system must have are a rational process by which an agent decides which intentions to follow depending on the current circumstances, and the level of commitment to the set of intentions to achieve a long-term goal. We think that BDI agents are important to give higher descriptive value on results of social studies, which a diffusion of innovation certainly is. They give more information on how agents behave and a deeper insight on how innovation diffuses in the population. As a future work, we will upgrade this model by expanding its agents to BDI agents, now that the model has been translated to GAMA, which allows BDI architecture. We will then add different human factors to these agents and observe how they affect the spread of the diffusion of an innovation and it’s speed and whether the results will stay in line with the original model. ## References * Adam and Gaudou [2016] Carole Adam and Benoit Gaudou. BDI agents in social simulations: a survey. The Knowledge Engineering Review, 31(3):207–238, Cambridge University Press, 2016. * Bass [1969] F. M Bass. A new product growth for model consumer durables. Management Science, 15(5):215–227. * Bohlman et al. [2010] J. Bohlman, R. Calantone and M. Zhao. The effects of market network heterogeneity on innovation diffusion: An agent-based modeling approach. Journal of Product Innovation Management, 27(5):741–60. * Delre et al. [2010] Sebastiano A. Delre. Will it spread or not?: The effects of social influences and network topology on innovation diffusion. The journal of product innovation management : an international publication of the Product Development & Management Association, 27(2), 2010. * Delre et al. [2007] Sebastiano A. Delre, W. Jager, T. H. A Bijmolt and M. A. Janssen. Targeting and timing promotional activities: An agent-based model for the takeoff of new products. Journal of Business Research, 60(8):826–35. * van Eck et al. [2011] Peter S. van Eck, Wander Jager and Peter S. H. Leeflang. Opinion leaders’ role in innovation diffusion : a simulation study. The journal of product innovation management : an international publication of the Product Development & Management Association., 28.2011(2):187–203, Oxford, Blackwell Publishing, 2011. * Goldenberg et al. [2009] J. Goldenberg, S. Han, D. R. Lehmann and J. W. Wong. The role of hubs in the adoption process. Journal of Marketing, 73(2):1–13. * Jensen and Chappin [2017] Thorben Jensen and Emile J.L. Chappin. Automating agent-based modeling: Data-driven generation and application of innovation diffusion models. Environmental Modelling & Software, 92:261–268, 2017. * Kiesling et al. [2012] Elmar Kiesling, Markus Günther, Christian Stummer and Lea M. Wakolbinger. Agent-based simulation of innovation diffusion: A review. Central European Journal of Operations Research, 20:183–230, 2012. * Laciana et al. [2017] Carlos E. Laciana, Gustavo Preyra and Santiago L. Rovere. Size invariance sector for an agent-based innovation diffusion model. ARXIV, 1706.03859, 2017. * Lyons and Henderson [2005] B. Lyons and K. Henderson. Opinion leadership in a computer-mediated environment. Journal of Consumer Behavior, 4(5):319–29. * Macal and North [2005] C. M. Macal and M. J. North. Tutorial on agent-based modeling and simulation. In 37th Winter Simulation Conference. Introductory Tutorials: Agent-Based Modeling, 2–15. * Mills and Schleich [2012] Bradford Mills and Joachim Schleich. Residential Energy-Efficient Technology Adoption, Energy Conservation, Knowledge, and Attitudes: An Analysis of European Countries. Energy Policy, 49, 2012. * Zhang and Vorobeychik [2016] Haifeng Zhang and Yevgeniy Vorobeychik. Empirically Grounded Agent-Based Models of Innovation Diffusion: A Critical Review. CoRR, 1608.08517, 2016.
On event-by-event pseudorapidity fluctuations in relativistic nuclear interactions M. Mohisin Khan1∗, Danish F. Meer1, Tahir Hussain2, N. Ahmad3 1\. Department of Applied Physics, ZHCET, Aligarh Muslim University, Aligarh, India 2\. Applied Sciences and Humanities Section, University Polytechnic, Aligarh Muslim University, Aligarh, India 3\. Department of Physics, Aligarh Muslim University, Aligarh, India <EMAIL_ADDRESS> Abstract Present study is an attempt to have a detailed look into event-by- event(e-by-e) pseudorapidity fluctuations of the relativistic charged particles produced in 28Si-nucleus interactions at incident momenta 4.5A and 14.5A GeV/c. The method used in the present study makes use of a kinematic variable which is derived in terms of the average pseudo-rapidity and the total number of particles produced in a single event. The multiplicity and pseudorapidity dependence of these fluctuations have also been studied. The results obtained for the experimental data are compared with HIJING simulation. Keywords: event-by-event pseudo-rapidity fluctuations, correlation, relativistic nuclear collisions. Introduction Relativistic nuclear collisions are the most fascinating and important tools to produce matter under extreme conditions of temperature and density. The key point to study and understand the behaviour of this produced matter is the copious production of secondary particles in these collisions.The global observable such as multiplicity and pseudo-rapidity of produced particle play an important role in understanding the particle production process in the relativistic hadron-hadron, hadron-nucleus and nucleus-nucleus collisions. The current interest in such studies are mainly to understand the characteristics of quark-gluon plasma (QGP) and the scenario of phase transition from QGP to the normal hadronic phase. Fluctuations in the values of global observable have always been considered as one of the possible signature of QGP formation1. Various nearly conclusive studies2-5 regarding QGP formation and its signatures have been made using the data on three experimental energy regime from the SPS, RHIC and LHC. As the matter produced in high energy heavy-ion collisions is a short live state and the hadronization after the collisions is very fast one has to rely on the observations made on the characteristics of the produced particles. The study of the particles coming out of the interaction region may provide important information regarding the underlying dynamics of collision process and the multi-particle production. Fluctuations in general and the event-by- event fluctuations of observable in particular are envisaged to give vital information about the phase transition5-9. The process of thermalization along-with the statistical behaviour of the produced particles can be understood by studying the fluctuation in particle multiplicity and momentum distribution10-15. Reference 6 stressed that the charge fluctuations may be an evidence of QGP formation. Many such studies about the fluctuations has been carried out. However, study of critical point of QGP phase transition can be well study using the concept of e-by-e fluctuations because in this case the fluctuations are predicted to be large enough 14-17. A study of each event produced in relativistic nuclear collision may reveal new physical phenomena occurring in some rare events for which conditions might have been created in these collisions. Nuclear collisions at high energies produce a large number of particles, the analysis of single event with large multiplicity can shed light on some different physics than the study of averages over a large events. Predictions have been made about the occurrence of critical density fluctuations in the vicinity of the phase transition and its manifestation as e e-by-e fluctuation of different physical observable 18. The e-by-e analysis may offer a possibility of observing a phase transition directly if the particle emitting source is hydro-chemical composition. The NA49 collaboration10 observed the fluctuations of transverse momentum and koans to pins ratio in central Pb-Pb collisions at 158A GeV. A. Bialas and V. Koch 9 and Belkacem et al19 reported the moments of e-by-e fluctuations are very nearly related to the correlation function. The ALICE collaboration20 has measured the e-by-e fluctuation in mean transverse momentum in p-p and Pb-Pb collisions at LHC. A number of papers are available in literature on e-by-e fluctuation analysis of different observable but a very few papers are there on e-by-e pseudo-rapidity fluctuations. The first such study was carried out by the KLM collaboration for 158A GeV Pb-AgBr interactions21, Recently S. Bhattacharya et al.18 and Gopa Bhoumic et al.22 has carried out the e-by-e pseudo-rapidity fluctuations analyses on various emulsions data having different projectiles and targets at 4.1A GeV, 4.5A GeV, 60A GeV and 200A GeV. In the present study we have carried out e-by-e fluctuations analysis for the data at 4.5A and 14.5A GeV 28Si-AgBr interactions for the experimental and HIJING simulated data. Following sections of this paper are devoted to the details of the data, analysis methods, results and discussion and the observations made on the basis of obtained results. Experimental details of the data The present analysis has been carried out on the experimental and simulated data. For the experimental data, two random samples consisting 555 events of 14.5A GeV/c 28Si-nucleus interactions and 530 events of 4.5A GeV/c 28Si- nucleus interactions with $N_{s}\geq$10 have been used where Ns represents the number of charged particles produced in an event with relative velocity ($\beta=v/c>0.7$). The emission angles of the relativistic charged particles were measured and their pseudo rapidities ($\eta=-ln(tan(\theta/2)$) are determined. All other details about the data may be found elsewhere23,24. Furthermore, for comparing the experimental results with the corresponding values obtained for the events generated by Monte Carlo code HIJING-1.3325 event generator, a similar sample of events was simulated. Method of analysis M. Gazdzicki and S. Mrowczynski7 proposed an excellent method to measure the fluctuations of any global kinematic variable. It is worth-mentioning that the second moment of distributions of global kinematic variables (multiplicity, rapidity, transverse momentum etc.) for individual events or all the events taken together may shed light on the extent of thermalization and randomization feature of high energy nuclear collisions. The basic idea used in this method is the fact that the correlated production of particles in each elementary interaction leads to large e-by-e fluctuations and these fluctuations in high energy nuclear collisions are believed to originate due to trivial variation in impact parameter of the interaction. It may also be aroused due to some statistical reasons or due to some dynamical reason of the underlying processes prevailing at the instant of the collisions. The method proposed here7 automatically filters the trivial contributions and provides a way to determine the remaining part contributing to the fluctuations. For this, a variable, $\Phi$ which is believed to be a measure of fluctuation is defined whose nonzero values points towards the correlation and fluctuation and a vanishing value of $\Phi$ points towards the independent particle emission(random emission) from a single source. The detailed procedure of calculating this variable is described below. As the global kinematic variable used in the present analysis to study the e-by-e fluctuation is the pseudo-rapidity,$\eta$, of the emitted particles we first define a single particle variable z in terms of $\eta$ as: $z=\eta-\bar{\eta},$ (1) where $\bar{\eta}$ represents the mean value of single particle inclusive pseudo-rapidity distribution that can be expressed as $\bar{\eta}=\frac{1}{N_{total}}\sum_{m=1}^{N_{evt}}\sum_{i=1}^{N_{m}}\eta_{m},$ (2) where Nm is the multiplicity of mth event. The second summation over i in the above equation runs over all the particle produced in the mth event and the first summation is performed over all the events Nevt in the sample. Ntotal in the denominator is the total number of particles produced in all the events. Further, a multi-particle analogue of z, Zk, is defined as $Z_{k}=\sum_{i=1}^{N_{k}}\eta_{i}-\bar{\eta}.$ (3) Finally, the measure of fluctuation parameter, $\Phi$ is defined as $\Phi=\sqrt{\frac{<Z^{2}>}{<N_{total}>}}-\sqrt{\bar{z^{2}}},$ (4) where the $<Z^{2}>$ and $<N_{total}>$ represents the event averaged of the variables therein and the $\sqrt{\bar{z^{2}}}$ is the square root values of the second moment of inclusive z distribution. As stated7, $\Phi$ vanishes when there is no correlation among the produced particles and its non vanishing values are a measure of correlations and fluctuation present in the system. This method has been extensively used with success to analyze many experimental data18,21 and to verify various aspects theoretically26,27. In the present analysis we have attempted to study e-by-e fluctuation in 4.5A and 14.5A GeV/c 28Si-nucleus interactions. As the present analysis is meant to study the e-by-e $\eta$ fluctuations, the variable z is defined as Results and discussion First of all the values of $\Phi$s are calculated for different groups of events selected on the basis of the average multiplicity of the relativistic charged particles, $<N_{s}>$. This calculation has been made for both the experimental and simulated data. These values along with statistical errors are tabulated in Table 1. The the different groups are selected in such a way to ensure that the average multiplicity of the group is greater than the average multiplicity of the sample of the data. Table 1: Calculated values of $\Phi$s for different multiplicity classes for the experimental and HIJING simulated data. Interactions | Multiplicity selection | Experimental | HIJING ---|---|---|--- | | $<N_{s}>$ | $\Phi$ | $<N_{s}>$ | $\Phi$ 4.5A GeV/c 28Si-nucleus | Ns $\geq$ 10 | 15.33 | 5.402 $\pm$ 0.073 | 17.85 | 4.491 $\pm$ 0.088 Ns $\geq$ 20 | 31.54 | 5.010 $\pm$ 0.071 | 33.55 | 3.923 $\pm$ 0.080 Ns $\geq$ 30 | 42.14 | 4.410 $\pm$ 0.082 | 39.74 | 3.108 $\pm$ 0.082 Ns $\geq$ 40 | 52.76 | 4.106 $\pm$ 0.088 | 51.23 | 2.213 $\pm$ 0.089 Ns $\geq$ 50 | 64.75 | 3.710 $\pm$ 0.069 | 63.25 | 1.984 $\pm$ 0.094 14.5A GeV/c 28Si-nucleus | Ns $\geq$ 10 | 24.98 | 5.281 $\pm$ 0.074 | 19.71 | 3.874 $\pm$ 0.089 Ns $\geq$ 20 | 33.99 | 5.010 $\pm$ 0.077 | 36.25 | 3.093 $\pm$ 0.101 Ns $\geq$ 30 | 47.99 | 4.740 $\pm$ 0.088 | 43.55 | 2.823 $\pm$ 0.113 Ns $\geq$ 40 | 51.86 | 3.901 $\pm$ 0.097 | 50.55 | 2.123 $\pm$ 0.134 Ns $\geq$ 50 | 64.66 | 2.558 $\pm$ 1.066 | 61.58 | 1.674 $\pm$ 0.149 Tabulated above table are the values of $\Phi$ and the averaged multiplicity of relativistic charged particles, $<N_{s}>$ for various multiplicity classes for the experimental and simulated events. It is observed from the table that the values of $\Phi$ are non-zero for all the multiplicity classes considered in the present study and this observation is same for both the experimental and simulated data at the two incident energies for 28Si-nucleus interactions. These non zero values of $\Phi$ supports the occurrence of dynamical fluctuations in the $\eta$-variable and the presence of correlation during particle production process in high energy nucleus-nucleus collisions. It is also observed from the table that the $\Phi$, which is considered to be the strength of correlation and fluctuation, shows a decreasing trend with increasing $<N_{s}>$. This dependence of $\Phi$ on $<N_{s}>$ is shown in Fig.1. The errors shown in the figure 1 are the statistical one. It is clear from Fig1. that the e-by-e $\eta$ fluctuation tends to decrease with increasing mean multiplicity of the produced relativistic charged particles. This may be due the fact that the contribution to particle production would have been taken place by several independent sources. These contribution might be masking the correlated production. One can argue that there are identical sources which are producing low multiplicity events which is resulting in the low $\Phi$ values. It means when source fluctuation tends to vanish, pseudo- rapidity fluctuation increases. This observed trend of variation of $\Phi$ with the average multiplicity at high energies is in agreement with observations made by other studies in high energy regime18,20,21. To compare the experimental results with HIJING simulation, sample of events are generated with statistics approximately 10 times that of the experimental statistics. It is clear from Fig.1 that the $\Phi$ values obtained for HIJING data are lower than the its values for the experimental data at both the energies but the trend of variation of $\Phi$ with $<N_{s}>$ is almost same for both experimental and simulated data. Another interesting aspect of the event-by-event pseudo-rapidity fluctuation or the fluctuation of any global observable describing high energy nuclear collision data is to see its dependence on phase space region, which in this study is the pseudo-rapidity space itself. For this, the values of $\Phi$s are determined for various pseudo-rapidity intervals, $\Delta\eta=\eta_{2}-\eta_{1}$, where $\eta_{1}$ and $\eta_{2}$ are the lower and upper limit of a chosen $\eta$-window. In the present study the chosen $\Delta\eta$ are 0.5,1.0,1.5,2.0,3.0,4.0,5.0,6.0 for both the data sets. The calculated values of $\Phi$s corresponding to these pseudo-rapidity regions are listed in Table 2. Table 2: Calculated values of $\Phi$s in different pseudo-rapidity windows for the experimental and HIJING simulated data. Interactions | $\Delta\eta$ | $\Phi$ ---|---|--- | | Experimental | HIJING 4.5A GeV/c 28Si-nucleus | 0.5 | 0.199 $\pm$ 0.009 | 0.116 $\pm$ 0.008 1.0 | 0.499 $\pm$ 0.018 | 0.362 $\pm$ 0.023 1.5 | 1.159 $\pm$ 0.028 | 0.905 $\pm$ 0.082 2.0 | 2.179 $\pm$ 0.041 | 1.937 $\pm$ 0.088 2.5 | 2.890 $\pm$ 0.098 | 2.987 $\pm$ 0.092 3.0 | 3.972 $\pm$ 0.101 | 3.257 $\pm$ 0.098 4.0 | 4.452 $\pm$ 0.161 | 4.096 $\pm$ 0.117 5.0 | 4.622 $\pm$ 0.201 | 4.362 $\pm$ 0.188 6.0 | 5.027 $\pm$ 0.211 | 4.674 $\pm$ 0.198 14.5A GeV/c 28Si-nucleus | 0.5 | 0.194 $\pm$ 0.006 | 0.102 $\pm$ 0.021 1.0 | 0.387 $\pm$ 0.009 | 0.341 $\pm$ 0.029 1.5 | 1.097 $\pm$ 0.021 | 0.891 $\pm$ 0.038 2.0 | 2.063 $\pm$ 0.082 | 1.912 $\pm$ 0.043 2.5 | 2.732 $\pm$ 0.111 | 2.889 $\pm$ 0.055 3.0 | 3.817 $\pm$ 0.128 | 3.172 $\pm$ 0.076 4.0 | 4.158 $\pm$ 0.141 | 3.995 $\pm$ 0.102 5.0 | 4.489 $\pm$ 0.188 | 4.355 $\pm$ 0.111 6.0 | 4.811 $\pm$ 0.214 | 4.788 $\pm$ 0.175 It is observed from Table 2 that as we widened the pseudo-rapidity space, we noticed larger e-by-e fluctuations. The values of $\Phi$s first increases with increasing $\Delta\eta$ and then tends to saturate for much larger $\Delta\eta$. This behaviour is observed at both the energies considered in this analysis. This might be due to the dominating long-range correlations as compared to short-range correlations as we explore a larger rapidity space. Based on phenomenological evidence, it has been argued that particle production in high energy hadron-hadron and nucleus-nucleus collisions have been carrying the signals of both the short and long range correlations. The average number of produced particles virtually depend on the size of the initiating cluster, this gives rise to the long range correlation, means the particles which are separated by relatively large $\eta$ shows some correlation. The values of $\Phi$s for HIJING simulated data are smaller as compared to its values for the experimental data but HIJING data too show the similar dependence of $\Phi$ on $\Delta\eta$. These observations are much more clearly depicted in Fig.2, where we plotted $\Phi$ against $\Delta\eta$ along- with statistical errors. Conclusions Event-by-event fluctuations of pseudo-rapidity of the relativistic charged particles produced in 28Si-nucleus interactions at 4.5A and 14.5A GeV/c have been studied in terms of the fluctuation and correlation quantifying parameter, $\Phi$. Analysis reveals the presence of e-by-e $\eta$-fluctuations and correlation amongst the produced particles in pseudo-rapidity space at both the incident momentum as the non vanishing values of $\Phi$s are obtained. It’s observed that these fluctuations decreases with increasing mean multiplicities of the produced particles. This might be due to smearing out of the existing correlation as the more and more independent particle emitting sources added up. E-by-e fluctuations are also found to depend on the pseudo- rapidity windows and shows an increasing behaviour with increasing $\Delta\eta$. Results obtained for HIJING data exhibit a similar trend as compared to the experimental data at both the energies. Correlation and fluctuation studies remain to be excellent tools to explore the behaviour of the system produced in heavy-ion collisions at relativistic and ultra- relativistic energies. Acknowledgment: Financial support from DST, Govt. of India is acknowledged with thanks. References 1. 1. M. A. Stephanov, K. Rajagopal, E.V. Shuryak, Phys. Rev. Lett. 81 4816(1998) 2. 2. M. Luzum eta al., J. Phys. G. 41 063102 (2004) 3. 3. Y.Yoki et al., Nature 443, 675 (2006) 4. 4. S. Jeon et al., Phys. Rev. C 73, 014905 (2006) 5. 5. L. F. Babichev, A.N. Khmialeuski, Proceeding of 15th Int. Conf.-School, September 20-23, 2010 6. 6. M. Weber for the ALICE collaboration, J. Phys.: Conf. Series 389, 012036 (2012) 7. 7. M. Gazdzicki, S. Mrowczynski, Z. Phys. C 54, 127 (1992) 8. 8. E.V. Shuryak, Phys. Lett B 423, 9 (1998) 9. 9. A. Bialas, V. Koch, Phys. Lett. B 456, 1(1999) 10. 10. NA49 Collaboration (H. Appelshauser et al.), Phys. Lett. B 459, 679 (1999) 11. 11. G. Baym, H. Heisenberg, Phys. Lett. B 469, 7 (1999) 12. 12. G. Danilov, E. Shuryak, nucl-th/9908027 13. 13. T. Anticic et al., Phys. Rev. C 70, 034902 (2004) 14. 14. T. K. Nayak, J. Phys. G. 32, S187 (2006) arXiv:nucl-ex/060802. 15. 15. H. Heiselberg, Phys. Rep. 351, 161(2001) 16. 16. M. Stephanov et al., Phys. Rev. Lett. 81, 4816 (1998) 17. 17. M. Stephanov et al., Phys. Rev. D 61, 114028 (1999) 18. 18. S. Bhattacharya et al., Phys. Lett. B 726, 194 (2013) 19. 19. M. Belkacem et al., arXiv:nucl-th/9903017v2, 22 April 1999 20. 20. B. Abelev et al. Eur. Phys. J. C 74, 3077 (2014) 21. 21. KLM Collaboration (M.L. Cherry et al.), Acta Phys. Pol. B 29, 2129(1998) 22. 22. Gopa Bhoumic, Swarnapratim Bhattacharya et al., Euro.Phys.J. A52 196(2016) 23. 23. Shafiq ahmad et al., J. Phys. Soc. Jpn., 75, 064604 (2006) 24. 24. Shakeel Ahmad et al., Acta Phys. Pol. B 35, 809 (2004) 25. 25. M. Gyulassy and X.N. Wang, Comp. Phys. Commun., G 25 1895 (1999) 26. 26. M. Gazdzicki et al., Eur. Phys. J. C6, 365 (1999) 27. 27. M. Mrowczynski, Phys. Lett. B439 6 (1998)
YHEP-COS-21-01 ††thanks: co-corresponding author††thanks: co-corresponding author # Cosmic-Neutrino-Boosted Dark Matter ($\nu$BDM) Yongsoo Jho<EMAIL_ADDRESS>Department of Physics and IPAP, Yonsei University, Seoul 03722, Republic of Korea Jong-Chul Park<EMAIL_ADDRESS>Department of Physics and Institute of Quantum Systems (IQS), Chungnam National University, Daejeon 34134, Republic of Korea Seong Chan Park <EMAIL_ADDRESS>Department of Physics and IPAP, Yonsei University, Seoul 03722, Republic of Korea Po-Yan Tseng<EMAIL_ADDRESS>Department of Physics and IPAP, Yonsei University, Seoul 03722, Republic of Korea ###### Abstract A novel mechanism of boosting dark matter by cosmic neutrinos is proposed. The new mechanism is so significant that the arriving flux of dark matter in the mass window $1~{}{\rm keV}\lesssim m_{\rm DM}\lesssim 1~{}{\rm MeV}$ on Earth can be enhanced by two to four orders of magnitude compared to one only by cosmic electrons. Thereby we firstly derive conservative but still stringent bounds and future sensitivity limits for such cosmic-neutrino-boosted dark matter ($\nu$BDM) from advanced underground experiments such as Borexino, PandaX, XENON1T, and JUNO. ## I Introduction Revealing the properties of dark matter (DM) is definitely one of the most pressing issues in particle physics, astrophysics, and cosmology. Direct detection experiments of DM have particular importance as they aim to probe interaction of DM with standard model (SM) particles [1]. However, there exists fundamental limitation in detecting a sub-MeV dark matter set by the maximum kinetic energy of the DM particle in halo: $K_{\rm DM}^{\rm max}\lower 3.01385pt\hbox{$\;\stackrel{{\scriptstyle\textstyle<}}{{\sim}}\;$}10^{-6}m_{\rm DM}\lower 3.01385pt\hbox{$\;\stackrel{{\scriptstyle\textstyle<}}{{\sim}}\;$}1~{}{\rm eV}$ (1) with the velocity $v\sim 10^{-3}$. This low kinetic energy causes a significant problem in detecting light dark matter since the recoil energy of scattered SM particle is also limited by the kinetic energy 111Several ideas have been suggested to detect signals with low recoil energies by lowering the threshold energies at detectors (see [2, 3] and references therein).. On the other hand, there still exists a chance to detect a subcomponent of DM, dubbed ‘boosted dark matter’ (BDM), which may carry much larger energy beyond threshold due to various mechanisms [4, 5, 6, 7, 8, 9, 10] including scattering by energetic cosmic-ray particles [11, 12, 13, 14, 15, 16, 17]. We note that focus has been given to cosmic-ray electron and proton so far even though the chance is not exclusively open for charged particles. In this letter, we focus on a noble class of cosmic-neutrino-boosted-dark matter ($\nu$BDM) extending previous studies: there exist a huge number of cosmic-ray neutrinos arriving at the solar system from various origins [18]. Our Sun is also generating a large number of neutrinos [19, 20, 21] so that they may boost DM within the solar system. We find that $\nu$BDM can be a dominant part of the whole BDM when DM-neutrino interaction is as strong as DM-electron interaction, which is indeed the case for gauged lepton number as mediator, for instance [22, 23]. The existing conclusions regarding cosmic- electron-induced BDM should be re-examined. ## II Boost mechanism by cosmic neutrino Cosmic neutrino inputs. Near Earth, our Sun provides the dominating neutrino flux $d\Phi^{\rm Sun}_{\nu}/dK_{\nu}$ in the neutrino energy $K_{\nu}\lesssim 10$ MeV reaching the maximum $\simeq\mathcal{O}(10^{8})~{}[{\rm cm^{-2}\,s^{-1}\,keV^{-1}}]$ around $K_{\nu}\simeq 0.3$ MeV [19, 20, 21], which gives the total number of neutrino emission rate per unit energy $\displaystyle\frac{d\dot{N}^{\rm Sun}_{\nu}}{dK_{\nu}}\equiv\frac{d\Phi^{\rm Sun}_{\nu}}{dK_{\nu}}\,(4\pi D^{2}_{\odot})\,,$ (2) where $D_{\odot}=1$ AU is the distance between Sun and Earth. The neutrinos can boost non-relativistic light DM, leaving distinctive signals at terrestrial experiments, e.g. XENON1T [24, 25]. The total contributions from all stars for $\nu$BDM could be significant compared to the BDM flux by the solar neutrinos. The overall neutrino flux from all stars in the Milky Way (say, cosmic-neutrino flux) has not been measured by astrophysical observations, and could be highly anisotropic, which is different from the isotropic diffused cosmic electrons. In general, DM particles can be boosted by the neutrino flux from the nearest star, instead of diffused neutrinos. Keep this philosophy in mind, we will compute the $\nu$BDM flux by starting with single star contribution in the following section, then integrate the entire star distribution in the Milky Way. Cosmic neutrino and DM scattering. The halo DM is boosted by neutrino through the process $\nu+\chi\to\nu+\chi$, which may originate from the exchange of the $U(1)_{L_{e}-L_{i}}$ gauge boson or dim-6 effective operators including $(\bar{\ell}\gamma^{\mu}\ell)(\bar{\chi}\gamma_{\mu}\chi)$ or $(\bar{\ell}\ell)(\bar{\chi}\chi)$. The resulting BDM kinetic energy $K_{\rm DM}$ can be determined from the kinetic energy of incoming neutrino $K_{\nu}$. At the halo DM rest frame, the allowed range of $K_{\rm DM}$ is given by [26] $\displaystyle 0\leq K_{\rm DM}\leq K^{\rm max}_{\rm DM}\equiv\frac{2m_{\rm DM}(K^{2}_{\nu}+2m_{\nu}K_{\nu})}{(m_{\rm DM}+m_{\nu})^{2}+2m_{\rm DM}K_{\nu}}\,.$ (3) Figure 1: [Top] Schematic description of BDM production by the neutrino from a single star. [Bottom] Areal density of unit-normalized distribution of the $\nu$BDM flux from stars in our galaxy $\mathcal{P}(\vec{y})\equiv\frac{1}{\Phi_{\rm DM}}\frac{d\Phi_{\rm DM}}{dA_{y}}$ [kpc-2], for two representative ranges of $K_{\rm DM}$: 10 – 100 keV (left) and 1 – 10 MeV (right). $dA_{y}$ is the areal element of the Galactic disk, defined by position of star, $\vec{y}$. The BDM flux by neutrinos from a Sun-like star is $\displaystyle\frac{d\Phi^{(1)}_{\rm DM}(\overrightarrow{y})}{dK_{\rm DM}}$ $\displaystyle\simeq$ $\displaystyle\frac{1}{8\pi^{2}}\left(\tilde{f}_{1}\frac{d\dot{N}^{\rm Sun}_{\nu}}{dK_{\nu}}\right)\int d^{3}\overrightarrow{z}\frac{\rho_{\rm DM}(|\overrightarrow{z}|)}{m_{\rm DM}}\frac{1}{|\overrightarrow{x}-\overrightarrow{z}|^{2}}$ (4) $\displaystyle\times\left(\left.\frac{dK_{\nu}}{d\bar{\theta}}\right|_{\bar{\theta}=\bar{\theta}_{0}}\right)\left(\left.\frac{d\sigma_{\nu{\rm DM}}}{dK_{\rm DM}}\right|_{\bar{\theta}=\bar{\theta}_{0}}\right)$ $\displaystyle\times\frac{1}{\sin\bar{\theta}_{0}}\frac{1}{|\overrightarrow{z}-\overrightarrow{y}|^{2}}\times\exp{\left(-\frac{|\overrightarrow{z}-\overrightarrow{y}|}{d_{\nu}}\right)}\,,$ where the schematic diagram of the coordinate system is shown in the top panel of Fig. 1, and $\overrightarrow{x},\overrightarrow{y}$, and $\overrightarrow{z}$ represent the positions of Earth, Star, and halo DM, respectively. The correction factor $\tilde{f}_{1}$ takes into account the variances of stellar properties from Sun [27] and $\rho_{\rm DM}$ is the DM halo density profile. The differential $\nu$-DM cross section depends on scattering angle $\bar{\theta}$, and $\bar{\theta}_{0}$ can be determined by $K_{\nu}$ and $K_{\rm DM}$ via kinematic relations: $\displaystyle K_{\nu}(K_{\rm DM},\bar{\theta})$ $\displaystyle=$ $\displaystyle\frac{p^{\prime 2}-K^{2}_{\rm DM}}{2\left(p^{\prime}\cos\bar{\theta}-K_{\rm DM}\right)}\,,$ (5) $\displaystyle\left.\frac{dK_{\nu}}{d\bar{\theta}}\right|_{\bar{\theta}=\bar{\theta}_{0}}$ $\displaystyle=$ $\displaystyle\frac{(p^{\prime 2}-K^{2}_{\rm DM})p^{\prime}}{2\left(p^{\prime}\cos\bar{\theta}_{0}-K_{\rm DM}\right)^{2}}\sin\bar{\theta}_{0}\,,$ (6) where $p^{\prime}\equiv\sqrt{2m_{\rm DM}K_{\rm DM}+K^{2}_{\rm DM}}$ is 3-momentum of BDM in the halo DM frame. $dK_{\nu}/d\bar{\theta}\propto 1/\cos^{2}\bar{\theta}$ and large scattering angle $\bar{\theta}\simeq\pi/2$ is favoured for $m_{\rm DM}\gg K_{\rm DM}$, whereas $dK_{\nu}/d\bar{\theta}\propto 1/(\cos\bar{\theta}-1)^{2}$ makes the forward scattering $\bar{\theta}\simeq 0$ dominate for $m_{\rm DM}\ll K_{\rm DM}$. The neutrino flux attenuation due to propagation is determined by the exponential function in Eq. (4), and the mean free path is obtained as $d_{\nu}\equiv 1/[(\rho_{\rm DM}/m_{\rm DM})\cdot\sigma_{\nu{\rm DM}}]$ where the total $\nu$-DM cross section is $\displaystyle\sigma_{\nu{\rm DM}}(K_{\nu})\equiv\int^{K^{\rm max}_{\rm DM}}_{0}dK_{\rm DM}\frac{d\sigma_{\nu{\rm DM}}}{dK_{\rm DM}}\,.$ (7) In the derivation of Eq. (4), we use point-like star approximation, by starting with finite star radius $R_{\rm star}$ then taking $R_{\rm star}\to 0$. The final result of $d\Phi^{(1)}_{\rm DM}(\overrightarrow{y})/dK_{\rm DM}$ is finite. Due to the distance-squared suppression, the dominating $\nu$BDM fluxes originate either from halo DM at the vicinity of Earth or the galatic center (GC). Figure 2: The unit-normalized arrival direction $\theta$ distributions of the $\nu$BDM spectral flux $\varphi_{\rm BDM}\equiv d\Phi_{\rm DM}/dK_{\rm DM}$ for two benchmark values of $K_{\rm DM}$: 10 keV (left) and 1 MeV (right) varying $m_{\rm DM}=$ 5 keV – 5 MeV with a fixed mediator mass, $m_{X}=700$ keV. From Eq. (4), we can calculate the BDM flux by neutrinos from Sun by taking $|\overrightarrow{x}-\overrightarrow{y}|=D_{\odot}$. Even though Sun provides the largest neutrino flux to Earth, only small volume of nearby DM halo compromises the BDM flux. Therefore, we need to consider the entire stellar contributions in the Milky Way by convolving Eq. (4) with stellar distribution $n_{\rm star}(\overrightarrow{y})$: $\displaystyle\frac{d\Phi_{\rm DM}}{dK_{\rm DM}}=\int d^{3}\overrightarrow{y}n_{\rm star}(\overrightarrow{y})\frac{d\Phi^{(1)}_{\rm DM}(\overrightarrow{y})}{dK_{\rm DM}}\,.$ (8) Here we assume stars distribute within the galactic disk, shown in the top panel of Fig. 1, with radius $R\leq 20$ kpc and thickness $|h|\leq 1$ kpc. Using the observation [28] and integrating out the $h$, the stellar distribution on 2-dimensional galactic disk is given by $\displaystyle n_{\rm star}(R)\simeq\tilde{f}_{2}\times 1.2\times 10^{11}/(R/{\rm kpc})^{3}~{}[{\rm kpc^{-2}}]\,,$ (9) where $\tilde{f}_{2}$ factor includes the uncertainties from detailed structures of the Milky Way, e.g. spiral arms and density fluctuations. The $\nu$BDM fluxes with two $K_{\rm DM}$ regimes are shown in the bottom panels of Fig. 1. Due to the high stellar and DM number densities around the GC, the BDM flux contribution from the GC region exceeds that from the vicinity of Earth. Fig. 2 shows the $\theta$ dependence of the $\nu$BDM fluxes $\frac{1}{\varphi_{\rm BDM}}\frac{\varphi_{\rm BDM}}{d\theta}$ at Earth where $\theta$ represents the angle between the $\nu$BDM arrival direction and the GC. For $K_{\rm DM}\gg m_{\rm DM}$ in the right panel, the forward scattering ($\bar{\theta}_{0}\simeq 0$) is preferred, so that $\nu$BDM from the GC dominates and thus $\theta\simeq 0$. In the left panel, $K_{\rm DM}\ll m_{\rm DM}$ prefers large-angle scattering ($\bar{\theta}_{0}\simeq 90^{\circ}$), which enhances the flux for $\theta\gtrsim 40^{\circ}$ originating relatively far from the GC. The $\theta$ dependence of the $\nu$BDM flux can be used to determine $m_{\rm DM}$ in the future. Figure 3: [Top] BDM fluxes by solar neutrinos, cosmic neutrinos, and cosmic electrons. We assume $\sigma_{\nu{\rm DM}}$ comes from a vector boson $X$ coupling to both DM and leptons ($g_{X}=g_{e}=g_{\nu}$) with $(m_{\rm DM},m_{X},g_{X}g_{\rm DM})=(5{\rm MeV},700{\rm keV},10^{-6})$. The uncertainty band for $\nu$BDM corresponds to $0.1\leq\tilde{f}\leq 10$. [Bottom] BDM fluxes for different $m_{X}$ and $m_{\rm DM}$ with $\tilde{f}=1$. Solid and dotted lines are $\nu$BDM and cosmic electron BDM fluxes, respectively. In the top panel of Fig. 3, we compare the BDM fluxes via solar neutrinos, cosmic neutrinos, and cosmic-ray electrons by fixing $\tilde{f}\equiv\tilde{f}_{1}\cdot\tilde{f}_{2}$ =1. The $\nu$BDM flux is three orders of magnitude larger than that by solar neutrinos, because the later is relevant to DM only within a few AUs around Earth. Three bumps of the $\nu$BDM flux correspond to the $pp$, 13N+15O, and 8B production processes of solar neutrinos [19]. Assuming $g_{e}=g_{\nu}$, the $\nu$BDM flux can be two to four orders of magnitude larger than that induced by cosmic electrons for $K_{\rm DM}\lesssim 50~{}{\rm keV}$. This feature is quite robust for other DM and mediator masses as shown in the bottom panel. There are several factors that can make our estimations different. i ) The DM halo profile, especially around the GC. We take the NFW profile. ii ) The $\nu$ flux varies with the type and age of stars [27]. iii ) The star distribution in the Milky Way. All the above uncertainties are hard to be included in the calculation. In order to show the robustness of the results, we conservatively vary $0.1\lesssim\tilde{f}\lesssim 10$ in Eq. (4) and (8), depicted as a blue band in the top panel of Fig. 3. Attenuation of the BDM flux. The attenuation effect of the cosmic-neutrino flux scattered by halo DM is taken into account by the exponential factor in Eq. 4. We estimate the mean free path of cosmic neutrino $d_{\nu}$ by taking $\sigma_{\nu{\rm DM}}\simeq 10^{-28}-10^{-34}~{}{\rm cm^{2}}$. For the $n_{\rm DM}\sim({\rm keV}/m_{\rm DM})\times 10^{6}~{}{\rm cm^{-3}}$, $d_{\nu}\simeq(m_{\rm DM}/{\rm keV})\times(10^{22}-10^{28})~{}{\rm cm}$, which is larger than the size of the Milky Way and results in negligible effect. Next, we estimate the mean free path of BDM inside Earth by assuming $\sigma_{e{\rm DM}}=\sigma_{\nu{\rm DM}}$. For $\sigma_{e{\rm DM}}=10^{-33}~{}{\rm cm^{2}}$ with electron number density of Earth $n_{e}\simeq 10^{24}~{}{\rm cm^{-3}}$, the mean free path $1/(n_{e}\cdot\sigma_{e{\rm DM}})\simeq 10^{4}~{}{\rm km}$ is comparable to the size of Earth. For $\sigma_{e{\rm DM}}=10^{-29}~{}{\rm cm^{2}}$, the BDM mean free path reduces to $\mathcal{O}({\rm km})$. Most of DM direct detection detectors locate a few kilometers underground, rendering the $\nu$BDM signal be substantially suppressed for $\sigma_{e{\rm DM}}\gtrsim 10^{-29}~{}{\rm cm^{2}}$. Thus, the attenuation of BDM inside Earth will provide upper limits on experimental sensitivities as shown in Fig. 4. ## III Experimental sensitivities To estimate experimental sensitivities, we use two approaches for DM models. i ) Heavy mediator: the interactions can be described by effective cross sections $\sigma^{\rm eff}_{\nu{\rm DM}}$ and $\sigma^{\rm eff}_{e{\rm DM}}$. ii ) Light mediator: the $X$ boson from the gauged $U(1)_{L_{e}-L_{i}}$ couples to DM and leptons. For the approach i ), the differential cross section is defined as $\displaystyle\frac{d\sigma^{\rm eff}_{\nu{\rm DM},e{\rm DM}}}{dK_{\rm DM}}\equiv\frac{\sigma^{\rm eff}_{\nu{\rm DM},e{\rm DM}}}{K^{\rm max}_{\rm DM}-K^{\rm min}_{\rm DM}}\,.$ (10) On the other hand, for the $U(1)_{L_{e}-L_{i}}$ model, the neutrino-DM scattering cross section is given by [29] $\displaystyle\frac{d\sigma_{\nu{\rm DM}}}{dK_{\rm DM}}=\frac{(g_{X}g_{\rm DM})^{2}}{4\pi}\frac{2m_{\rm DM}(m_{\nu}+K_{\nu})^{2}-K_{\rm DM}\left[(m_{\nu}+m_{\rm DM})^{2}+2m_{\rm DM}K_{\nu}\right]+m_{\rm DM}K^{2}_{\rm DM}}{(2m_{\nu}K_{\nu}+K^{2}_{\nu})(2m_{\rm DM}K_{\rm DM}+m^{2}_{X})^{2}}\,.$ (11) For $K_{\rm DM}\simeq\mathcal{O}({\rm keV})$ and $m_{\rm DM}\simeq m_{X}\simeq\mathcal{O}({\rm MeV})$, it makes $d\sigma_{\nu{\rm DM}}/dK_{\rm DM}$ almost independent of $K_{\rm DM}$. Figure 4: $\nu$BDM contributions to XENON1T electron recoil, assuming $\sigma^{\rm eff}_{\nu{\rm DM}}=\sigma^{\rm eff}_{e{\rm DM}}$, where the $1\sigma$ (green) and $2\sigma$ (white) regions from $\chi^{2}$ analysis, and the gray-shaded region is excluded more than $2\sigma$. The expected sensitivities from other underground detectors are depicted: Brexino [30], PandaX [31], XENONnT [32], and JUNO [33]. For comparison, existing limits are shown together: CDMS HVeV [34], DAMIC [35], EDELWEISS [36], and SENSEI [37]. The cosmic-electron-BDM constraints from Super-K and Hyper-K [14]. We perform the model-independent $\chi^{2}$ analysis for the effective cross section $(m_{\rm DM},\sigma^{\rm eff}_{\nu{\rm DM}}=\sigma^{\rm eff}_{e{\rm DM}})$ in Fig. 4. There are five disconnected $1\sigma$ regions for the XENON1T excess [25], which originate from the three bumps of the $\nu$BDM flux spectrum in Fig. 3. The $2\sigma$ exclusion region is gray-shaded. The $\nu$BDM provides stringent constraint on $\sigma^{\rm eff}_{\nu{\rm DM}}=\sigma^{\rm eff}_{e{\rm DM}}$ for unexplored small mass $m_{\rm DM}\lesssim{\rm MeV}$, compared with the current limits from DM direct detection experiments including CDMS HVeV [34], DAMIC [35], EDELWEISS [36], and SENSEI [37]. We evaluate the sensitivities of $\nu$BDM with other current (Brexino [30], PandaX [31]) and future experiments (XENONnT [32], JUNO [33]). To estimate the sensitivities, we take the four ton-year exposure for XENONnT and 20 kton-year exposure for JUNO assuming no excess above the expected background and dominance of statistical uncertainty. Borexino and JUNO have higher energy threshold above 100 keV but huge statistics. JUNO has the best sensitivity for $m_{\rm DM}\lesssim 0.5~{}{\rm MeV}$, while XENON1T/nT are better than JUNO for $m_{\rm DM}\gtrsim 0.5~{}{\rm MeV}$. PandaX has a slightly weaker limit due to the smaller 0.276 ton-year exposure than XENON1T of 0.65 tonne-year. For $\sigma_{e{\rm DM}}\gtrsim 10^{-29}~{}{\rm cm^{2}}$, the earth crust attenuates the BDM flux; specifically, XENON1T and XENONnT [24] are located underground at a depth of 3600 meter water equivalent (m.w.e.) and Borexino [38] is at 3800 m.w.e while PandaX [39] is shielded by 2400 m marble overburden ($\sim 6800$ m.w.e.). The most shallow JUNO detector [33], located at 700 m deep underground ($\sim 2000$ m.w.e.), has the best upper sensitive to $\nu$BDM with $\sigma_{e{\rm DM}}\simeq 10^{-28}~{}{\rm cm^{2}}$. ## IV Discussions The flux of the cosmic-neutrino-boosted-DM ($\nu$BDM) is substantially larger than the one of the cosmic-electron-boosted-DM so that it contributes dominantly in direct detection experiments on Earth. Due to the distributions of the sources of neutrinos in Milky Way and the dark matter in halo, the angular distribution of the $\nu$BDM is kinematically correlated with the DM mass. Therefore precise measurement of directional information helps in determination of the DM mass. The existing underground detectors probe the parameter region of neutrino-DM interaction and electron-DM interaction in $10^{-34}~{}{\rm cm^{2}}\lesssim\sigma_{\nu{\rm DM}}=\sigma_{e{\rm DM}}\lesssim 10^{-28}~{}{\rm cm^{2}}$ with $1~{}{\rm keV}\lesssim m_{\rm DM}\lesssim 100~{}{\rm MeV}$ based on the effective cross section approach. Since the DM flux is enhanced by neutrino-boost, we find parameter regions for the recent XENON1T anomaly (see Fig. 4). However, they are still hardly consistent with other DM searches. Finally, we discuss various factors of future refinement of the current study. Here we only assumed that nuclear activities inside each star are on average same as in our Sun, so that the neutrino fluxes from each star are all similar. Obviously, this is a crude estimation and actual neutrino fluxes differ from star to star. Also, the GC region has the largest population of main sequence stars and also red giants [40, 41], which enhances $\tilde{f}_{1}\cdot\tilde{f}_{2}$ factor over unity [27]. Last but not least, we point out the potential modification due to the extra galactic neutrinos. Even though extra galactic contributions in neutrino flux is subdominant in the energy range for $\nu$BDM [18] , it can lead modification in e.g. spatial and kinetic distributions of $\nu$BDM. All those factors of improvement are reserved for the future work. ## Acknowledgments The work is supported in part by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education, Science and Technology [NRF-2018R1A4A1025334, NRF-2019R1A2C1089334 (SCP), NRF-2019R1C1C1005073 (JCP) and NRF-2020R1I1A1A01066413 (PYT)]. ## References * Goodman and Witten [1985] M. W. Goodman and E. Witten, Phys. Rev. D31, 3059 (1985). * Battaglieri _et al._ [2017] M. Battaglieri _et al._ , (2017), arXiv:1707.04591 [hep-ph] . * Kim _et al._ [2020] D. Kim, J.-C. Park, K. C. Fong, and G.-H. Lee, (2020), arXiv:2002.07821 [hep-ph] . * Belanger and Park [2012] G. Belanger and J.-C. Park, JCAP 1203, 038, arXiv:1112.4491 [hep-ph] . * Agashe _et al._ [2014] K. Agashe, Y. Cui, L. Necib, and J. Thaler, JCAP 1410 (10), 062, arXiv:1405.7370 [hep-ph] . * Berger _et al._ [2015] J. Berger, Y. Cui, and Y. Zhao, JCAP 1502 (02), 005, arXiv:1410.2246 [hep-ph] . * Kong _et al._ [2015] K. Kong, G. Mohlabeng, and J.-C. Park, Phys. Lett. B743, 256 (2015), arXiv:1411.6632 [hep-ph] . * Kim _et al._ [2017] D. Kim, J.-C. Park, and S. Shin, Phys. Rev. Lett. 119, 161801 (2017), arXiv:1612.06867 [hep-ph] . * Giudice _et al._ [2018] G. F. Giudice, D. Kim, J.-C. Park, and S. Shin, Phys. Lett. B780, 543 (2018), arXiv:1712.07126 [hep-ph] . * D’Eramo and Thaler [2010] F. D’Eramo and J. Thaler, JHEP 06, 109, arXiv:1003.5912 [hep-ph] . * Bringmann and Pospelov [2019] T. Bringmann and M. Pospelov, Phys. Rev. Lett. 122, 171801 (2019), arXiv:1810.10543 [hep-ph] . * Ema _et al._ [2019] Y. Ema, F. Sala, and R. Sato, Phys. Rev. Lett. 122, 181802 (2019), arXiv:1811.00520 [hep-ph] . * Cappiello _et al._ [2019] C. V. Cappiello, K. C. Y. Ng, and J. F. Beacom, Phys. Rev. D99, 063004 (2019), arXiv:1810.07705 [hep-ph] . * Cappiello and Beacom [2019] C. Cappiello and J. F. Beacom, Phys. Rev. D 100, 103011 (2019), arXiv:1906.11283 [hep-ph] . * Dent _et al._ [2020] J. B. Dent, B. Dutta, J. L. Newstead, and I. M. Shoemaker, Phys. Rev. D 101, 116007 (2020), arXiv:1907.03782 [hep-ph] . * Cho _et al._ [2020] W. Cho, K.-Y. Choi, and S. M. Yoo, (2020), arXiv:2007.04555 [hep-ph] . * Jho _et al._ [2020] Y. Jho, J.-C. Park, S. C. Park, and P.-Y. Tseng, Phys. Lett. B811, 135863 (2020), arXiv:2006.13910 [hep-ph] . * Vitagliano _et al._ [2020] E. Vitagliano, I. Tamborra, and G. Raffelt, Rev. Mod. Phys. 92, 45006 (2020), arXiv:1910.11878 [astro-ph.HE] . * Bahcall _et al._ [2005] J. N. Bahcall, A. M. Serenelli, and S. Basu, Astrophys. J. Lett. 621, L85 (2005), arXiv:astro-ph/0412440 [astro-ph] . * Billard _et al._ [2014] J. Billard, L. Strigari, and E. Figueroa-Feliciano, Phys. Rev. D 89, 023524 (2014), arXiv:1307.5458 [hep-ph] . * Vitagliano _et al._ [2017] E. Vitagliano, J. Redondo, and G. Raffelt, JCAP 12, 010, arXiv:1708.02248 [hep-ph] . * Rajpoot [1989] S. Rajpoot, Phys. Rev. D 40, 2421 (1989). * He _et al._ [1991] X. He, G. C. Joshi, H. Lew, and R. Volkas, Phys. Rev. D 43, 22 (1991). * Aprile _et al._ [2017] E. Aprile _et al._ (XENON), Eur. Phys. J. C77, 881 (2017), arXiv:1708.07051 [astro-ph.IM] . * Aprile _et al._ [2020a] E. Aprile _et al._ (XENON), (2020a), arXiv:2006.09721 [hep-ex] . * Boschini _et al._ [2018] M. J. Boschini _et al._ , Astrophys. J. 854, 94 (2018), arXiv:1801.04059 [astro-ph.HE] . * Farag _et al._ [2020] E. Farag, F. X. Timmes, M. Taylor, K. M. Patton, and R. Farmer 10.3847/1538-4357/ab7f2c (2020), arXiv:2003.05844 [astro-ph.SR] . * de Jong _et al._ [2010] J. T. A. de Jong, B. Yanny, H.-W. Rix, A. E. Dolphin, N. F. Martin, and T. C. Beers (SDSS), Astrophys. J. 714, 663 (2010), arXiv:0911.3900 [astro-ph.GA] . * Cao _et al._ [2020] Q.-H. Cao, R. Ding, and Q.-F. Xiang, (2020), arXiv:2006.12767 [hep-ph] . * Bellini _et al._ [2011] G. Bellini _et al._ , Phys. Rev. Lett. 107, 141302 (2011), arXiv:1104.1816 [hep-ex] . * Zhou _et al._ [2020] X. Zhou _et al._ (PandaX-II), (2020), arXiv:2008.06485 [hep-ex] . * Aprile _et al._ [2020b] E. Aprile _et al._ (XENON), (2020b), arXiv:2007.08796 [physics.ins-det] . * An _et al._ [2016] F. An _et al._ (JUNO), J. Phys. G43, 030401 (2016), arXiv:1507.05613 [physics.ins-det] . * Agnese _et al._ [2018] R. Agnese _et al._ (SuperCDMS), Phys. Rev. Lett. 121, 051301 (2018), [erratum: Phys. Rev. Lett.122,no.6,069901(2019)], arXiv:1804.10697 [hep-ex] . * Aguilar-Arevalo _et al._ [2019] A. Aguilar-Arevalo _et al._ (DAMIC), Phys. Rev. Lett. 123, 181802 (2019), arXiv:1907.12628 [astro-ph.CO] . * Arnaud _et al._ [2020] Q. Arnaud _et al._ (EDELWEISS), (2020), arXiv:2003.01046 [astro-ph.GA] . * Barak _et al._ [2020] L. Barak _et al._ (SENSEI), (2020), arXiv:2004.11378 [astro-ph.CO] . * Back _et al._ [2012] H. Back _et al._ (Borexino), JINST 7, P10018, arXiv:1207.4816 [physics.ins-det] . * Cao _et al._ [2014] X. Cao _et al._ (PandaX), Sci. China Phys. Mech. Astron. 57, 1476 (2014), arXiv:1405.2882 [physics.ins-det] . * Robin _et al._ [2012] A. C. Robin, D. J. Marshall, M. Schultheis, and C. Reyle, Astron. Astrophys. 538, A106 (2012), arXiv:1111.5744 [astro-ph.GA] . * Valenti, E. _et al._ [2016] Valenti, E., Zoccali, M., Gonzalez, O. A., Minniti, D., Alonso-García, J., Marchetti, E., Hempel, M., Renzini, A., and Rejkuba, M., A&A 587, L6 (2016).
# Enquire One’s Parent and Child Before Decision: Fully Exploit Hierarchical Structure for Self-Supervised Taxonomy Expansion Suyuchen Wang<EMAIL_ADDRESS>0000-0003-0404-2921 Mila & DIRO, Université de MontréalMontréalQuébecCanada , Ruihui Zhao <EMAIL_ADDRESS>Tencent Jarvis LabShenzhenGuangdongChina , Xi Chen <EMAIL_ADDRESS>Tencent Jarvis LabShenzhenGuangdongChina , Yefeng Zheng<EMAIL_ADDRESS>Tencent Jarvis LabShenzhenGuangdongChina and Bang Liu<EMAIL_ADDRESS>Mila & DIRO, Université de MontréalMontréalQuébecCanada (2021) ###### Abstract. Taxonomy is a hierarchically structured knowledge graph that plays a crucial role in machine intelligence. The taxonomy expansion task aims to find a position for a new term in an existing taxonomy to capture the emerging knowledge in the world and keep the taxonomy dynamically updated. Previous taxonomy expansion solutions neglect valuable information brought by the hierarchical structure and evaluate the correctness of merely an added edge, which downgrade the problem to node-pair scoring or mini-path classification. In this paper, we propose the Hierarchy Expansion Framework (HEF), which fully exploits the hierarchical structure’s properties to maximize the coherence of expanded taxonomy. HEF makes use of taxonomy’s hierarchical structure in multiple aspects: i) HEF utilizes subtrees containing most relevant nodes as self-supervision data for a complete comparison of parental and sibling relations; ii) HEF adopts a coherence modeling module to evaluate the coherence of a taxonomy’s subtree by integrating hypernymy relation detection and several tree-exclusive features; iii) HEF introduces the Fitting Score for position selection, which explicitly evaluates both path and level selections and takes full advantage of parental relations to interchange information for disambiguation and self-correction. Extensive experiments show that by better exploiting the hierarchical structure and optimizing taxonomy’s coherence, HEF vastly surpasses the prior state-of-the-art on three benchmark datasets by an average improvement of 46.7% in accuracy and 32.3% in mean reciprocal rank. taxonomy expansion, self-supervised learning, hierarchical structure ††copyright: none††copyright: acmcopyright††journalyear: 2021††conference: WWW ’21: The Web Conference; April 19–23, 2021; Ljubljana, Slovenia††booktitle: WWW ’21: The Web Conference, April 19–23, Ljubljana, Slovenia††price: 15.00††isbn: 978-1-4503-XXXX-X/18/06 ## 1\. Introduction Figure 1. An illustration of the taxonomy expansion task and the contributions of the proposed HEF model. Taxonomy is a particular type of hierarchical knowledge graph that portrays the hypernym-hyponym relations or “is-A” relations of various concepts and entities. They have been adopted as the underlying infrastructure of a wide range of online services in various domains, such as product catalogs for e-commerce (Karamanolakis et al., 2020; Luo et al., 2020), scientific indices like MeSH (Lipscomb, 2000), and lexical databases like WordNet (Miller, 1995). A well-constructed taxonomy can assist various downstream tasks, including web content tagging (Liu et al., 2019; Peng et al., 2019), web searching (Yin and Shah, 2010), personalized recommendation (Huang et al., 2019) and helping users achieve quick navigation on web applications (Hua et al., 2017). Manually constructing and maintaining a taxonomy is laborious, expensive and time-consuming. It is also highly inefficient and detrimental for downstream tasks if we construct a taxonomy from scratch (Velardi et al., 2013; Gupta et al., 2017) as long as the taxonomy has new terms to be added. A more realistic strategy is to insert new terms (“query”) into an existing taxonomy, i.e., the seed taxonomy, as a child of an existing node in the taxonomy (“anchor”) without modifying its original structure to best preserve its design. This problem is called taxonomy expansion (Jurgens and Pilehvar, 2016). Early taxonomy expansion approaches use terms that do not exist in the seed taxonomy and its best-suited position in the seed taxonomy as training data (Jurgens and Pilehvar, 2015). However, it suffers from the insufficiency of training data and the shortage of taxonomy structure supervision. More recent solutions adopt self-supervision and try to exploit the information of nodes in the seed taxonomy (seed nodes) to perform node pair matching (Shen et al., 2020) or classification along mini paths in the taxonomy (Yu et al., 2020). However, these approaches do not fully utilize the taxonomy’s hierarchical structure’s characteristics, and neglect the coherence of the extended taxonomy which oughts to be the core of the taxonomy expansion task. More specifically, existing approaches do not model a hierarchical structure identical to the taxonomy. Instead, they use ego-nets (Shen et al., 2020) or mini-paths (Yu et al., 2020) and feature few or no tree-exclusive information, making them unable to extract or learn the complete hierarchical design of a taxonomy. Besides, they do not consider the coherence of a taxonomy. They manage to find the most suitable node in a limited subgraph and only evaluate the correctness of a single edge instead of the expanded taxonomy, which downgrades the taxonomy expansion task to a hypernymy detection task. Lastly, their scoring approach regards the anchor node as an individual node without considering the hierarchical context information. However, the hierarchical structure provides multi-aspect criteria to evaluate a node, such as its path or level correctness. The structure also marks the nodes that are most likely to be wrongly chosen to be a parent in a specific parental relation. To solve all the stated flaws in previous works, we propose the Hierarchy Expansion Framework (HEF), which aims to maximize the coherence of the expanded taxonomy instead of the fitness of a single edge by fully exploiting the hierarchical structure of a taxonomy for self-supervised training, as well as modeling and evaluating the structure of taxonomy. HEF’s designs and goals are illustrated in Fig. 1. Specifically, we make the following contributions. Firstly, we design an innovative hierarchical data structure for self- supervision to mimic how humans construct a taxonomy. Relations in a taxonomy include hypernymy relations along a root-path and similarity among siblings. To find the most suitable parent node for the query term, human experts need to compare an anchor node with all its ancestors to distinguish the most appropriate one and compare the query with its potential siblings to testify their similarity. For example, to choose the parent for query “black tea” in the food taxonomy, the most appropriate anchor “tea” can only be selected by distinguishing from its ancestors “beverage” and “food”, which are all “black tea” ’s hypernyms, as well as compare the query “black tea” with “tea”’s children like “iced tea” and “oolong” to guarantee similarity among siblings. Thus, we design a new structure called “ego-tree” for self-supervision, which contains all ancestors and sample of children of a node for taxonomy structure learning. Our ego-tree incorporates richer topological context information for attaching a query term to a candidate parent with minimal computation cost compared to previous approaches based on node pair matching or path information. Secondly, we design a new modeling strategy to perform explicit ego-tree coherence modeling apart from the traditional node-pair hypernymy detection. Instead of merely modeling the correctness of the added edge, we adopt a more comprehensive approach to detect whether the anchor’s ego-tree after adding the query maintains the original design of the seed taxonomy. The design of taxonomy includes natural hypernymy relations, which needs the representation of node-pair relations and expert-curated level configurations, such as species must be placed in the eighth level of biological taxonomy, or adding one more adjective to a term means exactly one level higher in the e-commerce taxonomy. We adopt a coherence modeling module to detect the two aspects of coherence: i) For natural hypernymy relations, we adopt a hypernymy detection module to represent the relation between the query and each node in the anchor’s ego-tree. ii) For expert-curated designs, we integrate hierarchy- exclusive features such as embeddings of a node’s absolute level and relative level to the anchor into the coherence modeling module. Thirdly, we design a multi-dimensional evaluation to score the coherence of the expanded taxonomy. The hierarchical structure of taxonomy allows the model to evaluate the correctness of path selection and level selection separately and the parental relationships in a hierarchy not only allow the model to disambiguate the most similar terms but also enables the model to self-correct its level selection by deciding the current anchor’s granularity is too high or too low. We introduce the Fitting Score for the coherence evaluation of the expanded ego-tree by using a Pathfinder and a Stopper to score path correctness and level correctness, respectively. The Fitting Score calculation also disambiguates the most appropriate anchor from its parent and children and self-correct its level selection by bringing the level suggestion from the anchor’s parent and one of its children into consideration. The Fitting Score’s optimization adopts a self-supervised multi-task training paradigm for the Pathfinder and Stopper, which automatically generates training data from the seed taxonomy to utilize its information fully. We conduct extensive evaluations based on three benchmark datasets to compare our method with state-of-the-art baseline approaches. The results suggest that the proposed HEF model significantly surpasses the previous solutions on all three datasets by an average improvement of 46.7% in accuracy and 32.3% in mean reciprocal rank. A series of ablation studies further demonstrate that HEF can effectively perform the taxonomy expansion task. ## 2\. Related Work Taxonomy Construction. Taxonomy construction aims to create a tree-structured taxonomy with a set of terms (such as concepts and entities) from scratch, integrating hypernymy discovery and tree structure alignment. It can be further separated into two subdivisions. The first focuses on topic-based taxonomy, where each node is a cluster of several terms sharing the same topic (Zhang et al., 2018; Shang et al., 2020b). The other subdivision tackles the problem of term-based taxonomy construction, in which each node represents the term itself (Cocos et al., 2018; Shen et al., 2018; Mao et al., 2018). A typical pipeline for this task is to extract “is-A” relations with a hypernymy detection model first using either a pattern-based model (Hearst, 1992; Agichtein and Gravano, 2000; Jiang et al., 2017; Roller et al., 2018) or a distributional model (Lin, 1998; Yin and Roth, 2018; Wang et al., 2019; Dash et al., 2020), then integrate and prune the mined hypernym-hyponym pairs into a single directed acyclic graph (DAG) or tree (Gupta et al., 2017). More recent solutions utilize hyperbolic embeddings (Le et al., 2019) or transfer learning (Shang et al., 2020a) to boost performance. Taxonomy Expansion. In the taxonomy expansion task, an expert-curated seed taxonomy like MeSH (Lipscomb, 2000) is provided as both the guidance and the base for adding new terms. The taxonomy expansion task is a ranking task to maximize a score of a node and its ground-truth parent in the taxonomy. Wang et al. (Wang et al., 2014) adopted Dirichlet distribution to model the parental relations. ETF (Vedula et al., 2018) trained a learning-to-rank framework with handcrafted structural and semantic features. Arborist (Manzoor et al., 2020) calculated the ranking score in a bi-linear form and adopted margin ranking loss. TaxoExpan (Shen et al., 2020) modeled the anchor node by passing messages from its egonet instead of considering a single node, and scored by feeding a concatenation of egonet representation and query embedding to a feed-forward layer. STEAM (Yu et al., 2020) transformed the scoring task into a classification task on mini-paths and performed model ensemble of three sub-models processing distributional, contextual, and lexical-syntactic features, respectively. However, existing approaches mostly neglect the characteristics of taxonomy’s hierarchical structure and only evaluate the correctness of a single edge from anchor to query. On the contrary, our method utilizes the features and relations brought by the hierarchical structure and aims to enhance the expanded taxonomy’s overall coherence. Modeling of Tree-Structured Data. Taxonomy expansion involves modeling a tree or graph structure. Plenty of works have been devoted to extending recurrent models to tree structures, like Tree-LSTM (Tai et al., 2015). For explicit tree-structure modeling, previous approaches include modeling the likelihood of a Bayesian network (Fountain and Lapata, 2012; Wang et al., 2014) or using graph neural net variants (Shen et al., 2020; Yu et al., 2020). Recently, Transformers (Vaswani et al., 2017) achieved state-of-the-art performance in the program translation task by designing a novel positional encoding related to paths in the tree (Shiv and Quirk, 2019) or merely transforming a tree to sequence by traversing its nodes (Kim et al., 2020). In our work, we model tree-structure by a Transformer encoder, which, to the best of our knowledge, is the first to use the Transformer for taxonomy modeling. We adopt a more natural setting than (Shiv and Quirk, 2019) by using two different embeddings for a node’s absolute and relative level to denote positions. Figure 2. Illustration of the HEF model. Each circle denotes a seed node or a query node. The “Anchor’s child*” in Fitting Score calculation denotes the anchor’s child with maximum Pathfinder Score $S_{p}$. ## 3\. Problem Definition In this section, we provide the formal definition of the taxonomy expansion task and the explanation of key concepts that will occur in the following sections. Definition and Concepts about Taxonomy. A taxonomy $\mathcal{T}=(\mathcal{N},\mathcal{E})$ is an arborescence that presents hypernymy relations among a set of nodes. Each node $n\in\mathcal{N}$ represents a term, usually a concept mined from a large corpus online or an artificially extracted phrase. Each edge $\left<n_{p},n_{c}\right>\in\mathcal{E}$ points to a node from its most exact hypernym node, where $n_{p}$ is $n_{c}$’s parent node, and $n_{c}$ is $n_{p}$’s child node. Since hypernymy relation is transitive (Sang, 2007), such relation exists not only in node pairs connected by a single edge, but also in node pairs connected by a path in the taxonomy. Thus, for a node $n$ in the taxonomy, its hypernym set and hyponym set consists of its ancestors $\mathcal{A}_{n}$, and its descendants $\mathcal{D}_{n}$ respectively. Definition of the Taxonomy Expansion Task. Given a seed taxonomy $\mathcal{T}^{0}=(\mathcal{N}^{0},\mathcal{E}^{0})$ and the set of terms $\mathcal{C}$ to be added to the seed taxonomy, The model outputs the taxonomy $\mathcal{T}=(\mathcal{N}^{0}\cup\mathcal{C},\mathcal{E}^{0}\cup\mathcal{R})$, where $\mathcal{R}$ is the newly added relations from seed nodes in $\mathcal{N}^{0}$ to new terms in $\mathcal{C}$. More specifically, during the inference phase of a taxonomy expansion model, when given a query node $q\in\mathcal{C}$, the model finds its best-suited parent node by iterating each node in the seed taxonomy as an anchor node $a\in\mathcal{N}^{0}$, calculating a score $f(a,q)$ representing the suitability for adding the edge $\left<a,q\right>$, and deciding $q$’s parent $p_{q}$ in the taxonomy by $p_{q}=\mathop{\arg\max}_{a\in\mathcal{N}^{0}}{f(a,q)}$. Accessible External Resources. As a term’s surface name is usually insufficient to convey the semantic information for hypernymy relationship detection, previous research usually utilizes term definitions (Jurgens and Pilehvar, 2016; Shen et al., 2020) or related web pages (Wang et al., 2014; Kozareva and Hovy, 2010) to learn term representations. Besides, existing hypernymy detection solutions usually use large external corpora to discover lexical or syntactic patterns (Shwartz et al., 2016; Yu et al., 2020). As for the SemEval-2016 Task 13 datasets (Bordea et al., 2016) used for our model’s evaluation, utilizing the WordNet (Miller, 1995) definitions is allowed by the original task, which guarantees a fair comparison with previous solutions. ## 4\. The Hierarchy Expansion Framework In this section, we introduce the design of the Hierarchy Expansion Framework (HEF). An illustration of HEF is shown in Fig. 2. We first introduce the way HEF models the coherence of a tree structure, including two components for node pair hypernymy detection and ego-tree coherence modeling, respectively. Then, we discuss how HEF further exploits the hierarchical structure for multi-dimensional evaluation by the modules of Pathfinder and Stopper, and the self-supervised paradigm to train the model for the Fitting Score calculation. ### 4.1. Node Pair Hypernymy Detection We first introduce the hypernymy detection module of HEF, which detects the hypernymy relationships between two terms. Unlike previous approaches that manually design a set of classical lexical-syntactic features, we accomplish the task more directly and automatically by expanding the surface names of terms to their descriptions and utilizing pre-trained language models to represent the relationship between two terms. Given a seed term $n\in\mathcal{N}^{0}$ and a query term $q\in\mathcal{N}^{0}$ during training or $q\in\mathcal{C}$ during inference, the hypernymy detection module outputs a representation $r_{n,q}$ suggesting how well these two terms form a hypernymy relation. Note that $n$ might not be identical to the anchor $a$. Since the surface names of terms do not contain sufficient information for relation detection, we expand the surface names to their descriptions, enabling the model to better understand the semantic of new terms. We utilize the WordNet (Miller, 1995) concept definitions for completing this task. However, WordNet cannot explain all terms in a taxonomy due to its low coverage. Besides, many terms used in taxonomies are complex phrases like “adaptation to climate change” or “bacon lettuce tomato sandwich”. Therefore, we further develop a description generation algorithm descr($\cdot$), which generates meaningful and domain-related descriptions for a given term based on WordNet. Specifically, descr($\cdot$) is a dynamic programming algorithm that tends to integrate tokens into longer and explainable noun phrases. It describes each noun phrase by the most relative description to the taxonomy’s root’s surface name for domain relevance. The details are shown in Alg. 2 in the appendix. The input for hypernymy detection is organized as the input format of a Transformer: $D_{n,q}=\left[\mbox{<CLS>}\oplus\texttt{descr(}n\texttt{)}\oplus\mbox{<SEP>}\oplus\texttt{descr(}q\texttt{)}\oplus\mbox{<SEP>}\right],$ where $\oplus$ represents concatenation, ¡CLS¿ and ¡SEP¿ are the special token for classification and sentence separation in the Transformer architecture, respectively. As shown in Fig. 2, the hypernymy detection module utilizes a pre-trained DistilBERT (Sanh et al., 2020), a lightweight variant of BERT (Devlin et al., 2019), to learn the representations of cross-text relationships. Specifically, we first encode $D_{n,q}$ by ${\rm DistilBERT}(\cdot)$ with positional encoding. Then we take the final layer representation of ¡CLS¿ as the representation of the node pair $\left<n,q\right>$: $r_{n,q}={\rm DistilBERT}\left(D_{n,q}\right)\left[0\right],$ where index $0$ represents the position of ¡CLS¿’s embedding. ### 4.2. Ego-Tree Coherence Modeling We further design a coherence modeling module to evaluate the coherence of the tree structure after attaching the query term $q$ into taxonomy $\mathcal{T}$ as the anchor $a\in\mathcal{N}^{0}$’s child. There are two different aspects for considering a taxonomy’s coherence: i) the natural hypernymy relations. Since a node’s ancestors in the taxonomy all hold hypernymy relations with it, an explicit comparison among a node’s ancestors is needed to distinguish the most appropriate one; ii) the expert-curated designs, which act as supplement information for maintaining the overall structure. Some taxonomies contain latent rules about a node’s absolute or relative levels in a taxonomy. For example, in the biological taxonomy, kingdoms and species are all placed in the second and eighth levels, respectively; in some e-commerce catalog taxonomies, terms that are one level higher than another term contain exactly one more adjective. Hence, the coherence modeling module needs to: i) model a subtree with the query as a node in it, rather than a single node pair, enabling the model to learn the design of a complete hierarchy; ii) add tree-exclusive features like absolute level or relative level compared to the anchor to assist learning the expert- curated designs of the taxonomy. We design the Ego-tree $\mathcal{H}_{a}$, a novel contextual structure of an anchor $a$, which consists of all the ancestors and children of $a$ (see Fig. 2). This structure contains all relevant nodes to both anchor and query, enabling the model to both compare all hypernymy relations along the root path and detect similarity among query and its potential siblings with minimal computation cost: (1) $\mathcal{H}_{a}=\mathcal{A}_{a}\cup\left\\{a\right\\}\cup{\rm sample\\_child}\left(a\right),$ where $\mathcal{A}_{a}$ is all ancestors of $a$ in the seed taxonomy $\mathcal{T}^{0}$, and sample_child$(\cdot)$ means sampling at most three children of the anchor based on surface name similarity. The 3-children sampling is a trade-off between accuracy and speed, for three potential siblings are empirically enough for a comprehensive similarity comparison with the query (especially when these potential siblings are quite different) while decreasing the computation cost. Since this procedure is to leverage the similarity brought by a hierarchy’s sibling relations, sampling by surface name similarity is intuitive and cost-saving given that similar surface names usually indicate similar terms. The input of the coherence modeling module includes the anchor’s ego-tree $\mathcal{H}_{a}$ and the query $q$ as the anchor’s child in $\mathcal{H}_{a}$. For each node $n\in\mathcal{H}_{a}$, we represent the node pair $\left<n,q\right>$ by the following representations: * • Ego-tree representations. The ego-tree representation $r_{n,q}$ is the output of the hypernymy detection module described in Sec. 4.1. It suggests the node pair’s relation. * • Absolute level embedding. The absolute level embedding $l_{n,q}=\mbox{AbsLvlEmb}\left(d_{n}\right)$, where $d_{n}$ is the depth of $n$ in the expanded taxonomy. When $n=q$, $l_{q,q}=\mbox{AbsLvlEmb}\left(d_{a}+1\right)$. It assists the modeling of the expert-curated designs about granularity of a certain level. * • Relative level embedding. The relative level embedding $e_{n,q}=\mbox{RelLvlEmb}\left(d_{n}-d_{q}\right)$, where $d_{n}$ is the depth of $n$ in the expanded taxonomy. It assists the modeling of expert-curated designs about the cross-level comparison. * • Segment embedding. The segment embedding of $\left<n,q\right>$ $g_{n,q}=\mbox{SegEmb}\left(\mbox{segment}\left(n\right)\right)$ distinguishes anchor and query with other nodes in the ego-tree, where: $\mbox{segment}\left(n,q\right)=\begin{cases}0,&\mbox{if }n\mbox{ is the anchor},\\\ 1,&\mbox{if }n\mbox{ is the query},\\\ 2,&\mbox{otherwise}.\end{cases}$ The input of the coherence modeling module $R_{a,q}\in\mathbb{R}^{\left(\mathopen{|}\mathcal{H}_{a}\mathclose{|}+3\right)\times d}$ is the sum of the above embeddings calculated with the anchor’s ego-tree and the query, organized as the input of a Transformer: (2) $R_{a,q}=\left[e_{<CLS>}\oplus e_{<CLS>}\bigoplus_{n\in\mathcal{H}_{a}\cup{\left\\{q\right\\}}}{\left(r_{n,q}+l_{n,q}+e_{n,q}+g_{n,q}\right)}\right],$ where $d$ is the dimension of embedding, $e_{<CLS>}$ is a randomly initialized placeholder vector for obtaining the ego-tree’s path and level coherence representations, and $\oplus$ denotes concatenation. We implement the coherence modeling module using a Transformer encoder. Transformers are powerful to model sequences, but they lack positional information to process the relations among nodes in graphs. However, in a hierarchy like taxonomy, the level of nodes can be used as positional information, which simultaneously eliminates the positional difference of nodes on the same level. Transformers are also strong enough to integrate multiple-source information by adding their embeddings, thus they are quite suitable for modeling tree structures. In our HEF model, as shown in Fig. 2, by using two ¡CLS¿s in the module’s input, we can obtain two different representations: $p_{a,q}$ representing the coherence of hypernymy relations (whether the path is correct), and $d_{a,q}$ representing the coherence of inter-level granularity (whether the level is correct), evaluating how well the query fits the current position in the taxonomy in both horizontal and vertical perspective: $\displaystyle p_{a,q}$ $\displaystyle=\mbox{TransformerEncoder}\left(R_{a,q}\right)\left[0\right]$ $\displaystyle d_{a,q}$ $\displaystyle=\mbox{TransformerEncoder}\left(R_{a,q}\right)\left[1\right],$ where $0$ and $1$ are the position indexes of the two ¡CLS¿s. ### 4.3. Fitting Score-based Training and Inference Figure 3. An illustration of the self-supervision data labels for Pathfinder and Stopper. The two representations $p_{a,q}$ and $d_{a,q}$ need to be transformed into scores indicating the fitness of placing the query $q$ on a particular path and a particular level. Thus, we propose the Pathfinder for path selection and the Stopper for level selection, as well as a new self-supervised learning algorithm for training and the Fitting Score calculation for inference. Pathfinder. The Pathfinder detects whether the query is positioned on the right path. This module performs a binary classification using $p_{a,q}$. The Pathfinder Score $S_{p}=1$ if and only if $a$ and $q$ are on the same root- path: (3) $S_{p}\left(a,q\right)=\sigma\left(\mathbf{W}_{p2}\tanh\left(\mathbf{W}_{p1}p_{a,q}+b_{p1}\right)+b_{p2}\right),$ where $\sigma$ is the sigmoid function, and $\mathbf{W}_{p1},\mathbf{W}_{p2},b_{p1},b_{p2}$ are trainable parameters for multi-layer perceptrons. Stopper. The Stopper detects whether the query $q$ is placed on the right level, i.e., under the most appropriate anchor $a$ on a particular path. Selecting the right level is nonidentical to selecting the right path since levels are kept in order. The order of nodes on a path enables us to design a more representative module for further classifying whether the current level is too high (anchor $a$ is a coarse-grained ancestor of $q$) or too low ($a$ is a descendant of $q$). Thus, the Stopper module uses $d_{a,q}$ to perform a $3$-class classification: searching for a better anchor node needs to go Forward, remain Current, or go Backward, in the taxonomy: $\displaystyle[S_{f}\left(a,q\right),S_{c}\left(a,q\right),S_{b}\left(a,q\right)]=$ (4) $\displaystyle\mbox{softmax}\left(\mathbf{W}_{s2}\tanh\left(\mathbf{W}_{s1}d_{a,q}+b_{s1}\right)+b_{s2}\right),$ where $\mathbf{W}_{p1},\mathbf{W}_{p2},b_{p1},b_{p2}$ are trainable parameters for multi-layer perceptrons. Forward Score $S_{f}$, Current Score $S_{c}$ and Backward Score $S_{b}$ are called Stopper Scores. Self-Supervised Training. Training the HEF model needs data labels for both the Pathfinder and the Stopper. The tagging scheme is illustrated in Fig. 3. There are totally four kinds of Pathfinder-Stopper label combinations since Pathfinder Score is always $1$ when Stopper Tag is Forward or Current. The training process of HEF is shown in Alg. 1. Specifically, we sample the ego- tree of all four types of nodes for a query: $q$’s parent $a$, $a$’s ancestors, $a$’s descendants and other nodes, as a mini-batch for training the Pathfinder and Stopper simultaneously. The optimization of Pathfinder and Stopper can be regarded as a multi-task learning process. The loss $\mathcal{L}_{q}$ in Alg. 1 is a linear combination of the loss from Pathfinder and Stopper: $\displaystyle\mathcal{L}_{q}=$ $\displaystyle-\eta\frac{1}{\mathopen{|}\mathcal{X}_{q}\mathclose{|}}\sum_{a\in\mathcal{X}_{q}}{{\rm BCELoss}\left(\hat{S_{p}}\left(a,q\right),S_{p}\left(a,q\right)\right)}$ (5) $\displaystyle-\left(1-\eta\right)\frac{1}{\mathopen{|}\mathcal{X}_{q}\mathclose{|}}\sum_{a\in\mathcal{X}_{q}}{\sum_{k\in\left\\{f,c,b\right\\}}{\hat{s_{k}}\left(a,q\right)\log s_{k}\left(a,q\right)}},$ where ${\rm BCELoss}\left(\cdot\right)$ denotes the binary cross entropy, and $\eta$ is the weight of multi-task learning. Algorithm 1 Self-Supervised Training Process of HEF. 1:procedure TrainEpoch($\mathcal{T}^{0},{\Theta^{0}}$) 2: $\Theta\leftarrow\Theta^{0}$ 3: for $q\leftarrow\mathcal{N}^{0}-{\rm root}\left(\mathcal{T}^{0}\right)$ do $\triangleright$ Root is not used as query 4: $\mathcal{X}_{q}=\\{\\}$ $\triangleright$ Initialize anchor set 5: $p\leftarrow{\rm parent}\left(q\right)$ $\triangleright$ Reference node of labeling 6: $\mathcal{X}_{q}\leftarrow\mathcal{X}_{q}\cup\left\\{p\right\\}$ $\triangleright$ Ground Truth Parent: $S_{p}=1,S_{c}=1$ 7: $\mathcal{X}_{q}\leftarrow\mathcal{X}_{q}\cup{\rm sample}\left(\mathcal{A}_{p}\right)$ $\triangleright$ Ancestors: $S_{p}=1,S_{f}=1$ 8: $\mathcal{X}_{q}\leftarrow\mathcal{X}_{q}\cup{\rm sample}\left(\mathcal{D}_{p}\right)$ $\triangleright$ Descendants: $S_{p}=1,S_{b}=1$ 9: $\mathcal{X}_{q}\leftarrow\mathcal{X}_{q}\cup{\rm sample}\left(\mathcal{N}^{0}-\left\\{p\right\\}-\mathcal{A}_{p}-\mathcal{D}_{p}\right)$ 10:$\triangleright$ Other nodes: $S_{p}=0,S_{b}=1$ 11: for $a\leftarrow\mathcal{X}_{q}$ do 12: Compute $S_{p}\left(a,q\right)$ using Eqn. 3 13: Compute $S_{f}\left(a,q\right),S_{c}\left(a,q\right),S_{b}\left(a,q\right)$ using Eqn. 4.3 14: end for 15: Compute $\mathcal{L}_{q}$ with $S_{p},S_{f},S_{c},S_{b}$ using Eqn. 5 16: $\Theta\leftarrow{\rm optimize}\left(\Theta,\mathcal{L}_{q}\right)$ 17: end for 18: return $\Theta$ 19:end procedure Fitting Score-based Inference. During inference, evaluation of an anchor-query pair $\left<a,q\right>$ should consider both Pathfinder’s path evaluation and Stopper’s level evaluation. However, instead of merely using $S_{p}$ and $S_{c}$, the multi-classifying Stopper also enables the HEF model to disambiguate the most suited anchor from its neighbors (its direct parent and children) and self-correct its level prediction by exchanging scores with its neighbors to find the best position for maintaining the taxonomy’s coherence. Thus, We introduce the Fitting Score function during inference. For a new query term $q\in\mathcal{C}$, we first obtain the Pathfinder Scores and Stopper Scores of all node pairs $\left<a,q\right>,a\in\mathcal{N}^{0}$. For each anchor node $a$, we assign its Fitting Score by multiplying the following four items: * • $a$’s Pathfinder Score: $S_{p}\left(a,q\right)$, which suggests whether $a$ is on the right path. * • $a$’s parent’s Forward Score: $S_{f}\left({\rm parent}\left(a\right),q\right)$, which distinguishes $a$ and $a$’s parent, and rectifies $a$’s Current Score. When $a$ is the root node, we assign this item as a small number like $1e-4$ since the first level of taxonomy is likely to remain unchanged. * • $a$’s Current Score: $S_{c}\left(a,q\right)$, which suggests whether $a$ is on the right level. * • $a$’s child with maximum Pathfinder Score’s Backward Score: $S_{b}\left(c_{a}^{*},q\right),c_{a}^{*}=\mathop{\arg\max}_{c_{a}\in{\rm child}\left(a\right)}{S_{p}\left(c_{a},q\right)}$, which distinguishes $a$ and $a$’s children, and rectifies $a$’s Current Score. Since $a$ might have multiple children, we pick the child with max Pathfinder Score, for larger $S_{p}$ indicates a better hypernymy relation to $q$. When $a$ is a leaf node, we assign this item as the proportion of leaf nodes in the seed taxonomy to keep its overall design. The Fitting Score of $\left<a,q\right>$ is given by: (6) $F\left(a,q\right)=S_{p}\left(a,q\right)\cdot S_{f}\left({\rm parent}\left(a\right),q\right)\cdot S_{c}\left(a,q\right)\cdot S_{b}\left(c_{a}^{*},q\right)$ $c_{a}^{*}=\mathop{\arg\max}_{c_{a}\in{\rm child}\left(a\right)}{S_{p}\left(c_{a},q\right)}.$ The Fitting Score can be computed using ordered $S_{p},S_{f},S_{c},S_{b}$ arrays and the seed taxonomy’s adjacency matrix. Since a tree’s adjacency matrix is sparse, the time complexity of Fitting Score computation is low. After calculating the Fitting Scores between all seed nodes and the query, we select the seed node with the highest Fitting Score as the query’s parent in the expanded taxonomy: (7) ${\rm parent}\left(q\right)\coloneqq\mathop{\arg\max}_{a\in\mathcal{N}^{0}}F\left(a,q\right).$ ## 5\. Experiments In this section, we first introduce our experimental setups, including datasets, our implementation details, evaluation criteria, and a brief description of the compared baseline methods. Then, we provide extensive evaluation results for overall model performance, performance contribution brought by each design, and sensitivity analysis of the multi-task learning weight $\eta$ in Equation 5. In-depth visualizations of hypernymy detection and coherence modeling modules are provided to analyze the model’s inner behavior. We also provide a case study in the appendix.111The code will be available at https://github.com/sheryc/HEF. ### 5.1. Experimental Setup #### 5.1.1. Datasets We evaluate HEF on three public benchmark datasets retrieved from SemEval-2016 Task 13 (Bordea et al., 2016). This task contains three taxonomies in the domain of Environment (SemEval16-Env), Science (SemEval16-Sci), and Food (SemEval16-Food), respectively. The statistics of the benchmark datasets are provided in Table 1. Note that the original dataset may not form a tree. In this case, we use a spanning tree of the taxonomy instead of the original graph to match the problem definition. The pruning process only removes less than 6% of the total edges, keeping the taxonomy’s information and avoiding multiple ground truth parents for a single node. Since HEF and the compared baselines (Shen et al., 2020; Yu et al., 2020) are all limited to adding new terms without modifying the seed taxonomy, nodes in the test and validation set can only sample from leaf nodes to guarantee that the parents of test or validation nodes exist in the seed taxonomy. This is also the sampling strategy of TaxoExpan (Shen et al., 2020). Following the previous state-of-the-art model STEAM (Yu et al., 2020), we exclude 20% of the nodes in each dataset, of which ten nodes of each dataset are separated as the validation set for early stopping, and the rest as the test set. The nodes not included in the validation set and test set are seed nodes for self- supervision in the training phase and potential anchor nodes in the inference phase. Note that pruning the dataset does not affect the node count, thus the scale of the dataset remains identical to our baselines’ settings. Table 1. Statistics of datasets. $\left|N\right|$ and $\left|E_{O}\right|$ are the numbers of nodes and edges in the original datasets, respectively. $D$ is the depth of the taxonomy. We adopt the spanning tree of each dataset, and $\left|E\right|$ is the number of remaining edges. Dataset | $\left|N\right|$ | $\left|E_{O}\right|$ | $\left|E\right|$ | $D$ ---|---|---|---|--- SemEval16-Env | 261 | 261 | 260 | 6 SemEval16-Sci | 429 | 452 | 428 | 8 SemEval16-Food | 1486 | 1576 | 1485 | 8 Table 2. Comparison of the proposed method against state-of-the-art methods. All metrics are presented in percentages (%). Best results for each metric of each dataset are marked in bold. Reported performance is the average of three runs using different random seeds. The MRR of TAXI (Panchenko et al., 2016) is inaccessible since it outputs the whole taxonomy instead of node rankings. The performance of baseline methods are retrieved from (Yu et al., 2020). Dataset | SemEval16-Env | SemEval16-Sci | SemEval16-Food ---|---|---|--- Metric | Acc | MRR | Wu&P | Acc | MRR | Wu&P | Acc | MRR | Wu&P BERT+MLP | 11.1 | 21.5 | 47.9 | 11.5 | 15.7 | 43.6 | 10.5 | 14.9 | 47.0 TAXI (Panchenko et al., 2016) | 16.7 | - | 44.7 | 13.0 | - | 32.9 | 18.2 | - | 39.2 HypeNet (Shwartz et al., 2016) | 16.7 | 23.7 | 55.8 | 15.4 | 22.6 | 50.7 | 20.5 | 27.3 | 63.2 TaxoExpan (Shen et al., 2020) | 11.1 | 32.3 | 54.8 | 27.8 | 44.8 | 57.6 | 27.6 | 40.5 | 54.2 STEAM (Yu et al., 2020) | 36.1 | 46.9 | 69.6 | 36.5 | 48.3 | 68.2 | 34.2 | 43.4 | 67.0 HEF | 55.3 | 65.3 | 71.4 | 53.6 | 62.7 | 75.6 | 47.9 | 55.5 | 73.5 #### 5.1.2. Baselines for Comparison We compare our proposed HEF model with the following baseline approaches: * • BERT+MLP: This method utilizes BERT (Devlin et al., 2019) to perform hypernym detection. This model’s input is the term’s surface name, and the representation of BERT’s classification token $\langle$CLS$\rangle$ is fed into a feed-forward layer to score whether the first sequence is the ground- truth parent. * • HypeNet (Shwartz et al., 2016): HypeNet is an LSTM-based hypernym extraction model that scores a term pair by representing node paths in the dependency tree. * • TAXI (Panchenko et al., 2016): TAXI was the top solution of SemEval-2016 Task 13. It explicitly splits the task into a pipeline of hypernym detection using substring matching and pattern extraction, and hypernym pruning to avoid multiple parents. * • TaxoExpan (Shen et al., 2020): TaxoExpan is a self-supervised taxonomy expansion model. The anchor’s representation is modeled by a graph network of its Egonet with consideration of relative levels, and the parental relationship is scored by a feed-forward layer. BERT embedding is used as its input instead of the model’s original configuration. * • STEAM (Yu et al., 2020): STEAM is the state-of-the-art self-supervised taxonomy expansion model, which scores parental relations by ensembling three classifiers considering graph, contextual, and hand-crafted lexical-syntactic features, respectively. #### 5.1.3. Implementation Details In our setting, the coherence modeling module is a 3-layer, 6-head, 768-dimensional Transformer encoder initialized from Gaussian distribution $\mathcal{N}\left(0,0.02\right)$. The first hidden layers of Pathfinder and Stopper are both 300-dimensional. The input to the hypernymy detection module is either truncated or padded to a length of 64. Each training step contains a set of 32 query-ego-tree pairs of 32 query nodes using gradient accumulation, with each query-ego-tree pair set containing one ground-truth parent ($S_{p}=1$, $S_{c}=1$), at most 6 ground-truth parent’s ancestors ($S_{p}=1$, $S_{f}=1$), at most 8 ground-truth parent’s descendants ($S_{p}=1$, $S_{b}=1$), and at least 16 other nodes ($S_{p}=0$, $S_{b}=1$). The hyperparameters above are empirically set, since our algorithm is not sensitive to the setting of splits. Each dataset is trained for 150 epochs. In a single epoch, each seed node is trained as the query exactly once. AdamW (Loshchilov and Hutter, 2019) is used for optimization with $\epsilon$ set to $1\times 10^{-6}$. A linear warm-up is adopted with the learning rate linearly rise from 0 to 5e-5 in the first 10% of total training steps and linearly drop to 0 at the end of 150 epochs. The multi-task learning weight $\eta$ is set to 0.9. After each epoch, we validate the model and save the model with the best performance on the validation set. These hyperparameters are tuned on on SemEval16-Env’s validation set, and are used across all datasets and experiments unless specified in the ablation studies or sensitivity analysis. #### 5.1.4. Evaluation Metrics Assume $k\coloneqq\mathopen{|}\mathcal{C}\mathclose{|}$ to be the term count of the test set, $\left\\{p_{1},p_{2},\cdots,p_{k}\right\\}$ to be the predicted parents for test set queries, and $\left\\{\hat{p_{1}},\hat{p_{2}},\cdots,\hat{p_{k}}\right\\}$ to be the ground truth parents accordingly. Following the previous solutions (Manzoor et al., 2020; Vedula et al., 2018; Yu et al., 2020), we adopt the following three metrics as evaluation criteria. * • Accuracy (Acc): It measures the proportion that the predicted parent for each node in the test set exactly matches the ground truth parent: $\mbox{Acc}=\mbox{Hit@1}=\frac{1}{k}\sum_{i=1}^{k}{\mathbb{I}\left(p_{i}=\hat{p_{i}}\right)},$ where $\mathbb{I}(\cdot)$ denotes the indicator function, * • Mean Reciprocal Rank (MRR): It calculates the average reciprocal rank of each test node’s ground truth parent: $\mbox{MRR}=\frac{1}{k}\sum_{i=1}^{k}{\frac{1}{\mbox{rank}\left(\hat{p_{i}}\right)}},$ * • Wu & Palmer Similarity (Wu&P) (Wu and Palmer, 1994): It is a tree-based measurement that judges how close the predicted and ground truth parents are in the seed taxonomy: $\mbox{Wu\&P}=\frac{1}{k}\sum_{i=1}^{k}{\frac{2\times\mbox{depth}\left(\mbox{LCA}\left(p_{i},\hat{p_{i}}\right)\right)}{\mbox{depth}\left(p_{i}\right)+\mbox{depth}\left(\hat{p_{i}}\right)}},$ where $\mbox{depth}(\cdot)$ is the node’s depth in the seed taxonomy and $\mbox{LCA}(\cdot,\cdot)$ is the least common ancestor of two nodes. ### 5.2. Main Results The performance of HEF is shown in Table 2. HEF achieves the best performance on all datasets and surpasses previous state-of-the-art models with a significant improvement on all metrics. From the table, we get an overview of how taxonomy expansion models evolve chronologically. The solution of BERT+MLP does not utilize any structural and lexical-syntactic features of terms, and the insufficiency of information attributes to its poor results. Models of the first generation like TAXI and HypeNet utilize lexical, syntactic, or contextual information to achieve better results, mainly for the task’s hypernymy detection part. However, these two models do not utilize any of the structural information of taxonomy; hence they are unable to maintain the taxonomy’s structural design. Models of the second generation, like TaxoExpan and STEAM, inherit handcrafted lexical- syntactic features for detecting hypernymy relations. They also utilize the structural information by self-supervision from seed taxonomy and graph neural networks on small subgraphs of the taxonomy. However, they neglect the hierarchical structure of taxonomies, and they do not consider the coherence of the whole expanded taxonomy. Thus, their usage of structural information is only an improvement for performing hypernymy detection rather than taxonomy expansion. HEF further improves both the previous two generations’ strength by proposing a new approach that better fits the taxonomy expansion task. Moreover, it introduces a new goal for the task: to best preserve the taxonomy’s coherence after expansion. We propose the description generation algorithm to generate accurate and domain-specific descriptions for complex and rare terms, to incorporate lexical-syntactic features for hypernymy detection. Aided by DistilBERT’s power of sentence-pair representation, HEF can mine hypernymy features more automatically and accurately. HEF also aims to fully exploit the information brought by the taxonomy’s hierarchical structure to boost performance. HEF uses ego-trees to perform thorough comparison along root path and among siblings, injects tree-exclusive features to assist modeling the expert-curated taxonomy designs and explicitly evaluates both path and level for the anchor node as well as its parent and child. Experiment results suggest that these designs are capable of modeling and maximizing the coherence of taxonomy in different aspects, which results in a vast performance increase in the taxonomy expansion task. | | ---|---|--- (a) Accuracy on all 3 datasets. | (b) MRR on all 3 datasets. | (c) Wu&P on all 3 datasets. Figure 4. Sensitivity analysis of model performance under different multi-task learning weight $\eta$. ### 5.3. Ablation Studies Table 3. Ablation experiment results on the SemEval16-Env dataset. All metrics are presented in percentages (%). For each ablation experiment setting, only the best result is reported. Abl. Type | Setting | Acc | MRR | Wu&P ---|---|---|---|--- Original | HEF | 55.3 | 65.3 | 71.4 Dataflows | \- WordNet Descriptions | 41.5 | 55.3 | 62.6 \- Ego-tree + Egonet | 45.3 | 60.6 | 69.9 \- Relative Level Emb. | 49.1 | 59.2 | 60.9 \- Absolute Level Emb. | 49.1 | 60.6 | 68.4 Scoring Function | Stopper Only | 52.8 | 62.5 | 68.7 Pathfinder + Current Only | 50.9 | 62.1 | 66.8 Current Only | 41.5 | 54.7 | 58.6 We discuss how exploiting different characteristics of taxonomy’s hierarchical structure brings performance increase by a series of ablation studies. We substitute some designs of HEF in dataflow and score function to a vanilla setting and rerun the experiments. The results of the ablation studies are shown in Table 3. * • \- WordNet Descriptions: WordNet descriptions are substituted with the term’s surface name as the hypernymy detection module’s input. * • \- Ego-tree + Egonet: the Egonet from TaxoExpan (Shen et al., 2020) is used instead of the ego-tree for modeling the tree structure. * • \- Relative Level Emb.: The relative level embedding for the coherence modeling module is removed. * • \- Absolute Level Emb.: The absolute level embedding for the coherence modeling module is removed. * • Stopper Only: Only the Stopper Scores are used for Fitting Score calculation. More specifically, $\eta=0$, and the Fitting Score in Equation 6 becomes: $F\left(a,q\right)=S_{f}\left({\rm parent}\left(a\right),q\right)\cdot S_{c}\left(a,q\right)\cdot S_{b}\left(c_{a}^{*},q\right),$ $c_{a}^{*}=\mathop{\arg\max}_{c_{a}\in{\rm child}\left(a\right)}{S_{p}\left(c_{a},q\right)}.$ * • Pathfinder + Current Only: Only the Pathfinder Score and the Current Score are used for Fitting Score calculation. More specifically, the Fitting Score in Equation 6 and the loss in Equation 5 become: $F\left(a,q\right)=S_{p}\left(a,q\right)\cdot S_{c}\left(a,q\right),$ $\displaystyle\mathcal{L}_{q}=$ $\displaystyle-\eta\frac{1}{\mathopen{|}\mathcal{X}_{q}\mathclose{|}}\sum_{a\in\mathcal{X}_{q}}{{\rm BCELoss}\left(\hat{S_{p}}\left(a,q\right),S_{p}\left(a,q\right)\right)}$ $\displaystyle-\left(1-\eta\right)\frac{1}{\mathopen{|}\mathcal{X}_{q}\mathclose{|}}\sum_{a\in\mathcal{X}_{q}}{{\rm BCELoss}\left(\hat{S_{c}}\left(a,q\right),S_{c}\left(a,q\right)\right)}.$ * • Current Only: Only the Current Score is used for Fitting Score calculation. This is the scoring strategy identical to prior arts (Shen et al., 2020; Yu et al., 2020). More specifically, the Fitting Score in Equation 6 and the loss in Equation 5 become: $F\left(a,q\right)=S_{c}\left(a,q\right),$ $\mathcal{L}_{q}=-\frac{1}{\mathopen{|}\mathcal{X}_{q}\mathclose{|}}\sum_{a\in\mathcal{X}_{q}}{{\rm BCELoss}\left(\hat{S_{c}}\left(a,q\right),S_{c}\left(a,q\right)\right)}.$ We notice that by changing the design of dataflows, the performance of the HEF model suffers from various deteriorations. Substituting WordNet descriptions with a term’s surface name surprisingly remains a relatively high performance, which might attribute to the representation power of the DistilBERT model. Using Egonets rather than ego-trees for coherence modeling also affects the performance. Although egonets can capture the local structure of taxonomy, ego-trees are more capable of modeling the complete construction of a hierarchy. For the introduction of level embeddings, the results show that removing one of the two level embeddings for the coherence modeling module hurts the learning of taxonomy’s design. This is in accordance with the previous research about the importance of using the information of both absolute and relative positions in Transformers (Shaw et al., 2018) and confirms our assumption that taxonomies have intrinsic designs about both absolute and relative levels. Changes to the score function bring a smaller negative impact on the model compared to the dataflow changes, except for the setting of using merely Current Score. When using only the Current Score, the model loses the ability to disambiguate with its neighbors and the capacity of directly choosing the right path, downgrading the problem to be a series of independent scoring problems like the previous solutions. Adding Backward Score and Forward Score into Fitting Score calculation allows the model to distinguish the ground truth from its neighbors, bringing a boost to accuracy. However, without the Pathfinder, the “Stopper Only” setting only explicitly focuses on choosing the right level without considering the path and is inferior to the original HEF model. However, we observe that although changing several designs of dataflow or scoring function deteriorates the performance, our method can still surpass the previous state-of-the-art in Acc and MRR, suggesting that the HEF model introduces improvements in multiple aspects, which also testifies that maximizing the taxonomy’s coherence is a better goal for the taxonomy expansion task. Figure 5. Illustration of one self-attention head in the last layer of the hypernymy detection module, showing how the hypernymy detection module detects hypernymy relations. In this example, the seed node is “tea”, and the query is “oolong”. ### 5.4. Impact of Multi-Task Learning Weight $\eta$ In this section, we discuss the impact of $\eta$ in Equation 5 through a sensitivity analysis. Since $\eta$ controls the proportion of loss from the path-oriented Pathfinder and the level-oriented Stopper, this hyperparameter affects HEF’s tendency to prioritize path or level selection. The results on all three datasets are shown in Fig. 4. From the result, we notice that $\eta$ cannot be set too low, which means that explicit path selection contributes a lot to the model’s performance. This is in accordance with the fact that taxonomies are based on hypernymy relations and selecting the right path is the essential guidance for anchor selection. A better setting of $\eta$ is $\left[0.4,0.6\right]$. In this setting, the model tends to balance path and level selections, which results in better performance. Surprisingly, setting $\eta$ to a high value like 0.9 also brings a performance boost, and sometimes even achieves the best result. This phenomenon consistently exists when changing random seeds. However, setting $\eta$ to 1 means using merely Pathfinder, which cannot distinguish the ground truth from other nodes and breaks the model. This discovery further testifies the importance of explicitly evaluating path selection in the taxonomy expansion task. ### 5.5. Visualization of Self-Attentions #### 5.5.1. Node Pair Hypernymy Detection Module To illustrate how the hypernymy detection module works, we show the weights of one of the attention heads of the hypernymy detection module’s last Transformer encoder layer in Fig. 5. By expanding a term to its description, the model is capable of understanding the term “oolong” by its description, which cannot be achieved by constructing rule-based lexical-syntactic features since “oolong” and “tea” have no similarity in their surface names. Furthermore, by adopting the pretrained DistilBERT, the hypernymy detection module can also discover more latent patterns such as the relation between “leaves” and “tree”, allowing the model to discover more in-depth hypernymy relations. #### 5.5.2. Ego-Tree Coherence Modeling Module Figure 6. Illustration of one self-attention head in the first layer of the coherence modeling module, showing how the coherence modeling module finds the most fitted node in an ego-tree. In this example, the anchor is “herb”, the query is “oolong”, and the query’s ground truth parent is “tea”. To illustrate how the coherence modeling module compares the nodes in the ego- tree to maintain the taxonomy’s coherence, we present the weights of an attention head of the module’s first Transformer encoder layer in Fig. 6. Since the last layer’s attention mostly focuses on the anchor node (“herb”), the first layer can better illustrate the model’s comparison among ego-tree nodes. Based on our observation, the two ¡CLS¿s are capable of finding the best- suited parent node in the ego-tree even if it is not the anchor. Since the coherence modeling module utilizes ego-trees for hierarchy modeling, the coherence modeling module can compare a node with all its ancestors and its children to find the most suited anchor, which makes the model more robust. Besides, the coherence modeling module is also able to assign a lower attention weight to the best-suited parent’s parent node when it is on the right path, suggesting that the coherence modeling module can achieve both path-wise and level-wise comparison. ## 6\. Conclusion We proposed HEF, a self-supervised taxonomy expansion model that fully exploits the hierarchical structure of a taxonomy for better hierarchical structure modeling and taxonomy coherence maintenance. Compared to previous methods that evaluate the anchor by merely a new edge in a normal graph neglecting the tree structure of taxonomy, we used extensive experiments to prove that, evaluating a tree structure for coherence maintenance, and mining multiple tree-exclusive features in the taxonomy, including hypernymy relations from parent-child relations, term similarity from sibling relations, absolute and relative levels, path+level based multi-dimensional evaluation, and disambiguation based on parent-current-child chains all brought performance boost. This indicates the importance of using the information of tree structure for the taxonomy expansion task. We also proposed a framework for injecting these features, introduced our implementation of the framework, and surpassed the previous state-of-the-art. We believe that these novel designs and their motivations will not only benefit the taxonomy expansion task, but also be influential for all tasks involving hierarchical or tree structure modeling and evaluation. Future works include how to model and utilize these or new tree-exclusive features to boost other taxonomy-related tasks, and better implementation of each module in HEF. ###### Acknowledgements. Thanks to everyone who helped me with this paper in the Tencent Jarvis Lab, my family, and my loved one. ## References * (1) * Agichtein and Gravano (2000) Eugene Agichtein and Luis Gravano. 2000. Snowball: Extracting Relations From Large Plain-Text Collections. In _Proceedings of JCDL_. 85–94. * Bordea et al. (2016) Georgeta Bordea, Els Lefever, and Paul Buitelaar. 2016\. SemEval-2016 Task 13: Taxonomy Extraction Evaluation (TExEval-2). In _Proceedings of the SemEval-2016_. 1081–1091. * Cocos et al. (2018) Anne Cocos, Marianna Apidianaki, and Chris Callison-Burch. 2018\. Comparing Constraints for Taxonomic Organization. In _Proceedings of NAACL_. 323–333. * Dash et al. (2020) Sarthak Dash, Md Faisal Mahbub Chowdhury, Alfio Gliozzo, Nandana Mihindukulasooriya, and Nicolas Rodolfo Fauceglia. 2020\. Hypernym Detection Using Strict Partial Order Networks. In _Proceedings of AAAI_. 7626–7633. * Devlin et al. (2019) Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-Training of Deep Bidirectional Transformers for Language Understanding. In _Proceedings of NAACL_. 4171–4186. * Fountain and Lapata (2012) Trevor Fountain and Mirella Lapata. 2012. Taxonomy Induction Using Hierarchical Random Graphs. In _Proceedings of NAACL_. 466–476. * Gupta et al. (2017) Amit Gupta, Rémi Lebret, Hamza Harkous, and Karl Aberer. 2017\. Taxonomy Induction Using Hypernym Subsequences. In _Proceedings of CIKM_. 1329–1338. * Hearst (1992) Marti A. Hearst. 1992\. Automatic Acquisition of Hyponyms From Large Text Corpora. In _Proceedings of ACL_. 539–545. * Hua et al. (2017) Wen Hua, Zhongyuan Wang, Haixun Wang, Kai Zheng, and Xiaofang Zhou. 2017. Understand Short Texts by Harvesting and Analyzing Semantic Knowledge. _IEEE Transactions on Knowledge and Data Engineering_ (2017), 499–512. * Huang et al. (2019) Jin Huang, Zhaochun Ren, Wayne Xin Zhao, Gaole He, Ji-Rong Wen, and Daxiang Dong. 2019\. Taxonomy-Aware Multi-Hop Reasoning Networks for Sequential Recommendation. In _Proceedings of WSDM_. 573–581. * Jiang et al. (2017) Meng Jiang, Jingbo Shang, Taylor Cassidy, Xiang Ren, Lance M. Kaplan, Timothy P. Hanratty, and Jiawei Han. 2017. MetaPAD: Meta Pattern Discovery From Massive Text Corpora. In _Proceedings of KDD_. 877–886. * Jurgens and Pilehvar (2015) David Jurgens and Mohammad Taher Pilehvar. 2015. Reserating the Awesometastic: An Automatic Extension of the WordNet Taxonomy for Novel Terms. In _Proceedings of NAACL_. 1459–1465. * Jurgens and Pilehvar (2016) David Jurgens and Mohammad Taher Pilehvar. 2016. SemEval-2016 Task 14: Semantic Taxonomy Enrichment. In _Proceedings of the SemEval-2016_. 1092–1102. * Karamanolakis et al. (2020) Giannis Karamanolakis, Jun Ma, and Xin Luna Dong. 2020. TXtract: Taxonomy-Aware Knowledge Extraction for Thousands of Product Categories. (2020). * Kim et al. (2020) Seohyun Kim, Jinman Zhao, Yuchi Tian, and Satish Chandra. 2020\. Code Prediction by Feeding Trees to Transformers. (2020). * Kozareva and Hovy (2010) Zornitsa Kozareva and Eduard Hovy. 2010. A Semi-Supervised Method to Learn and Construct Taxonomies Using the Web. In _Proceedings of EMNLP_. 1110–1118. * Le et al. (2019) Matthew Le, Stephen Roller, Laetitia Papaxanthos, Douwe Kiela, and Maximilian Nickel. 2019\. Inferring Concept Hierarchies From Text Corpora via Hyperbolic Embeddings. In _Proceedings of ACL_. 3231–3241. * Lin (1998) Dekang Lin. 1998\. An Information-Theoretic Definition of Similarity. In _Proceedings of ICML_. 296–304. * Lipscomb (2000) Carolyn E. Lipscomb. 2000\. Medical Subject Headings (MeSH). _Bulletin of the Medical Library Association_ (2000), 265–266. * Liu et al. (2019) Bang Liu, Weidong Guo, Di Niu, Chaoyue Wang, Shunnan Xu, Jinghong Lin, Kunfeng Lai, and Yu Xu. 2019\. A User-Centered Concept Mining System for Query and Document Understanding at Tencent. In _Proceedings of KDD_. 1831–1841. * Loshchilov and Hutter (2019) Ilya Loshchilov and Frank Hutter. 2019. Decoupled Weight Decay Regularization. (2019). * Luo et al. (2020) Xusheng Luo, Luxin Liu, Yonghua Yang, Le Bo, Yuanpeng Cao, Jinghang Wu, Qiang Li, Keping Yang, and Kenny Q. Zhu. 2020. AliCoCo: Alibaba E-Commerce Cognitive Concept Net. In _Proceedings of SIGMOD_. 313–327. * Manzoor et al. (2020) Emaad Manzoor, Rui Li, Dhananjay Shrouty, and Jure Leskovec. 2020\. Expanding Taxonomies With Implicit Edge Semantics. In _Proceedings of TheWebConf_. 2044–2054. * Mao et al. (2018) Yuning Mao, Xiang Ren, Jiaming Shen, Xiaotao Gu, and Jiawei Han. 2018. End-To-End Reinforcement Learning for Automatic Taxonomy Induction. In _Proceedings of ACL_. 2462–2472. * Miller (1995) George A. Miller. 1995\. WordNet: A Lexical Database for English. _Commun. ACM_ (1995), 39–41. * Panchenko et al. (2016) Alexander Panchenko, Stefano Faralli, Eugen Ruppert, Steffen Remus, Hubert Naets, Cédrick Fairon, Simone Paolo Ponzetto, and Chris Biemann. 2016. TAXI at SemEval-2016 Task 13: A Taxonomy Induction Method Based on Lexico-Syntactic Patterns, Substrings and Focused Crawling. In _Proceedings of SemEval-2016_. 1320–1327. * Peng et al. (2019) Hao Peng, Jianxin Li, Senzhang Wang, Lihong Wang, Qiran Gong, Renyu Yang, Bo Li, Philip Yu, and Lifang He. 2019. Hierarchical Taxonomy-Aware and Attentional Graph Capsule RCNNs for Large-Scale Multi-Label Text Classification. _IEEE Transactions on Knowledge and Data Engineering_ (2019). * Roller et al. (2018) Stephen Roller, Douwe Kiela, and Maximilian Nickel. 2018\. Hearst Patterns Revisited: Automatic Hypernym Detection From Large Text Corpora. In _Proceedings of ACL_. 358–363. * Sang (2007) Erik Tjong Kim Sang. 2007\. Extracting Hypernym Pairs From the Web. In _Proceedings of ACL_. Association for Computational Linguistics, 165–168. * Sanh et al. (2020) Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. 2020. DistilBERT, a Distilled Version of BERT: Smaller, Faster, Cheaper and Lighter. (2020). * Shang et al. (2020a) Chao Shang, Sarthak Dash, Md. Faisal Mahbub Chowdhury, Nandana Mihindukulasooriya, and Alfio Gliozzo. 2020a. Taxonomy Construction of Unseen Domains via Graph-Based Cross-Domain Knowledge Transfer. In _Proceedings of ACL_. 2198–2208. * Shang et al. (2020b) Jingbo Shang, Xinyang Zhang, Liyuan Liu, Sha Li, and Jiawei Han. 2020b. NetTaxo: Automated Topic Taxonomy Construction From Text-Rich Network. In _Proceedings of TheWebConf_. 1908–1919. * Shaw et al. (2018) Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. 2018\. Self-Attention With Relative Position Representations. In _Proceedings of NAACL_. 464–468. * Shen et al. (2020) Jiaming Shen, Zhihong Shen, Chenyan Xiong, Chi Wang, Kuansan Wang, and Jiawei Han. 2020\. TaxoExpan: Self-Supervised Taxonomy Expansion With Position-Enhanced Graph Neural Network. In _Proceedings of TheWebConf_. 486–497. * Shen et al. (2018) Jiaming Shen, Zeqiu Wu, Dongming Lei, Chao Zhang, Xiang Ren, Michelle T. Vanni, Brian M. Sadler, and Jiawei Han. 2018\. HiExpan: Task-Guided Taxonomy Construction by Hierarchical Tree Expansion. In _Proceedings of KDD_. 2180–2189. * Shiv and Quirk (2019) Vighnesh Shiv and Chris Quirk. 2019. Novel Positional Encodings to Enable Tree-Based Transformers. In _Advances in Neural Information Processing Systems 32_. 12081–12091. * Shwartz et al. (2016) Vered Shwartz, Yoav Goldberg, and Ido Dagan. 2016\. Improving Hypernymy Detection With an Integrated Path-Based and Distributional Method. In _Proceedings of ACL_. Association for Computational Linguistics, 2389–2398. * Tai et al. (2015) Kai Sheng Tai, Richard Socher, and Christopher D. Manning. 2015\. Improved Semantic Representations From Tree-Structured Long Short-Term Memory Networks. In _Proceedings of ACL/IJCNLP_. 1556–1566. * Vaswani et al. (2017) Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017\. Attention Is All You Need. In _Advances in Neural Information Processing Systems 30_. 5998–6008. * Vedula et al. (2018) Nikhita Vedula, Patrick K. Nicholson, Deepak Ajwani, Sourav Dutta, Alessandra Sala, and Srinivasan Parthasarathy. 2018. Enriching Taxonomies With Functional Domain Knowledge. In _Proceedings of SIGIR_. 745–754. * Velardi et al. (2013) Paola Velardi, Stefano Faralli, and Roberto Navigli. 2013\. OntoLearn Reloaded: A Graph-Based Algorithm for Taxonomy Induction. _Computational Linguistics_ (2013), 665–707. * Wang et al. (2019) Chengyu Wang, Yan Fan, Xiaofeng He, and Aoying Zhou. 2019\. A Family of Fuzzy Orthogonal Projection Models for Monolingual and Cross-Lingual Hypernymy Prediction. In _Proceedings of WWW_. 1965–1976. * Wang et al. (2014) Jingjing Wang, Changsung Kang, Yi Chang, and Jiawei Han. 2014\. A Hierarchical Dirichlet Model for Taxonomy Expansion for Search Engines. In _Proceedings of WWW_. 961–970. * Wu and Palmer (1994) Zhibiao Wu and Martha Palmer. 1994. Verbs Semantics and Lexical Selection. In _Proceedings of ACL_. 133–138. * Yin and Roth (2018) Wenpeng Yin and Dan Roth. 2018. Term Definitions Help Hypernymy Detection. In _Proceedings of *SEM_. 203–213. * Yin and Shah (2010) Xiaoxin Yin and Sarthak Shah. 2010. Building Taxonomy of Web Search Intents for Name Entity Queries. In _Proceedings of WWW_. 1001–1010. * Yu et al. (2020) Yue Yu, Yinghao Li, Jiaming Shen, Hao Feng, Jimeng Sun, and Chao Zhang. 2020\. STEAM: Self-Supervised Taxonomy Expansion With Mini-Paths. In _Proceedings of KDD_. 1026–1035. * Zhang et al. (2018) Chao Zhang, Fangbo Tao, Xiusi Chen, Jiaming Shen, Meng Jiang, Brian Sadler, Michelle Vanni, and Jiawei Han. 2018\. TaxoGen: Unsupervised Topic Taxonomy Construction by Adaptive Term Embedding and Clustering. In _Proceedings of KDD_. 2701–2709. ## Appendix A Case Study To understand how different Fitting Score components contribute to HEF’s performance, we conduct a case study on the SemEval16-Food dataset and show the detailed results in Table 4. The first two rows of Table 4 shows two cases where HEF successfully predicts the query’s parent. We can see that the Pathfinder Score and the three Stopper Scores all contribute to the correct selection, which testifies the effectiveness of the Fitting Score design. The last two rows of Table 4 provide situations when HEF fails to select the correct parent. In the third row, “bourguignon” is described as “reduced red wine”, thus the model attaches it to the node “wine”. However, “bourguignon” is also a sauce for cooking beef. Such ambiguation consequently affects the meaning of a term by assigning an incorrect description, which hurts the model’s performance. In the last row, although “hot fudge sauce”’s description contains ”chocolate sauce”, the node “chocolate sauce” still gets a low Current Score. In HEF, the design of Stopper Scores enables the model to self- correct the occasionally wrong Current Scores by assigning larger Forward Score from a node’s parent and larger Backward Score from one of the node’s children. However, since “chocolate sauce” is a leaf node, its child’s Backward Score is assigned to be the proportion of leaf nodes in the seed taxonomy, which is 0.07 in the SemEval16-Food dataset. This indicates that future work includes designing a more reasonable Backward Score function for leaf nodes to improve the model’s robustness. ## Appendix B Description Generation Algorithm Algorithm 2 shows the description generation algorithm descr($\cdot$) used in HEF’s hypernymy detection module. descr($\cdot$) utilizes WordNet descriptions to generate domain-related term descriptions by dynamic programming. In this algorithm, WordNetNounDescr($\cdot$) means the set of a concept’s noun descriptions from WordNet (Miller, 1995), and CosSimilarity($t,n_{\mbox{root}}$) means calculating the average token cosine similarity of word vectors between a candidate description $t$ and the surface name of a taxonomy’s root term $n_{\mbox{root}}$. Algorithm 2 Description Generation Algorithm for the Hypernymy Detection Module 1:procedure Descr($n$) $\triangleright$ Input: term $n$ 2: $N\leftarrow$split($n$) 3: for $i\leftarrow 0,\cdots,\texttt{length($N$)}$ do 4: $S[i]=0$ $\triangleright$ Initialize score array 5: $C[i]=0$ $\triangleright$ Initialize splitting positions 6: end for 7: for $i\leftarrow 0,\cdots,\texttt{length($N$)$-1$}$ do 8: for $j\leftarrow 0,\cdots,i$ do 9: if WordNetNounDescr($N[j:i+1]$)$>0$ then 10: $s_{ij}=\left(i-j+1\right)^{2}+1$ $\triangleright$ Prefer longer concepts 11: else 12: $s_{ij}=1$ 13: end if 14: if $S[j]+s_{ij}>S[i+1]$ then 15: $S[i+1]\leftarrow S[j]+s_{ij}$ $\triangleright$ Save max score 16: $C[i]=j$ $\triangleright$ Save splitting position 17: end if 18: end for 19: end for 20: $D\leftarrow$“” $\triangleright$ Initialize description 21: $p\leftarrow\texttt{length(}N\texttt{)}$ $\triangleright$ Generate split pointer 22: while $p\neq-1$ do 23: $D_{WN}=$WordNetNounDescr($N\left[C[p]:p+1\right]$) 24: if len($D_{WN}$)$>0$ then $\triangleright$ Noun or noun phrase 25: $d\leftarrow\mathop{\arg\max}_{t\in D_{WN}}{{\texttt{CosSimilarity(}}t,n_{\mbox{root}}{\texttt{)}}}$ 26: else$\triangleright$ Prep. or adj. 27: $d\leftarrow{\texttt{j}oin(}N[C[p]:p+1]\texttt{)}$ 28: end if 29: $D\leftarrow d+D$ $\triangleright$ Put new description in the front 30: $p\leftarrow C[p]-1$ $\triangleright$ Go to next split 31: end while 32:end procedure Table 4. Examples of HEF’s prediction, with detailed Fitting Score composition and comparison between the ground truth and the predicted parent. Scores in this table correspond to the node in the same tabular cell with the score. Query ($q$) | Ground Truth ($\hat{p}$) | Scores | Prediction ($p$) | Scores ---|---|---|---|--- $q$: paddy is rice in the husk either gathered or still in the field | $\hat{p}$: rice is grains used as food either unpolished or more often polished | $S_{p}=0.9997$ | $p$: rice is grains used as food either unpolished or more often polished | $S_{p}=0.9997$ | | $S_{c}=0.4599$ | | $S_{c}=0.4599$ | ${\rm parent}(\hat{p})$: starches is a commercial preparation of starch that is used to stiffen textile fabrics in laundering | $S_{f}=0.9755$ | ${\rm parent}(p)$: starches is a commercial preparation of starch that is used to stiffen textile fabrics in laundering | $S_{f}=0.9755$ $F(\hat{p},q)=0.4483$ | $c_{\hat{p}}^{*}$: white rice is having husk or outer brown layers removed | $S_{b}=0.9995$ | $c_{p}^{*}$: white rice is having husk or outer brown layers removed | $S_{b}=0.9995$ $\hat{p}$’s Ranking: 1 | | | | $q$: fish meal is ground dried fish used as fertilizer and as feed for domestic livestock | $\hat{p}$: feed is food for domestic livestock | $S_{p}=0.9993$ | $p$: feed is food for domestic livestock | $S_{p}=0.9993$ | | $S_{c}=0.3169$ | | $S_{c}=0.3169$ | ${\rm parent}(\hat{p})$: food is any substance that can be metabolized by an animal to give energy and build tissue | $S_{f}=0.9984$ | ${\rm parent}(p)$: food is any substance that can be metabolized by an animal to give energy and build tissue | $S_{f}=0.9984$ $F(\hat{p},q)=0.3158$ | $c_{\hat{p}}^{*}$: mash is mixture of ground animal feeds | $S_{b}=0.9988$ | $c_{p}^{*}$: mash is mixture of ground animal feeds | $S_{b}=0.9988$ $\hat{p}$’s Ranking: 1 | | | | $q$: bourguignon is reduced red wine with onions and parsley and thyme and butter | $\hat{p}$: sauce is flavorful relish or dressing or topping served as an accompaniment$\cdots$ | $S_{p}=0.0002$ | $p$: wine is a red as dark as red wine | $S_{p}=0.9997$ | | $S_{c}=0.0001$ | | $S_{c}=0.1399$ | ${\rm parent}(\hat{p})$: condiment is a preparation (a sauce or relish or spice) to enhance flavor or enjoyment | $S_{f}=0.0004$ | ${\rm parent}(p)$: alcohol is any of a series of volatile hydroxyl compounds that are made from hydrocarbons by distillation | $S_{f}=0.9812$ $F(\hat{p},q)=1e-11$ | $c_{\hat{p}}^{*}$: bercy is butter creamed with white wine and shallots and parsley | $S_{b}=0.9997$ | $c_{p}^{*}$: red wine is wine having a red color derived from skins of dark-colored grapes | $S_{b}=0.8784$ $\hat{p}$’s Ranking: 328 | | | | $q$: hot fudge sauce is hot thick chocolate sauce served hot | $\hat{p}$: chocolate sauce is sauce made with unsweetened chocolate or cocoa$\cdots$ | $S_{p}=0.9471$ | $p$: sauce is flavorful relish or dressing or topping served as an accompaniment$\cdots$ | $S_{p}=0.9995$ | | $S_{c}=9e-5$ | | $S_{c}=0.0172$ | ${\rm parent}(\hat{p})$: sauce is flavorful relish or dressing or topping served as an accompaniment$\cdots$ | $S_{f}=0.9617$ | ${\rm parent}(p)$: condiment is a preparation (a sauce or relish or spice) to enhance flavor or enjoyment | $S_{f}=0.9888$ $F(\hat{p},q)=6e-6$ | $c_{\hat{p}}^{*}$: None | $S_{b}=0.0700$ | $c_{p}^{*}$: lyonnaise sauce is brown sauce with sauteed chopped onions and parsley$\cdots$ | $S_{b}=0.9995$ $\hat{p}$’s Ranking: 20 | | | |
# Xova: Baseline-Dependent Time and Channel Averaging for Radio Interferometry Marcellin Atemkeng1, Simon Perkins2, Jonathan Kenyon1, Benjamin Hugo2,1, and Oleg Smirnov1,2 ###### Abstract Xova is a software package that implements baseline-dependent time and channel averaging on Measurement Set data. The $uv$-samples along a baseline track are aggregated into a bin until a specified decorrelation tolerance is exceeded. The degree of decorrelation in the bin correspondingly determines the amount of channel and timeslot averaging that is suitable for samples in the bin. This necessarily implies that the number of channels and timeslots varies per bin and the output data loses the rectilinear input shape of the input data. 1Rhodes University, Makhanda (Grahamstown), Eastern Cape, South Africa 2South African Radio Astronomy Observatory, Cape Town, Western Cape, South Africa ## 1 Effects of Time and Channel Averaging Consider $\mathcal{V}_{pq}=\mathcal{V}(\mathbf{u}_{pq}(t,\nu))$ as the visibility sampled by the baseline $pq$ at time $t$ and frequency $\nu$. An interferometer is non-ideal in the sense that the measured visibility is the average of this sampled visibility over the sampling bin, $B_{kr}^{[\Delta t\Delta\nu]}=[t_{k}-\Delta t/2,t_{k}+\Delta t/2]\times[\nu_{r}-\Delta\nu/2,\nu_{r}+\Delta\nu/2]$: $\widetilde{\mathcal{V}}_{pq}=\frac{1}{\Delta t\Delta\nu}\iint\limits_{B_{kr}^{[\Delta t\Delta\nu]}}\mathcal{V}(\mathbf{u}_{pq}(t,\nu))\text{d}t\text{d}\nu,$ (1) where $\Delta t$ and $\Delta\nu$ are the integration intervals. If $\Pi_{pq}$ represents a normalized 2D top-hat window for a baseline $pq$ then Eq. (1) is equivalent to the infinitesimal integral: $\widetilde{\mathcal{V}}_{pq}=\iint\limits_{\infty}\Pi_{pq}(t-t_{k},\nu-\nu_{r})\mathcal{V}_{pq}(t,\nu)\text{d}t\text{d}\nu,$ (2) which is a convolution in the Fourier space, i.e.: $\displaystyle\widetilde{\mathcal{V}}_{pq}$ $\displaystyle=[\Pi_{pq}\circ\mathcal{V}_{pq}](\mathbf{u}_{pq}(t_{k},\nu_{r}))$ (3) $\displaystyle=\delta_{pqkr}[\Pi_{pq}\circ\mathcal{V}_{pq}],$ (4) where $\delta_{pqkr}(\mathbf{u})=\delta(\mathbf{u}-\mathbf{u}_{pqkr})$ is the Delta function shifted to the sampled point $pqkr$. For an observation with frequency range $F=\Delta\nu\times N_{\nu}$ and total observing period $T=\Delta t\times N_{t}$, observing for long times and large frequency ranges leads to storage issues, as well as computation cost since $\\{T,F\\}\propto\\{N_{t},N_{\nu}\\}$ if $\Delta t$ and $\Delta\nu$ most remain sufficiently small. For aggressive averaging, $\Delta t$ and $\Delta\nu$ are large which makes $\Pi_{pq}$ to deviate significantly from $\delta_{pqkr}$ and therefore causes the visibility to decorrelate: $\mathcal{V}_{pq}\neq\widetilde{\mathcal{V}}_{pq}$. To derive the effect of averaging on the image, we can reformulate Eq. 4 as: $\displaystyle\widetilde{\mathcal{V}}_{pq}$ $\displaystyle=\mathcal{F}\\{\mathcal{P}_{pqkr}\\}\big{(}\Pi_{pq}\circ\mathcal{F}\\{\mathcal{I}\\}\big{)},$ (5) where the apparent sky $\mathcal{I}$ is the inverse Fourier transform of $\mathcal{V}_{pq}$ and $\mathcal{P}_{pqkr}$ is the inverse Fourier transform of $\delta_{pqkr}$. Here $\mathcal{F}$ represents the Fourier transform. Inverting the sum of Eq. 5 over all the baselines results in an estimate of the sky image: $\displaystyle\widetilde{\mathcal{I}}$ $\displaystyle=\sum_{pqkr}W_{pqkr}\mathcal{P}_{pqkr}\circ\big{(}\mathcal{D}_{pqkr}\mathcal{I}\big{)},$ (6) where $W_{pqkr}$ is the weight at the sampled point $pqkr$. We note that the apparent sky $\mathcal{I}$ is now tapered by the baseline-dependent distortion distribution $\mathcal{D}_{pqkr}$, the latter being the inverse Fourier transform of the baseline-dependent top-hat window: $\displaystyle\mathcal{D}_{pqkr}$ $\displaystyle=\mathcal{F}^{-1}\\{\Pi_{pq}\\}$ $\displaystyle=\mathrm{sinc}\left(\frac{\Delta\Psi}{2}\right)\mathrm{sinc}\left(\frac{\Delta\Phi}{2}\right).$ (7) For a source at the sky location $\mathbf{l}$, the $\Delta\Psi$ and $\Delta\Phi$ are the phase difference in time and frequency, respectively: $\displaystyle\Delta\Psi$ $\displaystyle=2\pi\Delta\mathbf{u}_{pq}(t,\nu_{r})\mathbf{l};\Delta\Phi$ $\displaystyle=2\pi\Delta\mathbf{u}_{pq}(t_{k},\nu)\mathbf{l}.$ (8) Assuming no other corruption effects apart from decorrelation and assuming naturally weighting a sky with a single source; with decorrelation in effect Eq. 6 becomes: $\displaystyle\widetilde{\mathcal{I}}$ $\displaystyle=\mathcal{P}_{pqkr}\circ\big{(}\mathcal{D}_{pqkr}\mathcal{I}\big{)}$ $\displaystyle=\mathcal{D}_{pqkr}\mathcal{I}.$ (9) Note that in this formulation, we have assumed that $\mathcal{P}_{pqkr}=\delta(\mathbf{l})$ is baseline independent as opposed to Atemkeng et al. (2020). Eq. 9 is simulated in Figure 1 using the MeerKAT telescope at 1.4 GHz showing the apparent intensity of a 1 Jy source, as seen by the shortest baseline, medium-length baseline, and longest baseline, as a function of distance from the phase center. We see that decorrelation is severe on the longest baseline then followed by the medium length baseline, and that decorrelation is a function of source position in the sky. Figure 1.: Effect of time averaging: The data is sampled at 1 s and 84 kHz frequency resolutions then averaged only in time across 15 s. ## 2 Baseline-Dependent Time and Channel Averaging (BDA) The distortion distribution, $\mathcal{D}_{pqkr}$ depends on each baseline and its rotation orientation in the Fourier space which makes the decorrelation to be baseline-dependent. For decorrelation to be baseline-independent, the rectangular sampling bin $B_{kr}^{[\Delta t\Delta\nu]}$ across which the data is averaged must be kept baseline-dependent as opposed to the fixed sampling bin currently employed in radio interferometer correlators: $\displaystyle B_{kr}^{[\Delta_{\mathbf{u}_{pq}}t\Delta_{\mathbf{u}_{pq}}\nu]}=[t_{k}-\Delta_{\mathbf{u}_{pq}}t/2,t_{k}+\Delta_{\mathbf{u}_{pq}}t/2]\times[\nu_{r}-\Delta_{\mathbf{u}_{pq}}\nu/2,\nu_{r}+\Delta_{\mathbf{u}_{pq}}\nu/2],$ (10) where the integration intervals $\Delta_{\mathbf{u}_{pq}}t$ and $\Delta_{\mathbf{u}_{pq}}\nu$ are now also baseline-dependant. In this case Eq. 6 becomes: $\displaystyle\widetilde{\mathcal{I}}$ $\displaystyle=\sum_{pqkr}\mathcal{W}_{pqkr}\mathcal{P}_{pqkr}\circ\big{(}\mathcal{D}\mathcal{I}\big{)},$ (11) where $\mathcal{D}=\mathcal{D}_{pqkr}=\mathcal{D}_{\alpha\beta kr}$ is the distortion distribution, which is now equal across all the baselines, $pq$ and $\alpha\beta$ no matter their orientation. We provide details on the implementation of $\mathcal{D}$ in Sections 3 and 4. ## 3 Technologies The core of Xova’s BDA algorithm is implemented using two recent parallelisation and acceleration frameworks: (1) Dask (Rocklin 2015) is a Python parallel computing library that expresses programs as Computational Graphs whose individual tasks are scheduled on multiple cores or nodes. Dask collections abstract underlying graphs with familiar Array and Dataframe interfaces. (2) Numba (Lam et al. 2015) a JIT compiler that translates the BDA algorithm, expressed as a subset of Python and NumPy code, to accelerated machine code. These are implemented in Xova as follows: dask-ms (Perkins et al. 2021) exposes Measurement Set columns as dask arrays for ingest by Xova then Codex Africanus (Perkins et al. 2021) a Radio Astronomy Algorithms Library, applies BDA, implemented in numba to dask arrays, producing averaged dask arrays and dask-ms writes the averaged dask arrays to a new Measurement Set. ## 4 Xova Figure 2.: The parts of the baseline closer to the phase centre are subject to greater averaging For each baseline (See Figure 2): Measurement Set timeslots are aggregated into averaging bins until $\textrm{sinc}\left(\Delta\Psi/2\right)$ falls below decorrelation tolerance $\mathcal{D}$. The acceptable corresponding change in frequency $\Delta\Phi=2\,\textrm{sinc}^{-1}\left(\mathcal{D}/\textrm{sinc}\left(\Delta\Psi/2\right)\right)$ is calculated and channel width $\Delta\nu$ is derived from $\Delta\Phi$ and used to divide the original band into a new channelisation. ## 5 Results Figure 3 shows the image of a high-resolution data set imaged without BDA (right panel) and with $95\%$ decorrelation tolerance BDA (left panel). We note that BDA does not distort the image when compared to the no BDA image. Figure 3.: BDA (left) vs. no BDA (right). ### Acknowledgments The research of Oleg Smirnov is supported by the South African Research Chairs Initiative of the Department of Science and Technology and National Research Foundation. ## References * Atemkeng et al. (2020) Atemkeng, M., Smirnov, O., Tasse, C., Foster, G., & Makhathini, S. 2020, Monthly Notices of the Royal Astronomical Society, 499, 292 * Lam et al. (2015) Lam, S. K., Pitrou, A., & Seibert, S. 2015, in Proceedings of the Second Workshop on the LLVM Compiler Infrastructure in HPC (New York, NY, USA: Association for Computing Machinery), LLVM ’15. URL https://doi.org/10.1145/2833157.2833162 * Perkins et al. (2021) Perkins, S. J., et al. 2021, in ADASS XXX, edited by J.-E. Ruiz, & F. Pierfederici (San Francisco: ASP), vol. TBD of ASP Conf. Ser., 999 TBD * Rocklin (2015) Rocklin, M. 2015, in Proceedings of the 14th Python in Science Conference, edited by K. Huff, & J. Bergstra, 130
# Response to Comment on: Tunneling in DNA with Spin Orbit coupling Solmar Varela Yachay Tech University, School of Chemical Sciences & Engineering, 100119-Urcuquí, Ecuador Yachay Tech University, School of Physical Sciences & Nanotechnology, 100119-Urcuquí, Ecuador Iskra Zambrano Yachay Tech University, School of Physical Sciences & Nanotechnology, 100119-Urcuquí, Ecuador Bertrand Berche Laboratoire de Physique et Chimie Théoriques, UMR Université de Lorraine-CNRS 7019 54506 Vandoeuvre les Nancy, France Vladimiro Mujica School of Molecular Sciences, Arizona State University, Tempe, Arizona 85287-1604, USA Ernesto Medina Yachay Tech University, School of Physical Sciences & Nanotechnology, 100119-Urcuquí, Ecuador Centro de Física, Instituto Venezolano de Investigaciones Cíentificas (IVIC), Apartado 21827, Caracas 1020 A, Venezuela ###### Abstract The comment in ref.[WohlmanAharony, ] makes a few points related to the validity of our model, especially in the light of the interpretation of Bardarson’s theorem: “in the presence of time reversal symmetry and for half- integral spin the transmission eigenvalues of the two terminal scattering matrix come in (Kramers) degenerate pairs”. The authors of ref.[WohlmanAharony, ] first propose an ansatz for the wave function in the spin active region and go on to show that the resulting transmission does not show spin dependence, reasoning that spin dependence would violate Bardarson’s assertion. Here we clearly show that the ansatz presented assumes spin- momentum independence from the outset and thus just addresses the spinless particle problem. We then find the appropriate eigenfunction contemplating spin-momentum coupling and show that the resulting spectrum obeys Bardarson’s theorem. Finally we show that the allowed wavevectors are the ones assumed in the original paper and thus the original conclusions follow. We recognize that the Hamiltonian in our paper written in local coordinates on a helix was deceptively simple and offer the expressions of how it should be written to more overtly convey the physics involved. The relation between spin polarization and torque becomes clear, as described in reference ref.[VarelaZambrano, ]. This response is a very important clarification in relation to the implications of Bardarson’s theorem concerning the possibility of spin polarization in one dimensional systems in the linear regime. In ref.[WohlmanAharony, ] Aharony et al discuss critical points of our model that allow an opportunity to clarify fine points about notions that have been pointed out in the literature regarding the possibility of spin polarization in one dimensional systems in the linear regimeBart ; BalseiroAharony . Ref. [1] begins to formulate an ansatz for the solution of our Eq.(5) ${\cal H}=\left[\frac{p_{x}^{2}}{2m}+V_{0}\right]{\bf 1}+\alpha\sigma_{y}p_{x},$ (1) by assuming a product wave function in basis of eigenspinors of $\sigma_{y}$ where $|\Psi_{\mu}(x)=\psi_{u}(x)|\mu\rangle$. Assuming $\psi_{u}(x)\propto e^{iQ_{\mu}x}$ they arrive at the energy $E=\frac{(Q_{\mu}+k_{so}\mu)^{2}-k_{so}^{2}}{2m}+V_{0},$ (2) with $Q_{\mu}^{\pm}=-k_{so}\mu\pm q~{}{\rm with}~{}q=\sqrt{k^{2}+k_{so}^{2}-q_{0}^{2}},$ (3) where $q_{0}=2m/\hbar^{2}$, $k_{so}=m\alpha/\hbar$ and $k^{2}=2mE/\hbar^{2}$. The aforementioned proposal clearly leads to the spinless particle solution, by substituting the proposed $Q_{\mu}$ into the energy, eliminates all dependence on $\mu$. This point makes their transmission computation redundant; they obtain the spinless particle transmission, not surprisingly independent of spin. Furthermore, the solution apparently satisfies Bardarson’s theorem since any spin orientation gives the same energy independently of the direction of propagation of the electron. The latter observation gives a first hint of what has been omitted from this solution which can be drawn from the Hamiltonian above i.e. that there are two sets of Kramers pairs with two different energies (which obey Bardarson’s conclusions). Thus, spin and momentum are coupled and the product solution is not forthcoming. The eigenfunction to Eq.1 is $\Psi_{s}=\left(\begin{array}[]{c}is\\\ 1\end{array}\right)e^{i\lambda|q|x},$ (4) where $s=\pm 1$ which denotes the two possible spin orientations (in $\sigma_{z}$ basis) and $\lambda=\pm 1$ the two momentum orientations (see ref.Birkholz, ). Substitution of this vector into the eigenvalue equation yields the following energy eigenvalues $E_{s}^{\lambda}=\frac{\hbar^{2}q^{2}}{2m}-s\lambda\hbar\alpha|q|+V_{0},$ (5) that now reflect two Kramers pairs $(s,\lambda)=(+,+)~{}{\rm and}~{}(-,-)$ with energy $E_{<}=\hbar^{2}q^{2}/2m-\hbar\alpha|q|+V_{0}$ and $(s,\lambda)=(+,-)~{}{\rm and}~{}(-,+)$ with energy $E_{>}=\hbar^{2}q^{2}/2m+\hbar\alpha|q|+V_{0}$. The $q$ vectors associated with the eigenfunctions of the Hamiltonian are $q=sk_{so}+\lambda\sqrt{k^{2}+k_{so}^{2}-q_{0}^{2}}.$ (6) Note that this is the wavevector of the comment but the form of the wavefunction (Eq.4) involves the quantum number $\lambda$ denoting the wave vector direction. This form is in agreement with our work. Another subtle detail that makes the proposed form of the comment suspect is that the superposition of waves under the barrier does not correspond to equal energies unless the momentum-spin relation is taken into account. The very subtle difference determines whether this model shows spin polarization under tunneling or not as demonstrated in the comment. For reference work on SO active rings, where these arguments are also applicable see refs.[Birkholz, ; Richter, ; Chatelain, ; Bolivar, ] Spin polarization in one dimensional systems has been argued to be feasible in the tunneling regimeBart in agreement with transport symmetry relations, where there is an energy dependence. In fact if the injection energy is in between $E_{<}$ and $E_{>}$ one expects spin polarization while for the injection energy above $E_{>}$ then both spin orientations are filled and no spin polarization ensues. The hamiltonian of Eq.(1) is deceivingly simple because we wrote in a local coordinate system that rotates because of the constraints imposed by the helix i.e. $k_{z}=k_{y}\tan\eta$ (taking $x\rightarrow z$ as in reference [Varela2016, ]) where $\eta$ is the chiral angle of the helix. The system is actually three dimensional with a hamiltonian close to full filling of the form $\displaystyle H$ $\displaystyle=$ $\displaystyle tR^{2}\cos^{2}\eta\left(q_{y}+q_{z}\tan\eta\right)^{2}\mathbbm{1}_{s}+$ (7) $\displaystyle 2R\cos\eta~{}\lambda_{SO}\left(q_{y}+q_{z}\tan\eta\right)~{}s_{y},$ applying the relation $q_{z}=q_{y}~{}\tan\eta$ we obtain Eq.(5) of the paper, the Hamiltonian in the paper $\displaystyle H$ $\displaystyle=$ $\displaystyle tR^{2}\csc^{2}\eta~{}q^{2}_{z}~{}\mathbbm{1}_{s}+2R\csc\eta~{}\lambda_{SO}~{}s_{y}q_{z},$ (8) where again we exchange $x$ in the paper with $z$ in the current notation for consistency ($z$ is along the axis of the helix). This is the Hamiltonian object of the comment of ref.[WohlmanAharony, ], that appears not to involve an orbital degree of freedom for the electron (no orbital angular momentum). To better understand the physics we will eliminate $q_{z}$ (the vector along the axis of the helix) in favor of $q_{y}$ using their relationship to obtain (consistent with ref. [Varela2016, ]) $\displaystyle H=tR^{2}\sec^{2}{\eta}~{}q^{2}_{y}~{}\mathbbm{1}_{s}+2R\sec\eta~{}\lambda_{SO}~{}s_{y}q_{y},$ (9) still in local, rotating, coordinates. Further rewriting the Hamiltonian in cylindrical coordinates, where we can identify orbital and spin angular momentum, we get $H=-\beta~{}\mathbbm{1}_{s}\partial_{\varphi}^{2}-i\alpha_{\eta}s_{\varphi}\partial_{\varphi},$ (10) where $\varphi$ is the angle around the helix axis, $s_{\varphi}$ and $\beta=t(R\sec\eta/a)^{2}$ and $\alpha_{\eta}=2(R/a)\sec\eta~{}\lambda_{SO}$. This problem is different from the closed ring since it describes the motion on a helix and thus, periodic boundary conditions do not apply. The Hamiltonian in the previous equation is nevertheless non-hermitian Morpurgo and can be made hermitian by symmetrizing the Hamiltonian in Eq.10. With the latter procedure we just have to change $\sigma_{\varphi}\partial_{\varphi}\rightarrow\sigma_{\varphi}\partial_{\varphi}-(1/2)\sigma_{\rho}$ so that the hermitian Hamiltonian is $H=-\beta~{}\mathbbm{1}_{s}\partial_{\varphi}^{2}-i\alpha_{\eta}(\sigma_{\varphi}\partial_{\varphi}-\frac{1}{2}\sigma_{\rho}).$ (11) The Hamiltonian in this form has a very revealing interpretation, and makes now an obvious connection to the conclusions of our paper, since the second term is the kinetic Hamiltonian for a graphene ring $\propto{\bm{\sigma}}\cdot{\bf p}$, except that here $\mathbbm{\sigma}$ describes real spin, not pseudo-spin. This term will exert a torque on the ring since momentum and spin cannot be kept at a fixed angle on the ring. The torque will disappear if the ring is rotating with the electron momentum. This manifests itself as a pseudo-spin angular momentum in grapheneBolivar and it describes the rotation of the ring. The same physics applies for the helix but for the real spin, on the helix, the term tends to align momentum (which now circulates around the helix) and spin. Since this is not possible without changing angular momentum then there is a torque on the helix. The latter term is precisely what was computed in the paper as $(1/i\hbar)[s_{z},H]={\cal T}$ the torque on the molecule. Addressing the problem in three dimensions brings about new features to the problem regarding the adiabatic/non-adiabatic following of the SO effective magnetic field, which does not arise in the simplified Hamiltonian of Eq.(1). In conclusion, we believe the comment in ref.[WohlmanAharony, ] does not capture the correct spin-momentum coupling present in the model of ref.[VarelaZambrano, ], treating only effectively spinless electron tunneling. We hope our presentation has clarified the issues. ###### Acknowledgements. We acknowledge fruitfull discussions with Alexander Lopez. ## References * (1) O. Entin-Wohlman, A. Aharony, and Y. Utsumi, Comment (2020). * (2) S. Varela, I. Zambrano, B. Berche, V. Mujica, and E. Medina, Phys. Rev. B 101, 241410(R) (2020). * (3) Xu Yang, Caspar H. van der Wal, and Bart J. van WeesNano Lett. 8, 6148 (2020). * (4) S. Matityahu, Y. Utsumi, A. Aharony, O. Entin-Wohlman, and C. A. Balseiro, Phys. Rev. B 93, 075407 (2016). * (5) J. E. Birkholz, and V. Meden J. Phys.: Condens. Matter 20, 085226 (2008). * (6) D. Frustaglia, and K. Richter, Phys. Rev. B 69, 235310 (2004). * (7) B. Berche, C. Chatelain, and E. Medina, Eur. J. Phys. 31, 1267 (2010). * (8) N. Bolivar, E. Medina, and B. Berche Phys. Rev. B 89, 125413 (2014). * (9) S. Varela, V. Mujica, and E. Medina, Phys. Rev. B 93, 155436 (2016). * (10) F. E. Meijer, A. F. Morpurgo, and T. M. Klapwijk, Phys. Rev. B 66, 033107 (2002).
# Nucleon-pair coupling scheme in Elliott’s SU(3) model G. J<EMAIL_ADDRESS>School of Physics Science and Engineering, Tongji University, Shanghai 200092, China Calvin W. <EMAIL_ADDRESS>Department of Physics, San Diego State University, 5500 Campanile Drive, San Diego, CA 92182-1233 P. Van <EMAIL_ADDRESS>Grand Accélérateur National d’Ions Lourds, CEA/DRF- CNRS/IN2P3, Boulevard Henri Becquerel, F-14076 Caen, France Zhongzhou <EMAIL_ADDRESS>School of Physics Science and Engineering, Tongji University, Shanghai 200092, China ###### Abstract Elliott’s SU(3) model is at the basis of the shell-model description of rotational motion in atomic nuclei. We demonstrate that SU(3) symmetry can be realized in a truncated shell-model space if constructed in terms of a sufficient number of collective $S$, $D$, $G$, …pairs (i.e., with angular momentum zero, two, four, …) and if the structure of the pairs is optimally determined either by a conjugate-gradient minimization method or from a Hartree-Fock intrinsic state. We illustrate the procedure for 6 protons and 6 neutrons in the $pf$ ($sdg$) shell and exactly reproduce the level energies and electric quadrupole properties of the ground-state rotational band with $SDG$ ($SDGI$) pairs. The $SD$-pair approximation without significant renormalization, on the other hand, cannot describe the full SU(3) collectivity. A mapping from Elliott’s fermionic SU(3) model to systems with $s$, $d$, $g$, …bosons provides insight into the existence of a decoupled collective subspace in terms of $S$, $D$, $G$, …pairs. Atomic nuclei exhibit a wide variety of behaviors, ranging from single- particle motion to superconducting-like pairing to vibrational and rotational modes. To a large extent the story of nuclear structure is the quest to encompass the widest range of behaviors within the fewest degrees of freedom. In the early stage of nuclear physics, the spherical nuclear shell model Mayer ; Jensen stressed the single-particle nature of the nucleons in a nucleus, while the geometric collective model BM1 ; BM2 and the Nilsson mean-field model Nilsson pointed the way to describing rotational bands by emphasizing permanent quadrupole deformations Rainwater in “intrinsic” states. The reconciliation between these two pictures has been one of the most important advances in our understanding of the structure of nuclei. It was in large part due to Elliott who showed, on the basis of an underlying SU(3) symmetry, how to obtain deformed “intrinsic” states in a finite harmonic-oscillator single- particle basis occupied by nucleons that interact through a quadrupole- quadrupole force Elliott58 . This major step forward provided a microscopic interpretation of rotational motion in the context of the spherical shell model and, more recently, led to the symmetry-adapted no-core shell model symmetryadapted . Although the spherical shell model does provide a general framework to reproduce rotational bands Caurier05 and shape coexistence Heyde11 in light- and medium-mass nuclei, it is computationally still extremely challenging to describe deformation in heavier-mass regions Otsuka19 . Approximations must be sought. A tremendous simplification of the shell model occurs by considering only pairs of nucleons with angular momentum 0 and 2, and treating them as ($s$ and $d$) bosons. This approximation, known as the interacting boson model (IBM) IBM1 ; IBM2 , is particularly attractive because of its symmetry treatment in terms of a U(6) Lie algebra, which allows a spherical U(5), a deformed SU(3), and an intermediate SO(6) limit. While the IBM has been connected to the shell model for spherical nuclei OAI ; GJ95 , such relation has never been established for deformed nuclei, in which case the IBM has rather been derived from mean-field models Nomura1 ; Nomura2 . The nucleon-pair approximation (NPA) NPA1 ; NPA2 is one possible truncation scheme of the shell-model configuration space. The building blocks of the NPA are fermion pairs with certain angular momenta. Calculations are carried out in a fully fermionic framework, albeit in a severely reduced model space defined by the most important degrees of freedom in terms of pairs. The NPA therefore can be considered as intermediate between the full-configuration shell model and models that adopt the same degrees of freedom as the nucleon pairs but in terms of bosons. While the NPA has been successful for nearly spherical nuclei NPAr ; gs1 ; gs2 ; gs3 ; bpa1 ; bpa2 ; gs4 ; Lei , previous studies for well-deformed nuclei are not satisfactory. For example, in the fermion dynamical symmetry model FDSM1 ; FDSM2 an SU(3) limit with Sp(6) symmetry can be constructed in terms of $S$ and $D$ pairs but their symmetry- determined structure is far removed from that of realistic pairs Halse89 . Also, the binding energy, moment of inertia, and electric quadrupole ($E2$) transitions calculated in an $SD$-pair approximation are much smaller than those obtained in Elliott’s SU(3) limit for the $pf$ and $sdg$ shells Zhao2000 . In this Letter we successfully apply the NPA of the shell model to well- deformed nuclei. We show that the low-energy excitations of many-nucleon systems in Elliott’s SU(3) limit can be exactly reproduced with a suitable choice of pairs in the NPA. We obtain an understanding of this observation through a mapping to a corresponding boson model. We consider an example system with even numbers of protons and neutrons in a degenerate $pf$ or $sdg$ shell, interacting through a quadrupole-quadrupole force of the form, $\displaystyle V_{Q}=-(Q_{\pi}+Q_{\nu})\cdot(Q_{\pi}+Q_{\nu}),$ (1) where $Q_{\pi}$ ($Q_{\nu}$) is the quadrupole operator for protons (neutrons), $\displaystyle Q=-\displaystyle\sum_{\alpha\beta}\displaystyle\frac{\langle n_{\alpha}l_{\alpha}j_{\alpha}\|r^{2}Y_{2}\|n_{\beta}l_{\beta}j_{\beta}\rangle}{\sqrt{5}r_{0}^{2}}\left(a_{\alpha}^{\dagger}\times\tilde{a}_{\beta}\right)^{(2)}.$ (2) Greek letters $\alpha$, $\beta,\ldots$ denote harmonic-oscillator single- particle orbits labeled by $n$, $l$, $j$, and $j_{z}$; $a_{\alpha}^{\dagger}$ and $\tilde{a}_{\beta}$ are the nucleon creation operator and its time- reversed form for the annihilation operator, respectively; and $r_{0}$ is the harmonic-oscillator length. As shown in Ref. Elliott58 , the interaction $V_{Q}$ is a combination of the Casimir operators of SU(3) and SO(3), and its eigenstates are therefore classified by (irreducible) representations of these algebras with eigenenergies given by $\displaystyle-\frac{5}{2\pi}\left[\frac{1}{2}(\lambda^{2}+\lambda\mu+\mu^{2}+3\lambda+3\mu)-\frac{3}{8}L(L+1)\right],$ (3) in terms of the SU(3) labels $(\lambda,\mu)$ and the SO(3) label $L$, the total orbital angular momentum. Several useful SU(3) representations for low- lying states can be found in Ref. Zhao2000 . In the following we discuss in detail the case of 6 protons and 6 neutrons (6p-6n) in the NPA of the shell model and subsequently generalize to other numbers of nucleons. A nucleon-pair state of 6 protons is written as $\displaystyle|\varphi^{(I_{\pi})}\rangle$ $\displaystyle=$ $\displaystyle\left(({{A}^{({J}_{1})}}^{{\dagger}}\times{{A}^{({J}_{2})}}^{{\dagger}})^{(I_{2})}\times{{A}^{({J}_{3})}}^{{\dagger}}\right)^{(I_{\pi})}|0\rangle,$ (4) where $I_{2}$ is an intermediate angular momentum and ${{A}^{(J)}}^{{\dagger}}$ is the creation operator of a collective pair with angular momentum $J$: $\displaystyle{{A}^{(J)}}^{{\dagger}}=\sum_{{\alpha}\leq{\beta}}y_{J}({\alpha}{\beta})\left(a_{{\alpha}}^{\dagger}\times a_{{\beta}}^{\dagger}\right)^{(J)},$ (5) where $y_{J}({\alpha}{\beta})$ is the pair-structure coefficient. For systems with protons and neutrons, we construct the basis by coupling the proton and neutron pair states to a state with total angular momentum $I$, i.e., $|\psi^{(I)}\rangle=\left(|\varphi^{(I_{\pi})}\rangle\times|\varphi^{(I_{\nu})}\rangle\right)^{(I)}$. Level energies and wave functions are obtained by diagonalization of the Hamiltonian matrix in the space spanned by $\left\\{|\psi^{(I)}\rangle\right\\}$, that is, from a configuration- interaction calculation. If a sufficient number of pair states are considered in Eq. (4), the NPA model space can be made exactly equivalent to the full shell-model space. The interest of the NPA, however, is to restrict to the relevant pairs and describe low-energy nuclear structure in a truncated shell- model space. The selection of relevant pairs with the correct structure in Eq. (5) has been a long standing problem in NPA calculations. Recent applications choose pairs by the generalized seniority scheme (GS). Specifically, one optimizes the structure coefficients of the $S$ pair by minimizing the expectation value of the Hamiltonian in the $S$-pair condensate and one obtains other pairs by diagonalizing the Hamiltonian matrix in the space spanned by GS-two (i.e., one-broken-pair) states gs2 ; Xu2009 . The collective pairs obtained with the GS approach provide a good description of nearly-spherical nuclei but, as recognized in Ref. Lei12un and as will also be shown below, they are inappropriate in deformed nuclei. Instead we use the conjugate gradient (CG) method CG1 ; CG2 , where the structure coefficients of all pairs considered in the basis are simultaneously optimized by minimizing the ground-state energy in a series of iterative NPA calculations for a given Hamiltonian. The initial pairs in this iterative procedure are SU(3) tensors, obtained by diagonalizing $V_{Q}$ in a two-particle basis and retaining the lowest-energy pair. Figure 1: (a) Excitation energy and (b) electric quadrupole reduced transition probability $B(E2;I\rightarrow I-2)$ for the ground rotational band of 6 protons and 6 neutrons in the $pf$ shell in Elliott’s SU(3) model. The subscript “GS” stands for generalized seniority and “CG” for conjugate gradient (see text). Figure 2: Same as Fig. 1 for the $sdg$ shell. Figure 1 shows, for a 6p-6n system in the $pf$ shell, the results of various NPA calculations concerning excitation energies and $E2$ reduced transition probabilities (with the standard effective charges $e_{\pi}=1.5$ and $e_{\nu}=0.5$) for the lowest rotational band. These are compared to the exact results of Elliott’s model, where the ground band belongs to the SU(3) representation $(\lambda,\mu)=(24,0)$. Surprisingly, the $SDG$-pair approximation of the shell model in the CG approach (denoted as $SDG_{\rm CG}$) reproduces the exact binding energy, $810/\pi$ MeV according to Eq. (3), to a precision of eight digits, as well as the exact excitation energies for the entire ground band. One can understand the occurrence of the $(24,0)$ representation from the coupling of $(12,0)$ for the six protons and six neutrons separately and, in fact, all bands contained in the product $(12,0)\times(12,0)$, i.e. $(24,0)$, $(22,1)$,…,$(0,12)$, are exactly reproduced in the $SDG_{\rm CG}$-pair truncated space. We also find that the results of the $SDG$-pair approximation are close to the exact results if the pairs are SU(3) tensors. For example, with such pairs the calculation reproduces 98% of the exact binding energy, 99% of the exact moment of inertia, and 97% of the exact $B(E2)$ values. On the other hand, the results of the $SDG$-pair approximation deteriorate if the pairs are obtained with the GS approach (denoted as $SDG_{\rm GS}$), which reproduces only 76% of the exact binding energy. Furthermore, $SDG_{\rm GS}$ fails to describe the quadrupole collectivity: The moment of inertia predicted by $SDG_{\rm GS}$ is only $\sim$43% of the exact one, the predicted $B(E2)$ values are too small, and the yrast states with angular momentum $I\geq 10$ do not follow the behavior of a quantum rotor. One concludes that the structure of the collective pairs, as determined by the GS approach, is not suitable for the description of well-deformed nuclei. It is also of interest to investigate the standard $SD$-pair approximation of the shell model and results of the $SD_{\rm GS}$-, $SD_{\rm CG}$-, and $SDS^{\prime}D^{\prime}_{\rm CG}$-pair approximations are shown in Fig. 1. Here $S^{\prime}$ and $D^{\prime}$ are collective pairs with angular momentum 0 and 2 but orthogonal to the $S$ and $D$ pairs, respectively. While the CG approach provides the numerically optimal solution in $SD_{\rm CG}$\- and $SDS^{\prime}D^{\prime}_{\rm CG}$-pair approximations, the results nonetheless are underwhelming. In the $SD_{\rm GS}$, $SD_{\rm CG}$, and $SDS^{\prime}D^{\prime}_{\rm CG}$ spaces only 76%, 83%, and 84% of the exact binding energy are reproduced, respectively, and the predicted moments of inertia and $B(E2)$ strengths are evidently smaller than the exact SU(3) results. We conclude that the collective $SD$ pairs cannot fully explain the quadrupole collectivity of the SU(3) states. Interestingly, the excitation energies of the yrast states predicted by the $SD$-pair approximations follow an $I(I+1)$ rule and the $B(E2)$ strength exhibits a nearly-parabolic shape [see Fig. 1(b)], two typical features of rotational motion. This raises the hope that an effective Hamiltonian and effective charges can be derived in the restricted $SD_{\rm CG}$ space, which takes into account the coupling with the excluded space. This conclusion is in line with a more phenomenological approach Nomura2 , in which an $L\cdot L$ term is added to the Hamiltonian, such that properties of low-lying states of well-deformed nuclei are reproduced in $sd$-IBM. Figure 2 shows the corresponding results of for the 6p-6n system in the $sdg$ shell. In this case the $SDGI_{\rm CG}$-pair approximation of the shell model reproduces exactly the SU(3) results and all states belonging to the coupled representation $(18,0)\times(18,0)$, i.e. $(36,0)$, $(34,1)$,…,$(0,18)$, are fully contained in the $SDGI_{\rm CG}$-pair truncated space. Again, if the pairs are SU(3) tensors, the $SDGI$-pair approximation is close to the exact result and reproduces 99% of the exact binding energy, 97% of the exact moment of inertia, and 99% of the exact $B(E2)$ values. The $SDG_{\rm CG}$-pair approximation yields 96% of the binding energy and 57% of the moment of inertia. The predicted $B(E2)$ strength in the $SDG_{\rm CG}$-pair approximation is close to the exact result for low angular momenta but deteriorates as angular momentum $I$ increases. The necessity of renormalization is even larger in the $SD_{\rm CG}$-pair approximation. Let us now try to understand the above results. Specifically, why is it that the SU(3) results in the $pf$ shell are exactly reproduced with $SDG$ but not with $SD$ pairs? Similarly, why is it that SU(3) in the $sdg$ shell cannot be represented with $SD$ or $SDG$ but requires $SDGI$ pairs? To explain these findings, we invoke a mapping to a system with corresponding $s$, $d$, $g$, and $i$ bosons (denoted as $sd$-, $sdg$-, or $sdgi$-IBM) and the bosonic realization of SU(3). The mapping is further specified by the fact that the quadrupole-quadrupole interaction $V_{Q}$ is an SU(4) invariant and, consequently, one aims to realize the symmetries associated with Wigner’s supermultiplet model Wigner in terms of bosons. An SU(4)-invariant boson model, known as IBM-4 Elliott81 , requires to assign to each boson a spin- isospin of $(s,t)=(0,1)$ or $(1,0)$, giving rise to a spin-isospin algebra ${\rm U}_{st}(6)$. The SU(3) limit can be realized in terms of bosons by first decoupling the orbital angular momentum from the spin-isospin of the bosons. For an $n_{b}$-boson state this leads to the classification $\displaystyle\begin{array}[]{ccccc}{\rm U}(6\Lambda)&\supset&{\rm U}(\Lambda)&\otimes&{\rm U}_{st}(6)\\\ \downarrow&&\downarrow&&\downarrow\\\ \left[n_{b}\right]&&\left[\bar{h}\right]\equiv\left[h_{1},...,h_{6}\right]&&\left[\bar{h}\right]\equiv\left[h_{1},...,h_{6}\right]\end{array},$ (9) with $\Lambda=6$, 15, and 28 for $sd$-, $sdg$-, and $sdgi$-IBM, respectively. The six labels $[\bar{h}]$ are a partition of $n_{b}$ such that $h_{1}\geq h_{2}\geq\cdots\geq h_{6}$; they specify the representations of ${\rm U}(\Lambda)$ and ${\rm U}_{st}(6)$, which by virtue of the overall ${\rm U}(6\Lambda)$ symmetry of the bosons must be identical. For all above values of $\Lambda$ (i.e., $\Lambda=6$, 15, and 28), Elliott’s SU(3) appears as a subalgebra of ${\rm U}(\Lambda)$, $\displaystyle\begin{array}[]{ccccccc}{\rm U}(\Lambda)&\supset&{\rm U}(3)&\supset&{\rm SU}(3)&\supset&{\rm SO}(3)\\\ \downarrow&&\downarrow&&\downarrow&&\downarrow\\\ \left[\bar{h}\right]&&\left[h_{1}^{\prime\prime},h_{2}^{\prime\prime},h_{3}^{\prime\prime}\right]&&(\lambda,\mu)&K&L\end{array},$ (13) while Wigner’s SU(4) occurs as a subalgebra of ${\rm U}_{st}(6)$, $\displaystyle\begin{array}[]{ccccccc}{\rm U}_{st}(6)&\supset&{\rm SU}_{st}(4)&\supset&{\rm SU}_{s}(2)&\otimes&{\rm SU}_{t}(2)\\\ \downarrow&&\downarrow&&\downarrow&&\downarrow\\\ \left[\bar{h}\right]&&(\lambda^{\prime},\mu^{\prime},\nu^{\prime})&&S&&T\end{array}.$ (17) The quantum numbers $(\lambda,\mu)$, $K$, and $L$ in Eq. (13) and $(\lambda^{\prime},\mu^{\prime},\nu^{\prime})$, $S$, and $T$ in Eq. (17) have an interpretation identical to that in Elliott’s fermionic SU(3) model Elliott58 ; Isacker2016 . The SU(3) labels $(\lambda,\mu)$ in the different versions of the IBM can be worked out with the following procedure Elliott1999 . For a given number of bosons $n_{b}$, one enumerates all possible Young diagrams $[\bar{h}]$ of ${\rm U}(\Lambda)$ or ${\rm U}_{st}(6)$. For each $[\bar{h}]$ one obtains the ${\rm SU}_{st}(4)$ labels $(\lambda^{\prime},\mu^{\prime},\nu^{\prime})$ from the branching rule ${\rm U}(6)\supset{\rm SU}(4)$, and retains only the ones that contain the favored supermultiplet. Finally, the SU(3) labels $(\lambda,\mu)$ for the above $[\bar{h}]$ are found from the ${\rm U}(\Lambda)\supset{\rm SU}(3)$ branching rule. Let us apply this procedure to the 6p-6n system in the $pf$ shell. The lowest eigenstates of the quadrupole-quadrupole interaction belong to the favored SU(4) supermultiplet $(\lambda^{\prime},\mu^{\prime},\nu^{\prime})=(0,0,0)$ and the leading (fermionic) SU(3) representation is $(\lambda,\mu)=(24,0)$. For $n_{b}=6$ bosons, the ${\rm U}_{st}(6)$ or ${\rm U}(\Lambda)$ representations containing this favored supermultiplet $(0,0,0)$ are $[\bar{h}]=[6]$, $[4,2]$, $[2^{3}]$, and $[1^{6}]$, which have the SU(3) labels $(\lambda,\mu)$ as listed in Table 1 for the $sd$-, $sdg$-, and $sdgi$-IBM. The leading SU(3) representation $(24,0)$ is not contained in $sd$-IBM but is present in the $[6]$ representation of U(15), and therefore it is contained in $sdg$-IBM. Similarly, 6p-6n in the $sdg$ shell give rise to the leading SU(3) representation $(36,0)$, which is not contained in $sd$\- nor $sdg$-IBM but present in $sdgi$-IBM. The generalization to the 2p-2n ($n=4$) and 4p-4n ($n=8$) systems in the $pf$ and $sdg$ shells is summarized in Table 2. The second column lists the leading fermionic SU(3) representations and the third, fourth, and fifth columns indicate whether this representation is contained in $sd$-, $sdg$-, and $sdgi$-IBM, respectively. A dash (—) indicates that it is not, in which case an NPA calculation adopting the corresponding $SD$, $SDG$, or $SDGI$ pairs does not reproduce the full collectivity of the ground-state band in the fermionic SU(3) model. For $n=4$ and $n=8$ nucleons in the $sdg$ shell no exact mapping can be realized to $sdgi$-IBM and bosons with even higher angular momentum are needed. It should be noted, however, that this generally occurs for low nucleon number (e.g., for $n=12$ nucleons in the $sdg$ shell the problem does not occur), for which NPA calculations with high angular momentum pairs are still feasible. Table 1: Leading SU(3) representations for 6 bosons in $sd$-, $sdg$-, and $sdgi$-IBM occurring in the ${\rm U}(\Lambda)$ and ${\rm U}_{st}(6)$ representations $[\bar{h}]$ containing the favored supermultiplet $(0,0,0)$. (bosons)${}^{n_{b}}$ | $[\bar{h}]$ | $(\lambda,\mu)$ ---|---|--- $(sd)^{6}$ | $[6]$ | $(12,0),(8,2),(4,4),(6,0),(0,6),\dots$ | $[4,2]$ | $(8,2),(6,3),(7,1),(4,4)^{2},(5,2),\dots$ | $[2^{3}]$ | $(6,0),(0,6),(3,3),(2,2)^{2},(0,0)$ | $[1^{6}]$ | $(0,0)$ $(sdg)^{6}$ | $[6]$ | $(24,0),(20,2),(18,3),(16,4)^{2},(18,0),\dots$ | $[4,2]$ | $(20,2),(18,3),(19,1),(16,4)^{3},(17,2),\dots$ | $[2^{3}]$ | $(18,0),(15,3),(12,6),(13,4),(14,2)^{3},\dots$ | $[1^{6}]$ | $(12,0),(8,5),(9,3),(3,9),(7,4),\dots$ $(sdgi)^{6}$ | $[6]$ | $(36,0),(32,2),(30,3),(28,4)^{2},(30,0),\dots$ | $[4,2]$ | $(32,2),(30,3),(31,1),(28,4)^{3},(29,2)^{2},\dots$ | $[2^{3}]$ | $(30,0),(27,3),(24,6),(25,4),(26,2)^{3},\dots$ | $[1^{6}]$ | $(24,0),(20,5),(21,3),(18,6),(19,4),\dots$ Table 2: Leading fermionic SU(3) representations $(\lambda,\mu)$ for $n$ nucleons in the $pf$ and $sdg$ shells and the U(6), U(15), and U(28) representations of the $n_{\rm b}=n/2$ boson system that contain this $(\lambda,\mu)$ in $sd$-, $sdg$-, and $sdgi$-IBM. (shell)n | $(\lambda,\mu)$ | $sd$-IBM | $sdg$-IBM | $sdgi$-IBM ---|---|---|---|--- $(pf)^{4}$ | $(12,0)$ | — | — | $[2]$ $(pf)^{8}$ | $(16,4)$ | — | — | $[4],[2^{2}]$ $(pf)^{12}$ | $(24,0)$ | — | $[6]$ | $[6],[4,2],[2^{3}],[1^{6}]$ $(sdg)^{4}$ | $(16,0)$ | — | — | — $(sdg)^{8}$ | $(24,4)$ | — | — | — $(sdg)^{12}$ | $(36,0)$ | — | — | $[6]$ While the best NPA solutions so far have been found by a numerically intensive optimization, it turns out they can also be obtained from a deformed “intrinsic” state. Again consider the 6p-6n system in the $pf$ shell. An unconstrained Hartree-Fock (HF) calculation in this single-particle shell- model space JohnsonSHERPA with a quadrupole-quadrupole interaction provides us with a HF state with an axially symmetric quadrupole deformed shape, a consequence of the spontaneous symmetry breaking Nambu of rotational symmetry. One can project out a $K=0$ band with good angular momentum from this HF state JohnsonLAMP , which exactly corresponds to the SU(3) representation $(24,0)$ Elliott58 . We use $a$ and $\bar{a}$ to denote the HF single-particle orbit and its time-reversal partner, respectively, and we write the creation operator of a nucleon as $c_{a}^{\dagger}$. A Slater determinant for an even number $2N$ of protons or neutrons can be written as a pair condensate: $\displaystyle\prod_{a=1}^{N}c_{a}^{\dagger}c_{\bar{a}}^{\dagger}|0\rangle=\mathcal{N}\left(\sum_{a}v_{a}~{}c^{\dagger}_{a}c_{\bar{a}}^{\dagger}\right)^{N}|0\rangle.$ (18) The pair in the deformed HF state is a superposition of collective pairs of good angular momentum in the shell model arXiv : $\displaystyle\sum_{a}v_{a}~{}c^{\dagger}_{a}c_{\bar{a}}^{\dagger}=\sum_{JM}{{A}^{(J)}_{M}}^{\dagger}.$ (19) For the appropriate $v_{a}$ one obtains $SDG$ pairs, which are the same as the $SDG$ pairs obtained by the CG-NPA calculations. Similarly, the $SDGI$ pairs responsible for (36,0) for 6p-6n in the $sdg$ shell can be also projected out from a deformed HF pair. The CG approach provides numerically optimal solutions in the NPA but is computationally heavy due to hundreds, even thousands of iterations. The HF approach derives pairs using an unconstrained HF calculation and the decomposition of pairs according to Eq. (19) has a very low computational cost. Figure 3: The ground rotational band of 52Fe. The experimental energies are taken from Ref. expt1 and the shell-model results are obtained with the GXPF1 interaction. $I^{\pi}$ | Expt. | SM | $SDG$ ---|---|---|--- $2^{+}$ | 14.2(19) | 19.2 | 17.0 $4^{+}$ | 26(6) | 25.0 | 21.6 $6^{+}$ | 10(3) | 17.4 | 20.0 $8^{+}$ | 9(4) | 11.5 | 15.5 $10^{+}$ | | 12.7 | 10.5 Table 3: $B(E2;I\rightarrow I-2)$ values (in W.u.) for the ground rotational band of 52Fe. The experimental values are taken from Ref. expt1 and the shell-model results are obtained with the GXPF1 interaction. Finally, we show that the NPA with CG-pairs provides a good description of low-lying states of rotational nuclei also if a realistic shell-model interaction is taken. We exemplify this with the nucleus 52Fe, considered as a 6p-6n system in the $pf$ shell with the GXPF1 effective interaction gxpf1 . Figure 3 and Table 3 compare, for the ground rotational band of 52Fe, the experimental data expt1 , the full configuration shell model (SM), and the $SDG_{\rm CG}$-pair approximation. Both the level energies and the $B(E2)$ values obtained with $SDG_{\rm CG}$ are in good agreement with the data and with the shell model. In summary, we construct in the NPA a collective subspace of the full shell- model space such that the former exactly reproduces, without any renormalization, the properties of the low-energy states of the latter. This construction is valid for an SU(3) quadrupole-quadrupole Hamiltonian and is achieved by determining the structure of the pairs with the conjugate-gradient minimization technique or on the basis of a deformed HF calculation. Exact correspondence is achieved only if a sufficient number of different pairs is considered. For example, a 6p-6n system in the $pf$ ($sdg$) shell is reproduced exactly with $SDG$ ($SDGI$) pairs; with just $SD$ pairs, an important renormalization of all operators is required. We have analytic understanding of this result: The collective subspace of the NPA exactly captures the collectivity of the full space if and only if the mapping to a model constructed with bosons corresponding to the pairs gives rise to a leading bosonic SU(3) representation that is also leading in fermionic SU(3). For many years a central problem in nuclear structure has been the construction of a collective subspace that decouples from the full shell-model space. With this work the conditions necessary for this decoupling to be exact are now understood for an SU(3) Hamiltonian. This understanding will pave the way for the construction of viable collective subspaces for more realistic shell-model interactions. It will also clarify the derivation of boson Hamiltonians appropriate for quadrupole deformed nuclei. Similar techniques conceivably might be applied elsewhere, such as to octupole-deformed nuclei with a Sp($\Omega$) or SO($\Omega$) symmetry Isacker2016 . ###### Acknowledgements. This material is based upon work supported by the National Key R&D Program of China under Grant No. 2018YFA0404403, the National Natural Science Foundation of China under Grants No. 12075169, 11605122, and 11535004, the U.S. Department of Energy, Office of Science, Office of Nuclear Physics, under Award No. DE-FG02-03ER41272, the CUSTIPEN (China-U.S. Theory Institute for Physics with Exotic Nuclei) funded by the U.S. Department of Energy, Office of Science grant number DE-SC0009971. ## References * (1) M. G. Mayer, Phys. Rev. 75, 1969 (1949). * (2) O. Haxel, J. H. D. Jensen, and H. E. Suess, Phys. Rev. 75, 1766 (1949). * (3) A. Bohr and B. R. Mottelson, Mat. Fys. Medd. K. Dan. Vidensk. Selsk 27, 16 (1953). * (4) A. Bohr and B. R. Mottelson, Nuclear Structure (World Scientific,1998). * (5) S. G. Nilsson, Mat. Fys. Medd. K. Dan. Vidensk. Selsk 29, 16 (1955). * (6) J. Rainwater, Phys. Rev. 79, 432 (1950). * (7) J. P. Elliott, Proc. R. Soc. A 245, 128 (1958); 245, 562 (1958). * (8) T. Dytrych, K. D. Launey, J. P. Draayer, P. Maris, J. P. Vary, E. Saule, U. Catalyurek, M. Sosonkina, D. Langr, and M. A. Caprio, Phys. Rev. Lett. 111, 252501 (2013); T. Dytrych, K. D. Launey, J. P. Draayer, D. J. Rowe, J. L. Wood, G. Rosensteel, C. Bahri, D. Langr, and R. B. Baker, Phys. Rev. Lett. 124, 042501 (2020). * (9) E. Caurier, G. Martínez-Pinedo, F. Nowacki, A. Poves, and A. P. Zuker, Rev. Mod. Phys. 77, 427 (2005). * (10) K. Heyde and J. L. Wood, Rev. Mod. Phys. 83, 1467 (2011). * (11) T. Otsuka, Y. Tsunoda, T. Abe, N. Shimizu, and P. Van Duppen, Phys. Rev. Lett. 123, 222502 (2019). * (12) A. Arima and F. Iachello, Phys. Rev. Lett. 35, 1069 (1975); Ann. Phys. 111, 201 (1978). * (13) F. Iachello and A. Arima, The Interacting Boson Model (Cambridge University Press, Cambridge, 1987). * (14) T. Otsuka, A. Arima, F. Iachello, and I. Talmi, Phys. Lett. B 76, 139 (1978); T. Otsuka, A. Arima, and F. Iachello, Nucl. Phys. A 309, 1 (1978). * (15) J. N. Ginocchio and C. W. Johnson, Phys. Rev. C 51, 1861 (1995). * (16) K. Nomura, N. Shimizu, and T. Otsuka, Phys. Rev. Lett. 101, 142501 (2008). * (17) K. Nomura, T. Otsuka, N. Shimizu, and L. Guo, Phys. Rev. C 83, 041302(R) (2011). * (18) J. Q. Chen, Nucl. Phys. A 626, 686 (1997). * (19) Y. M. Zhao, N. Yoshinaga, S. Yamaji, J. Q. Chen, and A. Arima, Phys. Rev. C 62, 014304 (2000). * (20) Y. M. Zhao and A. Arima, Phys. Rep. 545, 1 (2014). * (21) I. Talmi, Nucl. Phys. A 172, 1 (1971). * (22) Y. K. Gambhir, A. Rimini, and T. Weber, Phys. Rev. 188, 1573 (1969); Y. K. Gambir, S. Haq, and J. K. Suri, Ann. Phys.(N.Y.) 133, 154 (1981). * (23) K. Allaart, E. Boeker, G. Bonsignori, M. Savoia, and Y. K. Gambhir, Phys. Rep. 169, 209 (1988). * (24) Y. Lei, Z. Y. Xu, Y. M. Zhao, and A. Arima, Phys. Rev. C 80, 064316 (2009); 82, 034303 (2010). * (25) M. A. Caprio, F. Q. Luo, K. Cai, V. Hellemans, and C. Constantinou, Phys. Rev. C 85, 034324 (2012). * (26) Y. Y. Cheng, Y. M. Zhao, and A. Arima, Phys. Rev. C 94, 024307 (2016); Y. Y. Cheng, C. Qi, Y. M. Zhao, and A. Arima, Phys. Rev. C 94, 024321 (2016). * (27) C. Qi, L. Y. Jia, and G. J. Fu, Phys. Rev. C 94, 014312 (2016). * (28) J. N. Ginocchio, Phys. Lett. B 79, 173 (1978); 85, 9 (1979); Ann. Phys. 126, 234 (1980). * (29) C. L. Wu, D. H. Feng, X. G. Chen, J. Q. Chen, and M. W. Guidry, Phys. Rev. C 36, 1157 (1987); C. L. Wu, D. H. Feng, and M. Guidry, Adv. Nucl. Phys. 21, 227 (1994). * (30) P. Halse, Phys. Rev. C 39, 1104 (1989). * (31) Y. M. Zhao, N. Yoshinaga, S. Yamaji, and A. Arima, Phys. Rev. C 62, 014316 (2000). * (32) Z. Y. Xu, Y. Lei, Y. M. Zhao, S. W. Xu, Y. X. Xie, and A. Arima, Phys. Rev. C 79, 054315 (2009). * (33) Y. Lei, S. Pittel, G. J. Fu, and Y. M. Zhao, arXiv: 1207.2297v1; S. Pittel, Y. Lei, Y.M. Zhao, and G. J. Fu, AIP Conference Proceedings 1488, 300 (2012); S. Pittel, Y. Lei, G. J. Fu, and Y. M. Zhao, Journal of Physics: Conference Series 445, 012031 (2013). * (34) M. R. Hestenes and E. Stiefel, J. Res. Natl. Inst. Stan. 49, 409 (1952). * (35) R. Fletcher and C. M. Reeves, Comput. J. 7, 149 (1964). * (36) E. P. Wigner, Phys. Rev. 51, 106 (1937). * (37) J. P. Elliott and J. A. Evans, Phys. Lett. B 101, 216 (1981). * (38) P. Van Isacker and S. Pittel, Phys. Scr. 91, 023009 (2016). * (39) J. P. Elliott and J. A. Evans, J. Phys. G 25, 2071 (1999). * (40) I. Stetcu and C. W. Johnson, Phys. Rev. C 66, 034301 (2002). * (41) Y. Nambu, Phys. Rev. Lett. 4, 380 (1960). * (42) C. W. Johnson and K. D. O’Mara, Phys. Rev. C 96, 064304 (2017); C. W. Johnson and C. F. Jiao, Phys. G 46, 015101 (2019). * (43) G. J. Fu and C. W. Johnson, Phys. Lett. B 809, 135705 (2020). * (44) Y. Dong, H. Junde, Nucl. Data Sheets 128, 185 (2015). * (45) M. Honma, T. Otsuka, B. A. Brown, and T. Mizusaki, Phys. Rev. C 69, 034335 (2004).
# ”Can I Touch This?”: Survey of Virtual Reality Interactions via Haptic Solutions Revue de Littérature des Interactions en Réalité Virtuelle par le biais de Solutions Haptiques Elodie Bouzbib ISIR. Sorbonne UniversitéISCD. Sorbonne UniversitéParisFrance , Gilles Bailly , Sinan Haliyo ISIR. Sorbonne UniversitéParisFrance and Pascal Frey ISCD. Sorbonne UniversitéParisFrance (2021) ###### Abstract. Haptic feedback has become crucial to enhance the user experiences in Virtual Reality (VR). This justifies the sudden burst of novel haptic solutions proposed these past years in the HCI community. This article is a survey of Virtual Reality interactions, relying on haptic devices. We propose two dimensions to describe and compare the current haptic solutions: their degree of physicality, as well as their degree of actuation. We depict a compromise between the user and the designer, highlighting how the range of required or proposed stimulation in VR is opposed to the haptic interfaces flexibility and their deployment in real-life use-cases. This paper (1) outlines the variety of haptic solutions and provides a novel perspective for analysing their associated interactions, (2) highlights the limits of the current evaluation criteria regarding these interactions, and finally (3) reflects the interaction, operation and conception potentials of ”encountered-type of haptic devices”. haptics, Virtual Reality, human factors, haptic devices ††copyright: acmcopyright††journalyear: 2021††doi: 10.1145/XXXXXXXXX.XXXXXXXX††conference: 32e conférence Francophone sur l’Interaction Homme-Machine; April 13–16, 2021; Metz, France††booktitle: IHM ’20’21 : 32e conférence Francophone sur l’Interaction Homme-Machine, April 13–16, 2021, Metz, France††price: 15.00††isbn: XXX-X-XXXX-XXXX-X/XX/XX††ccs: Human-centered computing Virtual reality††ccs: Human-centered computing Haptic devices††ccs: Human-centered computing Interaction design theory, concepts and paradigms Le retour haptique est devenu essentiel pour améliorer l’expérience utilisateur en Réalité Virtuelle (RV). C’est pourquoi nous observons une explosion du nombre de solutions haptiques proposées ces dernières années en IHM. Cet article est une revue de littérature des interactions en RV s’appuyant sur des dispositifs haptiques. Nous proposons deux dimensions pour décrire et comparer les solutions haptiques : leur degré de physicalité ainsi que leur degré de robotisation. Nous formulons un compromis utilisateur/concepteur, reflétant la variété des stimulations requises/proposées en RV, en opposition à la flexibilité des interfaces et leur déploiement en situation réelle. Ce travail (1) offre un panorama des solutions haptiques en RV ainsi qu’un cadre d’analyse pour étudier les interactions associées, (2) souligne les limites des critères d’évaluation actuels pour ce type d’interactions, et finalement (3) reflète les potentiels interactionnel, opérationnel et conceptuel des interfaces haptiques ”à contacts intermittents”. haptique, Réalité Virtuelle, facteurs humains, dispositif haptique ## 1\. Introduction In the last few years, the terms ”Virtual Reality” and ”Haptics” have been amongst the most quoted keywords in HCI conferences such as ACM CHI or ACM UIST. Indeed, Head-Mounted Displays (HMDs) are now affordable and provide high quality visual and audio feedback, but augmenting the experience by enhancing VR through the sense of touch (haptic feedback) has become a main challenge. A large variety of haptic solutions has currently been proposed, nonetheless they have highly different scopes, due to the wide range of haptic features. It is hence difficult to compare their similarities and differences and have a clear understanding of the design possibilities. In this paper, we present a survey of existing haptic interactions in VR. We use the terms ”haptic interactions” to emphasize the focus on the users actions, and to analyse how the ”haptic devices” influence their behaviours in VR. We provide a synthesis of existing research on haptic interactions in VR and depict, from the required haptic features stimulation and interaction opportunities, a design space discussing and classifying the associated haptic solutions according to two dimensions: their degree of physicality, i.e. their physical consistency and level of resemblance as to replicating an object, and their degree of actuation, i.e. whether they rely on a motor-based hardware implementation enabling autonomous displacements of the interface (eg changing its shape or position) (Table 1). This design space is useful to characterize, classify and compare haptic interactions and the corresponding haptic solutions. We also propose two criteria, User experience and Conception costs, highlighting the implicit trade-offs between the quality of the user experience and the intricacy for the designer to implement these solutions. Both of the user’s and designer’s perspectives are hence considered in a novel framework to evaluate haptic interactions. Finally, we illustrate the utility of our design space by analyzing and comparing four haptic solutions. This analysis indicates that (1) the use of real props in a virtual environment benefits the user experience, but limits the interactions to the existing props available within the VR arena; (2) the use of robotised interfaces enables more various interactions; (3) combining them offers the best user experience/design cost trade-off; (4) current evaluation methods do not allow a fair representation and comparison of haptic solutions. We hence propose guidelines to evaluate haptic interactions from both the user and designer perspectives. We also outline how intertwining interfaces can expand haptic opportunities, by conducting a deeper investigation on Robotic Graphics interfaces (McNeely, 1993) . Indeed, in the quest of the Ultimate Display (Sutherland, 1965), these show (a) the largest variety of interactions, (b) the most reliable interfaces through their automation, and (c) the most natural interactions as they encounter the users at their positions of interest without further notice. ## 2\. Background Surveys in Virtual Reality consider the technology itself and its limits (Zhao, 2009; Zhou and Deng, 2009), or more specifically its use-case scenarios. VR is indeed used in industries (Berg and Vance, 2017; Zimmermann, 2008), healthcare (Moline, 1997), or in gaming. In gaming, the concerns are mainly regarding the evaluation protocols (Merino et al., 2020), ie the presence (Schuemie et al., 2001) and its related questionnaires (Schwind et al., 2019; Usoh et al., 2000). Surveys for instance compare the results whenever the questionnaires are asked in VR or in the real world (Alexandrovsky et al., 2020; Putze et al., 2020). The user behaviour in VR is also analysed, through gesture recognition (Sagayam and Hemanth, 2017) or system control techniques (eg menus) (Bowman and Wingrave, 2001). The research areas are coincidentally almost similar in haptics. Indeed, surveys analyse haptics themselves (Varalakshmi et al., 2012), haptic devices (Seifi et al., 2019; Rakkolainen et al., 2020; Talvas et al., 2014; Hayward and Maclean, 2007) or examine the scenarios which benefit from a stimulation of the haptic cues. Haptics are used in telemanipulation (Galambos, 2012), for training in the industry (Xia, 2016; Bloomfield et al., 2003) or for healthcare purposes (Coles et al., 2011; Rangarajan et al., 2020), or in gaming (Kim and Schneider, 2020). Finally, some surveys have been proposed at the intersection of VR and haptics and focus either on specific methods (pseudo-haptic feedback) (Lécuyer, 2009), technology according to stimulated haptic features (temperature, shape, skin stretch, pressure) (Wang et al., 2020d; Dominjon et al., 2007) or the motivations and applications of each haptic device category (Wang et al., 2020a). In contrast our survey outlines the variety of haptic interactions and technologies in VR and provides a framework to analyse them. ## 3\. Scope and Definitions The scope of this article is to analyse how a single user interacts and is provided with believable haptic feedback in Virtual Reality (Magnenat-Thalmann et al., 2005). We thus define the terms ”virtual reality” and ”haptics” and how they are related. ### 3.1. Virtual Reality Virtual reality corresponds to a 3D artificial numeric environment in which users are immersed in. The environment can be projected onto a large screen, in a simulation platform for instance, or multiple ones, such as with CAVE technology (where the image is projected onto at least 3 distinct walls of a room-scale arena). In this survey, we consider an artificial reality (Wexelblat, 1993) where users do not perceive their physical vicinity: the outside world is not noticeable and users are fully immersed through a head- mounted display (HMD). For instance, augmented reality (AR), where the physical environment is augmented with virtual artefacts, is out of our scope. Through a Head Mounted Display (HMD), Virtual reality creates immersive experiences for the users. These are only limited by the designers’ imagination, and are evaluated through presence. Presence is defined as the ”subjective experience of being in one place, even when one is physically situated in another” (Witmer and Singer, 1998; Slater, 1999). It quantifies the users’ involvement and naturalness of interactions through control, sensory, distraction and realism factors. This heavily relies on the sensory input and output channels, however, as VR was mainly integrating audio and visual cues, quantifying the haptic contribution in an experience remains difficult. ### 3.2. Haptics: Tactile vs Kinesthetic Perception Haptics is the general term for the sense of touch. They are a combination of two cues: tactile and kinesthetic. The tactile cues are developed through the skin, while the kinesthetic ones come from proprioception and are through the muscles and the tendons. #### 3.2.1. Tactile cues: The skin is composed of four types of mechanoreceptors (Lederman and Klatzky, 2009). The first ones, ”Merkel nerve endings”, transmit mechanical pressure, position and shapes or edges. They are stimulated whilst reading Braille for instance. The second ones, ”Ruffini corpuscle end-organ”, are sensitive to skin stretch and provide both pressure and slippage information. The third ones are the ”Pacinian corpuscles”, which are sensitive to vibration and pressure. The last ones, ”Meissner’s corpuscles”, are highly sensitive and provide light touch and vibrations information. It also contains thermoreceptors, which transmit information about temperature: the Ruffini endings respond to warmth, while the Krause ones detect cold. Through tactile cues, the human can hence feel shapes or edges, pressure, vibrations or temperature changes. #### 3.2.2. Kinesthetic cues: The kinesthetic cues rely on proprioception, ie the perception and the awareness of our own body parts positions and movements. Mechanoreceptors into the muscles, the ”spindles”, communicate to the nervous system information the forces muscle generate, as well as their length change (Jones, 2000). The primary type of spindle is sensitive to the velocity and acceleration of a muscle contraction or limb movement, while the second type provides information about static muscle length or limb positions. Kinesthetic cues hence allow to feel forces, as well as perceiving weights or inertia. ### 3.3. VR & Haptics Whenever we touch or manipulate an object, the combination of these two previous cues allows to understand its material, but also its shape and the constraints it implies to the user. On the one side, adding physical presence (Lepecq et al., 2008) through haptic feedback in VR enhances the users’ immersion, even at an emotional and physiological scale: the heart rate of a user can literally increase with the use of haptics through real objects (Insko, 2001). Haptics are also required for interacting with the environment: the user needs to control the changes in the environment (Held and Durlach, 1992) and to be aware of the modifications he physically has made (eg moving virtual objects, pushing a button). On the other side, haptics can benefit from VR. For instance, Lécuyer et al. leverage the users vision and analyse how it affects their haptic feedback (Lécuyer, 2009). This approach, ”pseudo- haptic feedback”, tricks the users’ perception into feeling virtual objects’ stiffness, texture, mass. Many more haptic features can be stimulated, such as temperature, shape, skin stretch, pressure. ## 4\. Analyzing haptic interactions The main objective of this survey is to provide analytical tools to evaluate and compare haptic interactions. ### 4.1. Design space We propose a two-dimension framework to discuss and classify haptic solutions in VR (see Table 1). The first dimension is their degree of physicality, ie how the haptic perception is tangible/physically consistent/resembling with the virtual objects. This dimension is drawn as a continuum, from ”no physicality” to ”real objects” (see Figure 2). We find that this continuum can be discretised as a two-category section: whether they use real objects or not. The second orthogonal dimension is their degree of actuation, ie whether haptic solutions rely on a motor-based hardware implementation enabling autonomous displacements (eg enabling to change its shape, position etc). ### 4.2. Analysis criteria We consider two main criteria to analyse haptic interactions in VR. They cover both the user and designer perspectives. The User experience is the first criterion and includes two aspects: interaction opportunities and visuo-haptic consistency/discrepancy. Interaction opportunities represent to which extent haptic solutions allow users to interact/act (e.g navigate, explore, manipulate) in a VR scene as opposed as in the real world. Visuo-haptic consistency/discrepancy refers to the tactile and kinesthetic perceptual rendering of these interactions. These two sub-criteria are complementary focusing on both action and perception. The second criterion is the conception cost, i.e. the challenges Designers should address when designing haptic interactions. We distinguish implementation and operation costs. Implementation costs include several technical aspects related to the acceptability of a haptic solution such as safety, robustness and ease-of-use (Dominjon et al., 2007). Operation costs include the financial and human costs required to deploy these technologies. ### 4.3. Application We rely on this design space and criteria to highlight and understand the trade-offs between the user’s interactions opportunities in VR, and the designers’ challenges in conception. This survey offers a novel perspective for researchers to study haptic interactions in VR. It can be used to compare and analytically evaluate existing haptic interactions. For a given application, designers can evaluate the most adapted haptic interaction. For a given technique, they can evaluate a haptic solution depending on their needs (tasks, workspace, use-cases etc). .We first discuss haptic interactions from the User perspective (Section 5 \- Interaction opportunities, Section 6 \- Visuo-Haptic Consistency/Discrepancy). We then adopt the designer perspective in Section 7. We use our design space on Sections 6 and 7, which emphasize haptic solutions. We propose two dimensions to classify current technologies: the degree of physicality as well as the degree of actuation. Four categories are hence drawn in this figure: Top Left: No robotics, No real objects; Top Right: Robotics, No real objects; Bottom Right: Robotics, Real objects; Bottom Left: No robotics, Real objects. We respectively displayed current technologies and interaction techniques in the category they belong to. Table 1. We propose two dimensions to classify current technologies: their degrees of physicality and actuation. Figure 1. Tasks in VR: (1) Navigation through Point & Teleport (Funk et al., 2019); (2) Navigation through a building, using redirection (Cheng, 2019); (3) Exploration with Bare-Hands: A user finds an invisible haptic code (Bouzbib et al., 2020); (4) Manipulation: Haptic proxies rearrange themselves to form a plane the user can manipulate (Zhao et al., 2017); (5) Edition: the user changes the shape of a haptic 2.5D tabletop (Nakagaki et al., 2016b); (6) The user is interacted with by a robotic arm to feel emotions (Teyssier et al., 2020). ## 5\. Interaction Opportunities In the real world, users move freely without constraints, pick any object of their environment and then interact with their bare-hands. They also can be interacted with, from the environment (wind, unexpected obstacles) or from other users, for instance to catch their attention or to lead them somewhere. A natural environment also naturally physically constrains users through their entire body. In this section, we discuss the interaction opportunities in VR and the methods available to provide them. In particular, we discuss them through four main tasks: navigation, exploration, manipulation and edition. ### 5.1. Navigation We qualify a navigation task as the exploration of the environment through the vision and the ability to navigate through it via the users displacements. We identify three main techniques to navigate in VR. The two firsts rely on controllers and push buttons, where the users do not necessary physically move. The last one is more natural as it allows the users to walk in the VR arena. #### 5.1.1. Panning: With grounded desktop haptic solutions, such as the Virtuose (Haption, 2019), users need to push a button to clutch the devices and hence move within the environment. #### 5.1.2. Point & Teleport: With ungrounded solutions, such as controllers, the common technique is teleportation. Users point their controllers (Baloup et al., 2018) to predetermined teleportation target areas, and are displaced in position but also in orientation (Funk et al., 2019) (Figure 1 \- 1). #### 5.1.3. Real Walking: Real walking in VR, ”perambulation”, has shown the best immersion and presence results (Usoh et al., 1999; Steinicke et al., 2013) because it relies on proprioception and kinesthetic feedback through the legs and gait awareness. Nonetheless, VR arenas are not infinite and HMD have a limited tracking space, hence methods need to be developed for the user to be able to move to any location of interest. One approach is to mount the previously discussed grounded desktop haptic solutions over mobile (Satler et al., 2011; Lee et al., 2007; Formaglio et al., 2005; Lee et al., 2009; Nitzsche et al., 2003; Pavlik et al., 2013) or wearable (Barnaby and Roudaut, 2019) interfaces. Users however still have to continuously maintain the handle in their palm. Other interfaces hence allow for free-hands Room-Scale VR (Bouzbib et al., 2020; Wang et al., 2020c; Yixian et al., 2020). For the users to perambulate in an infinite workspace, the virtual environment can also visually be warped for the users to unconsciously modify their trajectory or avoid obstacles (Cheng, 2019; Razzaque et al., 2001; Yang et al., 2019) (Figure 1 \- 2). This infinite redirection can also be provided from Electro-Muscle Stimulation (EMS) on the users’ legs (Auda et al., 2019), with wearable electrodes. The user can also wear actuated stilts to perceive staircases (Schmidt et al., 2015) or a vibrating shoe to perceive virtual materials (Strohmeier et al., 2020). To remain unencumbered from these wearable techniques, the VR arena can also include robotised techniques: users can for instance walk on treadmills (Vonach et al., 2017; Frissen et al., 2013), or on movable platforms that encounter their feet (Iwata, 2005, 2013). ### 5.2. Hand Interactions In the real world, bare-hands interaction is important to execute everyday tasks (exploration, manipulation, edition). However, in VR, users commonly have to hold controllers, wearables or handles, which create a discrepancy between what the users feel and see (Yokokohji et al., 1999). These exploit the God-object principle (Zilles and Salisbury, 1995), as opposed to bare- hands Real-touch interactions. #### 5.2.1. God-Object: The controller is considered as a continuity of the users’ hands, represented by a proxy that does not undergo physics or rigid collisions, and is attached to a complementary rigid object with a spring-damper model. This latter hence moves along with the proxy, but is constrained by the environment. Whenever it does collide with an object of interest, the users perceive the previous spring-damper stiffness through kinesthetic feedback. Users hence interact though a proxy, like a desktop mouse, which position is not co-located with the users’ vision. Bare-hands interactions are not necessarily needed depending on the use-cases. For instance, in healthcare and surgery training, users are more likely to interact with a tool, such as a scalpel or a clamp. Continuously holding the god-object is hence not a constrain, however the co-location of vision and haptics is recommended (Ortega and Coquillart, 2005). #### 5.2.2. Real Touch: In other scenarios, such as gaming, industry or tool training (Winther et al., 2020; Strandholt et al., 2020), using the appropriate tools through props and real objects is more natural. The users however need to be able to reach them whenever required. Some interfaces (e.g. Robotic Graphics; see Section 7.3) are hence developed in these regards, to encounter the users whenever they feel like interacting. ### 5.3. Exploration As opposed to the previous definition of ”navigation”, based on vision cues, an ”exploration” task consists in the ability to touch the environment and understand its constraints. Exploring thoroughly an environment in VR can be done through different haptic features, and can improve the users depth perception (Makin et al., 2019) or distances to an object. The different methods for exploring the environment are detailed in Section 6. Whenever a user is exploring the environment, shapes or textures are felt through his body displacements. He needs to move for his skin to stretch (through tactile cues) or his muscles to contract (through kinesthetic cues). #### 5.3.1. Through Tactile cues: Whenever real props or material patches are available, users can naturally interact with their fingertips to feel different materials (Degraen et al., 2019; Araujo et al., 2016), textures (Benko et al., 2016; Lo et al., 2018), temperatures (Ziat et al., 2014) or to feel shapes and patterns through their bare-hands (Bouzbib et al., 2020; Cheng et al., 2017) (Figure 1 \- 3). When no physicality is available, a stimulation can still be performed. As seen in Surface haptic displays (Bau et al., 2010), vibrations between 80 to 400 Hz are felt through the skin, hence users perceive stickiness, smoothness, pleasure, vibration or friction, and for instance explore a 3D terrain or volumetric data (Sinclair et al., 2014). Vibrations can then be combined with auditory and vision cues to render collisions in VR (Boldt et al., 2018). #### 5.3.2. Through Kinesthetic cues: Exploring the environment can also be done through kinesthetic cues: the users can literally be physically constrained to feel a wall, using electro-muscle stimulation (EMS) for instance (Lopes et al., 2017). With the god-object principle, users can also explore the environments’ constraints through force- feedback. In this configuration, the users’ arms are constrained by haptic desktop interfaces, providing strong enough forces to simulate a physical collision and discriminate shapes. ### 5.4. Manipulation A manipulation task is performed whenever modifying the position and orientation of an object. #### 5.4.1. Direct Manipulation: In VR, we distinguish the direct manipulation (Bryson, 2005), ”the ability for a user to control objects in a virtual environment in a direct and natural way, much as objects are manipulated in the real world” from pointing/selecting an object with controllers. A direct manipulation relies on the ability to hold an object with kinesthetic feedback, feel its weight (Lopes et al., 2017; Zenner and Krüger, 2019; Heo et al., 2018; Sagheb et al., 2019; Zenner and Kruger, 2017; Shigeyama et al., 2019), shape (Follmer et al., 2013; Sun et al., 2019; Kovacs et al., 2020), and constrains from the virtual environment, for instance when making objects interact with each other (Bouzbib et al., 2020). Changing a virtual object position or orientation can be used as an input in the virtual environment: in (Zhao et al., 2017) for instance, the user modifies a light intensity by moving a handle prop in the real environment. By transposing (Lopes et al., 2015) in VR, an object could even communicate its dynamic use to the user. #### 5.4.2. Pseudo-Haptic Manipulation: Leveraging vision over haptics allows to move an object with different friction, weights or force perceptions (Rietzler et al., 2018; Samad et al., 2019; Pusch and Lécuyer, 2011; Rietzler et al., 2019). For instance, visually reducing the speed of a virtual prop displacement leads to an increase in the users’ forces to move it, modifying their friction/weight perceptions. ### 5.5. Edition We qualify an Edition task as a modification of an object property, other than its orientation or position (for example through its scale (Xia et al., 2018) or shape). #### 5.5.1. Physical Edition: Editing an interface in VR requires it to be fully equipped with sensors. With wearables for instance, the hand phalanges positions are known, and can be tightly linked with an object property (Villa Salazar et al., 2020). Knowing their own position, modular interfaces can be rearranged to provide stretching or bending tasks (Feick et al., 2020), or be pushed on with a tool to reduce in size (Teng et al., 2019). Shape-changing interfaces have been developed to dynamically modify material properties (Nakagaki et al., 2016b) (Figure 1 \- 5) or augment the interactions in Augmented Reality (AR) (Leithinger et al., 2013), however these techniques only consider HMDs and VR as future work directions. These interfaces are relevant as 2.5D tabletops are already used in VR. Physically editing the virtual world through them could be implemented in a near future, by intertwining these interfaces with 3D modelling techniques (De Araújo et al., 2013). #### 5.5.2. Pseudo-Haptic Edition: The difficulty behind changing a real object property is to track it in real- time. This is why pseudo-techniques are relevant: they visually change the object properties such as their shape (Achibet et al., 2017), compliance (Lee et al., 2019; Sinclair et al., 2019), or their bending curvature (Heo et al., 2019) without physically editing the object. ### 5.6. Scenario-based Interactions In the real world, humans are free to interact with any object without further notice. In this regard, common controllers enable interactions with any object through pointing, but they display a high visuo-haptic discrepancy. In more advanced haptically rendered Virtual environments, users are often constrained to scenario-based interactions: only a few interactable objects are available, accordingly with the scenario’s progress. The greater the virtual:physical haptic consistency, the harder it is to enhance non-deterministic scenarios, where the user is free to interact with any object with no regards to the scenario’s progress. High quality haptic rendering in non-deterministic scenarios can be achieved through three methods: (a) numerous objects and primitives are available for interactions (Hettiarachchi and Wigdor, 2016); (b) the users’ intentions are to be predicted prior to interaction to make it occur (Bouzbib et al., 2020; Cheng et al., 2017); (c) props modify their own topology to match the users expected haptic rendering (Siu et al., 2018). ### 5.7. Environment-Initiated Interactions In both real and virtual environments with tangible interfaces, users usually are the decision makers and get to choose their points of contact during the next interaction. However, users themselves can be considered as tangible interfaces: uncontrolled interactions, such as being touched by a colleague, or feeling a temperature change in the environment (Shaw et al., 2019; Ziat et al., 2014), are part of everyday interactions that can be transposed in Virtual Reality. Replicating a social touch interaction in VR for instance increases presence (Hoppe et al., 2020) or invokes emotions (Teyssier et al., 2020). This type of interactions are recurrent in sports simulations, where the user is undergoing forces from his environment and perceiving impacts (jumping into space (Gugenheimer et al., 2016), shooting a soccer ball (Wang et al., 2020b), goalkeeping in a soccer game (Tsai and Chen, 2019), paragliding (Ye et al., 2019), intercepting a volleyball (Günther et al., 2020), flying (Cheng et al., 2014)). These interactions are involving multiple force types: tension, traction, reaction, resistance, impact that help enhancing the user experience in VR (Wang et al., 2020c). These can be strong enough to even lead the user through forces (Bouzbib et al., 2020). ### 5.8. Whole-Body Involvement All the previous subsections evoke interactions that mainly involve the hands or the fingers. This paradigm is revoked in (Zielasko and Riecke, 2020): a user should be able to choose his posture. This is currently only enabled in room-scale VR applications, where users experience sitting, standing, climbing or crouching (Teng et al., 2019; Suzuki et al., 2020; Bouzbib et al., 2020; Danieau et al., 2012) and interact with their whole-body. Figure 2. Degree of physicality continuum in VR.(1) Haptic desktop devices enable to explore the environment through a handle (Lee et al., 2009) with the god-object principle; (2) A controller (Benko et al., 2016) or (3) a wearable (Fang et al., 2020) simulate objects for exploration tasks; (4) Mid-air technology (Rakkolainen et al., 2020) create vibrations through the user’s hand to simulate an object; (5) Passive proxies are oriented for the user to feel objects’ primitives with their hands (Cheng et al., 2017); (6) Objects from the environment are assigned to virtual props with the same primitives (Hettiarachchi and Wigdor, 2016); (7) Real objects or passive props can be manipulated and interacted with each other (Bouzbib et al., 2020). Three physical devices simulating objects are shown with their virtual counterparts. (1) An inflatable prop in the user’s palm simulates holding a bomb. (2) A pin-based interface shaped as a ball interacts in the user palm to replicate a hamster. (3) Different primitives (ball, cube, pyramid) are displayed on a 2.5D tabletop. Figure 3. Simulating Objects. (1) A controller with an inflatable prop in the user’s palm simulates holding a bomb (Teng et al., 2018). (2) A pin-based interface shaped as a ball interacts in the user palm to replicate a hamster (Yoshida et al., 2020). (3) Different primitives (ball, cube, pyramid) are displayed on a 2.5D tabletop (Siu et al., 2018). ## 6\. Visuo-Haptic Consistency/Discrepancy Visuo-Haptic Consistency is the second aspect of the user experience. We exploit the dimension degree of physicality of our design space (Table 1) to discuss the different haptic solutions. In particular, we distinguish whether these solutions use real objects (exploiting real objects) or not (simulating objects). ### 6.1. Simulating Objects Object properties that need to be simulated are their shape, texture, temperature, weight. #### 6.1.1. No Physicality, (Figure 2 \- 1) Currently, grounded haptic devices such as the Virtuose (Haption, 2019) or the PHaNToM (Massie and Salisbury, 1994) simulate objects through their shapes (Figure 2 \- 1). The rendering is only done through kinesthetic feedback via a proxy. Conceptually, the ideal link between the users and this proxy is a massless, infinitely rigid stick, which would be an equivalent to moving the proxy directly (Hayward and Maclean, 2007; Sato, 2002). These solutions only provide stimulation at the hand-scale, with no regards to the rest of the body. #### 6.1.2. Shape Simulation, (Figure 2 \- 2-3-4) In the same regard, gloves or controllers provide some physicality (Figure 2 \- 2-3). Gloves or exoskeletons literally constrain the users hands for simulating shapes (Fang et al., 2020; Gu et al., 2016; noa, 2019a; Amirpour et al., 2019; Choi et al., 2017; Tsai and Rekimoto, 2018; Achibet et al., 2015; Choi et al., 2016; Nakagaki et al., 2016a; Achibet et al., 2014; Provancher et al., 2005), or stimulate other haptic features such as stiffness, friction (Villa Salazar et al., 2020) or slippage (Tsagarakis et al., 2005). These can be extended to overall body suits for users to feel impacts or even temperature changes (noa, 2019b; Danieau et al., 2018), or even intertwined with grounded devices to extend their use-cases (Steed et al., 2020). Customised controllers are currently designed to be either stimulating the palm (Sun et al., 2019; Yoshida et al., 2020; de Tinguy et al., 2020) (Figure 3 \- 1, 2), or held in the palm while providing haptic feedback on the fingertips. For instance, (Whitmire et al., 2018) proposes interchangeable haptic wheels with different textures or shapes, while (Benko et al., 2016) enables textures and shapes and (Lee et al., 2019) displays compliance changes. In these configurations, users hold a single controller, however bi- manual interactions can be created by combining two controllers. Their link transmits kinesthetic feedback, and constrain their respective positions to each other (Strasnick et al., 2018; Wei et al., 2020). Contactless technology has also been developed for simulating shapes. While studies demonstrated that interacting with bare-hands increased the user’s cognitive load (Galais et al., 2019), combining bare-hands interactions with haptic feedback actually enhances the users involvement. Since haptic feedback does require contact, ”contactless” technology defines an interaction where the users are unencumbered, as per Krueger’s postulate (Wexelblat, 1993), and ultrasounds are sent to their hands, for them to perceive shapes on their skin, without a physical prop contact (Rakkolainen et al., 2020) (Figure 2 \- 4). These unencumbered methods are also achieved through shape-changing interfaces, for instance with balloons arrays (Takizawa et al., 2017) or 2.5D tabletops (Figure 3 \- 3, Figure 1 \- 5) (Follmer et al., 2013; Siu et al., 2018; Iwata et al., 2001). These latter are constituted from pins, that raise and lower themselves to replicate different shapes. In the same regard, swarm interfaces rearrange themselves to display different shapes. These have mainly been developed in the real world (Le Goc et al., 2016; Suzuki et al., 2018; Kim et al., 2020; Suzuki et al., 2019; Ducatelle et al., 2011; Marquardt et al., 2009) but slowly take off as VR user interfaces (Zhao et al., 2017) (Figure 1 \- 4). Indeed, while these latter devices are used as desktop interfaces, the swarm robot idea has extended to the air, with drones for instance (Gomes et al., 2016; Rubens et al., 2015; Knierim et al., 2017; Hoppe et al., 2018; Tsykunov and Tsetserukou, 2019). All of these previous interfaces embrace the Roboxel principle enunciated in Robotic Graphics (McNeely, 1993): ”cellular robots that dynamically configure themselves into the desired shape and size”. #### 6.1.3. Object Primitives, (Figure 2 \- 5) Finally, a user can interact with object primitives. These represent the simplest geometries available: circle, cube, pyramid, cylinder, torus. Simply feeling an orientation through the fingertips provides the required information to understand an object shape, in an exploration task for instance. Panels with diverse orientations can hence be displayed for a user to explore various objects in a virtual environment (Cheng et al., 2017) (Figure 2 \- 5) or directly encounter the user at their position of interest (Yokokohji et al., 2005; Yokokohji et al., 2001). On the opposite, a bare-hands manipulation task requires multiple primitives to be available at the same time within the hand vicinity. This is why the exploitation of real objects is necessary. ### 6.2. Exploiting Real Objects Passive haptics (Insko, 2001), ie the use of passive props, consist in placing real objects corresponding to their exact virtual match at their virtual position. Insko demonstrated that passive haptics enhanced the virtual environment (Insko, 2001). Nonetheless, this does suffer from a main limitation: substituting the physical environment for a virtual one (Simeone et al., 2015) requires a thorough mapping of objects shapes, sizes, textures, and requires numerous props (Pair et al., 2003). This can be done with real objects in simulation rooms for instance (e.g plane cockpit, motorcycle), but cheaper methods need to be implemented to facilitate their use in other fields. #### 6.2.1. Object Primitives, (Figure 2 \- 6) One solution is to extract the primitives of the objects that are already available in the physical environment, to map virtual objects of the approximate same primitive over them (Hettiarachchi and Wigdor, 2016) (Figure 2 \- 6). #### 6.2.2. Visuo-Proprioceptive Illusions & Pseudo Haptics The number of props within the environment can also be reduced, while letting the users interact at different positions of the physical world. It is possible to leverage the vision over haptics and modify the users’ proprioception to redirect their trajectory (Kohli, 2010; Kohli et al., 2012, 2013; Azmandian et al., 2016; Gonzalez and Follmer, 2019; Han et al., 2018). A user might perceive multiple distinct cubes for instance, while interacting with a single one. On the same principle, the user hand displacement can be redirected at an angle, up-/down-scaled (Abtahi and Follmer, 2018; Bergström et al., 2019), or slowed down for friction or weight perception (Samad et al., 2019; Praveena et al., 2020). These techniques also allow for the exploration and manipulation of various shapes: models can for instance be added to enable complex virtual shapes to be mapped over real physical objects boundaries (Zhao and Follmer, 2018). The user can also be redirected to pinch a multi- primitive object (cubic, pyramidal and cylindrical) from different locations, which theoretically widens the variety of available props with a single one (de Tinguy et al., 2019). On the same principle, pseudo-haptics allow to modify the users’ shape (Ban et al., 2012a; Ban et al., 2012b) or texture (Degraen et al., 2019) perceptions when interacting with a physical prop. #### 6.2.3. Displacing Objects, (Figure 2 \- 7) Whenever objects are indeed available within the environment, various directions are available to displace them. This displacement allows for mapping one physical object over multiple ones, but also to display a multitude of props. These directions embrace the Robotic Shape Display principle from Robotic Graphics (McNeely, 1993): ”a robot that can reach any location of a virtual desktop with an end-effector” and matches the user’s object of interest. Their usability have been validated through a Wizard-of-Oz implementation, where human operators move real objects or even people around a Room-scale VR arena to encounter the users (Cheng et al., 2015) (Figure 4 \- 2). The users themselves can also reconfigure and actuate real props (Cheng et al., 2018). Robotic Shape Displays, RSDs, are also called encountered-type of haptic devices, as they literally encounter the users at their object of interest to provide haptic feedback. They allow to display real pieces of material (Araujo et al., 2016; Abtahi et al., 2019), physical props to simulate walls (Bouzbib et al., 2020; Kim et al., 2018; Yamaguchi et al., 2016), or even display furniture (Suzuki et al., 2020) or untethered objects (He et al., 2017a, b; Huang et al., 2020; Bouzbib et al., 2020), that can be interacted with each other. Figure 4. Degree of Actuation. (1) No actuation is available. The user’s hand is redirected to touch a passive prop that cannot move (Azmandian et al., 2016). The implementation of this technique relies exclusively on a software development leveraging the vision cues; (2) Human actuators are used to illustrate the Robotic Graphics (McNeely, 1993) principle with a Wizard of Oz technique (Cheng et al., 2015). They carry props for the user to feel a real continuous wall; Encountered-type of haptic devices (3-5): (3) A drone encounters the users’ hand for exploring passive props; (4) A cartesian robot displaces itself autonomously for users to interact with physical props (Bouzbib et al., 2020); (5) A robotic arm with multiple degrees of freedom displaces itself to encounter the users’ hand, and rotates its shape- approximation device to provide the right material (Araujo et al., 2016). ## 7\. Conception cost In practice, designers have to trade-off their interaction design space with implementation and operational costs in the conception phase. Implementation costs include technical aspects related to the acceptability of an haptic solution such as safety, robustness and ease-of-use (Dominjon et al., 2007). For instance, actuated haptic solutions require a special attention regarding this criterion. Operation costs include the financial and human cost for using a haptic solution. The financial cost is measured through the cost of the haptic device and additional elements such as motion capture systems to precisely track the users’ hand or the prior preparation of required props. Human cost refers to both labour time and number of human operators required during the user’s interactions. For instance, actuated haptic solutions generally do not require human operators (low human cost) but might be mechanically expensive. In this section, we use our two-dimension design space (Table 1) to discuss haptic solutions according to their conception cost. As non-actuated solutions globally share the same approaches and have a low implementation cost, we discuss them together in the ”No Robotics” subsection. ### 7.1. No Robotics Regarding implementation costs, all non-actuated haptic solutions are safe, robust and easy-to-use. We depict here an important design choice when opting for these solutions: either the designer relies on graphics solutions, leveraging vision cues over haptic ones, or needs operators to displace or change the interactable props (see Table 1). #### 7.1.1. Passive Props. Passive props (Insko, 2001) only consist in placing real objects corresponding to their exact virtual match at their virtual position. They provide a natural way of interacting through the objects’ natural affordances (Norman, 2013). They however are limited to the available objects within the scene as they are not actuated. They only can be used in a scenario-based experience, where the target is known in advance. The environment hence requires a prop for each available virtual object. #### 7.1.2. Shape Simulation, Pseudo-Haptics, Visuo-Haptic Illusions, Object Primitives. For graphics solutions, users are redirected towards their object of interest (Azmandian et al., 2016) using visuo-haptic illusions. However, physically overlaying a prop or primitive over a virtual object has a tracking cost, which usually relies on trackers which can be operationally costly (eg Optitrack (Optitrack, 2019) or HTC Trackers). Otherwise, the users intentions have to be predicted for the interaction to occur. The users hands are then redirected to the appropriate motionless prop, for them to explore their object of interest (Cheng et al., 2017). Operationally, the cost only relies on the proxy fabrication (Figure 2 \- 5). These implementations offer various scenarios in terms of interaction (even non-deterministic), at an affordable cost. #### 7.1.3. Surface Haptic Displays. These techniques exclusively allow for exploration through multiple haptic features such as friction or textures. They also can integrate a tablet or a smartphone (Savino, 2020), on which the user can interact at any location. #### 7.1.4. Human Actuators. This technique consists in using human operators to displace props in the VR arena. The designers however come across reliability and speed issues with these operators. Even though they only are used in scenario-based experiences, delay mechanisms based on graphics need to be implemented (Cheng et al., 2015) (Figure 4 \- 2) to overcome these issues. Conceptually, they broaden the interaction scope, however this solution is operationally very costly. #### 7.1.5. Real Props Reassignment. Instead of using a tracking system for passive props, a depth camera for instance allows to reassign props to different virtual objects of the same primitive (Hettiarachchi and Wigdor, 2016) (Figure 2 \- 6). The objects are hence all available to be interacted with. This drastically reduces the operational costs as they only rely on computer vision. This enables non- deterministic scenarios as the real world is literally substituted for a virtual one (Simeone et al., 2015) and objects can be reassigned with virtual:physical (He et al., 2017a) mappings. ### 7.2. Robotics & No Real Objects This section gathers technologies simulating the virtual environment through actuation: they replicate it to constrain the users. #### 7.2.1. Desktop Haptic Interfaces. The SPIDAR (Sato, 2002), the Virtuose (Haption, 2019) and other classic desktop haptic interfaces are already compared in multiple surveys (Dominjon et al., 2007; Seifi et al., 2019; Wang et al., 2020a) (see Figure 2 \- 1). They are safe as they are controlled by the user and only constrain their arm movements with kinesthetic feedback and adapt to any available object from the virtual scene (non-deterministic scenarios). They show a high perceived stiffness and robustness, but remain really expensive (¿10k$). #### 7.2.2. Shape-Changing Interfaces, Roboxels, 2.5D Tabletops. These technologies present a high perceived stiffness and change their shapes accordingly with the virtual environment (Fitzgerald and Ishii, 2018; Leithinger et al., 2013). They hence do not require any operator and allow for non-deterministic scenarios whenever their displacements are enabled (Siu et al., 2018) (see Figure 3 \- 3). They are however complex to build and require multiple motors as they are composed of arrays of numerous pins, which define their haptic fidelity resolution. Even though they present high voltages, they remain safe around the users. As they require bare-hands interactions, they hence show a high ease of use. #### 7.2.3. Wearables, Controllers, EMS. These rely on small torques, which are sufficient to constrain the users body parts. They are safe and easy to use, but in return are not robust enough to resist to users’ actions. As they are continuously changing the users’ haptic perception, they do allow non-deterministic scenarios and change their rendered stiffness and rigidity as a function of the distance to a virtual prop (de Tinguy et al., 2020; Kovacs et al., 2020). A customised controller usually relies on 3D printed parts and small servomotors and can be easily replicated (Sun et al., 2019) (Figure 2 \- 2,3; Figure 3 \- 1,2). #### 7.2.4. Mid-Air Haptics. Providing contactless interactions, mid-air haptics also provide a high level of safety around the user. They however do not allow to navigate the VR environment, and hence cannot consider non-deterministic scenarios. Their robustness is very low, as they send ultrasounds to the users and do not physically constrain them (Rakkolainen et al., 2020). #### 7.2.5. Inflatable Floor. The floor topology can be modified and inflated to create interactions at the body-scale (Teng et al., 2019). The users cannot inflate them, however they can push some tiles down and hence, edit them. These are safe, though they do not provide a wide range of interactions, but offer multiple static body postures. ### 7.3. Robotics & Real Objects In this subsection, we detail the different types of Robotic Shapes Displays - otherwise known as ”encountered-type of haptic devices”, mentioned in the Table 1. First, these interfaces move to encounter the users: this feature optimises their ease of use. Second, as these interfaces move within the user vicinity, safety concerns are raised in this section, depending on the interfaces robustness. Encountered-type of haptic devices combine different types of interaction techniques: they can provide the users with passive props, textures or primitives, and allow navigation, exploration, manipulation tasks. Their mechanical implementations offer a good repeatability and reliability. #### 7.3.1. Cartesian Robot: In (Bouzbib et al., 2020), CoVR, a physical column mounted over a Cartesian XY ceiling robot enables interactions at any height and any position of a room- scale VR arena (see Figure 1 \- 2; Figure 4 \- 4). This implementation presents a high perceived stiffness, and because it carries passive props around the arena, enables a high fidelity haptic rendering. It displays high accuracy and speed, and presents an algorithm which optimises the column’s displacements as a function of the users intentions. It hence enables non- deterministic scenarios. Safety measures have been validated in the field. In practice, the column’s celerity is decreasing around the user, as it is repulsed by this latter. Its software implementation ensures a safe environment for the user to perambulate in the arena without unexpected collision. However, in order to display many props in different scenarios, an operator is required to create panels and modify them. The materials however remain cheap, and even though its structure and motors are more expensive than 3D printed cases and servomotors, as per customised controllers for instance, this solution provides a wide range of interactions. #### 7.3.2. Robotic Arm: A robotic arm provides more degrees of freedom than the previous Cartesian robot. This primarily means a higher cost and a higher safety risk. For instance, H-Wall, using a Kuka LBR Iiwa robot, presents high motor torques and can hence increase the safety risks around the users. This implementation hence does not allow non-deterministic scenarios, and presents either a wall or a revolving door to the user, with a high robustness. Implementations with smaller torques, such as (Vonach et al., 2017; Araujo et al., 2016) are safer but display a reduced perceived stiffness. The use-cases for all these interactions are hence drastically different: H-Wall simulates a rigid wall while VRRobot (Vonach et al., 2017) and Snake Charmer (Araujo et al., 2016) (Figure 4 \- 5) present more interaction opportunities. This latter is also the single Robotic Shape Display that autonomously changes its end-effector, without an operator. #### 7.3.3. Drones, Swarm Robots, Mobile Platforms: With drones, the interactions are limited to the available props, for instance with a single wall at a given position (Yamaguchi et al., 2016). Going from an active mode (flying) to a passive one (graspable by the user) has a long delay (10s) (Abtahi et al., 2019), which on top of the safety concerns, does not allow non-deterministic scenarios. (Tsykunov et al., 2019) however allows the user to change the drone trajectory to fetch and magnetically recover an object of interest. Their accuracy and speed are limited (Gomes et al., 2016; Rubens et al., 2015) compared to the previous grounded interfaces, and can require dynamic redirection techniques to improve their performances (Abtahi et al., 2019). As they are ungrounded, they do not have a high robustness nor perceived stiffness. This is also valid for mobile robots, such as (He et al., 2017a; Gonzalez et al., 2020), which only display passive props. To decrease the conception cost, existing vacuuming robots are used as mobile platforms in (Wang et al., 2020c; Yixian et al., 2020). Designers can choose to duplicate them, as swarm robots, to enable non-deterministic scenarios (Suzuki et al., 2020). These are safe to use around the users, as their speed and robustness are limited. Instead of swarm mobile interfaces, a merry-go-round platform can also be designed to display various props at an equidistant position from the user (Huang et al., 2020). All of the previous interfaces require an operator cost on top of their mechanical and software ones, to modify the interactable props available, depending on the use-cases. On the opposite, (Zhao et al., 2017) proposes autonomous reconfigurable interfaces intertwining both Robotic Shape Displays and Roboxels (McNeely, 1993) principles to get rid of the operator cost (see Figure 1 \- 4). These small robotic volume elements reconfigure themselves into the users objects of interest. They have a sufficient perceived stiffness to represent objects, but are not robust enough to resist to body-scaled forces, for instance to simulate a rigid wall. ## 8\. Evaluation Protocols On top of choosing from the different trade-offs between conception and interaction opportunities, the designer also needs to pick-up an evaluation protocol. These protocols depend on the VR use-cases. For instance, the haptic benefits for medical or industrial assembly training can be evaluated against a real experience condition (Poyade et al., 2012), with criteria such as completion time, number of errors, user cognitive load (Gutierrez et al., 2010). On the opposite, the haptic benefits for a gaming experience are more likely to be evaluated through immersion and presence, comparing ”with/without haptics” conditions (Cheng et al., 2015). Although some papers do compare multiple haptic displays (Escobar-Castillejos et al., 2016; Ullrich, 2012), we point out the lack of referenced evaluation protocols for evaluating haptic solutions in VR. ### 8.1. Current Reference Evaluation Methods The most common evaluation methods in VR are the SUS or WS presence questionnaires (Witmer and Singer, 1998; Slater et al., 1994). These questionnaires mainly focus on graphics rendering and only two Likert-scale questions actually focus on haptic feedback: ”How well could you actively survey the VE using touch?” and ”How well could you manipulate objects in the VE?”. Besides, most of the above technologies are evaluated against ”no haptic feedback”, hence the results can seem biased and most of all, expected. This justifies why some implementations provide results on single parts of the questionnaire, or arbitrarily combine their results (Choi et al., 2018) with new subsections (eg ”ability to examine/act”) or tasks specific questions (eg ”How realistic was it to feel different textures?). Table 2. Comparison & Evaluation of 4 Encountered-type of Haptic Devices, according to the ”Evaluation section” parameters. ### 8.2. Evaluation Recommendations Haptics should be more incorporated into the different factors enunciated in (Witmer and Singer, 1998) (”Control, Sensory, Distraction, Realism”). In this direction, Kim et al. defined the Haptic Experience model (Kim and Schneider, 2020), where they take into account both of the designer and user experiences. It depicts how Design parameters (”timeliness, intensity, density and timbre”) impact Usability requirements (”utility, causality, consistency, saliency”) and target Experiential dimensions (”harmony, expressivity, autotelics, immersion, realism”) on the user’s side. In the same regards, we propose additional guidelines to evaluate haptic solutions in VR experiments (see Table 2). We believe that the different elements of interaction opportunities should be added to the users control parameters. In the sensory factors, the number of haptic features available should be added (eg shape, texture, friction, temperature), in line with their quality, in terms of ”timeliness, intensity, density and timbre”. The usability requirements should identify the use-cases and number of scenarios with the proposed solutions. Hence, a good evaluation of the interface timeliness and usability should anticipate future deployments and avoid unnecessary developments. ## 9\. Examples: Encountered-Type of Haptic Devices We propose in this section to compare four encountered-type of haptic devices: Beyond the Force (BTF) drone (Abtahi et al., 2019) (Figure 4 \- 3), ShapeShift (Siu et al., 2018) (Figure 3 \- 3), Snake Charmer (Araujo et al., 2016) (Figure 4 \- 5), and CoVR (Bouzbib et al., 2020) (Figure 4 \- 4). In terms of interactions and number of props, the drone is the most limited one. Indeed, because of both safety and implementation limitations, it only enables free navigation in a reduced workspace. It also allows exploration (through textures) and manipulation tasks. However, the manipulation task is at the moment limited to a single light object as BTF cannot handle large embedded masses yet. Whenever grabbed, it does not provide a haptic transparency (Hayward and Maclean, 2007) during the interactions because of its thrust and inertia. For the users to perform different tasks, an operator needs to manually change the drone configuration. Its mechanical implementation does not provide a sufficient speed for overlaying virtual props in non-deterministic scenarios, but its accuracy is also unsatisfactory and requires dynamic redirection techniques for the interactions to occur. It also provides unwanted noise and wind, which reduces the interaction realism. ShapeShift (Siu et al., 2018) is drastically different: it is a 2.5D desktop interface that displaces itself. Even though a drone is theoretically available in an infinite workspace, in practice they do share approximately the same one. As (Siu et al., 2018) relies on a shape-changing interface, no operator is required and it shape changes itself to overlay the users’ virtual objects of interest, in non-deterministic scenarios. It allows a free navigation at a desktop scale, as well as bimanual manipulation and exploration. Both of these devices haptic transparency are limited as they are ungrounded solutions. We believe that ShapeShift could be updated to allow Edition tasks, by synchronising the users force actions with the actuated pins stiffness. In terms of haptic features, it simulates shapes and stimulates both tactile and kinesthetic cues. As per all 2.5D tabletops, it can be used in various applications: 3D terrain exploration, volumetric data etc. Its resolution seems promising as its studies shows successful object recognition and haptic search. The same interactions are available at a desktop scale with Snake Charmer (Araujo et al., 2016), which provides a wide range of props and stimulation, as each of its end-effector include 6 faces with various interaction opportunities (textures to explore, buttons to push, heater and fan to perceive temperature, handle and lightbulb to grasp and manipulate…). It also can change its shape approximation device, SAD (ie its end-effector), autonomously, using magnets. It follows the user hand and orient the expected interaction face of its SAD prior to the interactions: it hence enables non- deterministic scenarios. Besides, Snake Charmer has a promising future regarding its deployment: LobbyBot (noa, [n.d.]), is already in the Renault industry research lab, to enable VR haptic feedback in the automotive industry. Finally, CoVR (Bouzbib et al., 2020) enables the largest workspace as well as the highest range of interactions. The user is free to navigate in a 30 $m^{3}$ VR arena, and CoVR predicts and physically overlays his object of interest prior to interaction. These interactions include tactile exploration, manipulation of untethered objects (full haptic transparency), body postures. Indeed, CoVR is robust enough to resist body-scaled users, and shows over a 100N perceived stiffness and can carry over 80kg of embedded mass. CoVR can also initiate the interactions with the users, and is strong enough to lead the users through forces or even to transport them. Moreover, with the appropriate physical:virtual mapping (He et al., 2017a), one physical prop can overlay multiple virtual ones of the same approximate primitive without redirection techniques. It however requires an operator to create, assemble and display panels on its sides. Room-scale VR becomes more and more relevant, and Snake Charmer could benefit from being attached to an interface such as CoVR. Similarly, intertwining CoVR with a robotic arm autonomously changing its SAD like Snake Charmer or with a shape-changing interface could reduce its operational costs. This would display all of the Robotics Graphics concept capabilities. ## 10\. Conclusion We analysed in this paper haptic interactions in VR and their corresponding haptic solutions. We analyzed them from both the user and designer perspectives by considering interaction opportunities and visuo-haptic consistency, as well as implementation and operation costs. We proposed a novel framework to classify haptic displays, through a two-dimension design space: the interfaces’ degree of physicality and degree of actuation. We then evaluated these latter solutions from an interaction and conception perspectives. Implementation-wise, we evaluated the interfaces robustness, their ease of use as well as their safety considerations. From an operation perspective, we also evaluated the costs of the proposed solutions. This survey highlights the variety of props, tasks and haptic features that a haptic solution can potentially provide in VR. This survey can be used to analytically evaluate the existing haptic interactions. It can also help VR designers to choose the desired haptic interaction and/or haptic solution depending on their needs (tasks, workspace, use-cases etc). We believe that combining multiple haptic solutions benefits the user experience, as it optimises the above criteria. Encountered-type of haptic interfaces were then highlighted as they already combine multiple interaction techniques: they displace passive props in potentially large VR arenas and allow for numerous tasks, such as navigation, exploration, manipulation, and even allow the user to be interacted with. ## References * (1) * noa ([n.d.]) [n.d.]. renault. https://www.clarte-lab.fr/component/tags/tag/renault * noa (2019a) 2019a. CyberGrasp. http://www.cyberglovesystems.com/cybergrasp * noa (2019b) 2019b. Teslasuit | Full body haptic VR suit for motion capture and training. https://teslasuit.io/ * Abtahi and Follmer (2018) Parastoo Abtahi and Sean Follmer. 2018. Visuo-Haptic Illusions for Improving the Perceived Performance of Shape Displays. In _Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems - CHI ’18_. ACM Press, Montreal QC, Canada, 1–13. https://doi.org/10.1145/3173574.3173724 * Abtahi et al. (2019) Parastoo Abtahi, Benoit Landry, Jackie (Junrui) Yang, Marco Pavone, Sean Follmer, and James A. Landay. 2019. Beyond The Force: Using Quadcopters to Appropriate Objects and the Environment for Haptics in Virtual Reality. In _Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems - CHI ’19_. ACM Press, Glasgow, Scotland Uk, 1–13. https://doi.org/10.1145/3290605.3300589 * Achibet et al. (2015) Merwan Achibet, Adrien Girard, Anthony Talvas, Maud Marchal, and Anatole Lecuyer. 2015. Elastic-Arm: Human-scale passive haptic feedback for augmenting interaction and perception in virtual environments. In _2015 IEEE Virtual Reality (VR)_. IEEE, Arles, Camargue, Provence, France, 63–68. https://doi.org/10.1109/VR.2015.7223325 * Achibet et al. (2017) Merwan Achibet, Benoit Le Gouis, Maud Marchal, Pierre-Alexandre Leziart, Ferran Argelaguet, Adrien Girard, Anatole Lecuyer, and Hiroyuki Kajimoto. 2017. FlexiFingers: Multi-finger interaction in VR combining passive haptics and pseudo-haptics. In _2017 IEEE Symposium on 3D User Interfaces (3DUI)_. IEEE, Los Angeles, CA, USA, 103–106. https://doi.org/10.1109/3DUI.2017.7893325 * Achibet et al. (2014) Merwan Achibet, Maud Marchal, Ferran Argelaguet, and Anatole Lecuyer. 2014. The Virtual Mitten: A novel interaction paradigm for visuo-haptic manipulation of objects using grip force. In _2014 IEEE Symposium on 3D User Interfaces (3DUI)_. IEEE, MN, USA, 59–66. https://doi.org/10.1109/3DUI.2014.6798843 * Alexandrovsky et al. (2020) Dmitry Alexandrovsky, Susanne Putze, Michael Bonfert, Sebastian Höffner, Pitt Michelmann, Dirk Wenig, Rainer Malaka, and Jan David Smeddinck. 2020. Examining Design Choices of Questionnaires in VR User Studies. In _Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems_. ACM, Honolulu HI USA, 1–21. https://doi.org/10.1145/3313831.3376260 * Amirpour et al. (2019) E. Amirpour, M. Savabi, A. Saboukhi, M. Rahimi Gorii, H. Ghafarirad, R. Fesharakifard, and S. Mehdi Rezaei. 2019. Design and Optimization of a Multi-DOF Hand Exoskeleton for Haptic Applications. In _2019 7th International Conference on Robotics and Mechatronics (ICRoM)_. 270–275. https://doi.org/10.1109/ICRoM48714.2019.9071884 ISSN: 2572-6889. * Araujo et al. (2016) Bruno Araujo, Ricardo Jota, Varun Perumal, Jia Xian Yao, Karan Singh, and Daniel Wigdor. 2016\. Snake Charmer: Physically Enabling Virtual Objects. In _Proceedings of the TEI ’16: Tenth International Conference on Tangible, Embedded, and Embodied Interaction - TEI ’16_. ACM Press, Eindhoven, Netherlands, 218–226. https://doi.org/10.1145/2839462.2839484 * Auda et al. (2019) Jonas Auda, Max Pascher, and Stefan Schneegass. 2019. Around the (Virtual) World: Infinite Walking in Virtual Reality Using Electrical Muscle Stimulation. In _Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems - CHI ’19_. ACM Press, Glasgow, Scotland Uk, 1–8. https://doi.org/10.1145/3290605.3300661 * Azmandian et al. (2016) Mahdi Azmandian, Mark Hancock, Hrvoje Benko, Eyal Ofek, and Andrew D. Wilson. 2016. Haptic Retargeting: Dynamic Repurposing of Passive Haptics for Enhanced Virtual Reality Experiences. In _Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems - CHI ’16_. ACM Press, Santa Clara, California, USA, 1968–1979. https://doi.org/10.1145/2858036.2858226 * Baloup et al. (2018) Marc Baloup, Veïs Oudjail, Thomas Pietrzak, and Géry Casiez. 2018. Pointing techniques for distant targets in virtual reality. In _Proceedings of the 30th Conference on l’Interaction Homme-Machine - IHM ’18_. ACM Press, Brest, France, 100–107. https://doi.org/10.1145/3286689.3286696 * Ban et al. (2012a) Y. Ban, T. Kajinami, T. Narumi, T. Tanikawa, and M. Hirose. 2012a. Modifying an identified curved surface shape using pseudo-haptic effect. In _2012 IEEE Haptics Symposium (HAPTICS)_. 211–216. https://doi.org/10.1109/HAPTIC.2012.6183793 * Ban et al. (2012b) Yuki Ban, Takuji Narumi, Tomohiro Tanikawa, and Michitaka Hirose. 2012b. Modifying an identified position of edged shapes using pseudo-haptic effects. In _Proceedings of the 18th ACM symposium on Virtual reality software and technology - VRST ’12_. ACM Press, Toronto, Ontario, Canada, 93. https://doi.org/10.1145/2407336.2407353 * Barnaby and Roudaut (2019) Gareth Barnaby and Anne Roudaut. 2019. Mantis: A Scalable, Lightweight and Accessible Architecture to Build Multiform Force Feedback Systems. In _Proceedings of the 32nd Annual ACM Symposium on User Interface Software and Technology - UIST ’19_. ACM Press, New Orleans, LA, USA, 937–948. https://doi.org/10.1145/3332165.3347909 * Bau et al. (2010) Olivier Bau, Ivan Poupyrev, Ali Israr, and Chris Harrison. 2010. TeslaTouch: electrovibration for touch surfaces. In _Proceedings of the 23nd annual ACM symposium on User interface software and technology - UIST ’10_. ACM Press, New York, New York, USA, 283\. https://doi.org/10.1145/1866029.1866074 * Benko et al. (2016) Hrvoje Benko, Christian Holz, Mike Sinclair, and Eyal Ofek. 2016\. NormalTouch and TextureTouch: High-fidelity 3D Haptic Shape Rendering on Handheld Virtual Reality Controllers. In _Proceedings of the 29th Annual Symposium on User Interface Software and Technology - UIST ’16_. ACM Press, Tokyo, Japan, 717–728. https://doi.org/10.1145/2984511.2984526 * Berg and Vance (2017) Leif P. Berg and Judy M. Vance. 2017. Industry use of virtual reality in product design and manufacturing: a survey. _Virtual Reality_ 21, 1 (March 2017), 1–17. https://doi.org/10.1007/s10055-016-0293-9 * Bergström et al. (2019) Joanna Bergström, Aske Mottelson, and Jarrod Knibbe. 2019\. Resized Grasping in VR: Estimating Thresholds for Object Discrimination. In _Proceedings of the 32nd Annual ACM Symposium on User Interface Software and Technology_. ACM, New Orleans LA USA, 1175–1183. https://doi.org/10.1145/3332165.3347939 * Bloomfield et al. (2003) A. Bloomfield, Yu Deng, J. Wampler, P. Rondot, D. Harth, M. McManus, and N. Badler. 2003. A taxonomy and comparison of haptic actions for disassembly tasks. In _IEEE Virtual Reality, 2003\. Proceedings._ IEEE Comput. Soc, Los Angeles, CA, USA, 225–231. https://doi.org/10.1109/VR.2003.1191143 * Boldt et al. (2018) Mette Boldt, Boxuan Liu, Tram Nguyen, Alina Panova, Ramneek Singh, Alexander Steenbergen, Rainer Malaka, Jan Smeddinck, Michael Bonfert, Inga Lehne, Melina Cahnbley, Kim Korschinq, Loannis Bikas, Stefan Finke, Martin Hanci, and Valentin Kraft. 2018\. You Shall Not Pass: Non-Intrusive Feedback for Virtual Walls in VR Environments with Room-Scale Mapping. In _2018 IEEE Conference on Virtual Reality and 3D User Interfaces (VR)_. IEEE, Reutlingen, 143–150. https://doi.org/10.1109/VR.2018.8446177 * Bouzbib et al. (2020) Elodie Bouzbib, Gilles Bailly, Sinan Haliyo, and Pascal Frey. 2020. CoVR: A Large-Scale Force-Feedback Robotic Interface for Non-Deterministic Scenarios in VR. In _Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology_. ACM, Virtual Event USA, 209–222. https://doi.org/10.1145/3379337.3415891 * Bowman and Wingrave (2001) D.A. Bowman and C.A. Wingrave. 2001. Design and evaluation of menu systems for immersive virtual environments. In _Proceedings IEEE Virtual Reality 2001_. IEEE Comput. Soc, Yokohama, Japan, 149–156. https://doi.org/10.1109/VR.2001.913781 * Bryson (2005) Steve Bryson. 2005\. Direct Manipulation in Virtual Reality. In _Visualization Handbook_. Elsevier, 413–430. https://doi.org/10.1016/B978-012387582-2/50023-X * Cheng (2019) Lung-Pan Cheng. 2019\. VRoamer: Generating On-The-Fly VR Experiences While Walking inside Large, Unknown Real-World Building Environments. (2019), 8. * Cheng et al. (2018) Lung-Pan Cheng, Li Chang, Sebastian Marwecki, and Patrick Baudisch. 2018. iTurk: Turning Passive Haptics into Active Haptics by Making Users Reconfigure Props in Virtual Reality. In _Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems - CHI ’18_. ACM Press, Montreal QC, Canada, 1–10. https://doi.org/10.1145/3173574.3173663 * Cheng et al. (2014) Lung-Pan Cheng, Patrick Lühne, Pedro Lopes, Christoph Sterz, and Patrick Baudisch. 2014. Haptic Turk: a Motion Platform Based on People. (2014), 11. * Cheng et al. (2017) Lung-Pan Cheng, Eyal Ofek, Christian Holz, Hrvoje Benko, and Andrew D. Wilson. 2017. Sparse Haptic Proxy: Touch Feedback in Virtual Environments Using a General Passive Prop. In _Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems - CHI ’17_. ACM Press, Denver, Colorado, USA, 3718–3728. https://doi.org/10.1145/3025453.3025753 * Cheng et al. (2015) Lung-Pan Cheng, Thijs Roumen, Hannes Rantzsch, Sven Köhler, Patrick Schmidt, Robert Kovacs, Johannes Jasper, Jonas Kemper, and Patrick Baudisch. 2015. TurkDeck: Physical Virtual Reality Based on People. In _Proceedings of the 28th Annual ACM Symposium on User Interface Software & Technology - UIST ’15_. ACM Press, Daegu, Kyungpook, Republic of Korea, 417–426. https://doi.org/10.1145/2807442.2807463 * Choi et al. (2017) Inrak Choi, Heather Culbertson, Mark R. Miller, Alex Olwal, and Sean Follmer. 2017. Grabity: A Wearable Haptic Interface for Simulating Weight and Grasping in Virtual Reality. In _Proceedings of the 30th Annual ACM Symposium on User Interface Software and Technology - UIST ’17_. ACM Press, Qu&#233;bec City, QC, Canada, 119–130. https://doi.org/10.1145/3126594.3126599 * Choi et al. (2016) Inrak Choi, Elliot W. Hawkes, David L. Christensen, Christopher J. Ploch, and Sean Follmer. 2016. Wolverine: A wearable haptic interface for grasping in virtual reality. In _2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)_. IEEE, Daejeon, South Korea, 986–993. https://doi.org/10.1109/IROS.2016.7759169 * Choi et al. (2018) Inrak Choi, Eyal Ofek, Hrvoje Benko, Mike Sinclair, and Christian Holz. 2018. CLAW: A Multifunctional Handheld Haptic Controller for Grasping, Touching, and Triggering in Virtual Reality. In _Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems - CHI ’18_. ACM Press, Montreal QC, Canada, 1–13. https://doi.org/10.1145/3173574.3174228 * Coles et al. (2011) Timothy R. Coles, Dwight Meglan, and Nigel W. John. 2011\. The Role of Haptics in Medical Training Simulators: A Survey of the State of the Art. _IEEE Transactions on Haptics_ 4, 1 (Jan. 2011), 51–66. https://doi.org/10.1109/TOH.2010.19 * Danieau et al. (2012) Fabien Danieau, Julien Fleureau, Philippe Guillotel, Nicolas Mollet, Anatole Lécuyer, and Marc Christie. 2012. HapSeat: producing motion sensation with multiple force-feedback devices embedded in a seat. In _Proceedings of the 18th ACM symposium on Virtual reality software and technology - VRST ’12_. ACM Press, Toronto, Ontario, Canada, 69\. https://doi.org/10.1145/2407336.2407350 * Danieau et al. (2018) Fabien Danieau, Philippe Guillotel, Olivier Dumas, Thomas Lopez, Bertrand Leroy, and Nicolas Mollet. 2018\. HFX studio: haptic editor for full-body immersive experiences. In _Proceedings of the 24th ACM Symposium on Virtual Reality Software and Technology - VRST ’18_. ACM Press, Tokyo, Japan, 1–9. https://doi.org/10.1145/3281505.3281518 * De Araújo et al. (2013) Bruno R. De Araújo, Géry Casiez, Joaquim A. Jorge, and Martin Hachet. 2013. Mockup Builder: 3D modeling on and above the surface. _Computers & Graphics_ 37, 3 (May 2013), 165–178. https://doi.org/10.1016/j.cag.2012.12.005 * de Tinguy et al. (2020) Xavier de Tinguy, Thomas Howard, Claudio Pacchierotti, Maud Marchal, and Anatole Lécuyer. 2020\. WeATaViX: WEarable Actuated TAngibles for VIrtual reality eXperiences. (2020), 9. * de Tinguy et al. (2019) Xavier de Tinguy, Claudio Pacchierotti, Maud Marchal, and Anatole Lecuyer. 2019. Toward Universal Tangible Objects: Optimizing Haptic Pinching Sensations in 3D Interaction. In _2019 IEEE Conference on Virtual Reality and 3D User Interfaces (VR)_. IEEE, Osaka, Japan, 321–330. https://doi.org/10.1109/VR.2019.8798205 * Degraen et al. (2019) Donald Degraen, André Zenner, and Antonio Krüger. 2019\. Enhancing Texture Perception in Virtual Reality Using 3D-Printed Hair Structures. In _Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems - CHI ’19_. ACM Press, Glasgow, Scotland Uk, 1–12. https://doi.org/10.1145/3290605.3300479 * Dominjon et al. (2007) Lionel Dominjon, Jérôme Perret, and Anatole Lécuyer. 2007\. Novel devices and interaction techniques for human-scale haptics. _The Visual Computer_ 23, 4 (March 2007), 257–266. https://doi.org/10.1007/s00371-007-0100-4 * Ducatelle et al. (2011) Frederick Ducatelle, Gianni A. Di Caro, Carlo Pinciroli, and Luca M. Gambardella. 2011\. Self-organized cooperation between robotic swarms. _Swarm Intelligence_ 5, 2 (June 2011), 73–96. https://doi.org/10.1007/s11721-011-0053-0 * Escobar-Castillejos et al. (2016) David Escobar-Castillejos, Julieta Noguez, Luis Neri, Alejandra Magana, and Bedrich Benes. 2016\. A Review of Simulators with Haptic Devices for Medical Training. _Journal of Medical Systems_ 40, 4 (April 2016), 1–22. https://doi.org/10.1007/s10916-016-0459-8 * Fang et al. (2020) Cathy Fang, Yang Zhang, Matthew Dworman, and Chris Harrison. 2020\. Wireality: Enabling Complex Tangible Geometries in Virtual Reality with Worn Multi-String Haptics. (2020), 10. * Feick et al. (2020) Martin Feick, Scott Bateman, Anthony Tang, André Miede, and Nicolai Marquardt. 2020. TanGi: Tangible Proxies for Embodied Object Exploration and Manipulation in Virtual Reality. _arXiv:2001.03021 [cs]_ (Jan. 2020). http://arxiv.org/abs/2001.03021 arXiv: 2001.03021. * Fitzgerald and Ishii (2018) Daniel Fitzgerald and Hiroshi Ishii. 2018. Mediate: A Spatial Tangible Interface for Mixed Reality. In _Extended Abstracts of the 2018 CHI Conference on Human Factors in Computing Systems_. ACM, Montreal QC Canada, 1–6. https://doi.org/10.1145/3170427.3188472 * Follmer et al. (2013) Sean Follmer, Daniel Leithinger, Alex Olwal, Akimitsu Hogge, and Hiroshi Ishii. 2013. inFORM: dynamic physical affordances and constraints through shape and object actuation. In _Proceedings of the 26th annual ACM symposium on User interface software and technology - UIST ’13_. ACM Press, St. Andrews, Scotland, United Kingdom, 417–426. https://doi.org/10.1145/2501988.2502032 * Formaglio et al. (2005) A. Formaglio, A. Giannitrapani, M. Franzini, D. Prattichizzo, and F. Barbagli. 2005\. Performance of Mobile Haptic Interfaces. In _Proceedings of the 44th IEEE Conference on Decision and Control_. 8343–8348. https://doi.org/10.1109/CDC.2005.1583513 * Frissen et al. (2013) Ilja Frissen, Jennifer L. Campos, Manish Sreenivasa, and Marc O. Ernst. 2013. Enabling Unconstrained Omnidirectional Walking Through Virtual Environments: An Overview of the CyberWalk Project. In _Human Walking in Virtual Environments: Perception, Technology, and Applications_ , Frank Steinicke, Yon Visell, Jennifer Campos, and Anatole Lécuyer (Eds.). Springer, New York, NY, 113–144. https://doi.org/10.1007/978-1-4419-8432-6_6 * Funk et al. (2019) Markus Funk, Florian Müller, Marco Fendrich, Megan Shene, Moritz Kolvenbach, Niclas Dobbertin, Sebastian Günther, and Max Mühlhäuser. 2019. Assessing the Accuracy of Point & Teleport Locomotion with Orientation Indication for Virtual Reality using Curved Trajectories. In _Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems - CHI ’19_. ACM Press, Glasgow, Scotland Uk, 1–12. https://doi.org/10.1145/3290605.3300377 * Galais et al. (2019) Thomas Galais, Alexandra Delmas, and Rémy Alonso. 2019\. Natural interaction in virtual reality: impact on the cognitive load. In _Proceedings of the 31st Conference on l’Interaction Homme-Machine Adjunct - IHM ’19_. ACM Press, Grenoble, France, 1–9. https://doi.org/10.1145/3366551.3370342 * Galambos (2012) Péter Galambos. 2012\. Vibrotactile Feedback for Haptics and Telemanipulation: Survey, Concept and Experiment. _Acta Polytechnica Hungarica_ 9, 1 (2012), 25\. * Gomes et al. (2016) Antonio Gomes, Calvin Rubens, Sean Braley, and Roel Vertegaal. 2016. BitDrones: Towards Using 3D Nanocopter Displays as Interactive Self-Levitating Programmable Matter. In _Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems - CHI ’16_. ACM Press, Santa Clara, California, USA, 770–780. https://doi.org/10.1145/2858036.2858519 * Gonzalez et al. (2020) Eric J. Gonzalez, Parastoo Abtahi, and Sean Follmer. 2020\. REACH+: Extending the Reachability of Encountered-type Haptics Devices through Dynamic Redirection in VR. In _Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology_. ACM, Virtual Event USA, 236–248. https://doi.org/10.1145/3379337.3415870 * Gonzalez and Follmer (2019) Eric J. Gonzalez and Sean Follmer. 2019. Investigating the Detection of Bimanual Haptic Retargeting in Virtual Reality. In _25th ACM Symposium on Virtual Reality Software and Technology on - VRST ’19_. ACM Press, Parramatta, NSW, Australia, 1–5. https://doi.org/10.1145/3359996.3364248 * Gu et al. (2016) Xiaochi Gu, Yifei Zhang, Weize Sun, Yuanzhe Bian, Dao Zhou, and Per Ola Kristensson. 2016\. Dexmo: An Inexpensive and Lightweight Mechanical Exoskeleton for Motion Capture and Force Feedback in VR. In _Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems - CHI ’16_. ACM Press, Santa Clara, California, USA, 1991–1995. https://doi.org/10.1145/2858036.2858487 * Gugenheimer et al. (2016) Jan Gugenheimer, Dennis Wolf, Eythor R. Eiriksson, Pattie Maes, and Enrico Rukzio. 2016. GyroVR: Simulating Inertia in Virtual Reality using Head Worn Flywheels. In _Proceedings of the 29th Annual Symposium on User Interface Software and Technology_. ACM, Tokyo Japan, 227–232. https://doi.org/10.1145/2984511.2984535 * Gutierrez et al. (2010) T. Gutierrez, J. Rodriguez, Y. Velaz, S. Casado, A. Suescun, and E. J. Sanchez. 2010\. IMA-VR: A multimodal virtual training system for skills transfer in Industrial Maintenance and Assembly tasks. _19th International Symposium in Robot and Human Interactive Communication_ (2010). https://www.academia.edu/15623406/IMA_VR_A_multimodal_virtual_training_system_for_skills_transfer_in_Industrial_Maintenance_and_Assembly_tasks * Günther et al. (2020) Sebastian Günther, Dominik Schön, Florian Müller, Max Mühlhäuser, and Martin Schmitz. 2020\. PneumoVolley: Pressure-based Haptic Feedback on the Head through Pneumatic Actuation. (2020), 10. * Han et al. (2018) Dustin T. Han, Mohamed Suhail, and Eric D. Ragan. 2018\. Evaluating Remapped Physical Reach for Hand Interactions with Passive Haptics in Virtual Reality. _IEEE Transactions on Visualization and Computer Graphics_ 24, 4 (April 2018), 1467–1476. https://doi.org/10.1109/TVCG.2018.2794659 * Haption (2019) Haption. 2019. Virtuose™ 6D - HAPTION SA. https://www.haption.com/en/products-en/virtuose-6d-en.html * Hayward and Maclean (2007) Vincent Hayward and Karon Maclean. 2007. Do it yourself haptics: part I. _IEEE Robotics & Automation Magazine_ 14, 4 (Dec. 2007), 88–104. https://doi.org/10.1109/M-RA.2007.907921 * He et al. (2017b) Zhenyi He, Fengyuan Zhu, Aaron Gaudette, and Ken Perlin. 2017b. Robotic Haptic Proxies for Collaborative Virtual Reality. _arXiv:1701.08879 [cs]_ (Jan. 2017). http://arxiv.org/abs/1701.08879 arXiv: 1701.08879. * He et al. (2017a) Zhenyi He, Fengyuan Zhu, and Ken Perlin. 2017a. PhyShare: Sharing Physical Interaction in Virtual Reality. _arXiv:1708.04139 [cs]_ (Aug. 2017). http://arxiv.org/abs/1708.04139 arXiv: 1708.04139. * Held and Durlach (1992) Richard M. Held and Nathaniel I. Durlach. 1992. Telepresence. _Presence: Teleoperators and Virtual Environments_ 1, 1 (Jan. 1992), 109–112. https://doi.org/10.1162/pres.1992.1.1.109 * Heo et al. (2018) Seongkook Heo, Christina Chung, Geehyuk Lee, and Daniel Wigdor. 2018. Thor’s Hammer: An Ungrounded Force Feedback Device Utilizing Propeller-Induced Propulsive Force. In _Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems - CHI ’18_. ACM Press, Montreal QC, Canada, 1–11. https://doi.org/10.1145/3173574.3174099 * Heo et al. (2019) Seongkook Heo, Jaeyeon Lee, and Daniel Wigdor. 2019\. PseudoBend: Producing Haptic Illusions of Stretching, Bending, and Twisting Using Grain Vibrations. In _Proceedings of the 32nd Annual ACM Symposium on User Interface Software and Technology - UIST ’19_. ACM Press, New Orleans, LA, USA, 803–813. https://doi.org/10.1145/3332165.3347941 * Hettiarachchi and Wigdor (2016) Anuruddha Hettiarachchi and Daniel Wigdor. 2016. Annexing Reality: Enabling Opportunistic Use of Everyday Objects as Tangible Proxies in Augmented Reality. In _Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems - CHI ’16_. ACM Press, Santa Clara, California, USA, 1957–1967. https://doi.org/10.1145/2858036.2858134 * Hoppe et al. (2018) Matthias Hoppe, Pascal Knierim, Thomas Kosch, Markus Funk, Lauren Futami, Stefan Schneegass, Niels Henze, Albrecht Schmidt, and Tonja Machulla. 2018. VRHapticDrones: Providing Haptics in Virtual Reality through Quadcopters. In _Proceedings of the 17th International Conference on Mobile and Ubiquitous Multimedia - MUM 2018_. ACM Press, Cairo, Egypt, 7–18. https://doi.org/10.1145/3282894.3282898 * Hoppe et al. (2020) Matthias Hoppe, Daniel Neumann, Stephan Streuber, Albrecht Schmidt, and Tonja-Katrin Machulla. 2020\. _A Human Touch: Social Touch Increases the Perceived Human-likeness of Agents in Virtual Reality_. https://doi.org/10.1145/3313831.3376719 * Huang et al. (2020) Hsin-Yu Huang, Chih-Wei Ning, Po-Yao Wang, Jen-Hao Cheng, and Lung-Pan Cheng. 2020. Haptic-go-round: A Surrounding Platform for Encounter-type Haptics in Virtual Reality Experiences. In _Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems_. ACM, Honolulu HI USA, 1–10. https://doi.org/10.1145/3313831.3376476 * Insko (2001) Brent Edward Insko. 2001\. Passive Haptics Significantly Enhances Virtual Environments. (2001), 111. * Iwata (2005) Hiroo Iwata. 2005\. CirculaFloor. https://ieeexplore.ieee.org/abstract/document/1381227 * Iwata (2013) Hiroo Iwata. 2013\. Locomotion Interfaces. In _Human Walking in Virtual Environments: Perception, Technology, and Applications_ , Frank Steinicke, Yon Visell, Jennifer Campos, and Anatole Lécuyer (Eds.). Springer, New York, NY, 199–219. https://doi.org/10.1007/978-1-4419-8432-6_9 * Iwata et al. (2001) Hiroo Iwata, Hiroaki Yano, Fumitaka Nakaizumi, and Ryo Kawamura. 2001. Project FEELEX: adding haptic surface to graphics. In _Proceedings of the 28th annual conference on Computer graphics and interactive techniques - SIGGRAPH ’01_. ACM Press, Not Known, 469–476. https://doi.org/10.1145/383259.383314 * Jones (2000) Lynette Jones. 2000\. Kinesthetic Sensing. _Human and Machine Haptics_ (2000). http://bdml.stanford.edu/twiki/pub/Haptics/PapersInProgress/jones00.pdf * Kim and Schneider (2020) Erin Kim and Oliver Schneider. 2020. Defining Haptic Experience: Foundations for Understanding, Communicating, and Evaluating HX. In _Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems_. ACM, Honolulu HI USA, 1–13. https://doi.org/10.1145/3313831.3376280 * Kim et al. (2020) Lawrence H. Kim, Daniel S. Drew, Veronika Domova, and Sean Follmer. 2020. User-defined Swarm Robot Control. In _Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems_. ACM, Honolulu HI USA, 1–13. https://doi.org/10.1145/3313831.3376814 * Kim et al. (2018) Yaesol Kim, Hyun Jung Kim, and Young J. Kim. 2018. Encountered-type haptic display for large VR environment using per-plane reachability maps: Encountered-type Haptic Display for Large VR Environment. _Computer Animation and Virtual Worlds_ 29, 3-4 (May 2018), e1814. https://doi.org/10.1002/cav.1814 * Knierim et al. (2017) Pascal Knierim, Thomas Kosch, Valentin Schwind, Markus Funk, Francisco Kiss, Stefan Schneegass, and Niels Henze. 2017. Tactile Drones - Providing Immersive Tactile Feedback in Virtual Reality through Quadcopters. In _Proceedings of the 2017 CHI Conference Extended Abstracts on Human Factors in Computing Systems - CHI EA ’17_. ACM Press, Denver, Colorado, USA, 433–436. https://doi.org/10.1145/3027063.3050426 * Kohli (2010) Luv Kohli. 2010\. Redirected touching: Warping space to remap passive haptics. In _2010 IEEE Symposium on 3D User Interfaces (3DUI)_. IEEE, Waltham, MA, USA, 129–130. https://doi.org/10.1109/3DUI.2010.5444703 * Kohli et al. (2012) L. Kohli, M. C. Whitton, and F. P. Brooks. 2012. Redirected touching: The effect of warping space on task performance. In _2012 IEEE Symposium on 3D User Interfaces (3DUI)_. IEEE, Costa Mesa, CA, 105–112. https://doi.org/10.1109/3DUI.2012.6184193 * Kohli et al. (2013) Luv Kohli, Mary C. Whitton, and Frederick P. Brooks. 2013\. Redirected Touching: Training and adaptation in warped virtual spaces. In _2013 IEEE Symposium on 3D User Interfaces (3DUI)_. IEEE, Orlando, FL, 79–86. https://doi.org/10.1109/3DUI.2013.6550201 * Kovacs et al. (2020) Robert Kovacs, Eyal Ofek, Mar Gonzalez Franco, Alexa Fay Siu, Sebastian Marwecki, Christian Holz, and Mike Sinclair. 2020. Haptic PIVOT: On-Demand Handhelds in VR. In _Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology_ _(UIST ’20)_. Association for Computing Machinery, New York, NY, USA, 1046–1059. https://doi.org/10.1145/3379337.3415854 * Le Goc et al. (2016) Mathieu Le Goc, Lawrence H. Kim, Ali Parsaei, Jean-Daniel Fekete, Pierre Dragicevic, and Sean Follmer. 2016. Zooids: Building Blocks for Swarm User Interfaces. In _Proceedings of the 29th Annual Symposium on User Interface Software and Technology - UIST ’16_. ACM Press, Tokyo, Japan, 97–109. https://doi.org/10.1145/2984511.2984547 * Lederman and Klatzky (2009) S. J. Lederman and R. L. Klatzky. 2009. Haptic perception: A tutorial. _Attention, Perception & Psychophysics_ 71, 7 (Oct. 2009), 1439–1459. https://doi.org/10.3758/APP.71.7.1439 * Lee et al. (2007) Chaehyun Lee, Min Sik Hong, In Lee, Oh Kyu Choi, Kyung-Lyong Han, Yoo Yeon Kim, Seungmoon Choi, and Jin S Lee. 2007\. Mobile Haptic Interface for Large Immersive Virtual Environments: PoMHI v0.5. (2007), 2. * Lee et al. (2009) In Lee, Inwook Hwang, Kyung-Lyoung Han, Oh Kyu Choi, Seungmoon Choi, and Jin S. Lee. 2009\. System improvements in Mobile Haptic Interface. In _World Haptics 2009 - Third Joint EuroHaptics conference and Symposium on Haptic Interfaces for Virtual Environment and Teleoperator Systems_. IEEE, Salt Lake City, UT, USA, 109–114. https://doi.org/10.1109/WHC.2009.4810834 * Lee et al. (2019) Jaeyeon Lee, Mike Sinclair, Mar Gonzalez-Franco, Eyal Ofek, and Christian Holz. 2019\. TORC: A Virtual Reality Controller for In-Hand High-Dexterity Finger Interaction. In _Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems - CHI ’19_. ACM Press, Glasgow, Scotland Uk, 1–13. https://doi.org/10.1145/3290605.3300301 * Leithinger et al. (2013) Daniel Leithinger, Sean Follmer, Alex Olwal, Samuel Luescher, Akimitsu Hogge, Jinha Lee, and Hiroshi Ishii. 2013. Sublimate: state-changing virtual and physical rendering to augment interaction with shape displays. In _Proceedings of the SIGCHI Conference on Human Factors in Computing Systems - CHI ’13_. ACM Press, Paris, France, 1441. https://doi.org/10.1145/2470654.2466191 * Lepecq et al. (2008) Jean-Claude Lepecq, Lionel Bringoux, Jean-Marie Pergandi, Thelma Coyle, and Daniel Mestre. 2008\. Afforded Actions as a Behavioral Assessment of Physical Presence. (2008), 8. * Lo et al. (2018) Jo-Yu Lo, Da-Yuan Huang, Chen-Kuo Sun, Chu-En Hou, and Bing-Yu Chen. 2018. RollingStone: Using Single Slip Taxel for Enhancing Active Finger Exploration with a Virtual Reality Controller. In _The 31st Annual ACM Symposium on User Interface Software and Technology - UIST ’18_. ACM Press, Berlin, Germany, 839–851. https://doi.org/10.1145/3242587.3242627 * Lopes et al. (2015) Pedro Lopes, Patrik Jonell, and Patrick Baudisch. 2015\. Affordance++: Allowing Objects to Communicate Dynamic Use. In _Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems - CHI ’15_. ACM Press, Seoul, Republic of Korea, 2515–2524. https://doi.org/10.1145/2702123.2702128 * Lopes et al. (2017) Pedro Lopes, Sijing You, Lung-Pan Cheng, Sebastian Marwecki, and Patrick Baudisch. 2017. Providing Haptics to Walls & Heavy Objects in Virtual Reality by Means of Electrical Muscle Stimulation. In _Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems - CHI ’17_. ACM Press, Denver, Colorado, USA, 1471–1482. https://doi.org/10.1145/3025453.3025600 * Lécuyer (2009) Anatole Lécuyer. 2009\. Simulating Haptic Feedback Using Vision: A Survey of Research and Applications of Pseudo-Haptic Feedback. _Presence: Teleoperators and Virtual Environments_ 18, 1 (Feb. 2009), 39–53. https://doi.org/10.1162/pres.18.1.39 * Magnenat-Thalmann et al. (2005) N. Magnenat-Thalmann, HyungSeok Kim, A. Egges, and S. Garchery. 2005. Believability and Interaction in Virtual Worlds. In _11th International Multimedia Modelling Conference_. IEEE, Honolulu, HI, USA, 2–9. https://doi.org/10.1109/MMMC.2005.24 * Makin et al. (2019) Lawrence Makin, Gareth Barnaby, and Anne Roudaut. 2019\. Tactile and kinesthetic feedbacks improve distance perception in virtual reality. In _Proceedings of the 31st Conference on l’Interaction Homme-Machine - IHM ’19_. ACM Press, Grenoble, France, 1–9. https://doi.org/10.1145/3366550.3372248 * Marquardt et al. (2009) Nicolai Marquardt, Miguel A. Nacenta, James E. Young, Sheelagh Carpendale, Saul Greenberg, and Ehud Sharlin. 2009. The Haptic Tabletop Puck: tactile feedback for interactive tabletops. In _Proceedings of the ACM International Conference on Interactive Tabletops and Surfaces - ITS ’09_. ACM Press, Banff, Alberta, Canada, 85. https://doi.org/10.1145/1731903.1731922 * Massie and Salisbury (1994) Thomas H Massie and J K Salisbury. 1994. The PHANTOM Haptic Interface: A Device for Probing Virtual Objects. (1994), 5. * McNeely (1993) W. A. McNeely. 1993\. Robotic graphics: a new approach to force feedback for virtual reality. In _Proceedings of IEEE Virtual Reality Annual International Symposium_. 336–341. https://doi.org/10.1109/VRAIS.1993.380761 * Merino et al. (2020) Leonel Merino, Magdalena Schwarzl, Matthias Kraus, Michael Sedlmair, Dieter Schmalstieg, and Daniel Weiskopf. 2020. Evaluating Mixed and Augmented Reality: A Systematic Literature Review (2009-2019). _arXiv:2010.05988 [cs]_ (Oct. 2020). http://arxiv.org/abs/2010.05988 arXiv: 2010.05988. * Moline (1997) Judi Moline. 1997\. _Virtual reality for health care: a survey_. Technical Report. * Nakagaki et al. (2016a) Ken Nakagaki, Artem Dementyev, Sean Follmer, Joseph A. Paradiso, and Hiroshi Ishii. 2016a. ChainFORM: A Linear Integrated Modular Hardware System for Shape Changing Interfaces. In _Proceedings of the 29th Annual Symposium on User Interface Software and Technology - UIST ’16_. ACM Press, Tokyo, Japan, 87–96. https://doi.org/10.1145/2984511.2984587 * Nakagaki et al. (2016b) Ken Nakagaki, Luke Vink, Jared Counts, Daniel Windham, Daniel Leithinger, Sean Follmer, and Hiroshi Ishii. 2016b. Materiable: Rendering Dynamic Material Properties in Response to Direct Physical Touch with Shape Changing Interfaces. In _Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems - CHI ’16_. ACM Press, Santa Clara, California, USA, 2764–2772. https://doi.org/10.1145/2858036.2858104 * Nitzsche et al. (2003) Norbert Nitzsche, Uwe D. Hanebeck, and G. Schmidt. 2003\. Design issues of mobile haptic interfaces. _Journal of Robotic Systems_ 20, 9 (Sept. 2003), 549–556. https://doi.org/10.1002/rob.10105 * Norman (2013) Donald A. Norman. 2013\. _The design of everyday things_ (revised and expanded edition ed.). Basic Books, New York, New York. * Optitrack (2019) Optitrack. 2019\. Motion Capture Systems. http://optitrack.com/index.html * Ortega and Coquillart (2005) M. Ortega and S. Coquillart. 2005. Prop-based haptic interaction with co-location and immersion: an automotive application. In _IREE International Worksho on Haptic Audio Visual Environments and their Applications, 2005._ IEEE, Ottawa, Canada, 23–28. https://doi.org/10.1109/HAVE.2005.1545646 * Pair et al. (2003) J. Pair, U. Neumann, D. Piepol, and B. Swartout. 2003\. FlatWorld: combining Hollywood set-design techniques with VR. _IEEE Computer Graphics and Applications_ 23, 1 (Jan. 2003), 12–15. https://doi.org/10.1109/MCG.2003.1159607 * Pavlik et al. (2013) Ryan A. Pavlik, Judy M. Vance, and Greg R. Luecke. 2013\. Interacting With a Large Virtual Environment by Combining a Ground-Based Haptic Device and a Mobile Robot Base. In _Volume 2B: 33rd Computers and Information in Engineering Conference_. ASME, Portland, Oregon, USA, V02BT02A029. https://doi.org/10.1115/DETC2013-13441 * Poyade et al. (2012) M Poyade, L Molina-Tanco, A Reyes-Lecuona, A Langley, E Frutos, and S Flores. 2012\. Validation of a haptic virtual reality simulation in the context of industrial maintenance. (2012), 4. * Praveena et al. (2020) Pragathi Praveena, Daniel Rakita, Bilge Mutlu, and Michael Gleicher. 2020. Supporting Perception of Weight through Motion-induced Sensory Conflicts in Robot Teleoperation. In _Proceedings of the 2020 ACM/IEEE International Conference on Human-Robot Interaction_. ACM, Cambridge United Kingdom, 509–517. https://doi.org/10.1145/3319502.3374841 * Provancher et al. (2005) William R. Provancher, Mark R. Cutkosky, Katherine J. Kuchenbecker, and Günter Niemeyer. 2005\. Contact Location Display for Haptic Perception of Curvature and Object Motion. _The International Journal of Robotics Research_ 24, 9 (Sept. 2005), 691–702. https://doi.org/10.1177/0278364905057121 * Pusch and Lécuyer (2011) Andreas Pusch and Anatole Lécuyer. 2011. Pseudo-haptics: from the theoretical foundations to practical system design guidelines. In _Proceedings of the 13th international conference on multimodal interfaces - ICMI ’11_. ACM Press, Alicante, Spain, 57\. https://doi.org/10.1145/2070481.2070494 * Putze et al. (2020) Susanne Putze, Dmitry Alexandrovsky, Felix Putze, Sebastian Höffner, Jan David Smeddinck, and Rainer Malaka. 2020. Breaking The Experience: Effects of Questionnaires in VR User Studies. In _Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems_. ACM, Honolulu HI USA, 1–15. https://doi.org/10.1145/3313831.3376144 * Rakkolainen et al. (2020) Ismo Rakkolainen, Euan Freeman, Antti Sand, Roope Raisamo, and Stephen Brewster. 2020. A Survey of Mid-Air Ultrasound Haptics and Its Applications. _IEEE Transactions on Haptics_ (2020), 1–1. https://doi.org/10.1109/TOH.2020.3018754 * Rangarajan et al. (2020) Karan Rangarajan, Heather Davis, and Philip H. Pucher. 2020\. Systematic Review of Virtual Haptics in Surgical Simulation: A Valid Educational Tool? _Journal of Surgical Education_ 77, 2 (March 2020), 337–347. https://doi.org/10.1016/j.jsurg.2019.09.006 * Razzaque et al. (2001) Sharif Razzaque, Zachariah Kohn, and Mary C. Whitton. 2001\. _EUROGRAPHICS 2001 / Jonathan C. Roberts Short Presentation © The Eurographics Association 2001. Redirected Walking_. * Rietzler et al. (2018) Michael Rietzler, Florian Geiselhart, Jan Gugenheimer, and Enrico Rukzio. 2018. Breaking the Tracking: Enabling Weight Perception using Perceivable Tracking Offsets. In _Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems - CHI ’18_. ACM Press, Montreal QC, Canada, 1–12. https://doi.org/10.1145/3173574.3173702 * Rietzler et al. (2019) Michael Rietzler, Gabriel Haas, Thomas Dreja, Florian Geiselhart, and Enrico Rukzio. 2019. Virtual Muscle Force: Communicating Kinesthetic Forces Through Pseudo-Haptic Feedback and Muscle Input. In _Proceedings of the 32nd Annual ACM Symposium on User Interface Software and Technology - UIST ’19_. ACM Press, New Orleans, LA, USA, 913–922. https://doi.org/10.1145/3332165.3347871 * Rubens et al. (2015) Calvin Rubens, Sean Braley, Antonio Gomes, Daniel Goc, Xujing Zhang, Juan Pablo Carrascal, and Roel Vertegaal. 2015. BitDrones: Towards Levitating Programmable Matter Using Interactive 3D Quadcopter Displays. In _Proceedings of the 28th Annual ACM Symposium on User Interface Software & Technology - UIST ’15 Adjunct_. ACM Press, Daegu, Kyungpook, Republic of Korea, 57–58. https://doi.org/10.1145/2815585.2817810 * Sagayam and Hemanth (2017) K. Martin Sagayam and D. Jude Hemanth. 2017. Hand posture and gesture recognition techniques for virtual reality applications: a survey. _Virtual Reality_ 21, 2 (June 2017), 91–107. https://doi.org/10.1007/s10055-016-0301-0 * Sagheb et al. (2019) Shahabedin Sagheb, Frank Wencheng Liu, Alireza Bahremand, Assegid Kidane, and Robert LiKamWa. 2019\. SWISH: A Shifting-Weight Interface of Simulated Hydrodynamics for Haptic Perception of Virtual Fluid Vessels. In _Proceedings of the 32nd Annual ACM Symposium on User Interface Software and Technology - UIST ’19_. ACM Press, New Orleans, LA, USA, 751–761. https://doi.org/10.1145/3332165.3347870 * Samad et al. (2019) Majed Samad, Elia Gatti, Anne Hermes, Hrvoje Benko, and Cesare Parise. 2019. Pseudo-Haptic Weight: Changing the Perceived Weight of Virtual Objects By Manipulating Control-Display Ratio. In _Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems - CHI ’19_. ACM Press, Glasgow, Scotland Uk, 1–13. https://doi.org/10.1145/3290605.3300550 * Satler et al. (2011) Massimo Satler, Carlo A. Avizzano, and Emanuele Ruffaldi. 2011\. Control of a desktop mobile haptic interface. In _2011 IEEE World Haptics Conference_. IEEE, Istanbul, 415–420. https://doi.org/10.1109/WHC.2011.5945522 * Sato (2002) M. Sato. 2002. SPIDAR and virtual reality. In _Proceedings of the 5th Biannual World Automation Congress_ , Vol. 13. 17–23. https://doi.org/10.1109/WAC.2002.1049515 * Savino (2020) Gian-Luca Savino. 2020\. Virtual Smartphone: High Fidelity Interaction with Proxy Objects in Virtual Reality. _arXiv:2010.00942 [cs]_ (Oct. 2020). http://arxiv.org/abs/2010.00942 arXiv: 2010.00942. * Schmidt et al. (2015) Dominik Schmidt, Rob Kovacs, Vikram Mehta, Udayan Umapathi, Sven Köhler, Lung-Pan Cheng, and Patrick Baudisch. 2015. Level-Ups: Motorized Stilts that Simulate Stair Steps in Virtual Reality. In _Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems - CHI ’15_. ACM Press, Seoul, Republic of Korea, 2157–2160. https://doi.org/10.1145/2702123.2702253 * Schuemie et al. (2001) Martijn J. Schuemie, Peter van der Straaten, Merel Krijn, and Charles A.P.G. van der Mast. 2001\. Research on Presence in Virtual Reality: A Survey. _CyberPsychology & Behavior_ 4, 2 (April 2001), 183–201. https://doi.org/10.1089/109493101300117884 * Schwind et al. (2019) Valentin Schwind, Pascal Knierim, Nico Haas, and Niels Henze. 2019\. Using Presence Questionnaires in Virtual Reality. In _Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems - CHI ’19_. ACM Press, Glasgow, Scotland Uk, 1–12. https://doi.org/10.1145/3290605.3300590 * Seifi et al. (2019) Hasti Seifi, Farimah Fazlollahi, Michael Oppermann, John Andrew Sastrillo, Jessica Ip, Ashutosh Agrawal, Gunhyuk Park, Katherine J. Kuchenbecker, and Karon E. MacLean. 2019. Haptipedia: Accelerating Haptic Device Discovery to Support Interaction & Engineering Design. In _Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems - CHI ’19_. ACM Press, Glasgow, Scotland Uk, 1–12. https://doi.org/10.1145/3290605.3300788 * Shaw et al. (2019) Emily Shaw, Tessa Roper, Tommy Nilsson, Glyn Lawson, Sue V. G. Cobb, and Daniel Miller. 2019\. The Heat is On: Exploring User Behaviour in a Multisensory Virtual Environment for Fire Evacuation. _Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems - CHI ’19_ (2019), 1–13. https://doi.org/10.1145/3290605.3300856 arXiv: 1902.04573. * Shigeyama et al. (2019) Jotaro Shigeyama, Takeru Hashimoto, Shigeo Yoshida, Takuji Narumi, Tomohiro Tanikawa, and Michitaka Hirose. 2019. Transcalibur: A Weight Shifting Virtual Reality Controller for 2D Shape Rendering based on Computational Perception Model. In _Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems - CHI ’19_. ACM Press, Glasgow, Scotland Uk, 1–11. https://doi.org/10.1145/3290605.3300241 * Simeone et al. (2015) Adalberto L. Simeone, Eduardo Velloso, and Hans Gellersen. 2015\. Substitutional Reality: Using the Physical Environment to Design Virtual Reality Experiences. In _Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems - CHI ’15_. ACM Press, Seoul, Republic of Korea, 3307–3316. https://doi.org/10.1145/2702123.2702389 * Sinclair et al. (2019) Mike Sinclair, Eyal Ofek, Mar Gonzalez-Franco, and Christian Holz. 2019. CapstanCrunch: A Haptic VR Controller with User-supplied Force Feedback. In _Proceedings of the 32nd Annual ACM Symposium on User Interface Software and Technology - UIST ’19_. ACM Press, New Orleans, LA, USA, 815–829. https://doi.org/10.1145/3332165.3347891 * Sinclair et al. (2014) Mike Sinclair, Michel Pahud, and Hrvoje Benko. 2014\. TouchMover 2.0 - 3D touchscreen with force feedback and haptic texture. In _2014 IEEE Haptics Symposium (HAPTICS)_. IEEE, Houston, TX, USA, 1–6. https://doi.org/10.1109/HAPTICS.2014.6775425 * Siu et al. (2018) Alexa F. Siu, Eric J. Gonzalez, Shenli Yuan, Jason B. Ginsberg, and Sean Follmer. 2018\. shapeShift: 2D Spatial Manipulation and Self-Actuation of Tabletop Shape Displays for Tangible and Haptic Interaction. In _Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems - CHI ’18_. ACM Press, Montreal QC, Canada, 1–13. https://doi.org/10.1145/3173574.3173865 * Slater (1999) Mel Slater. 1999\. Measuring Presence: A Response to the Witmer and Singer Presence Questionnaire. _Presence: Teleoperators and Virtual Environments_ 8, 5 (Oct. 1999), 560–565. https://doi.org/10.1162/105474699566477 Publisher: MIT Press. * Slater et al. (1994) Mel Slater, Martin Usoh, and Anthony Steed. 1994. Depth of Presence in Virtual Environments. _Presence: Teleoperators and Virtual Environments_ 3, 2 (Jan. 1994), 130–144. https://doi.org/10.1162/pres.1994.3.2.130 * Steed et al. (2020) Anthony Steed, Sebastian Friston, Vijay Pawar, and David Swapp. 2020. Docking Haptics: Extending the Reach of Haptics by Dynamic Combinations of Grounded and Worn Devices. _arXiv:2002.06093 [cs]_ (Feb. 2020). http://arxiv.org/abs/2002.06093 arXiv: 2002.06093. * Steinicke et al. (2013) Frank Steinicke, Visell Yon, Jennifer Campos, and Anatole Lecuyer (Eds.). 2013. _Human walking in virtual environments: perception, technology, and applications_. Springer, New York,NY. OCLC: 856865949. * Strandholt et al. (2020) Patrick L. Strandholt, Oana A. Dogaru, Niels C. Nilsson, Rolf Nordahl, and Stefania Serafin. 2020\. Knock on Wood: Combining Redirected Touching and Physical Props for Tool-Based Interaction in Virtual Reality. In _Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems_. ACM, Honolulu HI USA, 1–13. https://doi.org/10.1145/3313831.3376303 * Strasnick et al. (2018) Evan Strasnick, Christian Holz, Eyal Ofek, Mike Sinclair, and Hrvoje Benko. 2018. Haptic Links: Bimanual Haptics for Virtual Reality Using Variable Stiffness Actuation. In _Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems - CHI ’18_. ACM Press, Montreal QC, Canada, 1–12. https://doi.org/10.1145/3173574.3174218 * Strohmeier et al. (2020) Paul Strohmeier, Seref Güngör, Luis Herres, Dennis Gudea, Bruno Fruchard, and Jürgen Steimle. 2020\. bARefoot: Generating Virtual Materials using Motion Coupled Vibration in Shoes. In _Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology_. ACM, Virtual Event USA, 579–593. https://doi.org/10.1145/3379337.3415828 * Sun et al. (2019) Yuqian Sun, Shigeo Yoshida, Takuji Narumi, and Michitaka Hirose. 2019. PaCaPa: A Handheld VR Device for Rendering Size, Shape, and Stiffness of Virtual Objects in Tool-based Interactions. In _Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems - CHI ’19_. ACM Press, Glasgow, Scotland Uk, 1–12. https://doi.org/10.1145/3290605.3300682 * Sutherland (1965) Ivan Sutherland. 1965\. The Ultimate Display. (1965), 2. * Suzuki et al. (2020) Ryo Suzuki, Hooman Hedayati, Clement Zheng, James Bohn, Daniel Szafir, Ellen Yi-Luen Do, Mark D Gross, and Daniel Leithinger. 2020\. RoomShift: Room-scale Dynamic Haptics for VR with Furniture-moving Swarm Robots. (2020), 11. * Suzuki et al. (2018) Ryo Suzuki, Junichi Yamaoka, Daniel Leithinger, Tom Yeh, Mark D. Gross, Yoshihiro Kawahara, and Yasuaki Kakehi. 2018. Dynablock: Dynamic 3D Printing for Instant and Reconstructable Shape Formation. In _The 31st Annual ACM Symposium on User Interface Software and Technology - UIST ’18_. ACM Press, Berlin, Germany, 99–111. https://doi.org/10.1145/3242587.3242659 * Suzuki et al. (2019) Ryo Suzuki, Clement Zheng, Yasuaki Kakehi, Tom Yeh, Ellen Yi-Luen Do, Mark D Gross, and Daniel Leithinger. 2019. ShapeBots: Shape-changing Swarm Robots. (2019), 13. * Takizawa et al. (2017) N. Takizawa, H. Yano, H. Iwata, Y. Oshiro, and N. Ohkohchi. 2017. Encountered-Type Haptic Interface for Representation of Shape and Rigidity of 3D Virtual Objects. _IEEE Transactions on Haptics_ 10, 4 (Oct. 2017), 500–510. https://doi.org/10.1109/TOH.2017.2740934 * Talvas et al. (2014) Anthony Talvas, Maud Marchal, and Anatole Lecuyer. 2014\. A Survey on Bimanual Haptic Interaction. _IEEE Transactions on Haptics_ 7, 3 (July 2014), 285–300. https://doi.org/10.1109/TOH.2014.2314456 * Teng et al. (2018) Shan-Yuan Teng, Tzu-Sheng Kuo, Chi Wang, Chi-huan Chiang, Da-Yuan Huang, Liwei Chan, and Bing-Yu Chen. 2018. PuPoP: Pop-up Prop on Palm for Virtual Reality. In _The 31st Annual ACM Symposium on User Interface Software and Technology - UIST ’18_. ACM Press, Berlin, Germany, 5–17. https://doi.org/10.1145/3242587.3242628 * Teng et al. (2019) Shan-Yuan Teng, Cheng-Lung Lin, Chi-huan Chiang, Tzu-Sheng Kuo, Liwei Chan, Da-Yuan Huang, and Bing-Yu Chen. 2019. TilePoP: Tile-type Pop-up Prop for Virtual Reality. (2019), 11. * Teyssier et al. (2020) Marc Teyssier, Gilles Bailly, Catherine Pelachaud, and Eric Lecolinet. 2020. Conveying Emotions Through Device-Initiated Touch. _IEEE Transactions on Affective Computing_ (2020), 1–1. https://doi.org/10.1109/TAFFC.2020.3008693 * Tsagarakis et al. (2005) N. G. Tsagarakis, T. Horne, and D. G. Caldwell. 2005\. SLIP AESTHEASIS: a portable 2D slip/skin stretch display for the fingertip. In _First Joint Eurohaptics Conference and Symposium on Haptic Interfaces for Virtual Environment and Teleoperator Systems. World Haptics Conference_. 214–219. https://doi.org/10.1109/WHC.2005.117 * Tsai and Chen (2019) Hsin-Ruey Tsai and Bing-Yu Chen. 2019. ElastImpact: 2.5D Multilevel Instant Impact Using Elasticity on Head-Mounted Displays. In _Proceedings of the 32nd Annual ACM Symposium on User Interface Software and Technology_. ACM, New Orleans LA USA, 429–437. https://doi.org/10.1145/3332165.3347931 * Tsai and Rekimoto (2018) Hsin-Ruey Tsai and Jun Rekimoto. 2018. ElasticVR: Providing Multi-level Active and Passive Force Feedback in Virtual Reality Using Elasticity. In _Extended Abstracts of the 2018 CHI Conference on Human Factors in Computing Systems - CHI ’18_. ACM Press, Montreal QC, Canada, 1–4. https://doi.org/10.1145/3170427.3186540 * Tsykunov et al. (2019) Evgeny Tsykunov, Roman Ibrahimov, Derek Vasquez, and Dzmitry Tsetserukou. 2019. SlingDrone: Mixed Reality System for Pointing and Interaction Using a Single Drone. In _25th ACM Symposium on Virtual Reality Software and Technology on - VRST ’19_. ACM Press, Parramatta, NSW, Australia, 1–5. https://doi.org/10.1145/3359996.3364271 * Tsykunov and Tsetserukou (2019) Evgeny Tsykunov and Dzmitry Tsetserukou. 2019. WiredSwarm: High Resolution Haptic Feedback Provided by a Swarm of Drones to the User’s Fingers for VR interaction. In _25th ACM Symposium on Virtual Reality Software and Technology on - VRST ’19_. ACM Press, Parramatta, NSW, Australia, 1–2. https://doi.org/10.1145/3359996.3364789 * Ullrich (2012) Sebastian Ullrich. 2012\. Haptic Palpation for Medical Simulation in Virtual Environments. _IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS_ 18, 4 (2012), 9. * Usoh et al. (1999) Martin Usoh, Kevin Arthur, Mary C. Whitton, Rui Bastos, Anthony Steed, Mel Slater, and Frederick P. Brooks. 1999. Walking > walking-in-place > flying, in virtual environments. In _Proceedings of the 26th annual conference on Computer graphics and interactive techniques - SIGGRAPH ’99_. ACM Press, Not Known, 359–364. https://doi.org/10.1145/311535.311589 * Usoh et al. (2000) Martin Usoh, Ernest Catena, Sima Arman, and Mel Slater. 2000\. Using Presence Questionnaires in Reality. _Presence: Teleoperators and Virtual Environments_ 9, 5 (Oct. 2000), 497–503. https://doi.org/10.1162/105474600566989 * Varalakshmi et al. (2012) Varalakshmi, Thriveni, Venugopal, and Patnaik. 2012\. Haptics: State of the Art Survey. _IJCSI International Journal of Computer Science Issues_ (2012). https://core.ac.uk/download/pdf/25725449.pdf * Villa Salazar et al. (2020) David Steeven Villa Salazar, Claudio Pacchierotti, Xavier De Tinguy De La Girouliere, Anderson Maciel, and Maud Marchal. 2020. Altering the Stiffness, Friction, and Shape Perception of Tangible Objects in Virtual Reality Using Wearable Haptics. _IEEE Transactions on Haptics_ (2020), 1–1. https://doi.org/10.1109/TOH.2020.2967389 * Vonach et al. (2017) Emanuel Vonach, Clemens Gatterer, and Hannes Kaufmann. 2017\. VRRobot: Robot actuated props in an infinite virtual environment. In _2017 IEEE Virtual Reality (VR)_. IEEE, Los Angeles, CA, USA, 74–83. https://doi.org/10.1109/VR.2017.7892233 * Wang et al. (2020b) Chi Wang, Da-Yuan Huang, Shuo-Wen Hsu, Cheng-Lung Lin, Yeu-Luen Chiu, Chu-En Hou, and Bing-Yu Chen. 2020b. Gaiters: Exploring Skin Stretch Feedback on the Legs for Enhancing Virtual Reality Experiences. (2020), 14. * Wang et al. (2020a) Dangxiao Wang, Yuan Guo, Zhang Yuru, XY Weiliang, and WWIA Jing. 2020a. Haptic display for virtual reality: progress and challenges | Elsevier Enhanced Reader. https://doi.org/10.3724/SP.J.2096-5796.2019.0008 ISSN: 2096-5796. * Wang et al. (2020d) Dangxiao Wang, Kouhei Ohnishi, and Weiliang Xu. 2020d. Multimodal Haptic Display for Virtual Reality: A Survey. _IEEE Transactions on Industrial Electronics_ 67, 1 (Jan. 2020), 610–623. https://doi.org/10.1109/TIE.2019.2920602 * Wang et al. (2020c) Yuntao Wang, Hanchuan Li, Zhengyi Cao, Huiyi Luo, Ke Ou, John Raiti, Chun Yu, Shwetak Patel, and Yuanchun Shi. 2020c. MoveVR: Enabling Multiform Force Feedback in Virtual Reality using Household Cleaning Robot. (2020), 12. * Wei et al. (2020) Tzu-Yun Wei, Hsin-Ruey Tsai, Yu-So Liao, Chieh Tsai, Yi-Shan Chen, Chi Wang, and Bing-Yu Chen. 2020. ElastiLinks: Force Feedback between VR Controllers with Dynamic Points of Application of Force. In _Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology_. ACM, Virtual Event USA, 1023–1034. https://doi.org/10.1145/3379337.3415836 * Wexelblat (1993) Alan Wexelblat. 1993\. Virtual reality: applications and explorations. http://libertar.io/lab/wp-content/uploads/2016/02/Virtual.Reality.-.Applications.And_.Explorations.pdf/page=164 Myron Krueger, Artificial reality 2 An easy entry to Virtual reality Chap 7. * Whitmire et al. (2018) Eric Whitmire, Hrvoje Benko, Christian Holz, Eyal Ofek, and Mike Sinclair. 2018. Haptic Revolver: Touch, Shear, Texture, and Shape Rendering on a Reconfigurable Virtual Reality Controller. In _Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems - CHI ’18_. ACM Press, Montreal QC, Canada, 1–12. https://doi.org/10.1145/3173574.3173660 * Winther et al. (2020) Frederik Winther, Linoj Ravindran, Kasper Paabol Svendsen, and Tiare Feuchtner. 2020. Design and Evaluation of a VR Training Simulation for Pump Maintenance Based on a Use Case at Grundfos. In _2020 IEEE Conference on Virtual Reality and 3D User Interfaces (VR)_. IEEE, Atlanta, GA, USA, 738–746. https://doi.org/10.1109/VR46266.2020.1580939036664 * Witmer and Singer (1998) Bob G. Witmer and Michael J. Singer. 1998. Measuring Presence in Virtual Environments: A Presence Questionnaire. _Presence: Teleoperators and Virtual Environments_ 7, 3 (June 1998), 225–240. https://doi.org/10.1162/105474698565686 * Xia et al. (2018) Haijun Xia, Sebastian Herscher, Ken Perlin, and Daniel Wigdor. 2018. Spacetime: Enabling Fluid Individual and Collaborative Editing in Virtual Reality. In _The 31st Annual ACM Symposium on User Interface Software and Technology - UIST ’18_. ACM Press, Berlin, Germany, 853–866. https://doi.org/10.1145/3242587.3242597 * Xia (2016) Pingjun Xia. 2016\. Haptics for Product Design and Manufacturing Simulation. _IEEE Transactions on Haptics_ 9, 3 (July 2016), 358–375. https://doi.org/10.1109/TOH.2016.2554551 * Yamaguchi et al. (2016) Kotaro Yamaguchi, Ginga Kato, Yoshihiro Kuroda, Kiyoshi Kiyokawa, and Haruo Takemura. 2016\. A Non-grounded and Encountered-type Haptic Display Using a Drone. In _Proceedings of the 2016 Symposium on Spatial User Interaction - SUI ’16_. ACM Press, Tokyo, Japan, 43–46. https://doi.org/10.1145/2983310.2985746 * Yang et al. (2019) Jackie (Junrui) Yang, Christian Holz, Eyal Ofek, and Andrew D. Wilson. 2019. DreamWalker: Substituting Real-World Walking Experiences with a Virtual Reality. In _Proceedings of the 32nd Annual ACM Symposium on User Interface Software and Technology - UIST ’19_. ACM Press, New Orleans, LA, USA, 1093–1107. https://doi.org/10.1145/3332165.3347875 * Ye et al. (2019) Yuan-Syun Ye, Hsin-Yu Chen, and Liwei Chan. 2019. Pull-Ups: Enhancing Suspension Activities in Virtual Reality with Body-Scale Kinesthetic Force Feedback. In _Proceedings of the 32nd Annual ACM Symposium on User Interface Software and Technology - UIST ’19_. ACM Press, New Orleans, LA, USA, 791–801. https://doi.org/10.1145/3332165.3347874 * Yixian et al. (2020) Yan Yixian, Kazuki Takashima, Anthony Tang, Takayuki Tanno, Kazuyuki Fujita, and Yoshifumi Kitamura. 2020. ZoomWalls: Dynamic Walls that Simulate Haptic Infrastructure for Room-scale VR World. In _Proceedings of the 33rd Annual ACM Symposium on User Interface Software and Technology_ _(UIST ’20)_. Association for Computing Machinery, New York, NY, USA, 223–235. https://doi.org/10.1145/3379337.3415859 * Yokokohji et al. (1999) Yasuyoshi Yokokohji, Ralph L. Hollis, and Takeo Kanade. 1999\. WYSIWYF Display: A Visual/Haptic Interface to Virtual Environment. _Presence: Teleoperators and Virtual Environments_ 8, 4 (Aug. 1999), 412–434. https://doi.org/10.1162/105474699566314 * Yokokohji et al. (2001) Y. Yokokohji, J. Kinoshita, and T. Yoshikawa. 2001\. Path planning for encountered-type haptic devices that render multiple objects in 3D space. In _Proceedings IEEE Virtual Reality 2001_. 271–278. https://doi.org/10.1109/VR.2001.913796 * Yokokohji et al. (2005) Yasuyoshi Yokokohji, Nobuhiko Muramori, Yuji Sato, and Tsuneo Yoshikawa. 2005. _Haptic Display for Multiple Fingertip Contacts Based on the Observation of Human Grasping Behaviors_. * Yoshida et al. (2020) Shigeo Yoshida, Yuqian Sun, and Hideaki Kuzuoka. 2020\. PoCoPo: Handheld Pin-based Shape Display for Haptic Rendering in Virtual Reality. In _Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems_. ACM, Honolulu HI USA, 1–13. https://doi.org/10.1145/3313831.3376358 * Zenner and Kruger (2017) Andre Zenner and Antonio Kruger. 2017. Shifty: A Weight-Shifting Dynamic Passive Haptic Proxy to Enhance Object Perception in Virtual Reality. _IEEE Transactions on Visualization and Computer Graphics_ 23, 4 (April 2017), 1285–1294. https://doi.org/10.1109/TVCG.2017.2656978 * Zenner and Krüger (2019) André Zenner and Antonio Krüger. 2019. Drag:on: A Virtual Reality Controller Providing Haptic Feedback Based on Drag and Weight Shift. In _Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems - CHI ’19_. ACM Press, Glasgow, Scotland Uk, 1–12. https://doi.org/10.1145/3290605.3300441 * Zhao (2009) QinPing Zhao. 2009\. A survey on virtual reality. _Science in China Series F: Information Sciences_ 52, 3 (March 2009), 348–400. https://doi.org/10.1007/s11432-009-0066-0 * Zhao and Follmer (2018) Yiwei Zhao and Sean Follmer. 2018. A Functional Optimization Based Approach for Continuous 3D Retargeted Touch of Arbitrary, Complex Boundaries in Haptic Virtual Reality. In _Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems - CHI ’18_. ACM Press, Montreal QC, Canada, 1–12. https://doi.org/10.1145/3173574.3174118 * Zhao et al. (2017) Yiwei Zhao, Lawrence H. Kim, Ye Wang, Mathieu Le Goc, and Sean Follmer. 2017. Robotic Assembly of Haptic Proxy Objects for TangibleInteraction and Virtual Reality. In _Proceedings of the Interactive Surfaces and Spaces on ZZZ - ISS ’17_. ACM Press, Brighton, United Kingdom, 82–91. https://doi.org/10.1145/3132272.3134143 * Zhou and Deng (2009) Ning-Ning Zhou and Yu-Long Deng. 2009. Virtual reality: A state-of-the-art survey. _International Journal of Automation and Computing_ 6, 4 (Nov. 2009), 319–325. https://doi.org/10.1007/s11633-009-0319-9 * Ziat et al. (2014) Mounia Ziat, Taylor Rolison, Andrew Shirtz, Daniel Wilbern, and Carrie Anne Balcer. 2014. Enhancing virtual immersion through tactile feedback. In _Proceedings of the adjunct publication of the 27th annual ACM symposium on User interface software and technology - UIST’14 Adjunct_. ACM Press, Honolulu, Hawaii, USA, 65–66. https://doi.org/10.1145/2658779.2659116 * Zielasko and Riecke (2020) Daniel Zielasko and Bernhard E Riecke. 2020. Either Give Me a Reason to Stand or an Opportunity to Sit in VR. (2020), 3. * Zilles and Salisbury (1995) C. B. Zilles and J. K. Salisbury. 1995. A constraint-based god-object method for haptic display. In _In International Conference on Intelligent Robots and Systems_. 146–151. * Zimmermann (2008) Peter Zimmermann. 2008\. Virtual Reality Aided Design. A survey of the use of VR in automotive industry. (Jan. 2008). https://doi.org/10.1007/978-1-4020-8200-9_13
# Sending or not sending twin-field quantum key distribution with distinguishable decoy states Yi-Fei Lu Mu-Sheng Jiang<EMAIL_ADDRESS>Yang Wang Xiao-Xu Zhang Fan Liu Chun Zhou Hong-Wei Li Wan-Su Bao<EMAIL_ADDRESS>Henan Key Laboratory of Quantum Information and Cryptography, SSF IEU, Zhengzhou, Henan 450001, China Synergetic Innovation Center of Quantum Information and Quantum Physics, University of Science and Technology of China, Hefei, Anhui 230026, China ###### Abstract Twin-field quantum key distribution (TF-QKD) and its variants can overcome the fundamental rate-distance limit of QKD which has been demonstrated in the laboratory and field while their physical implementations with side channels remains to be further researched. We find the external modulation of different intensity states through the test, required in those TF-QKD with post-phase compensation, shows a side channel in frequency domain. Based on this, we propose a complete and undetected eavesdropping attack, named passive frequency shift attack, on sending or not-sending (SNS) TF-QKD protocol given any difference between signal and decoy states in frequency domain which can be extended to other imperfections with distinguishable decoy states. We analyze this attack by giving the formula of upper bound of real secure key rate and comparing it with lower bound of secret key rate under Alice and Bob’s estimation with the consideration of actively odd-parity pairing (AOPP) method and finite key effects. The simulation results show that Eve can get full information about the secret key bits without being detected at long distance. Our results emphasize the importance of practical security at source and might provide a valuable reference for the practical implementation of TF- QKD. ††preprint: APS/123-QED ## I Introduction Quantum key distribution (QKD) allows two distant parties, Alice and Bob, to share secret keys securely in the presence of an eavesdropper, Eve, by harnessing the laws of physics [1, 2, 3]. Combined with one-time pad, Alice and Bob can achieve unconditionally secure private communication. Notable progress has been made to improve performance, such as the communication distance and secret key rate, and bridge the gap between the idealized device models assumed in security proofs and the functioning of realistic devices in practical systems. The measurement-device-independent (MDI) QKD [4] can remove both known and unknown security loopholes, or so-called side channels, in the measurement unit perfectly which shift the focus of quantum attacks to the source. Photon- number-splitting (PNS) attack [5, 6], the major threat at source since single- photon source is not available at present and weak laser light are widely used in practical QKD systems, has been overcome by the decoy-state method [7, 8]. Combining these two methods, the decoy-state MDI-QKD equipped with some security patches performs well with imperfect single-photon sources. However, the key rate and communication distance are two implementation bottlenecks. To exceed the linear scale of key rate [9, 10], twin-field QKD (TF-QKD) was proposed by Lucamarini _et al._ [11] whose key rate scales linearly with square-root of the channel transmittance $\eta$ by harnessing single-photon interference over long distance. Though the security is not completed and the security loophole is caused by the later announcement of the phase information [12], then many variants of TF-QKD [12, 13, 14, 15, 16, 17] have been proposed to deal with this security loophole and each has its advantages. Many effects have been considered in real-life implementation to accelerate its application, including finite key effects, the number of states with different intensities, the phase slice of appropriate and asymmetric transmission distance, etc. [18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34]. Meanwhile, several experiments of TF-QKD have been completed in laboratory and field to demonstrate its ability to overcome the rate-distance limit [35, 36, 37, 38, 39, 40, 41, 42]. However, the physical implementations of TF-QKD protocols with side channels remains to be further researched at present. Since TF-QKD retains the MDI characteristic, we should just focus on light source. Ideally, it is assumed that the sending devices are placed in a protected laboratory, and can prepare and encode quantum states correctly. Unfortunately, these conditions are not met in practical systems, and state preparation flaws (SPFs) and leakage may be induced from imperfect devices or Eve’s disturbance. A small imperfection at source does not necessarily mean a small impact on the secret key rate, because Eve could enhance such imperfection by exploiting channel loss. Therefore, Eve can steal secret information actively by performing Trojan- horse attack [43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55] on modulators, lasers or attenuators and so on, or passively by harnessing the SPFs and leakage caused by imperfect devices [5, 6, 56, 57, 58, 59]. In those QKD protocols with imperfect single-photon sources, the decoy-state method is vital and employed to monitor the channel eavesdropping whose security is based on the fact that Eve could not distinguish between decoy and signal states. In practice, however, it does not with the behavior of the real apparatuses or Eve’s disturbance. For instance, the probability distributions of signal states and decoy states do not overlap in time domain totally with pump-current modulation [58]. Besides, signal states and decoy states can be distinguished with external intensity modulation in frequency domain when Eve applying wavelength-selected photon-number-splitting attack actively in ’plug- and-play’ systems [46]. This loophole is introduced by decoy-state method and caused by the imperfections of modulators and modulation voltage. As the decoy-state method is employed in TF-QKD protocols, it is of significance to analyze its practical security in this aspect. In this paper, we concentrate on the sending or not-sending (SNS) TF-QKD protocol [12] with actively odd-parity pairing (AOPP) method and finite-key effects taken into consideration [20, 21], and propose a complete and undetected eavesdropping attack, named passive frequency shift attack, which can take advantage of the most general side channels in frequency domain and can be extended to other cases with distinguishable decoy states. In Sec. II, we recap the frequency shift of intensity modulators (IMs) and test experimentally the spectral distribution of signal pulses with external modulation method which shows a side channel in frequency domain. In Sec. III, we propose a passive frequency shift attack on SNS protocol and analyze its adverse impact by giving the formula of upper bound of secure key rate and comparing it with the lower bound of secret key rate under Alice and Bob’s estimation with consideration of AOPP method and finite-key effects. In Sec. IV, we present our simulation results. Last, we give some discussion about the countermeasure of side channels in Sec. V and conclude in Sec. VI. ## II Frequency shift of intensity modulators In this section, we will recap the frequency shift of IMs and test experimentally to show a side channel in frequency domain. There are several kinds of intensity modulators, such as the Mach-Zehnder type electro-optical intensity modulators (EOIMs), electro-absorption modulators (EAMs), and acousto-optical modulators (AOMs). EOIMs, especially LiNbO3-based devices, possess the excellent performance of wavelength-independent modulation characteristics, excellent extinction performance (typically 20 dB) and low insertion losses (typically 5 dB) [60]. LiNbO3-based EOIMs work by the principle of interference, controlled by modulating the optical phase. The incoming light is coupled into a waveguide and then split into two paths of a Mach-Zehnder interferometer equally and interfere at an output coupler. The two arms made of lithium niobate will induce a phase change when applying voltages. Accordingly, the intensity and phase of output light will be modulated after interference depending on the applied electrical voltages. Assuming voltages $V_{1}(t)$ and $V_{2}(t)$ are applied to two arms separately with the input field of intensity $E_{0}$ and frequency $\omega_{0}$, the output field can be written as [46] $\displaystyle E_{\rm out}(t)=E_{0}{\rm cos}[\Delta\varphi(t)]e^{i[\omega_{0}t+\varphi(t)]},$ (1) where $\Delta\varphi(t)=[\gamma V_{1}(t)+\varphi_{1}-\gamma V_{2}(t)-\varphi_{2}]/2$ and $\varphi(t)=[\gamma V_{1}(t)+\varphi_{1}+\gamma V_{2}(t)+\varphi_{2}]/2$, $\gamma=\pi/V_{\pi}$ is the voltage-to-phase conversion cofficients for two arms, and $\varphi_{1}$ and $\varphi_{2}$ are the static phases which we will omit for simplicity. Here, $V_{\pi}$ is half- wave voltage that is required to change the phase in one modulator arm by $\pi$ radians. The output intensity is given by $\displaystyle P_{\rm out}(t)=|E_{\rm out}(t)|^{2}=\frac{P_{0}}{2}\bigl{[}1+{\rm cos}[\gamma V(t)]\bigr{]},$ (2) where $V(t)=V_{1}(t)-V_{2}(t)$ and $P_{0}$ is the input optical power. The phase maintains and intensity is determined by Eq. 2 on condition that the two modulator arms are driven by the same amount, but in opposite directions (i.e. $V_{1}(t)=-V_{2}(t)$), which is known as balanced driving or push–pull operation. When $V(t)$ is constant we will get pure intensity modulation without frequency shift. However, once $V(t)$ is not a constant any more, something unexpected will arise to the output field. Specifically, frequency shift will be induced as we can see, for example, if $V_{1}(t)=-V_{2}(t)=V_{0}+kt$, the output field can be expressed as [46] $\displaystyle E_{\rm out}=\frac{E_{0}}{2}\Bigl{[}e^{i[(\omega_{0}+\gamma k)t+\gamma V_{0}]}+e^{i[(\omega_{0}-\gamma k)t+\gamma V_{0}]}\Bigr{]}.$ (3) From Eq. 3, we can see a frequency shift of the light pulses with $\pm\omega_{m}=\pm\gamma k$ compared with the original $\omega_{0}$, where $k$ is the slope of modulation voltage. Moreover, the frequency shift of the output field will be more confusing when the modulation voltages are more complicated in practical systems. We can analyze the spectrum of the output field using the fast Fourier transform method. To evaluate the frequency shift of different intensity pulses, we test it in principle. Optical pulses with 1 ns pulse width are produced with a constant intensity from a laser diode (Keysight 8164B) first, modulated by an IM driven by an arbitrary waveform generator (Keysight M9505A) and measured by an optical spectrum analyzer (Yokogawa AQ637D) last. Note that the measurement is taken before the fixed attenuation as implemented in actual systems since the probability of emitting pulses at the single-photon level follows the same distribution. Fig. 1 illustrates the wavelength spectrum of three states with intensity ratio taken from the SNS experiment [39] as 0.1: 0.384: 0.447 ($\mu_{a}=0.1$, $\mu_{b}=0.384$, $\mu_{z}=0.447$). And the normalized intensity probability distributions are shown in Fig. 2 to distinguish the difference. Figure 1: The wavelength spectrum of three different states with intensity ratio as 0.1: 0.384: 0.447 ($\mu_{a}=0.1$, $\mu_{b}=0.384$, $\mu_{z}=0.447$). Figure 2: The normalized intensity distribution of three different states with intensity ratio as 0.1: 0.384: 0.447 ($\mu_{a}=0.1$, $\mu_{b}=0.384$, $\mu_{z}=0.447$). Two dashed lines at 1550.1125 nm and 1550.131 nm are boundaries that can be used to distinguish states with intensity $\mu_{z}$ ($\mu_{b}$) and $\mu_{a}$. There are also some subtle differences between states $\mu_{z}$ and $\mu_{b}$ in frequency domain. Obviously, the states modulated by IMs of different intensities do not overlap totally in frequency domain. The distinction of signal states (also strong decoy states) and weak decoy states is evident as expected because the amplitude of modulation voltages of signal and strong decoy states are higher which will induce more frequency shift, and thus the peaks of the signal and strong decoy states are lower than weak decoy states. More precisely, the normalized probability of weak decoy states is higher than signal and strong decoy states between 1550.112 nm and 1550.131 nm. There are also slight differences between signal and strong decoy states. The difference will be bigger when narrowing pulses or the line width as TF-QKD requires, or when the intensity difference of the different states becomes larger. On this foundation, Eve can apply a passive frequency shift attack by harnessing this side channel. ## III Passive frequency shift attack on SNS protocol In this section, we propose a passive frequency shift attack scheme on practical SNS TF-QKD systems by exploiting the side channels in frequency domain and analyze its adverse impact by fiving the formula of upper bound of secure key rate with consideration of AOPP method and finite-key effects. But we note that this attack can be applied with other side channels except in frequency domain. In fact, the signal pulses (including signal states and decoy states) and reference pulses must be modulated with a stable continuous-wave laser source with external modulation method so as to estimate and compensate phase noise in those TF-QKD protocols which need post-phase compensation, such as SNS TF- QKD [12] and phase-matching (PM) TF-QKD [13]. Even if the synchronization can not be controlled by Eve unlike discussed in Ref. [46, 47], the probability distributions of signal states and decoy states may do not overlap inevitably in frequency domain. In the 4 intensity SNS TF-QKD protocol (see Appendix A for details), Alice and Bob need to modulate the continuous light to 5 different intensities, the maximum intensity pulses are used as phase reference pulses, the minimum as vacuum states, while others as signal states, weak and strong decoy states. In practical SNS systems [38, 39], three IMs are used to modulate these 5 different pulses to ensure that the output signal intensities are in agreement with the theoretical requirements and the reference detections are high enough for phase compensation. There will be side channels in frequency domain inevitably. In what follows, we introduce the passive frequency shift attack. Suppose Eve intercepts all signal pulses at Alice and Bob’s output ports where the signal pulses haven’t been attenuated by channels, and then distinguishes signal and decoy states with a wavelength division multiplexer (WDM) and three single-photon detectors (SPDs) as illustrated in Fig. 3. Set internals $T_{\alpha}$ properly, where $\alpha\in\\{z,a,b\\}$, according to the wavelength spectrum of different states to distinguish the states with intensity $\mu_{\alpha}$ but may fail with a certain probability. Theoretically, $T_{\alpha}$ can be set as the union of two symmetric intervals according to Eq. 3. Figure 3: Schematic of the passive frequency shift attack. PM: phase modulator, IM: intensity modulator, ATT: attenuator, WDM: wavelength division multiplexer, SPD: single-photon detector, S: light path selector, BS: beam splitter, SNSPD: superconducting nanowire single-photon detectors. The light path selector S1 (S2) is controlled by SPD1 (SPD1 and SPD2) and Bob’s device is the same as Alice’s which we have omitted in figure. Suppose the four ports of WDM 1, 2, 3 and 4 can export photons with frequency located in $T_{z}$, $T_{a}$, $T_{b}$ and others. The light path selector S1 (S2) is controlled by SPD1 (SPD1 and SPD2). Denote as 1 or 0 when SPD (i.e. SPD1, SPD2 or SPD3) clicks or not, and 1 or 0 when the light path selector (i.e. S1 or S2) selects the up or down path. Then we set $\displaystyle{\rm S1}$ $\displaystyle=\overline{{\rm SPD1}},$ $\displaystyle{\rm S2}$ $\displaystyle={\rm SPD1}\vee{\rm SPD2}.$ (4) Note that only one SPD at most will click under this principle. According to the response of SPDs, set the total transmittance as $\eta_{z}$, $\eta_{a}$, $\eta_{b}$ or $\eta_{k}$ when SPD1, SPD2, SPD3 clicks or none of them click, respectively. In this process, Eve could get partial raw key bits after Alice (Bob) announces their signal and decoy windows, which can be understood in this way that Eve can conclude their key bits as 1 (0) when ${\rm SPD1}\vee{\rm SPD2}\vee{\rm SPD3}=1$ in $Z$ windows. There is no bit-flip error between Alice (Bob) and Eve because Eve can intercepts photons at output ports without stray photons. Only the raw bits are secure in $Z$ windows when ${\rm SPD1}\vee{\rm SPD2}\vee{\rm SPD3}=0$ on both sides. On the one hand, once Eve detects photons successfully on one side, the bit is either the same or a bit-flip error with the other side which will be revealed in the error correction step (and pre-error correction process when AOPP is performed). On the other hand, the bits are balanced (i.e. random for Eve) in one-detector heralded events with ${\rm SPD1}\vee{\rm SPD2}\vee{\rm SPD3}=0$ which means the raw bits are secure in these windows. Though Eve could not distinguish the decoy states and signal states without errors, once the transmittance of signal and decoy states differ from others, the decoy-state method may not estimate the count rate and phase-flip error rate of single-photon states correctly. When the actual secure key rate is lower than the estimated one, the final secret string is insecure partially. We emphasize that this attack will not introduce unnecessary errors as the beam splitting and measurement by Eve can be seen as a loss and the rest of the pulses are coherent states with no phase noise. What is more, Eve could control errors completely expect the inherent errors of the protocol through channels, superconducting nanowire single-photon detectors (SNSPDs) and classic information he announces. In the following, we will analyze the negative effect of this passive frequency shift attack. Consider the most general case, we assume that the envelope of wavelength spectrum can be written as $f_{\beta}(\lambda)$, where $\beta\in\\{z,a,b\\}$. The wavelength spectrum of signal states and decoy states could not totally overlap. By setting the internals $T_{\alpha}$ properly, Eve can distinguish the decoy states and siganl states with errors. The proportion of state $\mu_{\beta}$ in internals $T_{\alpha}$ can be shown as $\displaystyle r_{\beta|\alpha}=\int_{T_{\alpha}}f_{\beta}(\lambda)d\lambda,$ (5) where $\alpha$, $\beta\in\\{z,a,b\\}$. Which one detector will click obeys a certain probability distribution. Therefore, the states of intensity $\mu_{\beta}$ would be transformed with one of four different transmittance $\varOmega_{\beta}=\\{\eta_{\beta z},\eta_{\beta a},\eta_{\beta b},\eta_{\beta k}\\}$ with a finite probability, where $\displaystyle\eta_{\beta z}=$ $\displaystyle\eta_{z}(1-r_{\beta|z}),$ $\displaystyle\eta_{\beta a}=$ $\displaystyle\eta_{a}(1-r_{\beta|z}-r_{\beta|a}),$ $\displaystyle\eta_{\beta b}=$ $\displaystyle\eta_{b}(1-r_{\beta|z}-r_{\beta|a}-r_{\beta|b}),$ $\displaystyle\eta_{\beta k}=$ $\displaystyle\eta_{k}(1-r_{\beta|z}-r_{\beta|a}-r_{\beta|b}).$ (6) Table 1: List of experimental parameters. Here, $\gamma$ is the fiber loss coefficient (dB/km), $\eta_{d}$ is the detection efficiency of detectors, $e_{d}$ is the misalignment-error probability, $f_{\rm EC}$ is the error correction inefficiency, $\xi$ is the failure probability of statiscal fluctuations analysis, $p_{d}$ is the dark count rate and $M$ is the number of phase slices. $\gamma$ | $\eta_{d}$ | $e_{d}$ | $f_{\rm EC}$ | $\xi$ | $p_{d}$ | $M$ ---|---|---|---|---|---|--- 0.2 | 56% | 0.1 | 1.1 | $2.2\times 10^{-9}$ | $10^{-10}$ | 16 Table 2: List of experimental parameters about intensity and probability Alice and Bob select. $\mu_{a}$ | $\mu_{b}$ | $\mu_{z}$ | $p_{z}$ | $p_{a}$ | $p_{b}$ | $p_{z0}$ ---|---|---|---|---|---|--- 0.1 | 0.384 | 0.447 | 0.776 | 0.85 | 0.073 | 0.732 Table 3: Five groups of the proportion $r_{\beta|\alpha}$ of state $\mu_{\beta}$ in internals $T_{\alpha}$. | $r_{z|z}$ | $r_{a|z}$ | $r_{b|z}$ | $r_{z|a}$ | $r_{a|a}$ | $r_{b|a}$ | $r_{z|b}$ | $r_{a|b}$ | $r_{b|b}$ ---|---|---|---|---|---|---|---|---|--- G1 | 0.01 | 0.008 | 0.01 | 0.01 | 0.1 | 0.01 | 0.01 | 0.008 | 0.01 G2 | 0.01 | 0.008 | 0.01 | 0.005 | 0.1 | 0.005 | 0.01 | 0.008 | 0.01 G3 | 0.01 | 0.008 | 0.01 | 0.01 | 0.2 | 0.01 | 0.01 | 0.008 | 0.01 G4 | 0.01 | 0.005 | 0.01 | 0.005 | 0.2 | 0.005 | 0.01 | 0.005 | 0.01 G5 | 0.01 | 0.008 | 0.01 | 0 | 0.1 | 0 | 0.01 | 0.008 | 0.01 Here, the attenuation coefficient comes from Eve’s interfection and detection, and the total transmittance can be controlled by Eve completely which means that Eve is allowed to use a lower-loss or even lossless channel and perfect detectors with 100% detection efficiency and no dark count, and could select internals freely to obtain a satisfactory results. For states with intensity $\mu_{\beta}$, the probability of being transmitted with $\eta_{\beta\gamma}\in\varOmega_{\beta}$, $\gamma\in M=\\{z,a,b,k\\}$ can be shown as $\displaystyle p_{\beta z}$ $\displaystyle=(1-e^{-\mu_{\beta|z}}),$ $\displaystyle p_{\beta a}$ $\displaystyle=e^{-\mu_{\beta|z}}(1-e^{-\mu_{\beta|a}}),$ $\displaystyle p_{\beta b}$ $\displaystyle=e^{-\mu_{\beta|z}}e^{-\mu_{\beta|a}}(1-e^{-\mu_{\beta|b}}),$ $\displaystyle p_{\beta k}$ $\displaystyle=e^{-\mu_{\beta|z}}e^{-\mu_{\beta|a}}e^{-\mu_{\beta|b}},$ (7) where $\mu_{\beta|\alpha}=\mu_{\beta}r_{\beta|\alpha}$. Here, $e^{-\mu_{\beta|\alpha}}$ is the probatility of zero photon in internals $T_{\alpha}$ with intensity $\mu_{\beta}$. Since TF-QKD protocols are proposed for the implementation of long optical fiber communications, Eve’s best target is to acquire more percentage of keys by minimizing $\eta_{k}$ as far as possible while maintaining the key rate and communication distance under Alice and Bob’s estimation. When the communication distance is long enough, Eve may steal all secret key bits. There are two key rates matter: the lower bound of secret key rate under Alice and Bob’s estimation $R_{e}$ and the upper bound of real secure key rate $R_{u}$. Note that Alice and Bob could not estimate $R_{e}$ correctly under this attack since it is impossible to pick out the decoy states that have undergone the same operation as the signal states, i.e. the decoy-state method does not work properly. When AOPP method is performed with partial bits leaked to Eve, Bob only chooses odd-parity bit pairs and they will keep the second bit if Alice’s bits pairs are odd, too. In this process, Eve’s information on raw bits does not reduce as the parity information is public. Therefore, the upper bound of secure key rate $R_{u}$ can be shown as $\displaystyle R_{u}=n_{1,{\rm sec}}[1-H(e_{\rm in})]-n_{t}^{\prime}fH(E_{Z}^{\prime}),$ (8) where $n_{1,{\rm sec}}$ is the upper bound of single photons which can be used to distill secure key bits and $e_{\rm in}$ is the inherent phase-flip error rate (i.e. the lower bound of untagged bits) determined by the number of phase slices $M$. These two parameters will be discussed in Appendix B. This attack can be applied as long as the wavelength spectrum distributions of signal and decoy states are different. The effect of this attack varies based on the discernibility of different states. However, this attack can be extended to other imperfections with distinguishable decoy states, like the polarization, temporal shape, etc. by replacing the WDM with appropriate devices. ## IV Numerical simulations Figure 4: The estimated lower bound of secret key rate $R_{e}$ and upper bound of real secure key rates $R_{ui}$ in logarithmic scale versus tranmission distance (between Alice and Bob) under passive frequency shift attack, where $i\in\\{1,2,3,4,5\\}$, with experimental parameters listed in Table 1, 2 and 3. The solid line corresponds to the etimated key rate $R_{e}$ which is the same as that without attack, while the dashed lines represent the upper bound of real secure key rate $R_{ui}$. And the real secure key rate $R_{u5}$ not shown in the figure is identically 0. We numerically simulate the behaviour of SNS protocol under passive frequency shift attack in this section. There are nine parameters obtained by statistic in practical systems before AOPP, including $n_{\alpha\beta}$, $n_{\Delta^{+}}^{R}$, $n_{\Delta^{-}}^{L}$, $n_{t}$ and $E_{z}$, where $n_{t}=n_{\rm sig}+n_{\rm err}$ is the length of raw keys, $\alpha\beta\in S=\\{vv,va,av,vb,bv\\}$. Here, $E_{z}=n_{\rm err}/n_{t}$, $n_{\rm sig}$ and $n_{\rm err}$ is the number of right and wrong raw bits, respectively. Under passive frequency shift attack, the parameters can be simulated as discussed in Appendix C. The promotion to AOPP is trival as the attack is not applied. For simulation purposes, the experimental parameters listed in Tables 1 and 2 are taken according to the SNS experiment [39] with a little modification. Then we simulate the normal secret key rate without frequency shift attack, the key rate under Alice and Bob’s estimation and the upper bound of secure key rate under frequency shift attack with AOPP and finite-key effects by selecting five groups of reasonable parameters about $r_{\beta|\alpha}$ listed in Tables 3 following the principle that the difference between signal states (strong decoy states) and weak decoy states is significant while small between signal and strong decoy states. In Fig. 4, the estimated key rate $R_{e}$ under passive frequency shift attack represented by the solid line is the same as that without attack which means this attack will not be detected by Alice and Bob with the list of experimental parameters in Table 1, 2 and 3. In comparison, the dashed lines represent the upper bound of secure key rates $R_{u}$ under frequency shift attack. Denote the the upper bound of secure key rate under parameters in Group $i$ (G$i$) in Table 3 as $R_{ui}$, where $i\in\\{1,2,3,4,5\\}$. Compared $R_{u1}$ with $R_{u2}$, we can find that the difference between weak decoy states and signal states (strong decoy states) affects the effects of attack significantly i.e. the communication distance is limited from 472 km to 368 km when the difference changes from 10 times to 20 times. The negative effects will be more greater when $r_{z|a}:r_{a|a}:r_{b|a}$ is constant but the absolute values increase by comparing $R_{u2}$ with $R_{u3}$. The secure distance will be limited to shorter when the difference becomes larger by contrasting $R_{u2}$ and $R_{u4}$. Besides, Alice and Bob could not distribute keys when Eve can distinguish weak decoy states correctly with parameters of Group 5. On the one hand, if Alice and Bob attach importance to these side channels and know Eve’s action, this attack will limit the communication distance. Note that the distance and key rate will be more pessimistic since the key rates $R_{ui}$ is only the upper bound. Otherwise, the secret key bits will be insecure over long distance especially. For example, the key bits are all insecure when the estimated key rate is $10^{-8}$ per pulse at 479 km (an acceptable value at long distance). In this attack, we have assumed that Eve intercepts photons in the order of $T_{z}$, $T_{a}$ and $T_{b}$, but we find there is no obvious difference when changing this order through numerical simulations which can be understood in this way that only the photons are secure with ${\rm SPD1}\vee{\rm SPD2}\vee{\rm SPD3}=0$, i.e. only the pulses that are not detected by Eve in all internals $T_{\alpha}$ are secured, or there will be no difference in any order when $r_{\alpha|\beta}=0$ with $\alpha\neq\beta$. ## V Discussion This eavesdropping attack proposed above is a passive attack harnessing the side channels and hard to be detected. To guarantee security in practical systems with side channels, the first potential way is to improve experimental techniques or modulation methods to restrain side channels but may be unavoidable when one is closed, but another appears. The second alternative is to develop mathematical models in theory to maximize the secure key rates under attacks, like the loss-tolerant method [61, 62, 63, 64, 65] but needs an accurate characterization of real apparatuses. It may be an ongoing search for side channels to guarantee the practical security of QKD systems. ## VI Conclusion The goal of QKD at present is to provide long-distance and high-speed key distribution, which will induce side channels inevitably. Increasing repetition rate and narrowing pulses to improve speed will make the pulses complex and distinguishable, like the frequency, polarization, temporal shape, and so on. Any small imperfections may be exploited and enhanced utilizing channel loss by Eve, especially at long distance. Therefore, it is necessary to pay more attenuation to the practical security of TF-QKD systems. In this paper, we have investigated and tested the side channels with external modulation which is required in those TF-QKD protocols with post-phase compensation, like SNS TF-QKD [12] and PM TF-QKD [13]. Based on this, we propose a complete and undetected eavesdropping attack named passive frequency shift attack on SNS protocol which can be applied once there are differences between different states in frequency domain and can be extended to other imperfections with distinguishable decoy states. Normally, Alice and Bob could estimate the lower bound of secret key rate correctly no matter what Eve does. But this estimation is not accurate once Eve’s operation on signal and decoy states is different which may cause insecure bits when the upper bound of secure key rate is lower than the estimated lower one. According to the numerical results, Eve can get full information about the secret key bits at long distance if Alice and Bob neglect this distinguishability. For example, the key bits are all insecure when the estimated key rate is $10^{-8}$ per pulse at 479 km under the five selected groups of parameters. As there is a variety of potentially exploitable loopholes at source, our results emphasize the practical security of the light source. It is a constant search to build hardened implementations of practical QKD systems. ###### Acknowledgements. This work is supported by National Key Research and Development Program of China (Grant No. 2020YFA0309702) and National Natural Science Foundation of China (Grants N0. 61605248, No. 61675235 and No. 61505261). ## Appendix A SNS TF-QKD protocol We make a review of the SNS protocol and the key rate formula with AOPP method and finite-key effects [12, 20, 21] in this section. (1) Preparation and measurement. At any time window $i$, Alice (Bob) randomly determines whether it is a signal window or a decoy window with probabilities $p_{z}$ and $p_{x}=1-p_{z}$. If it is a signal window, Alice (Bob) sends a phase-randomized coherent state with intensity $\mu_{z}$ and denotes it as 1 (0), or a vacuum state $|0\rangle$ and denotes it as 0 (1) with probabilities $p_{z1}=1-p_{z0}$ and $p_{z0}$, seperately. If it is a decoy window, Alice (Bob) sends a phase-randomized coherent state $|\sqrt{\mu_{a}}e^{i\theta_{A}}\rangle$, $|\sqrt{\mu_{b}}e^{i\theta_{A}^{\prime}}\rangle$ or $|0\rangle$ ($|\sqrt{\mu_{a}}e^{i\theta_{B}}\rangle$, $|\sqrt{\mu_{b}}e^{i\theta_{B}^{\prime}}\rangle$ or $|0\rangle$) with probabilities $p_{a}$, $p_{b}$ and $p_{v}=1-p_{a}-p_{b}$, where $\mu_{a}<\mu_{b}$. The third party, Chrelie, renamed as Eve is supposed to perform interferometic measurements on the incoming pulses and announce the results. (2) Different types of time windows. Suppose Alice and Bob repeat the above process $N$ times, then they announce their signal windows and decoy windows through public channels. If both Alice and Bob determine a signal window, it is a $Z$ window. And the effective events in $Z$ windows are defined as one- detector heralded events no matter which detector clicks. Alice and Bob will get two raw $n_{t}$-bit strings $Z_{A}$ and $Z_{B}$ according to effective events in $Z$ windows. Note that the phase-randomized coherent state of intensity $\mu$ is equivalent to a probabilistic mixture of different photon- number states $\sum_{k=0}^{\infty}\frac{e^{-\mu}\mu^{k}}{k!}|k\rangle\langle k|$. Therefore, we can define $Z_{1}$ windows as a subset of $Z$ windows when only one party determines to send and she (he) actually sends a single-photon state $|1\rangle$. The bits from effective $Z_{1}$ windows are regarded as untagged bits by the tagged model [66]. Then the intensity of pulses would be announced to each other expect the intensity in $Z$ windows. If both commit to a decoy window, it is an $X$ window. Alice and Bob also announce their phase information $\theta_{A}$, $\theta_{B}$ when they choose the same intensity $\mu_{a}$ in an $X$ window denoted as an $X_{a}$ window. And if only one detector clicks in $X_{a}$ windows with phases satisfying $\displaystyle|\theta_{A}-\theta_{B}-\varphi_{AB}|\leq\Delta/2$ (9) or $\displaystyle|\theta_{A}-\theta_{B}-\pi-\varphi_{AB}|\leq\Delta/2,$ (10) it is an effective event in $X_{a}$ windows. All effective events in $X_{a}$ windows can be divided into two subsets as $C_{\Delta^{+}}$ and $C_{\Delta^{-}}$ according Eq. 9 and Eq. 10, respectively. And the number of events in $C_{\Delta^{+}}$ and $C_{\Delta^{-}}$ can be defined as $N_{\Delta^{+}}$ and $N_{\Delta^{-}}$. Here, $\varphi_{AB}$ is set properly to obtain a satisfactory key rate which will be different over time due to phase drift and can be obtained with reference pulses. In the following, we will omit the phase drift without loss of generality and set $\varphi_{AB}=0$. (3) Parameter estimation. They can estimate parameters, including the bit-flip error rate of the raw bits $E_{Z}$, the lower bound of untagged bits $\underline{n}_{1}$ (or the lower bound of the counting rate $\underline{s}_{1}$ equivalently) and the upper bound of the phase-flip error rate of untagged bits $\overline{e}_{1}^{ph}$. The bit-flip error rate $E_{Z}$ can be obtained by error test, $\underline{s}_{1}$ and $\overline{e}_{1}^{ph}$ can be estimated with decoy state method as follows. Denote $\rho_{v}=|0\rangle\langle 0|$, $\rho_{a}=\sum_{k=0}^{\infty}e^{-\mu_{a}}\mu_{a}^{k}/k!|k\rangle\langle k|$ and $\rho_{b}=\sum_{k=0}^{\infty}e^{-\mu_{b}}\mu_{b}^{k}/k!|k\rangle\langle k|$, where $\rho_{a}$ and $\rho_{b}$ are density operators of the phase- randomized coherent states used in $X$ windows in which the phase is not announced. Let $N_{\alpha\beta}$ be the number of intsnces Alice sends state $\rho_{\alpha}$ and Bob sends state $\rho_{\beta}$ and $n_{\alpha\beta}$ be the number of corresponding one-detector heralded events, where $\alpha\beta\in S=\\{vv,va,av,vb,bv\\}$. Thus, the counting rate can be defined as $S_{\alpha\beta}=n_{\alpha\beta}/N_{\alpha\beta}$. And $\underline{s}_{1}$ can be estimated with decoy-state method as [67, 18] $\displaystyle\underline{s}_{1}\geq$ $\displaystyle\frac{1}{2\mu_{a}\mu_{b}(\mu_{b}-\mu_{a})}[\mu_{b}^{2}e^{\mu_{a}}(S_{va}+S_{av})$ (11) $\displaystyle-\mu_{a}^{2}e^{\mu_{b}}(S_{vb}+S_{bv})-2(\mu_{b}^{2}-\mu_{a}^{2})S_{vv}].$ Denote the bit-flip errors in $C_{\Delta^{+}}$ ($C_{\Delta^{-}}$) as the effective events when the right (left) detector clicks and its total number as $n_{\Delta^{+}}^{R}$ ($n_{\Delta^{-}}^{L}$). The bit-flip error rate in $C_{\Delta}=C_{\Delta^{+}}\bigcup C_{\Delta^{-}}$ can be shown as $\displaystyle T_{\Delta}=\frac{n_{\Delta^{+}}^{R}+n_{\Delta^{-}}^{L}}{N_{\Delta^{+}}+N_{\Delta^{-}}}.$ (12) Therefore $\overline{e}_{1}^{ph}$ can be estimated with decoy-state method as [12, 18] $\displaystyle\overline{e}_{1}^{ph}\leq\frac{T_{\Delta}-1/2e^{-2\mu_{a}}S_{vv}}{2\mu_{a}e^{-2\mu_{a}}\underline{s}_{1}}.$ (13) (4) Key rate formula. With these quantities, the final key length can be expressed as [68, 12] $\displaystyle R=$ $\displaystyle 2p_{z0}(1-p_{z0})\mu_{z}e^{-\mu_{z}}\underline{s}_{1}[1-H(\overline{e}_{1}^{ph})]$ (14) $\displaystyle-n_{t}fH(E_{Z})/N.$ where $N_{f}$ is the number of final bits, $H(x)=-x{\rm log}_{2}x-(1-x){\rm log}_{2}(1-x)$ is the binary entropy function, and $f$ is the error correction efficiency factor. (5) AOPP method. AOPP method [20, 21] is a pre-error correction process on raw strings $Z_{A}$ and $Z_{B}$ proposed to improve the direct tranmission key rate. In AOPP method, Bob randomly select two unequal bits as pairs and will obtain $n_{p}={\rm min}(n_{t0},n_{t1})$ pairs, where $n_{t0}$ ($n_{t1}$) is the number of bits 0 (1) in raw string $Z_{B}$. There will be only two types of pairs can be survived when Alice make exactly the same or opposite decision as Bob for two bits, and denote the number as $n_{vd}$ or $n_{cc}$, respectively. Therefore, the bit error after AOPP is shown as $\displaystyle E_{Z}^{\prime}=\frac{n_{vd}}{n_{cc}+n_{vd}}.$ (15) The lower bound of the number of untagged bits is $\displaystyle\underline{n}_{1}^{\prime}=n_{p}\frac{\underline{n}_{1}^{0}}{n_{t0}}\frac{\underline{n}_{1}^{1}}{n_{t1}},$ (16) where $\underline{n}_{1}^{0}$ and $\underline{n}_{1}^{1}$ is the lower bound of untagged bits when they make the opposite decision and obtain bits 0 and 1, correspondingly. And the phase-flip error rate changed into ${\overline{e}^{\prime}}_{1}^{ph}=2\overline{e}_{1}^{ph}(1-\overline{e}_{1}^{ph})$. Besides, finite-key effects should be considered in practical systems using Chernoff bound [69, 70] and the parameters can be estimated as $n_{1}^{\prime}=\varphi^{L}(\underline{n}_{1}^{\prime})$ and $e_{1}^{\prime ph}=\varphi^{U}(\underline{n}_{1}^{\prime}\overline{e}_{1}^{\prime ph})/\underline{n}_{1}^{\prime}$. Finally, the improved key length can be shown as [21, 20, 68] $\displaystyle N_{f}^{\prime}=$ $\displaystyle n_{1}^{\prime}[1-H({e^{\prime}}_{1}^{ph})]-n_{t}^{\prime}fH(E_{Z}^{\prime})-{\rm log}_{2}\frac{2}{\varepsilon_{cor}}$ (17) $\displaystyle-2{\rm log}_{2}\frac{1}{\sqrt{2}\varepsilon_{PA}\hat{\varepsilon}}.$ ## Appendix B Details of the upper bound of secure key rate To obtain the upper bound of secure key rate, we should consider the ideal situation, i.e. the upper bound of the number of secure single photons and the lower bound of phase-flip error rate without finite-key effects. Ideally, the single photons which could be used to distill secure key bits (i.e. the upper bound of secure untagged bits) after AOPP can be shown as $\displaystyle n_{1,{\rm sec}}=n_{p}\frac{n_{1s}^{0}}{n_{t0}}\frac{n_{1s}^{1}}{n_{t1}},$ (18) where $\displaystyle n_{1s}^{0}=n_{1s}^{1}=Np_{z}^{2}p_{z0}(1-p_{z0})p_{zk}u_{z}\eta_{zk}e^{-u_{z}\eta_{zk}}.$ (19) In the rest, we will analyze this inherent phase-flip error rate from the perspective of virtual protocol. In the virtual protocol [12], Alice and Bob will prepare an extended state $\displaystyle|\varPsi\rangle=\frac{1}{\sqrt{2}}(e^{i\delta_{B}}|01\rangle\otimes|01\rangle+e^{i\delta_{A}}|10\rangle\otimes|10\rangle),$ (20) with restriction of Eq. (9) or (10) which is equivalent to the state $[|01\rangle\langle 01|\otimes|01\rangle\langle 01|+|10\rangle\langle 10|\otimes|10\rangle\langle 10|]/2$ when Alice and Bob measure ancillary photons in photon-number basis in advance. Those states left of $\otimes$ are real states which will be sent to Charlie, while the right are local ancillary states with bit value encoded. Local state $|0\rangle$ corresponds to a bit 0 (1), and state $|1\rangle$ corresponds to a bit 1 (0) for Alice (Bob). In order to obtain lower bound of phase-flip error rate, consider the ideal situation in which the phase shift can be compensated perfectly which is equivalent to no phase shift. After interference and detection, the local states change into $\displaystyle\rho_{l}=\frac{1}{2}[|\varphi_{1}\rangle\langle\varphi_{1}|+|\varphi_{2}\rangle\langle\varphi_{2}|],$ (21) where $\displaystyle|\varphi_{1}\rangle$ $\displaystyle=[|01\rangle+e^{i\delta}|10\rangle]/\sqrt{2},$ $\displaystyle|\varphi_{2}\rangle$ $\displaystyle=[|01\rangle-e^{i\delta}|10\rangle]/\sqrt{2},$ (22) with $\delta=\delta_{A}-\delta_{B}$. When the local ancillary states measured virtually with basis $|\Phi^{0}\rangle=[|01\rangle+|10\rangle]/\sqrt{2}$ and $|\Phi^{1}\rangle=[|01\rangle-|10\rangle]/\sqrt{2}$, the phase-flip error rate before AOPP can be shown as $\displaystyle e_{\rm in}^{\prime}=\frac{e_{\rm in}^{0}+e_{\rm in}^{1}}{2},$ (23) where $e_{\rm in}^{0}$ and $e_{\rm in}^{1}$ are phase-flip error rate when $\delta_{A}$ and $\delta_{B}$ satisfy Eq. (9) and (10), respectively, $\displaystyle e_{\rm in}^{0}$ $\displaystyle={\rm Tr}[|\Phi^{1}\rangle\langle\Phi^{1}|\rho_{l}],$ $\displaystyle e_{\rm in}^{1}$ $\displaystyle={\rm Tr}[|\Phi^{0}\rangle\langle\Phi^{0}|\rho_{l}].$ (24) On average, the inherent phase-flip error will be $\overline{e}_{\rm in}^{\prime}=\int_{-\pi/M}^{\pi/M}e_{\rm in}^{\prime}dM\delta/2\pi$ and is approximately equal to 0.0032 when $M=16$. Thus, we can obtain the inherent phase-flip error rate with AOPP method as $e_{\rm in}=2\overline{e}_{\rm in}^{\prime}(1-\overline{e}_{\rm in}^{\prime})$. ## Appendix C Details of numerical simulations Under passive frequency shift attack, the parameters obtained by statistic can be shown as follows $\displaystyle n_{\rm sig}=$ $\displaystyle 4Np_{z}^{2}p_{z0}p_{z1}\sum_{\gamma\in M}p_{z\gamma}\bigl{[}\overline{p}_{d}e^{-\frac{\mu_{z\gamma}}{2}}-\overline{p}_{d}^{2}e^{-\mu_{z\gamma}}\bigr{]},$ (25) $\displaystyle n_{\rm err}=$ $\displaystyle 2Np_{z}^{2}\Big{[}p_{z1}^{2}\sum_{\gamma_{1},\gamma_{2}\in M}p_{z\gamma_{1}}p_{z\gamma_{2}}\bigl{[}-\overline{p}_{d}^{2}e^{-(\mu_{z\gamma_{1}}+\mu_{z\gamma_{2}})}$ (26) $\displaystyle+\overline{p}_{d}e^{-\frac{\mu_{z\gamma_{1}}+\mu_{z\gamma_{2}}}{2}}I_{0}(\sqrt{\mu_{z\gamma_{1}}\mu_{z\gamma_{2}}})\bigr{]}+p_{z0}^{2}p_{d}\overline{p}_{d}\Bigr{]},$ $\displaystyle n_{\alpha v}=$ $\displaystyle 2N_{\alpha v}\sum_{\gamma\in M}p_{\alpha\gamma}[\overline{p}_{d}e^{-\mu_{\alpha\gamma}/2}-\overline{p}_{d}^{2}e^{-\mu_{\alpha\gamma}}],$ (27) $\displaystyle n_{v\beta}=$ $\displaystyle 2N_{v\beta}\sum_{\gamma\in M}p_{\beta\gamma}[\overline{p}_{d}e^{-\mu_{\beta\gamma}/2}-\overline{p}_{d}^{2}e^{-\mu_{\beta\gamma}}],$ (28) $\displaystyle n_{vv}=$ $\displaystyle 2N_{vv}p_{d}\overline{p}_{d},$ (29) where $p_{d}$ is the dark count rate and $\overline{p}_{d}=1-p_{d}$, and $\mu_{\beta\gamma}=\mu_{\beta}\eta_{\beta\gamma}$. Here, $\beta\in\\{a,b,z\\}$ and $\gamma\in M$. Note that the intensities of state $|e^{i\theta_{A}}\sqrt{\mu_{a}\eta_{a\gamma_{1}}}\rangle$ and $|e^{i\theta_{B}}\sqrt{\mu_{a}\eta_{a\gamma_{2}}}\rangle$ from Alice and Bob in $X_{a}$ windows may be different after passive frequency shift attack where $\gamma_{1}$, $\gamma_{2}\in M$, but this does not mean it could not cause right detection. After interference, the intensity of left and right detectors will be $\displaystyle\mu_{l}(\gamma_{1},\gamma_{2})=$ $\displaystyle\frac{1}{2}\Bigl{[}\mu_{a\gamma_{1}}+\mu_{a\gamma_{2}}+2\sqrt{\mu_{a\gamma_{1}}\mu_{a\gamma_{2}}}{\rm cos}\delta\Bigr{]},$ (30) $\displaystyle\mu_{r}(\gamma_{1},\gamma_{2})=$ $\displaystyle\frac{1}{2}\Bigl{[}\mu_{a\gamma_{1}}+\mu_{a\gamma_{2}}-2\sqrt{\mu_{a\gamma_{1}}\mu_{a\gamma_{2}}}{\rm cos}\delta\Bigr{]},$ (31) where $\delta=\theta_{B}-\theta_{A}$. We can see that the difference of two output intensities does not determined by the distinction of two input intensities but phase difference. Then the number of error events in $C_{\Delta^{\pm}}$ can be shown as $\displaystyle n_{\Delta^{+}}^{R}$ $\displaystyle=N_{\Delta^{+}}\sum_{\gamma_{1},\gamma_{2}\in W}p_{a\gamma_{1}}p_{a\gamma_{2}}\Big{[}-\overline{p}_{d}^{2}e^{-\mu_{a\gamma_{1}}-\mu_{a\gamma_{2}}}$ (32) $\displaystyle+\overline{p}_{d}\int_{-\Delta/2}^{\Delta/2}e_{d}e^{-\mu_{r}(\gamma_{1},\gamma_{2})}+\overline{e}_{d}e^{-\mu_{l}(\gamma_{1},\gamma_{2})}d\frac{\delta}{\Delta}\Bigr{]},$ where $e_{d}$ is the misalignment-error probability and $\overline{e}_{d}=1-e_{d}$. Similarly, we can obtain $n_{\Delta^{-}}^{L}$. ## References * Bennett and Brassard [2014] C. H. Bennett and G. Brassard, Quantum cryptography: Public key distribution and coin tossing, Theor. Comput. Sci. 560, 7 (2014). * Shor and Preskill [2000] P. W. Shor and J. Preskill, Simple proof of security of the bb84 quantum key distribution protocol, Phys. Rev. Lett. 85, 441 (2000). * Xu _et al._ [2020a] F. Xu, X. Ma, Q. Zhang, H.-K. Lo, and J.-W. Pan, Secure quantum key distribution with realistic devices, Rev. Mod. Phys. 92, 025002 (2020a). * Lo _et al._ [2012] H. K. Lo, M. Curty, and B. Qi, Measurement-device-independent quantum key distribution, Phys. Rev. Lett. 108, 130503 (2012). * Brassard _et al._ [2000] G. Brassard, N. Lütkenhaus, T. Mor, and B. C. Sanders, Limitations on practical quantum cryptography, Phys. Rev. Lett. 85, 1330 (2000). * Lütkenhaus and Jahma [2002] N. Lütkenhaus and M. Jahma, Quantum key distribution with realistic states: photon-number statistics in the photon-number splitting attack, New J. Phys. 4, 44 (2002). * Hwang [2003] W. Y. Hwang, Quantum key distribution with high loss: toward global secure communication, Phys. Rev. Lett. 91, 057901 (2003). * Lo _et al._ [2005] H. K. Lo, X. Ma, and K. Chen, Decoy state quantum key distribution, Phys. Rev. Lett. 94, 230504 (2005). * Takeoka _et al._ [2014] M. Takeoka, S. Guha, and M. M. Wilde, Fundamental rate-loss tradeoff for optical quantum key distribution, Nat. Commun. 5, 5235 (2014). * Pirandola _et al._ [2017] S. Pirandola, R. Laurenza, C. Ottaviani, and L. Banchi, Fundamental limits of repeaterless quantum communications, Nat. Commun. 8, 15043 (2017). * Lucamarini _et al._ [2018] M. Lucamarini, Z. L. Yuan, J. F. Dynes, and A. J. Shields, Overcoming the rate-distance limit of quantum key distribution without quantum repeaters, Nature 557, 400 (2018). * Wang _et al._ [2018] X.-B. Wang, Z.-W. Yu, and X.-L. Hu, Twin-field quantum key distribution with large misalignment error, Phys. Rev. A 98, 062323 (2018). * Ma _et al._ [2018] X. Ma, P. Zeng, and H. Zhou, Phase-matching quantum key distribution, Phys. Rev. X 8, 031043 (2018). * Curty _et al._ [2019] M. Curty, K. Azuma, and H.-K. Lo, Simple security proof of twin-field type quantum key distribution protocol, npj Quant. Inf. 5, 64 (2019). * Cui _et al._ [2019] C. Cui, Z.-Q. Yin, R. Wang, W. Chen, S. Wang, G.-C. Guo, and Z.-F. Han, Twin-field quantum key distribution without phase postselection, Phys. Rev. Appl. 11, 034053 (2019). * Lin and Lütkenhaus [2018] J. Lin and N. Lütkenhaus, Simple security analysis of phase-matching measurement-device-independent quantum key distribution, Phys. Rev. A 98, 042332 (2018). * Tamaki _et al._ [2018] K. Tamaki, H. Lo, W. Wang, and M. Lucamarini, Information theoretic security of quantum key distribution overcoming the repeaterless secret key capacity bound, arXiv: 1805.05511v1 (2018). * Yu _et al._ [2019] Z. W. Yu, X. L. Hu, C. Jiang, H. Xu, and X. B. Wang, Sending-or-not-sending twin-field quantum key distribution in practice, Sci. Rep. 9, 3080 (2019). * Jiang _et al._ [2019] C. Jiang, Z.-W. Yu, X.-L. Hu, and X.-B. Wang, Unconditional security of sending or not sending twin-field quantum key distribution with finite pulses, Phys. Rev. Appl. 12, 024061 (2019). * Xu _et al._ [2020b] H. Xu, Z.-W. Yu, C. Jiang, X.-L. Hu, and X.-B. Wang, Sending-or-not-sending twin-field quantum key distribution: Breaking the direct transmission key rate, Phys. Rev. A 101, 042330 (2020b). * Jiang _et al._ [2020] C. Jiang, X.-L. Hu, H. Xu, Z.-W. Yu, and X.-B. Wang, Zigzag approach to higher key rate of sending-or-not-sending twin field quantum key distribution with finite-key effects, New J. Phys. 22, 053048 (2020). * Hu _et al._ [2019] X.-L. Hu, C. Jiang, Z.-W. Yu, and X.-B. Wang, Sending-or-not-sending twin-field protocol for quantum key distribution with asymmetric source parameters, Phys. Rev. A 100, 062337 (2019). * Zhou _et al._ [2019] X.-Y. Zhou, C.-H. Zhang, C.-M. Zhang, and Q. Wang, Asymmetric sending or not sending twin-field quantum key distribution in practice, Phys. Rev. A 99, 062316 (2019). * Yin and Fu [2019] H.-L. Yin and Y. Fu, Measurement-device-independent twin-field quantum key distribution, Sci. Rep. 9, 3045 (2019). * Zeng _et al._ [2020] P. Zeng, W. Wu, and X. Ma, Symmetry-protected privacy: Beating the rate-distance linear bound over a noisy channel, Phys. Rev. Appl. 13, 064013 (2020). * Currás-Lorenzo _et al._ [2021] G. Currás-Lorenzo, L. Wooltorton, and M. Razavi, Twin-field quantum key distribution with fully discrete phase randomization, Phys. Rev. Appl. 15, 014016 (2021). * Zhang _et al._ [2020] C.-M. Zhang, Y.-W. Xu, R. Wang, and Q. Wang, Twin-field quantum key distribution with discrete-phase-randomized sources, Phys. Rev. Appl. 14, 064070 (2020). * Grasselli and Curty [2019] F. Grasselli and M. Curty, Practical decoy-state method for twin-field quantum key distribution, New J. Phys. 21, 073001 (2019). * Teng _et al._ [2020] J. Teng, F.-Y. Lu, Z.-Q. Yin, G.-J. Fan-Yuan, R. Wang, S. Wang, W. Chen, W. Huang, B.-J. Xu, G.-C. Guo, and Z.-F. Han, Twin-field quantum key distribution with passive-decoy state, New J. Phys. 22, 103017 (2020). * Lu _et al._ [2019] F.-Y. Lu, Z.-Q. Yin, R. Wang, G.-J. Fan-Yuan, S. Wang, D.-Y. He, W. Chen, W. Huang, B.-J. Xu, G.-C. Guo, and Z.-F. Han, Practical issues of twin-field quantum key distribution, New J. Phys. 21, 123030 (2019). * Wang _et al._ [2020] R. Wang, Z.-Q. Yin, F.-Y. Lu, S. Wang, W. Chen, C.-M. Zhang, W. Huang, B.-J. Xu, G.-C. Guo, and Z.-F. Han, Optimized protocol for twin-field quantum key distribution, Commun. Phys. 3, 149 (2020). * Wang and Lo [2020] W. Wang and H.-K. Lo, Simple method for asymmetric twin-field quantum key distribution, New J. Phys. 22, 013020 (2020). * Lorenzo _et al._ [2019] G. C. Lorenzo, Á. Navarrete, K. Azuma, G. Kato, M. Curty, and M. Razavi, Tight finite-key security for twin-field quantum key distribution, arXiv: 1910.11407v4 (2019). * Mao _et al._ [2021] Y. Mao, P. Zeng, and T.-Y. Chen, Recent advances on quantum key distribution overcoming the linear secret key capacity bound, Adv. Quantum Technol. 4, 2000084 (2021). * Minder _et al._ [2019] M. Minder, M. Pittaluga, G. L. Roberts, M. Lucamarini, J. F. Dynes, Z. L. Yuan, and A. J. Shields, Experimental quantum key distribution beyond the repeaterless secret key capacity, Nat. Photon. 13, 334 (2019). * Wang _et al._ [2019] S. Wang, D.-Y. He, Z.-Q. Yin, F.-Y. Lu, C.-H. Cui, W. Chen, Z. Zhou, G.-C. Guo, and Z.-F. Han, Beating the fundamental rate-distance limit in a proof-of-principle quantum key distribution system, Phys. Rev. X 9, 021046 (2019). * Zhong _et al._ [2019] X. Zhong, J. Hu, M. Curty, L. Qian, and H. K. Lo, Proof-of-principle experimental demonstration of twin-field type quantum key distribution, Phys. Rev. Lett. 123, 100506 (2019). * Liu _et al._ [2019] Y. Liu, Z. W. Yu, W. Zhang, J. Y. Guan, J. P. Chen, C. Zhang, X. L. Hu, H. Li, C. Jiang, J. Lin, T. Y. Chen, L. You, Z. Wang, X. B. Wang, Q. Zhang, and J. W. Pan, Experimental twin-field quantum key distribution through sending or not sending, Phys. Rev. Lett. 123, 100505 (2019). * Chen _et al._ [2020] J. P. Chen, C. Zhang, Y. Liu, C. Jiang, W. Zhang, X. L. Hu, J. Y. Guan, Z. W. Yu, H. Xu, J. Lin, M. J. Li, H. Chen, H. Li, L. You, Z. Wang, X. B. Wang, Q. Zhang, and J. W. Pan, Sending-or-not-sending with independent lasers: Secure twin-field quantum key distribution over 509 km, Phys. Rev. Lett. 124, 070501 (2020). * Fang _et al._ [2020] X.-T. Fang, P. Zeng, H. Liu, M. Zou, W. Wu, Y.-L. Tang, Y.-J. Sheng, Y. Xiang, W. Zhang, H. Li, Z. Wang, L. You, M.-J. Li, H. Chen, Y.-A. Chen, Q. Zhang, C.-Z. Peng, X. Ma, T.-Y. Chen, and J.-W. Pan, Implementation of quantum key distribution surpassing the linear rate-transmittance bound, Nat. Photon. 14, 422 (2020). * Liu _et al._ [2021] H. Liu, C. Jiang, H.-T. Zhu, M. Zou, Z. Yu, X.-L. Hu, H. Xu, S. Ma, Z. Han, J. Chen, Y. Dai, S.-B. Tang, W. Zhang, H. Li, L. You, Z. Wang, F. Zhou, Q. Zhang, X.-b. Wang, T.-Y. Chen, and J. Pan, Field test of twin-field quantum key distribution through sending-or-not-sending over 428 km, arXiv: 2101.00276v1 (2021). * Chen _et al._ [2021] J.-P. Chen, C. Zhang, Y. Liu, C. Jiang, W.-J. Zhang, Z.-Y. Han, S.-Z. Ma, X.-L. Hu, Y.-H. Li, H. Liu, F. Zhou, H.-F. Jiang, T.-Y. Chen, H. Li, L.-X. You, Z. Wang, X.-B. Wang, Q. Zhang, and J.-W. Pan, Twin-field quantum key distribution over 511 km optical fiber linking two distant metropolitans, arXiv:2102.00433v1 (2021). * Vakhitov _et al._ [2001] A. Vakhitov, V. Makarov, and D. R. Hjelme, Large pulse attack as a method of conventional optical eavesdropping in quantum cryptography, J. Mod. Opt. 48, 2023 (2001). * Gisin _et al._ [2006] N. Gisin, S. Fasel, B. Kraus, H. Zbinden, and G. Ribordy, Trojan-horse attacks on quantum-key-distribution systems, Phys. Rev. A 73, 022320 (2006). * Fung _et al._ [2007] C.-H. F. Fung, B. Qi, K. Tamaki, and H.-K. Lo, Phase-remapping attack in practical quantum-key-distribution systems, Phys. Rev. A 75, 032314 (2007). * Jiang _et al._ [2012] M.-S. Jiang, S.-H. Sun, C.-Y. Li, and L.-M. Liang, Wavelength-selected photon-number-splitting attack against plug-and-play quantum key distribution systems with decoy states, Phys. Rev. A 86, 032310 (2012). * Jiang _et al._ [2014] M.-S. Jiang, S.-H. Sun, C.-Y. Li, and L.-M. Liang, Frequency shift attack on ‘plug-and-play’ quantum key distribution systems, J. Mod. Opt. 61, 147 (2014). * Jain _et al._ [2014] N. Jain, E. Anisimova, I. Khan, V. Makarov, C. Marquardt, and G. Leuchs, Trojan-horse attacks threaten the security of practical quantum cryptography, New J. Phys. 16, 123030 (2014). * Lucamarini _et al._ [2015] M. Lucamarini, I. Choi, M. B. Ward, J. F. Dynes, Z. L. Yuan, and A. J. Shields, Practical security bounds against the trojan-horse attack in quantum key distribution, Phys. Rev. X 5, 031030 (2015). * Bugge _et al._ [2014] A. N. Bugge, S. Sauge, A. M. M. Ghazali, J. Skaar, L. Lydersen, and V. Makarov, Laser damage helps the eavesdropper in quantum cryptography, Phys. Rev. Lett. 112, 070503 (2014). * Sun _et al._ [2015] S.-H. Sun, F. Xu, M.-S. Jiang, X.-C. Ma, H.-K. Lo, and L.-M. Liang, Effect of source tampering in the security of quantum cryptography, Phys. Rev. A 92, 022304 (2015). * Makarov _et al._ [2016] V. Makarov, J.-P. Bourgoin, P. Chaiwongkhot, M. Gagné, T. Jennewein, S. Kaiser, R. Kashyap, M. Legré, C. Minshull, and S. Sajeed, Creation of backdoors in quantum communications via laser damage, Phys. Rev. A 94, 030302 (2016). * Huang _et al._ [2019] A. Huang, l. Navarrete, S.-H. Sun, P. Chaiwongkhot, M. Curty, and V. Makarov, Laser-seeding attack in quantum key distribution, Phys. Rev. Appl. 12, 064043 (2019). * Huang _et al._ [2020] A. Huang, R. Li, V. Egorov, S. Tchouragoulov, K. Kumar, and V. Makarov, Laser-damage attack against optical attenuators in quantum key distribution, Phys. Rev. Appl. 13, 034017 (2020). * Pang _et al._ [2020] X.-L. Pang, A.-L. Yang, C.-N. Zhang, J.-P. Dou, H. Li, J. Gao, and X.-M. Jin, Hacking quantum key distribution via injection locking, Phys. Rev. Appl. 13, 034008 (2020). * Tang _et al._ [2013] Y.-L. Tang, H.-L. Yin, X. Ma, C.-H. F. Fung, Y. Liu, H.-L. Yong, T.-Y. Chen, C.-Z. Peng, Z.-B. Chen, and J.-W. Pan, Source attack of decoy-state quantum key distribution using phase information, Phys. Rev. A 88, 022308 (2013). * Tamaki _et al._ [2016] K. Tamaki, M. Curty, and M. Lucamarini, Decoy-state quantum key distribution with a leaky source, New J. Phys. 18, 065008 (2016). * Huang _et al._ [2018] A. Huang, S.-H. Sun, Z. Liu, and V. Makarov, Quantum key distribution with distinguishable decoy states, Phys. Rev. A 98, 012330 (2018). * Sajeed _et al._ [2015] S. Sajeed, I. Radchenko, S. Kaiser, J.-P. Bourgoin, A. Pappa, L. Monat, M. Legré, and V. Makarov, Attacks exploiting deviation of mean photon number in quantum key distribution and coin tossing, Phys. Rev. A 91, 032326 (2015). * Winzer and Essiambre [2006] P. J. Winzer and R. Essiambre, Advanced optical modulation formats, Proc. IEEE 94, 952 (2006). * Tamaki _et al._ [2014] K. Tamaki, M. Curty, G. Kato, H.-K. Lo, and K. Azuma, Loss-tolerant quantum cryptography with imperfect sources, Phys. Rev. A 90, 052314 (2014). * Pereira _et al._ [2019] M. Pereira, M. Curty, and K. Tamaki, Quantum key distribution with flawed and leaky sources, npj Quant. Inf. 5, 62 (2019). * Mizutani _et al._ [2019] A. Mizutani, T. Sasaki, Y. Takeuchi, K. Tamaki, and M. Koashi, Quantum key distribution with simply characterized light sources, npj Quant. Inf. 5, 87 (2019). * Navarrete _et al._ [2020] A. Navarrete, M. Pereira, M. Curty, and K. Tamaki, Practical quantum key distribution secure against side-channels, arXiv: 2007.03364v1 (2020). * Pereira _et al._ [2020] M. Pereira, G. Kato, A. Mizutani, M. Curty, and K. Tamaki, Quantum key distribution with correlated sources, Sci. Adv. 6, 4487 (2020). * Inamori _et al._ [2007] H. Inamori, N. Lütkenhaus, and D. Mayers, Unconditional security of practical quantum key distribution, Eur. Phys. J. D 41, 599 (2007). * Yu _et al._ [2013] Z.-W. Yu, Y.-H. Zhou, and X.-B. Wang, Three-intensity decoy-state method for measurement-device-independent quantum key distribution, Phys. Rev. A 88, 062339 (2013). * Tomamichel _et al._ [2012] M. Tomamichel, C. C. W. Lim, N. Gisin, and R. Renner, Tight finite-key analysis for quantum cryptography, Nat. Commun. 3, 634 (2012). * Chernoff [1952] H. Chernoff, A measure of asymptotic efficiency for tests of a hypothesis based on the sum of observations, Ann. Math. Stat. 23, 493 (1952). * Curty _et al._ [2014] M. Curty, F. Xu, W. Cui, C. C. Lim, K. Tamaki, and H. K. Lo, Finite-key analysis for measurement-device-independent quantum key distribution, Nat. Commun. 5, 3732 (2014).
# Ghost distributions on supersymmetric spaces I: Koszul induced superspaces, branching, and the full ghost centre Alexander Sherman ###### Abstract. Given a Lie superalgebra $\mathfrak{g}$, Gorelik defined the anticentre $\mathcal{A}$ of its enveloping algebra, which consists of certain elements that square to the center. We seek to generalize and enrich the anticentre to the context of supersymmetric pairs $(\mathfrak{g},\mathfrak{k})$, or more generally supersymmetric spaces $G/K$. We define certain invariant distributions on $G/K$, which we call ghost distributions, and which in some sense are induced from invariant distributions on $G_{0}/K_{0}$. Ghost distributions, and in particular their Harish-Chandra polynomials, give information about branching from $G$ to a symmetric subgroup $K^{\prime}$ which is related (and sometimes conjugate) to $K$. We discuss the case of $G\times G/G$ for an arbitrary quasireductive supergroup $G$, where our results prove the existence of a polynomial which determines projectivity of irreducible $G$-modules. Finally, a generalization of Gorelik’s ghost centre is defined called the full ghost centre, $\mathcal{Z}_{full}$. For type I basic Lie superalgebras $\mathfrak{g}$ we fully describe $\mathcal{Z}_{full}$, and prove that if $\mathfrak{g}$ contains an internal grading operator, $\mathcal{Z}_{full}$ consists exactly of those elements in $\mathcal{U}\mathfrak{g}$ acting by $\mathbb{Z}$-graded constants on every finite-dimensional irreducible representation. ## 1\. Introduction Let $\mathfrak{g}$ be a Lie superalgebra over an algebraically closed field $k$ of characteristic zero. In [Gor00], Gorelik defined a certain natural ’twisted’ adjoint action of a Lie superalgebra $\mathfrak{g}$ on its enveloping algebra $\mathcal{U}\mathfrak{g}$. The action was originally considered by [ABF97] for $\mathfrak{o}\mathfrak{s}\mathfrak{p}(1|2n)$, where a certain element $T$ in the enveloping algebra was constructed, called Casimir’s ghost, which squared to the center. The action defined by Gorelik in general is remarkable in that the structure of $\mathcal{U}\mathfrak{g}$ becomes that of an induced module, $\operatorname{Ind}_{\mathfrak{g}_{\overline{0}}}^{\mathfrak{g}}\mathcal{U}\mathfrak{g}_{\overline{0}}$. Further the invariants of this action, denoted $\mathcal{A}$ and called the anticentre in [Gor00], are both a module over the center $\mathcal{Z}:=\mathcal{Z}(\mathcal{U}\mathfrak{g})$ and multiply into the center, so that $\tilde{\mathcal{Z}}:=\mathcal{Z}+\mathcal{A}$ is an algebra which Gorelik called the ghost centre of $\mathfrak{g}$. If $\mathfrak{g}$ is quasireductive, i.e. $\mathfrak{g}_{\overline{0}}$ is reductive and acts semisimply on $\mathfrak{g}$, and $\Lambda^{top}\mathfrak{g}_{\overline{1}}$ is a trivial $\mathfrak{g}_{\overline{0}}$-module, Gorelik obtained an explicit identification as vector spaces of $\mathcal{A}$ with the center of $\mathcal{U}\mathfrak{g}_{\overline{0}}$. Further, for basic classical Lie superalgebras, Gorelik gave a complete description of the Harish-Chandra image of $\mathcal{A}$ and the structure of $\tilde{\mathcal{Z}}$. In particular she showed that $\tilde{\mathcal{Z}}$ consists of exactly those elements of $\mathcal{U}\mathfrak{g}$ that act by superconstants on every irreducible representation. We also note that Gorelik fully computed the $\mathcal{A}$ for $Q$-type superalgebras. ### 1.1. Generalization to supersymmetric spaces We seek to understand Gorelik’s original results from the geometric perspective, and thereby understand how they may be generalized. Consider the setting of symmetric supervarieties (or supersymmetric spaces, if one prefers that term) $G/K$ and their corresponding supersymmetric pairs $(\mathfrak{g},\mathfrak{k})$, corresponding to an involution $\theta$. We have a decomposition $\mathfrak{g}=\mathfrak{k}\oplus\mathfrak{p}$ into the $\pm 1$ eigenspaces of $\theta$. We define a new subalgebra $\mathfrak{k}^{\prime}:=\mathfrak{k}_{\overline{0}}\oplus\mathfrak{p}_{\overline{1}},$ which itself is the fixed points of the involution $\delta\circ\theta$, where $\delta(x)=(-1)^{\overline{x}}x$ is the grading operator on $\mathfrak{g}$. Let $K^{\prime}$ be the subgroup of $G$ with $K^{\prime}_{0}=K_{0}$ and $\operatorname{Lie}(K^{\prime})=\mathfrak{k}^{\prime}$. Then because $\mathfrak{k}_{\overline{1}}^{\prime}\oplus\mathfrak{k}_{\overline{1}}=\mathfrak{g}_{\overline{1}}$, we will see that the action of $K^{\prime}$ on $G/K$ enjoys many nice properties, in particular many of which are not explicitly enjoyed by the action of $K$ on $G/K$. The first result indicating this is the following. Recall that for an affine supervariety $X$ with a closed point $x$ and maximal ideal $\mathfrak{m}_{x}\subseteq k[X]$ in the space of functions, we may consider the super vector space of distributions $\operatorname{Dist}(X,x)=\\{\psi:k[X]\to k:\psi(\mathfrak{m}_{x}^{n})=0\text{ for }n\gg 0\\}.$ If we consider the point $eK$ on $G/K$, $K_{0}^{\prime}$ fixes it, and thus $K^{\prime}$ has a natural action on $\operatorname{Dist}(G/K,eK)$. ###### Theorem 1.1. We have an isomorphism of $K^{\prime}$-modules $\operatorname{Dist}(G/K,eK)\cong\operatorname{Ind}_{\mathfrak{k}_{\overline{0}}}^{\mathfrak{k}}\operatorname{Dist}(G_{0}/K_{0},eK_{0}).$ In particular, if $K^{\prime}$ is quasireductive and $\Lambda^{top}\mathfrak{p}_{\overline{1}}$ is a trivial $K_{0}$-module, then we have an explicit isomorphism of vector spaces $\operatorname{Dist}(G/K,eK)^{K^{\prime}}\cong\operatorname{Dist}(G_{0}/K_{0},eK_{0})^{K_{0}}.$ For the symmetric supervariety $G\times G/G\cong G$, we have identifications $G^{\prime}\cong G$ and $\operatorname{Dist}(G,eG)\cong\mathcal{U}\mathfrak{g}$, and the action we obtain in this case of $G^{\prime}$ on $\operatorname{Dist}(G,eG)$ is exactly Gorelik’s twisted adjoint action, thus reproducing her results from this context. The proof of 1.1 uses a construction due to Koszul ([Kos82]) which allows one to take the variety $G_{0}/K_{0}$, which has an action of $K_{0}^{\prime}=K_{0}$, and induce it to a $K^{\prime}$-supervariety denoted $(G_{0}/K_{0})^{\mathfrak{k}^{\prime}}$. The latter supervariety has the special property that the vector fields $\mathfrak{k}_{\overline{1}}$ are everywhere non-vanishing, and its algebra of functions and spaces of distributions, respectively, are (co)induced from $G_{0}/K_{0}$. Then by a general result, $G/K$ and $(G_{0}/K_{0})^{\mathfrak{k}}$ are locally isomorphic as $K^{\prime}$-supervarieties, implying an isomorphism of their spaces of distributions as in 1.1. ### 1.2. The Harish-Chandra morphism Now consider a symmetric supervariety $G/K$ where we assume $G$ is quasireductive and Cartan-even (i.e. its Lie superalgebra has an even Cartan subalgebra), $\Lambda^{top}\mathfrak{p}_{\overline{1}}$ is the trivial $K_{0}$-module, and we have an Iwasawa decomposition $\mathfrak{g}=\mathfrak{k}\oplus\mathfrak{a}\oplus\mathfrak{n}$. We write $\mathcal{A}_{G/K}:=\operatorname{Dist}(G/K,eK)^{K^{\prime}},$ for the $K^{\prime}$-invariant distributions, and call elements of $\mathcal{A}_{G/K}$ ghost distributions on $G/K$. One important problem is to compute the image of $\mathcal{A}_{G/K}$ under the Harish-Chandra homomorphism $HC:\operatorname{Dist}(G/K,eK)\to S(\mathfrak{a}).$ In this case $G/K$ is also spherical, and admits rational functions $f_{\lambda}$ indexed by a full rank lattice $\Lambda\subseteq\mathfrak{a}^{*}$ which are $\mathfrak{n}$-invariant and eigenfunctions for $\mathfrak{a}$, where we normalize by requiring that $f_{\lambda}(eK)=1$. If $\lambda$ is a dominant weight, then generically $f_{\lambda}$ will be a regular function. Then for a distribution $\psi$, $HC(\psi)(\lambda)=\psi(f_{\lambda}).$ Using this, we have the following representation-theoretic interpretation of the values of $\mathcal{A}_{G/K}$ on such eigenfunctions. ###### Theorem 1.2. Let $f_{\lambda}\in k[G/K]$ be a highest weight vector of weight $\lambda$. Then the $K^{\prime}$-module generated by $f_{\lambda}$ contains a copy of $I_{K^{\prime}}(k)$ if and only if there exists $\gamma\in\mathcal{A}_{G/K}$ such that $HC(\gamma)(\lambda)\neq 0$. Here $I_{K^{\prime}}(k)$ denote the injective indecomposable $K^{\prime}$-module with socle $k$. From this we obtain, as one corollary: ###### Corollary 1.3. Keep the hypotheses of 1.2, and suppose further that the $G$-module generated by $f_{\lambda}$, $L$, is irreducible. Then $I_{G}(L)$ is a submodule of $k[G/K^{\prime}]$. Thus obtaining $HC(\mathcal{A}_{G/K})$ is of interest, and this will be taken up in more detail for the case of basic classical Lie superalgebras in a subsequent article. However even for such Lie superalgebras the answer is not known in general (outside the case $G\times G/G$, originally done by Gorelik). ### 1.3. The case of $G\times G/G$ For $G\times G/G$, we have the following nice application of 1.3. Note that in the following we remove the restriction that $G$ be Cartan-even. For this one needs to generalize the Harish-Chandra morphism to the non-Cartan-even case, which is done in the appendix. ###### Theorem 1.4. Let $G$ be a quasireductive supergroup with Cartan subalgebra $\mathfrak{h}\subseteq\mathfrak{g}$, and choose a Borel subgroup $B$ with Lie superalgebra $\mathfrak{b}$ containing $\mathfrak{h}$. Then there exists a polynomial $p_{G,B}\in S(\mathfrak{h})$ of degree less than or equal to $\operatorname{dim}\mathfrak{b}_{\overline{1}}$, such that for a $\mathfrak{b}$-dominant weight $\lambda$, $p_{G,B}(\lambda)\neq 0\ \text{ if and only if }\ L_{B}(\lambda)\text{ is projective},$ where $L_{B}(\lambda)$ is the irreducible $G$-module of $B$-highest weight $\lambda$. In particular the above result implies that if one simple $G$-module is projective then this is a generic property of simple $G$-modules. However it is possible that $p_{G,B}$ is the zero polynomial, so that no simple $G$-modules are projective. Although this was already know for many Lie superalgebras by direct study, this gives a general explanation for this phenomenon. ###### Remark 1.5. In every example the author is aware of, $p_{G,B}$ is a product of linear polynomials. We do not see an a priori reason for this, and it would be interesting to find an example where this does not occur (or to prove it always happens). Recall that $\mathcal{A}$ denotes the ghost center of $\mathcal{U}\mathfrak{g}$. Then $\mathcal{A}$ contains an element of least degree, which we write as $T_{\mathfrak{g}}$. This operator can test whether a given semisimple $G$-module is projective in the following sense: ###### Proposition 1.6. Let $L$ be a simple $G$-module. Then $T_{\mathfrak{g}}$ acts by $0$ on $L$ if and only if $L$ is not projective. If $L$ is projective, then $T_{\mathfrak{g}}$ acts by one of the following automorphisms: * • if $T_{\mathfrak{g}}$ is even, then up to scalar it acts by $\delta_{L}$, the parity operator on $L$; * • if $T_{\mathfrak{g}}$ is odd, then it acts by $\delta_{L}\circ\sigma_{L}$, where $\sigma_{L}:L\to L$ is a $G$-equivariant odd automorphism of $L$. For a different application of 1.4, we can prove the following general sufficient criteria for having projective irreducible modules. ###### Theorem 1.7. Let $G$ be a Cartan-even quasireductive supergroup with a chosen Cartan subalgebra $\mathfrak{h}$ such that the following conditions hold on its Lie superalgebra: 1. (1) for a root $\alpha\in\mathfrak{h}^{*}$, $\operatorname{dim}[\mathfrak{g}_{\alpha},\mathfrak{g}_{-\alpha}]=1$; and 2. (2) for a root $\alpha\in\mathfrak{h}^{*}$, the pairing $[-,-]:\mathfrak{g}_{\alpha}\otimes\mathfrak{g}_{-\alpha}\to[\mathfrak{g}_{\alpha},\mathfrak{g}_{-\alpha}]$ is nondegenerate. Choose an arbitrary Borel subgroup $B$ whose Lie superalgebra contains $\mathfrak{h}$, and let $\alpha_{1},\dots,\alpha_{n}$ denote the odd positive roots with $r_{i}=\operatorname{dim}\mathfrak{g}_{\alpha_{i}}$. Write $h_{\alpha_{i}}$ for a nonzero element of $[\mathfrak{g}_{\alpha_{i}},\mathfrak{g}_{-\alpha_{i}}]$. Then we have (up to a scalar) $p_{G,B}=h_{\alpha_{1}}^{r_{1}}\cdots h_{\alpha_{n}}^{r_{n}}+l.o.t.$ In particular $p_{G,B}\neq 0$, so $G$ admits irreducible projective modules. Note that the above conditions hold in particular for a Cartan-even quadratic quasireductive supergroup, i.e. a Cartan-even quasireductive supergroup whose Lie superalgebra admits a nondegenerate, invariant, and even supersymmetric form. See [Ben00] for a classification of quadratic quasireductive Lie superalgebras. ### 1.4. The full ghost centre $\mathcal{Z}_{full}$ We may generalize Gorelik’s results in a different way as follows to produce an interesting subalgebra of $\mathcal{U}\mathfrak{g}$ which contains Gorelik’s ghost centre. Let $\operatorname{Aut}(\mathfrak{g},\mathfrak{g}_{\overline{0}})$ be those automorphisms of $\mathfrak{g}$ that fix $\mathfrak{g}_{\overline{0}}$ pointwise. For $\phi\in\operatorname{Aut}(\mathfrak{g},\mathfrak{g}_{\overline{0}})$, define the $\phi$-twisted adjoint action $\operatorname{ad}_{\phi}$ of $\mathfrak{g}$ on $\mathcal{U}\mathfrak{g}$ by $\operatorname{ad}_{\phi}(u)(v)=uv-(-1)^{\overline{u}\overline{v}}v\phi(u).$ When $\phi=\delta$, we obtain the twisted adjoint action studied by Gorelik. Then using 1.1 we can prove that: ###### Theorem 1.8. If $\phi(x)=x$ implies that $x\in\mathfrak{g}_{\overline{0}}$, then $\mathcal{U}\mathfrak{g}\cong\operatorname{Ind}_{\mathfrak{g}_{\overline{0}}}^{\mathfrak{g}}\mathcal{U}\mathfrak{g}_{\overline{0}}$ under the $\phi$-twisted adjoint action. Write $\mathcal{A}_{\phi}\subseteq\mathcal{U}\mathfrak{g}$ for the $\operatorname{ad}_{\phi}$-invariant elements in $\mathcal{U}\mathfrak{g}$, for any $\phi\in\operatorname{Aut}(\mathfrak{g},\mathfrak{g}_{\overline{0}})$. Then $\mathcal{A}_{\operatorname{id}}=\mathcal{Z}$, and $\mathcal{A}_{\delta}=\mathcal{A}$. Further, for $\phi,\psi\in\operatorname{Aut}(\mathfrak{g},\mathfrak{g}_{\overline{0}})$ multiplication induces a morphism $\mathcal{A}_{\phi}\otimes\mathcal{A}_{\psi}\to\mathcal{A}_{\psi\phi}.$ Therefore if we set $\mathcal{Z}_{full}:=\sum\limits_{\phi\in\operatorname{Aut}(\mathfrak{g},\mathfrak{g}_{\overline{0}})}\mathcal{A}_{\phi},$ we obtain a subalgebra of $\mathcal{U}\mathfrak{g}$, which also contains $\tilde{\mathcal{Z}}$. For $\mathfrak{g}$ one of the type I basic classical Lie superalgebras $\mathfrak{g}\mathfrak{l}(m|n)$, $\mathfrak{s}\mathfrak{l}(m|n)$, $\mathfrak{p}\mathfrak{s}\mathfrak{l}(n|n)$ with $n>2$, or $\mathfrak{o}\mathfrak{s}\mathfrak{p}(2|2n)$) we have an explicit description of this algebra. Here, $\operatorname{Aut}(\mathfrak{g},\mathfrak{g}_{\overline{0}})\cong k^{\times}$, so we write $\phi_{c}\in\operatorname{Aut}(\mathfrak{g},\mathfrak{g}_{\overline{0}})$ for the automorphism corresponding to $c\in k^{\times}$ and $\mathcal{A}_{c}$ for the invariants of the $\phi_{c}$-twisted adjoint action. Then $\phi_{c}$ satisfies the conditions of 1.8 exactly if $c\neq 1$. ###### Theorem 1.9. Let $N=\operatorname{dim}\mathfrak{g}_{\overline{1}}/2$. Then $HC(\mathcal{A}_{c})=HC(\mathcal{A}_{-1})$ for all $c\neq 1$, and $\mathcal{Z}_{full}=\bigoplus\limits_{\zeta^{N}=1}\mathcal{A}_{\zeta}.$ Further, for $\mathfrak{g}\mathfrak{l}(m|n)$, $\mathfrak{s}\mathfrak{l}(m|n)$ with $m\neq n$, and $\mathfrak{o}\mathfrak{s}\mathfrak{p}(2|2n)$, $\mathcal{Z}_{full}$ consists of exactly the set of elements in $\mathcal{U}\mathfrak{g}$ which act by $\mathbb{Z}$-graded constants on all finite-dimensional irreducible representations of $\mathfrak{g}$. We exclude $\mathfrak{s}\mathfrak{l}(n|n)$ and $\mathfrak{p}\mathfrak{s}\mathfrak{l}(n|n)$ in the last statement due to the lack of an internal grading operator. Further, $\mathfrak{p}\mathfrak{s}\mathfrak{l}(2|2)$ is fully excluded because in this case $\operatorname{Aut}(\mathfrak{g},\mathfrak{g}_{\overline{0}})=SL_{2}$ (see section 5.5 of [Mus12]). For this reason it would be interesting to compute the full ghost center for this superalgebra. ### 1.5. Future work This article is the first of two on ghost distributions and their applications. In the subsequent article we study in more detail two questions of interest: (1) when is it possible to produce an algebra using ghost distributions on a general supersymmetric space? In particular can we form an algebra of differential operators, and if not can we at least produce an algebra of polynomials using the Harish-Chandra homomorphism? (2) Computing $HC(\mathcal{A}_{G/K})$ as much as possible in the case when $\mathfrak{g}$ is an almost simple basic classical Lie superalgebra and the involution under consideration preserves the form on $\mathfrak{g}$. In this case we give some general properties of $HC(\mathcal{A}_{G/K})$, and seek to compute it for all rank one supersymmetric pairs. ### 1.6. Outline of paper In section 2 we introduce the basic algebraic supergeometry we need, in particular the algebra of differential operators and the space of distributions at a given point. Section 3 recalls basic facts about algebraic supergroups and their actions, and gives a description of invariant differential operators on homogeneous spaces. In section 4, we explain the main technical construction, the induced superspace as defined by Koszul. Section 5 applies the ideas of section 4 to homogeneous superspaces, and deduces what we need to generalize the results of [Gor00]. Section 6 studies the case of $G/G_{0}$, looking in particular at a certain invariant distribution, $v_{\mathfrak{g}}$, which will play an important role in the theory of ghost distributions. Section 7 looks at applications to a symmetric supervariety $G/K$, and section 8 studies more closely the case when an Iwasawa decomposition is present, giving the definition of the Harish-Chandra map and its interpretations. In section 9 we take a special look at the case of $G\times G/G$, where the theory of ghost distributions is most developed, and prove 1.4. Finally in section 10 we define and study the full ghost centre $\mathcal{Z}_{full}$, and look especially at the cases of type I algebras. The appendix proves 1.4 by generalizing the Harish-Chandra homomorphism to the case when $G$ is not Cartan-even, and studies the image of the Harish-Chandra homomorphism in this case. ### 1.7. Acknowledgments The author would like to thank Alexander Alldridge for many enlightening discussions about this project. Thanks to Vera Serganova for her patience, tremendous support and guidance throughout my PhD, during which much of this work was done. Thank you to Maria Gorelik for patiently explaining many aspects of her relevant papers to me. Finally thank you to Siddartha Sahi, Johannes Flakes, and Inna Entova-Aizenbud for helpful comments and suggestions. This research was partially supported by ISF grant 711/18 and NSF-BSF grant 2019694. ## 2\. Preliminaries from algebraic supergeometry ### 2.1. Linear super algebra notation We work throughout over an algebraically closed field $k$ of characteristic zero. For a super vector space $V$ we write $V=V_{\overline{0}}\oplus V_{\overline{1}}$ for its parity decomposition. Even though we precisely consider $V_{\overline{0}}$ and $V_{\overline{1}}$ to be even vector spaces, we will occasionally (and abusively) view them as super vector spaces where $V_{\overline{0}}$ is purely even and $V_{\overline{1}}$ is purely odd. We write $\operatorname{dim}V=\operatorname{dim}V_{\overline{0}}+\operatorname{dim}V_{\overline{1}}$ for the dimension of the underlying vector space of $V$, and the superdimension of $V$ is given by $\operatorname{sdim}V=\operatorname{dim}V_{\overline{0}}-\operatorname{dim}V_{\overline{1}}$. ### 2.2. Algebraic supergeometry notation We will use the symbols $X,Y,\dots$ for supervarieties with even subschemes $X_{0},Y_{0},\dots.$ We will be considering supervarieties in the sense of [She19], however all spaces of interest will be smooth and affine. A smooth affine supervariety is always given by $\Lambda^{\bullet}E$, the exterior algebra of a vector bundle $E$ on a smooth affine variety $X_{0}$ (see [VMP90]). Thus one will lose almost nothing if one simply works with supervarieties of this form in this article. Note that affine supervarieties and morphisms between them are entirely determined by their spaces of global functions and maps between them, just as with affine varieties. See [CCF11] for more on the basics of algebraic supergeometry. If $X$ is a supervariety, there is a canonical closed embedding $i_{X}:X_{0}\to X$, and this is a homeomorphism of underlying topological spaces. The closed points of $X$ are the $k$-points, which we write as $X(k)$, and they are canonically identified with the closed points of $X_{0}$ via $i_{X}$. If $x$ is a closed point of $X$ and $\mathcal{F}$ is a sheaf on $X$, we write $\mathcal{F}_{x}$ for the stalk of $\mathcal{F}$ at $x$. Then $\mathcal{O}_{X,x}$ is a local superalgebra, and we write $\mathfrak{m}_{x}$ for its corresponding maximal ideal and ${}_{0}\mathfrak{m}_{x}$ for the maximal ideal in $\mathcal{O}_{X_{0},x}$. For affine supervarieties we will also write $\mathfrak{m}_{x}$ for the maximal ideal of $k[X]$ corresponding to $x$, and similarly ${}_{0}\mathfrak{m}_{x}$ for the maximal ideal of $k[X_{0}]$ corresponding to $x$. ### 2.3. Differential operators and distributions ###### Definition 2.1. For a supervariety $X$, let $\mathcal{D}_{X}$ denote the sheaf of filtered algebras which is the subsheaf of $\mathcal{E}nd(\mathcal{O}_{X})$ defined inductively as follows. We set $\mathcal{D}_{X}^{n}=0$ for $n<0$, and for $n\geq 0$ and an open subset $U$ of $X$, set $\Gamma(U,\mathcal{D}_{X}^{n}):=\\{D\in\operatorname{End}(\mathcal{O}_{U})):[D,f]\in\Gamma(U,\mathcal{D}_{X}^{n-1})\text{ for all }f\in\mathcal{O}_{X}(U)\\}$ We call $\mathcal{D}_{X}$ the sheaf of differential operators on $X$, and refer to its sections as differential operators. If $\mathcal{F}$ is a sheaf on $X$, we say that it is a left, resp. right $\mathcal{D}_{X}$-module if it is exactly that. ###### Definition 2.2. If $x$ is a closed point of $X$, define the super vector space of distributions at $x$ to be all (not necessarily even) linear maps $\psi:\mathcal{O}_{X,x}\to k$ such that for some $n\in\mathbb{N}$ we have $\psi(\mathfrak{m}_{x}^{n})=0$. We denote this super vector space by $\operatorname{Dist}(X,x)$. Define $\operatorname{Dist}^{n}(X,x)\subseteq\operatorname{Dist}(X,x)$ to be those distributions vanishing on $\mathfrak{m}_{x}^{n+1}$ so that $\operatorname{Dist}(X,x)$ obtains a filtration. Note that $\operatorname{Dist}^{0}(X,x)$ is one-dimensional and consists of the distinguished even distribution given by evaluation at $x$, which we denote by $\operatorname{ev}_{x}$. We may give $\operatorname{Dist}(X,x)$ the structure of a right $\mathcal{D}_{X}$-module as follows. We view $\operatorname{Dist}(X,x)$ as a sheaf on $X$ supported on $x$. Given a differential operator $D$ defined in a neighborhood of $x$, and a distribution $\psi$, define $(\psi D)(f):=\psi(Df)$ This action respects the filtration on $\operatorname{Dist}(X,x)$, so it becomes a filtered right $\mathcal{D}_{X}$-module. The following lemma is proved in the same way as in the classical setting, so we omit the proof. ###### Lemma 2.3. Let $X$ be a supervariety with a closed point $x$. 1. (1) Given a map of supervarieties $\phi:X\to Y$, we have a natural map of filtered super vector spaces $d\phi_{x}:\operatorname{Dist}(X,x)\to\operatorname{Dist}(Y,\phi(x))$. 2. (2) The chain rule holds: if $\phi:X\to Y$ and $\psi:Y\to Z$, then $d(\psi\circ\phi)=d\psi\circ d\phi$. 3. (3) If $X$ is affine, then the natural pairing $\operatorname{Dist}(X,x)\otimes k[X]\to k$ has the property that if $\psi(f)=0$ for all $f\in k[X]$, then $\psi=0$. 4. (4) There is a natural restriction morphism $\operatorname{res}_{x}:\Gamma(U,\mathcal{D}_{X})\to\operatorname{Dist}(X,x)$ for any open subscheme $U$ containing $x$, given by $\operatorname{res}_{x}(D)(f)=D(f)(x)$. This is a morphism of filtered right $\mathcal{D}_{X}$-modules, where $\mathcal{D}_{X}$ acts on itself by right multiplication. ###### Remark 2.4. We have the following identifications: $\operatorname{Dist}^{n}(X,x)=(\mathcal{O}_{X,x}/\mathfrak{m}_{x}^{n+1})^{*},\ \ \operatorname{Dist}(X,x)=\lim\limits_{\rightarrow}(\mathcal{O}_{X,x}/\mathfrak{m}_{x}^{n+1})^{*}.$ Furthermore when $X$ is affine, we have: $\operatorname{Dist}^{n}(X,x)=(k[X]/\mathfrak{m}_{x}^{n+1})^{*},\ \ \operatorname{Dist}(X,x)=\lim\limits_{\rightarrow}(k[X]/\mathfrak{m}_{x}^{n+1})^{*}.$ In general we have an isomorphism of $\mathcal{D}_{X}$-modules $\operatorname{Dist}(X,x)=\Gamma_{\mathfrak{m}_{x}}(\mathcal{O}_{X,x})^{*}$ and for $X$ affine an isomorphism of $\Gamma(X,\mathcal{D}_{X})$-modules $\operatorname{Dist}(X,x)=\Gamma_{\mathfrak{m}_{x}}k[X]^{*}.$ We recall the definition of the functor $\Gamma_{\mathfrak{m}_{x}}$: $\Gamma_{\mathfrak{m}_{x}}M=\\{m\in M:\mathfrak{m}_{x}^{n}m=0\text{ for }n\gg 0\\}.$ ###### Definition 2.5. Define the sheaf $\mathcal{T}_{X}$ of vector fields on $X$ by setting $\Gamma(\text{Spec}A,\mathcal{T}_{X})=\operatorname{Der}(A)$ for an affine open subscheme $\text{Spec}A$ of $X$. In this way $\mathcal{T}_{X}$ becomes a subsheaf of $\mathcal{D}_{X}^{1}$, and a sheaf of Lie superalgebras under supercommutator. ###### Definition 2.6. For a supervariety $X$ and a closed point $x$ of $X$, we define the tangent space of $X$ at $x$ to be the super vector space $T_{x}X:=(\mathfrak{m}_{x}/\mathfrak{m}_{x}^{2})^{*}$. In this way $T_{x}X$ is exactly the subspace of $\operatorname{Dist}^{1}(X,x)$ given by functionals $\psi:\mathcal{O}_{X,x}/\mathfrak{m}_{x}^{2}\to k$ such that $\psi(1)=0$. Observe the restriction morphism $\operatorname{res}_{x}:\mathcal{D}_{X,x}\to\operatorname{Dist}(X,x)$ restricts to a morphism $\mathcal{T}_{X,x}\to T_{x}X$. ###### Definition 2.7. A supervariety $X$ is smooth if for all $x\in X(k)$ the morphism $\operatorname{res}_{x}:\mathcal{T}_{X,x}\to T_{x}X$ is surjective. See the appendix of [She19] for more equivalent conditions of smoothness. We have the following standard theorem for $\mathcal{D}_{X}$-modules which we will use later on. The proof is almost verbatim from the classical case, so we omit the proof. ###### Proposition 2.8. Suppose that $X$ is a smooth supervariety. Then if a left (or right) $\mathcal{D}_{X}$-module $\mathcal{F}$ is coherent over $\mathcal{O}_{X}$, then it is locally free over $\mathcal{O}_{X}$. ###### Lemma 2.9. Let $X$ be a supervariety and $x\in X(k)$ a closed point at which $X$ is smooth. Suppose $V$ is a subspace of vector fields defined in a neighborhood $U$ of $x$, such that the restriction map $V\to T_{x}X$ is an isomorphism. If $v_{1},\dots,v_{n}$ is a homogeneous basis of $V$, then the restriction of the set of all monomials in $v_{1},\dots,v_{n}$, in any order, is a spanning set of $\operatorname{Dist}(X,x)$ ###### Proof. Choose homogeneous functions $f_{1},\dots,f_{n}$ defined in a neighborhood of $x$ which vanish at $x$ such that they project to a basis of $\mathfrak{m}_{x}/\mathfrak{m}_{x}^{2}$. Then $\mathcal{O}_{X,x}/\mathfrak{m}_{x}^{\ell}$ is isomorphic to the set of polynomials in $f_{1},\dots,f_{n}$ of degree less than or equal to $\ell$. By applying a linear automorphism, we may assume that $\operatorname{res}_{x}(v_{i})(f_{j})=\delta_{ij}$. Then we see that $\operatorname{res}_{x}(v_{1}^{r_{1}}\cdots v_{n}^{r_{n}})(f_{1}^{s_{1}}\cdots f_{n}^{s_{n}})=\pm s_{1}!\cdots s_{n}!\delta_{r_{1}s_{1}}\cdots\delta_{r_{n}s_{n}},$ where the sign is determined by the number of parity changes which occurs in the computation. From this the statement follows. ∎ ###### Definition 2.10. Suppose $\mathcal{F}$ is a quasi-coherent sheaf on $X$. For a closed point $x$ of $X$, define distributions on $\mathcal{F}$ at $x$ to be all linear maps $\psi:\mathcal{F}_{x}\to k$ such that for some $n\in\mathbb{N}$, $\psi(\mathfrak{m}_{x}^{n}\mathcal{F}_{x})=0$. We write this as $\operatorname{Dist}(\mathcal{F},x)$. ###### Remark 2.11. Notice that if $\mathcal{F}$ has a connection, hence an action by vector fields, then $\operatorname{Dist}(\mathcal{F},x)$ naturally admits a right action by vector fields, given by $(\psi v)(s)=\psi(vs).$ If the connection is flat and $X$ is smooth, then this action extends in the natural way to a right action by all differential operators. As before, we have an identification for each $x\in X(k)$: $\operatorname{Dist}(\mathcal{F},x)\cong\Gamma_{\mathfrak{m}_{x}}(\mathcal{F}_{x})^{*}$ and this identification respects the right action by vector fields when $\mathcal{F}$ admits a connection. ## 3\. Differential Operators on a $G$-variety ### 3.1. Preliminaries on algebraic supergroups Recall that an algebraic supergroup is a group object in the category of algebraic supervarieties. Morphisms of supergroups are those which respect the multiplication morphisms. We only consider affine algebraic supergroups, or equivalently those algebraic supergroups that are linear, i.e. have a faithful finite-dimensional representation. The letters $G,H,\dots$ will denote an affine algebraic supergroup. To avoid cumbersome language, we will not write the adjectives affine and algebraic when referring to affine algebraic supergroups, and instead simply call them supergroups. Similarly, we use the term subgroup instead of subsupergroup. We refer again to [CCF11] for basics on algebraic supergroups. A supergroup has a Lie superalgebra which we will denote with the letters $\mathfrak{g},\mathfrak{h},\dots$. The Lie superalgebra may be defined as the super vector space of left-invariant (or right-invariant) vector fields on $G$. As in the classical case, the Lie superalgebra is canonically identified with the tangent space of $G$ at the identity, $T_{e}G$. Morphisms of algebraic supergroups induce morphisms of the corresponding Lie superalgebras. If $G$ is a supergroup, then $G_{0}$ is an algebraic group, and the morphism $i_{G}:G_{0}\to G$ is a morphisms of supergroups. Further, $i_{G}$ induces an isomorphism of Lie algebras $\operatorname{Lie}(G_{0})\cong\operatorname{Lie}(G)_{\overline{0}}$. ### 3.2. Representations of supergroups In this article we will be using some basic facts about the representation theory of quasireductive supergroups. We refer to [Ser11] for further background. Recall that for an affine algebraic supergroup $G$, $k[G]$ has the natural structure of a supercommutative Hopf superalgebra. We define a left $G$-module to be a left $k[G]$-comodule, and a right $G$-module to be a right $k[G]$-comodule. It will be necessary for us to consider both left and right $G$-modules due to the fact that distributions form a right module over the algebra of differential operators, as we have seen. The category of left $G$-modules is equivalent to the category of right $G$-modules because $k[G]$ has a Hopf structure. We will sometimes call a left or right $G$-module simply a representation of $G$, or even a $G$-module, without specifying whether it has a left or right action. In this case either the type of action is apparent or is of no importance. A left (resp. right) $G$-module induces in a natural way a left (resp. right) representation of the Lie superalgebra. Recall that the category of representations of $G$ is equivalent to the category of $(G_{0},\mathfrak{g}=\operatorname{Lie}(G))$-modules such that the action of $G_{0}$ and $\mathfrak{g}_{\overline{0}}\subseteq\mathfrak{g}$ are compatible (see [CCF11]). Given a representation $V$ of a Lie supergroup $G$ such that $\operatorname{dim}V_{\overline{1}}=n$, we write $\operatorname{Ber}(V)$ for the Berezinian of $V$, which is the one-dimensional $G$-module with the same parity as $n$, where the action by $G$ is given by the Berezinian morphism $G\to GL(V)\xrightarrow{\operatorname{Ber}}\mathbb{G}_{m}$ (see chapter 3 of [Man13]). If $V$ is purely even (resp. purely odd) then $\operatorname{Ber}(V)$ coincides with the top exterior power (resp. top symmetric power) of $V$. We write $\operatorname{ber}_{V}:G\to\mathbb{G}_{m}$ for the character of $G$ determined by $\operatorname{Ber}(V)$, and by abuse of notation we also write $\operatorname{ber}_{V}$ for the character of $\mathfrak{g}$ that this determines. Then we have $\operatorname{ber}_{V}=\det_{V_{\overline{0}}}\cdot\det_{V_{\overline{1}}}^{-1}$ as a character of $G_{0}$, and $\operatorname{ber}_{V}=\text{tr}_{V_{\overline{0}}}-\text{tr}_{V_{\overline{1}}}$ as a character of $\mathfrak{g}_{\overline{0}}$. If $\chi:G\to\mathbb{G}_{m}$ is a character of $G$, and $V$ is a representation of $G$, we will write $V^{\chi}$ for the subspace of $V$ where $G$ acts by $\chi$. If $\chi$ is the trivial character, we just write $V^{G}:=V^{\chi}$. If $V$ is a $G$-module and $v\in V$ is a homogeneous element, the $G$-module generated by $v$, which we write as $\langle G\cdot v\rangle$, is given by $\mathcal{U}\mathfrak{g}\cdot\langle G_{0}\cdot v\rangle$. That is, we first take the $G_{0}$-module generated by $v$, and then take the $\mathcal{U}\mathfrak{g}$-module which that generates. Finally, if $V$ is a $G$-module then we will write (when they exist) $I_{G}(V)$, resp. $P_{G}(V)$ (or $I(V)$, resp. $P(V)$ when the context is clear) for the injective hull, resp projective cover of $V$. ### 3.3. Actions of supergroups If $X$ is a supervariety, $G$ a supergroup, and $G$ acts on (the left on) $X$, then we call $X$ a $G$-supervariety. We will usually reserve the letter $a_{X}=a$ for the action morphism, i.e. $a:G\times X\to X$. In this case we will consider $k[X]$ as a right $G$-module via translation. Explicitly, $g\in G_{0}(k)$ acts by pullback $L_{g}^{*}$ along the left translation morphism $L_{g}:X\to X$. The Lie superalgebra acts, for $u\in T_{e}G$, by $u\mapsto(u\otimes 1)\circ a^{*}.$ This induces an map of superalgebras $\mathcal{U}\mathfrak{g}\to\Gamma(X,\mathcal{D}_{X})^{op}$, and in this way $\operatorname{Dist}(X,x)$ becomes a left $\mathcal{U}\mathfrak{g}$-module for any $x\in X(k)$. In general, we will say $\mathfrak{g}$ acts on a supervariety $X$ if it admits a homomorphism of algebras $\mathcal{U}\mathfrak{g}\to\Gamma(X,\mathcal{D}_{X})^{op}$ such that $\mathfrak{g}$ maps into $\Gamma(X,\mathcal{T}_{X})$. ###### Remark 3.1. Suppose that $G$ acts on $X$, and $x\in X(k)$ is a closed point which is fixed by $G_{0}$. Then $G_{0}$ acts on $\operatorname{Dist}(X,x)$. However $\mathfrak{g}$ also acts on $\operatorname{Dist}(X,x)$, and in this way $\operatorname{Dist}(X,x)$ obtains the structure of a $G$-module. Notice that this will happen even if $x$ is not stabilized by all of $G$. ### 3.4. Differential operators on a $G$-supervariety Let $X$ be an affine $G$-supervariety, with action morphism $a:G\times X\to X$, and consider $D_{X}=\Gamma(X,\mathcal{D}_{X})$. ###### Lemma 3.2. For a differential operator $D\in D_{X}$, the following are equivalent: 1. (1) The map $D:k[X]\to k[X]$ is $G$-equivariant; 2. (2) We have $a^{*}\circ D=\operatorname{id}\otimes D\circ a^{*}$; and in the case when $G$ is connected, we have the third equivalent condition: * (3) For all $u\in\mathfrak{g}$, $[u,D]=0$. In this case we say that $D$ is $G$-invariant. ###### Proof. (1)$\iff$(3) is obvious when $G$ is connected. And (2) says that $D$ is a $k[G]$-comodule homomorphism, equivalently a $G$-module homomorphism giving (1)$\iff$(2). ∎ ###### Definition 3.3. Write $D_{X}^{G}$ for the superalgebra of $G$-invariant differential operators on $X$. For the meaning of an open orbit of an algebraic supergroup, see [She19]. ###### Proposition 3.4. Let $X$ be a $G$-supervariety, and $x$ a point of $X$ with stabilizer subgroup $K\subseteq G$. If $G$ has an open orbit at $x$, then the morphism $\operatorname{res}_{x}:D_{X}\to\operatorname{Dist}(X,x)$ restricts to an injection $D_{X}^{G}\to\operatorname{Dist}(X,x)^{K}$. ###### Proof. The map $\operatorname{res}_{x}$ is $K$-equivariant, so certain $\operatorname{res}_{x}(D_{X}^{G})\subseteq\operatorname{Dist}(X,x)^{K}$. To see that it is injective, let $a_{x}:G\to X$ be the orbit map at $x$, and $D\in D_{X}^{G}$. Then by $a_{x}^{*}:\mathcal{O}_{X}\to(a_{x})_{*}\mathcal{O}_{G}$ is an injective morphism of sheaves by Prop. 3.11 of [She19]. Therefore, if $\operatorname{res}_{x}(D)=0$, we have $D(f)(x)=0$ for all $f$ defined in an open neighborhood of $x$. Equivalently, $a_{x}^{*}(D(f))(e)=0$ for all such $f$. But we have the factorization $a_{x}=a\circ(\operatorname{id}_{G}\times i_{x})$, so this says that $\displaystyle(\operatorname{id}_{G}\times i_{x})^{*}\circ a^{*}(D(f))$ $\displaystyle=$ $\displaystyle(\operatorname{id}_{G}\times i_{x})^{*}(\operatorname{id}\otimes D)(a^{*}(f))$ $\displaystyle=$ $\displaystyle(\operatorname{id}\otimes\operatorname{res}_{x}(D))(a^{*}(f))=0.$ This implies $D(f)=0$ for all $f$ defined in an open neighborhood of $X$. Since by definition the restriction morphism on functions is injective for a supervariety, this implies that $D=0$. ∎ ###### Proposition 3.5. In the context of the previous proposition, if $X\cong G/K$ via the orbit map at $x$ then the map $D_{X}^{G}\to\operatorname{Dist}(X,x)^{K}$ is an isomorphism. ###### Proof. It remains to show it is surjective. For this, if $\psi\in\operatorname{Dist}(X,x)^{K}$, define $D_{\psi}$ by $f\mapsto(\operatorname{id}\otimes\psi)\circ a^{*}(f)$. Indeed, $\displaystyle a^{*}\circ D_{\psi}$ $\displaystyle=$ $\displaystyle a^{*}\circ(\operatorname{id}\otimes\psi)\circ a^{*}$ $\displaystyle=$ $\displaystyle(\operatorname{id}\otimes\operatorname{id}\otimes\psi)\circ(a^{*}\otimes\operatorname{id})\circ a^{*}$ $\displaystyle=$ $\displaystyle(\operatorname{id}\otimes\operatorname{id}\otimes\psi)\circ(\operatorname{id}\otimes a^{*})\circ a^{*}$ $\displaystyle=$ $\displaystyle[\operatorname{id}\otimes((\operatorname{id}\otimes\psi)\circ a^{*})]\circ a^{*}$ $\displaystyle=$ $\displaystyle(\operatorname{id}\otimes D_{\psi})\circ a^{*}.$ ∎ ## 4\. Induced supervarieties in the sense of Koszul We present a construction that is originally due to Koszul in [Kos82]. Although it may be defined for non-affine supervarieties, for simplicity we stick to the affine case since that is all we need. Throughout, we let $H$ be a supergroup and write $\mathfrak{h}=\operatorname{Lie}(H)$. ### 4.1. Induced and coinduced modules ###### Definition 4.1. Let $V_{0}$ be an $\mathfrak{h}_{\overline{0}}$-module, and define the $\mathfrak{h}$-module $\operatorname{Ind}_{\mathfrak{h}_{\overline{0}}}^{\mathfrak{h}}V_{0}$ to be $\operatorname{Ind}_{\mathfrak{h}_{\overline{0}}}^{\mathfrak{h}}V_{0}:=\mathcal{U}\mathfrak{h}\otimes_{\mathcal{U}\mathfrak{h}_{\overline{0}}}V_{0}.$ The action by $\mathfrak{h}$ is left multiplication. If the action of $\mathfrak{h}_{\overline{0}}$ on $V_{0}$ integrates to an action of $H_{0}$, then the $\mathfrak{h}$ action on $\operatorname{Ind}_{\mathfrak{h}_{\overline{0}}}^{\mathfrak{h}}V_{0}$ integrates to an action of $H$, where $H_{0}$ acts by $h\cdot(u\otimes v)=\operatorname{Ad}(h)(u)\otimes h\cdot v.$ Similarly we define the $\mathfrak{h}$-module $\operatorname{Coind}_{\mathfrak{h}_{\overline{0}}}^{\mathfrak{h}}V_{0}$ by $\operatorname{Coind}_{\mathfrak{h}_{\overline{0}}}^{\mathfrak{h}}V_{0}:=\operatorname{Hom}_{\mathfrak{h}_{\overline{0}}}(\mathcal{U}\mathfrak{h},V_{0}).$ Here we consider $\mathcal{U}\mathfrak{h}$ as a left $\mathcal{U}\mathfrak{h}_{\overline{0}}$-module. The action by $\mathfrak{h}$ on this module is $(u\eta)(v)=(-1)^{\overline{u}(\overline{\eta}+\overline{v})}\eta(vu).$ Once again, if $V_{0}$ is actually an $H_{0}$-module then $\operatorname{Coind}_{\mathfrak{h}_{\overline{0}}}^{\mathfrak{h}}V_{0}$ will be an $H$-module, where the action by $H_{0}$ is given by $(h\cdot\eta)(v)=h\cdot\eta(\operatorname{Ad}(h^{-1})(v)).$ ###### Remark 4.2. If $A_{0}$ is an algebra on which $\mathfrak{h}_{\overline{0}}$ acts by derivations, then $\operatorname{Coind}_{\mathfrak{h}_{\overline{0}}}^{\mathfrak{h}}A_{0}$ is naturally a superalgebra with multiplication given by $(\eta\xi)(u)=m_{A_{0}}\circ(\eta\otimes\xi)(\Delta(u)).$ In this case, the action of $\mathfrak{h}$ on $\operatorname{Coind}_{\mathfrak{h}_{\overline{0}}}^{\mathfrak{h}}A_{0}$ is by super derivations. A similar statement holds if $A_{0}$ is a Hopf algebra which $\mathfrak{h}_{\overline{0}}$ acts on by derivations preserving the Hopf algebra structure. ###### Lemma 4.3. For an $\mathfrak{h}_{\overline{0}}$-module $V_{0}$, we have a canonical isomorphism of $\mathfrak{h}$-modules $(\operatorname{Coind}_{\mathfrak{h}_{\overline{0}}}^{\mathfrak{h}}V_{0})^{*}\cong\operatorname{Ind}_{\mathfrak{h}_{\overline{0}}}^{\mathfrak{h}}(V_{0})^{*}.$ This extends to an isomorphism of $H$-modules when $V_{0}$ is an $H_{0}$-module. ###### Proof. We always have a canonical map of $\mathfrak{h}$-modules (or $H$-modules when applicable) $\operatorname{Ind}_{\mathfrak{h}_{\overline{0}}}^{\mathfrak{h}}(V_{0})^{*}\to(\operatorname{Coind}_{\mathfrak{h}_{\overline{0}}}^{\mathfrak{h}}V_{0})^{*}$ given by $(u\otimes\varphi)(\eta)=(-1)^{\overline{u}(\overline{\varphi}+\overline{\eta})}\varphi(\eta(s(u))),$ where $s:\mathcal{U}\mathfrak{h}\to\mathcal{U}\mathfrak{h}$ is the antipode. Here it is an isomorphism by the PBW theorem, since $\mathcal{U}\mathfrak{h}$ is a finitely generated free $\mathcal{U}\mathfrak{h}_{\overline{0}}$-module. ∎ ### 4.2. Koszul’s induced superspace ###### Definition 4.4. For an affine variety $X_{0}$ with an action of $\mathfrak{h}_{\overline{0}}$, define $(X_{0})^{\mathfrak{h}}$ to be the affine variety with coordinate ring $k[(X_{0})^{\mathfrak{h}}]:=\operatorname{Hom}_{\mathfrak{h}_{\overline{0}}}(\mathcal{U}\mathfrak{h},k[X_{0}])=\operatorname{Coind}_{\mathfrak{h}_{0}}^{\mathfrak{h}}k[X_{0}].$ By remark 4.2, $(X_{0})^{\mathfrak{h}}$ is an affine supervariety with an action by $\mathfrak{h}$. If the action of $\mathfrak{h}_{\overline{0}}$ on $X_{0}$ comes from an action of $H_{0}$ on $X_{0}$, then $(X_{0})^{\mathfrak{h}}$ will be an $H$-supervariety. We have that $((X_{0})^{\mathfrak{h}})_{0}=X_{0}$, and the natural projection $k[(X_{0})^{\mathfrak{h}}]\to k[X_{0}]$ is given by $\eta\mapsto\eta(1)$. ###### Remark 4.5. The construction of the induced superspace given above can be done for any supervariety $X_{0}$ with an $\mathfrak{h}_{\overline{0}}$-action as follows. We define $(X_{0})^{\mathfrak{h}}$ to have underlying space $X_{0}$ and sheaf of functions given by, for an open subset $U_{0}\subseteq X_{0}$, $\Gamma(U_{0},\mathcal{O}_{(X_{0})^{\mathfrak{h}}})=\operatorname{Coind}_{\mathfrak{h}_{0}}^{\mathfrak{h}}k[U_{0}]=k[(U_{0})^{\mathfrak{h}}].$ One can check this construction respects localization and thus gives a well- defined supervariety. ###### Remark 4.6. If $X_{0}$ is smooth, then $(X_{0})^{\mathfrak{h}}$ is also smooth. In any case, $\mathfrak{h}_{\overline{1}}$ defines everywhere non-vanishing vector fields on $(X_{0})^{\mathfrak{h}}$ so that $\mathfrak{h}_{\overline{1}}\to T_{x}(X_{0})^{\mathfrak{h}}$ is an isomorphism for all $x\in X_{0}(k)$. ###### Remark 4.7. In [Kos82], it was shown that $G=(G_{0})^{\mathfrak{g}}$. Thus we have an explicit description of the algebra of functions on $k[G]$ given by $k[G]=\operatorname{Hom}_{\mathfrak{g}_{\overline{0}}}(\mathcal{U}\mathfrak{g},k[G_{0}]).$ From this one can also determine the Hopf algebra structure on $k[G]$ from the Hopf algebra structures on $\mathcal{U}\mathfrak{g}$ and $k[G_{0}]$. ###### Proposition 4.8. Let $X_{0}$ be an affine variety with an action of $\mathfrak{h}_{\overline{0}}$. Then $(X_{0})^{\mathfrak{h}}$ has the following universal property: given an affine supervariety $Y$ with an action of $\mathfrak{h}$, and an $\mathfrak{h}_{\overline{0}}$-equivariant map $\overline{\phi}:X_{0}\to Y_{0}$, there exists a unique $\mathfrak{h}$-equivariant map $\phi:(X_{0})^{\mathfrak{h}}\to Y$ such that $\phi_{0}=\overline{\phi}$ and the following diagram commutes: $\textstyle{(X_{0})^{\mathfrak{h}}\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{\phi}$$\textstyle{Y}$$\textstyle{X_{0}\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{i_{X_{0}}}$$\scriptstyle{\overline{\phi}}$$\textstyle{Y_{0}\ignorespaces\ignorespaces\ignorespaces\ignorespaces.}$$\scriptstyle{i_{Y_{0}}}$ If the actions of $\mathfrak{h}_{\overline{0}}$ and $\mathfrak{h}$ integrate to actions of $H_{0}$ and $H$, respectively, then $\phi$ is a morphism of $H$-supervarieties. ###### Proof. Since everything is affine, we may work instead with the algebras of functions. Then the commutativity of the square says that for such a map $\phi$, we must have that for $f\in k[Y]$, $\phi^{*}(f)(1)=(\overline{\phi})^{*}\circ(i_{Y_{0}})^{*}(f)$ The property of begin an $\mathfrak{h}$-equivariant map then forces, for $u\in\mathcal{U}\mathfrak{h}$, $\phi^{*}(f)(u)=(-1)^{\overline{u}\overline{f}}(u\phi^{*}(f))(1)=(-1)^{\overline{u}\overline{f}}\phi^{*}(uf)(1)=(-1)^{\overline{u}\overline{f}}(\overline{\phi})^{*}((i_{Y_{0}})^{*}(uf))$ so the definition of $\phi^{*}$ is forced on us. One can check that the above definition defines an algebra homomorphism, and so we are done. ∎ For the following, observe that if $X_{0}$ is an $H_{0}$-variety and $x$ is a closed point of $X_{0}$ which is fixed by $H_{0}$, then $H_{0}$ preserves ${}_{0}\mathfrak{m}_{x}$ and thus acts on $\operatorname{Dist}(X_{0},x)$. ###### Proposition 4.9. Let $X_{0}$ be affine $H_{0}$-variety, and $x\in X_{0}(k)$ a closed point which is fixed by $H_{0}$. Then we have an isomorphism of $H$-modules $\operatorname{Dist}((X_{0})^{\mathfrak{h}},x)\cong\operatorname{Ind}_{\mathfrak{h}_{\overline{0}}}^{\mathfrak{h}}\operatorname{Dist}(X_{0},x).$ (See remark 3.1 for the $H$-module structure on $\operatorname{Dist}((X_{0})^{\mathfrak{h}},x)$.) ###### Proof. We have the following string of isomorphisms of $H$-modules: $\displaystyle\operatorname{Dist}((X_{0})^{\mathfrak{h}},x)$ $\displaystyle\cong$ $\displaystyle\Gamma_{\mathfrak{m}_{x}}(k[(X_{0})^{\mathfrak{h}}])^{*}$ $\displaystyle=$ $\displaystyle\Gamma_{\mathfrak{m}_{x}}(\operatorname{Coind}_{\mathfrak{h}_{\overline{0}}}^{\mathfrak{h}}k[X_{0}])^{*}$ $\displaystyle\cong$ $\displaystyle\Gamma_{\mathfrak{m}_{x}}\operatorname{Ind}_{\mathfrak{h}_{\overline{0}}}^{\mathfrak{h}}(k[X_{0}])^{*}$ $\displaystyle=$ $\displaystyle\operatorname{Ind}_{\mathfrak{h}_{\overline{0}}}^{\mathfrak{h}}\Gamma_{{}_{0}\mathfrak{m}_{x}}(k[X_{0}])^{*}$ $\displaystyle=$ $\displaystyle\operatorname{Ind}_{\mathfrak{h}_{\overline{0}}}^{\mathfrak{h}}\operatorname{Dist}(X_{0},x)$ The non-trivial step here is that we can move $\Gamma$ past $\operatorname{Ind}$. To see this, observe that $\mathfrak{m}_{x}=\\{\eta\in\operatorname{Hom}_{\mathfrak{h}_{\overline{0}}}(\mathcal{U}\mathfrak{h},k[X_{0}]):\eta(1)\in{{}_{0}\mathfrak{m}_{x}}\\}.$ It follows that for any $d\in\mathbb{N}$, for $n\gg d$, if $\eta\in\mathfrak{m}_{x}^{n}$, then $\eta(u)\in$ ${}_{0}\mathfrak{m}_{x}^{d}$ for all $u\in\mathcal{U}\mathfrak{h}$, because $\mathfrak{h}_{\overline{0}}$ preserves the ideal ${}_{0}\mathfrak{m}_{x}$. The claim now follows. ∎ ### 4.3. Induced vector bundles Let us extend the above framework to more general vector bundles. Note that this subsection can largely be skipped, as it is only used to prove the $\operatorname{Ind}-\operatorname{Coind}$ isomorphism in section 6.1, which is already known via other methods. See section 4.6 of [She20a] for the definition of $\mathfrak{h}$-equivariant vector bundle. Again assume that $X_{0}$ is affine, and suppose that $F_{0}$ is an $\mathfrak{h}_{\overline{0}}$-equivariant vector bundle over $X_{0}$. Then $\mathfrak{h}_{\overline{0}}$ acts on $\Gamma(X_{0},F_{0})$ and satisfies, for $f_{0}\in k[X_{0}]$, $s_{0}\in\Gamma(X_{0},F_{0})$, and $u_{0}\in\mathfrak{h}_{\overline{0}}$, $u_{0}\cdot(f_{0}s_{0})=u_{0}(f_{0})s_{0}+f_{0}u_{0}(s_{0}),$ where the action of $u_{0}$ on $f_{0}$ is coming from the action on $k[X_{0}]$. If $F_{0}$ is actually an $H_{0}$-equivariant vector bundle, then for $h\in H_{0}$ we have $h\cdot(f_{0}s_{0})=L_{h}^{*}(f_{0})h\cdot(s_{0}).$ Then we define an $\mathfrak{h}$-equivariant vector bundle $(F_{0})^{\mathfrak{h}}$ over $(X_{0})^{\mathfrak{h}}$ to have the space of sections given by $\Gamma((X_{0})^{\mathfrak{h}},(F_{0})^{\mathfrak{h}}):=\operatorname{Coind}_{\mathfrak{h}_{\overline{0}}}^{\mathfrak{h}}\Gamma(X_{0},F_{0})=\operatorname{Hom}_{\mathfrak{h}_{\overline{0}}}(\mathcal{U}\mathfrak{h},\Gamma(X_{0},F_{0})).$ Then this admits a natural $\mathfrak{h}$-action by virtue of being a coinduced module, and if $F_{0}$ is $H_{0}$-equivariant, $(F_{0})^{\mathfrak{h}}$ will be $H$-equivariant. The $k[(X_{0})^{\mathfrak{h}}]$-module structure is defined by $(\varphi s)(u)=a_{0,F_{0}}^{*}\circ(\varphi\otimes s)(\Delta(u))$ where $a_{0,F_{0}}^{*}:k[X_{0}]\otimes\Gamma(X_{0},F_{0})\to\Gamma(X_{0},F_{0})$ is the action map. We check that this makes sense: for $v_{0}\in\mathfrak{h}_{\overline{0}}$, writing $\Delta(u)=\sum u_{i}\otimes u^{i}$, $\displaystyle(\varphi s)(v_{0}u)$ $\displaystyle=$ $\displaystyle a_{0}\circ(\varphi\otimes s)((v_{0}\otimes 1+1\otimes v_{0})\Delta(u))$ $\displaystyle=$ $\displaystyle(-1)^{\overline{s}\overline{u_{i}}}(\varphi(v_{0}u_{i})s(u^{i})+\varphi(u_{i})s(v_{0}u^{i}))$ $\displaystyle=$ $\displaystyle(-1)^{\overline{s}\overline{u_{i}}}((v_{0}\varphi)(u_{i})s(u^{i})+\varphi(u_{i})(v_{0}s)(u^{i}))$ $\displaystyle=$ $\displaystyle v_{0}((\varphi s)(u)).$ It’s also straightforward to show that for $v\in\mathfrak{h}$ we have $v(\varphi s)=v(\varphi)s+(-1)^{\overline{v}\overline{\varphi}}\varphi v(s),$ and (when applicable) for $h\in H_{0}$ we have $h\cdot(\varphi s)=L_{h}^{*}(\varphi)h\cdot s.$ This construction is local in the sense that we have, for an open subvariety $U_{0}\subseteq X_{0}$, $\Gamma(U_{0},(F_{0})^{\mathfrak{h}})=\operatorname{Coind}_{\mathfrak{h}_{\overline{0}}}^{\mathfrak{h}}\Gamma(U_{0},F_{0})=\operatorname{Hom}_{\mathfrak{h}_{\overline{0}}}(\mathcal{U}\mathfrak{h},\Gamma(U_{0},F_{0})).$ From this it is not hard to show that $(F_{0})^{\mathfrak{h}}$ is an even rank vector bundle on $(X_{0})^{\mathfrak{h}}$ of the same rank as $F_{0}$. ###### Proposition 4.10. There is a natural bijection between even rank $\mathfrak{h}$-equivariant vector bundles on $(X_{0})^{\mathfrak{h}}$ and $\mathfrak{h}_{\overline{0}}$-equivariant vector bundles on $X_{0}$, given explicitly as follows: * • given an $\mathfrak{h}_{\overline{0}}$-equivariant vector bundle $F_{0}$ on $X_{0}$, we may produce the $\mathfrak{h}$-equivariant vector bundle $(F_{0})^{\mathfrak{h}}$ on $X$; and * • given an $\mathfrak{h}$-equivariant vector bundle $F$ on $(X_{0})^{\mathfrak{h}}$ we may take the $\mathfrak{h}_{\overline{0}}$-equivariant vector bundle on $X_{0}$ gotten by pulling back along embedding $i:X_{0}\hookrightarrow(X_{0})^{\mathfrak{h}}$. This restricts to a bijection of even rank $H$-equivariant vector bundles on $(X_{0})^{\mathfrak{h}}$ and $H_{0}$-equivariant vector bundles on $X_{0}$, when applicable. ###### Proof. If we start with an $\mathfrak{h}_{\overline{0}}$-equivariant vector bundle $F_{0}$, then we have a natural surjection $\Gamma(X_{0},i^{*}(F_{0})^{\mathfrak{h}})\to\Gamma(X_{0},F_{0})$ given by evaluating an element of $\operatorname{Hom}_{\mathfrak{h}_{\overline{0}}}(\mathcal{U}\mathfrak{h},\Gamma(X_{0},F_{0}))$ at $1\in\mathcal{U}\mathfrak{h}$. It is an $\mathfrak{h}_{\overline{0}}$-equivariant map, and since the vector bundles have the same rank, it must be an isomorphism. In the other direction, suppose we have an $\mathfrak{h}$-equivariant bundle $F$. Then we have a natural map of $\mathfrak{h}$-equivariant vector bundles $F\to(i^{*}F)^{\mathfrak{h}}$, given on sections $\Gamma(X,F)\to\operatorname{Hom}_{\mathfrak{h}_{\overline{0}}}(\mathcal{U}\mathfrak{h},\Gamma(X,i^{*}F))$ by $s\mapsto(u\mapsto i^{*}(us))$ To see this is an isomorphism, it suffices to look at the map on fibers, which are isomorphisms. ∎ ###### Proposition 4.11. Under the same hypothesis of 4.9, let $F_{0}$ be an $H_{0}$-equivariant vector bundle on $X_{0}$. Then we have a natural isomorphism of $H$-modules $\operatorname{Dist}((F_{0})^{\mathfrak{h}},x)\cong\operatorname{Ind}_{\mathfrak{h}_{\overline{0}}}^{\mathfrak{h}}\operatorname{Dist}(F_{0},x)$ ###### Proof. The proof extends almost verbatim from 4.9. ∎ ## 5\. Induced spaces and homogeneous spaces ### 5.1. A local isomorphism for certain homogeneous supervarieties Given a supergroup $G$ and a closed algebraic subgroup $K$, we may form the homogeneous supervariety $G/K$. For technical aspects of such spaces see [MT18] and [MZ10]. In particular $G/K$ is always smooth and $(G/K)_{0}=G_{0}/K_{0}$, so that $G/K$ is affine if and only if $G_{0}/K_{0}$ is. Let $X=G/K$ be a homogeneous affine supervariety, and suppose that there exists a subgroup $H$ of $G$ such that $\mathfrak{h}_{\overline{1}}\oplus\mathfrak{k}_{\overline{1}}=\mathfrak{g}_{\overline{1}}$. Then consider the $H$-supervariety $(G_{0}/K_{0})^{\mathfrak{h}}$. By its universal property it admits a canonical $H$-equivariant morphism to $G/K$. ###### Proposition 5.1. The canonical $H$-equivariant map $\phi:(G_{0}/K_{0})^{\mathfrak{h}}\to G/K$ induces an isomorphism of supervarieties in a Zariski open neighborhood of $eK_{0}$. In particular, the map on functions $\phi^{*}:k[G/K]\to k[(G_{0}/K_{0})^{\mathfrak{h}}]=\operatorname{Hom}_{\mathfrak{h}_{\overline{0}}}(\mathcal{U}\mathfrak{h},k[G_{0}/K_{0}])$ is an injective $H$-module homomorphism First we need a lemma. ###### Lemma 5.2. Suppose that $f:X\to Y$ is a morphism of smooth affine supervarieties such that $f_{0}:X_{0}\to Y_{0}$ is an isomorphism and $df_{x}$ is an isomorphism for all closed points $x\in X(k)$. Then $f$ is an isomorphism. ###### Proof. Because $X$ and $Y$ are smooth and affine, we may present them as exterior algebras of vector bundles $E_{X_{0}}$, $E_{Y_{0}}$ on $X_{0}$, $Y_{0}$. Working on a cover, we may assume these vector bundles are trivial. Let $\xi_{1},\dots,\xi_{n}\in\Gamma(Y_{0},E_{Y_{0}})\subseteq k[Y]_{\overline{1}}$ be a $k[Y_{0}]$-basis for $\Gamma(Y_{0},E_{Y_{0}})$ so these elements project to a basis of $(T_{y}Y)_{\overline{1}}$ for each $y\in Y(k)$. Then $f^{*}(\xi_{1}),\dots,f^{*}(\xi_{n})\in k[X]_{\overline{1}}$ must project to a basis of $T_{x}X$ for all $x\in X(k)$. It follows that $f^{*}(\xi_{1}),\dots,f^{*}(\xi_{n})$ project to a $k[X_{0}]$-basis of $\Gamma(X_{0},E_{X_{0}})$ in the associated graded of $k[X]$. Hence the associated graded morphism of $f^{*}$ is an isomorphism, which implies that $f^{*}$ is an isomorphism, and we are done. ∎ ###### Proof of 5.1. Each space admits an action by $\mathfrak{h}$ as vector fields acting on functions, and since $\phi^{*}$ is a $\mathfrak{h}$-homomorphism we have $\phi^{*}(uf)=u\phi^{*}(f)$ for $u\in\mathfrak{h},f\in k[G/K]$. Hence for any closed point $x$ of $G/K$ the following diagram is commutative: $\textstyle{\mathfrak{h}\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\textstyle{T_{x}(G_{0}/K_{0})^{\mathfrak{h}}\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{df_{x}}$$\textstyle{T_{x}(G/K)}$ In particular, wherever $\mathfrak{h}_{\overline{1}}$ spans the odd part of the tangent space of $G/K$ at $x$, $df_{x}$ will be an isomorphism of vector spaces when restricted to the odd part. However, we see that $f_{0}^{*}$ is the identity map by construction, hence we get a commutative diagram: $\textstyle{(G_{0}/K_{0})^{\mathfrak{h}}\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{f}$$\textstyle{G/K}$$\textstyle{G_{0}/K_{0}\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{f_{0}=\operatorname{id}}$$\textstyle{G_{0}/K_{0}.\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$ If we restrict to the open neighborhood of $eK_{0}$ upon which $\mathfrak{h}_{\overline{1}}$ spans the odd tangent space of $G/K$, $f$ will be an isomorphism by lemma 5.2. ∎ ###### Corollary 5.3. Maintain the assumptions of 5.1. and suppose further that $H_{0}\subseteq K_{0}$. Then we have a canonical isomorphism of $H$-modules $\operatorname{Dist}(G/K,eK)\cong\operatorname{Ind}_{\mathfrak{h}_{\overline{0}}}^{\mathfrak{h}}\operatorname{Dist}(G_{0}/K_{0},eK_{0}).$ Or in terms of enveloping algebras, $\mathcal{U}\mathfrak{g}/(\mathcal{U}\mathfrak{g}\mathfrak{k})\cong\operatorname{Ind}_{\mathfrak{h}_{\overline{0}}}^{\mathfrak{h}}\mathcal{U}\mathfrak{g}_{\overline{0}}/(\mathcal{U}\mathfrak{g}_{\overline{0}}\mathfrak{k}_{\overline{0}})$ ###### Proof. The first statement follows from 4.9. The second statement follows from the usual identification of $\mathcal{U}\mathfrak{g}$-modules $\operatorname{Dist}(G/K,eK)\cong\mathcal{U}\mathfrak{g}/(\mathcal{U}\mathfrak{g}\mathfrak{k}).$ which comes from the composition $\mathcal{U}\mathfrak{g}\to\Gamma(G/K,\mathcal{D}_{G/K})\xrightarrow{\operatorname{res}_{eK}}\operatorname{Dist}(G/K,eK).$ ∎ ### 5.2. Supersymmetric spaces From now on, we assume that $G$ is a connected supergroup, i.e. an affine algebraic supergroup such that $G_{0}$ is connected. Let $\theta$ be an involution of $G$, and let $K$ be a closed subgroup of $G$ such that $(G^{\theta})^{0}\subseteq K\subseteq G^{\theta}$. In particular $K$ need not be connected. From this we may consider the homogeneous supervariety $G/K$, and we call $G$-supervarieties of this form symmetric supervarieties (or supersymmetric spaces). On the level of Lie superalgebras, $\theta$ induces an involution on $\mathfrak{g}$, which by abuse of notation we again write as $\theta$, giving rise to the decomposition $\mathfrak{g}=\mathfrak{k}\oplus\mathfrak{p}$, where $\mathfrak{k}=\operatorname{Lie}K$ is the fixed subspace of $\theta$ and $\mathfrak{p}$ the $(-1)$-eigenspace of $\theta$. ###### Definition 5.4. Define $\mathfrak{k}^{\prime}:=\mathfrak{k}_{\overline{0}}\oplus\mathfrak{p}_{\overline{1}}$, which is the fixed points of the involution $\delta\circ\theta$, where $\delta(x)=(-1)^{\overline{x}}x$ is the grading operator on $\mathfrak{g}$. Let $K^{\prime}$ denote the closed algebraic subgroup of $G$ such that $K^{\prime}_{0}=K_{0}$ and $\operatorname{Lie}K^{\prime}=\mathfrak{k}^{\prime}$. Notice that $\delta\circ\theta$ is an involution on $G$ such that $(G^{\delta\circ\theta})^{0}\subseteq K^{\prime}\subseteq G^{\delta\circ\theta}$, hence $G/K^{\prime}$ is another symmetric supervariety. ###### Proposition 5.5. We have a $K^{\prime}$-equivariant morphism $(G_{0}/K_{0})^{\mathfrak{k}^{\prime}}\to G/K$ which is an isomorphism in a neighborhood of $eK_{0}$. In particular, the pullback morphism of functions $k[G/K]\to k[(G_{0}/K_{0})^{\mathfrak{k}^{\prime}}]=\operatorname{Coind}_{\mathfrak{k}_{\overline{0}}}^{\mathfrak{k}^{\prime}}k[G_{0}/K_{0}]$ is injective. ###### Proof. This follows immediately from 5.1. ∎ ###### Corollary 5.6. We have a natural injective morphism of algebras $k[K^{\prime}\backslash G/K]=k[G/K]^{K^{\prime}}\to k[K_{0}\backslash G_{0}/K_{0}]=k[G_{0}/K_{0}]^{K_{0}}.$ In particular, $k[K^{\prime}\backslash G/K]$ is an integral domain. ###### Proof. Taking $K^{\prime}$ invariants of the pullback morphism we obtain an injection $k[G/K]^{K^{\prime}}\to\left(\operatorname{Coind}_{\mathfrak{k}_{\overline{0}}}^{\mathfrak{k}^{\prime}}k[G_{0}/K_{0}]\right)^{K^{\prime}}.$ Now one may use Frobenius reciprocity to identify $\left(\operatorname{Coind}_{\mathfrak{k}_{\overline{0}}}^{\mathfrak{k}^{\prime}}k[G_{0}/K_{0}]\right)^{K^{\prime}}$ with $k[G_{0}/K_{0}]^{K_{0}}$ as an algebra. ∎ The following result, which now is proven easily from 5.3, is an appropriate generalization of the fundamental observation made in [Gor00]. ###### Proposition 5.7. We have a canonical isomorphism of $K^{\prime}$-modules $\operatorname{Dist}(G/K,eK)\cong\operatorname{Ind}_{\mathfrak{k}_{\overline{0}}}^{\mathfrak{k}^{\prime}}\operatorname{Dist}(G_{0}/K_{0},eK_{0}).$ ## 6\. The symmetric space $G/G_{0}$ Consider the involution $\theta=\delta$, the canonical grading operator on $\mathfrak{g}$ which is defined by $\delta(x)=-x$. In this case $G^{\delta}=G_{0}$, and the local isomorphism of 5.5 becomes a global isomorphism of $G$-supervarieties (both consisting of just one point): $G/G_{0}\cong(G_{0}/G_{0})^{\mathfrak{g}}.$ ### 6.1. $\operatorname{Ind}$-$\operatorname{Coind}$ isomorphism The homogeneous space $G/G_{0}$ has one closed point, which we will call $e$. Let $V$ be a $G$-equivariant vector bundle on $G/G_{0}$, and write $V$ again for its space of sections. Since $\mathfrak{m}_{e}$ is nilpotent, we have the identification $\operatorname{Dist}(V,e)=V^{*}$. A $G_{0}$-equivariant vector bundle on $G_{0}/G_{0}$ is the same data as a finite-dimensional $G_{0}$-representation $V_{0}$, and its sections are again $V_{0}$. Then 4.11 tells us that $\operatorname{Dist}((V_{0})^{\mathfrak{g}},e)=(\operatorname{Coind}_{\mathfrak{g}_{\overline{0}}}^{\mathfrak{g}}V_{0})^{*}\cong\operatorname{Ind}_{\mathfrak{g}_{\overline{0}}}^{\mathfrak{g}}(V_{0})^{*},$ which we already showed in lemma 4.3. We will write $(V_{0})^{\mathfrak{g}}$ once again for the sections of $(V_{0})^{\mathfrak{g}}$ on $G/G_{0}$ since this space has only one point. Since $D_{G/G_{0}}$ is generated by $\mathcal{U}\mathfrak{g}$ and $k[G/G_{0}]$ as an algebra, $(V_{0})^{\mathfrak{g}}$ will be a $D_{G/G_{0}}$-module on $G/G_{0}$. Thus $\operatorname{Dist}((V_{0})^{\mathfrak{g}},e)$ is a $D_{G/G_{0}}$-module, and, being finite-dimensional, must also be coherent. By 2.8, $\operatorname{Dist}((V_{0})^{\mathfrak{g}},e)$ must be a vector bundle on $G/G_{0}$, and is $G$-equivariant via its $\mathcal{D}_{G/G_{0}}$-module structure. Further it is of even rank since since $(V_{0})^{\mathfrak{g}}$ is of even rank, and therefore by 4.10 there exists a $G_{0}$-equivariant vector bundle on $G_{0}/G_{0}$, that is a $G_{0}$-representation $W_{0}$, such that $\operatorname{Dist}((V_{0})^{\mathfrak{g}},e)\cong(W_{0})^{\mathfrak{g}}.$ By 4.10, $W_{0}$ is obtained by taking the sections of the restriction of $(V_{0})^{\mathfrak{g}}$ to $G_{0}/G_{0}$, i.e. $W_{0}\cong\operatorname{Dist}((V_{0})^{\mathfrak{g}},e)/\mathfrak{m}_{e}\operatorname{Dist}((V_{0})^{\mathfrak{g}},e).$ Now let $n=\operatorname{dim}\mathfrak{g}_{\overline{1}}$ so that $\mathfrak{m}_{e}^{n+1}=0$ but $\mathfrak{m}_{e}^{n}\neq 0$. Then $\operatorname{Dist}(V,e)\cong((V_{0})^{\mathfrak{g}}/\mathfrak{m}_{e}^{n+1}(V_{0})^{\mathfrak{g}})^{*}$, and therefore $((V_{0})^{\mathfrak{g}}/\mathfrak{m}_{e}^{n+1}(V_{0})^{\mathfrak{g}})^{*}/\mathfrak{m}_{e}((V_{0})^{\mathfrak{g}}/\mathfrak{m}_{e}^{n+1}(V_{0})^{\mathfrak{g}})^{*}\cong(\mathfrak{m}_{e}^{n}(V_{0})^{\mathfrak{g}})^{*}.$ Now as a $G_{0}$-module, $(V_{0})^{\mathfrak{g}}=\operatorname{Hom}_{\mathfrak{g}_{\overline{0}}}(\mathcal{U}\mathfrak{g},V_{0})\cong\operatorname{Hom}(\Lambda\mathfrak{g}_{\overline{1}},V_{0})\cong\Lambda\mathfrak{g}_{\overline{1}}^{*}\otimes V_{0}$ and $\mathfrak{m}_{e}^{n}V$ sits inside as the $\mathfrak{g}_{\overline{0}}$-submodule $\Pi^{n}\Lambda^{n}\mathfrak{g}_{\overline{1}}^{*}\otimes V_{0}$. Therefore, $W_{0}\cong\Pi^{n}\Lambda^{n}\mathfrak{g}_{\overline{1}}\otimes V_{0}^{*}=\operatorname{Ber}(\mathfrak{g}_{\overline{1}})\otimes V_{0}^{*}$. Putting everything together and replacing $V_{0}$ by $V_{0}^{*}$, we have reproduced the following well-known result (see for instance section 9 of [Ser11] for an algebraic proof): ###### Proposition 6.1. For a $G_{0}$-module $V_{0}$, we have a canonical isomorphism of $G$-modules $\operatorname{Ind}_{\mathfrak{g}_{\overline{0}}}^{\mathfrak{g}}V_{0}\cong\operatorname{Coind}_{\mathfrak{g}_{\overline{0}}}^{\mathfrak{g}}(\operatorname{Ber}(\mathfrak{g}_{\overline{1}})\otimes V_{0}).$ ### 6.2. The invariant differential ghost operator Now we study differential operators on $G/G_{0}$. By 3.5, the invariant differential operators are given by $G_{0}$-invariant distributions, i.e. $(\mathcal{U}\mathfrak{g}/(\mathcal{U}\mathfrak{g})\mathfrak{g}_{\overline{0}})^{G_{0}}$ which can be identified, via symmetrization, with $S(\mathfrak{g}_{\overline{1}})^{G_{0}}$. However, we also have $\operatorname{Dist}(G/G_{0},e)\cong\operatorname{Ind}_{\mathfrak{g}_{\overline{0}}}^{\mathfrak{g}}k$ as a $\mathfrak{g}$-module. Therefore by 6.1 we have $\operatorname{Dist}(G/G_{0},e)\cong\operatorname{Coind}_{\mathfrak{g}_{\overline{0}}}^{\mathfrak{g}}(\operatorname{Ber}(\mathfrak{g}_{\overline{1}}))$ Recall that $\operatorname{ber}_{\mathfrak{g}}$ is the character of $G$ determined determined by $\operatorname{Ber}(\mathfrak{g})$. If $V$ is a $G$-module, we write $V^{\operatorname{ber}_{\mathfrak{g}}}$ for the subspace of eigenvectors for $G$ of weight $\operatorname{ber}_{\mathfrak{g}}$. ###### Corollary 6.2. Suppose that $\Lambda^{top}\mathfrak{g}_{\overline{0}}$ is a trivial $G_{0}$-module. Then the subspace $\operatorname{Dist}(G/G_{0},e)^{\operatorname{ber}_{\mathfrak{g}}}$ is one- dimensional. ###### Proof. Under this assumption, $\operatorname{Ber}(\mathfrak{g})=\operatorname{Ber}(\mathfrak{g}_{\overline{1}})$, and so the result follows by Frobenius reciprocity. ∎ Observe that the conditions of 6.2 hold if $G_{0}$ is reductive. When it exists, we write $v_{\mathfrak{g}}\in\operatorname{Dist}(G/G_{0},e)^{\operatorname{ber}_{\mathfrak{g}}}$ for a chosen non-zero element. If $\beta=0$, i.e. $\operatorname{Ber}(\mathfrak{g}_{\overline{1}})$ is a trivial $G_{0}$-module, then $v_{\mathfrak{g}}\in\operatorname{Dist}(G/G_{0},e)^{G}\subseteq\operatorname{Dist}(G/G_{0},e)^{G_{0}}\cong D_{G/G_{0}}^{G}$. Therefore in this case $v_{\mathfrak{g}}$ corresponds to a $G$-invariant differential operator on $G/G_{0}$, which we write as $D_{\mathfrak{g}}$. ###### Definition 6.3. We say a supergroup $G$ is quasireductive if $G_{0}$ is reductive. We say a Lie superalgebra $\mathfrak{g}$ is quasireductive if it is the Lie superalgebra of a quasireductive supergroup. Assumption: We assume for the rest of the section that $G$ is quasireductive and $\operatorname{Ber}(\mathfrak{g})=\operatorname{Ber}(\mathfrak{g}_{\overline{1}})$ is the trivial $G_{0}$-module. Under this assumption, $v_{\mathfrak{g}}$ corresponds, as we said, to a certain $G$-invariant differential operator. We now determine what it is. ###### Lemma 6.4. $D_{G/G_{0}}=\operatorname{End}_{k}(k[G/G_{0}])$. ###### Proof. This easily follows from the definition of differential operators. ∎ Recall that since $G_{0}$ is reductive, $k[G/G_{0}]=\operatorname{Coind}_{\mathfrak{g}_{\overline{0}}}^{\mathfrak{g}}k$ is projective, and hence is a sum of injective indecomposable modules $I(L)$ for $L$ a simple $G$-module in the socle of $\operatorname{Coind}_{\mathfrak{g}_{\overline{0}}}^{\mathfrak{g}}k$. We have that the trivial module, $k$, shows up exactly once in $\operatorname{Coind}_{\mathfrak{g}_{\overline{0}}}^{\mathfrak{g}}k$, hence we can write: $k[G/G_{0}]=\operatorname{Coind}_{\mathfrak{g}_{\overline{0}}}^{\mathfrak{g}}k=I(k)\oplus V.$ Since $\operatorname{Ber}(\mathfrak{g}_{\overline{1}})$ is trivial, we have that $I(k)=P(k)$ (see [Ser11]), and thus the head and tail of $I(k)$ are both the trivial module. It follows that $\operatorname{End}(I(k))$ contains a unique up to scalar endomorphism $\phi$ taking the head to the tail, which is nilpotent exactly if $G$ does not have semisimple representations (i.e. $k$ is not projective). ###### Proposition 6.5. Up to a non-zero scalar, $D_{\mathfrak{g}}$ is the endomorphism of $k[G/G_{0}]=I(k)\oplus V$ given by $\phi\oplus 0_{V}$. ###### Proof. Since $D_{\mathfrak{g}}$ is $G$-invariant by construction, it suffices to show that $\operatorname{res}_{e}(D_{\mathfrak{g}})$ is $\mathfrak{g}$-invariant. However, we observe that $uD_{\mathfrak{g}}=D_{\mathfrak{g}}u=0$ for all $u\in\mathfrak{g}$, so since $D_{G/G_{0}}\to\operatorname{Dist}(G/G_{0},e)$ is right $D_{G/G_{0}}$-equivariant, we are done. ∎ The following is now easy to show: ###### Corollary 6.6. For a $G_{0}$-module $V_{0}$, $D_{\mathfrak{g}}$ acts on $\operatorname{Coind}_{\mathfrak{g}_{\overline{0}}}^{\mathfrak{g}}V_{0}$ by $\phi$ on each summand isomorphic to $I(k)$, and zero otherwise. We have the following characterization of $v_{\mathfrak{g}}$: ###### Corollary 6.7. Let $\mathfrak{g}$ be a quasireductive Lie superalgebra such that $\operatorname{Ber}(\mathfrak{g}_{\overline{1}})$ is the trivial $G_{0}$-module, and write $v_{\mathfrak{g}}$ for a non-zero element of $(\mathcal{U}\mathfrak{g}/\mathcal{U}\mathfrak{g}\mathfrak{g}_{\overline{0}})^{G}$. Then $v_{\mathfrak{g}}$ is the unique non-zero element up to scalar with the property that $uv_{\mathfrak{g}}\in\mathcal{U}\mathfrak{g}\mathfrak{g}_{\overline{0}}$ for all $u\in\mathfrak{g}$. ### 6.3. Relation to Gorelik’s element $v_{\emptyset}$ ###### Proposition 6.8. Let $\mathfrak{g}$ be a Lie superalgebra such that $\Lambda^{top}\mathfrak{g}_{\overline{1}}$ is the trivial $G_{0}$-module. Then for a $G_{0}$-module $V_{0}$ we have a natural isomorphism $(V_{0})^{G}\to(\operatorname{Ind}_{\mathfrak{g}_{\overline{0}}}^{\mathfrak{g}}V_{0})^{\operatorname{ber}_{\mathfrak{g}}}$ given by $z\mapsto v_{\mathfrak{g}}z.$ ###### Proof. This easily follows from the work in already done in this section. ∎ In [Gor00], it was proven that if $\operatorname{Ber}(\mathfrak{g}_{\overline{1}})$ is trivial then there exists an element $v_{\emptyset}\in\mathcal{U}\mathfrak{g}$ with the property that for a $\mathfrak{g}_{\overline{0}}$-module $V$, the map $z\mapsto v_{\emptyset}z$ defines an isomorphism $V^{\mathfrak{g}_{\overline{0}}}\to(\operatorname{Ind}_{\mathfrak{g}_{\overline{0}}}^{\mathfrak{g}}V)^{\mathfrak{g}}.$ ###### Corollary 6.9. The element $v_{\mathfrak{g}}$ agrees with the construction of such an element given by Gorelik in [Gor00]. ###### Proof. Gorelik’s element has the property that it defined a nonzero element of $(\mathcal{U}\mathfrak{g}/\mathcal{U}\mathfrak{g}\mathfrak{g}_{\overline{0}})^{\mathfrak{g}}$, so we are done. ∎ ### 6.4. Computations of $v_{\mathfrak{g}}$ for some Lie superalgebras We compute explicitly the element $v_{\mathfrak{g}}\in\left(\mathcal{U}\mathfrak{g}/\mathcal{U}\mathfrak{g}\mathfrak{g}_{\overline{0}}\right)^{\operatorname{ber}_{\mathfrak{g}}}$ for certain quasireductive Lie superalgebras. First, an easy example: ###### Lemma 6.10. If $[\mathfrak{g}_{\overline{1}},\mathfrak{g}_{\overline{1}}]$ is central in $\mathfrak{g}$, then $v_{\mathfrak{g}}=v_{1}\cdots v_{n}$ for any choice of basis $v_{1},\dots,v_{n}$ of $\mathfrak{g}_{\overline{1}}$. ###### Proof. This element is acted on by $\mathfrak{g}_{\overline{0}}$ according to its action on $\operatorname{Ber}(\mathfrak{g})$. Observe that $v_{i}v_{\mathfrak{g}}=(-1)^{i-1}v_{1}\cdots v_{i}^{2}\cdots v_{n}+\sum\limits_{1\leq j<i}(-1)^{j-1}v_{1}\dots v_{j-1}[v_{i},v_{j}]v_{j+1}\cdots v_{n}.$ Since both $[v_{j},v_{i}]$ and $v_{i}^{2}$ are central, we may rewrite this sum as $(-1)^{i-1}v_{1}\cdots\widehat{v_{i}}\cdots v_{n}v_{i}^{2}+(-1)^{j-1}v_{1}\dots v_{j-1}\widehat{v_{j}}v_{j+1}\cdots v_{n}[v_{j},v_{i}]\in\mathcal{U}\mathfrak{g}\mathfrak{g}_{\overline{0}}.$ ∎ For type I algebras, we have the following: ###### Proposition 6.11. Suppose that $\mathfrak{g}$ is a quasireductive Lie superalgebra with a $\mathbb{Z}$-grading $\mathfrak{g}=\mathfrak{g}_{-1}\oplus\mathfrak{g}_{0}\oplus\mathfrak{g}_{1}$ such that $[\mathfrak{g}_{i},\mathfrak{g}_{j}]\subseteq\mathfrak{g}_{i+j}$, $\mathfrak{g}_{\overline{0}}=\mathfrak{g}_{0}$ and $\mathfrak{g}_{\overline{1}}=\mathfrak{g}_{-1}\oplus\mathfrak{g}_{1}$. Suppose further that for an odd weight $\alpha$, of a Cartan subalgebra of $\mathfrak{g}_{\overline{0}}$, $[\mathfrak{g}_{\alpha},\mathfrak{g}_{-\alpha}]$ acts trivially on $\Lambda^{top}\mathfrak{g}_{1}$. Let $v_{1},\dots v_{n}$ be any basis of $\mathfrak{g}_{-1}$ and $w_{1},\dots,w_{m}$ any basis of $\mathfrak{g}_{1}$. Then $v_{\mathfrak{g}}=(v_{1}\cdots v_{n})\cdot(w_{1}\cdots w_{m}).$ ###### Proof. Clearly $\mathfrak{g}_{\overline{0}}$ acts on it as the top exterior power of $\mathfrak{g}_{\overline{1}}$, and $\mathfrak{g}_{-1}$ annihilates it. Therefore it remains to show that $\mathfrak{g}_{1}$ also annihilates it. Let $w\in\mathfrak{g}_{1}$ be a weight vector, of weight $\alpha$ (note that $\alpha$ could be 0). The above expression is seen to be independent of the choice of bases up to a nonzero scalar, so let us assume that $v_{1},\dots,v_{n}$ are weight vectors and that $v_{1},\dots,v_{i}$ are a basis of $\mathfrak{g}_{-\alpha}$ (and if $\mathfrak{g}_{-\alpha}=0$ then this condition is vacuous). We see that (working up to $\mathcal{U}\mathfrak{g}\mathfrak{g}_{\overline{0}}$) $\displaystyle w(v_{1}\cdots v_{n})\cdot(w_{1}\cdots w_{m})$ $\displaystyle=$ $\displaystyle\sum\limits_{j}(-1)^{j-1}v_{1}\cdots v_{j-1}[w,v_{j}]v_{j+1}\cdots v_{n}w_{1}\cdots w_{m}$ $\displaystyle=$ $\displaystyle\sum\limits_{j\leq i}(-1)^{j-1}v_{1}\cdots v_{j-1}[[w,v_{j}],v_{j+1}\cdots v_{n}w_{1}\cdots w_{m}]$ $\displaystyle+$ $\displaystyle\sum\limits_{j>i,k>j}(-1)^{j-1}v_{1}\cdots v_{j-1}v_{j+1}\cdots v_{k-1}[[w,v_{j}],v_{k}]v_{k+1}\cdots v_{n}w_{1}\cdots w_{m}$ $\displaystyle+$ $\displaystyle\sum\limits_{j>i}(-1)^{j-1}v_{1}\cdots v_{j-1}v_{j+1}\cdots v_{n}[[w,v_{j}],w_{1}\cdots w_{m}].$ In the first sum, we have that action of $[w,v_{j}]\in[\mathfrak{g}_{\alpha},\mathfrak{g}_{-\alpha}]$ on the vector $v_{j+1}\cdots v_{n}w_{1}\cdots w_{m}$ which has weight $j\alpha+\sum\limits_{\beta\in\Delta}\beta.$ By assumption, $\sum\limits_{\beta\in\Delta}\beta([\mathfrak{g}_{\alpha},\mathfrak{g}_{-\alpha}])=0$, and by the Jacobi identity we have $\alpha([\mathfrak{g}_{\alpha},\mathfrak{g}_{-\alpha}])=0$ as well. Therefore this first term is zero. For the second sum, the weight of $[[w,v_{j}],v_{k}]$ is $\alpha+\alpha_{j}+\alpha_{k}$, where $\alpha_{j}$, resp. $\alpha_{k}$ is the weight of $v_{j}$, resp. $v_{k}$. Since $\alpha\neq-\alpha_{j},-\alpha_{k}$, we have that $\alpha+\alpha_{j}+\alpha_{k}\neq\alpha_{j},\alpha_{k}$, and therefore $[[w,v_{j}],v_{k}]$ is either zero or a root vector in $\mathfrak{g}_{-1}$ which we may assume already appears in the product $v_{1}\cdots v_{j-1}v_{j+1}\cdots v_{k-1}v_{k+1}\cdots v_{n}$, giving zero. For the final sum, we know that $[w,v_{j}]$ is a nilpotent element of $\mathfrak{g}_{\overline{0}}$ and thus acts trivially on $\Lambda^{top}\mathfrak{g}_{1}$, so we once again get zero. ∎ ###### Corollary 6.12. For $\mathfrak{g}\mathfrak{l}(m|n)$, $(\mathfrak{p})\mathfrak{s}\mathfrak{l}(m|n)$, $\mathfrak{o}\mathfrak{s}\mathfrak{p}(2|2n)$, and $(\mathfrak{s})\mathfrak{p}(n)$ the formula for $v_{\mathfrak{g}}$ is given by 6.11. ###### Proof. It is straightforward to check the conditions of 6.11 for these superalgebras. ∎ ###### Proposition 6.13. Let $\pm\delta_{1},\dots,\pm\delta_{n}$ denote the odd roots of $\mathfrak{g}=\mathfrak{o}\mathfrak{s}\mathfrak{p}(1|2n)$ for the standard presentation of $\mathfrak{o}\mathfrak{s}\mathfrak{p}(1|2n)$ as given for example in [Mus12]. Choose elements $u_{1},\dots,u_{n}$ of weight $\delta_{1},\dots,\delta_{n}$ and $v_{1},\dots,v_{n}$ of weight $-\delta_{1},\dots,-\delta_{n}$ such that if $h_{i}=[u_{i},v_{i}]$, then $\delta_{i}(h_{i})=1$. Write $t_{i}=u_{i}v_{i}$. Then we have $v_{\mathfrak{g}}=(1+t_{1})(3+t_{2})\cdots((2n-1)+t_{n}).$ ###### Proof. This is in fact proven in section 4 of [DH+76]. There they prove that the above element is equal, mod $\mathcal{U}\mathfrak{g}\mathfrak{g}_{\overline{0}}$, to $(1+t_{\sigma(1)})(3+t_{\sigma(2)})\cdots(2n-1+t_{\sigma(n)}).$ for any permutation $\sigma$. Now $u_{i}(1+t_{i})=-v_{i}u_{i}^{2},\ \ \ \ \ v_{i}(1+t_{i})=-u_{i}v_{i}^{2}+v_{i}h_{i}.$ Since $u_{i}^{2},v_{i}^{2},$ and $h_{i}$ commute with $t_{j}$ for $j\neq i$, we may move them all the way to the right and obtain elements of $\mathcal{U}\mathfrak{g}\mathfrak{g}_{\overline{0}}$. ∎ It would be interesting to obtain formulas for $v_{\mathfrak{g}}$ when $\mathfrak{g}=\mathfrak{o}\mathfrak{s}\mathfrak{p}(m|2n)$ with $m>2$, $\mathfrak{g}=\mathfrak{q}(n)$, or when $\mathfrak{g}$ is exceptional simple. Explicit formulas are important in computing the image of ghost distributions under the Harish-Chandra homomorphism, which we discuss later. ### 6.5. Semisimplicity criteria We give a brief application of the ideas above. ###### Theorem 6.14. Let $G$ be a quasireductive supergroup. Then the following are equivalent: 1. (1) The category $\operatorname{Rep}(G)$ of representations of $G$ is semisimple; 2. (2) $\operatorname{Ber}(\mathfrak{g}_{\overline{1}})$ is trivial and $D_{\mathfrak{g}}$ is not nilpotent. 3. (3) $\operatorname{Ber}(\mathfrak{g}_{\overline{1}})$ is trivial and $\varepsilon(v_{\mathfrak{g}})\neq 0$, where $\varepsilon$ is the counit on $\mathcal{U}\mathfrak{g}$ (and is well-defined on $\mathcal{U}\mathfrak{g}/\mathfrak{g}_{\overline{0}}\mathcal{U}\mathfrak{g}$). Note that the condition that $G$ be quasireductive is necessary in order for $\operatorname{Rep}(G)$ to be semisimple. ###### Proof. The equivalence $(3)\iff(2)$ is clear, so we show $(2)\iff(1)$. Since $G$ is quasireductive, $k[G/G_{0}]=\operatorname{Coind}_{\mathfrak{g}_{\overline{0}}}^{\mathfrak{g}}k$ is projective. Since $\operatorname{Rep}(G)$ is semisimple if and only if the trivial module $k$ is projective, it is equivalent to show that $k$ splits off of $k[G/G_{0}]$. For this it is equivalent that $D_{\mathfrak{g}}(1)\neq 0$, which gives $(2)\iff(1)$. ∎ It is well-known, going back to a result of Djokovic and Hochschild in [DH+76], that if $G$ is a connected algebraic supergroup such that $\operatorname{Rep}(G)$ is semisimple, then $G\cong K\times SOSp(1|2n_{1})\times\dots\times SOSp(1|2n_{k})$, where $K$ is a reductive Lie group. Using 6.14 and 6.13 one can obtain a simple proof of this statement, and this has been carried out in [She20b]. ## 7\. General symmetric space $G/K$ We come back to the general case of symmetric supervarieties $G/K$. For the rest of the article we assume that $G$ is quasireductive, so that the connected component of the identity of $K$ is quasireductive, and in particular $K_{0}$ has semisimple representation theory. ### 7.1. Ghost distributions $\mathcal{A}_{G/K}$ We know by 5.7 that $\operatorname{Dist}(G/K,eK)\cong\operatorname{Ind}_{\mathfrak{k}_{0}}^{\mathfrak{k}^{\prime}}\operatorname{Dist}(G_{0}/K_{0},eK_{0})$ as $K^{\prime}$-modules. Since $K_{0}$ has semisimple representation theory, $\operatorname{Dist}(G_{0}/K_{0},eK_{0})$ is a sum of finite-dimensional $K_{0}$-modules. Hence by 6.1 we have that $\operatorname{Ind}_{\mathfrak{k}_{0}}^{\mathfrak{k}^{\prime}}\operatorname{Dist}(G_{0}/K_{0},eK_{0})\cong\operatorname{Coind}_{\mathfrak{k}_{0}}^{\mathfrak{k}^{\prime}}(\operatorname{Dist}(G_{0}/K_{0},eK_{0})\otimes\operatorname{Ber}(\mathfrak{k}^{\prime}))$ where we have used that $\mathcal{U}\mathfrak{k}^{\prime}$ is a finite rank, free left $\mathcal{U}\mathfrak{k}_{0}$-module. Now we have $\displaystyle\operatorname{Dist}(G/K,ek)^{\operatorname{ber}_{\mathfrak{k}^{\prime}}}$ $\displaystyle\cong$ $\displaystyle(\operatorname{Coind}_{\mathfrak{k}_{0}}^{\mathfrak{k}^{\prime}}(\operatorname{Dist}(G_{0}/K_{0},eK_{0})\otimes\operatorname{Ber}(\mathfrak{k}^{\prime}))^{\operatorname{ber}_{\mathfrak{k}^{\prime}}}$ $\displaystyle\cong$ $\displaystyle(\operatorname{Dist}(G_{0}/K_{0},eK_{0})\otimes\operatorname{Ber}(\mathfrak{k}^{\prime}))^{\operatorname{ber}_{\mathfrak{k}^{\prime}}}$ $\displaystyle\cong$ $\displaystyle\operatorname{Dist}(G_{0}/K_{0},eK_{0})^{K_{0}}$ Write $v_{\mathfrak{k}^{\prime}}$ for an element of $\mathcal{U}\mathfrak{k}^{\prime}$ which projects to a nonzero element of $(\mathcal{U}\mathfrak{k}^{\prime}/\mathcal{U}\mathfrak{k}^{\prime}\mathfrak{k}_{\overline{0}})^{\operatorname{ber}_{\mathfrak{k}^{\prime}}}$. Then by 6.8 we have that ###### Proposition 7.1. The isomorphism $\eta:\operatorname{Dist}(G_{0}/K_{0},eK_{0})^{K_{0}}\to\operatorname{Dist}(G/K,eK)^{\operatorname{ber}_{\mathfrak{k}^{\prime}}}$ is given by $z\mapsto v_{\mathfrak{k}^{\prime}}\cdot z.$ ###### Remark 7.2. Recall we have an identification $\operatorname{Dist}(G_{0}/K_{0},eK_{0})^{K_{0}}\cong D^{G_{0}}(G_{0}/K_{0})$, and this algebra is identified with $S(\mathfrak{a})^{W_{\mathfrak{a}}}$, where $\mathfrak{a}\subseteq\mathfrak{p}_{\overline{0}}$ is a Cartan subspace and $W_{\mathfrak{a}}$ is the little Weyl group associated to the symmetric space. Stated in terms of enveloping algebras, we have shown: ###### Corollary 7.3. We have an isomorphism $(\mathcal{U}\mathfrak{g}_{\overline{0}}/\mathcal{U}\mathfrak{g}_{\overline{0}}\mathfrak{k}_{\overline{0}})^{K_{0}}\to(\mathcal{U}\mathfrak{g}/\mathcal{U}\mathfrak{g}\mathfrak{k})^{\operatorname{ber}_{\mathfrak{k}^{\prime}}}$ given by $\phi(z)=v_{\mathfrak{k}^{\prime}}z$. ###### Definition 7.4. We define $\mathcal{A}_{G/K}$ to be $\operatorname{Dist}(G/K,eK)^{\operatorname{ber}_{\mathfrak{k}^{\prime}}}$, and refer to elements of $\mathcal{A}_{G/K}$ as ghost distributions on $G/K$. We will often use the letter $\gamma$ to denote such a distribution. ###### Remark 7.5 (Caution). In [Gor00], $\mathcal{A}$ is used to denote the $G^{\prime}$ invariants in $\operatorname{Dist}(G\times G/G,eG)$ as we will see later on. However in our notation, $\mathcal{A}_{G\times G/G}$ denotes the $\operatorname{ber}_{\mathfrak{g}^{\prime}}$ semi-invariants of $G^{\prime}$ acting on $\operatorname{Dist}(G\times G/G,eG)$. Thus these will agree only when $\operatorname{Ber}(\mathfrak{g})$ is the trivial module. In section 10, we will introduce another object, $\mathcal{A}_{\phi}$, which is a subspace of $\mathcal{U}\mathfrak{g}$ that is invariant under a certain twisted adjoint action depending on an automorphism $\phi$. For this notation, we have that $\mathcal{A}=\mathcal{A}_{\delta}$. ### 7.2. Module structure of $\mathcal{A}_{G/K}$ Write $\mathcal{Z}_{G/K}$ for $\operatorname{Dist}(G/K,eK)^{K}=(\mathcal{U}\mathfrak{g}/\mathcal{U}\mathfrak{g}\mathfrak{k})^{K}$. This is identified with the algebra of invariant differential operators on $G/K$, as explained in 3.5. ###### Proposition 7.6. We have a natural map $\mathcal{A}_{G/K}\otimes\mathcal{Z}_{G/K}=(\mathcal{U}\mathfrak{g}/\mathcal{U}\mathfrak{g}\mathfrak{k})^{\operatorname{ber}_{\mathfrak{k}^{\prime}}}\otimes(\mathcal{U}\mathfrak{g}/\mathcal{U}\mathfrak{g}\mathfrak{k})^{K}\to(\mathcal{U}\mathfrak{g}/\mathcal{U}\mathfrak{g}\mathfrak{k})^{\operatorname{ber}_{\mathfrak{k}^{\prime}}}=\mathcal{A}_{G/K}$ making $\mathcal{A}_{G/K}$ into a right module over $\mathcal{Z}_{G/K}$. ###### Proof. We define $(\gamma+\mathcal{U}\mathfrak{g}\mathfrak{k})(z+\mathcal{U}\mathfrak{g}\mathfrak{k}):=\gamma z+\mathcal{U}\mathfrak{g}\mathfrak{k}$ Since $z$ is $K$-invariant, it is easy to check this is well-defined. ∎ ###### Lemma 7.7. Suppose that $\operatorname{Ber}(\mathfrak{k}_{\overline{1}})$ and $\operatorname{Ber}(\mathfrak{p}_{\overline{1}})$ are trivial $K_{0}$-modules. Then we have a natural map $\mathcal{A}_{G/K}\otimes\mathcal{A}_{G/K^{\prime}}\to\mathcal{Z}_{G/K}$, or $(\mathcal{U}\mathfrak{g}/\mathcal{U}\mathfrak{g}\mathfrak{k})^{K^{\prime}}\otimes(\mathcal{U}\mathfrak{g}/\mathcal{U}\mathfrak{g}\mathfrak{k}^{\prime})^{K}\to(\mathcal{U}\mathfrak{g}/\mathfrak{k}\mathcal{U}\mathfrak{g})^{K},$ given by $(\gamma+\mathcal{U}\mathfrak{g}\mathfrak{k})\otimes(\gamma^{\prime}+\mathcal{U}\mathfrak{g}\mathfrak{k}^{\prime})\mapsto\gamma\gamma^{\prime}+\mathfrak{k}\mathcal{U}\mathfrak{g}.$ ###### Proof. The proof is straightforward. ∎ ### 7.3. $k[G/K]$ as a $K^{\prime}$-module Observe that via pullback and the isomorphism of distributions we have the commutative diagram $\textstyle{\operatorname{Dist}(G/K,eK)^{\operatorname{ber}_{\mathfrak{k}^{\prime}}}\otimes k[G/K]\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\textstyle{\operatorname{Dist}((G_{0}/K_{0})^{\mathfrak{k}^{\prime}},eK_{0})^{\operatorname{ber}_{\mathfrak{k}^{\prime}}}\otimes k[(G_{0}/K_{0})^{\mathfrak{k}^{\prime}}]\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\textstyle{k.}$ Now let $\gamma$ be an element of $\operatorname{Dist}((G_{0}/K_{0})^{\mathfrak{k}^{\prime}},eK_{0})^{\operatorname{ber}_{\mathfrak{k}^{\prime}}}$, and suppose that $\gamma(f)\neq 0$ for $f\in k[(G_{0}/K_{0})^{\mathfrak{k}^{\prime}}]$. Then necessarily the $K^{\prime}$-module generated by $f$ generates a copy of $I_{K^{\prime}}(k)$, the injective hull of $k$ for $K^{\prime}$. It follows that the same must be true for the space $G/K$, and so we have: ###### Proposition 7.8. Let $\gamma\in\mathcal{A}_{G/K}$, and suppose that $f\in k[G/K]$ is such that $\gamma(f)\neq 0$. Then the $K^{\prime}$-module generated by $f$ contains a copy of $I_{K^{\prime}}(k)$. In particular, if $\mathfrak{k}_{\overline{0}}\neq\mathfrak{g}_{\overline{0}}$ then $k[G/K]$ contains $I_{K^{\prime}}(k)$ with infinite multiplicity. ###### Proof. For the last statement, we observe that since $G/K$ is affine the pairing of distributions with functions is nondegenerate. Using the fact that the $K^{\prime}$-semi-invariant distributions form an infinite-dimensional vector space (being isomorphic, as a vector space, to $S(\mathfrak{a})^{W_{\mathfrak{a}}}$), it is not hard to prove that the multiplicity must be infinite. ∎ ###### Remark 7.9. We observe that since $k[G_{0}/K_{0}]^{K_{0}}$ is a subalgebra of $k[G_{0}/K_{0}]$, $A:=\operatorname{Coind}_{\mathfrak{k}_{\overline{0}}}^{\mathfrak{k}^{\prime}}k[G_{0}/K_{0}]^{K_{0}}$ is a subalgebra of $k[(G_{0}/K_{0})^{\mathfrak{k}^{\prime}}]$. In particular $A$ is the sum of all copies of $I_{K^{\prime}}(k)$ appearing in $k[(G_{0}/K_{0})^{\mathfrak{k}^{\prime}}]$. It follows that $k[G/K]\cap A$ is a subalgebra of $k[G/K]$ which contains all copies of $I_{K^{\prime}}(k)$ in $k[G/K]$, as well as all of $k[G/K]^{K^{\prime}}$. The author has a rather limited understanding of $A\cap k[G/K]$. Further we do not have a good answer, or even good formulation of the question of ‘how many’ copies of $I_{K^{\prime}}(k)$ are within $k[G/K]$. The copies of $I_{K^{\prime}}(k)$ in $k[(G_{0}/K_{0})^{\mathfrak{k}^{\prime}}]$ may be indexed by a basis of $k[G_{0}/K_{0}]^{K_{0}}$, which itself is indexed by the irreducible summands of $k[G_{0}/K_{0}]$, i.e. by certain dominant weights in $\mathfrak{a}$. For each copy of $I_{K^{\prime}}(k)$ in $k[G/K]$, one could record which dominant weights of $\mathfrak{a}$ it is supported on in $k[(G_{0}/K_{0})^{\mathfrak{k}^{\prime}}]$. One could then look at all weights that appear in the supports of such copies of $I_{K^{\prime}}(k)$ in $k[G/K]$. This collection of weights would be infinite and could not, for example, lie within any hyperplane. We also observe that since each copy of $I_{K^{\prime}}(k)$ contains a $K^{\prime}$-invariant function, we can deduce there are many $K^{\prime}$-invariant functions on $k[G/K]$ (recall we already know from 5.6 that $k[G/K]^{K^{\prime}}$ is an integral domain). Again the structure of $K[G/K]^{K^{\prime}}$ is not generally understood, although it has been partially computed in certain examples, i.e. $G\times G/G$ and the superspheres $OSP(m|2n)/OSP(m-1|2n)$. ###### Remark 7.10. Suppose again the $\mathfrak{k}_{\overline{0}}\neq\mathfrak{g}_{\overline{0}}$. Following a similar argument to the one above, one may deduce the existence of many projective $K^{\prime}$-submodules of $k[G/K]$ as follows: given an irreducible $K^{\prime}$-submodule $L$ of $\operatorname{Dist}(G/K,eK)$, it defines an irreducible submodule of $\operatorname{Dist}((G_{0}/K_{0})^{\mathfrak{k}^{\prime}},eK_{0})$ via our isomorphism. Thus $L$ must pair nontrivially with some projective indecomposable summand $P=P_{K^{\prime}}(V)$ of $k[(G_{0}/K_{0})^{\mathfrak{k}^{\prime}}]$. However this is only possible if $V\cong L^{*}$, and $L$ pairs with $P$ via $L\otimes P\to L\otimes L^{*}\to k$. It follows that $k[G/K]$ must contain a copy of $P_{K^{\prime}}(L^{*})$ as well. ## 8\. Pairs that have an Iwasawa decomposition ### 8.1. The Iwasawa decomposition We continue to assume $G$ is quasireductive. Due to technical difficulties with Lie superalgebras like $\mathfrak{q}(n)$, which have Cartan subalgebras that are not purely even, we will from now on also assume $G$ is Cartan-even. ###### Definition 8.1. A quasireductive supergroup $G$ is Cartan-even if for a Cartan subalgebra $\mathfrak{h}\subseteq\mathfrak{g}$ we have $\mathfrak{h}=\mathfrak{h}_{\overline{0}}$ (i.e. $\mathfrak{h}_{\overline{0}}$ is self-centralizing in $\mathfrak{g}$). Let $(\mathfrak{g},\mathfrak{k})$ be a supersymmetric pair, and let $\mathfrak{a}\subseteq\mathfrak{p}_{\overline{0}}$ be a Cartan subspace, i.e. a maximal subalgebra of $\mathfrak{p}_{\overline{0}}$ consisting only of semisimple elements. Then we may decompose $\mathfrak{g}$ into weight spaces according to the action of $\mathfrak{a}$ as $\mathfrak{g}=\mathfrak{m}\oplus\bigoplus\limits_{\overline{\alpha}\in\mathfrak{a}^{*}}\mathfrak{g}_{\overline{\alpha}}$ where $\overline{\alpha}\neq 0$ and $\mathfrak{m}$ is the centralizer of $\mathfrak{a}$ in $\mathfrak{g}$. We write $\overline{\Delta}$ for the set of nonzero weights of this action, and call them restricted roots. Note that the weights of the action are exactly the restriction to $\mathfrak{a}$ of the roots under the action of a maximal torus in $\mathfrak{g}_{\overline{0}}$ which contains $\mathfrak{a}$. We say that the pair $(\mathfrak{g},\mathfrak{k})$ admits an Iwasawa decomposition if there is some choice of positive roots in $\overline{\Delta}$, $\overline{\Delta}=\overline{\Delta}^{+}\sqcup\overline{\Delta^{-}}$, such that $\text{if }\ \mathfrak{n}=\bigoplus\limits_{\alpha\in\overline{\Delta}^{+}}\mathfrak{g}_{\alpha},\ \text{ then }\ \mathfrak{g}=\mathfrak{k}\oplus\mathfrak{a}\oplus\mathfrak{n}.$ In this case, we may extend the choice of positive restricted roots to a choice of positive roots $\Delta=\Delta^{+}\sqcup\Delta^{-}$ for $\mathfrak{g}$. In that case, if $\mathfrak{b}$ is the Borel subalgebra determined by this choice of positive roots then we have $\mathfrak{a}\oplus\mathfrak{n}\subseteq\mathfrak{b}$. In particular, $\mathfrak{b}+\mathfrak{k}=\mathfrak{g}$, i.e. $\mathfrak{k}$ has a complimentary Borel subalgebra. ###### Definition 8.2. We call a Borel subalgebra arising in this way an Iwasawa Borel subalgebra for $\mathfrak{k}$ with respect to $\mathfrak{a}$. If $B\subseteq G$ integrates $\mathfrak{b}$, then we call $B$ an Iwasawa Borel subgroup of $G$. ### 8.2. $G/K$ is spherical; the rational functions $f_{\lambda}$ Recall that a $G$-supervariety $X$ is spherical if $X$ admits an open $B$-orbit for some Borel $B$. For $G/K$, this is equivalent to $\mathfrak{k}$ admitting a complementary Borel subalgebra. Hence if $\mathfrak{g}$ admits an Iwasawa decomposition, $G/K$ is spherical. Choose a Cartan subspace $\mathfrak{a}$ and let $B$ be an Iwasawa Borel subgroup with respect to $\mathfrak{a}$. In this case, $G_{0}/K_{0}$ is spherical with respect to $B_{0}$. Write $\Lambda_{0}^{+}\subseteq\mathfrak{a}^{*}$ for the dominant weights that appear as $B_{0}$-highest weights in $k[G_{0}/K_{0}]$. Then the lattice generated by $\Lambda_{0}^{+}$ is a full rank lattice in $\mathfrak{a}^{*}$ (see, for instance, [Tim11]). By the general theory of spherical supervarieties (see [She19]), the $B$-highest weights in $k[G/K]$, which we will call $\Lambda^{+}$, are a subset of $\Lambda_{0}^{+}$, and there is at most one function $f_{\lambda}\in k[G/K]$ of highest weight $\lambda$ for any $\lambda\in\Lambda_{0}^{+}$. Further, $f_{\lambda}$ is even and non-nilpotent, hence it maps down to a non- zero function of highest weight $\lambda$ on $G_{0}/K_{0}$. Further, $\Lambda^{+}$ generates the same lattice as $\Lambda_{0}^{+}$ in $\mathfrak{a}^{*}$, which we call $\Lambda$, and in particular is Zariski dense. It follows that by inverting the functions $f_{\lambda}$ for $\lambda\in\Lambda^{+}$ and taking arbitrary products, we obtain rational functions $f_{\lambda}$ for all $\lambda\in\Lambda$ which are $\mathfrak{a}\oplus\mathfrak{n}$-eigenfunctions and are regular in a neighborhood of $eK$. ### 8.3. A different perspective via induced spaces We may gain another perspective on the sphericity of $G/K$ using the notion of induced spaces. Write $A$ and $N$ for the subgroups of $G$ integrating $\mathfrak{a}$ and $\mathfrak{n}$, respectively. Then $AN$ is a subgroup of $G$ which integrates $\mathfrak{a}\oplus\mathfrak{n}$. Since $(\mathfrak{a}\oplus\mathfrak{n})_{\overline{1}}$ is complimentary to $\mathfrak{k}_{\overline{1}}$ by the Iwasawa decomposition, we may apply 5.1 to obtain the existence of a canonical map of $AN$-supervarieties $(G_{0}/K_{0})^{\mathfrak{a}\oplus\mathfrak{n}}\to G/K$ which is an isomorphism in a neighborhood of $eK_{0}$. In particular the map on functions $k[G/K]\to k[(G_{0}/K_{0})^{\mathfrak{a}\oplus\mathfrak{n}}]$ is injective map of $AN$-modules. It follows that we have an injective morphism $k[G/K]^{N}\to k[(G_{0}/K_{0})^{\mathfrak{a}\oplus\mathfrak{n}}]^{N},$ and the right hand side is isomorphic to the subalgebra of highest weight vectors in $k[G_{0}/K_{0}]$. Further, in a neighborhood of the $eK_{0}$ in $(G_{0}/K_{0})^{\mathfrak{a}\oplus\mathfrak{n}}$, we have functions $f_{\lambda}$ annihilated by $\mathfrak{n}$ and of weight $\lambda\in\mathfrak{a}^{*}$ for every $\lambda\in\Lambda$. Using the local isomorphism, we find that such functions also exist in a neighborhood of $eK$ in $G/K$, and in particular they are the highest weight functions of weight $\lambda\in\Lambda$. . ###### Definition 8.3. In the setup above, for $\lambda\in\Lambda$ we let $f_{\lambda}$ denote the rational $\mathfrak{a}\oplus\mathfrak{n}$-eigenfunction on $G/K$ such that $f_{\lambda}(eK)=1$. Note we can always normalize as above, because the $\mathfrak{a}\oplus\mathfrak{n}$ eigenvectors are all non-nilpotent and if they vanished at $eK$ then necessarily they would vanish everywhere. ### 8.4. Highest weight submodules of $k[G/K]$ Recall that for a homogeneous vector $v$ we write $\langle K^{\prime}\cdot v\rangle$ for the $K^{\prime}$-module generated by $v$. ###### Lemma 8.4. If $\lambda\in\Lambda^{+}$, then $\mathcal{U}\mathfrak{k}^{\prime}f_{\lambda}$ is stable under $K^{\prime}$, and thus $\langle K^{\prime}\cdot f_{\lambda}\rangle=\mathcal{U}\mathfrak{k}^{\prime}f_{\lambda}.$ ###### Proof. It suffices to prove that $\mathcal{U}\mathfrak{k}^{\prime}f_{\lambda}$ is stable under $K_{0}$. For this we notice that by the Iwasawa decomposition for $\mathfrak{g}_{\overline{0}}$ and the fact that $G_{0}$ is connected, $\langle G_{0}\cdot f_{\lambda}\rangle=\mathcal{U}\mathfrak{g}_{\overline{0}}f_{\lambda}=\mathcal{U}\mathfrak{k}_{\overline{0}}f_{\lambda}\subseteq\langle K_{0}\cdot f_{\lambda}\rangle.$ Since $\langle K_{0}\cdot f_{\lambda}\rangle\subseteq\langle G_{0}\cdot f_{\lambda}\rangle$ it follows that $\langle K_{0}\cdot f_{\lambda}\rangle=\mathcal{U}\mathfrak{k}_{\overline{0}}f_{\lambda}$. From here it is easy to check that $\mathcal{U}\mathfrak{k}^{\prime}f_{\lambda}$ is $K_{0}$-stable. ∎ ###### Lemma 8.5. For $\lambda\in\Lambda^{+}$, $\langle K^{\prime}\cdot f_{\lambda}\rangle$ contains at most one summand isomorphic to $I_{K^{\prime}}(k)$. ###### Proof. By the classical Iwasawa decomposition, we have $\langle K_{0}\cdot f_{\lambda}\rangle=\mathcal{U}\mathfrak{g}_{\overline{0}}f_{\lambda}\cong L_{0}(\lambda)$, since $f_{\lambda}$ is a $B_{0}$-highest weight vector. Hence we have a surjection of $K^{\prime}$-modules $\operatorname{Ind}_{\mathfrak{k}_{0}}^{\mathfrak{k}^{\prime}}L_{0}(\lambda)\to\langle K^{\prime}\cdot f_{\lambda}\rangle.$ Each copy of $I_{K^{\prime}}(k)$ showing up in $\langle K^{\prime}\cdot f_{\lambda}\rangle$ would have to split disjointly back into $\operatorname{Ind}_{\mathfrak{k}_{0}}^{\mathfrak{k}^{\prime}}L_{0}(\lambda)$. However, since $K_{0}$ is a spherical subgroup of $G_{0}$, $L_{0}(\lambda)^{K_{0}}$ is one-dimensional, so the induced module has only one copy of $I_{K^{\prime}}(k)$. ∎ ###### Remark 8.6. If $(\mathfrak{g},\mathfrak{k}^{\prime})$ also satisfies the Iwasawa decomposition then we have $\langle K^{\prime}\cdot f_{\lambda}\rangle=\langle G\cdot f_{\lambda}\rangle$. Also, it will be shown in the subsequent article that if $\mathfrak{g}$ is basic classical and the involution $\theta$ preserves the nondegenerate form on $\mathfrak{g}$, then we always have $\langle K^{\prime}\cdot f_{\lambda}\rangle=\langle G\cdot f_{\lambda}\rangle$. ### 8.5. Harish-Chandra Homomorphism The Iwasawa decomposition $\mathfrak{g}=\mathfrak{k}\oplus\mathfrak{a}\oplus\mathfrak{n}$ implies that $\mathcal{U}\mathfrak{g}=S(\mathfrak{a})\oplus(\mathfrak{n}\mathcal{U}\mathfrak{g}+\mathcal{U}\mathfrak{g}\mathfrak{k})$. Thus we have a vector space isomorphism $\operatorname{Dist}(G/K,eK)\cong\mathcal{U}\mathfrak{g}/\mathcal{U}\mathfrak{g}\mathfrak{k}\cong S(\mathfrak{a})\oplus\mathfrak{n}\mathcal{U}\mathfrak{g}/(\mathfrak{n}\mathcal{U}\mathfrak{g}\cap\mathcal{U}\mathfrak{g}\mathfrak{k}).$ ###### Definition 8.7. The Harish-Chandra morphism $HC:\operatorname{Dist}(G/K,eK)\to S(\mathfrak{a})$ is given by the projection onto $S(\mathfrak{a})$ along $\mathfrak{n}\mathcal{U}\mathfrak{g}/(\mathfrak{n}\mathcal{U}\mathfrak{g}\cap\mathcal{U}\mathfrak{g}\mathfrak{k})$. Choose a homogeneous basis $p_{1},\dots,p_{r}$ of $\mathfrak{p}$. Recall from lemma 2.9 that $\operatorname{Dist}(G/K,eK)$ is spanned by monomials $p_{1}^{k_{1}}\cdots p_{r}^{k_{r}},$ where we abusively identify them with their restrictions to $\operatorname{Dist}(G/K,eK)$. Define a filtration $F^{\bullet}$ on $\operatorname{Dist}(G/K,eK)$ by setting $F^{0}=\langle\operatorname{ev}_{eK}\rangle$, and set $F^{i}$ to be the span of all monomials as above where the degree of $p_{i}$ is 2 if $p_{i}$ is even and 1 if $p_{i}$ is odd. The following is a generalization of 4.2.2 in [Gor00]. ###### Lemma 8.8. $HC(F^{r})\subseteq\sum\limits_{j\leq r/2}S^{j}\mathfrak{a}$. ###### Proof. Since the Harish-Chandra projection is linear, it suffices to prove this on monomials. We induct on the length of the monomial. If the monomial is length zero, the result is clear. Suppose we have a monomial $b_{1}\cdots b_{t}\in F^{r}$. Using the Iwasawa decomposition, we may write $b_{1}=k+a+n$ where $k\in\mathfrak{k}$, $a\in\mathfrak{a}$, and $n\in\mathfrak{n}$. Since $nb_{2}\cdots b_{t}\in\mathfrak{n}\mathcal{U}\mathfrak{g}$ it vanishes under the Harish-Chandra projection, so we have $\displaystyle HC(b_{1}\dots b_{t})$ $\displaystyle=$ $\displaystyle HC(ab_{2}\cdots b_{t})+HC(kb_{2}\cdots b_{t})$ $\displaystyle=$ $\displaystyle HC(ab_{2}\cdots b_{t})+\sum\limits_{i}HC(b_{2}\cdots[k,b_{i}]\cdots b_{t})+HC(b_{2}\cdots b_{t}k).$ Since $b_{2}\cdots b_{t}k\in\mathcal{U}\mathfrak{g}\mathfrak{k}$, the last term vanishes. Now notes that $a\neq 0$ only if $b_{t}$ is even, and in that case $b_{2}\cdots b_{t}\in F^{r-2}$, and $HC(ab_{2}\cdots b_{t})=HC(a)HC(b_{2}\cdots b_{t})$. Thus by induction, $\deg HC(ab_{2}\cdots b_{t})\leq\deg HC(b_{2}\cdots b_{t})+1\leq\frac{r-2}{2}+1=\frac{r}{2}.$ As for $HC(b_{2}\cdots[k,b_{i}]\cdots b_{t})$, if $b_{1}$ is even then the parity of $[k,b_{i}]$ is the same as the parity of $b_{i}$, and thus $b_{2}\cdots[k,b_{i}]\cdots b_{t}\in F^{r-2}$. Therefore $\deg HC(b_{2}\cdots[k,b_{i}]\cdots b_{t})\leq(r-2)/2\leq r/2.$ If $b_{1}$ is odd, then the parity of $[k,b_{i}]$ is opposite the parity of $b_{i}$, so $b_{2}\cdots[k,b_{i}]\cdots b_{t}\in F^{r}$, and by induction we have $HC(b_{2}\cdots[k,b_{i}]\cdots b_{t})\leq r/2.$ ∎ ###### Corollary 8.9. Let $z\in\operatorname{Dist}^{r}(G_{0}/K_{0},eK_{0})^{K_{0}}$ lie in the $r$th part of the standard filtration on $\operatorname{Dist}(G_{0}/K_{0},eK_{0})$ defined in definition 2.2. Then $v_{\mathfrak{k}^{\prime}}\cdot z\in\mathcal{A}_{G/K}$ has $HC(v_{\mathfrak{k}^{\prime}}\cdot z)\leq r+\operatorname{dim}\mathfrak{p}_{\overline{1}}/2.$ ###### Proof. We have that $z\in F^{2r}$ and $v_{\mathfrak{k}^{\prime}}\in F^{\operatorname{dim}\mathfrak{p}_{\overline{1}}}$, so $v_{\mathfrak{k}^{\prime}}\cdot z=zv_{\mathfrak{k}^{\prime}}\in F^{2r+\operatorname{dim}\mathfrak{p}_{\overline{1}}}$. ∎ ###### Lemma 8.10. $HC(\mathcal{A}_{G/K})$ is naturally a module over $HC(\mathcal{Z}_{G/K})$ such that $HC:\mathcal{A}_{G/K}\to S(\mathfrak{a})$ induces a morphism of $\mathcal{Z}_{G/K}$-modules. ###### Proof. Let $\gamma\in\mathcal{A}_{G/K}$ and $z\in\mathcal{Z}_{G/K}$. Write $z=n+HC(z)+\mathcal{U}\mathfrak{g}\mathfrak{k}$ and $\gamma=n^{\prime}+HC(\gamma)+\mathcal{U}\mathfrak{g}\mathfrak{k}$. Then we see that $\gamma z=(n^{\prime}+HC(\gamma)+\mathcal{U}\mathfrak{g}\mathfrak{k})(n+HC(z)+\mathcal{U}\mathfrak{g}\mathfrak{k})=n^{\prime}(n+HC(z))+HC(\gamma)n+HC(z)HC(\gamma)+\mathcal{U}\mathfrak{g}\mathfrak{k}.$ Clearly $n^{\prime}(n+HC(z))\in\mathfrak{n}\mathcal{U}\mathfrak{g}$. And since $HC(\gamma)\in S(\mathfrak{a})$, it preserves $\mathfrak{n}\mathcal{U}\mathfrak{g}$ under commutator and thus $HC(\gamma)n\in\mathfrak{n}\mathcal{U}\mathfrak{g}$ as well. Hence we find that $HC(\gamma z)=HC(\gamma)HC(z),$ as desired. ∎ ###### Lemma 8.11. For $\gamma\in\operatorname{Dist}(G/K,eK)$ and $\lambda\in\Lambda$ we have that $HC(\gamma)(\lambda)=\gamma(f_{\lambda})$. ###### Proof. Since $\mathcal{U}\mathfrak{g}\mathfrak{n}$ annihilates $f_{\lambda}$, the result follows from our normalization of $f_{\lambda}$. ∎ The following illustrates the importance of understanding $HC(\mathcal{A}_{G/K})$, especially in light of remark 8.6. ###### Corollary 8.12. Let $\lambda\in\Lambda^{+}$. Then $\langle K^{\prime}\cdot f_{\lambda}\rangle$ contains a copy of $I_{K^{\prime}}(k)$ if and only if there exists $\gamma\in\mathcal{A}_{G/K}$ such that $HC(\gamma)(\lambda)=\gamma(f_{\lambda})\neq 0$. ###### Proof. These results follow from the work in section 7.3. ∎ For the next lemma, let $\mathfrak{h}\subseteq\mathfrak{g}$ be a $\theta$-stable Cartan subalgebra which contains $\mathfrak{a}$, so that $\mathfrak{h}=\mathfrak{t}\oplus\mathfrak{a}$, where $\mathfrak{t}$ are the fixed points of $\theta$ in $\mathfrak{h}$. In particular, $\mathfrak{t}\subseteq\mathfrak{k}_{\overline{0}}$. ###### Proposition 8.13. $HC(\mathcal{A}_{G/K})=0$ if one of the following two conditions holds. 1. (1) $\operatorname{dim}\mathfrak{p}_{\overline{1}}$ is odd; 2. (2) The restriction of $\operatorname{ber}_{\mathfrak{k}^{\prime}}$ to $\mathfrak{t}$ is non-zero. Note this holds if $\operatorname{ber}_{\mathfrak{k}^{\prime}}\neq 0$ and the map $\mathfrak{t}\to\mathfrak{k}^{\prime}/[\mathfrak{k}^{\prime},\mathfrak{k}^{\prime}]$ is surjective. ###### Proof. If $\mathfrak{p}_{\overline{1}}$ is odd-dimensional then $v_{\mathfrak{k}^{\prime}}$ is odd, and thus all elements of $\mathcal{A}_{G/K}$ are odd. Since $S(\mathfrak{a})$ is purely even and $HC$ is an even map, we must have $HC(\mathcal{A}_{G/K})=0$. Now suppose condition (2) holds. Let $\gamma\in\mathcal{A}_{G/K}=\operatorname{Dist}(G/K,eK)^{\operatorname{ber}_{\mathfrak{k}^{\prime}}}$, and let $\lambda\in\Lambda^{+}$. Then since $\lambda\in\mathfrak{a}^{*}$, $\lambda(\mathfrak{t})=0$. If $t\in\mathfrak{t}$ we see that $\operatorname{ber}_{\mathfrak{k}^{\prime}}(t)\gamma(f_{\lambda})=(t\cdot\gamma)(f_{\lambda})=\gamma tf_{\lambda}=\lambda(t)\gamma(f_{\lambda})=0.$ However by assumption there exists $t\in\mathfrak{t}$ such that $\operatorname{ber}_{\mathfrak{k}^{\prime}}(t)\neq 0$, so if we use this $t$ we find that $\gamma(f_{\lambda})=0$ for all $\lambda$, so that $HC(\gamma)=0$. ∎ ### 8.6. The polynomial $p_{G/K,B}$ We now interpret the polynomial $p_{G/K,B}:=HC(\operatorname{ev}_{eK}v_{\mathfrak{k}^{\prime}})$, noting that it depends on our Iwasawa Borel subgroup $B$. Recall that $\operatorname{ev}_{eK}v_{\mathfrak{k}^{\prime}}=v_{\mathfrak{k}^{\prime}}\cdot\operatorname{ev}_{eK}$ is a ghost distribution, and is the one of minimal degree. Observe that we have a $K^{\prime}$-equivariant embedding $i:K^{\prime}/K_{0}\to G/K.$ This induces a $K^{\prime}$-equivariant surjective morphism on functions $i^{*}:k[G/K]\to k[K^{\prime}/K_{0}].$ Write $\operatorname{ev}_{eK}$ for the evaluation distribution at $eK$ on $G/K$. Then on $K^{\prime}$ distributions we obtain a map which sends $v_{\mathfrak{k}^{\prime}}\mapsto v_{\mathfrak{k}^{\prime}}\cdot\operatorname{ev}_{eK}=\operatorname{ev}_{eK}v_{\mathfrak{k}^{\prime}}.$ Thus for $\lambda\in\Lambda^{+}$ we have $p_{G/K,B}(\lambda)=(\operatorname{ev}_{eK}v_{\mathfrak{k}^{\prime}})f_{\lambda}=v_{\mathfrak{k}^{\prime}}i^{*}(f_{\lambda})=\operatorname{ev}_{eK_{0}}\circ D_{\mathfrak{k}^{\prime}}\circ i^{*}(f_{\lambda}).$ It follows that: ###### Lemma 8.14. For $\lambda\in\Lambda^{+}$, $p_{G/K,B}(\lambda)\neq 0$ if and only if $\langle K^{\prime}\cdot i^{*}f_{\lambda}\rangle$ contains $I_{K^{\prime}}(k)$. ###### Corollary 8.15. If $\lambda\in\Lambda^{+}$ and $\langle K^{\prime}\cdot f_{\lambda}\rangle$ contains a copy of $I_{K^{\prime}}(k)$, then $p_{G/K,B}(\lambda)\neq 0$ if and only if the $K^{\prime}$-invariant in the copy of $I_{K^{\prime}}(k)$ is non- zero at $eK$. We have now shown that: ###### Corollary 8.16. We have $p_{G/K,B}(\lambda)\neq 0$ if and only if $\langle K^{\prime}\cdot f_{\lambda}\rangle$ contains a copy of $I_{K^{\prime}}(k)$ and the $K^{\prime}$-invariant of $I_{K^{\prime}}(k)$ does not vanish at $eK$. The following gives one reason for the importance of the polynomial $p_{G/K,B}$. ###### Proposition 8.17. Suppose that whenever $\langle K^{\prime}\cdot f_{\lambda}\rangle$ contains a copy of $I_{K^{\prime}}(k)$ for $\lambda\in\Lambda^{+}$, the $K^{\prime}$-invariant in it does not vanish at $eK$. Then, 1. (1) For $\lambda\in\Lambda^{+}$, $p_{G/K,B}(\lambda)\neq 0$ if and only if $\langle K^{\prime}\cdot f_{\lambda}\rangle$ contains a copy of $I_{K^{\prime}}(k)$. 2. (2) For any $\lambda\in\Lambda^{+}$, if $p_{G/K,B}(\lambda)=0$, then $HC(\gamma)(\lambda)=0$ for all $\gamma\in\mathcal{A}_{G/K}$. ###### Proof. (1) follows from 8.16. (2) follows from 8.12. ∎ ### 8.7. Relationship between $G/K$ and $G/K^{\prime}$ We continue to suppose that $(\mathfrak{g},\mathfrak{k})$ satisfies the Iwasawa decomposition. ###### Proposition 8.18. For $\lambda\in\Lambda^{+}$, suppose that 1. (1) $\langle K^{\prime}\cdot f_{\lambda}\rangle$ contains a copy of $I_{K^{\prime}}(k)$ (equivalently $HC(\gamma)(\lambda)\neq 0$ for some $\gamma\in\mathcal{A}_{G/K}$); and 2. (2) $\langle G\cdot f_{\lambda}\rangle\cong L(\lambda)$ is an irreducible $G$-module. Then $I_{G}(L(\lambda))$ is a submodule of $k[G]^{\operatorname{ber}_{\mathfrak{k}^{\prime}}}$. Note that $k[G]^{\operatorname{ber}_{\mathfrak{k}^{\prime}}}$ is equal to the induced bundle $\operatorname{Ind}_{K^{\prime}}^{G}\operatorname{Ber}(\mathfrak{k}^{\prime})$, and if $\operatorname{Ber}(\mathfrak{k}^{\prime})$ is trivial then $k[G]^{\operatorname{ber}_{\mathfrak{k}^{\prime}}}=k[G/K^{\prime}]$. ###### Proof. Since $I_{K^{\prime}}(k)\cong P_{K^{\prime}}(\operatorname{Ber}(\mathfrak{k}^{\prime}))$, $I_{K^{\prime}}(k)$ has a morphism $\varphi:I_{K^{\prime}}(k)\to\operatorname{Ber}(\mathfrak{k}^{\prime})$ determined by the projection onto its head. Since $I_{K^{\prime}}(k)$ splits off $L(\lambda)$ as a $K^{\prime}$-module, it also must split off $I_{G}(L(\lambda))$ as a $K^{\prime}$-module. Therefore we may extend $\varphi$ to a $K^{\prime}$-equivariant morphism $\phi:I_{G}(L(\lambda))\to\operatorname{Ber}(\mathfrak{k}^{\prime})$. By construction $\phi$ is non-zero on $L(\lambda)$, the socle of $I_{G}(L(\lambda))$. Thus by Frobenius reciprocity, $\phi$ defines an injective morphism of $G$-modules $\Phi:I_{G}(L(\lambda))\to k[G]^{\operatorname{ber}_{\mathfrak{k}^{\prime}}}.$ ∎ ###### Remark 8.19. The above theorem is especially useful when $G/K\cong G/K^{\prime}$, as it helps to determine the structure of $k[G/K]$ as a $G$-module. We will see one nice application of this in the next section. ## 9\. The Group Case $G\times G/G$ We now consider the case of the supersymmetric space $G\times G/G$, where $G$ is embedded diagonally. This space is isomorphic to $G$ with the $G\times G$ action given by left and right translation. Since $G$ is assumed to be quasireductive, it must be connected in particular, so we can work with the Lie superalgebra without losing much. The involution we take in this case, $\theta$, is given on the Lie superalgebra $\mathfrak{g}\times\mathfrak{g}$ by $\theta(x,y)=(y,x)$. ###### Lemma 9.1. We have a $\mathfrak{g}\times\mathfrak{g}$-module isomorphism $\mathcal{U}\mathfrak{g}\cong\operatorname{Dist}(G\times G/G,eG)=\mathcal{U}(\mathfrak{g}\times\mathfrak{g})/\mathcal{U}(\mathfrak{g}\times\mathfrak{g})\mathfrak{g}$ given by $u\mapsto u\otimes 1.$ ###### Proof. Let $(v_{1},v_{2})\in\mathfrak{g}\times\mathfrak{g}$ and $u\in\mathcal{U}\mathfrak{g}$. We see that $(v_{1},v_{2})\cdot u=v_{1}u-(-1)^{\overline{u}\overline{v_{2}}}uv_{2}$, and this will map to (before modding out by $\mathcal{U}(\mathfrak{g}\times\mathfrak{g})\mathfrak{g}$) $\displaystyle v_{1}u\otimes 1-(-1)^{\overline{u}\overline{v_{2}}}uv_{2}\otimes 1$ $\displaystyle=$ $\displaystyle v_{1}u\otimes 1+(-1)^{\overline{v_{2}}\overline{u}}u\otimes v_{2}-(-1)^{\overline{v_{2}}\overline{u}}uv_{2}\otimes 1-(-1)^{\overline{v_{2}}\overline{u}}u\otimes v_{2}$ $\displaystyle=$ $\displaystyle(v_{1}\otimes 1+1\otimes v_{2})(u\otimes 1)-(-1)^{\overline{v_{2}}\overline{u}}(u\otimes 1)(v_{2}\otimes 1+1\otimes v_{2})$ which shows the $\mathfrak{g}\times\mathfrak{g}$-equivariance of the map. It is an isomorphism by the PBW theorem. ∎ From now on we work with $\mathcal{U}\mathfrak{g}$ as our space of distributions. Observe that in this case, $\mathfrak{g}^{\prime}=\\{(u,\delta(u)):u\in\mathfrak{g}\\}\cong\mathfrak{g},$ and so we can identify $\mathfrak{g}^{\prime}$ with $\mathfrak{g}$ as a Lie superalgebra. With this setup, the action of $\mathfrak{g}^{\prime}$ on $\operatorname{Dist}(G\times G/G,eG)$ is given by, for $u\in\mathfrak{g}$ and $v\in\mathcal{U}\mathfrak{g}$, $u\cdot v=uv-(-1)^{\overline{u}\overline{v}+\overline{u}}vu.$ This is exactly Gorelik’s twisted adjoint action defined in [Gor00]. There, she proved algebraically that $\mathcal{U}\mathfrak{g}$ is an induced module from $\mathfrak{g}_{\overline{0}}$ under this action. However this follows from our geometric perspective via 5.7. Gorelik defined $\mathcal{A}\subseteq\mathcal{U}\mathfrak{g}$, the anticentre of $\mathcal{U}\mathfrak{g}$, to be the $\mathfrak{g}^{\prime}$-invariant distributions on $G$. However please note remark 7.5 regarding notation. ### 9.1. Structure of $k[G]$ as a $G\times G$-module We assume that $G$ is Cartan-even. The following is proven in [She19]. ###### Proposition 9.2. 1. (1) As a $G\times G$-module, we may write $k[G]=\bigoplus\limits_{\mathcal{B}}M_{\mathcal{B}},$ where each $M_{\mathcal{B}}$ is an indecomposable $G\times G$-module, and the sum runs over all blocks of $\operatorname{Rep}(G)$ up to parity. 2. (2) The socle of $k[G]$ is the submodule generated by all highest weight vectors with respect to an Iwasawa Borel, and is equal to $\bigoplus\limits_{L}L^{*}\boxtimes L$, where $L$ runs over all simple modules of $G$ up to parity. 3. (3) The socle of $M_{\mathcal{B}}$ is $\bigoplus\limits_{L}L^{*}\boxtimes L$, where the sum runs over all simple $G$-modules $L$ in $\mathcal{B}$, up to parity. In particular $M_{\mathcal{B}}$ is simple if and only if its socle is simple. 4. (4) For a simple $G$-module $L$, the submodule $L^{*}\boxtimes L$ is given by the image of the $\operatorname{End}(L)^{*}$ in $k[G]$ under the pullback morphism $k[\operatorname{End}(L)]\to k[G]$ coming from the representation $G\to\operatorname{End}(L)$. ###### Proposition 9.3. Let $L$ be a simple $G$-module. Then $\operatorname{End}(L)$, as a $G\times G$-module, admits a unique, up to scalar, $G^{\prime}$-invariant given by $\delta_{L}$, the parity involution. Consequently, $\text{tr}_{L}\in\operatorname{End}(L)^{*}$ defines a nonzero $G^{\prime}$-invariant function on $G$. ###### Proof. Clearly $\delta_{L}$ is $\mathfrak{g}_{\overline{0}}$-invariant. For $u\in\mathfrak{g}_{1}$, we see that $((u,-u)\delta)(v)=u\delta(v)+\delta(uv)=(-1)^{\overline{v}}uv+(-1)^{\overline{u}+\overline{v}}uv=0$ so $\gamma$ is $\mathfrak{g}^{\prime}$-invariant. Now as a tensor, $\delta_{L}=\sum\limits_{i}(-1)^{\overline{e_{i}}}e_{i}\otimes\varphi_{i}$, where $\varphi_{i}(e_{i})=1$, so after applying the braiding to switch the order of tensors, we obtain the trace in $\operatorname{End}(L)^{*}$. ∎ ###### Corollary 9.4. If $L$ is a simple $G$-module, $L^{*}\boxtimes L\subseteq k[G]$ has a unique $\mathfrak{g}^{\prime}$-invariant, given by $\text{tr}_{L}$. In particular, $\text{tr}_{L}(eG)=\operatorname{dim}L\neq 0$, so that the hypotheses of 8.17 apply. The following is well-known: ###### Lemma 9.5. Write $\mathfrak{a}=\\{(h,-h):h\in\mathfrak{h}\\}$. Then $\mathfrak{a}$ is a Cartan subspace for $(\mathfrak{g}\times\mathfrak{g},\mathfrak{g})$. Now choose a positive system for $\mathfrak{g}$, say $\Delta^{+}$, with Borel $\mathfrak{b}^{+}$ and opposite Borel $\mathfrak{b}^{-}$. Write $\rho$ for the Weyl vector. Then we obtain a positive system for $\mathfrak{g}\times\mathfrak{g}$ with Borel subalgebra $\mathfrak{b}^{-}\times\mathfrak{b}^{+}$. ###### Corollary 9.6. The above choice of Borel subalgebra is an Iwasawa Borel subalgebra. Further, we have that $\mathfrak{n}=\mathfrak{u}^{-}\times\mathfrak{u}^{+}$, where $\mathfrak{u}^{\pm}\subseteq\mathfrak{b}^{\pm}$ are the nilpotent radicals of $\mathfrak{b}^{\pm}$, and the Iwasawa decompositions: $\mathfrak{g}\times\mathfrak{g}=\mathfrak{g}\oplus\mathfrak{a}\oplus\mathfrak{n}=\mathfrak{g}^{\prime}\oplus\mathfrak{a}\oplus\mathfrak{n}$ ###### Lemma 9.7. For the above choice of $\mathfrak{a}$ and positive system, $\Lambda^{+}=\\{(-\lambda,\lambda):\lambda\text{ is a }\mathfrak{b}\text{-dominant weight}\\}.$ Given this lemma, we identify $\Lambda^{+}$ with the set of $\mathfrak{b}$-dominant weights, and write $f_{\lambda}\in k[G]$ for the function of highest weight $(-\lambda,\lambda)$ such that $f_{\lambda}(eG)=1$. ###### Proposition 9.8. We have $\mathcal{U}(\mathfrak{g}\times\mathfrak{g})f_{\lambda}=\mathcal{U}\mathfrak{g}f_{\lambda}=\mathcal{U}\mathfrak{g}^{\prime}f_{\lambda}=L_{\lambda}^{*}\boxtimes L_{\lambda}$. ###### Proof. This follows from the Iwasawa decompositions. ∎ Recall that for this supersymmetric pair, the Harish-Chandra morphism on distributions corresponds to the usual Harish-Chandra morphism $HC:\mathcal{U}\mathfrak{g}\to S(\mathfrak{h})$ on the enveloping algebra. ###### Corollary 9.9. $p_{G\times G/G,B}(\lambda)\neq 0$ if and only if $L_{\lambda}^{*}\boxtimes L_{\lambda}$ has a copy of $I_{G^{\prime}}(k)$ as a $G^{\prime}$-module. ###### Proof. This follows from 9.8, 9.4, and 8.17. ∎ ### 9.2. Projectivity criteria for irreducible modules For this section we work with a fixed Borel subalgebra $\mathfrak{b}$ of $\mathfrak{g}$, which as stated above determines a fixed Iwasawa Borel subalgebra of $\mathfrak{g}\times\mathfrak{g}$. We observe that $\operatorname{id}\times\delta$ defines an automorphism of $G\times G$ which takes $G$ to $G^{\prime}$ and vice-versa. In particular it defines an isomorphism $G\times G/G^{\prime}\to G\times G/G$ which is $\operatorname{id}\times\delta$-equivariant, meaning that the pullback morphism defines a $G\times G$-equivariant isomorphism $k[G]=k[G\times G/G]\to k[G\times G/G^{\prime}]^{\operatorname{id}\times\delta}.$ Here we use the notation $V^{\phi}$ for the $\phi$ twist of a $G$-module $V$, where $\phi$ is an automorphism of $G$. Notice that since $\operatorname{id}\times\delta$ is the identity on $G_{0}\times G_{0}$, we have that $(L^{*}\boxtimes L)^{\operatorname{id}\times\delta}\cong L^{*}\boxtimes L$ for a simple $G$-module $L$, by highest weight theory. ###### Proposition 9.10. Let $(-\lambda,\lambda)\in\Lambda^{+}$. Then $L(\lambda)^{*}\boxtimes L(\lambda)$ contains a copy of $I_{G^{\prime}}(k)$ if and only if $L(\lambda)$ is a projective $G$-module. ###### Proof. For the backward direction, if $L(\lambda)$ is projective then $L(\lambda)^{*}\boxtimes L(\lambda)$ is too, thus its restriction to $G^{\prime}$ must be projective since $G$ is quasireductive. By 9.4, $L(\lambda)^{*}\boxtimes L(\lambda)$ contains a $G^{\prime}$-invariant, and thus it must contain a copy of $I_{G^{\prime}}(k)$. Conversely, suppose that $L(\lambda)^{*}\boxtimes L(\lambda)$ contains a copy of $I_{G^{\prime}}(k)$. Since $L(\lambda)^{*}\boxtimes L(\lambda)$ is irreducible, we may apply 8.18 to obtain that $I_{G\times G}(L(\lambda)^{*}\boxtimes L(\lambda))$ is a $G\times G$-submodule of $k[G\times G/G^{\prime}]$. Using the isomorphism $k[G\times G/G^{\prime}]^{\operatorname{id}\times\delta}\to k[G\times G/G]$, we obtain that $I_{G\times G}(L(\lambda)^{*}\boxtimes L(\lambda))$ is a submodule $k[G]$. Since $I_{G\times G}(L(\lambda)^{*}\boxtimes L(\lambda))$ is projective, it splits off $k[G]$ as a $G\times G$-module. By 9.2, we find that $L(\lambda)$ must be in its own block, and is therefore projective. ∎ From this, we obtain the following remarkable result. We write $p_{G,B}:=p_{G\times G/G,B}=HC(\operatorname{ev}_{eG}v_{\mathfrak{g}^{\prime}}).$ ###### Theorem 9.11. For a $\mathfrak{b}$-dominant weight $\lambda$, the simple $G$-module $L(\lambda)$ is projective if and only if $p_{G,B}(\lambda)\neq 0$. In particular, the dominant weights $\lambda$ for which $L(\lambda)$ is projective is given by the intersection of $\Lambda^{+}$ with a Zariski open subset of $\mathfrak{h}^{*}$ given by the non-vanishing set of the polynomial $p_{G,B}$. In particular if $\operatorname{dim}\mathfrak{g}_{\overline{1}}$ is odd or $\operatorname{Ber}(\mathfrak{g}_{\overline{1}})$ is not the trivial $\mathfrak{g}$-module then no irreducible $G$-modules are projective. ###### Proof. The first statement follows from 9.10 and 9.9. The last statement follows from 8.13. Note that the last statement could be proven more easily using the result in [Ser11] that $I_{G}(L)\cong P_{G}(L\otimes\operatorname{Ber}(\mathfrak{g}_{\overline{1}}))$. ∎ In the appendix, we will prove 9.11 in full generality, i.e. without the assumption the $G$ is Cartan-even. ###### Corollary 9.12. If $HC(\mathcal{Z})=k$ (e.g. if $\mathcal{Z}=k$) and $\mathfrak{g}\neq 0$, then no finite-dimensional irreducible $G$-modules are projective. ###### Proof. By 9.11, if there exists a finite-dimensional irreducible projective $G$-module, then $p_{G,B}\in S(\mathfrak{h})$ is a nonzero polynomial. By 9.11, $\operatorname{Ber}(\mathfrak{g})$ must be trivial and thus $\mathcal{A}_{G\times G/G}=\mathcal{A}$. Thus we know from [Gor00] that $p_{G,B}^{2}\in HC(\mathcal{Z})=k$, and is nonzero, so $p_{G,B}$ must be a constant polynomial. Therefore $\operatorname{Rep}(G)$ is semisimple, and thus $HC(\mathcal{Z})$ must be larger than constants by the classification of algebraic supergroups with semisimple representation theory (see [She20b]). ∎ ###### Corollary 9.13. If there are no finite-dimensional irreducible projective $G$-modules, then $HC(\mathcal{A}_{G/K})=0$. ###### Proof. If for $\lambda\in\Lambda^{+}$ we have $HC(\gamma)(\lambda)\neq 0$, then $L(\lambda)^{*}\boxtimes L(\lambda)$ contains a copy of $I_{G^{\prime}}(k)$, so that $L(\lambda)$ is projective, a contradiction. ∎ We now can prove a nice sufficient criteria for when $G$ admits projective irreducible modules. ###### Theorem 9.14. Suppose that $G$ is a Cartan-even quasireductive supergroup such that the following conditions on its Lie superalgebra, $\mathfrak{g}$, hold: 1. (1) $\operatorname{dim}[\mathfrak{g}_{\alpha},\mathfrak{g}_{-\alpha}]=1$ for all $\alpha\in\Delta$; and 2. (2) for all $\alpha\in\Delta$, the pairing $[-,-]:\mathfrak{g}_{\alpha}\otimes\mathfrak{g}_{-\alpha}\to[\mathfrak{g}_{\alpha},\mathfrak{g}_{-\alpha}]$ is nondegenerate. Make a choice of positive roots and thus a Borel subalgebra $\mathfrak{b}$ containing $\mathfrak{h}$, and let $\Delta_{\overline{1}}^{+}=\\{\alpha_{1},\dots,\alpha_{n}\\}$. Let $h_{\alpha_{i}}\in[\mathfrak{g}_{\alpha_{i}},\mathfrak{g}_{-\alpha_{i}}]$ be a chosen nonzero element, and write $r_{i}=\operatorname{dim}\mathfrak{g}_{\alpha_{i}}$. Then up to a scalar we have: $p_{G,B}=h_{\alpha_{1}}^{r_{1}}\cdots h_{\alpha_{n}}^{r_{n}}+l.o.t.$ In particular $p_{G,B}\neq 0$, so that $G$ admits projective irreducible modules. ###### Proof. We may choose a weight basis of $\mathfrak{g}_{\overline{1}}$ as follows. Let $u_{0},\dots,u_{m}$ be a root vector basis of $\mathfrak{n}_{\overline{1}}$, where $u_{i}$ has weight $\beta_{i}$, and the $\beta_{i}$’s need not be distinct. Then there exists a root vector basis $v_{0},\dots,v_{m}$ of $\mathfrak{n}^{-}_{\overline{1}}$ such that $v_{i}$ has weight $-\beta_{i}$ and $[u_{i},v_{j}]=\delta_{ij}h_{\beta_{i}}$ whenever $\beta_{i}=\beta_{j}$. Write $u_{I}$ and $v_{I}$ for the corresponding ordered products when $I\subseteq\\{0,\dots,m\\}$, and set $u_{\emptyset}=v_{\emptyset}=1$. For a subset $J\subseteq\\{0,\dots,m\\}$ with $J=\\{j_{1}<\dots<j_{l}\\}$ we set $\widetilde{v_{J}}=v_{j_{l}}\cdots v_{j_{1}}.$ Then we may take as a basis for $\mathcal{U}\mathfrak{g}/\mathcal{U}\mathfrak{g}\mathfrak{g}_{\overline{0}}$ the elements $u_{I}\widetilde{v_{J}}$ for $I,J\subseteq\\{0,\dots,m\\}$. Therefore we may write $v_{\mathfrak{g}}=u_{0}\cdots u_{m}v_{m}\cdots v_{0}+v^{\prime},$ where $v^{\prime}\in\mathcal{F}^{\operatorname{dim}\mathfrak{g}_{\overline{1}}-2}$. Thus $\operatorname{ad}^{\prime}(v_{\mathfrak{g}})(1)=\operatorname{ad}^{\prime}(u_{0}\cdots u_{m}v_{m}\cdots v_{0})(1)+\operatorname{ad}^{\prime}(v^{\prime})(1).$ By lemma 8.8 $HC(\operatorname{ad}^{\prime}(v^{\prime})(1))\leq m$, so it suffices to show that (up to a scalar) $\operatorname{ad}^{\prime}(u_{0}\cdots u_{m}v_{m}\cdots v_{0})(1)=h_{\beta_{0}}\cdots h_{\beta_{m}}+l.o.t.$ Observe that $\operatorname{ad}^{\prime}(v_{n}\cdots v_{0})(1)=\sum\limits_{I\subseteq\\{1,\dots,m\\}}(-1)^{i_{1}+\dots+i_{j}}\widetilde{v_{I^{c}}}v_{I},$ (note the $\\{1,\dots,m\\}$ in the sum is not a typo) where $I=\\{i_{1},\dots,i_{j}\\}$ and $I^{c}$ denotes the complement of $I$. Now if we apply $\operatorname{ad}^{\prime}(u_{0}\cdots u_{m})$ to $\operatorname{ad}^{\prime}(v_{m}\cdots v_{0})(1)$, the only term that does not vanish under the Harish-Chandra morphism is $\sum\limits_{I\subseteq\\{1,\dots,m\\}}(-1)^{i_{1}+\dots+i_{j}}u_{0}\cdots u_{m}\widetilde{v_{I^{c}}}v_{I}.$ Now it suffices to show that $HC((-1)^{i_{1}+\dots+i_{j}}u_{0}\cdots u_{m}\widetilde{v_{I^{c}}}v_{I})=h_{\beta_{0}}\cdots h_{\beta_{m}}+l.o.t.$ The proof of this is inductive. If $m\notin I$, then we may write $u_{0}\cdots u_{m}\widetilde{v_{I^{c}}}v_{I}=u_{0}\cdots u_{m}v_{m}\widetilde{v_{I^{c}\setminus\\{m\\}}}v_{I}.$ This is equal to (after removing the term in $v_{m}\mathcal{U}\mathfrak{g}$) $\displaystyle\sum\limits_{j}(-1)^{m-j}u_{0}\cdots[u_{j},v_{m}]\cdots u_{m}\widetilde{v_{I^{c}\setminus\\{n\\}}}v_{I\setminus\\{m\\}}$ $\displaystyle+$ $\displaystyle u_{0}\cdots u_{m-1}\widetilde{v_{I^{c}\setminus\\{m\\}}}v_{I}h_{\beta_{m}}$ $\displaystyle+$ $\displaystyle cu_{0}\cdots u_{m-1}\widetilde{v_{I^{c}\setminus\\{m\\}}}v_{I}$ In the last summand $c$ is the constant given by $(-\beta_{0}-\dots-\beta_{m-1})(h_{\beta_{m}})$, but its precise value does not matter as this term will have Harish-Chandra projection of degree $m$ or less. For the first summand, let $e_{j,m}=[u_{j},v_{m}]\in\mathfrak{g}_{\overline{0}}$, and suppose that $e_{j,m}$ lies in $\mathfrak{n}^{-}$ (note that it cannot lie in $\mathfrak{h}$ by our choice of basis). The argument for when it lies in $\mathfrak{n}^{+}$ is entirely analogous. Then we have (after removing the term in $e_{j,m}\mathcal{U}\mathfrak{g}$.) $u_{0}\cdots e_{j,m}\cdots u_{m}\widetilde{v_{I^{c}\setminus\\{m\\}}}v_{I}=\sum\limits_{l}u_{0}\cdots[u_{l},e_{j,m}]\cdots\hat{u_{j}}\cdots u_{m}\widetilde{v_{I^{c}\setminus\\{m\\}}}v_{I}$ Thus we end up with an element in $\mathcal{F}^{2m}$ so that its Harish- Chandra projection is of degree at most $m$, and so does not contribute to the top degree. Now suppose instead $m\in I$, so that we can write $u_{0}\cdots u_{m}\widetilde{v_{I^{c}}}v_{I}=u_{0}\cdots u_{m}\widetilde{v_{I^{c}}}v_{I\setminus\\{m\\}}v_{m}.$ Write $I=\\{i_{1},\dots,i_{s},m\\}$ and $I^{c}=\\{j_{1},\dots,j_{t}\\}$. Then the above expression is equal to $\displaystyle(-1)^{n}u_{0}\cdots u_{m-1}\widetilde{v_{I^{c}}}v_{I\setminus\\{m\\}}h_{\beta_{m}}$ $\displaystyle+$ $\displaystyle\sum\limits_{p}u_{0}\cdots u_{m-1}v_{j_{t}}\cdots[u_{m},v_{j_{p}}]\cdots v_{j_{t}}v_{I}$ $\displaystyle+$ $\displaystyle\sum\limits_{q}u_{0}\cdots u_{m-1}\widetilde{v_{I^{c}}}v_{i_{1}}\cdots[u_{m},v_{i_{q}}]\cdots v_{n}.$ Now in the last two summands we may apply a similar argument as above to prove that their Harish-Chandra projections are of degree less than or equal to $m$. Therefore, we have shown that the following two elements have the same top degree term: $HC((-1)^{i_{1}+\dots+i_{j}}u_{0}\cdots u_{m}\widetilde{v_{I^{c}}}v_{I}),\ \ \ \ HC((-1)^{j_{1}+\cdot+j_{l}}u_{0}\cdots u_{m-1}\widetilde{v_{J^{c}}}v_{J})h_{\alpha_{m}},$ where $\\{j_{1},\dots,j_{l}\\}=J=I\cap\\{0,\dots,m-1\\}$ and the complement of $J$ is taken in $\\{0,\dots,m-1\\}$. Now we may apply an induction argument to show that $HC((-1)^{i_{1}+\dots+i_{j}}u_{0}\cdots u_{m}\widetilde{v_{I^{c}}}v_{I})=h_{\beta_{0}}\cdots h_{\beta_{m}}+l.o.t.$ which completes the proof. ∎ We say a Lie superalgebra is quadratic if it admits a nondegenerate, invariant, and even supersymmetric form $(-,-)$. The following is straightforward to show from the properties of being quadratic. ###### Corollary 9.15. Let $G$ be a Cartan-even quasireductive supergroup whose Lie superalgebra is quadratic. Then the conditions of 9.14 hold. In particular $G$ admits projective irreducible modules. ### 9.3. $\mathfrak{g}$ basic classical We now additionally assume that $\mathfrak{g}$ is almost simple and basic classical, by which we mean it is either basic simple, $\mathfrak{s}\mathfrak{l}(n|n)$, or $\mathfrak{g}\mathfrak{l}(m|n)$. The following two results could be proven more easily using 9.11 and facts about the representation theory of basic Lie superalgebras, but we present the proofs below as they demonstrate a different approach which generalizes to other supersymmetric pairs. ###### Lemma 9.16. Suppose that $\alpha\in\Delta^{+}$ is a simple isotropic root, and that we have $(\lambda,\alpha)=0$. Then $p_{G,B}(\lambda)=0$. ###### Proof. Let $u_{\alpha}=(e_{\alpha},0)-(0,e_{\alpha})\in\mathfrak{g}^{\prime}$. Then $u_{\alpha}f_{\lambda}=0$ because $\alpha$ is simple isotropic and $(\lambda,\alpha)=0$. Now suppose that $\mathcal{U}\mathfrak{g}^{\prime}f_{\lambda}$ contained a copy of $I_{G^{\prime}}(k)$, so that we may write $\mathcal{U}\mathfrak{g}^{\prime}f_{\lambda}=I(k)\oplus M$, for some complimentary module $M$. Then with respect the this direct sum decomposition we can write $f_{\lambda}=g+m$, where $g\in I_{G^{\prime}}(k)$ generates. But then we must have $u_{\alpha}g=0$, implying by the projectivity of $I_{G^{\prime}}(k)$ that $g=u_{\alpha}h$ for some $h\in I(k)$. However this contradicts $g$ begin non-zero on some $\mathfrak{g}^{\prime}$-coinvariant. This finishes the proof. ∎ ###### Proposition 9.17. If $\alpha$ is a positive isotropic root such that $(\lambda+\rho,\alpha)=0$, then $p_{G,B}(\lambda)=0$. ###### Proof. If $\alpha$ is simple isotropic, we are done by lemma 9.16. Otherwise, there exists a sequence of odd reflections transforming $\alpha$ into a simple root. If $r_{\beta}$ is a such an odd reflection, then we see that $\rho$ becomes $\rho+\beta$ and $\lambda$ changes to $\lambda-\beta$ if $(\lambda,\beta)\neq 0$, and if $(\lambda,\beta)=0$ then by lemma 9.16 we are done anyway, so we may assume this doesn’t happen. Writing $\rho^{\prime}$ for the new Weyl vector and $\lambda^{\prime}$ for the new highest weight, we see that we still have $(\lambda^{\prime}+\rho^{\prime},\alpha)=(\lambda+\rho,\alpha)=0$ The proof follows by induction. ∎ ###### Corollary 9.18. $\prod\limits_{\alpha\in\Delta^{+}_{\overline{1}},(\alpha,\alpha)=0}(h_{\alpha}+(\rho,\alpha))$ divides $p_{G,B}$ and thus divides $HC(\gamma)$ for any $\gamma\in\operatorname{Dist}(G/K,eK)^{\mathfrak{k}^{\prime}}$. ###### Proof. The first statement follows by a density argument, and the second statement follows from 8.17. ∎ The above result recovers part of Gorelik’s result in [Gor00], where she proved that $HC(\mathcal{A})=S(\mathfrak{h})^{W_{.}}\prod\limits_{\alpha\in\Delta^{+}_{\overline{1}}}(h_{\alpha}+(\rho,\alpha)).$ Here $S(\mathfrak{h})^{W_{.}}$ denotes the $\rho$-shifted $W$-invariant polynomials in $\mathfrak{h}$, where $W$ has a slightly different action than usual (see [Gor00] for details). Gorelik proved this result by studying the action of $\mathcal{A}$ on Verma modules of $\mathfrak{g}$, and using known results on highest weight vectors in Verma modules. Thus far the author does not know how to reprove her results from a purely geometric standpoint. In particular the invariance under an even Weyl group remains elusive. ## 10\. Full ghost centre As explained previously, in [Gor00] an algebra called the ghost centre $\tilde{\mathcal{Z}}\subseteq\mathcal{U}\mathfrak{g}$ was defined as $\tilde{\mathcal{Z}}:=\mathcal{Z}+\mathcal{A},$ where $\mathcal{A}=(\mathcal{U}\mathfrak{g})^{\mathfrak{g}^{\prime}}$. One can check that for $a,b\in\mathcal{A}$ we have $ab\in\mathcal{Z}$, so this indeed forms an algebra. We seek to expand this algebra to what we will call the full ghost centre, $\mathcal{Z}_{full}$, of $\mathfrak{g}$. It is a subalgebra of $\mathcal{U}\mathfrak{g}$ which we now describe. ### 10.1. Twisted adjoint actions Let $\operatorname{Aut}(\mathfrak{g},\mathfrak{g}_{\overline{0}})$ denote the set of automorphisms of $\mathfrak{g}$ which fix $\mathfrak{g}_{\overline{0}}$ pointwise. For $\phi\in\operatorname{Aut}(\mathfrak{g},\mathfrak{g}_{\overline{0}})$, define the $\phi$-twisted adjoint action on $\mathcal{U}\mathfrak{g}$ by $\operatorname{ad}_{\phi}(u)(v)=uv-(-1)^{\overline{u}\overline{v}}v\phi(u)$ where $u\in\mathfrak{g}$, $v\in\mathcal{U}\mathfrak{g}$. ###### Definition 10.1. Define $\mathcal{A}_{\phi}$ to be the $\operatorname{ad}_{\phi}$-invariant elements of $\mathcal{U}\mathfrak{g}$. Observe that $\mathcal{A}_{\operatorname{id}}=\mathcal{Z}$ and $\mathcal{A}_{\delta}=\mathcal{A}$, where $\delta(u)=(-1)^{\overline{u}}u$. Another way to understand this action is as follows: consider the subalgebra $\mathfrak{g}_{\phi}$ of $\mathfrak{g}\times\mathfrak{g}$ given by $\\{(u,\phi(u)):u\in\mathfrak{g}\\}$. Then $\mathfrak{g}_{\phi}\cong\mathfrak{g}$, and its even part is $(\mathfrak{g}_{\phi})_{\overline{0}}=\\{(u,u):u\in\mathfrak{g}_{\overline{0}}\\}$. Then the action of $\mathfrak{g}_{\phi}$ on $\operatorname{Dist}(G\times G/G,eG)\cong\mathcal{U}\mathfrak{g}$ exactly induces the $\phi$-twisted adjoint action. Let $G_{\phi}\subseteq G\times G$ be the connected subgroup which integrates $\mathfrak{g}_{\phi}$. ###### Lemma 10.2. Suppose that $\phi(u)=u$ implies $u\in\mathfrak{g}_{\overline{0}}$. Then $\mathcal{U}\mathfrak{g}\cong\operatorname{Ind}_{\mathfrak{g}_{\overline{0}}}^{\mathfrak{g}}\mathcal{U}\mathfrak{g}_{\overline{0}}$ with respect to the $\phi$-twisted action. In particular if $\operatorname{Ber}(\mathfrak{g}_{\overline{1}})$ is trivial, then there is a natural isomorphism of vector spaces $\mathcal{Z}(\mathcal{U}\mathfrak{g}_{\overline{0}})\to\mathcal{A}_{\phi}$ given by $z\mapsto\operatorname{ad}_{\phi}(v_{\mathfrak{g}})(z).$ ###### Proof. If $\phi(u)=u$ implies $u\in\mathfrak{g}_{\overline{0}}$, then $\mathfrak{g}_{\phi}\to(T_{eG}(G\times G/G))_{\overline{1}}$ is an isomorphism. Therefore, 5.3 applies. ∎ Observe that $\operatorname{id}\times\phi$ defines an isomorphism of $G\times G$ taking $G$ to $G_{\phi}$, and thus an isomorphism $G\times G/G\to G\times G/G_{\phi}$ which is $\operatorname{id}\times\phi$-equivariant. Since $\operatorname{id}\times\phi$ is the identity on $\mathfrak{g}_{\overline{0}}\times\mathfrak{g}_{\overline{0}}$, we obtain the following generalization of 9.11. ###### Proposition 10.3. Suppose that $\operatorname{Ber}(\mathfrak{g})$ is trivial and that $\phi\in\operatorname{Aut}(\mathfrak{g},\mathfrak{g}_{\overline{0}})$ satisfies the conditions of lemma 10.2. Then for $\gamma\in\mathcal{A}_{\phi}$ and $\lambda\in\Lambda^{+}$, $HC(\gamma)(\lambda)=0$ if $L(\lambda)$ is not projective. Further if the unique $\operatorname{ad}_{\phi}$-invariant element of $\operatorname{End}(L)^{*}$ always is nonzero at $eG$ for a simple $G$-module $L$, then $HC(\operatorname{ad}_{\phi}(v_{\mathfrak{g}})1)(\lambda)\neq 0$ if and only if $L(\lambda)$ is projective. The unique $\operatorname{ad}_{\phi}$-invariant element of $\operatorname{End}(L)^{*}$ can be thought of as a $\phi$-twisted trace function. We describe these elements for basic type I algebras in section 10.6. ### 10.2. The full ghost centre For $\phi,\psi\in\operatorname{Aut}(\mathfrak{g},\mathfrak{g}_{\overline{0}})$, it is easy to check that multiplication in $\mathcal{U}\mathfrak{g}$ induces a morphism $\mathcal{A}_{\phi}\otimes\mathcal{A}_{\psi}\to\mathcal{A}_{\psi\phi}.$ ###### Definition 10.4. Define the full ghost centre of $\mathcal{U}\mathfrak{g}$ to be the algebra given by $\mathcal{Z}_{full}:=\sum\limits_{\phi\in\operatorname{Aut}(\mathfrak{g},\mathfrak{g}_{\overline{0}})}\mathcal{A}_{\phi}.$ Notice that $\tilde{\mathcal{Z}}\subseteq\mathcal{Z}_{full}\subseteq(\mathcal{U}\mathfrak{g})^{\mathfrak{g}_{\overline{0}}}$. ### 10.3. The case of an abelian Lie superalgebra As a toy case, we consider the above constructions for the abelian superalgebra $\mathfrak{g}=k^{m|n}$. Then $\operatorname{Aut}(\mathfrak{g},{\mathfrak{g}_{\overline{0}}})\cong GL(n)$. Let $A\in GL(n)$, and let $\xi_{A}$ be a nonzero element of $\Lambda^{top}\operatorname{im}(A-\operatorname{id})$, where $\xi_{\operatorname{id}}=1$. ###### Lemma 10.5. $\mathcal{A}_{A}=\mathcal{U}\mathfrak{g}\xi_{A}$. ###### Proof. For $u\in\mathcal{U}\mathfrak{g}$ and $\eta\in\mathfrak{g}$, we see that $\displaystyle\operatorname{ad}_{A}(\eta)(u)$ $\displaystyle=$ $\displaystyle\eta u-(-1)^{\overline{u}}uA\eta$ $\displaystyle=$ $\displaystyle(\eta-A\eta)u.$ From here the result is straightforward. ∎ In this case we see that the ghost centres for different automorphisms in $\operatorname{Aut}(\mathfrak{g},\mathfrak{g}_{\overline{0}})$ may overlap in myriad ways. Now suppose $A\in\operatorname{Aut}(\mathfrak{g},\mathfrak{g}_{\overline{0}})$ doesn’t fix any non-zero odd vectors. Then under the twisted adjoint action determined by $A$, $\mathcal{U}\mathfrak{g}$ is isomorphic to a sum of the left (or right) regular representation of $\mathcal{U}\mathfrak{g}_{\overline{1}}$, and we have $\mathcal{A}_{A}=S(\mathfrak{g}_{\overline{0}})\Lambda^{n}\mathfrak{g}_{\overline{1}}$. By lemma 6.10, we have $v_{\mathfrak{g}}=\xi_{1}\dots\xi_{n}$, where $\xi_{1},\dots,\xi_{n}$ is any basis of $\mathfrak{g}_{\overline{1}}$. Thus if $A$ is fixed point free, the following element must be non-zero: $\operatorname{ad}_{A}(\xi_{1}\cdots\xi_{n})(1)=(1-A)\xi_{1}\cdots(1-A)\xi_{n}=\det(1-A)\xi_{1}\cdots\xi_{n}.$ So this is nonzero if and only if $A-1$ is invertible, i.e. $A$ is not an eigenvalue of $1$. ###### Remark 10.6. Using the above computation, it is possible to give a purely algebraic proof of lemma 10.2 with the same proof as given in [Gor00]. Namely, given $u_{1},\dots,u_{n}\in\mathfrak{g}_{\overline{1}}$, $v\in\mathcal{U}\mathfrak{g}$, and $\phi\in\operatorname{Aut}(\mathfrak{g},\mathfrak{g}_{\overline{0}})$, we have $\operatorname{gr}(\operatorname{ad}_{\phi}(u_{1}\cdots u_{n})(v))=\operatorname{ad}_{\phi}(\operatorname{gr}(u_{1}\cdots u_{n}))(\operatorname{gr}v),$ where $\operatorname{gr}$ is the associated graded morphism $\mathcal{U}\mathfrak{g}\to S(\mathfrak{g})$, and we view $S(\mathfrak{g})$ as the enveloping algebra of $\mathfrak{g}$ with trivial bracket. Thus if $L_{0}\subseteq\mathcal{U}\mathfrak{g}_{\overline{0}}$ is a $\mathfrak{g}_{\overline{0}}$-submodule and $\\{v_{j}\\}\subseteq L_{0}$ is a basis such that $\\{\operatorname{gr}v_{j}\\}$ is linearly independent in $\mathcal{U}\mathfrak{g}_{\overline{0}}$, then by passing to the associated graded we find that $\mathcal{U}\mathfrak{g}L_{0}\cong\operatorname{Ind}_{\mathfrak{g}_{0}}^{\mathfrak{g}}L_{0}$. ### 10.4. $\mathcal{Z}_{full}$ for basic classical Lie superalgebras Let $\mathfrak{g}$ be either $\mathfrak{g}\mathfrak{l}(m|n)$, $\mathfrak{s}\mathfrak{l}(m|n)$, $\mathfrak{p}\mathfrak{s}\mathfrak{l}(n|n)$ for $n>2$, $\mathfrak{o}\mathfrak{s}\mathfrak{p}(m|2n)$, or a basic exceptional simple Lie superalgebra. Recall that such a superalgebra is called type I if it admits a $\mathbb{Z}$-grading $\mathfrak{g}=\mathfrak{g}_{-1}\oplus\mathfrak{g}_{0}\oplus\mathfrak{g}_{1}$ compatible with the $\mathbb{Z}_{2}$-grading. Of these, $\mathfrak{g}\mathfrak{l}(m|n)$, $\mathfrak{s}\mathfrak{l}(m|n)$, $\mathfrak{p}\mathfrak{s}\mathfrak{l}(n|n)$ and $\mathfrak{o}\mathfrak{s}\mathfrak{p}(2|2n)$ are the type I algebras. ###### Remark 10.7 (Caution). We do not consider $\mathfrak{p}\mathfrak{s}\mathfrak{l}(2|2)$ here because lemma 10.8 is false (see section 5.5 of [Mus12]). Further, in section 10.5, we will not consider $\mathfrak{s}\mathfrak{l}(n|n)$ or $\mathfrak{p}\mathfrak{s}\mathfrak{l}(n|n)$ due to the lack of an internal grading operator. ###### Lemma 10.8. $\operatorname{Aut}(\mathfrak{g},\mathfrak{g}_{\overline{0}})\cong k^{\times}$ if $\mathfrak{g}$ is of type I, and otherwise $\operatorname{Aut}(\mathfrak{g},\mathfrak{g}_{\overline{0}})=\langle\delta\rangle\cong\mathbb{Z}/2$. In particular $\tilde{\mathcal{Z}}=\mathcal{Z}_{full}$ for $\mathfrak{g}$ not of type I. ###### Proof. We refer to [Mus12]. In the type I case, we identify $\operatorname{Aut}(\mathfrak{g},\mathfrak{g}_{\overline{0}})\cong k^{\times}$ via defining for $c\in k^{\times}$ the automorphism $c\mathbf{1}_{\mathfrak{g}_{-1}}\oplus\mathbf{1}_{\mathfrak{g}_{0}}\oplus c^{-1}\mathbf{1}_{\mathfrak{g}_{1}}$. ∎ So for these algebras we only get something new in the type I case. We study this case now. Since $\operatorname{Aut}(\mathfrak{g},\mathfrak{g}_{\overline{0}})\cong k^{\times}$, we label the automorphisms by complex numbers $c\in k^{\times}$ according to the identification given in the proof of lemma 10.8. Write $\mathfrak{g}=\mathfrak{g}_{-1}\oplus\mathfrak{g}_{0}\oplus\mathfrak{g}_{1}$ for the $\mathbb{Z}$-grading on $\mathfrak{g}$. Choose a Cartan subalgebra $\mathfrak{h}\subseteq\mathfrak{g}_{\overline{0}}$ and consider a Borel subalgebra $\mathfrak{b}=\mathfrak{b}_{0}\oplus\mathfrak{g}_{1}$ where $\mathfrak{b}_{0}$ is a Borel of $\mathfrak{g}_{0}$ containing $\mathfrak{h}_{\overline{0}}$. We call this a standard Borel subalgebra of $\mathfrak{g}$. Then we obtain a Harish-Chandra morphism with respect to this Borel. Let $W$ denote the Weyl group of $\mathfrak{g}_{\overline{0}}$ and consider its action $\rho$-shifted action on $\mathfrak{h}$, where $\rho$ is the Weyl vector. ###### Lemma 10.9. Every $\mathfrak{b}$-highest weight module $M$ admits a $\mathbb{Z}$-grading that is compatible with the $\mathbb{Z}$-grading on $\mathfrak{g}$. ###### Proof. We may set $M_{0}=M^{\mathfrak{g}_{1}}$, and then define $M_{-i}=\Lambda^{i}\mathfrak{g}_{-1}M_{0}$. ∎ We will use the $\mathbb{Z}$-grading defined in the proof of lemma 10.9 many times in what follows. ###### Theorem 10.10. 1. (1) $\mathcal{Z}_{full}$ is a commutative algebra. 2. (2) For $c\neq 1$, $HC$ is injective on $\mathcal{A}_{c}$, and we have $HC(\mathcal{A}_{c})=HC(\mathcal{A}_{-1})=S(\mathfrak{h})^{W}\prod\limits_{\alpha\in\Delta^{+}_{\overline{1}}}(h_{\alpha}+(\rho,\alpha)).$ ###### Remark 10.11. In this case we see that $HC(\mathcal{A}_{c})\subseteq HC(\mathcal{Z})$. See for instance [CW12] for a description of the Harish-Chandra image of the centre for $\mathfrak{g}\mathfrak{l}(m|n)$ and $\mathfrak{o}\mathfrak{s}\mathfrak{p}(m|2n)$. ###### Proof. Let $t_{\mathfrak{g}}:=\prod\limits_{\alpha\in\Delta_{1}^{+}}(h_{\alpha}+(\alpha,\rho)).$ By 10.3, if $u\in\mathcal{A}_{c}$ then we must have $HC(u)(\lambda)=0$ for $\lambda$ an atypical dominant integral weight, i.e. $\lambda$ such that $(\lambda+\rho,\alpha)=0$ for some odd root $\alpha$, or equivalently $t_{\mathfrak{g}}(\lambda)=0$. The set of such $\lambda$ are dense in the zero set of $t_{\mathfrak{g}}$, and thus we have that $t_{\mathfrak{g}}$ divides $HC(u)$. To obtain Weyl group invariance, let $\lambda$ be a dominant integral weight, and consider the Verma module $M_{\mathfrak{b}}(\lambda)$ of highest weight $\lambda$ and highest weight vector $v$. Choose a simple even root $\alpha$, and write $f_{\alpha}$ for a root vector of weight $-\alpha$. Then $f_{\alpha}^{(\lambda,\alpha)+1}v$ will be a highest weight vector of weight $\lambda-((\lambda,\alpha)+1)\alpha=s_{\alpha}(\lambda)$. Since $f_{\alpha}\in\mathfrak{g}_{0}$, $uf_{\alpha}^{(\lambda,\alpha)+1}v=f_{\alpha}^{(\lambda,\alpha)+1}uv=HC(u)(\lambda)f_{\alpha}^{(\lambda,\alpha)+1}v.$ Thus $HC(u)(\lambda)=HC(u)(s_{\alpha}(\lambda))$. Since such $\lambda$ are dense, and the reflections in the even simple roots of this Borel subalgebra generate the Weyl group, it follows that $HC(u)\in S(\mathfrak{h})^{W}$. Since $t_{\mathfrak{g}}$ is itself $W$-invariant, it follows that we’ve shown $HC(\mathcal{A}_{c})\subseteq S(\mathfrak{h})^{W}t_{\mathfrak{g}}.$ To finish the proof, first observe that $HC$ is injective on $\mathcal{A}_{c}$, as if $HC(u)=0$ for $u\in\mathcal{A}_{c}$ then it must act by zero on every Verma module, implying it is zero by [LM94]. Now if we apply lemma 8.8, we can make the same degree arguments as in [Gor00] to obtain the equality $HC(\mathcal{A}_{c})=S(\mathfrak{h})^{W}t$. ∎ With 10.10 we can now completely describe the structure of $\mathcal{Z}_{full}$ for type I algebras. Let $N=\operatorname{dim}\mathfrak{g}_{1}=\operatorname{dim}\mathfrak{g}_{\overline{1}}/2$, and let $\zeta_{N}\in k$ be a primitive $N$th root of unity. ###### Theorem 10.12. $\mathcal{Z}_{full}=\bigoplus\limits_{i=0}^{N-1}\mathcal{A}_{\zeta_{N}^{i}}.$ ###### Proof. Let $u\in\mathcal{A}_{c}$ for $c\neq 1$, and write $p=HC(u)$. Define $u_{i}\in\mathcal{A}_{\zeta_{N}^{i}}$ be the unique element such that $HC(u_{i})=p$. We want to solve the equation $u=a_{0}u_{0}+\cdots+a_{N-1}u_{N-1}=\sum\limits_{i=0}^{N-1}a_{i}u_{i}.$ Let $M:=M_{\mathfrak{b}}(\lambda)$ be the Verma module of highest weight $\lambda$ for a standard Borel $\mathfrak{b}$, and write $M_{j}$ for the $j$th graded part according to the $\mathbb{Z}$-grading defined in the proof of lemma 10.9. Then the scalar action of each side of the above equation applied to $M_{j}$ gives the equation $c^{j}p(\lambda)=\sum\limits_{i=0}^{N-1}a_{i}\zeta_{N}^{ji}p(\lambda).$ If $p(\lambda)=0$ then this equation always holds. If $p(\lambda)\neq 0$, we divide by it to get the system of equations $c^{j}=\sum\limits_{i=0}^{N-1}a_{i}\zeta_{N}^{ji}$ This has a unique solution for $a_{0},\dots,a_{N-1}$ since it is a linear system for the Vandermonde matrix determined by $1,\zeta_{N},\dots,\zeta_{N}^{N-1}$. Since $u-a_{0}u_{0}-\dots-a_{0}u_{0}$ then acts trivially on every Verma module, it is zero in $\mathcal{U}\mathfrak{g}$. It follows that $\mathcal{Z}_{full}\subseteq\bigoplus\limits_{i=0}^{N-1}\mathcal{A}_{\zeta_{N}^{i}}.$ The sum is direct by the nonsingularity of the Vandermonde matrix, so we are done. ∎ ###### Remark 10.13. If $c\in k^{\times}$, we have an inclusion $\mathcal{A}_{c}\to\mathcal{Z}_{full}=\bigoplus\limits_{i=0}^{N-1}\mathcal{A}_{\zeta_{N}^{i}}$. Assume that $c\neq 1$, and for $u\in\mathcal{A}_{c}$ write $p=HC(u)$. Let $u_{i}\in\mathcal{A}_{\zeta_{N}^{i}}$ such that $HC(u_{i})=p$. Then by inverting the Vandermonde matrix, we obtain the decomposition $u=\frac{1}{N}\sum\limits_{i=0}^{N-1}\left(\sum\limits_{j=0}^{N-1}c^{j}\zeta_{N}^{-ij}\right)u_{i}.$ ### 10.5. Characterization of $Z_{full}$ We now give an intrinsic description of $\mathcal{Z}_{full}$ for the type I algebras $\mathfrak{g}\mathfrak{l}(m|n)$, $\mathfrak{s}\mathfrak{l}(m|n)$ for $m\neq n$, and $\mathfrak{o}\mathfrak{s}\mathfrak{p}(2|2n)$. These are distinguished by their having an internal grading operator, that is an element $h\in\mathfrak{g}_{0}$ such that $[h,u]=\deg(u)u$ for $u\in\mathfrak{g}$. Thus we assume for this subsection that $\mathfrak{g}$ is one of these superalgebras. Given a $\mathfrak{g}$-module $V$, recall that $V$ is $\mathbb{Z}$-graded if there exists a $\mathbb{Z}$-grading of $V$ which is compatible with the $\mathbb{Z}$-grading on $\mathfrak{g}$. In lemma 10.9 it was shown that a highest weight module for the Borel $\mathfrak{b}_{0}\oplus\mathfrak{g}_{1}$ is $\mathbb{Z}$-graded. In particular all finite-dimensional irreducible representations of $\mathfrak{g}$ are $\mathbb{Z}$-graded. If $V$ is $\mathbb{Z}$-graded, we say an element $u\in\mathcal{U}\mathfrak{g}$ acts on it by a $\mathbb{Z}$-graded constant if $u$ acts by a scalar on each component of the $\mathbb{Z}$-grading. We now seek to prove the following analogue of corollary 4.4.4 of [Gor00]. ###### Theorem 10.14. $\mathcal{Z}_{full}$ consists exactly of all elements of $\mathcal{U}\mathfrak{g}$ which act by $\mathbb{Z}$-graded constants on all finite-dimensional irreducible representations of $\mathfrak{g}$. Our proof follows the same strategy as taken in [Gor00]. Observe that the $\mathbb{Z}$-grading on $\mathfrak{g}$ induces a $\mathbb{Z}$-grading on $\mathcal{U}\mathfrak{g}$: $\mathcal{U}\mathfrak{g}=\bigoplus\limits_{n\in\mathbb{Z}}(\mathcal{U}\mathfrak{g})_{n}.$ Then $(\mathcal{U}\mathfrak{g})_{0}$ is a subalgebra of $\mathcal{U}\mathfrak{g}$. Let $C$ denote its centralizer, and let $\mathfrak{b}=\mathfrak{b}_{0}\oplus\mathfrak{g}_{1}$ denote a standard Borel subalgebra. Clearly $\mathcal{Z}_{full}\subseteq C$. We would like the show that $\mathcal{Z}_{full}=C$. ###### Lemma 10.15. $C\subseteq(\mathcal{U}\mathfrak{g})_{0}$. ###### Proof. Recall that $\mathfrak{g}$ has an internal grading operator, $h$, in $\mathfrak{g}_{\overline{0}}$. Thus $[h,C]=0$, implying that $C\subseteq(\mathcal{U}\mathfrak{g})_{0}$. ∎ Now let $c\in C$. ###### Lemma 10.16. If $L(\lambda)$ is an irreducible $\mathfrak{b}$-highest weight module, then $c$ acts on $L(\lambda)$ by a $\mathbb{Z}$-graded constant. ###### Proof. Write $L(\lambda)=\bigoplus\limits_{n}L(\lambda)_{n}$ for a $\mathbb{Z}$-grading on $L(\lambda)$. Then $L(\lambda)_{n}$ is a $(\mathcal{U}\mathfrak{g})_{0}$-module, and it suffices to prove it is irreducible. If $W\subseteq L(\lambda)_{n}$ is a $(\mathcal{U}\mathfrak{g})_{0}$-submodule, then $\mathcal{U}\mathfrak{g}W\cap L(\lambda)_{n}=W$. But since $L(\lambda)$ is irreducible this forces either $W=L(\lambda)_{n}$ or $W=0$, so we are done. ∎ Let $S\subseteq\mathfrak{h}^{*}$ denote the collection of weights $\lambda$ such that $M_{\mathfrak{b}}(\lambda)=L_{\mathfrak{b}}(\lambda)$. Note that $S$ is a Zariski dense subset of $\mathfrak{h}^{*}$. We give all Verma modules for $\mathfrak{g}$ with respect to the standard Borel the $\mathbb{Z}$-grading defined in the proof of lemma 10.9. Then for our fixed $c\in C$, we write $f_{i}(\lambda)$ for the constant by which $c$ acts on $M(\lambda)_{-i}$, where $f_{i}:S\to k$ is some function determined by $c$. ###### Proposition 10.17. The functions $f_{i}$ are polynomials on $\mathfrak{h}^{*}$, i.e. $f_{i}\in S(\mathfrak{h})$. For this, we need a lemma: ###### Lemma 10.18. Let $1\leq n\leq\operatorname{dim}\mathfrak{g}_{1}$, and let $\alpha_{1},\dots,\alpha_{n}$ be the $n$ smallest distinct positive odd roots of $\mathfrak{g}$ with respect to our choice of Borel. Then $\operatorname{dim}(\mathcal{U}\mathfrak{n}^{-})_{-\alpha_{1}-\dots-\alpha_{n}}$ is one-dimensional. ###### Proof. The dimension of $(\mathcal{U}\mathfrak{n}^{-})_{-\alpha_{1}-\dots-\alpha_{n}}$ is given by the number of ways to write $-\alpha_{1}-\dots-\alpha_{n}$ as a sum of negative roots of $\mathfrak{g}$, where each odd root can show up at most once. Suppose that we have $-\alpha_{1}-\dots-\alpha_{n}=-\beta_{1}-\dots-\beta_{m}-r_{1}\gamma_{1}-\dots- r_{m^{\prime}}\gamma_{m^{\prime}},$ where $\beta_{1},\dots,\beta_{m}$ are positive odd roots are $\gamma_{1},\dots,\gamma_{m^{\prime}}$ are positive even roots. Writing again $h\in\mathfrak{g}_{0}$ for the internal grading operator, on $\mathfrak{g}$, if we apply $h$ to the above equality of roots we learn that $n=m$. However by our choice of $\alpha_{1},\dots,\alpha_{n}$, this clearly forces $r_{i}=0$ for all $i$ and $\\{\beta_{1},\dots,\beta_{m}\\}=\\{\alpha_{1},\dots,\alpha_{n}\\}$, so we are done. ∎ ###### Proof of 10.17. Observe that $f_{0}=HC(c)$, so this is a polynomial. For $1\leq n\leq\operatorname{dim}\mathfrak{g}_{1}$, let $\alpha_{1},\dots,\alpha_{n}$ be the $n$ smallest distinct positive odd roots with root vectors $u_{1},\dots,u_{n}$, and write $v_{1},\dots,v_{n}$ for the root vectors of weight $-\alpha_{1},\dots,-\alpha_{i}$, where we assume $[u_{i},v_{i}]=h_{\alpha_{i}}$, and $h_{\alpha_{i}}$ is the coroot of $\alpha_{i}$. Let $\lambda\in S$, and write $v_{\lambda}$ for the highest weight vector of $L(\lambda)$. Then observe that $u_{1}\cdots u_{n}cv_{n}\cdots v_{1}v_{\lambda}=f_{n}(a)HC(u_{1}\cdots u_{n}v_{n}\cdots v_{1})v_{\lambda}.$ On the other hand, $u_{1}\cdots u_{n}cv_{n}\cdots v_{1}v_{\lambda}=HC(u_{1}\cdots u_{n}cv_{n}\cdots v_{1})v_{\lambda}.$ Thus on $S$ we have an equality of functions: $HC(u_{1}\cdots u_{n}v_{n}\cdots v_{1})f_{n}(c)=HC(u_{1}\cdots u_{n}cv_{n}\cdots v_{1}).$ Now let $\lambda\in\mathfrak{h}^{*}$ be arbitrary, and write $v_{\lambda}$ for the highest weight vector of $M_{\mathfrak{b}}(\lambda)$. If $HC(u_{1}\cdots u_{n}v_{n}\cdots v_{1})(\lambda)=0$, then $u_{1}\cdots u_{n}v_{n}\cdots v_{1}v_{\lambda}=0$. However by lemma 10.18, $cv_{n}\cdots v_{1}v_{\lambda}$ is again a multiple of $v_{n}\cdots v_{1}v_{\lambda}$, so in this case we also have $u_{1}\cdots u_{n}cv_{n}\cdots v_{1}v_{\lambda}=0$, and thus $HC(u_{1}\cdots u_{n}cv_{n}\cdots v_{1})=0$. Therefore, wherever $HC(u_{1}\cdots u_{n}v_{n}\cdots v_{1})$ vanishes, $HC(u_{1}\cdots u_{n}cv_{n}\cdots v_{1})$ also vanishes. Further, $HC(u_{1}\cdots u_{n}v_{n}\cdots v_{1})$ will have top degree term given by $h_{\alpha_{1}}\cdots h_{\alpha_{n}}$, and thus this polynomial is a product of distinct irreducible polynomials. These facts together imply it divides $HC(u_{1}\cdots u_{n}cv_{n}\cdots v_{1})$ so that $f_{n}(c)=HC(u_{1}\cdots u_{n}cv_{n}\cdots v_{1})/HC(u_{1}\cdots u_{n}v_{n}\cdots v_{1})\in S(\mathfrak{h}).$ ∎ Thank you to Maria Gorelik for helping me work through the following argument. ###### Proposition 10.19. $C$ acts by $\mathbb{Z}$-graded constants on every Verma module $M(\lambda)$. ###### Proof. Let $M=\mathcal{U}\mathfrak{g}\otimes_{\mathcal{U}\mathfrak{b}}S(\mathfrak{h})$ denote the universal Verma module, where $\mathfrak{n}^{+}$ acts trivially on $S(\mathfrak{h})$ and $\mathfrak{h}$ acts by multiplication. This module admits a $\mathbb{Z}$-grading given $M_{-i}=\Lambda^{i}\mathfrak{g}_{-1}\mathcal{U}\mathfrak{g}_{0}S(\mathfrak{h}).$ For each $\lambda\in\mathfrak{h}^{*}$, we have a surjective $\mathfrak{g}$-equivariant morphism $M\to M(\lambda)$ given by evaluation at $\lambda$ on $S(\mathfrak{h})$. For $u\in(\mathcal{U}\mathfrak{n}^{-})_{i}$, consider the element $[c,u]+(HC(c)(\lambda)-f_{i}(\lambda))u\in\mathcal{U}\mathfrak{g}$, and then consider $\left([c,u]+(HC(c)(\lambda)-f_{i}(\lambda))u\right)\otimes 1\in M.$ Then we have shown that this element evaluates to $0$ on $S$, a Zariski dense subset of $\mathfrak{h}^{*}$, and therefore it must vanish under every evaluation. Since $c\cdot uv_{\lambda}=[c,u]v_{\lambda}+HC(c)(\lambda)uv_{\lambda},$ it follows that for every $\lambda$, $c$ acts by $f_{i}(\lambda)$ on $M(\lambda)_{-i}$, so that it acts by $\mathbb{Z}$-graded constants. ∎ ###### Proposition 10.20. $Z_{full}=C$. ###### Proof. Define polynomials $p_{0},\dots,p_{N-1}\in S(\mathfrak{h})$ by: $\begin{bmatrix}f_{0}\\\ f_{1}\\\ \vdots\\\ f_{N-1}\end{bmatrix}=p_{0}\begin{bmatrix}1\\\ 1\\\ \vdots\\\ 1\end{bmatrix}+p_{1}\begin{bmatrix}1\\\ \zeta_{N}\\\ \vdots\\\ \zeta_{N}^{N-1}\end{bmatrix}+\dots+p_{N-1}\begin{bmatrix}1\\\ \zeta_{N}^{N-1}\\\ \vdots\\\ \zeta_{N}^{(N-1)(N-1)}\end{bmatrix}.$ Using the same argument as in 10.10, we must have that $p_{i}\in S(\mathfrak{h})^{W}$ for all $i$. Let $\alpha$ be the unique simple odd isotropic root of this Borel subalgebra, and write $f_{\alpha}$ for a root vector of weight $-\alpha$. Then when $(\lambda,\alpha)=0$, $f_{\alpha}v$ will be a highest weight vector in $M(\lambda)$. Thus $p_{i}(\lambda-\alpha)=\zeta_{N}^{i}p_{i}(\lambda).$ For $i=0$ this implies that $p_{0}(\lambda+r\alpha)=p_{0}(\lambda)$ for all $r\in k$, and for $i>0$ this forces $p_{i}(\lambda-n\alpha)=\zeta_{N}^{ni}p_{i}(\lambda).$ Since $p_{i}$ is a polynomial, this forces $p_{i}(\lambda)=0$, so that $p_{i}$ vanishes on the hyperplane $(\lambda,\alpha)=0$. Using the $\rho$-shifted $W$-invariance of these polynomials, these conditions imply that $p_{0}$ is constant along all hyperplanes of the form $(\lambda+\rho,\alpha)=0$, so that $p_{0}\in HC(\mathcal{Z})$ by Sergeev’s description of $HC(\mathcal{Z})$. On the other hand we find that for $i>0$, $t_{\mathfrak{g}}$ divides $p_{i}$ (see the proof of 10.10), so that $p_{i}\in HC(\mathcal{A}_{-1})$. Now we may take, for each $i$, $u_{i}\in\mathcal{A}_{\zeta_{N}^{i}}$ such that $HC(u_{i})=p_{i}$ for all $i$, and consider the element: $c-u_{0}-\cdots-u_{N-1}.$ By construction this acts by 0 on all Verma modules, and thus is zero so that $c\in\mathcal{Z}_{full}$, finishing the proof. ∎ ###### Theorem 10.21. The following subalgebras of $\mathcal{U}\mathfrak{g}$ agree with each other: 1. (1) $\mathcal{Z}_{full}$; 2. (2) the centralizer of $(\mathcal{U}\mathfrak{g})_{0}$; 3. (3) the center of $(\mathcal{U}\mathfrak{g})_{0}$; and 4. (4) the collection of elements in $\mathcal{U}\mathfrak{g}$ which act by $\mathbb{Z}$-graded constants on every irreducible finite-dimensional representation of $\mathfrak{g}$. ###### Proof. The only nontrivial equality is (2)$\iff$(4). However by [LM94], the set of finite-dimensional irreducible representations of $\mathfrak{g}$ form a complete set, meaning their collective annihilator is trivial. Therefore if an element of $\mathcal{U}\mathfrak{g}$ acts by $\mathbb{Z}$-graded constants on every irreducible finite-dimensional representation of $\mathfrak{g}$, it commutes with $(\mathcal{U}\mathfrak{g})_{0}$ under every such representation, and thus commutes with $(\mathcal{U}\mathfrak{g})_{0}$. ∎ ### 10.6. Twisted trace functions We continue to let $\mathfrak{g}$ be a basic algebra of type I as in the previous section, and let $c\in k^{\times}$. Then for a simple $G$-module $L$, let $L_{0}=L^{\mathfrak{g}_{1}}\subseteq L$ be the invariants of $\mathfrak{g}_{1}$ so that $L=L_{0}\oplus\mathfrak{g}_{-1}L_{0}\oplus\Lambda^{2}\mathfrak{g}_{-1}L_{0}\oplus\cdots\oplus\Lambda^{top}\mathfrak{g}_{-1}L_{0}$ defines $\mathbb{Z}$-grading on $L$. Write $L_{-i}=\Lambda^{i}\mathfrak{g}_{-1}L_{0}$, and define the operator $T_{c}\in\operatorname{End}(L)$ by declaring that $T_{c}$ preserves the $\mathbb{Z}$-grading and $T_{c}$ acts on $L_{-i}$ by the scalar $c^{-i}$. ###### Lemma 10.22. The submodule $L^{*}\boxtimes L\subseteq k[G]$ contains a unique $G_{c}$-invariant function $f_{c}$ such that $f_{c}(eG)=\sum\limits_{i\geq 0}(-1)^{i}c^{i}\operatorname{dim}L_{-i}.$ ###### Proof. The operator $T_{c}$ defined above is $G_{c^{-1}}$-invariant in $\operatorname{End}(L)$. If we apply the braiding isomorphism $L\otimes L^{*}\cong L^{*}\otimes L$, $T_{c}$ becomes a $G_{c}$-invariant element, which we write as $f_{c}$. It is now straightforward to check the above formula for $f_{c}(eG)$. ∎ ###### Definition 10.23. Let $L$ be a simple $G$-module with the $\mathbb{Z}$-grading as above. Define the polynomial $p_{L}\in\mathbb{Z}[c]$ to be $p_{L}(c)=f_{c}(eG)=\sum\limits_{i\geq 0}(-1)^{i}c^{i}\operatorname{dim}L_{-i}.$ Observe that $0\leq\deg p_{L}\leq\operatorname{dim}\mathfrak{g}_{-1}$. ###### Lemma 10.24. If $L$ is projective, then $p_{L}(c)\neq 0$ if and only if $c\neq 1$. In fact, $p_{L}=\operatorname{dim}L_{0}(1-c)^{\operatorname{dim}\mathfrak{g}_{1}}.$ ###### Proof. The second statement clearly implies the first, and it follows from the fact that $L=\operatorname{Ind}_{\mathfrak{g}_{0}\oplus\mathfrak{g}_{1}}^{\mathfrak{g}}L_{0}$ is a Kac-module (see for instance [Kac77]). Thus we have $p_{L}(c)=\sum\limits_{i}(-1)^{i}c^{i}\operatorname{dim}L_{0}\begin{pmatrix}\operatorname{dim}\mathfrak{g}_{1}\\\ i\end{pmatrix}=\operatorname{dim}L_{0}(1-c)^{\operatorname{dim}\mathfrak{g}_{1}}.$ However we may give another proof which generalizes to other situations. If $L$ is projective of highest weight $\lambda$, then by 10.3 we have $HC(\operatorname{ev}_{eK}v_{\mathfrak{g}_{c}})(\lambda)\neq 0$. Thus $f_{c}\in L^{*}\boxtimes L$ must not vanish at $eG$, i.e. we must have $p_{L}(c)\neq 0$. ∎ ###### Remark 10.25. It is now possible to define, for $c\in k^{\times}$ and a simple $G$-module $L$, the $c$-graded dimension of $L$ to be $p_{L}(c)$. This definition also naturally arises if we consider type $I$ algebras as Lie algebra objects in the tensor category of $\mathbb{Z}$-graded vector spaces with the symmetric structure lifted from the category of super vector spaces. Observe that $p_{L}(1)=\operatorname{dim}L$ and $p_{L}(-1)=\operatorname{sdim}(L)$. ###### Remark 10.26. It would be interesting to understand the roots of $p_{L}$ for irreducible $L$, and in particular the order of vanishing at $c=1$, in terms of the representation theory of $L$. For instance, $L$ is maximally atypical if and only if $p_{L}(1)\neq 0$, while $L$ is projective if and only if the order of vanishing at 1 is $\operatorname{dim}\mathfrak{g}_{-1}$. ### 10.7. Limiting to the center Write $\operatorname{Aut}:=\operatorname{Aut}(\mathfrak{g},\mathfrak{g}_{\overline{0}})$, which is an algebraic group, and let $S\subseteq\operatorname{Aut}$ denote the subset of automorphisms without nonzero fixed vectors in $\mathfrak{g}_{\overline{1}}$. Then $S$ is open in $\operatorname{Aut}$, and further is nonempty because $\delta\in S$. If $\operatorname{Aut}$ has dimension bigger than 0 and $\operatorname{Aut}^{0}\cap S$ is nonempty, $\operatorname{Aut}^{0}\cap S$ will be Zariski dense. Thus if we choose $(\phi_{c})_{c\in k^{\times}}\subseteq S$ such that $\lim\limits_{c\to 0}\phi_{c}=\operatorname{id}_{\mathfrak{g}}$, it is reasonable to consider, for $u\in\mathcal{Z}(\mathcal{U}\mathfrak{g}_{\overline{0}})$, the element $\lim\limits_{c\to 0}\operatorname{ad}_{\phi_{c}}(v_{\mathfrak{g}})(u).$ This limit exists and is equal to $\operatorname{ad}(v_{\mathfrak{g}})(u)\in\mathcal{Z}$. However it is quite possibly zero, and in particular the above limit need not preserve Harish- Chandra polynomials. A more fruitful approach is to choose elements (if they exist) $u_{c}\in\mathcal{A}_{\phi_{c}}$ for each $c$ such that $HC(u_{c})$ is constant. Then we may consider the limit (if it exists): $u_{0}:=\lim\limits_{c\to 0}u_{c}.$ If $u_{0}$ does exist then it must be in $\mathcal{Z}$ and have $HC(u_{0})=HC(u_{c})$ for all $c$. However in general such a limit need not exist, for example in the case that $HC(\mathcal{A}_{c})=0$ for all $c$. However for type I basic algebras this limit does exist, which we now prove. Thus let $\mathfrak{g}$ be a type $I$ basic Lie superalgebra, and let $N=\operatorname{dim}\mathfrak{g}_{1}$. ###### Definition 10.27. Consider the filtration by degree on $S(\mathfrak{h})$, and pull this back under $HC$ to each $\mathcal{A}_{\zeta_{N}^{i}}$ to obtain a filtration $K^{\bullet}\mathcal{A}_{\zeta_{N}^{i}}$. Then let $G^{\bullet}\mathcal{Z}_{full}$ be the filtration on $\mathcal{Z}_{full}$ given by $G^{n}\mathcal{Z}_{full}:=\sum\limits_{i}K^{n}\mathcal{A}_{\zeta_{N}^{i}}.$ This defines an algebra filtration on $\mathcal{Z}_{full}$ such that $G^{n}\mathcal{Z}_{full}$ is finite-dimensional for each $n$. Now we have the following: ###### Lemma 10.28. For each $n\in\mathbb{N}$, there exists typical integral dominant weights $\lambda_{1},\dots,\lambda_{s}$ such that map $G^{n}\mathcal{Z}_{full}\to\operatorname{End}(L(\lambda_{1})\oplus\cdots\oplus L(\lambda_{s}))$ is injective. ###### Proof. Choose a basis $p_{1},\dots,p_{k}$ of $HC(K^{n}\mathcal{A}_{-1})$, and extend this to a basis of $HC(K^{n}\mathcal{Z})$, $p_{1},\dots,p_{k},p_{k+1},\dots,p_{\ell}$. The integral typical dominant weights are Zariski dense in $\mathfrak{h}^{*}$, so necessarily there exists typical integral dominant weights $\lambda_{1},\dots,\lambda_{s}$ such that evaluation at these points induces an injective map on $HC(K^{n}\mathcal{Z})$. Now consider the map $\phi:G^{n}\mathcal{Z}_{full}\to\operatorname{End}(L(\lambda_{1})\oplus\cdots\oplus L(\lambda_{s})),$ and suppose it is not injective. We may write an arbitrary element in $G^{n}\mathcal{Z}_{full}$ as $\sum\limits_{0\leq j\leq N-1}\sum\limits_{1\leq i\leq k}\alpha_{i,j}a_{i,j}+\sum\limits_{k<i\leq\ell}\beta_{i}z_{i}$ where $z_{i}\in K^{n}\mathcal{Z}$ and $a_{i,j}\in K^{n}\mathcal{A}_{\zeta_{N}^{j}}$ such that $HC(z_{i})=p_{i}$ and $HC(a_{i,j})=p_{i}$ for all valid $i$. Now suppose that $\sum\limits_{0\leq j\leq N-1}\sum\limits_{1\leq i\leq k}\alpha_{i,j}\phi(a_{i,j})+\sum\limits_{k<i\leq\ell}\beta_{i}\phi(z_{i})=0$ Looking at the action on the highest weight vectors, this implies by our choice of $\lambda_{1},\dots,\lambda_{s}$ that $\sum\limits_{1\leq i\leq k,\zeta}\alpha_{i,j}p_{i}+\sum\limits_{k<i\leq\ell}\beta_{i}p_{i}=0.$ Thus we must have $\beta_{i}=0$ for all $i$, and $\sum\limits_{0\leq j\leq N-1}\alpha_{i,j}=0$ for all $i$. Looking further at the action on the $(-r)$th graded component of $L(\lambda_{1})\oplus\cdots\oplus L(\lambda_{s})$ according to the grading defined in 10.10, we find that $\sum\limits_{i,j}\alpha_{i,j}\zeta^{rj}p_{i}=0$ which implies that $\sum\limits_{j}\alpha_{i,j}\zeta^{jr}=0$ for all $i,r$. By the nonsingularity of the Vandermonde matrix, this implies $\alpha_{i,j}=0$, and we are done. ∎ ###### Corollary 10.29. Choose an element $p\in HC(\mathcal{A}_{-1})$, and let $a_{\lambda,p}\in\mathcal{A}_{\lambda}$ satisfy $HC(a_{\lambda,p})=p$ for all $\lambda\in k^{\times}$. Then as $\lambda\to 1$, $a_{\lambda,p}$ converges in $G^{2\deg p}\mathcal{Z}_{full}$ to the unique central element $z_{p}$ with $HC(z_{p})=p$. ###### Proof. Choose an embedding $\phi:G^{2\deg p}\mathcal{Z}_{full}\to\operatorname{End}(L(\lambda_{1})\oplus\cdots\oplus L(\lambda_{s}))$ using lemma 10.28. Now in $\operatorname{End}(L(\lambda_{i}))$ we have that $a_{\lambda,p}-z_{p}$ acts on $\Lambda^{j}\mathfrak{g}_{-1}\otimes L_{0}(\lambda_{i})$ as $p(\lambda_{i})(\lambda^{i}-1).$ Thus if we take $\lambda\to 1$ we get convergence as elements of $\operatorname{End}(L(\lambda_{1})\oplus\cdots\oplus L(\lambda_{s}))$. Since $\phi$ is an injective linear map, we are done. ∎ ###### Remark 10.30. It is now possible, in principle, to obtain explicit formulas for the elements of $\mathcal{Z}$ whose Harish-Chandra image lies in $HC(\mathcal{A}_{-1})$. In the case of $\mathfrak{g}\mathfrak{l}(1|1)$ for instance, this would give all elements of the center with trivial constant term. For example, we may produce the known formula for the element of the center whose Harish-Chandra polynomial is a scalar multiple of $t_{\mathfrak{g}}=\prod\limits_{\alpha\in\Delta_{1}^{+}}(h_{\alpha}+(\alpha,\rho)).$ Let $u_{1},\dots,u_{N}$ be a basis of $\mathfrak{g}_{1}$ and $v_{1},\dots,v_{N}$ a basis of $\mathfrak{g}_{-1}$, and write $V=v_{1}\cdots v_{N}\in\mathcal{U}\mathfrak{g}$. Then by 6.11, $v_{\mathfrak{g}}=u_{1}\cdots u_{N}v_{1}\cdots v_{N}$, and we see that $\displaystyle\operatorname{ad}_{c}(v_{\mathfrak{g}})(1)$ $\displaystyle=$ $\displaystyle(1-c)^{N}\operatorname{ad}_{c}(u_{1}\cdots u_{N})(V)$ $\displaystyle=$ $\displaystyle(1-c)^{N}\sum\limits_{I\subseteq\\{1,\dots,N\\}}(-1)^{Nl+i_{1}+\dots+i_{l}}c^{-l}u_{I^{c}}V\tilde{u_{I}}.$ Here the sum runs over all subsets $I$ of $\\{1,\dots,N\\}$, and we write $l=|I|$. Here $I^{c}$ is the complement of $I$ as a set. Further, we define for a subset $J=\\{j_{1}<\dots<j_{l}\\}\subseteq\\{1,\dots,N\\}$, $v_{J}=v_{j_{1}}\cdots v_{j_{l}},\ \ \ \ \tilde{v_{J}}=v_{j_{l}}\cdots v_{j_{1}}.$ If we divide $\operatorname{ad}_{c}(v_{\mathfrak{g}})(1)$ by $(1-c)^{N}$, the Harish-Chandra projection of the resulting element is constant and equal to $HC(u_{1}\cdots u_{N}v_{1}\cdots v_{N})$, which is $t_{\mathfrak{g}}$ up to a constant. Taking the limit $c\to 1$ we obtain the following element of the center: $\sum\limits_{I\subseteq\\{1,\dots,n\\}}(-1)^{nl+i_{1}+\dots+i_{l}}u_{I^{c}}V\tilde{u_{I}}.$ In general the above process will not be as straightforward, as for a general $z_{0}\in\mathcal{Z}(\mathcal{U}\mathfrak{g}_{\overline{0}})$ we have that $\operatorname{ad}_{c}(v_{\mathfrak{g}})(z_{0})=(1-c)^{N}\sum\limits_{I\subseteq\\{1,\dots,n\\}}(-1)^{nl+i_{1}+\dots+i_{l}}c^{-l}u_{I^{c}}Vz_{0}\tilde{u_{I}}+l.o.t$ where $l.o.t.$ denotes terms of lower order in $(1-c)$. Thus we cannot divide by $(1-c)^{N}$. An (albeit tedious) way to overcome this is to take $k[c,(1-c)^{-1}]$ linear combinations of elements $\operatorname{ad}_{c}(v_{\mathfrak{g}})(z_{0})$ in order to obtain an element in $\mathcal{A}_{c}$ with constant Harish-Chandra polynomial. For $\mathfrak{g}\mathfrak{l}(1|1)$ this could certainly be worked out, but for higher rank superalgebras the elements of the center of $\mathcal{U}\mathfrak{g}_{\overline{0}}$ are more complicated, making this process more challenging. However we note that in [Gor04] a method for algorithmically computing any element of $\mathcal{Z}$ with a given Harish-Chandra projection is given, based off an idea originally due to Kac. ## 11\. Appendix In this section111Thanks so much to Maria Gorelik and Vera Serganova for discussions that inspired the results of this section. we generalize the results of section 8 and section 9 to the case where $G$ is an arbitrary quasireductive supergroup, i.e. we remove the assumption that it is Cartan- even. We will define Cartan subspaces and the Harish-Chandra homomorphism, with the ultimate goal of proving 9.11. ### 11.1. Cartan subspaces and the Iwasawa decomposition Let $G$ be an arbitrary quasireductive supergroup with involution $\theta$ and a corresponding symmetric subgroup $K$. As always write $\mathfrak{g}=\mathfrak{k}\oplus\mathfrak{p}$ for the eigenspace decomposition of $\theta$ on $\mathfrak{g}$. Choose a Cartan subspace $\mathfrak{a}_{\overline{0}}\subseteq\mathfrak{p}_{\overline{0}}$, and extend it to a $\theta$-stable Cartan subalgebra $\mathfrak{h}_{\overline{0}}\subseteq\mathfrak{g}_{\overline{0}}$. Then let $\mathfrak{h}=\mathfrak{c}(\mathfrak{h}_{\overline{0}})$, so that $\mathfrak{h}$ is a $\theta$-stable Cartan subalgebra of $\mathfrak{g}$. We may then write $\mathfrak{h}=\mathfrak{t}\oplus\mathfrak{a}$ for the eigenspace decomposition of $\theta$, where $\mathfrak{t}$ is fixed and $\mathfrak{a}$ is the $(-1)$-eigenspace. ###### Definition 11.1. We define $\mathfrak{a}$ to be a Cartan subspace of $\mathfrak{p}$ for the pair $(\mathfrak{g},\mathfrak{k})$. ###### Remark 11.2. It is known that all choices of $\mathfrak{h}_{\overline{0}}$ constructed in this way are conjugate under $K_{0}$, thus all Cartan subalgebras constructed in this way are too, so that a Cartan subspace $\mathfrak{a}$ is well-defined up to conjugation by $K_{0}$. ###### Remark 11.3 (Caution). If $\mathfrak{a}\neq\mathfrak{a}_{\overline{0}}$ then $\mathfrak{a}$ need not be a subalgebra of $\mathfrak{g}$. Indeed, $[\mathfrak{a}_{\overline{1}},\mathfrak{a}_{\overline{1}}]\subseteq\mathfrak{t}_{\overline{0}}$. We may decompose $\mathfrak{g}$ into eigenspaces under the action of $\mathfrak{a}_{\overline{0}}$, and write $\overline{\Delta}$ for the non-zero weights of this action. Then we may choose a decomposition $\overline{\Delta}=\overline{\Delta}^{+}\sqcup\overline{\Delta}^{-}$ into positive and negative roots. Define $\mathfrak{n}=\bigoplus\limits_{\overline{\alpha}\in\overline{\Delta}^{+}}\mathfrak{g}_{\overline{\alpha}}.$ ###### Definition 11.4. We say that the supersymmetric pair $(\mathfrak{g},\mathfrak{k})$ admits an Iwasawa decomposition if we have a decomposition $\mathfrak{g}=\mathfrak{k}\oplus\mathfrak{a}\oplus\mathfrak{n}$ for some choice of $\mathfrak{n}$ as above. ### 11.2. Harish-Chandra homomorphism We suppose from now on that $(\mathfrak{g},\mathfrak{k})$ admits an Iwasawa decomposition. Consider the Lie superalgebra $\mathfrak{t}_{\overline{0}}\oplus\mathfrak{a}$. Then $\mathfrak{t}_{\overline{0}}$ is a central subalgebra and contains the derived subalgebra, so we may take the quotient by it to obtain the abelian Lie superalgebra $\mathfrak{t}_{\overline{0}}\oplus\mathfrak{a}/\mathfrak{t}_{\overline{0}}$. We write $\mathfrak{A}:=\mathcal{U}(\mathfrak{t}_{\overline{0}}\oplus\mathfrak{a}/\mathfrak{t}_{\overline{0}})$, and we view it as the supersymmetric polynomial algebra on $\mathfrak{a}$. If we restrict the natural map $\mathcal{U}\mathfrak{g}\to\mathcal{U}\mathfrak{g}/\mathcal{U}\mathfrak{g}\mathfrak{k}$ to $\mathcal{U}(\mathfrak{t}_{\overline{0}}\oplus\mathfrak{a})$, we obtain the projection $\mathcal{U}(\mathfrak{t}_{\overline{0}}\oplus\mathfrak{a})\to\mathcal{U}(\mathfrak{t}_{\overline{0}}\oplus\mathfrak{a})/(\mathfrak{t}_{\overline{0}})\cong\mathfrak{A},$ so that $\mathfrak{A}$ is naturally subspace of $\mathcal{U}\mathfrak{g}/\mathcal{U}\mathfrak{g}\mathfrak{k}$. Further, by the PBW theorem we have a decomposition $\mathcal{U}\mathfrak{g}/\mathcal{U}\mathfrak{g}\mathfrak{k}=\mathfrak{A}\oplus\mathfrak{n}\mathcal{U}\mathfrak{g}/(\mathfrak{n}\mathcal{U}\mathfrak{g}\cap\mathcal{U}\mathfrak{g}\mathfrak{k}).$ ###### Definition 11.5. We define the Harish-Chandra homomorphism $HC:\mathcal{U}\mathfrak{g}/\mathcal{U}\mathfrak{g}\mathfrak{k}\to\mathfrak{A}$ to be the projection along $\mathfrak{n}\mathcal{U}\mathfrak{g}/(\mathfrak{n}\mathcal{U}\mathfrak{g}\cap\mathcal{U}\mathfrak{g}\mathfrak{k})$. Notice that this agrees with the usual Harish-Chandra map when $\mathfrak{h}=\mathfrak{h}_{\overline{0}}$, i.e. $G$ is Cartan-even. Now $\mathfrak{t}_{\overline{0}}\oplus\mathfrak{a}_{\overline{1}}$ is a subalgebra of $\mathfrak{k}^{\prime}$, and this acts on $\mathfrak{A}$ by left multiplication on the quotient, and $\mathfrak{t}_{\overline{0}}$ acts trivially while $\mathfrak{a}_{\overline{1}}$ acts freely. Thus we obtain a free action of the exterior algebra on $\mathfrak{a}_{\overline{1}}$ on $\mathfrak{A}$, and therefore the invariants of this action are given by $\mathfrak{A}^{\mathfrak{a}_{\overline{1}}}=S(\mathfrak{a}_{\overline{0}})\Lambda^{top}\mathfrak{a}_{\overline{1}}.$ Since the decomposition $\mathcal{U}\mathfrak{g}/\mathcal{U}\mathfrak{g}\mathfrak{k}=\mathfrak{A}\oplus\mathfrak{n}\mathcal{U}\mathfrak{g}/(\mathfrak{n}\mathcal{U}\mathfrak{g}\cap\mathcal{U}\mathfrak{g}\mathfrak{k})$ is $\mathfrak{t}_{\overline{0}}\oplus\mathfrak{a}_{\overline{1}}$-invariant, we clearly obtain the following lemma. ###### Lemma 11.6. The Harish-Chandra morphism restricts to a map $HC:\mathcal{A}_{G/K}\to\mathfrak{A}^{\mathfrak{a}_{\overline{1}}}=S(\mathfrak{a}_{\overline{0}})\Lambda^{top}\mathfrak{a}_{\overline{1}}.$ ###### Remark 11.7. By lemma 11.6, it follows that given a ghost distribution $\gamma\in\mathcal{A}_{G/K}$, we may obtain a polynomial in $S(\mathfrak{a}_{\overline{0}})$ by writing $HC(\gamma)=p_{\gamma}\xi$, where $\xi\in\Lambda^{top}\mathfrak{a}_{\overline{1}}$ is some chosen nonzero element. Thus $p_{\gamma}$ is well-defined up to a scalar, so that its vanishing set is well-defined. The following lemma and corollary are proven in the same way as in lemma 8.8, although note that the statement is slightly different. ###### Lemma 11.8. Give $\mathcal{U}\mathfrak{g}/\mathcal{U}\mathfrak{g}\mathfrak{k}$ the same filtration $F^{\bullet}$ as defined in lemma 8.8. Viewing $\mathfrak{A}$ as the supersymmetric polynomial algebra on $\mathfrak{a}$, give it a grading by degree, where elements of $\mathfrak{a}_{\overline{0}}$ have degree one and elements of $\mathfrak{a}_{\overline{1}}$ have degree $1/2$. Then we have $HC(F^{r})\subseteq\sum\limits_{s\leq r/2}\mathfrak{A}^{s}.$ ###### Corollary 11.9. Let $z\in\operatorname{Dist}^{r}(G_{0}/K_{0},eK_{0})^{K_{0}}$ lie in the $r$th part of the standard filtration on $\operatorname{Dist}(G_{0}/K_{0},eK_{0})$ defined in definition 2.2. Then $v_{\mathfrak{k}^{\prime}}\cdot z\in\mathcal{A}_{G/K}$ has $HC(v_{\mathfrak{k}^{\prime}}\cdot z)\in\mathfrak{A}^{r+\operatorname{dim}\mathfrak{p}_{\overline{1}}/2}.$ Further, in the notation of remark 11.7, $\deg p_{v_{\mathfrak{k}^{\prime}}\cdot z}\leq\operatorname{dim}\mathfrak{n}_{\overline{1}}/2+r$. ###### Proof. The first statement follows from lemma 11.8. The second statement follows from remark 11.7, where we showed that $HC(v_{\mathfrak{k}^{\prime}}\cdot z)=p_{\gamma}\xi$ for some $p_{\gamma}\in S(\mathfrak{a}_{\overline{0}})$ and a non-zero element $\xi\in\Lambda^{top}\mathfrak{a}_{\overline{1}}$. Since $\deg\xi=\operatorname{dim}\mathfrak{a}_{\overline{1}}/2$, and $\operatorname{dim}\mathfrak{p}_{\overline{1}}=\operatorname{dim}\mathfrak{a}_{\overline{1}}+\operatorname{dim}\mathfrak{n}_{\overline{1}}$, the bound follows. ∎ ### 11.3. Ghost distributions Since $(\mathfrak{g},\mathfrak{k})$ admits an Iwasawa decomposition, it will admit an open orbit at $eK$ under an Iwasawa Borel subgroup $B$, whose Lie superalgebra contains $\mathfrak{a}\oplus\mathfrak{n}$. We write $\Lambda^{+}\subseteq\mathfrak{a}_{\overline{0}}^{*}$ for the set of $B$-dominant weights $\lambda$ such that there exists an irreducible $B$-submodule of highest weight $\lambda$ in $k[G/K]$. Let $\lambda\in\Lambda^{+}$, and let $L_{\lambda}$ be an irreducible $B$-submodule of $k[G/K]$ of highest weight $\lambda$. In particular $L_{\lambda}$ is an irreducible $\mathfrak{h}$-module. Because $B$ admits an open orbit at $eK$ we must have that $\operatorname{ev}_{eK}:L_{\lambda}\to k$ is non-zero. ###### Remark 11.10. Unlike in the Cartan-even case, there is no multiplicity-free statement for irreducible $B$-submodules of $k[G/K]$. This is due to the fact that the Cartan subalgebra is no longer abelian so its irreducible representations need not be dimension 1. Therefore our choice of $L_{\lambda}$ is not unique. Although in the case $G\times G/G$ irreducible submodules we will have multiplicity-freenes, due to the extraordinary amount of symmetry present. Let $\gamma\in\mathcal{A}_{G/K}$. Then $HC(\gamma)$ defines a functional on $L_{\lambda}$. For what follows, note that lemma 8.4 still holds here, i.e. $\langle K^{\prime}\cdot L_{\lambda}\rangle=\mathcal{U}\mathfrak{k}^{\prime}L_{\lambda}$. ###### Lemma 11.11. If $HC(\gamma):L_{\lambda}\to k$ is nonzero then $\mathcal{U}\mathfrak{k}^{\prime}L_{\lambda}$ contains a copy of $I_{K^{\prime}}(k)$. Further, $\mathcal{U}\mathfrak{k}^{\prime}L_{\lambda}$ contains a copy of $I_{K^{\prime}}(k)$ if and only if $HC(\gamma):L_{\lambda}\to k$ is non-zero for some $\gamma\in\mathcal{A}_{G/K}$. ###### Proof. Follows from the work done in section 7.3. ∎ Consider the irreducible $\mathfrak{h}$-module $L_{\lambda}$. We have already noted that $\mathfrak{t}_{\overline{0}}$ acts trivially, thus it is really a $\mathfrak{h}/\mathfrak{t}_{\overline{0}}$-module. Both $\mathfrak{a}_{\overline{1}}$ and $\mathfrak{t}_{\overline{1}}$ sit inside the quotient superalgebra as commutative subalgebras. The action of $\mathfrak{h}/\mathfrak{t}_{\overline{0}}$ on $L_{\lambda}$ is given by the unique irreducible super representation of the Clifford superalgebra $Cl(\mathfrak{h}_{\overline{1}},(-,-)_{\lambda})$, where $(u,v)_{\lambda}=\lambda([u,v]).$ Thus $W_{\lambda}:=\operatorname{ker}(-,-)_{\lambda}$ acts trivially on $L_{\lambda}$, and $\mathfrak{a}_{\overline{1}}/(W_{\lambda}\cap\mathfrak{a}_{\overline{1}})$ and $\mathfrak{t}_{\overline{1}}/(W_{\lambda}\cap\mathfrak{t}_{\overline{1}})$ define complimentary maximal isotropic subspaces of $\mathfrak{h}_{\overline{1}}/\operatorname{ker}(-,-)_{\lambda}$. ###### Lemma 11.12. 1. (1) As an $\mathfrak{a}_{\overline{1}}$-module, $L_{\lambda}$ is isomorphic to $\Lambda^{\bullet}\left(\mathfrak{a}_{\overline{1}}/(W_{\lambda}\cap\mathfrak{a}_{\overline{1}})\right)$; 2. (2) as an $\mathfrak{t}_{\overline{1}}$-module, $L_{\lambda}$ is isomorphic to $\Lambda^{\bullet}\left(\mathfrak{t}_{\overline{1}}/(W_{\lambda}\cap\mathfrak{t}_{\overline{1}})\right)$. Further, the socle of $L_{\lambda}$ as an $\mathfrak{a}_{\overline{1}}$-module generates $L_{\lambda}$ as a $\mathfrak{t}_{\overline{1}}$-module, and the socle as a $\mathfrak{t}_{\overline{1}}$-module generates it as an $\mathfrak{a}_{\overline{1}}$-module. ###### Proof. Both of these lemma follow from the fact that we may realize the irreducible representation of an even-dimensional Clifford algebra as an exterior algebra $\Lambda^{\bullet}\langle\xi_{1},\dots,\xi_{n}\rangle$. Then if $V_{1},V_{2}$ are complimentary maximal isotropic subspaces, we may realize $V_{1}$ as acting by multiplication by $\xi_{1},\dots,\xi_{n}$, and $V_{2}$ as acting by $\partial_{\xi_{1}},\dots,\partial_{\xi_{n}}$. ∎ ###### Proposition 11.13. $\mathcal{U}\mathfrak{k}^{\prime}L_{\lambda}$ contains at most one copy of $I_{K^{\prime}}(k)$. ###### Proof. The $\mathfrak{k}^{\prime}$-semicoinvariants of $\mathcal{U}\mathfrak{k}^{\prime}L_{\lambda}$ are determined by their values on $L_{\lambda}$, and give rise to an $\mathfrak{a}_{\overline{1}}$-coinvariant on $L_{\lambda}$. However as an $\mathfrak{a}_{\overline{1}}$-module $L_{\lambda}$ admits a unique $\mathfrak{a}_{\overline{1}}$-coinvariant by lemma 11.12, so there can be at most one $\mathfrak{k}^{\prime}$-semicoinvariant on $\mathcal{U}\mathfrak{k}^{\prime}L_{\lambda}$ up to scalar. Thus at most one copy of $I_{K^{\prime}}(k)$ can arise. ∎ ###### Lemma 11.14. If $\mathcal{U}\mathfrak{k}^{\prime}L_{\lambda}$ contains a copy of $I_{K^{\prime}}(k)$, then $\operatorname{ker}(-,-)_{\lambda}\cap\mathfrak{a}_{\overline{1}}=0$, or equivalently $L_{\lambda}\cong\Lambda^{\bullet}\mathfrak{a}_{\overline{1}}$ as an $\mathfrak{a}_{\overline{1}}$-module. ###### Proof. Write $\mathcal{U}\mathfrak{k}^{\prime}L_{\lambda}=I_{K^{\prime}}(k)\oplus M$ for some complimentary submodule $M$. Then the restriction of $I_{K^{\prime}}(k)$ to $\mathfrak{t}_{\overline{0}}\oplus\mathfrak{a}_{\overline{1}}$ must remain injective. In particular if $g\in I_{K^{\prime}}(k)$ generates the head, then we may assume it is a weight vector of weight $0$ so that $\mathfrak{t}_{\overline{0}}g=0$, and further we will have $\Lambda^{top}\mathfrak{a}_{\overline{1}}g\neq 0$. If we choose $f\notin\mathfrak{a}_{\overline{1}}L_{\lambda}$, then $\mathcal{U}\mathfrak{k}^{\prime}f$ will generate the copy of $I_{K^{\prime}}(k)$ and thus we must also have $\Lambda^{top}\mathfrak{a}_{\overline{1}}f\neq 0$. The statement follows. ∎ ###### Proposition 11.15. Let $\gamma\in\mathcal{A}_{G/K}$ and write $HC(\gamma)=p_{\gamma}\xi$ as in remark 11.7. Then the following are equivalent: 1. (1) $HC(\gamma):L_{\lambda}\to k$ is nonzero; 2. (2) $\mathcal{U}\mathfrak{k}^{\prime}L_{\lambda}$ contains a copy of $I_{K^{\prime}}(k)$ and $p_{\gamma}(\lambda)\neq 0$; 3. (3) $L_{\lambda}$ is a projective $\mathfrak{a}_{\overline{1}}$-module and $p_{\gamma}(\lambda)\neq 0$. ###### Proof. Clearly $(2)\Rightarrow(3)$. If either $(3)$ or (1) hold, then by lemma 11.14 $L_{\lambda}$ is projective over $\mathfrak{t}_{\overline{0}}\oplus\mathfrak{a}_{\overline{1}}$, so in these cases let $f\notin\mathfrak{a}_{\overline{1}}L_{\lambda}$ so that $\mathcal{U}\mathfrak{k}^{\prime}f$ generates $I_{K^{\prime}}(k)$ whenever $I_{K^{\prime}}(k)$ is a submodule of $\mathcal{U}\mathfrak{k}^{\prime}L_{\lambda}$. Then $HC(\gamma)$ is non-zero on $L_{\lambda}$ if and only if $HC(\gamma)(f)\neq 0$. However $HC(\gamma)(f)=p_{\gamma}(\lambda)(\xi f)(eK).$ Since $\xi f$ spans the socle of $L_{\lambda}$ as an $\mathfrak{a}_{\overline{1}}$-module, it must be non-vanishing under the unique (up to scalar) $\mathfrak{t}_{\overline{1}}$-coinvariant on $L_{\lambda}$ by lemma 11.12. Since $\operatorname{ev}_{eK}:L_{\lambda}\to k$ is non-zero and is a $\mathfrak{t}_{\overline{1}}$-coinvariant, necessarily $(\xi f)(eK)\neq 0$. Thus $HC(\gamma)(f)\neq 0$ if and only if $p_{\gamma}(\lambda)\neq 0$. From these arguments $(3)\Rightarrow(1)$ and $(1)\Rightarrow(2)$ follow. ∎ We have a linear map $\mathfrak{a}_{\overline{1}}\otimes\mathfrak{a}_{\overline{0}}^{*}\to\mathfrak{t}_{\overline{1}}^{*}$ given by $(u\otimes\lambda)\mapsto(u,-)_{\lambda}.$ Let $U_{reg}\subseteq\mathfrak{a}_{\overline{0}}^{*}$ denote the locus where this defines an injective morphism $\mathfrak{a}_{\overline{1}}\to\mathfrak{t}_{\overline{1}}^{*}$. Clearly $U_{reg}$ is Zariski open, although it need not be nonempty. Then we have shown: ###### Corollary 11.16. If $\lambda\in U_{reg}\cap\Lambda^{+}$, then $HC(\gamma)$ is non-zero on $L_{\lambda}$ if and only if $p_{\gamma}(\lambda)\neq 0$. Further, we clearly have the following sufficient condition of when $U_{reg}$ is empty. ###### Corollary 11.17. If $\operatorname{dim}\mathfrak{a}_{\overline{1}}>\operatorname{dim}\mathfrak{t}_{\overline{1}}$, or if there exists a nonzero element $u\in\mathfrak{a}_{\overline{1}}$ such that $[u,\mathfrak{t}_{\overline{1}}]=0$, then $U_{reg}=\emptyset$. We also note the following proposition, which is proven in the exact same way as 8.18. ###### Proposition 11.18. Suppose that $L$ is an irreducible $G$-submodule of $k[G/K]$ of $B$-highest weight $\lambda$, and suppose that $L$ contains a copy of $I_{K^{\prime}}(k)$ (equivalently $HC(\gamma):L_{\lambda}\to k$ is nonzero for some $\gamma\in\mathcal{A}_{G/K}$). Then $I_{G}(L)$ is a submodule of $k[G]^{\operatorname{ber}_{\mathfrak{k}^{\prime}}}$. ### 11.4. The distribution $\operatorname{ev}_{eK}v_{\mathfrak{k}^{\prime}}$ The distribution $\operatorname{ev}_{eK}v_{\mathfrak{k}^{\prime}}$ arises in the same way as in section 8.6, and the analogous results of section 8.6 still hold. We write them out now. Let $HC(\operatorname{ev}_{eK}v_{\mathfrak{k}^{\prime}})=p_{G/K}\xi$ as in the notation of remark 11.7. ###### Proposition 11.19. The following are equivalent: * • $\operatorname{ev}_{eK}v_{\mathfrak{k}^{\prime}}:L_{\lambda}\to k$ is nonzero; * • $\mathcal{U}\mathfrak{k}^{\prime}L_{\lambda}$ contains a copy of $I_{K^{\prime}}(k)$, and the $K^{\prime}$-invariant in $I_{K^{\prime}}(k)$ is non-vanishing at $eK$; * • $L_{\lambda}$ is a projective $\mathfrak{a}_{\overline{1}}$-module and $p_{G/K}(\lambda)\neq 0$. ###### Proof. These follow from 11.15. ∎ ###### Proposition 11.20. Suppose that whenever a $B$-irreducible submodule $L_{\lambda}\subseteq k[G/K]$ has $\mathcal{U}\mathfrak{k}^{\prime}L_{\lambda}$ containing $I_{K^{\prime}}(k)$, the $K^{\prime}$-invariant is non-vanishing at $eK$. Then 1. (1) $\mathcal{U}\mathfrak{k}^{\prime}L_{\lambda}$ contains $I_{K^{\prime}}(k)$ if and only if $L_{\lambda}$ is projective over $\mathfrak{a}_{\overline{1}}$ and $p_{G/K}(\lambda)\neq 0$; and 2. (2) if $HC(\operatorname{ev}_{eK}v_{\mathfrak{k}^{\prime}}):L_{\lambda}\to k$ is zero, then $HC(\gamma):L_{\lambda}\to k$ is zero for all $\gamma\in\mathcal{A}_{G/K}$. ### 11.5. The case $G\times G/G$ Many of the results in section 9 still hold, but we will try to explicitly mention this as much as possible. To begin with, lemma 9.1 still holds. Choose a Cartan subalgebra $\mathfrak{h}\subseteq\mathfrak{g}$. Then $\mathfrak{h}\times\mathfrak{h}$ is a Cartan subalgebra of $\mathfrak{g}\times\mathfrak{g}$ such that $\mathfrak{a}=\\{(h,-h):h\in\mathfrak{h}\\}$ is a Cartan subspace of $\mathfrak{p}$. If we choose a Borel subalgebra $\mathfrak{b}$ of $\mathfrak{g}$ containing $\mathfrak{h}$, then $\mathfrak{b}^{-}\times\mathfrak{b}^{+}$ becomes an Iwasawa Borel subalgebra of $\mathfrak{g}\times\mathfrak{g}$. Again for this choice of Borel subalgebra, both $(\mathfrak{g}\times\mathfrak{g},\mathfrak{g})$ and $(\mathfrak{g}\times\mathfrak{g},\mathfrak{g}^{\prime})$ admit an Iwasawa decomposition for any such choice of Borel subalgebra as in section 9. Further, if $L_{\lambda}\subseteq k[G]$ is an irreducible $B^{-}\times B^{+}$ submodule, then $\mathcal{U}(\mathfrak{g}\times\mathfrak{g})L_{\lambda}=\mathcal{U}\mathfrak{g}L_{\lambda}=\mathcal{U}\mathfrak{g}^{\prime}L_{\lambda}.$ We again have $\Lambda^{+}=\\{(-\lambda,\lambda):\lambda\text{ is a }\mathfrak{b}\text{-dominant weight}\\}$. By abuse of notation we will also write $\Lambda^{+}$ for the set of $B$-dominant weights in $\mathfrak{h}_{\overline{0}}^{*}$. ###### Lemma 11.21. Let $L(\lambda):=L_{B}(\lambda)$ be the irreducible $G$-module of highest weight $\lambda\in\Lambda^{+}$. Then one of the following two must occur: 1. (1) $L(\lambda)^{*}\boxtimes L(\lambda)$ is an irreducible $G\times G$-module and admits a unique even $G^{\prime}$-invariant. 2. (2) $L(\lambda)^{*}\boxtimes L(\lambda)=L\oplus\Pi L$ is a sum the two irreducible $G\times G$-modules $L$ and $\Pi L$ which are parity shifts of one another. In this case, both $L$ and $\Pi L$ admit a unique $G^{\prime}$-invariant, one even and one odd. ###### Proof. If $L(\lambda)\not\cong\Pi L(\lambda)$ then the first case happens. Otherwise the second case happens. ∎ ###### Definition 11.22. For $\lambda\in\Lambda^{+}$, define $d(\lambda)\in\\{0,1\\}$ to be $0$ if $L(\lambda)\not\cong\Pi L(\lambda)$, and $1$ otherwise. If $d(\lambda)=1$, set $\frac{1}{2}L(\lambda)^{*}\boxtimes L(\lambda)$ to be the irreducible $G\times G$-submodule of $L(\lambda)^{*}\boxtimes L(\lambda)$ which contains an even $G^{\prime}$-invariant. ###### Proposition 11.23. We have the following decomposition: $\operatorname{soc}k[G]=\bigoplus\limits_{\lambda\in\Lambda^{+}}\frac{1}{2^{d(\lambda)}}L(\lambda)^{*}\boxtimes L(\lambda).$ Further, the unique $G^{\prime}$-invariant of $\frac{1}{2^{d(\lambda)}}L(\lambda)^{*}\boxtimes L(\lambda)$ evaluates (up to a scalar) at $eG$ to $\text{tr}(L)$. ###### Proof. The irreducible $G\times G$ modules are all of the form $\frac{1}{2^{d(\lambda)d(\mu)}}L(\lambda)\boxtimes L(\mu)$, where $\lambda,\mu\in\Lambda^{+}$. By Frobenius reciprocity, this admits an even $G\times G$-equivariant morphism into $k[G]$ if and only if $\lambda=-\mu$ and the unique $G^{\prime}$-invariant of it is even. The evaluation of the unique $G^{\prime}$-invariant is the same computation as done in 9.4. ∎ ###### Remark 11.24. We have shown that the irreducible $B^{-}\times B^{+}$-submodules of $k[G]$ are given by $\frac{1}{2^{d(\lambda)}}L(\lambda)_{-\lambda}^{*}\boxtimes L(\lambda)_{\lambda}$ and $\mathcal{U}\mathfrak{g}^{\prime}\frac{1}{2^{d(\lambda)}}L(\lambda)_{-\lambda}^{*}\boxtimes L(\lambda)_{\lambda}=\frac{1}{2^{d(\lambda)}}L(\lambda)^{*}\boxtimes L(\lambda).$ Notice that in this setting $\operatorname{dim}\mathfrak{t}_{\overline{1}}=\operatorname{dim}\mathfrak{a}_{\overline{1}}$, so by lemma 11.14 if $\frac{1}{2^{d(\lambda)}}L(\lambda)^{*}\boxtimes L(\lambda)$ contains $I_{G^{\prime}}(k)$ then necessarily $\frac{1}{2^{d(\lambda)}}L(\lambda)_{-\lambda}^{*}\boxtimes L(\lambda)_{\lambda}$ is a projective $\mathfrak{h}\times\mathfrak{h}$-module, and it is not difficult to show this is equivalent to $L(\lambda)_{\lambda}$ being a projective $\mathfrak{h}$-module. ###### Lemma 11.25. $L(\lambda)$ is a projective $G$-module if and only if $\frac{1}{2^{d(\lambda)}}L(\lambda)^{*}\boxtimes L(\lambda)$ contains a copy of $I_{G^{\prime}}(k)$. ###### Proof. The forward direction is clear. Conversely, if this module contains $I_{G^{\prime}}(k)$ then we know by 11.18 that $I_{G\times G}(\frac{1}{2^{d(\lambda)}}L(\lambda)^{*}\boxtimes L(\lambda))$ is a $G\times G$ submodule of $k[G]$. (As in section 9, we once again use that $G\times G/G\cong G\times G/G^{\prime}$ via $\operatorname{id}\times\delta$.) Further, $L(\lambda)_{\lambda}$ must be a projective $H$-module by remark 11.24, so that $L(\lambda)$ has no self- extensions. If there was a non-trivial extension $V$ between $L(\lambda)$ and $L(\mu)$, where $\mu\neq\lambda$, then the matrix coefficients morphism would induce a $G\times G$ morphism $V^{*}\boxtimes V\to k[G]$ such that the image is an indecomposable module with socle $\frac{1}{2^{d(\lambda)}}L(\lambda)^{*}\boxtimes L(\lambda)\oplus\frac{1}{2^{d(\mu)}}L(\mu)^{*}\boxtimes L(\mu).$ However $I_{G\times G}(\frac{1}{2^{d(\lambda)}}L(\lambda)^{*}\boxtimes L(\lambda))$ splits off $k[G]$, so this cannot happen. It follows that $L(\lambda)$ cannot have nontrivial extensions with any modules, and is therefore projective. ∎ Let us look at the Harish-Chandra homomorphism. Translating to the enveloping algebra, it defines a homomorphism $HC:\mathcal{U}\mathfrak{g}\to\mathcal{U}\mathfrak{h}.$ Note that $\mathfrak{h}$ is no longer abelian in general. However this induces a morphism $HC:\mathcal{A}_{G\times G/G}\to\mathcal{A}_{H\times H/H}.$ Since $H_{0}$ is a central subgroup, $\Lambda^{top}\mathfrak{h}_{\overline{1}}$ is a trivial $H_{0}$-module, and so $\mathcal{A}_{H\times H/H}=(\mathcal{U}\mathfrak{h})^{\mathfrak{h}^{\prime}}$. Write $T_{\mathfrak{h}}:=\operatorname{ad}^{\prime}(v_{\mathfrak{h}})(1)$. Then since $\mathfrak{h}_{\overline{0}}$ is central, we clearly have $\mathcal{A}_{H\times H/H}=S(\mathfrak{h}_{\overline{0}})T_{\mathfrak{h}}.$ Since $HC(\operatorname{ad}^{\prime}(v_{\mathfrak{g}})(1))\in(\mathcal{U}\mathfrak{h})^{\mathfrak{h}^{\prime}}$, we may write $HC(\operatorname{ad}^{\prime}(v_{\mathfrak{g}})(1))=p_{1}T_{\mathfrak{h}}$. For each $\lambda\in\mathfrak{h}_{\overline{0}}^{*}$ we have a bilinear form $(-,-)_{\lambda}$ on $\mathfrak{h}_{\overline{1}}$, in other words we have a linear map $(-,-):\mathfrak{h}_{\overline{0}}^{*}\to S^{2}\mathfrak{h}_{\overline{1}}^{*}$. If we choose a basis for $\mathfrak{h}_{\overline{1}}$, then we obtain a map $S^{2}\mathfrak{h}_{\overline{1}}^{*}\to k$ given by taking the determinant of the corresponding bilinear form. The composition defines a degree $\operatorname{dim}\mathfrak{h}_{\overline{1}}$-degree polynomial $b_{H}\in S(\mathfrak{h}_{\overline{0}})$. Note that $b_{H}$ is well-defined only up to scalar. Further, $b_{H}(\lambda)\neq 0$ if and only if the irreducible $H$-module of weight $\lambda$ is projective. ###### Definition 11.26. Define the projectivity polynomial of $G$ with respect to $B$ and $H$ to be $p_{G,B}:=p_{1}b_{H}$. Notice that if $G=H$ we have $p_{G,B}=b_{H}$. The following theorem justifies the name. ###### Theorem 11.27. Let $G$ be a quasireductive supergroup with the Cartan subgroup $H$ and Borel subgroup $B$ containing $H$. Then for a $B$-dominant weight $\lambda$, $p_{G,B}(\lambda)\neq 0$ if and only if $L_{B}(\lambda)$ is projective. Further $p_{G,B}$ is a polynomial of degree at most $\operatorname{dim}\mathfrak{b}_{\overline{1}}$. ###### Proof. If $p_{G,B}(\lambda)\neq 0$, then $b_{H}(\lambda)\neq 0$ which implies that $\frac{1}{2^{d(\lambda)}}L(\lambda)^{*}_{-\lambda}\boxtimes L(\lambda)_{\lambda}$ is projective. Since $p_{1}(\lambda)\neq 0$, by 11.20 $\frac{1}{2^{d(\lambda)}}L(\lambda)^{*}\boxtimes L(\lambda)$ contains $I_{G^{\prime}}(k)$ and so is projective by lemma 11.25. Conversely if $L(\lambda)$ is projective then $L(\lambda)_{\lambda}$ is projective over $H$ so that $b_{H}(\lambda)\neq 0$. Further, $\frac{1}{2^{d(\lambda)}}L(\lambda)^{*}_{-\lambda}\boxtimes L(\lambda)_{\lambda}$ contains $I_{G^{\prime}}(k)$, and so by 11.20 and 11.23 $p_{1}(\lambda)\neq 0$. It follows that $p_{G,B}(\lambda)\neq 0$. ∎ Define $T_{\mathfrak{g}}:=\operatorname{ad}^{\prime}(v_{\mathfrak{g}})(1)\in\mathcal{U}\mathfrak{g}$. Then if $L$ is an irreducible $G$-module, $T_{\mathfrak{g}}$ acts on $L$ by a twisted-invariant operator, and thus is either equal, up to a scalar, to $\delta_{L}$ if $T_{\mathfrak{g}}$ is even, or $\delta_{L}\circ\sigma$ if $T_{\mathfrak{g}}$ is odd, where $\sigma$ is an odd $G$-equivariant automorphism of $L$. In particular it either acts by a linear automorphism or by 0. By 11.27, we have: ###### Corollary 11.28. If $L$ is an irreducible $G$-module, then $T_{\mathfrak{g}}$ acts by an automorphism on $L$ if and only if $L$ is projective, and otherwise $T_{\mathfrak{g}}$ acts by $0$. ## References * [ABF97] Daniel Arnaudon, Michel Bauer, and L Frappat, _On casimir’s ghost_ , Communications in mathematical physics 187 (1997), no. 2, 429–439. * [Ben00] Said Benayadi, _Quadratic Lie superalgebras with the completely reducible action of the even part on the odd part_ , Journal of Algebra 223 (2000), no. 1, 344–366. * [CCF11] Claudio Carmeli, Lauren Caston, and Rita Fioresi, _Mathematical foundations of supersymmetry_ , vol. 15, European Mathematical Society, 2011. * [CW12] Shun-Jen Cheng and Weiqiang Wang, _Dualities and representations of Lie superalgebras_ , American Mathematical Soc., 2012. * [DH+76] DŽ Djoković, G Hochschild, et al., _Semisimplicity of $2$-graded Lie algebras, ii_, Illinois Journal of Mathematics 20 (1976), no. 1, 134–143. * [Gor00] Maria Gorelik, _On the ghost centre of Lie superalgebras_ , Annales de l’institut Fourier, vol. 50, 2000, pp. 1745–1764. * [Gor04] by same author, _The Kac construction of the centre of for Lie superalgebras_ , Journal of Nonlinear Mathematical Physics 11 (2004), no. 3, 325–349. * [Kac77] Victor G Kac, _Lie superalgebras_ , Advances in mathematics 26 (1977), no. 1, 8–96. * [Kos82] Jean-Louis Koszul, _Graded manifolds and graded Lie algebras_ , Proceedings of the International Meeting on Geometry and Physics (Bologna), Pitagora, 1982, pp. 71–84. * [LM94] Edward S Letzter and Ian M Musson, _Complete sets of representations of classical Lie superalgebras_ , letters in mathematical physics 31 (1994), no. 3, 247–253. * [Man13] Yuri I Manin, _Gauge field theory and complex geometry_ , vol. 289, Springer Science & Business Media, 2013. * [MT18] Akira Masuoka and Yuta Takahashi, _Geometric construction of quotients ${G/H}$ in supersymmetry_, arXiv preprint arXiv:1808.05753 (2018). * [Mus12] Ian Malcolm Musson, _Lie superalgebras and enveloping algebras_ , vol. 131, American Mathematical Soc., 2012. * [MZ10] Akira Masuoka and Alexander N Zubkov, _Quotient sheaves of algebraic supergroups are superschemes_ , arXiv preprint arXiv:1007.2236 (2010). * [Ser11] Vera Serganova, _Quasireductive supergroups_ , New developments in Lie theory and its applications 544 (2011), 141–159. * [She19] Alexander Sherman, _Spherical supervarieties_ , arXiv preprint arXiv:1910.09610 (2019). * [She20a] by same author, _Spherical and symmetric supervarieties_ , Ph.D. thesis, UC Berkeley, 2020. * [She20b] by same author, _Two geometric proofs of the classification of algebraic supergroups with semisimple representation theory_ , arXiv preprint arXiv:2012.11317 (2020). * [Tim11] Dmitry A Timashev, _Homogeneous spaces and equivariant embeddings_ , vol. 138, Springer Science & Business Media, 2011. * [VMP90] AA Voronov, Yu I Manin, and IB Penkov, _Elements of supergeometry_ , Journal of Soviet Mathematics 51 (1990), no. 1, 2069–2083. Dept. of Mathematics, Ben Gurion University, Beer-Sheva, Israel Email address<EMAIL_ADDRESS>
# A Note on the Representation Power of GHHs Zhou Lu111This work is done during LZ’s visit to SQZ institution. Princeton University <EMAIL_ADDRESS> (January 2021) ###### Abstract In this note we prove a sharp lower bound on the necessary number of nestings of nested absolute-value functions of generalized hinging hyperplanes (GHH) to represent arbitrary CPWL functions. Previous upper bound states that $n+1$ nestings is sufficient for GHH to achieve universal representation power, but the corresponding lower bound was unknown. We prove that $n$ nestings is necessary for universal representation power, which provides an almost tight lower bound. We also show that one-hidden-layer neural networks don’t have universal approximation power over the whole domain. The analysis is based on a key lemma showing that any finite sum of periodic functions is either non- integrable or the zero function, which might be of independent interest. ## 1 Introduction We consider the complexity of representing continuous piecewise linear functions using the generalized hinging hyperplane model Wang and Sun (2005). We begin with a short review on these two notions. ### 1.1 Continuous Piecewise Linear (CPWL) Functions Continuous piecewise linear (CPWL) functions play an important role in non- linear function approximation, such as nonlinear circuit or neural networks. We introduce the definition of CPWL functions borrowed from Chua and Deng (1988). ###### Definition 1.1 (CPWL function). A function $f(x):R^{n}\to R$ is said to be a CPWL function iff it satisfies: 1):The domain space $R^{n}$ is divided into a finite number of polyhedral regions by a finite number of disjunct boundaries. Each boundary is a subset of a hyperplane and takes non-zero measure (standard lebesgue measure) on the hyperplane (as $R^{n-1}$). 2):The restriction of $f(x)$ on each polyhedral region is an affine function. 3):$f(x)$ is continuous on $R^{n}$. ### 1.2 Generalized Hinging Hyperplanes (GHH) The model of hinging hyperplanes (HH) is a sum of hinges like $\pm\max\\{w_{1}^{\top}x+b_{1},w_{2}^{\top}x+b_{2}\\}$ (1) where $w_{1},w_{2}\in R^{n}$ and $b_{1},b_{2}\in R$ are parameters. The HH model (in fact equivalent to a one hidden-layer ReLU network) can approximate any continuous function over a compact domain to arbitrary precision as the number of hinges go infinity Breiman (1993). However, this model can’t exactly represent all CPWL function as pointed out in He et al. (2018), which brings doubt on its approximation efficiency. To overcome this problem, Wang and Sun (2005) first proposed a generalization of HH model, called GHH which allows more than 2 affine functions within the nested maximum operator: ###### Definition 1.2 ($n$-order hinge). A $n$-order hinge is a function of the following form: $\pm\max\\{w_{1}^{\top}x+b_{1},w_{2}^{\top}x+b_{2},\cdots,w_{n+1}^{\top}x+b_{n+1}\\}$ (2) where $w_{i}\in R^{n}$ and $b_{i}\in R$ are parameters. A linear combination of a finite number of $n$-order hinges is called a $n$-order hinging hyperplane ($n$-HH) model. Such model has universal representation power over all CPWL functions, as formalized in the theorem below: ###### Theorem 1.3 (Theorem 1 in Wang and Sun (2005)). For any positive integer $n$ and CPWL function $f(x):R^{n}\to R$, there exists a $n$-HH which exactly represents $f(x)$. The question is whether we can give a sharp lower bound on the necessary number of affine functions within the nested maximum operator. Wang and Sun (2005) conjected that $(n-1)$-HH can’t represent all CPWL functions, but this open problem is left unanswered for more than a decade. In the following section we will prove our main result that $(n-2)$-HH can’t represent all CPWL functions, yielding an almost tight lower bound. ## 2 Main Result Observe that any $(n-2)$-order hinge depends on only $n-1$ affine transforms of $x$, thus there always exists a direction in which the value of the $(n-2)$-order hinge remains the same. We make such observation precise by introducing the definition of low-dimensional and periodic functions. ###### Definition 2.1 (Low-dimensional/periodic function). A function $f(x):R^{n}\to R$ is said to be low-dimensional, if there exists a vector $v\neq 0$, such that for any $x\in R^{n}$ and $c\in R$, we have that $f(x)=f(x+cv)$. If we have only $f(x)=f(x+v)$ then $f$ is said to be periodic (a weaker notion). $v$ is called an invariant vector of $f$. Any $(n-2)$-order hinge is a low-dimensional function on $R^{n}$, so our problem is reduced to proving the class of finite sum of low-dimensional functions has limited representation power. The following key lemma actually proves (a stronger result) that finite sum of periodic functions can’t represent any non-trivial integrable functions. ###### Lemma 2.2. Any finite sum of periodic functions is either non-integrable or the zero function, i.e. given periodic functions $f_{i}(x)$, $i=1,...,m$, then $f(x)\triangleq\sum_{i=1}^{m}f_{i}(x)$ satisfies $\int_{R^{n}}|f|=\infty\quad or\quad f\equiv 0$ (3) ###### Proof. We prove Lemma 2.2 by induction. Suppose each $f_{i}$ has an invariant vector $v_{i}$, base case $m=1$ is trivial since if we denote the orthogonal hyperplane $H_{i}=\\{x|x^{\top}v_{i}=0\\}$, we have $\int_{R^{n}}|f|=\int_{R}\int_{H_{1}}|f|$ (4) thus $\int_{R^{n}}|f|<\infty$ if and only if $\int_{H}|f|=0$. Assume $f=\sum_{i=1}^{m}f_{i}$ is integrable, then $g(x)\triangleq f(x+v_{m})-f(x)$ is also integrable. We make the following decomposition of $g$: $g(x)=\sum_{i=1}^{m}f_{i}(x+v_{m})-f_{i}(x)=\sum_{i=1}^{m-1}f_{i}(x+v_{m})-f_{i}(x)$ (5) where each $f_{i}(x+v_{m})-f_{i}(x)$ is periodic (with invariant vector $v_{i}$) as well. By induction we have $g\equiv 0$ and $f$ is also a periodic function (with invariant vector $v_{m}$). Using the base case on $f$ again concludes our proof. ∎ Our main result is a direct corollary of Lemma 2.2, as stated below: ###### Theorem 2.3. For any positive integer $n\geq 2$, there exists a CPWL function $g(x):R^{n}\to R$, such that no $(n-2)$-HH can exactly represent $g(x)$. ###### Proof. Let $g(x)\triangleq\max\\{0,1-||x||_{\infty}\\}$. It’s straightfoward to check that $g(x)$ is a CPWL function with at most $2^{n+1}$ affine polyhedral regions, and meanwhile is an integrable function with positive integral. As any $(n-2)$-HH can be written as a finite sum of low-dimensional functions, it can’t represent $g(x)$ by Lemma 2.2. ∎ Theorem 2.3 implies that in order to achieve universal representation power over all CPWL functions, a $(n-1)$-HH model is necessary which provides an almost tight lower bound corresponding to the upper bound in Theorem 1.3. ## 3 Implications on Universal Approximation of ANNs Traditional universal approximation theorems of artifical neural networks (ANN) Cybenko (1989); Hornik et al. (1989); Barron (1994) typically states that an ANN with one hidden layer and unbounded width can approximate any measurable function with arbitrary precision on a compact set. Our result demonstrates that the compact set assumption is indeed necessary for ANNs with traditional activation (composition of an affine transform and a fixed univariate function $\sigma$): ###### Corollary 3.1. Given an integrable function $f$ on $R^{n}$ ($n\geq 2$), for any one-hidden- layer neural network $g$ with traditional activation $\sigma(w^{\top}x+b)$, we have that $\int_{R^{n}}|f-g|=\infty\quad or\quad\int_{R^{n}}|f-g|=\int_{R^{n}}|f|$ (6) ###### Proof. Any unit $\sigma(w^{\top}x+b)$ is obviously a low-dimensional function when $n\geq 2$, thus by Lemma 2.2 we finish our proof. ∎ Corollary 3.1 reveals a fundamental gap of representation power between one- hidden layer neural networks and deeper ones, as Theorem 1.3 indicates a neural network with $\lceil log_{2}(n+1)\rceil$ hidden layers can represent any CPWL function He et al. (2018), showing the benefits of depth in universal approximation Lu et al. (2017). ## 4 Conclusion In this note we give a sharp lower bound on the necessary number of nestings of nested absolute-value functions of generalized hinging hyperplanes (GHH) to represent arbitrary CPWL functions, which is the first non-trivial lower bound to the best of our knowledge. Our results fully characterizes the representation power (and limit) of the GHH model. Our result also has implications on ANNs, a much more popular model in machine learning. It shows that one-hidden-layer neural networks with traditional activation can’t control the approximation error on the whole domain despite existing universal approximation theorems, a fundamental gap between one- hidden-layer networks and deeper ones. We conject similar depth-separation results should hold for deeper networks and the $\lceil log_{2}(n+1)\rceil$ bound should be tight in representing CPWL functions. Instead of low- dimensional (periodic), other properties need to be discovered for deeper networks. ## Acknowledgements The author would like to thank Fedor Petrov for giving an elegant proof of Lemma 2.2 on Mathoverflow. ## References * Barron (1994) Andrew R Barron. Approximation and estimation bounds for artificial neural networks. _Machine learning_ , 14(1):115–133, 1994. * Breiman (1993) Leo Breiman. Hinging hyperplanes for regression, classification, and function approximation. _IEEE Transactions on Information Theory_ , 39(3):999–1013, 1993. * Chua and Deng (1988) Leon O Chua and A-C Deng. Canonical piecewise-linear representation. _IEEE Transactions on Circuits and Systems_ , 35(1):101–111, 1988. * Cybenko (1989) George Cybenko. Approximation by superpositions of a sigmoidal function. _Mathematics of control, signals and systems_ , 2(4):303–314, 1989. * He et al. (2018) Juncai He, Lin Li, Jinchao Xu, and Chunyue Zheng. Relu deep neural networks and linear finite elements. _arXiv preprint arXiv:1807.03973_ , 2018. * Hornik et al. (1989) Kurt Hornik, Maxwell Stinchcombe, and Halbert White. Multilayer feedforward networks are universal approximators. _Neural networks_ , 2(5):359–366, 1989. * Lu et al. (2017) Zhou Lu, Hongming Pu, Feicheng Wang, Zhiqiang Hu, and Liwei Wang. The expressive power of neural networks: A view from the width. _arXiv preprint arXiv:1709.02540_ , 2017. * Wang and Sun (2005) Shuning Wang and Xusheng Sun. Generalization of hinging hyperplanes. _IEEE Transactions on Information Theory_ , 51(12):4425–4431, 2005.
# Which Nilpotent Groups are Self-Similar? 111Research supported by CNRS-UMR 5028 and Labex MILYON/ANR-10-LABX-0070. Olivier Mathieu 222Institut Camille Jordan, Université de Lyon. Email: <EMAIL_ADDRESS> ###### Abstract Let $\Gamma$ be a finitely generated torsion free nilpotent group, and let $A^{\omega}$ be the space of infinite words over a finite alphabet $A$. We investigate two types of self-similar actions of $\Gamma$ on $A^{\omega}$, namely the faithfull actions with dense orbits and the free actions. A criterion for the existence of a self-similar action of each type is established. Two corollaries about the nilmanifolds are deduced. The first involves the nilmanifolds endowed with an Anosov diffeomorphism, and the second about the existence of an affine structure. Then we investigate the virtual actions of $\Gamma$, i.e. actions of a subgroup $\Gamma^{\prime}$ of finite index. A formula, with some number theoretical content, is found for the minimal cardinal of an alphabet $A$ endowed with a virtual self-similar action on $A^{\omega}$ of each type. Mathematics Subject Classification 37B10-20G30-53C30 ## Introduction 1\. General introduction Let $A$ be a finite alphabet and let $A^{\omega}$ be the topological space of infinite words $a_{1}a_{2}\dots$ over $A$, where the topology of $A^{\omega}=\varprojlim A^{n}$ is the pro-finite topology. An action of a group $\Gamma$ on $A^{\omega}$ is called self-similar iff for any $\gamma\in\Gamma$ and $a\in A$ there exists $\gamma_{a}\in\Gamma$ and $b\in A$ such that $\gamma(aw)=b\gamma_{a}(w)$ for any $w\in A^{\omega}$. The group $\Gamma$ is called self-similar (respectively densely self-similar, respectively freely self-similar, respectively freely densely self-similar) if $\Gamma$ admits a faithfull self-similar action (respectively a faithfull self-similar action with dense orbits, respectively a free self-similar action, respectively a free self-similar action with dense orbits) on $A^{\omega}$ for some finite alphabet $A$. Self-similar groups appeared in the early eighties, in the works of R. Grigorchuk [10] [11] and in the joint works of N. Gupta and S. Sidki [13] [14]. See also the monography [24] for an extensive account before 2005 and [25] [2] [9] [16] [12] for more recent works. A general question is which groups $\Gamma$ are (merely, or densely …) self-similar? This paper brings an answer for finitely generated torsion-free nilpotent groups $\Gamma$, called FGTF nilpotent groups in the sequel. Then we will connect the main result with topics involving differential geometry and arithmetic groups. The systematic study of self-similar actions of nilpotent groups started with [4], and the previous question has been raised in some talks of S. Sidki. 2\. The main results A few definitions are now required. A grading of a Lie algebra $\mathfrak{m}$ is a decomposition $\mathfrak{m}=\oplus_{n\in\mathbb{Z}}\,\mathfrak{m}_{n}$ such that $[\mathfrak{m}_{n},\mathfrak{m}_{m}]\subset\mathfrak{m}_{n+m}$ for all $n,\,m\in\mathbb{Z}$. It is called special if $\mathfrak{m}_{0}\cap\mathfrak{z}=0$, where $\mathfrak{z}$ is the center of $\mathfrak{m}$. It is called very special if $\mathfrak{m}_{0}=0$. Let’s assume now that $\Gamma$ is a FGTF nilpotent group. By Malcev Theory [18][26], $\Gamma$ is a cocompact lattice in a unique connected, simply connected (or CSC in what follows) nilpotent Lie group $N$. Let $\mathfrak{n}^{\mathbb{R}}$ be the Lie algebra of $N$ and set $\mathfrak{n}^{\mathbb{C}}=\mathbb{C}\otimes_{\mathbb{R}}\mathfrak{n}^{\mathbb{R}}$. The main results, proved in Section 7, are the following ###### Theorem 2. The group $\Gamma$ is densely self-similar iff the Lie algebra $\mathfrak{n}^{\mathbb{C}}$ admits a special grading. ###### Theorem 3. The following assertions are equivalent (i) The group $\Gamma$ is freely self-similar, (ii) the group $\Gamma$ is freely densely self-similar, and (iii) the Lie algebra $\mathfrak{n}^{\mathbb{C}}$ admits a very special grading. As a consequence, let’s mention ###### Corollary 4. Let $M$ be a nilmanifold endowed with an Anosov diffeomorphism. Then there a free self-similar action of $\pi_{1}(M)$ with dense orbits on $A^{\omega}$, for some finite $A$. ###### Corollary 8. Let $M$ be a nilmanifold. If $\pi_{1}(M)$ is freely self-similar, then $M$ is affine complete. Among FGTF nilpotent groups, some of them are self-similar but not densely self-similar. Some of them are not even self-similar, since Theorem 2 implies the next ###### Corollary 7. Let $M$ be one of the non-affine nilmanifolds appearing in [3]. Then $\pi_{1}(M)$ is not self-similar. 3\. A concrete version of Theorems 2 and 3 Let $N$ be a CSC nilpotent Lie group, with Lie algebra $\mathfrak{n}^{\mathbb{R}}$. Let’s assume that $N$ contains some cocompact lattices $\Gamma$. By definition, the degree of a self-similar action of $\Gamma$ on $A^{\omega}$ is $\mathrm{Card\,}\,A$. We ask the following question For a given cocompact lattice $\Gamma\subset N$, what is the minimal degree degree of a faithfull (or a free) self-similar action with dense orbits? More notions are now defined. Recall that the commensurable class $\xi$ of a cocompact lattice $\Gamma_{0}\subset N$ is the set of all cocompact lattices of $N$ which share with $\Gamma_{0}$ a subgroup of finite index. The complexity $\mathrm{cp}\,\xi$ (respectively the free complexity $\mathrm{fcp}\,\xi$) of the class $\xi$ is the minimal degree of a self- similar action of $\Gamma$ with dense orbits (respectively a free self-similar action of $\Gamma$), for some $\Gamma\in\xi$. For any algebraic number $\lambda\neq 0$, set $d(\lambda)=\mathrm{Card\,}\,{\cal O}(\lambda)/\pi_{\lambda}$, where ${\cal O}(\lambda)$ is the ring of integers of $\mathbb{Q}(\lambda)$ and $\pi_{\lambda}=\\{x\in{\cal O}(\lambda)|x\lambda\in{\cal O}(\lambda)\\}$. For any isomorphism $h$ of a finite dimensional vector space over $\mathbb{Q}$, set ${\mathrm{ht}}\,h=\prod_{\lambda\in{\mathrm{Spec}}\,h/{\mathrm{Gal}}(\mathbb{Q})}\,d(\lambda)^{m_{\lambda}}$, where ${\mathrm{Spec}}\,h/{\mathrm{Gal}}(\mathbb{Q})$ is the list of eigenvalues of $h$ up to conjugacy by ${\mathrm{Gal}}(\mathbb{Q})$ and where $m_{\lambda}$ is the multiplicity of the eigenvalue $\lambda$. By Malcev’s Theory, the commensurable class $\xi$ determines a canonical $\mathbb{Q}$-form $\mathfrak{n}(\xi)$ of the Lie algebra $\mathfrak{n}^{\mathbb{R}}$. Let ${\cal S}(\mathfrak{n}(\xi))$ (respectively ${\cal V}(\mathfrak{n}(\xi))$) be the set of all $f\in\mathrm{Aut}\,\mathfrak{n}(\xi)$ such that ${\mathrm{Spec}}\,\,f|\mathfrak{z}^{\mathbb{C}}$ (respectively ${\mathrm{Spec}}\,\,f$) contains no algebraic integer. ###### Theorem 9. We have $\mathrm{cp}\,\xi=\mathrm{Min}_{h\in{\cal S}(\mathfrak{n}(\xi))}\,{\mathrm{ht}}\,h$, and $\mathrm{fcp}\,\xi=\mathrm{Min}_{h\in{\cal V}(\mathfrak{n}(\xi))}\,{\mathrm{ht}}\,h$. If, in the previous statement, ${\cal S}(\mathfrak{n}(\xi))$ is empty, then the equality $\mathrm{cp}\,\xi=\infty$ means that no $\Gamma\in\xi$ admits a faithfull self-similar action with dense $\Gamma$-orbits. Theorem 9 answers the previous question only for the commensurable classes $\xi$. For an individual $\Gamma\in\xi$, it provides some ugly estimates for the minimal degree of $\Gamma$-actions, and nothing better can be expected. The framework of nonabelian Galois cohomology shows the concreteness of Theorem 9. Up to conjugacy, the commensurable classes in $N$ are classified by the $\mathbb{Q}$-forms of some classical objects with a prescribed $\mathbb{R}$-form, see Corollary 4 of ch. 9, and their complexity is an invariant of the arithmetic group $\mathrm{Aut}\,\mathfrak{n}(\xi)$. As an illustration of the previous vague sentence, we investigate a class ${\cal N}$ of CSC nilpotent Lie groups $N$, with Lie algebra $\mathfrak{n}^{\mathbb{R}}$. The commensurable classes $\xi(q)$ in $N$ are classified, up to conjugacy, by the positive definite quadratic forms $q$ on $\mathbb{Q}^{2}$. Then, we have $\mathrm{cp}\,\xi(q)=F(d)^{e(N)}$ where $e(N)$ is an invariant of the special grading of $\mathbb{C}\otimes\mathfrak{n}^{\mathbb{R}}$, where $-d$ is the discriminant of $q$, and where $F(d)$ is the norm of a specific ideal in $\mathbb{Q}(\sqrt{-d})$, see Theorem 11 and Lemma 28. In particular, $N$ contains some commensurable classes of arbitrarily high complexity. In a forthcoming paper [22], more complicated examples are investigated, but the formulas are less explicit. 4\. About the proofs. The proofs of the paper are based on different ideas. Theorem 1, which is a statement about rational points of algebraic tori, is the key step in the proof of Theorems 2, 3 and 11. It is based on standard results of number theory, including the Cebotarev’s Theorem. It is connected with the density of rational points for connected groups proved by A. Borel [6], see also [27]. Also, the proof of Corollary 4 is based on a paper of A. Manning [20] about Anosov diffeomorphisms. The proof of Corollary 7 is based on very difficult computations, which, fortunately, were entirely done in [3]. ## 1 Self-similar actions and self-similar data Let $\Gamma$ be a group. This section explains the correspondence between the faithfull transitive self-similar $\Gamma$-actions and some virtual endomorphisms of $\Gamma$, called self-similar data. Usually self-similar actions are actions on a rooted tree $A^{*}$, see [24]. Here the groups are acting on the boundary $A^{\omega}$ of $A^{*}$. This equivalent viewpoint is better adapted to our setting. 1.1 Transitive self-similar actions In addition of the definitions of the introduction, the following technical notion of transitivity will be used. A self-similar action of $\Gamma$ on $A^{\omega}$ induces an action of $\Gamma$ on $A$. Indeed, for $a,\,b\in A$ and $\gamma\in\Gamma$, we have $\gamma(a)=b$ if $\gamma(aw)=b\gamma_{a}(w)$, for all $w\in A^{\omega}$. A self-similar action is called transitive if the induced action on $A$ is transitive. The group $\Gamma$ is called transitive self-similar if it admits a faithfull transitive self-similar action. Similarly the self-similar action of $\Gamma$ on $A^{\omega}$ induces an action of $\Gamma$ on each level set $A^{n}$. Such an action is often called level transitive if $\Gamma$ acts transitively on each level $A^{n}$. Obviously, the level transitive actions on $A^{*}$ of [24] corresponds with the actions on $A^{\omega}$ with dense orbits. 1.2 Core and $f$-core Let $\Gamma$ be a group and $\Gamma^{\prime}$ be a subgroup. The core of $\Gamma^{\prime}$ is the biggest normal subgroup $K\triangleleft G$ with $K\subset\Gamma^{\prime}$. Equivalently the core is the kernel of the left action of $\Gamma$ on $\Gamma/\Gamma^{\prime}$. Now let $f:\Gamma^{\prime}\to\Gamma$ be a group morphism. By defintion the $f$-core is the biggest normal subgroup $K\triangleleft G$ with $K\subset\Gamma^{\prime}$ and $f(K)\subset K$. 1.3 Self-similar data Let $\Gamma$ be a group. A virtual endomorphism of $\Gamma$ is a pair $(\Gamma^{\prime},f)$, where $\Gamma^{\prime}$ is a subgroup of finite index and $f:\Gamma^{\prime}\to\Gamma$ is a group morphism. A self-similar datum is a virtual endomorphism $(\Gamma^{\prime},f)$ with a trivial $f$-core. Assume given a faithfull transitive self-similar action of $\Gamma$ on $A^{\omega}$. Let $a\in A$, and let $\Gamma^{\prime}$ be the stabilizer of $a$. By definition, for each $\gamma\in\Gamma^{\prime}$ there is a unique $\gamma_{a}\in\Gamma$ such that $\gamma(aw)=a\gamma_{a}(w)$, for any $w\in A^{\omega}$. Let $f:\Gamma^{\prime}\rightarrow\Gamma$ be the map $\gamma\mapsto\gamma_{a}$. Since the action is faithfull, $\gamma_{a}$ is uniquely determined and $f$ is a group morphism. Also it follows from Proposition 2.7.4 and 2.7.5 of [24] that the $f$-core of $\Gamma^{\prime}$ is the kernel of the action, therefore it is trivial. Hence $(\Gamma^{\prime},f)$ is a self-similar datum. Conversely, a virtual endomorphism $(\Gamma^{\prime},f)$ determines a transitive self-similar action of $\Gamma$ on $A^{\omega}$, where $A\simeq\Gamma/\Gamma^{\prime}$. Moreover the $f$-core is the kernel of the corresponding action, see ch 2 of [24] for details, especially subsection 2.5.5 of [24]). In conclusion, we have ###### Lemma 1. Let $\Gamma$ be a group. There is a correspondence between the self-similar data $(\Gamma^{\prime},f)$ and the faithfull transitive self-similar actions of $\Gamma$ on $A^{\omega}$, where $A\simeq\Gamma/\Gamma^{\prime}$. This correspondence is indeed a bijection up to conjugacy, see [24] for a precise statement. 1.4 Good self-similar data Let $\Gamma$ be a group, and let $(\Gamma^{\prime},f)$ be a virtual endomorphism. Let $\Gamma_{n}$ be the subgroups of $\Gamma$ inductively defined by $\Gamma_{0}=\Gamma$, $\Gamma_{1}=\Gamma^{\prime}$ and for $n\geq 2$ $\Gamma_{n}=\\{\gamma\in\Gamma_{n-1}|\,f(\gamma)\in\Gamma_{n-1}\\}$ ###### Lemma 2. The sequence $n\mapsto[\Gamma_{n}:\Gamma_{n+1}]$ is not increasing. ###### Proof. For $n>0$, the map $f$ induces an injection of the set $\Gamma_{n}/\Gamma_{n+1}$ into $\Gamma_{n-1}/\Gamma_{n}$, thus we have $[\Gamma_{n}:\Gamma_{n+1}]\leq[\Gamma_{n-1}:\Gamma_{n}]$. ∎ The virtual endomorphism $(\Gamma^{\prime},f)$ is called good if $[\Gamma_{n}:\Gamma_{n+1}]=[\Gamma/\Gamma^{\prime}]$ for all $n$. Let $(\Gamma^{\prime},f)$ be a self-similar datum, and let $A^{*}$ be the corresponding tree on which $\Gamma$ acts. If $a$ is the distinguished point in $A\simeq\Gamma/\Gamma^{\prime}$, then $\Gamma_{n}$ is the stabilizer of $a^{n}$. If the self-similar datum $(\Gamma^{\prime},f)$ is good, then $[\Gamma:\Gamma_{n}]=\mathrm{Card\,}A^{n}$ and therefore $\Gamma$ acts transitively on $A^{n}$. Exactly as before, we have ###### Lemma 3. Let $\Gamma$ be a group. There is a correspondence between the good self- similar data $(\Gamma^{\prime},f)$ and the faithfull self-similar actions of $\Gamma$ on $A^{\omega}$ with dense orbits, where $A\simeq\Gamma/\Gamma^{\prime}$. 1.5 Fractal self-similar data Let $\Gamma$ be a group. A self-similar datum $(\Gamma^{\prime},f)$ is called fractal (or recurrent) if $f(\Gamma^{\prime})=\Gamma$. A self-similar action of $\Gamma$ on some $A^{\omega}$ is called fractal if it is transitive and the corresponding self-similar datum is fractal, see [24] section 2.8. Obviously a fractal action has dense orbits. The group $\Gamma$ is called fractal (respectively freely fractal) if $\Gamma$ admits a faithfull (respectively free) fractal action on some $A^{\omega}$. ## 2 Rational points of a torus We are going to prove Theorem 1, about the rational points of algebraic tori. For the whole chapter, let $\bf H$ be an algebraic torus defined over $\mathbb{Q}$ and let $X({\bf H})$ be the group of characters of $\bf H$. For a number field $K$, let’s denote by ${\mathrm{Gal}}(K):={\mathrm{Gal}}(\overline{\mathbb{Q}}/K)$ its absolute Galois group. The group $X({\bf H})$ is a ${\mathrm{Gal}}(\mathbb{Q})$-module which is isomorphic to $\mathbb{Z}^{r}$ as an abelian group, where $r=\dim{\bf H}$. The splitting field of $\bf H$ is the smallest Galois extension $L$ of $\mathbb{Q}$ such that ${\mathrm{Gal}}(L)$ acts trivially on $X({\bf H})$, or, equivalently such that $\bf H$ is $L$-isomorphic to $\mathbb{G}_{m}^{r}$, where $\mathbb{G}_{m}$ denotes the multiplicative group. Moreover, we have $\chi(h)\in L^{*}$ for any $\chi\in X({\bf H})$ and any $h\in{\bf H}(\mathbb{Q})$. Let ${\cal O}$ be the ring of integers of $L$. Recall that a fractional ideal is a nonzero finitely generated ${\cal O}$-submodule of $K$. A fractional ideal $I$ is called integral if $I\subset{\cal O}$. If the fractional ideal $I$ is integral and $I\neq{\cal O}$, then $I$ is merely an ideal of ${\cal O}$. Let ${\cal I}$ be the set of all fractional ideals and ${\cal I}^{+}$ be the subset of all integral ideals. Given $I$ and $J$ in ${\cal I}$, their product is the ${\cal O}$-module generated by all products $ab$ where $a\in I$ and $b\in J$. Since ${\cal O}$ is a Dedekind ring, we have ${\cal I}\simeq\oplus_{\pi\in{\cal P}}\,\mathbb{Z}\,[\pi]$ ${\cal I}^{+}\simeq\oplus_{\pi\in{\cal P}}\,\mathbb{Z}_{\geq 0}\,[\pi]$, where ${\cal P}$ is the set of prime ideals of ${\cal O}$. Indeed the additive notation is used for for the group ${\cal I}$ and the monoid ${\cal I}^{+}$: view as an element of ${\cal I}$ the fractional ideal $\pi_{1}^{m_{1}}\dots\pi_{n}^{m_{n}}$ will be denoted as $m_{1}[\pi_{1}]+\dots+m_{n}[\pi_{n}]$. Since ${\mathrm{Gal}}(L/\mathbb{Q})$ acts by permutation on ${\cal P}$, ${\cal I}$ is a $\mathbb{Z}{\mathrm{Gal}}(L/\mathbb{Q})$-module. For $S\subset{\cal P}$, set ${\cal I}_{S}=\oplus_{\pi\in{\cal P}\setminus S}\,\mathbb{Z}\,[\pi]$. ###### Lemma 4. Let $S\subset{\cal P}$ be a finite subset and let $r>0$ be an integer. The ${\mathrm{Gal}}(L/\mathbb{Q})$-module ${\cal I}$ contains a free $\mathbb{Z}{\mathrm{Gal}}(L/\mathbb{Q})$-module $M(r)$ of rank $r$ such that (i) $M(r)\cap{\cal I}^{+}=\\{0\\}$, and (ii) $M(r)\subset{\cal I}_{S}$. ###### Proof. Let $S^{\prime}$ be the set of all prime numbers which are divisible by some $\pi\in S$. Let $\Sigma$ be the set of prime numbers $p$ that are completely split in $K$, i.e. such that ${\cal O}/p{\cal O}\simeq\mathbb{F}_{p}^{[L:\mathbb{Q}]}$. For $p\in\Sigma$, let $\pi\in{\cal P}$ be a prime ideal over $p$. When $\sigma$ runs over ${\mathrm{Gal}}(L/\mathbb{Q})$ the ideals $\pi^{\sigma}$ are all distinct, and therefore $[\pi]$ generates a free $\mathbb{Z}{\mathrm{Gal}}(L/\mathbb{Q})$-submodule of rank one in ${\cal I}$. By Cebotarev theorem, the set $\Sigma$ is infinite. Choose $r+1$ distinct prime numbers $p_{0},\dots p_{r}$ in $\Sigma\setminus S^{\prime}$, and let $\pi_{0},\dots,\pi_{r}\in{\cal P}$ such that ${\cal O}/\pi_{i}=\mathbb{F}_{p_{i}}$. For $1\leq i\leq r$, set $\tau_{i}=[\pi_{i}]-[\pi_{0}]$ and let $M(r)$ be the $\mathbb{Z}{\mathrm{Gal}}(L/\mathbb{Q})$-module generated by $\tau_{1},\dots,\tau_{r}$. Obviously, the ${\mathrm{Gal}}(L/\mathbb{Q})$-module $M(r)$ is free of rank $r$ and $M(r)\subset{\cal I}_{S}$. It remains to prove that $M(r)\cap{\cal I}^{+}=\\{0\\}$. Let $A=\sum\limits_{1\leq i\leq r,\sigma\in{\mathrm{Gal}}(L/\mathbb{Q})}\,m_{i,\sigma}\,\tau_{i}^{\sigma}$ be an element of $M(r)\cap{\cal I}^{+}$. We have $A=B-C$, where $B=\sum\limits_{1\leq i\leq r,\sigma\in{\mathrm{Gal}}(L/\mathbb{Q})}\,m_{i,\sigma}\,[\pi_{i}^{\sigma}]$, and $C=\sum\limits_{\sigma\in{\mathrm{Gal}}(L/\mathbb{Q})}\,(\sum\limits_{1\leq i\leq r}\,m_{i}^{\sigma})[\pi_{0}^{\sigma}]$. Thus the condition $A\in{\cal I}^{+}$ implies that $m_{i}^{\sigma}\geq 0$, for any $1\leq i\leq k$ and $\sigma\in{\mathrm{Gal}}(L/\mathbb{Q})$, and $\sum\limits_{1\leq i\leq r}\,m_{i}^{\sigma}\leq 0$, for any $\sigma\in{\mathrm{Gal}}(L/\mathbb{Q})$. Thus all the integers $m_{i}^{\sigma}$ vanish. Therefore $M(r)$ intersects ${\cal I}^{+}$ trivially. ∎ For $\pi\in{\cal P}$, let $v_{\pi}:L^{*}\to\mathbb{Z}$ be the corresponding valuation. ###### Lemma 5. Let $S\subset{\cal P}$ be a finite ${\mathrm{Gal}}(L/\mathbb{Q})$-invariant subset and let $r>0$ be an integer. The ${\mathrm{Gal}}(L/\mathbb{Q})$-module $L^{*}$ contains a free $\mathbb{Z}{\mathrm{Gal}}(L/\mathbb{Q})$-module $N(r)$ of rank $r$ such that (i) $N(r)\cap{\cal O}=\\{1\\}$, and (ii) $v_{\pi}(x)=0$ for any $x\in N(r)$ and any $\pi\in S$. ###### Proof. Set $L^{*}_{S}=\\{x\in L^{*}|v_{\pi}(x)=0,\,\forall\pi\in S\\}$ and let $\theta:L^{*}_{S}\rightarrow{\cal I}_{S}$ be the map $x\mapsto\sum_{\pi\in{\cal P}}\,v_{\pi}(x)\,[\pi]$. By Lemma 4, ${\cal I}_{S}$ contains a free $\mathbb{Z}{\mathrm{Gal}}(L/\mathbb{Q})$-module $M(r)$ of rank $r$ such that $M(r)\cap{\cal I}^{+}=\\{0\\}$. Let’s remark that $\mathrm{Coker}\,\theta$ is a subgroup of the class group $\mathrm{Cl}(L)$ of $L$. Since, by Dirichelet Theorem, $\mathrm{Cl}(L)$ is finite, there is a positive integer $d$ such that $d.M(r)$ lies in the image of $\theta$. Since $M(r)$ is free, there is a free $\mathbb{Z}{\mathrm{Gal}}(K/\mathbb{Q})$-module $N(r)\subset L^{*}_{S}$ of rank $r$ which is a lift of $dM(r)$, i.e. such that $\theta$ induces an isomorphism $N(r)\simeq d.M(r)$. Since $\theta({\cal O}\setminus 0)$ lies in ${\cal I}^{+}$, we have $\theta(N(r)\cap{\cal O})=\\{0\\}$. It follows that $N(r)\cap{\cal O}=\\{1\\}$. The second assertion follows from the fact that $N(r)$ lies in $L^{*}_{S}$. ∎ For $\pi\in{\cal P}$, let ${\cal O}_{\pi}$ and $L_{\pi}$ be the $\pi$-adic completions of ${\cal O}$ and $L$. Let $x,\,y\in L$ and let $n>0$ be an integer. In what follows, the congruence $x\equiv y$ modulo $n{\cal O}_{\pi}$ means $x_{\pi}\equiv y_{\pi}\,\mathrm{mod}\,n{\cal O}_{\pi}$, where $x_{\pi}$ and $y_{\pi}$ are the images of $x$ and $y$ in $L_{\pi}$. The case $n=1$ of the next statement will be used in further sections. In such a case, Assertion (ii) is tautological. ###### Theorem 1. Let ${\bf H}$ be an algebraic torus defined over $\mathbb{Q}$, and let $L$ be its splitting field. Let $n>0$ be an integer and let $S\subset{\cal P}$ be the set of prime divisors of $n$. There exists $h\in{\bf H}(\mathbb{Q})$ such that (i) $\chi(h)$ is not an algebraic integer, for any non-trivial $\chi\in X({\bf H})$, and (ii) $\chi(h)\equiv 1\,\mathrm{mod}\,n{\cal O}_{\pi}$ for any $\chi\in X({\bf H})$ and any $\pi\in S$. ###### Proof. Step 1. First an element $h^{\prime}\in{\bf H}(\mathbb{Q})$ satisfying Assertion (i) and (iii) $v_{\pi}(\chi(h^{\prime}))=0$, for any $\pi\in S$ and any $\chi\in X({\bf H})$ is found. The abelian group $X({\bf H})$ is free of rank $r$ where $r=\dim\,{\bf H}$. Therefore, the comultiplication $\Delta:X({\bf H})\rightarrow X({\bf H})\otimes\mathbb{Z}{\mathrm{Gal}}(L/\mathbb{Q})$ provides an embedding of $X({\bf H})$ into a free $\mathbb{Z}{\mathrm{Gal}}(L/\mathbb{Q})$-module of rank $r$. By lemma 5, there a free $\mathbb{Z}{\mathrm{Gal}}(L/\mathbb{Q})$-module $N(r)\subset L^{*}_{S}$ of rank $r$ with $N(r)\cap{\cal O}=\\{1\\}$. Let $\mu:X({\bf H})\otimes\mathbb{Z}{\mathrm{Gal}}(L/\mathbb{Q})\to N(r)$, be an isomorphism of $\mathbb{Z}{\mathrm{Gal}}(L/\mathbb{Q})$-modules, and set $h^{\prime}=\mu\circ\Delta$. Since ${\bf H}(\mathbb{Q})={\mathrm{Hom}}_{{\mathrm{Gal}}(L/\mathbb{Q})}(X({\bf H}),L^{*})$, the embedding $h^{\prime}$ is indeed an element of ${\bf H}(\mathbb{Q})$. Viewed as a map from $X({\bf H})$ to $L^{*}$, $h^{\prime}$ is the morphism $\chi\in X({\bf H})\mapsto\chi(h^{\prime})$. Since $\mathrm{Im}\,h^{\prime}\cap{\cal O}=1$ and $h^{\prime}$ is injective, $\chi(h^{\prime})$ is not an algebraic integer if $\chi$ is a non-trivial character. Since $\mathrm{Im}\,h^{\prime}\subset L^{*}_{S}$, we have $v_{\pi}(\chi(h^{\prime}))=0$ for any $\chi\in X({\bf H})$. Therefore $h^{\prime}$ satisfies Assertions (i) and (iii). Step 2. Let $\chi_{1},\dots,\chi_{r}$ be a basis of $X({\bf H})$. Since $v_{\pi}(\chi_{i}(h^{\prime}))=0$ for any $\pi\in S$, the element $\chi_{i}(h^{\prime})\,\mathrm{mod}\,n{\cal O}_{\pi}$ is an inversible element in the finite ring ${\cal O}_{\pi}/n{\cal O}_{\pi}$. Therefore there are positive integers $m_{i,\pi}$ such that $\chi_{i}(h^{\prime})^{m_{i,\pi}}\equiv 1\,\mathrm{mod}\,n{\cal O}_{\pi}$, for all $1\leq i\leq r$ and all $\pi\in S$. Set $m=\mathrm{lcm}(m_{i,\pi})$ and set $h=h^{\prime m}$. Obviously $h$ satisfies Assertion (i). Moreover we have $\chi_{i}(h)\equiv 1\,\mathrm{mod}\,n{\cal O}_{\pi}$, for all $\pi\in S$ and all $1\leq i\leq r$, and therefore $h$ satisfies Assertion (ii) as well. ∎ ## 3 Special Gradings Let $\mathfrak{n}$ be a finite dimensional Lie algebra defined over $\mathbb{Q}$ and let $\mathfrak{z}$ be its center. The relations between the gradings of $\mathbb{C}\otimes\mathfrak{n}$ and the automorphisms of $\mathfrak{n}$ are investigated now. The following important definitions will be used in the whole paper. Let ${\cal S}(\mathfrak{n})$ (respectively ${\cal V}(\mathfrak{n})$) be the set of all $f\in\mathrm{Aut}\,\mathfrak{n}$ such that ${\mathrm{Spec}}\,\,f|_{\mathfrak{z}}$ (respectively ${\mathrm{Spec}}\,\,f$) contains no algebraic integers. Moreover let ${\cal F}(\mathfrak{n})$ be the set of all $f\in{\cal S}(\mathfrak{n})$ such that all eigenvalues of $f^{-1}$ are algebraic integers. Also set ${\cal F}^{+}(\mathfrak{n})={\cal F}(\mathfrak{n})\cap{\cal V}(n)$. Here, by eigenvalues of a $\mathbb{Q}$-linear endomorphism $F$, we always mean the eigenvalues of $F$ in $\overline{\mathbb{Q}}$. For any field $K$ of characteristic zero, set $\mathfrak{n}^{K}=K\otimes\mathfrak{n}$ and $\mathfrak{z}^{K}=K\otimes\mathfrak{z}$. Let ${\bf G}={\bf Aut}\,\mathfrak{n}$ be the algebraic group of automorphisms of $\mathfrak{n}$. By definition, ${\bf G}$ is defined over $\mathbb{Q}$, and we have ${\bf G}(K)=\mathrm{Aut}\,\mathfrak{n}^{K}$ for any field $K$ of characteristic zero. The notation $\mathfrak{n}$ underlines that $\mathfrak{n}$ can be viewed as the functor in Lie algebras $K\mapsto\mathfrak{n}^{K}$. Let ${\bf H}\subset{\bf G}$ be a maximal torus defined over $\mathbb{Q}$, whose existence is proved in [7], see also [6], Theorem 18.2. By definition, a $K$-grading of $\mathfrak{n}$ is is a decomposition of $\mathfrak{n}^{K}$ $\mathfrak{n}^{K}=\oplus_{n\in\mathbb{Z}}\,\mathfrak{n}^{K}_{n}$ such that $[\mathfrak{n}^{K}_{n},\mathfrak{n}^{K}_{m}]\subset\mathfrak{n}^{K}_{n+m}$ for all $n,\,m\in\mathbb{Z}$. A grading is called special (respectively very special) if $\mathfrak{z}^{K}\cap\mathfrak{n}^{K}_{0}=0$ (respectively if $\mathfrak{n}^{K}_{0}=0$). A grading is called non-negative (respectively positive) if $\mathfrak{n}^{K}_{n}=0$ for $n<0$ (respectively $\mathfrak{n}^{K}_{n}=0$ for $n\leq 0$). For any field $K$ of characteristic zero, a $K$-grading of $\mathfrak{n}$ can be identified with an algebraic group morphism $\rho:\mathbb{G}_{m}\rightarrow{\bf G}$ defined over $K$, where $\mathbb{G}_{m}$ denotes the multiplicative group. Consider the following two hypotheses (${\cal H}_{K}$) The Lie algebra $\mathfrak{n}$ admits a special $K$-grading, (${\cal H}^{0}_{K}$) The Lie algebra $\mathfrak{n}$ admits a very special $K$-grading. ###### Lemma 6. Let $K$ be the splitting field of ${\bf H}$. Up to conjugacy, any grading of $\mathfrak{n}^{\mathbb{C}}$ is defined over $K$. In particular (i) The hypotheses ${\cal H}_{\mathbb{C}}$ and ${\cal H}_{\overline{\mathbb{Q}}}$ are equivalent. (ii) The hypotheses ${\cal H}^{0}_{\mathbb{C}}$ and ${\cal H}^{0}_{\overline{\mathbb{Q}}}$ are equivalent. ###### Proof. Let $\mathfrak{n}^{\mathbb{C}}=\oplus_{n\in\mathbb{Z}}\,\mathfrak{n}_{n}^{\mathbb{C}}$ be a grading of $\mathfrak{n}^{\mathbb{C}}$ and let $\rho:\mathbb{G}_{m}\rightarrow{\bf G}$ be the corresponding algebraic group morphism. Since any maximal torus of ${\bf G}$ is ${\bf G}(\mathbb{C})$-conjugate to $\bf H$, it can be assumed that $\rho(\mathbb{G}_{m})\subset{\bf H}$. Let $X({\bf H})$ be the character group of ${\bf H}$. The group morphism $\rho$ is determined by the dual morphism $L:X({\bf H})\rightarrow\mathbb{Z}=X(\mathbb{G}_{m})$. However, ${\mathrm{Gal}}(K)$ acts trivially on $X({\bf H})$. Thus $\rho$ is automaticaly defined over $K$. ∎ ###### Lemma 7. Let $\Lambda$ be a finitely generated abelian group and let $S\subset\Lambda$ be a finite subset containing no element of finite order. Then there exists a morphism $L:\Lambda\rightarrow\mathbb{Z}$ such that $L(\lambda)\neq 0$ for any $\lambda\in S$. ###### Proof. Let $F$ be the subgroup of finite order elements in $\Lambda$. Using $\Lambda/F$ instead of $\Lambda$, it can be assumed that $\Lambda=\mathbb{Z}^{d}$ for some $d$ an $0\notin S$. Let’s choose a positive integer $N$ such that $S\subset\,]-N,N[^{d}$ and let $L:\Lambda\rightarrow\mathbb{Z}$ be the function defined by $L(a_{1},\dots,a_{d})=\sum_{1\leq i\leq d}a_{i}N^{i-1}$. For any $\lambda=(a_{1},\dots,a_{d})\in S$, there is a smallest index $i$ with $a_{i}\neq 0$. We have $L(\lambda)=a_{i}N^{i-1}$ modulo $N^{i}$. Since $|a_{i}|<N$, it follows that $L(\lambda)\neq 0\,\mathrm{mod}\,N^{i}$ and therefore $L(\lambda)\neq 0$. ∎ ###### Lemma 8. Let $f\in{\bf G}(\mathbb{Q})$. There is a $f$-invariant $\mathbb{Z}$-grading of $\mathfrak{n}^{\overline{\mathbb{Q}}}$ such that all eigenvalues of $f$ on $\mathfrak{n}_{0}^{{\overline{\mathbb{Q}}}}$ are roots of unity. In particular, if ${\mathrm{Spec}}\,\,f$ contains no root of unity, then $\mathfrak{n}^{\overline{\mathbb{Q}}}$ admits a very special grading. ###### Proof. Let $\Lambda\subset\overline{\mathbb{Q}}^{*}$ be the subgroup generated by the ${\mathrm{Spec}}\,\,f$. For any $\lambda\in\Lambda$ denote by $E_{(\lambda)}\subset\mathfrak{n}^{\overline{\mathbb{Q}}}$ the corresponding generalized eigenspace of $f$. Let $R$ be the set of all roots of unity in ${\mathrm{Spec}}\,\,f$ and set $S={\mathrm{Spec}}\,\,f\setminus R$. By Lemma 7, there is a morphism $L:\Lambda\rightarrow\mathbb{Z}$ such that $L(\lambda)\neq 0$ for any $\lambda\in S$. Let ${\cal G}$ be the decomposition $\mathfrak{n}^{\overline{\mathbb{Q}}}=\oplus_{k\in\mathbb{Z}}\,\mathfrak{n}^{\overline{\mathbb{Q}}}_{k}$ of $\mathfrak{n}^{\overline{\mathbb{Q}}}$ defined by $\mathfrak{n}^{\overline{\mathbb{Q}}}_{k}=\oplus_{L(\lambda)=k}\,E_{(\lambda)}$. Since $[E_{(\lambda)},E_{(\mu)}]\subset E_{(\lambda\mu)}$ and $L(\lambda\mu)=L(\lambda)+L(\mu)$ for any $\lambda,\,\mu\in\Lambda$, it follows that ${\cal G}$ is a grading of the Lie algebra $\mathfrak{n}^{\overline{\mathbb{Q}}}$. Moreover we have $\mathfrak{n}^{\overline{\mathbb{Q}}}_{0}=\oplus_{\lambda\in R}\,E_{(\lambda)}$, from which the lemma follows. ∎ ###### Lemma 9. With the previous notations (i) the Lie algebra $\mathfrak{n}^{\mathbb{C}}$ admits a special grading iff ${\cal S}(\mathfrak{n})\neq\emptyset$. (ii) the Lie algebra $\mathfrak{n}^{\mathbb{C}}$ admits a very special grading iff ${\cal V}(\mathfrak{n})\neq\emptyset$. ###### Proof. In order to prove Assertion (i), let’s consider the following assertion (${\cal A}$) $H^{0}({\bf H}(\overline{\mathbb{Q}}),\mathfrak{z}^{\overline{\mathbb{Q}}})=0$. The proof is based on the following ”cycle” of implications $\mathfrak{n}^{\mathbb{C}}$ has a special grading $\Rightarrow({\cal A})$ $\Rightarrow{\cal S}(\mathfrak{n})\neq\emptyset\Rightarrow\mathfrak{n}^{\mathbb{C}}$ has a special grading. Step 1: the existence of a special grading of $\mathfrak{n}^{\mathbb{C}}$ implies (${\cal A}$). By hypothesis and Lemma 6, $\mathfrak{n}^{\overline{\mathbb{Q}}}$ admits a special grading. Let $\rho:\mathbb{G}_{m}\rightarrow{\bf G}$ be the corresponding group morphism. Since all maximal tori of ${\bf G}$ are conjugate to ${\bf H}$, we can assume that $\rho(\mathbb{G}_{m})\subset{\bf H}$. Therefore we have $H^{0}({\bf H}({\overline{\mathbb{Q}}}),\mathfrak{z}^{{\overline{\mathbb{Q}}}})\subset H^{0}(\rho({\overline{\mathbb{Q}}}^{*}),\mathfrak{z}^{{\overline{\mathbb{Q}}}})=0$. Thus Assertion ${\cal A}$ is proved. Step 2: proof that (${\cal A}$) implies ${\cal S}(\mathfrak{n})\neq\emptyset$. By Theorem 1, there exists $f\in{\bf H}(\mathbb{Q})$ such that $\chi(f)$ is not an algebraic integer for any non-trivial character $\chi\in X({\bf H})$. If we assume (${\cal A}$), then ${\mathrm{Spec}}\,\,f|_{\mathfrak{z}}$ contains no algebraic integers and therefore ${\cal S}(\mathfrak{n})\neq\emptyset$. Step 3: proof that ${\cal S}(\mathfrak{n})\neq\emptyset$ implies the existence of a special grading. For any $f\in{\cal S}(\mathfrak{n})$, Since ${\mathrm{Spec}}\,\,f|_{\mathfrak{z}}$ contains no roots of unity. It follows from Lemma 8 that the Lie algebra $\mathfrak{n}^{\overline{\mathbb{Q}}}$ (and therefore $\mathfrak{n}^{\mathbb{C}}$) admits a special grading. Therefore ${\cal S}(\mathfrak{n})\neq\emptyset$ implies the existence of a special grading. The proof of Assertion (ii) is almost identical. Instead of $({\cal A})$, the ”cycle” of implications uses the following assertion (${\cal A}^{0}$) $H^{0}({\bf H}(\overline{\mathbb{Q}}),\mathfrak{n}^{\overline{\mathbb{Q}}})=0$. ∎ ###### Lemma 10. The following are equivalent: (i) the Lie algebra $\mathfrak{n}^{\mathbb{Q}}$ admits a non-negative special grading, (ii) the Lie algebra $\mathfrak{n}^{\mathbb{C}}$ admits a non-negative special grading, and (iii) The set ${\cal F}(\mathfrak{n})$ is not empty. ###### Proof. Proof that $(ii)\Rightarrow(iii)$. Let $\mathfrak{n}^{\mathbb{C}}=\oplus_{k\geq 0}\,\mathfrak{n}^{\mathbb{C}}_{k}$ be a non-negative special grading of $\mathfrak{n}^{\mathbb{C}}$ and let $\rho:\mathbb{G}_{m}\to{\bf G}$ be the corresponding group morphism. Up to conjugacy, we can assume that $\rho(\mathbb{G}_{m})\subset{\bf H}$. It follows that the grading is defined over the splitting field $K$ of ${\bf H}$. Let $g_{1}\in{\bf H}(K)$ be the isomorphism defined by $g_{1}x=2^{k}x$ if $x\in\mathfrak{n}_{k}^{\mathbb{C}}$. Set $n=[K:\mathbb{Q}]$ and let $g_{1},g_{2}\dots g_{n}$ be the ${\mathrm{Gal}}(L/\mathbb{Q})$-conjugates of $g_{1}$. Since all $g_{i}$ belongs to ${\bf H}(K)$, the automorphisms $g_{i}$ commute. Hence the product $g:=g_{1}\dots g_{n}$ is well defined and $g$ belongs to ${\bf H}(\mathbb{Q})$. By hypotheses, all eigenvalues of $g_{i}$ are power of $2$, and all eigenvalues of $g_{i}|_{\mathfrak{z}^{\mathbb{C}}}$ are distinct from $1$. Therefore all eigenvalues of $g$ are integers, and all eigenvalues of $g|_{\mathfrak{z}^{\mathbb{C}}}$ are $\neq\pm 1$. It follows that $g^{-1}$ belongs to ${\cal F}(\mathfrak{n})$. Therefore ${\cal F}(\mathfrak{n})\neq\emptyset$ Proof that $(iii)\Rightarrow(i)$. Let $f\in{\cal F}(\mathfrak{n})$ and set $g=f^{-1}$. Set $K=\mathbb{Q}({\mathrm{Spec}}\,\,g)$ and let $L:K^{*}\rightarrow\mathbb{Z}$ be the map defined by $L(x)=\sum_{p}\,v_{p}(N_{K/\mathbb{Q}}(x))$ where the sum runs over all prime numbers $p$ and where $N_{K/\mathbb{Q}}:K^{*}\rightarrow\mathbb{Q}^{*}$ denotes the norm map. For any integer $k$, set $\mathfrak{n}_{k}^{\overline{\mathbb{Q}}}=\bigoplus\limits_{L(x)=k}E_{(x)}$ where $E_{(x)}\subset\mathfrak{n}^{\overline{\mathbb{Q}}}$ denotes the generalized eigenspace associated to $x\in{\mathrm{Spec}}\,\,g$. We have $[E_{(x)},E_{(y)}]\subset E_{(xy)}$ and $L(xy)=L(x)+L(y)$, for any $x,\,y\in K$. Therefore the decomposition $\mathfrak{n}^{K}=\oplus_{k\in\mathbb{Z}}\,\mathfrak{n}_{k}^{\overline{\mathbb{Q}}}$ is a grading $\cal G$ of the Lie algebra $\mathfrak{n}^{\overline{\mathbb{Q}}}$. Since the function $L$ is ${\mathrm{Gal}}(\mathbb{Q})$-invariant, the grading $\cal G$ is indeed defined over $\mathbb{Q}$. It remains to prove that $\cal G$ is non-negative and special. Since any $x\in{\mathrm{Spec}}\,\,g$ is an algebraic integer, we have $L(x)\geq 0$ and the grading is non-negative. Since no $x\in{\mathrm{Spec}}\,\,g|_{\mathfrak{z}}$ is an algebraic unit, we have $N_{K/\mathbb{Q}}(x)\neq\pm 1$ and $L(x)>0$. Thus the grading is special, what proves that $(iii)\implies(i)$. ∎ ###### Lemma 11. The following are equivalent: (i) the Lie algebra $\mathfrak{n}^{\mathbb{Q}}$ admits a positive grading, (ii) the Lie algebra $\mathfrak{n}^{\mathbb{C}}$ admits a positive grading, and (iii) The set ${\cal F}^{+}(\mathfrak{n})$ is not empty. Since the proof is almost identical to the previous proof, it will be skipped. The equivalence $(i)\Leftrightarrow(ii)$ also appears in [8]. ## 4 Height and relative complexity For the whole chapter, $V$ denotes a finite dimensional vector space over $\mathbb{Q}$. In this section, we define the notion of the height of the isomorphisms $h\in GL(V)$ and the notion of a minimal lattice. 4.1 Height, complexity and minimality Let $h\in GL(V)$. Recall that a lattice of $V$ is a finitely generated subgroup $\Lambda$ which contains a basis of $V$. Let ${\cal D}(h)$ be the set of all couple of lattices $(\Lambda,E)$ of $V$ such that $E\subset\Lambda$ and $h(E)\subset\Lambda$. By definition, the height of $h$, is the integer ${\mathrm{ht}}(h):=\mathrm{Min}_{(\Lambda,E)\in{\cal D}(h)}\,[\Lambda:E]$. Let ${\cal D}_{min}(h)$ be the set of all couples $(\Lambda,E)\in{\cal D}(h)$ such that $[\Lambda:E]={\mathrm{ht}}(h)$. Similarly, for a lattice $\Lambda$ of $V$, the $h$-complexity of $\Lambda$ is the integer $\mathrm{cp}_{h}(\Lambda):=\mathrm{Min}_{(\Lambda,E)\in{\cal D}(h)}\,[\Lambda:E]$. It is clear that $\mathrm{cp}_{h}(\Lambda)=[\Lambda:E]$, where $E=\Lambda\cap h^{-1}(\Lambda)$. The lattice $\Lambda$ is called minimal relative to $h$ if $\mathrm{cp}_{h}(\Lambda)={\mathrm{ht}}(h)$. For the proofs, a technical notion of relative height is needed. Let $\mathrm{End}_{h}(V)$ be the commutant of $h$ and let let $A\subset C(h)\subset\mathrm{End}_{h}(V)$ be a subring. By definition, an $A$-lattice $\Lambda$ means a lattice $\Lambda$ which is an $A$-module. Let ${\cal D}^{A}(h)$ be the set of all couple of $A$-lattices $(\Lambda,E)$ in ${\cal D}(h)$. The $A$-height of $h$ is the integer ${\mathrm{ht}}_{A}(h):=\mathrm{Min}_{(\Lambda,E)\in{\cal D}^{A}(h)}\,[\Lambda:E]$. Obviously, we have ${\mathrm{ht}}_{A}(h)\geq{\mathrm{ht}}(h)={\mathrm{ht}}_{\mathbb{Z}}(h)$. Let ${\cal D}^{A}_{min}(h)$) be the set of all couples $(\Lambda,E)\in{\cal D}^{A}(h)$) such that $[\Lambda:E]={\mathrm{ht}}_{A}(h)$. 4.2 Height and filtrations Let $V$ be a finite dimensional vector space over $\mathbb{Q}$ and let $h\in GL(V)$. Let $A$ be a subring of $\mathrm{End}_{h}(V)$ and let $A[h]$ be the subring of $\mathrm{End}_{h}(V)$ generated by $A$ and $h$. ###### Lemma 12. Let $0=V_{0}\subset V_{1}\subset\dots\subset V_{n}=V$ be a fitration of $V$, where each vector space $V_{i}$ is a $A[h]$-submodule. For $i=1$ to $n$, set $h_{i}=h_{V_{i}/V_{i-1}}$. Then we have ${\mathrm{ht}}_{A}(h)\geq\prod_{1\leq i\leq n}\,{\mathrm{ht}}_{A}(h_{i})$. Moreover if $V\simeq\oplus V_{i}/V_{i-1}$ as a $A[h]$-module, we have ${\mathrm{ht}}_{A}(h)=\prod_{1\leq i\leq n}\,{\mathrm{ht}}_{A}(h_{i})$. ###### Proof. Clearly it is enough to prove the lemma for $n=2$. Let $(\Lambda,E)\in{\cal D}^{A}_{min}(h)$. Set $\Lambda_{1}=\Lambda\cap V_{1}$, $E_{1}=E\cap V_{1}$, $\Lambda_{2}=\Lambda/\Lambda_{1}$ and $E_{2}=E/E_{1}$. We have $[\Lambda:E]=[\Lambda_{1}:E_{1}][\Lambda_{2}:E_{2}]$. Since $(\Lambda_{1},E_{1})\in{\cal D}^{A}(h_{1})$ and $(\Lambda_{2},E_{2})\in{\cal D}^{A}(h_{2})$, we have ${\mathrm{ht}}_{A}(h)\geq{\mathrm{ht}}_{A}(h_{1})\,{\mathrm{ht}}_{A}(h_{2})$, what proves the first assertion. Next, we assume that $V\simeq V_{1}\oplus V_{2}$ as a $A[h]$-module. Let $(\Lambda_{1},E_{1})\in{\cal D}_{min}^{A}(h_{1})$, $(\Lambda_{2},E_{2})\in{\cal D}_{min}^{A}(h_{2})$ and set $\Lambda=\Lambda_{1}\oplus\Lambda_{2}$ and $E=E_{1}\oplus E_{2}$. We have $[\Lambda:E]=[\Lambda_{1}:E_{1}][\Lambda_{2}:E_{2}]={\mathrm{ht}}_{A}(h_{1})\,{\mathrm{ht}}_{A}(h_{2})$. Therefore ${\mathrm{ht}}_{A}(h)\leq{\mathrm{ht}}_{A}(h_{1})\,{\mathrm{ht}}_{A}(h_{2})$. Hence ${\mathrm{ht}}_{A}(h)={\mathrm{ht}}_{A}(h_{1})\,{\mathrm{ht}}_{A}(h_{2})$. ∎ Let $h\in GL(V)$ as before. Its Chevalley decomposition $h=h_{s}h_{u}$ is uniquely defined by the following three conditions: $h_{s}$ and $h_{u}$ commutes, $h_{s}$ is semi-simple and $h_{u}$ is unipotent. ###### Lemma 13. We have ${\mathrm{ht}}(h)={\mathrm{ht}}(h_{s})$. ###### Proof. By Lemma 12, it can be assumed that the $\mathbb{Q}[h]$-module $V$ is indecomposable. Therefore there is a vector space $V_{0}$, a semi-simple endomorphism $h_{0}\in\mathrm{End}(V_{0})$ and an isomorphism $V\simeq V_{0}\otimes\mathbb{Q}[t]/(t^{n})$, relative to which $h_{s}$ acts as $h_{0}\otimes 1$ and $h_{u}$ acts as $1\otimes t$. Let $(\Lambda_{0},E_{0})\in{\cal D}_{min}(h_{0})$ and set $\Lambda=\Lambda_{0}\otimes\mathbb{Z}[t]/(t^{n})$ and $E=E_{0}\otimes\mathbb{Z}[t]/(t^{n})$. By Lemma 12, we have ${\mathrm{ht}}(h)\geq{\mathrm{ht}}(h_{s})={\mathrm{ht}}(h_{0})^{n}$. Since $(\Lambda,E)\in{\cal D}(h)$ and $[\Lambda:E]=[\Lambda_{0}:E_{0}]^{n}={\mathrm{ht}}(h_{0})^{n}$, it follows that ${\mathrm{ht}}(h)={\mathrm{ht}}(h_{s})$ ∎ 4.3 Complexity of ${\cal O}(h)$-lattices For any algebraic number $\lambda$, let ${\cal O}(\lambda)$ be the ring of integers of the number field $\mathbb{Q}(\lambda)$. Set $\pi_{\lambda}=\\{x\in{\cal O}(\lambda)|\,x\lambda\in{\cal O}(\lambda)\\}$. Then $\pi_{\lambda}$ is an integral ideal and its norm is the integer $d(\lambda):=\mathrm{N}_{\mathbb{Q}(\lambda)/\mathbb{Q}}(\pi_{\lambda})=\mathrm{Card\,}{\cal O}(\lambda)/\pi_{\lambda}$. Let $h\in GL(V)$ be semi-simple. Let $P(t)$ be its minimal polynomial, let $P=P_{1}\dots P_{k}$ be its factorization into irreducible factors. For $1\leq i\leq k$, set $K_{i}=\mathbb{Q}[t]/(P_{i}(t))$ and let ${\cal O}_{i}$ be the ring of integers of the number field $K_{i}$. Set ${\cal O}(h)=\oplus_{1\leq i\leq k}\,{\cal O}_{i}$. For each $\lambda\in{\mathrm{Spec}}\,\,h$, let $m_{\lambda}$ be its multiplicity. Note that the functions $\lambda\mapsto m_{\lambda}$ and $\lambda\mapsto d(\lambda)$ are ${\mathrm{Gal}}(\mathbb{Q})$-invariant, so they can be viewed as functions defined over ${\mathrm{Spec}}\,\,h/{\mathrm{Gal}}(\mathbb{Q})$. ###### Lemma 14. Let $\Lambda$ be an ${\cal O}(h)$-lattice of $V$. Then $\mathrm{cp}_{h}(\Lambda)=\prod\,d(\lambda)^{m_{\lambda}}$, where the product runs over ${\mathrm{Spec}}\,h/{\mathrm{Gal}}(\mathbb{Q})$. ###### Proof. With the previous notations, let $e_{i}$ be the unit of ${\cal O}_{i}$ and set $\Lambda_{i}=e_{i}\Lambda$. Since $\Lambda=\oplus_{1\leq i\leq k}\,\Lambda_{i}$, it is enough to prove the lemma for $k=1$, i.e. when the minimal polynomial of $h$ is irreducible. Let $\lambda$ be one eigenvalue of $h$. With these new hypotheses, we have $\mathbb{Q}[h]/(P(t))\simeq\mathbb{Q}(\lambda)$, ${\cal O}(h)\simeq{\cal O}(\lambda)$ and $V$ is a vector space of dimension $m_{\lambda}$ over $\mathbb{Q}(\lambda)$, relative to which $h$ is identified with the multiplication by $\lambda$. We have $r_{\lambda}\Lambda=\Lambda\cap h^{-1}\Lambda$. Since $\Lambda/r_{\lambda}{\cal I}\simeq({\cal O}(\lambda)/r_{\lambda})^{m_{\lambda}}$, it follows that $\mathrm{cp}_{h}(\Lambda)=d(\lambda)^{m_{\lambda}}$. ∎ 4.4 Computation of the height Let $h\in GL(V)$ be semi-simple. ###### Lemma 15. We have ${\mathrm{ht}}(h)=\prod\,d(\lambda)^{m_{\lambda}}$, where the product runs over ${\mathrm{Spec}}\,h/{\mathrm{Gal}}(\mathbb{Q})$. ###### Proof. Using Lemmas 13 and Lemma 12, we can be assumed $V$ is a simple $\mathbb{Q}[h]$-module, and let $n$ be its dimension. The eigenvalues $\lambda_{1},\dots,\lambda_{n}$ of $h$ are conjugate by ${\mathrm{Gal}}(\mathbb{Q})$. Under these simplifying hypotheses, the formula to be proved is ${\mathrm{ht}}(h)=d(\lambda_{1})$. Step 1: scalar extension. Set $K=\mathbb{Q}(\lambda_{1},\dots,\lambda_{n})$, let $U=K\otimes V$, let $\tilde{h}=1\otimes h$ be the extension of $h$ to $U$ and let $\\{v_{1},\dots,v_{n}\\}$ be a $K$ basis of $U$ such that $\tilde{h}.v_{i}=\lambda_{i}\,v_{i}$. We have $U=\oplus_{1\leq i\leq n}U_{i}$, where $U_{i}=K\,v_{i}$. Let $\cal O$ be the ring of integers of $K$. For each $1\leq i\leq n$, set $\tilde{h}_{i}=h|_{U_{i}}$. Since each $U_{i}$ is a ${\cal O}[\tilde{h}]$-module, Lemma 12 shows that ${\mathrm{ht}}_{\cal O}(\tilde{h})=\prod_{1\leq i\leq n}\,{\mathrm{ht}}_{\cal O}(\tilde{h}_{i})$. Next, the integers ${\mathrm{ht}}_{\cal O}(\tilde{h}_{i})$ are computed. Let $\Lambda_{i}\subset U_{i}$ be any ${\cal O}$-lattice. Since ${\cal O}$ contains ${\cal O}(\lambda_{i})={\cal O}(\tilde{h}_{i})$, it follows from Lemma 14 that $\mathrm{cp}_{\tilde{h}_{i}}(\Lambda_{i})=d(\lambda_{i})^{r}$ where $r=\mathrm{rk}_{\cal O}(\Lambda_{i})=[K:\mathbb{Q}(\lambda_{i})]$. Hence we have ${\mathrm{ht}}_{\cal O}(\tilde{h}_{i})=d(\lambda_{i})^{[K:\mathbb{Q}(\lambda_{i})]}$. It follows that ${\mathrm{ht}}_{\cal O}(\tilde{h})=\prod_{1\leq i\leq n}\,d(\lambda_{i})^{[K:\mathbb{Q}(\lambda_{i})]}=d(\lambda_{1})^{[K:\mathbb{Q}]}$ Step 2: end of the proof. Now let $(\Lambda,E)\in{\cal D}_{min}(h)$. Set $\tilde{\Lambda}={\cal O}\otimes\Lambda$ and $\tilde{E}={\cal O}\otimes E$. Since $\tilde{E}$ is an ${\cal O}$-module, we have $[\tilde{\Lambda}:\tilde{E}]\geq{\mathrm{ht}}_{\cal O}(\tilde{h})$. It follows that ${\mathrm{ht}}(h)^{[K:\mathbb{Q}]}=[\Lambda:E]^{[K:\mathbb{Q}]}=[\tilde{\Lambda}:\tilde{E}]\geq{\mathrm{ht}}_{\cal O}(\tilde{h})=d(\lambda)^{[K:\mathbb{Q}]}$. Thus we have $d(\lambda_{1})\leq{\mathrm{ht}}(h)$. By Lemma 14, we have ${\mathrm{ht}}_{{\cal O}(h)}(h)=d(\lambda_{1})$. It follows that $d(\lambda_{1})\leq{\mathrm{ht}}(h)\leq{\mathrm{ht}}_{\cal O(h)}(h)=d(\lambda_{1})$, what proves the formula. ∎ Remark: In number theory, the Weil height of an algebraic number $\lambda$ is $H(\lambda)=\theta d(\lambda)^{1/n}$, where $\theta$ involves the norms at infinite places. Therefore ${\mathrm{ht}}(h)$ is essentially the Weil’s height of $h$, up to the factor at infinite places. 4.5 A simple criterion of minimality An obvious consequence of Lemmas 14 and 15 is ###### Lemma 16. Let $h\in GL(V)$ be semi-simple and let $\Lambda$ be an ${\cal O}(h)$-lattice of $V$. Then $\Lambda$ is minimal relative to $h$. ## 5 Malcev’s Theorem and self-similar data In this chapter, we recall Malcev’s Theorem. Then we collect some related results, which are due to Malcev or viewed as folklore results. Then it is easy to characterize the self-similar data for FGTF nilpotent groups. 5.1 Three types of lattices Let $\mathfrak{n}$ be a finite dimensional be a nilpotent Lie algebra over $\mathbb{Q}$. The Lie algebra $\mathfrak{n}$ is endowed with two group structures, the addition and the the Campbell-Hausdorff product. To avoid confusion, the Campbell-Hausdorff product is called the multiplication and it is denoted accordingly. A multiplicative subgroup $\Gamma$ of $\mathfrak{n}$ means a subgroup relative to the Campbell-Hausdorff product. In general, a multiplicative subgroup $\Gamma$ is not an additive subgroup of $\mathfrak{n}$. However, notice that $\mathbb{Z}.x\subset\Gamma$ for any $x\in\Gamma$, because $x^{n}=nx$ for any $n\in\mathbb{Z}$. A finitely generated multiplicative subgroup $\Gamma$ is called a multiplicative lattice if $\Gamma\,\mathrm{mod}\,[\mathfrak{n},\mathfrak{n}]$ generates the $\mathbb{Q}$-vector space $\mathfrak{n}/[\mathfrak{n},\mathfrak{n}]$, or, equivalently, if $\Gamma$ generates the Lie algebra $\mathfrak{n}$. Let $N$ be the CSC nilpotent Lie group with Lie algebra $\mathfrak{n}^{R}=\mathbb{R}\otimes\mathfrak{n}$. A discrete subroup $\Gamma$ of $N$ is called a cocompact lattice if $N/\Gamma$ is compact. It should be noted that three distinct notions of lattices will be used in the sequel: the additive lattices, the multiplicative lattices and the cocompact lattices. When it is used alone, a lattice is always an additive lattice. This very commoun terminology could be confusing: the reader should read ”multiplicative lattice” or ”cocompact lattice” as single words. 5.2 Malcev’s Theorem Any multiplicative lattice $\Gamma$ of a finite dimensional nilpotent Lie algebra over $\mathbb{Q}$ is a FGTF nilpotent group. Conversely, Malcev proved in [18] ###### Malcev’s Theorem. Let $\Gamma$ be a FGTF nilpotent group. 1\. There exists a unique nilpotent Lie algebra $\mathfrak{n}$ over $\mathbb{Q}$ wich contains $\Gamma$ as a multiplicative lattice. 2\. There exists a unique CSC nilpotent Lie group $N$ which contains $\Gamma$ as a cocompact lattice. 3\. The Lie algebra of $N$ is $\mathbb{R}\otimes\mathfrak{n}$. The Lie algebra $\mathfrak{n}$ of the previous theorem will be called the Malcev Lie algebra of $\Gamma$. 5.3 The coset index From now on, let $\mathfrak{n}$ will be a finite dimensional nilpotent Lie algebra. The coset index, which is defined now, generalizes the notions of indices for additive lattices and for multiplicative lattices. A subset $X$ of $\mathfrak{n}$ is called a coset union if $X$ is a finite union of $\Lambda$-coset for some additive lattice $\Lambda$. Recall that the nilpotency index of $\mathfrak{n}$ is the smallest integer $n$ such that $C^{n+1}\mathfrak{n}=0$, where $(C^{n}\,\mathfrak{n})_{n\geq 0}$ is its descending central series. The following lemma is easily proved by induction on the nilpotency index of $\mathfrak{n}$. ###### Lemma 17. Any multiplicative lattice $\Gamma$ of $\mathfrak{n}$ is a coset union. Let $X\supset Y$ be two coset unions in $\mathfrak{n}$. Obviously, there is a lattice $\Lambda$ such that $X$ and $Y$ are both a finite union of $\Lambda$-coset. The coset index of $Y$ in $X$ is the number $[X:Y]_{coset}={\mathrm{Card\,}\,X/\Lambda\over\mathrm{Card\,}\,Y/\Lambda}$ The numerator and denominator of the previous expression depends on the choice of $\Lambda$, but $[X:Y]_{coset}$ is well defined. In general, the coset index is not an integer. Obviously if $\Lambda\supset\Lambda^{\prime}$ are additive lattices in $\mathfrak{n}$, we have $[\Lambda:\Lambda^{\prime}]_{coset}=[\Lambda:\Lambda^{\prime}]$. Similarly, for multiplicative lattices there is ###### Lemma 18. Let $\Gamma\supset\Gamma^{\prime}$ be multiplicative lattices in $\mathfrak{n}$, we have $[\Gamma:\Gamma^{\prime}]_{coset}=[\Gamma:\Gamma^{\prime}]$. The proof, done by induction on the nipotency index of $\mathfrak{n}$, is skipped. 5.4 Morphims of FGTF nilpotent groups ###### Lemma 19. Let $\Gamma$, $\Gamma^{\prime}\subset\mathfrak{n}$ be multiplicative lattices in $\mathfrak{n}$ and let $f:\Gamma^{\prime}\to\Gamma$ be a group morphism. Then $f$ extends uniquely to a Lie algebra morphism $\tilde{f}:\mathfrak{n}\to\mathfrak{n}$. Moreover $\tilde{f}$ is an isomorphism if $f$ is injective. When $f$ is an isomorphism, the result is due to Malcev, see [18], Theorem 5. In general, the lemma is a folklore result and it is implicitely used in Homotopy Theory, see e.g. [1]. Since we did not found a precise reference, a proof, essentially based on Hall’s collecting formula (see Theorem 12.3.1 in [15]), is now provided. ###### Proof. Let $x\in\mathfrak{n}$. Since $\Gamma$ contains an additive lattice by Lemma 17, we have $m\mathbb{Z}x\subset\Gamma$ for some $m>0$. Thus there is a unique map $\tilde{f}:\mathfrak{n}\to\mathfrak{n}$ extending $f$ such that $\tilde{f}(nx)=n\tilde{f}(x)$ for any $x\in\mathfrak{n}$ and $n\in\mathbb{Z}$. It remains to prove that $\tilde{f}(x+y)=\tilde{f}(x)+\tilde{f}(y)$, and $\tilde{f}([x,y])=[\tilde{f}(x),\tilde{f}(y)]$, for any $x,\,y\in\mathfrak{n}$. Let $n$ be the nilpotency index of $\mathfrak{n}$. Set ${\cal L}(2,n)={\cal L}(2)/C^{n+1}{\cal L}(2)$, where ${\cal L}(2)$ denotes the free Lie algebra over $\mathbb{Q}$ freely generated by $X$ and $Y$. Let $\Gamma(2,n)\subset{\cal L}(2,n)$ be the multiplicative subgroup generated by $X$ and $Y$. As before, $m(X+Y)$ and $m[X,Y]$ belongs to $\Gamma(2,n)$ for some $m>0$. Thus there are $w_{1},\,w_{2}$ in the free group over two generators, such that $w_{1}(X,Y)=m(X+Y)$ and $w_{2}(X,Y)=m[X,Y]$. Since ${\cal L}(2,n)$ is a free in the category of nilpotent Lie algebras of nilpotency index $\leq n$, we have $w_{1}(x,y)=m(x+y)$ and $w_{2}(x,y)=m[x,y]$ for any $x,y\in\mathfrak{n}$. From this it follows easily that $\tilde{f}$ is a Lie algebra morphism. ∎ 5.6 Self-similar data for FGTF nilpotent groups Let $\mathfrak{z}$ be the center of $\mathfrak{n}$. Recall that ${\cal S}(\mathfrak{n})$ (respectively ${\cal V}(\mathfrak{n})$) is the set of all $f\in\mathrm{Aut}\,\mathfrak{n}$ such that ${\mathrm{Spec}}\,\,f|_{\mathfrak{z}}$ (respectively ${\mathrm{Spec}}\,\,f$) contains no algebraic integers. Let $\Gamma\supset\Gamma^{\prime}$ be multiplicative lattices of $\mathfrak{n}$, let $f:\Gamma^{\prime}\to\Gamma$ be a morphism and let $\tilde{f}:\mathfrak{n}\to\mathfrak{n}$ be its extension. ###### Lemma 20. Let’s assume that $f$ is injective. Then (i) $(\Gamma^{\prime},f)$ is a self-similar datum iff $\tilde{f}$ belongs to ${\cal S}(\mathfrak{n})$, (ii) $(\Gamma^{\prime},f)$ is a free self-similar datum iff $\tilde{f}$ belongs to ${\cal V}(\mathfrak{n})$ (iii) if $(\Gamma^{\prime},f)$ is a fractal datum, then $f$ belongs to ${\cal F}(\mathfrak{n})$. ###### Proof. Let $V$ be a finite dimensional vector space over $\mathbb{Q}$ and let $f\in GL(V)$. We will repeatedly use the fact that ${\mathrm{Spec}}\,\,f$ contains an algebraic integer iff $V$ contains a finitely generated subgroup $E\neq 0$ such that $f(E)\subset E$. Proof of Assertion (i). Since $\Gamma^{\prime}$ contains a set of generators of $\mathfrak{n}$, the subgroup $Z(\Gamma^{\prime}):=\Gamma^{\prime}\cap\mathfrak{z}$ is the center of $\Gamma^{\prime}$. Let $K$ be the $f$-core of the virtual endomorphism $(\Gamma^{\prime},f)$. Let’s assume that $(\Gamma^{\prime},f)$ is not a self-similar datum. Since $K\neq 1$, the additive group $K\cap Z(\Gamma^{\prime})$ is non-trivial, finitely generated and $\tilde{f}$-invariant. Therefore ${\tilde{f}}\notin{\cal S}(\mathfrak{n})$. Conversely let’s assume that ${\tilde{f}}\notin{\cal S}(\mathfrak{n})$. Then there is a nonzero finitely generated subgroup $E\subset\mathfrak{z}$ such that $\tilde{f}(E)\subset E$. By Lemma 17, $Z(\Gamma^{\prime})$ is an additive lattice of $\mathfrak{z}$. Therefore we have $mE\subset Z(\Gamma^{\prime})$ for some $m>0$. Since $K$ contains $mE$, it follows that $(\Gamma^{\prime},f)$ is not a self-similar datum. Proof of Assertion (ii). Let $A\subset\Gamma$ be a set of representatives of $\Gamma/\Gamma^{\prime}$. Let’s consider the action of $\Gamma$ on $A^{\omega}$ associated with the virtual endomorphism $(\Gamma^{\prime},f)$. Let’s assume that ${\tilde{f}}\notin{\cal V}(\mathfrak{n})$. Then there is a nonzero finitely generated abelian subgroup $F\subset\mathfrak{n}$ such that $\tilde{f}(F)\subset F$. As before, it can be assumed $F$ lies in $\Gamma^{\prime}$. Let $e\in A$ be the representative of the trivial coset and let $e^{\omega}=ee\dots$ be the infinite word over the single letter $e$. Since $f(F)\subset F$, it follows that $\gamma(e^{\omega})=e^{\omega}$ for any $\gamma\in F$. Hence $\Gamma$ does not act freely on $A^{\omega}$. Conversely, let assume that $\Gamma$ does not act freely on $A^{\omega}$. Let’s define inductively the subsets ${\cal H}(n)\subset\Gamma$ by ${\cal H}(1)=\cup_{a\in A}\,a\Gamma^{\prime}a^{-1}$ and ${\cal H}(n+1)=\\{\gamma\in\Gamma|\,\exists\,a\in A:a\gamma a^{-1}\in\Gamma^{\prime}\land f(a\gamma a^{-1})\in{\cal H}(n)\\}$, for $n\geq 1$. Indeed ${\cal H}(n)$ is the set of all $\gamma\in\Gamma$ which have at least one fixed point on $A^{n}$. It follows easily that ${\cal H}:=\cap_{n\geq 1}\,{\cal H}(n)$ is the set of all $\gamma\in\Gamma$ which have at least one fixed point on $A^{\omega}$. There is an integer $k$ such that ${\cal H}\subset C^{k}\mathfrak{n}$ but ${\cal H}\not\subset C^{k+1}\mathfrak{n}$. Let $\overline{\cal H}$ be the image of ${\cal H}$ in $C^{k}\mathfrak{n}/C^{k+1}\mathfrak{n}$ and let $F$ be the additive subgroup of $C^{k}\mathfrak{n}/C^{k+1}\mathfrak{n}$ generated by $\overline{\cal H}$. Since $\Gamma$ lies in a lattice, $F$ is finitely generated. Moreover we have $axa^{-1}\equiv x\,\mathrm{mod}\,\,C^{k+1}\mathfrak{n}$, for any $x\in C^{k}\mathfrak{n}$ and $a\in A$. It follows that $\tilde{f}_{k}(\overline{\cal H})\subset\overline{\cal H}$, where $\tilde{f}_{k}$ is the linear map induced by $\tilde{f}$ on $C^{k}\mathfrak{n}/C^{k+1}\mathfrak{n}$. Hence $\tilde{f}_{k}(F)\subset F$ and ${\mathrm{Spec}}\,\,\tilde{f}_{k}$ contains an algebraic integer. Therefore ${\tilde{f}}\notin{\cal V}(\mathfrak{n})$. Proof of Assertion (iii). Let $(\Gamma^{\prime},f)$ be a fractal datum. Let $\Lambda$ be the additive lattice generated by $\Gamma$. Since $\tilde{f}^{-1}(\Lambda)\subset\Lambda$, all $x\in{\mathrm{Spec}}\,\tilde{f}^{-1}$ are algebraic integers. Therefore $\tilde{f}$ belongs to ${\cal F}(\mathfrak{n})$. ∎ ## 6 Relative complexity of multiplicative lattices This chapter is the mutiplicative analogue of ch. 4. The main result is the refined criterion of minimality. Together with Theorem 1, it is the main ingredient of the proof of Theorem 2 and 3. Throughout the whole chapter, $\mathfrak{n}$ is finite dimensional nilpotent Lie algebra over $\mathbb{Q}$, and $\mathfrak{z}$ is its center. 6.1 Complexity of multiplicative lattices Let $f\in\mathrm{Aut}\,\mathfrak{n}$ and let $\Gamma$ be a multiplicative lattice of $\mathfrak{n}$. The complexity of $\Gamma$ relative to $f$ is the integer $\mathrm{cp}_{f}(\Gamma)=[\Gamma:\Gamma^{\prime}]$, where $\Gamma^{\prime}=\Gamma\cap f^{-1}(\Gamma)$. The multiplicative lattice $\Gamma$ is called minimal relative to $f$ if $\mathrm{cp}_{f}(\Gamma)={\mathrm{ht}}(f)$. Thanks to Lemma 18 the notation $\mathrm{cp}_{f}(\Gamma)$ is unambiguous. ###### Lemma 21. Let $\Gamma$ be multiplicative lattices of $\mathfrak{n}$. Then we have $\mathrm{cp}_{f}(\Gamma)\geq{\mathrm{ht}}(f)$. ###### Proof. The proof goes by induction on the nilpotency index of $\mathfrak{n}$. Let $Z$ be the center of $\Gamma$. Set $\Gamma^{\prime}=\Gamma\cap f^{-1}(\Gamma)$, $Z^{\prime}=Z\cap f^{-1}(Z)$, $\overline{\Gamma}=\Gamma/Z$, $\overline{\Gamma^{\prime}}=\Gamma^{\prime}/Z^{\prime}$. Also set ${\overline{\mathfrak{n}}}=\mathfrak{n}/\mathfrak{z}$ and let $\overline{f}:{\overline{\mathfrak{n}}}\to{\overline{\mathfrak{n}}}$ and $f_{0}:\mathfrak{z}\to\mathfrak{z}$ be the isomorphisms induced by $f$. By induction hypothesis, we have $\mathrm{cp}_{\overline{f}}(\overline{\Gamma})\geq{\mathrm{ht}}(\overline{f})$ and therefore $[\overline{\Gamma}:\overline{\Gamma^{\prime}}]\geq{\mathrm{ht}}(\overline{f})$. By definition, we have $[Z:Z^{\prime}]=\mathrm{cp}_{f_{0}}\,Z\geq{\mathrm{ht}}(f_{0})$. Moreover by Lemma 15 we have ${\mathrm{ht}}(f)={\mathrm{ht}}(f_{0}){\mathrm{ht}}(\overline{f})$. It follows that $\mathrm{cp}_{f}\,\Gamma=[\Gamma:\Gamma^{\prime}]=[Z:Z^{\prime}]\,[\overline{\Gamma}:\overline{\Gamma^{\prime}}]\geq{\mathrm{ht}}(f_{0}){\mathrm{ht}}(\overline{f})={\mathrm{ht}}(f)$, and the statement is proved. ∎ 6.2 A property of the minimal multiplicative lattices Let $\Gamma$ be a multiplicative lattice of $\mathfrak{n}$ and let $h\in\mathrm{Aut}\,\mathfrak{n}$. For simplicity, let’s assume that $h$ is semi-simple. ###### Lemma 22. The following assertions are equivalent (i) $\Gamma$ is minimal relative to $h$, and (ii) the virtual morphism $(\Gamma^{\prime},h)$ is good, where $\Gamma^{\prime}=\Gamma\cap h^{-1}(\Gamma)$. In particular, there is a multiplicative lattice $\tilde{\Gamma}\subset\Gamma$ which is minimal relative to $h$. ###### Proof. By Lemma 17, $\Gamma$ is a coset union. Any additive lattice contains a ${\cal O}(h)$-module of finite index. Therefore there is an ${\cal O}(h)$-lattice $\Lambda$ such that $\Gamma$ is an union of $\Lambda$-cosets. Let $\Gamma_{0},\Gamma_{1},\dots$ be the multiplicative lattices inductively defined by $\Gamma_{0}=\Gamma$, $\Gamma_{1}=\Gamma^{\prime}$ and $\Gamma_{n+1}=\Gamma_{n}\cap h^{-1}(\Gamma_{n})$ for $n\geq 1$. Similarly let $\Lambda_{0},\Lambda_{1},\dots$ be the additive lattices defined by $\Lambda_{0}=\Lambda$, and $\Lambda_{n+1}=\Lambda_{n}\cap h^{-1}(\Lambda_{n})$ for $n\geq 0$. By Lemma 2, the sequence $[\Gamma_{n}:\Gamma_{n+1}]$ is not increasing. By Lemma 21, we have $[\Gamma_{n}:\Gamma_{n+1}]\geq{\mathrm{ht}}(f)$. Moreover, it follows from Lemma 16 that $[\Lambda_{n}:\Lambda_{n+1}]={\mathrm{ht}}(h)$ for all $n$. Let’s assume now that $\Gamma$ is minimal relative to $h$. We have $[\Gamma_{n}:\Gamma_{n+1}]={\mathrm{ht}}(f)$ for all $n$, and therefore the virtual morphism $(\Gamma^{\prime},h)$ is good. Conversely, let’s assume that the virtual morphism $(\Gamma^{\prime},h)$ is good. By hypotheses we have $[\Gamma_{0}:\Gamma_{n}]=[\Gamma_{0}:\Gamma_{1}]^{n}$ and $[\Lambda_{0}:\Lambda_{n}]={\mathrm{ht}}(h)^{n}$ for all $n\geq 1$. It follows that $[\Gamma_{0}:\Lambda_{n}]_{coset}=[\Gamma_{0}:\Lambda_{0}]_{coset}\,{\mathrm{ht}}(h)^{n}$. Since $\Gamma_{n}\supset\Lambda_{n}$, we have $[\Gamma_{0}:\Gamma_{n}]\leq[\Gamma_{0}:\Lambda_{n}]_{coset}$. and therefore $[\Gamma_{0}:\Gamma_{1}]^{n}\leq[\Gamma_{0}:\Lambda_{0}]_{coset}\,{\mathrm{ht}}(h)^{n}$, for all $n\geq 0$. Hence $[\Gamma_{0}:\Gamma_{1}]\leq{\mathrm{ht}}(f)$. It follows from Lemma 21 that $[\Gamma_{0}:\Gamma_{1}]={\mathrm{ht}}(f)$, thus $\Gamma$ is minimal relative to $h$. In order to prove the last assertion, notice that the sequence $[\Gamma_{n}:\Gamma_{n+1}]$ is stationary for $n\geq N$, for some $N>0$. Therefore $(\Gamma_{N+1},h)$ is a good virtual morphism of $\Gamma_{N}$. Thus the subgroup $\tilde{\Gamma}=\Gamma_{N}$ is minimal relative to $h$. ∎ 6.3 A refined criterion of minimality A refined version of Lemma 16 is now provided. Let $\Gamma$ be a multiplicative lattice in $\mathfrak{n}$ and let $h\in\mathrm{Aut}\,\mathfrak{n}$ be semi-simple. Let $L$ be the field generated by ${\mathrm{Spec}}\,h$, let ${\cal O}$ be its ring of integers and let ${\cal P}$ be the set of prime ideals of ${\cal O}$. Let $\Lambda$ be an ${\cal O}(h)$-lattice and let $n>0$ be an integer. Let’s assume that $\Lambda\supset\Gamma$ and $\Gamma$ is an union of $n\Lambda$-cosets. ###### Lemma 23. Let $S$ be the set of divisors of $n$ in ${\cal P}$. Assume that $\lambda\equiv 1\,\mathrm{mod}\,n{\cal O}_{\pi}$, for any $\lambda\in{\mathrm{Spec}}\,\,h$ and any $\pi\in S$. Then $\Gamma$ is minimal relative to $h$. ###### Proof. Step 1. Since ${\mathrm{Spec}}\,\,h$ lies in ${\cal O}_{\pi}$ for all $\pi\in S$, there exists a positive integer $d$, which is prime to $n$, such that $d\lambda\in{\cal O}$ for all $\lambda\in{\cal O}$. Moreover we can assume that $d\equiv 1\,\mathrm{mod}\,n$. Let $\lambda\in{\mathrm{Spec}}\,\,h$. We have $d\lambda\equiv 1\,\mathrm{mod}\,n{\cal O}_{\pi}$ for all $\pi\in S$. Therefore we have $d\lambda\in 1+n{\cal O}$, for all $\lambda\in{\mathrm{Spec}}\,\,h$. Set $H=dh$. Since ${\mathrm{Spec}}\,\,dH$ and ${\mathrm{Spec}}\,\,(H-1)/n$ lie in ${\cal O}$, it follows that $H\in{\cal O}(h)$ and $H\in 1+n{\cal O}(h)$. Step 2. Set $\Lambda^{\prime}=\Lambda\cap h^{-1}\Lambda$. Since all eigenvalues of $h$ are units in ${\cal O}_{\pi}$ whenever $\pi$ divides $n$, the height of $h$ is prime to $n$. By Lemma 16, we have $[\Lambda:\Lambda^{\prime}]={\mathrm{ht}}(h)$. Therefore we get $\Lambda=\Lambda^{\prime}+n\Lambda$. It follows that $\Gamma=\coprod\limits_{1\leq i\leq k}\,g_{i}+n\Lambda$ for some $g_{1},...,g_{k}\in\Lambda^{\prime}$, where $k=[\Gamma:n\Lambda]$ and where $\coprod$ is the symbol of the disjoint union. Since $H(g_{i})\equiv g_{i}\,\mathrm{mod}\,n\Lambda$, we get that $h(g_{i})\in g_{i}+n\Lambda\subset\Gamma$. Therefore we have $\Gamma^{\prime}\supset\coprod\limits_{1\leq i\leq k}\,g_{i}+n\Lambda^{\prime}$, Therefore we have $[\Gamma^{\prime}:n\Lambda^{\prime}]\geq k=[\Gamma:n\Lambda]$. It follows that $[\Gamma:\Gamma^{\prime}]\leq[n\Lambda:n\Lambda^{\prime}]={\mathrm{ht}}(h)$. By Lemma 21, we have $[\Gamma:\Gamma^{\prime}]={\mathrm{ht}}(h)$. Thus $\Gamma$ is minimal relative to $h$. ∎ ## 7 Proof of Theorems 2 and 3 7.1 Proof of Theorem 2 and 3. Let $\mathfrak{n}$ be a finite dimensional nilpotent Lie algebra over $\mathbb{Q}$ and let $\mathfrak{z}$ be its center and let $\Gamma$ be a multiplicative lattice of $\mathfrak{n}$. ###### Theorem 2. The following assertions are equivalent (i) The group $\Gamma$ is transitive self-similar, (ii) the group $\Gamma$ is densely self-similar, and (iii) the Lie algebra $\mathfrak{n}^{\mathbb{C}}$ admits a special grading. ###### Proof. Let’s consider the following assertion $({\cal A})$ ${\cal S}(\mathfrak{n})\neq\emptyset$. The implication $(ii)\Rightarrow(i)$ is tautological. Together with the Lemmas 6(i) and 9(i), the following implications are already proved $(ii)\Rightarrow(i)\Rightarrow({\cal A})\Leftrightarrow(iii)$. Therefore, it is enough to prove that $({\cal A})\Rightarrow(ii)$. Step 1. Definition of some $h\in{\bf G}(\mathbb{Q})$. Let’s assume that ${\cal S}(\mathfrak{n})\neq\emptyset$, and let $f\in{\cal S}(\mathfrak{n})$. Since the semi-simple part of $f$ is also in ${\bf G}(\mathbb{Q})$, it can be assumed that $f$ is semi-simple. Let ${\bf K}\subset{\bf G}$ be the Zariski closure of the subgroup generated by $f$ and set ${\bf H}={\bf K}^{0}$. Let $\Lambda$ be the ${\cal O}(f)$-module generated by $\Gamma$. By Lemma 17, $\Gamma$ is a coset union. Therefore $\Lambda$ is a lattice and $\Gamma$ is an union of $n\Lambda$-coset for some positive integer $n$. Let $X({\bf H})$ be the group of characters of ${\bf H}$, let $K$ be the splitting field of ${\bf H}$, let ${\cal O}$ be the ring of integers of $K$, let ${\cal P}$ be the set of prime ideals of ${\cal O}$ and let $S$ be set set of all $\pi\in{\cal P}$ dividing $n$. By Theorem 1, there exists $h\in{\bf H}(\mathbb{Q})$ such that, for any non- trivial $\chi\in X$ we have (i) $\chi(h)$ is not an algebraic integer, and (ii) $\chi(h)\equiv 1\,\mathrm{mod}\,n{\cal O}_{\pi}$ for any $\pi\in S$. Step 2. Let $\Gamma^{\prime}=\Gamma\cap h^{-1}(\Gamma)$. We claim that the virtual morphism $(\Gamma^{\prime},h)$ is a good self-similar datum. Since ${\bf K}\subset{\bf G}$ is the Zariski closure of the subgroup generated by $f$, we have $\mathbb{Q}[h]\subset\mathbb{Q}[f]$ and therefore $\Lambda$ is a ${\cal O}(h)$-lattice. It follows from Lemma 23 that the virtual endomorphism $(\Gamma^{\prime},h)$ is good. Moreover, let $\Omega_{0}$ be the set of weights of ${\bf H}$ over $\mathfrak{z}^{\overline{\mathbb{Q}}}$. There is an integer $l$ such that $f^{l}\in{\bf K}^{0}={\bf H}$. The spectrum of $f^{l}$ on $\mathfrak{z}^{\overline{\mathbb{Q}}}$ are the numbers $\chi(f^{l})$ when $\chi$ runs over $\Omega_{0}$. Thus it follows that $\Omega_{0}$ does not contain the trivial character, hence $h$ belongs to ${\cal S}(\mathfrak{n})$. Therefore by Lemma 20, the virtual endomorphism $(\Gamma^{\prime},h)$ is a good self-similar datum. Thus by Lemma 3, $\Gamma$ is a densely self-similar group. ∎ ###### Theorem 3. The following assertions are equivalent (i) The group $\Gamma$ is freely self-similar, (ii) the group $\Gamma$ is freely densely self-similar, and (iii) the Lie algebra $\mathfrak{n}^{\mathbb{C}}$ admits a very special grading. ###### Proof. Let’s assume Assertion (i). Let’s consider a free self-similar action of $\Gamma$ on some $A^{\omega}$ and let $A^{\prime}$ be any $\Gamma$-orbit in $A$. Then the action of $\Gamma$ on $A^{\prime\omega}$ is free transitive self-similar, thus $\Gamma$ is freely transitive self-similar. The rest of the proof is identical to the previous proof, except that 1) the assertion $({\cal A})$ is replaced by $({\cal A^{\prime}})$: ${\cal V}(\mathfrak{n})\neq\emptyset$, 2) the Lemmas 6(ii) and 9(ii) are used instead of Lemmas 6(i) and 9(i) in order to prove that $(ii)\Rightarrow(i)\Rightarrow({\cal A^{\prime}})\Leftrightarrow(iii)$, 3) the proof that ${\cal A^{\prime}}\Rightarrow(ii)$ uses the weights of ${\bf H}$ and the eigenvalues of $f$ on $\mathfrak{n}$ instead of $\mathfrak{z}$. ∎ 7.2 Manning’s Theorem Let $N$ be a CSC nilpotent Lie group $N$ and let $\Gamma$ be a cocompact lattice. The manifold $M=N/\Gamma$ is called a nilmanifold. A diffeomorphism $f:M\to M$ is called an Anosov diffeomorphism if (i) there is a continuous splitting of the tangent bundle $TM$ as $TM=E_{u}\oplus E_{s}$ which is invariant by $df$, and (ii) there is a Riemannian metric relative to which $df|_{E_{s}}$ and $df^{-1}|_{E_{u}}$ are contracting. For any $x\in M$, $f$ induces a group automorphism $f_{*}$ of $\Gamma\simeq\pi_{1}(M)$. By Lemma 19, $f_{*}$ extends to an isomorphism $\tilde{f}_{*}:\mathfrak{n}^{\mathbb{R}}\to\mathfrak{n}^{\mathbb{R}}$, where $\mathfrak{n}^{\mathbb{R}}$ is the Lie algebra of $N$. Strictly speaking, $\tilde{f}_{*}$ is only defined up to an inner automorphism. Since $f_{*}$ is well defined modulo the unipotent radical of $\mathrm{Aut}\,\mathfrak{n}^{\mathbb{R}}$, the set ${\mathrm{Spec}}\,f_{*}$ is unambiguously defined. ###### Manning’s Theorem. The set ${\mathrm{Spec}}\,f_{*}$ contains no root of unity. See [18]. Later on, A. Manning proved a much stronger result. Namely ${\mathrm{Spec}}\,f_{*}$ contains no eigenvalues of absolute value $1$, and $f$ is topologically conjugated to an Anosov automorphism, see [19]. 7.3 A Corollary for nilmanifolds with an Anosov diffeomorphim ###### Corollary 4. Let $M$ be a nilmanifold endowed with an Anosov diffeomorphism. Then $\pi_{1}(M)$ is freely densely self-similar. ###### Proof. By definition, we have $M=N/\Gamma$, where $N$ is a CSC nilpotent Lie group and $\Gamma\simeq\pi_{1}(M)$ is a cocompact lattice. Set $\mathfrak{n}^{\mathbb{R}}=\mathrm{Lie}\,N$ and $\mathfrak{n}^{\mathbb{C}}=\mathbb{C}\otimes\mathfrak{n}^{\mathbb{R}}$. By Manning’s Theorem and Lemma 8, $\mathfrak{n}^{\mathbb{C}}$ has a very special grading. Therefore $\Gamma$ is freely densely self-similar by Theorem 3. ∎ 7.4 Characterisation of fractal FGTF nilpotent groups For completeness purpose, we will now investigate the non-negative gradings of $\mathfrak{n}^{C}$. Unlike Theorems 2 and 3, the proof of Propositions 5 and 6 are quite obvious. Let $\mathfrak{n}^{\mathbb{Q}}$ be a finite dimensional nilpotent Lie algebra and let $\Gamma$ be a multiplicative lattice in $\mathfrak{n}$. Set $\mathfrak{n}^{\mathbb{C}}=\mathbb{C}\otimes\mathfrak{n}^{\mathbb{Q}}$. ###### Proposition 5. The following assertions are equivalent (i) The group $\Gamma$ is fractal (ii) $\mathfrak{n}^{\mathbb{C}}$ admits a non-negative special grading. (iii) $\mathfrak{n}^{\mathbb{Q}}$ admits a non-negative special grading. ###### Proof. It follows from Lemma 10 that Assertions (ii) and (iii) are equivalent. Proof that (i) $\Rightarrow$ (ii). By assumption, there is a fractal datum $(\Gamma^{\prime},f)$. Let $g:\Gamma\rightarrow\Gamma^{\prime}$ be the inverse of $f$ and let $\tilde{g}\in\mathrm{Aut}\,\mathfrak{n}$ be its unique extension. Let $\Lambda\subset\mathfrak{n}$ be the additive subgroup generated by $\Gamma$. By Lemmas 17, $\Lambda$ is an additive lattice. Since we have $\tilde{g}(\Lambda)\subset\Lambda$, it follows that all eigenvalues of $\tilde{g}$ are algebraic integers. Moreover $(\Gamma^{\prime},g^{-1})$ is a self-similar datum, thus ${\mathrm{Spec}}\,\,\tilde{g}^{-1}|_{\mathfrak{z}}$ contains no root of unity. Therefore, by Lemma 10, Assertion (ii) holds. Proof that (iii) $\Rightarrow$ (i). Let’s assume Assertion (iii) and let $\mathfrak{n}^{\mathbb{Q}}=\oplus_{k\geq 0}\,\mathfrak{n}_{k}^{\mathbb{Q}}$ be a non-negative special grading of $\mathfrak{n}^{\mathbb{Q}}$. By Lemma 17, $\Gamma$ lies in a lattice $\Lambda$. Since it is possible to enlarge $\Lambda$, we can assume that $\Lambda=\oplus_{k\geq 0}\,\Lambda_{k}$, where $\Lambda_{k}=\Lambda\cap\mathfrak{n}_{k}^{\mathbb{Q}}$. Since $\Gamma$ is a coset union, there is an integer $d\geq 1$ such that $\Gamma$ is an union of $d\Lambda$-cosets. Let $g$ be the automorphism of $\mathfrak{n}^{\mathbb{Q}}$ defined by $g(x)=(d+1)^{k}\,x$ if $x\in\mathfrak{n}_{k}^{\mathbb{Q}}$. We claim that $g(\Gamma)\subset\Gamma$. Let $x\in\Gamma$ and let $x=\sum_{k\geq 0}\,x_{k}$ be its decomposition into homogenous components. We have $g(x)=x+\sum_{k\geq 1}\,((d+1)^{k}-1)x_{k}$. By hypothesis each homogenous component $x_{k}$ belongs to $\Lambda$. Since $(d+1)^{k}-1$ is divisible by $d$, we have $g(x)\in x+d\Lambda\subset\Gamma$ and the claim is proved. Set $\Gamma^{\prime}=g(\Gamma)$ and let $f:\Gamma^{\prime}\rightarrow\Gamma$ be the inverse of $g$. It is clear that $(\Gamma^{\prime},f)$ is a fractal datum for $\Gamma$, what proves Assertion (i). ∎ ###### Proposition 6. The following assertions are equivalent (i) The group $\Gamma$ is freely fractal (ii) $\mathfrak{n}^{\mathbb{C}}$ admits a positive grading. (iii) $\mathfrak{n}^{\mathbb{Q}}$ admits a positive grading. Since the proof is strictly identical, it will be skipped. ## 8 Not self-similar FGTF nilpotent groups and affine nilmanifolds This section provides an example of a FGTF nilpotent group which is not even self-similar, see subsection 8.6. The end of the section is about the Milnor- Scheuneman conjecture. 8.1 FGTF nilpotent groups with rank one center Let $\Gamma$ be a FGTF nilpotent group and let $Z(\Gamma)$ be its center. ###### Lemma 24. Let’s asssume that $\Gamma$ is self-similar and $Z(\Gamma)\simeq\mathbb{Z}$. Then $\Gamma$ is transitive self-similar. ###### Proof. Assume that $\Gamma$ admits a faithful self-similar action on some $A^{\omega}$, where $A$ is a finite alphabet. Let $a_{1},\dots,a_{k}$ be a set of representatives of $A/\Gamma$, where $k$ is the number of $\Gamma$-orbits on $A$. For each $1\leq i\leq k$, let $\Gamma_{i}$ be the stabilizer of $a_{i}$. For any $h\in\Gamma_{i}$, there is $h_{i}\in\Gamma$ such that $h(a_{i}w)=a_{i}h_{i}(w)$, for all $w\in A^{\omega}$. Since the action is faithfull $h_{i}$ is uniquely determined and the map $f_{i}:\Gamma_{i}\to\Gamma,h\mapsto h_{i}$ is a group morphism. Let $\mathfrak{n}^{\mathbb{Q}}$ be the Malcev Lie algebra of $\Gamma$, and let $\mathfrak{z}$ be its center, and let $z\neq 0$ be a generator of $\cap_{i}\,Z(\Gamma_{i})$. By Lemma 19, the group morphism $f_{i}$ extends to a Lie algebra morphism $\tilde{f}_{i}:\mathfrak{n}^{\mathbb{Q}}\to\mathfrak{n}^{\mathbb{Q}}$. Since $\mathfrak{z}=\mathbb{Q}\otimes Z(\Gamma)$ is one dimensional, it follows that either $\tilde{f}_{i}$ is an isomorphism or $\tilde{f}_{i}(\mathfrak{z})=0$. In any case, we have $\tilde{f}_{i}(z)=x_{i}z$, for some $x_{i}\in\mathbb{Q}$. However $\mathbb{Z}z$ is not invariant by all $\tilde{f}_{i}$, otherwise it would be in the kernel of the action. It follows that at least one $x_{i}$ is not an integer. For such an index $i$, the $f_{i}$-core of $\Gamma_{i}$ is trivial, and the virtual morphism $(\Gamma_{i},f_{i})$ is a self-similar datum for $\Gamma$. Thus $\Gamma$ is transitive self-similar. ∎ 8.2 Small representations Let $N$ be a CSC nilpotent Lie group with Lie algebra $\mathfrak{n}^{\mathbb{R}}$ and let $\Gamma$ be a cocompact lattice. ###### Lemma 25. If $\Gamma$ is transitive self-similar, then there exists a faithfull $\mathfrak{n}^{\mathbb{R}}$-module of dimension $1+\dim\mathfrak{n}^{\mathbb{R}}$. ###### Proof. By hypothesis, $\Gamma$ is transitive self-similar. By Theorem 2, $\mathfrak{z}^{\mathbb{C}}$ admits a special grading $\mathfrak{n}^{\mathbb{C}}=\oplus_{n\in\mathbb{Z}}\,\mathfrak{n}^{\mathbb{C}}_{n}$. Let $\delta:\mathfrak{n}^{\mathbb{C}}\to\mathfrak{n}^{\mathbb{C}}$ be the derivation defined by $\delta(x)=nx$ if $x\in\mathfrak{n}_{n}$. Since $\delta|\mathfrak{z}^{\mathbb{C}}$ is injective, it follows that there is some $\partial\in\mathrm{Der}\,\mathfrak{n}^{\mathbb{R}}$ such that $\partial|\mathfrak{z}^{\mathbb{R}}$ is injective. Set $\mathfrak{m}^{\mathbb{R}}=\mathbb{R}\partial\ltimes\mathfrak{n}^{\mathbb{R}}$. Relative to the adjoint action, $\mathfrak{m}^{\mathbb{R}}$ is a faithfull $\mathfrak{z}^{\mathbb{R}}$-module. Therefore $\mathfrak{m}^{\mathbb{R}}$ is a faithfull $\mathfrak{n}^{\mathbb{R}}$-module with the prescribed dimension. ∎ 8.3 Filiform nilpotent Lie algebras Let $\mathfrak{n}$ be a nilpotent Lie algebra over $\mathbb{Q}$. Let $C^{n}\mathfrak{n}$ be the decreasing central series, which is inductively defined by $C^{1}\mathfrak{n}=\mathfrak{n}$ and $C^{n+1}\mathfrak{n}=[\mathfrak{n},C^{n}\mathfrak{n}]$. The nilpotent Lie algebra $\mathfrak{n}$ is called filiform if $\dim C^{1}\mathfrak{n}/C^{2}\mathfrak{n}=2$ and $\dim C^{k}\mathfrak{n}/C^{k+1}\mathfrak{n}\leq 1$ for any $k>1$. Set $n=\dim\mathfrak{n}$. It follows from the definition that $\dim C^{k}\mathfrak{n}/C^{k+1}\mathfrak{n}=1$ for any $0<k\leq n-1$ and $C^{k}\mathfrak{n}=0$ for any $k\geq n$. ###### Lemma 26. Let $\mathfrak{n}$ be a filiform nilpotent Lie algebra over $\mathbb{Q}$, with $\dim\mathfrak{n}\geq 3$. Then its center $\mathfrak{z}$ has dimension one. ###### Proof. Let $z\in\mathfrak{n}$ be nonzero. Let $k$ be the integer such that $z\in C^{k}\mathfrak{n}\setminus C^{k+1}\mathfrak{n}$. Since $C^{k}\mathfrak{n}=C^{k+1}\mathfrak{n}\oplus\mathbb{Q}z$ $C^{k+1}\mathfrak{n}=[\mathfrak{n},C^{k}\mathfrak{n}]=[\mathfrak{n},C^{k+1}\mathfrak{n}]+[\mathfrak{n},z]=C^{k+2}\mathfrak{n}$. It follows that $C^{k+1}\mathfrak{n}=0$. Therefore $\mathfrak{z}$ lies in $C^{k}\mathfrak{n}$, which is a one dimensional ideal. ∎ 8.4 Benoist Theorem ###### Benoist’s Theorem. There is a nilpotent Lie algebra $\mathfrak{n}_{B}^{\mathbb{R}}$ of dimension $11$ over $\mathbb{R}$, with the following properties (i) The Lie algebra $\mathfrak{n}_{B}^{\mathbb{R}}$ has no faithfull representations of dimension $12$, (ii) the Lie algebra $\mathfrak{n}_{B}^{\mathbb{R}}$ is defined over $\mathbb{Q}$, and (iii) the Lie algebra $\mathfrak{n}_{B}^{\mathbb{R}}$ is filiform. The three assertions appear in different places of [3]. Indeed Assertion (i), which is explicitely stated in Theorem 2 of [3], hold for a one-parameter family of eleven dimensional Lie algebras, which are denoted $\mathfrak{a}_{-2,1,t}$ in section 2.1 of [3]. These Lie algebras are filiform by Lemma 4.2.2 of [3]. Moreover, when $t$ is rational, $\mathfrak{a}_{-2,1,t}$ is defined over $\mathbb{Q}$. Therefore the Benoist Theorem holds for the Lie algebras $\mathfrak{n}_{B}=\mathfrak{a}_{-2,1,t}$ where $t$ is any rational number. 8.5 A FGTF group which is not self-similar Let $N_{B}$ the CSC nilpotent Lie group with Lie algebra $\mathfrak{n}_{B}^{\mathbb{R}}$. Since $\mathfrak{n}_{B}^{\mathbb{R}}$ is defined over $\mathbb{Q}$, $N_{B}$ contains some cocompact lattice. ###### Corollary 7. Let $\Gamma$ be any cocompact lattice in $N_{B}$. Then $\Gamma$ is not self- similar. ###### Proof. Let’s assume otherwise. By Benoist Theorem and Lemma 26, the center of $\mathfrak{n}_{B}^{\mathbb{R}}$ is one dimensional. Thus the center of $\Gamma$ has rank one, and by Lemma 24, $\Gamma$ is transitive self-similar. By Lemma 25, $\mathfrak{n}_{B}^{\mathbb{R}}$ admits a faithfull representation of dimension 12, which contradicts Benoist Theorem. Therefore $\Gamma$ is not self-similar. ∎ 8.6 On the Scheuneman-Milnor conjecture A smooth manifold $M$ is called affine if it admits a torsion-free and flat connection. Scheuneman [28] and Milnor [23] asked the following question is any nilmanifold $M$ affine? The story of the Scheuneman-Milnor conjecture is quite interesting. For many years, there are been a succession of proofs followed by refutations, but there was no doubts that the conjecture should be ultimalely proved… until a counterexample has been found by Benoist [3]. Indeed it is an easy corollary of his previously mentionned Theorem. The following question is a refinement of the previous conjecture if $\pi_{1}(M)$ is densely self-similar, is the nilmanifold $M$ affine? A positive result in that direction is ###### Corollary 8. Let $M$ be a nilmanifold. If $\pi_{1}(M)$ is freely self-similar, then $M$ is affine complete. ###### Proof. Set $M=N/\Gamma$, where $N$ is a CSC nilpotent Lie group and $\Gamma$ is a cocompact lattice. Let $\mathfrak{n}^{\mathbb{R}}$ be the Lie algebra of $N$. By Theorem 3, $\mathbb{C}\otimes\mathfrak{n}^{\mathbb{R}}$ admits a very special grading, what implies that a generic derivation is injective. Therefore there is a derivation $\delta$ of $\mathfrak{n}^{\mathbb{R}}$ which is injective. Set $\mathfrak{m}^{\mathbb{R}}=\mathbb{R}\delta\ltimes\mathfrak{n}^{R}$. Then $N$ is equivariantly diffeomorphic to the affine space $\delta+\mathfrak{n}^{\mathbb{R}}\subset\mathfrak{m}^{\mathbb{R}}$. Therefore $M$ is affine complete. ∎ ## 9 Absolute Complexities For the whole chapter, $N$ will be a CSC nilpotent Lie groups, with Lie algebra $\mathfrak{n}^{\mathbb{R}}$. Let’s assumethat that $N$ contains some cocompact lattices. Under the condition of Theorem 2 or 3, any cocompact lattice $\Gamma$ in $N$ admits a transitive or free self-similar action on some $A^{\omega}$. In this section, we try to determine the minimal degree of these actions. 9.1 Three type of absolute complexities The complexity of a cocompact lattice $\Gamma\subset N$, denoted by $\mathrm{cp}\,\Gamma$, is the smallest degree of a faithfull transitive self- similar action of $\Gamma$ on some $A^{\omega}$, with the convention that $\mathrm{cp}\,\Gamma=\infty$ if $\Gamma$ is not transitive self-similar. Similarly, the free complexity of $\Gamma$, denoted by $\mathrm{fcp}\,\Gamma$, is the smallest degree of a free self-similar action of $\Gamma$. Two cocompact lattices are called commensurable if they share a commoun subgroup of finite index. The complexity and the free complexity of a commensurable class $\xi$ are the integers $\mathrm{cp}\,\xi=\mathrm{Min}_{\Gamma\in\xi}\,\mathrm{cp}\,\Gamma$, and $\mathrm{fcp}\,\xi=\mathrm{Min}_{\Gamma\in\xi}\,\mathrm{fcp}\,\Gamma$. Then, the complexity of the nilpotent group $N$ is $\mathrm{cp}\,N=\mathrm{Max}_{\xi}\,\mathrm{cp}\,\xi$, where $\xi$ runs over all commensurable classes in $N$. In what follows, we will provide a formula for the complexity of commensurable classes. The question under which condition $\mathrm{cp}N<\infty$? is not solved, but it is a deep question. In chapter 10, a class of CSC nilpotent Lie groups of infinite complexity is investigated. 9.2 Theorem 9 Let $\xi$ be a commensurable class of cocompact lattices in $N$, and let $\Gamma\in\xi$. The Malcev Lie algebra $\Gamma$ is a $\mathbb{Q}$-form of the Lie algebra $\mathfrak{n}^{\mathbb{R}}$. Since it depends only on $\xi$, it will be denoted by $\mathfrak{n}(\xi)$. ###### Theorem 9. We have $\mathrm{cp}\,\xi=\mathrm{Min}_{h\in{\cal S}(\mathfrak{n}(\xi))}\,{\mathrm{ht}}(h)$, and $\mathrm{fcp}\,\xi=\mathrm{Min}_{h\in{\cal V}(\mathfrak{n}(\xi))}\,{\mathrm{ht}}(h)$. ###### Proof. Let $h\in{\cal S}(\mathfrak{n}(\xi))$ be an isomorphism of minimal height. In order to show that $\mathrm{cp}\,\xi={\mathrm{ht}}(h)$, we can assume that $h$ is semi-simple, by Lemma 13. Further, let $\Gamma$ be any cocompact lattice in $\xi$. By Lemma 20, we have $\mathrm{cp}\,\Gamma=\mathrm{Min}_{f\in{\cal S}(\mathfrak{n}(\xi))}\,\mathrm{cp}_{f}\,\Gamma$. By lemma 21, we have $\mathrm{cp}_{f}\,\Gamma\geq{\mathrm{ht}}(f)$, therefore we have $\mathrm{cp}\,\Gamma\geq{\mathrm{ht}}(h)$. In particular $\mathrm{cp}\,\xi\geq{\mathrm{ht}}(h)$. By Lemma 22, $\Gamma$ contains a finite index subgroup $\tilde{\Gamma}$ which is minimal relative to $h$. Since $\mathrm{cp}_{h}\,\tilde{\Gamma}={\mathrm{ht}}(h)$, it follows that $\mathrm{cp}\,\xi\leq{\mathrm{ht}}(h)$. Therefore $\mathrm{cp}\,\xi={\mathrm{ht}}(h)$ and the first assertion is proved. For the second assertion, let’s notice that an free action of minimal degree is automatically transitive, see the proof of Theorem 3. Then the rest of the proof is strictly identical to the previous proof. ∎ 9.3 Classification of lattices in a CSC nilpotent Lie groups Obviously Malcev’s Theorem implies the following ###### Malcev’s Corollary. The map $\xi\mapsto\mathfrak{n}(\xi)$ establishes a bijection between the commensurable classes of lattices and the $\mathbb{Q}$-forms of the Lie algebra $\mathfrak{n}^{\mathbb{R}}$. For the next chapter, it is interesting to translate this into the framework of non-abelian Galois cohomology. Somehow, it is more concrete, since the non- abelian Galois cohomology classifies $\mathbb{Q}$-forms of classical objects. Set ${\bf G}=\mathrm{Aut}\mathfrak{n}^{\mathbb{C}}$, let ${\bf U}$ be its unipotent radical and set $\overline{\bf G}={\bf G}/{\bf U}$. From now on, fix once for all a commensurable class $\xi_{0}$ of cocompact lattices. Then $\mathfrak{n}(\xi_{0})$ is a $\mathbb{Q}$-form of $\mathfrak{n}^{\mathbb{C}}$, what provides a $\mathbb{Q}$-form of the algebraic groups ${\bf G}$ and $\overline{\bf G}$. It induces an action of ${\mathrm{Gal}}(\mathbb{Q})$ over $\overline{\bf G}(\overline{\mathbb{Q}})$. Set ${\overline{\mathbb{Q}}}_{re}=\overline{\mathbb{Q}}\cap\mathbb{R}$ and let $\overline{\pi}:H^{1}({\mathrm{Gal}}(\mathbb{Q}),{\bf\overline{G}}(\overline{\mathbb{Q}}))\rightarrow H^{1}({\mathrm{Gal}}({\overline{\mathbb{Q}}}_{re}),{\bf\overline{G}}(\overline{\mathbb{Q}}))$ be the natural map. Recall that these two non-abelian cohomologies are pointed sets, where the distinguished point $*$ comes from the given $\mathbb{Q}$-form and the induced ${\overline{\mathbb{Q}}}_{re}$-form. Denote by $\mathrm{Ker}\,\overline{\pi}$ the kernel of $\overline{\pi}$, i.e. the fiber ${\overline{\pi}}^{-1}(*)$ of the distinguished point. Let ${\cal L}(N)$ be the set of all commensurable classes classes of lattices of $N$, up to conjugacy. ###### Corollary 10. There is a natural identification ${\cal L}(N)\simeq\,\mathrm{Ker}\,\overline{\pi}$. ###### Proof. For any field $K\subset\mathbb{C}$, set $\mathfrak{n}^{K}=K\otimes\mathfrak{n}(\xi_{0})$. For any two fields $K\subset L\subset\mathbb{C}$, let ${\cal F}(L/K)$ be the set of $K$-forms of $\mathfrak{n}^{L}$, up to conjugacy. Then ${\cal F}(L/K)$ is a pointed set, whose distinguished point is the $K$-form $\mathfrak{n}^{K}$. By the Lefschetz principle, the $\mathbb{Q}$-forms of $\mathfrak{n}^{\mathbb{C}}$ (up to conjugacy) are in bijection with the $\mathbb{Q}$-forms of $\mathfrak{n}^{\overline{\mathbb{Q}}}$. Similarly by the Tarski-Seidenberg principle the real forms (up to conjugacy) of $\mathfrak{n}^{\mathbb{C}}$ are in bijection with the ${\overline{\mathbb{Q}}}_{re}$-forms of $\mathfrak{n}^{\overline{\mathbb{Q}}}$. So we have ${\cal F}(\mathbb{C}/\mathbb{Q})\simeq{\cal F}(\overline{\mathbb{Q}}/\mathbb{Q})$ and ${\cal F}(\mathbb{C}/\mathbb{R})\simeq{\cal F}(\overline{\mathbb{Q}}/{\overline{\mathbb{Q}}}_{re})$ Since a Lie algebra is a vector space endowed with a tensor (its Lie bracket), it follows from [29], III-2, Proposition 1 that ${\cal F}(\overline{\mathbb{Q}}/\mathbb{Q})=H^{1}({\mathrm{Gal}}(\mathbb{Q}),{\bf G}(\overline{\mathbb{Q}}))$, and ${\cal F}(\overline{\mathbb{Q}}/{\overline{\mathbb{Q}}}_{re})=H^{1}({\mathrm{Gal}}({\overline{\mathbb{Q}}}_{re}),{\bf G}(\overline{\mathbb{Q}}))$. Moreover since ${\bf U}$ is unipotent, we have $H^{1}({\mathrm{Gal}}(\mathbb{Q}),{\bf G}(\overline{\mathbb{Q}}))\simeq H^{1}({\mathrm{Gal}}(\mathbb{Q}),{\bf\overline{G}}(\overline{\mathbb{Q}}))$, and $H^{1}({\mathrm{Gal}}(\mathbb{Q}_{re}),{\bf G}(\overline{\mathbb{Q}}))\simeq H^{1}({\mathrm{Gal}}(\mathbb{Q}_{re}),{\bf\overline{G}}(\overline{\mathbb{Q}}))$. There is a commutative diagram of pointed sets ${\cal F}(\mathbb{C}/\mathbb{Q})$ | $\overset{\theta}{\longrightarrow}$ | ${\cal F}(\mathbb{C}/\mathbb{R})$ ---|---|--- $\downarrow$ | | $\downarrow$ ${\cal F}(\overline{\mathbb{Q}}/\mathbb{Q})$ | $\overset{\theta^{\prime}}{\longrightarrow}$ | ${\cal F}(\overline{\mathbb{Q}}/{\overline{\mathbb{Q}}}_{re})$ $\downarrow$ | | $\downarrow$ $H^{1}({\mathrm{Gal}}(\mathbb{Q}),{\bf G}(\overline{\mathbb{Q}}))$ | $\overset{\pi}{\longrightarrow}$ | $H^{1}({\mathrm{Gal}}({\overline{\mathbb{Q}}}_{re}),{\bf G}(\overline{\mathbb{Q}}))$ $\downarrow$ | | $\downarrow$ $H^{1}({\mathrm{Gal}}(\mathbb{Q}),{\bf\overline{G}}(\overline{\mathbb{Q}}))$ | $\overset{\overline{\pi}}{\longrightarrow}$ | $H^{1}({\mathrm{Gal}}({\overline{\mathbb{Q}}}_{re}),{\bf\overline{G}}(\overline{\mathbb{Q}}))$ where $\theta$ is the map $\mathbb{R}\otimes_{\mathbb{Q}}\,-$ , $\theta^{\prime}$ is the map ${\overline{\mathbb{Q}}}_{re}\otimes_{\mathbb{Q}}\,-$, $\pi$ and $\overline{\pi}$ are restrictions maps. It is tautological that ${\cal L}(N)=\mathrm{Ker}\,\theta$. Since all vertical maps are bijective, it follows that ${\cal L}(N)$ is isomorphic to $\mathrm{Ker}\,\overline{\pi}$. ∎ ## 10 Some Nilpotent Lie groups of infinite complexity. This chapter is devoted to the analysis to a class of CSC nilpotent Lie groups ${\cal N}$, for which the classification of commensurable classes and the computation of their complexity are very explicitely connected with the arithmetic of complex quadratic fields. For $K=\mathbb{R}$ or $\mathbb{C}$, let $O(2,K)$ be the group of linear automorphisms of $\mathbb{R}^{2}$ preserving the quadratic form $x^{2}+y^{2}$. Let ${\cal L}$ be the class of nilpotent Lie algebras $\mathfrak{n}^{R}$ over $\mathbb{R}$ satisfying the following properties (i) $\mathfrak{n}^{R}$ has a $\mathbb{Q}$-form (ii) $\mathfrak{n}^{R}/[\mathfrak{n}^{R},\mathfrak{n}^{R}]\simeq\mathbb{R}^{2}$ has dimension two (iii) the Lie algebra $\mathfrak{n}^{\mathbb{C}}:=\mathbb{C}\otimes\mathfrak{n}^{\mathbb{R}}$ has a special grading (iv) for $K=\mathbb{R}$ or $\mathbb{C}$, the image of $\mathrm{Aut}\,\mathfrak{n}^{K}$ in $GL(\mathfrak{n}^{K}/[\mathfrak{n}^{K},\mathfrak{n}^{K}])$ is $O(2,K)$. Let be the class of CSC nilpotent Lie groups $N$ whose Lie algebra $\mathfrak{n}^{\mathbb{R}}$ is in ${\cal L}$. It should be noted that the class ${\cal N}$ is not empty. There is one Lie group $N_{112}\in{\cal N}$ of dimension $112$, see [22]. Indeed [22] contains a general method to find nilpotent Lie algebras with a prescribed group of automorphisms, modulo its unipotent radical. For the group $O(2,\mathbb{R})$, $N_{112}$ is the Lie group of minimal dimension obtained with this method. However it is difficult to provide more details, without going to very long explanations. From now on, $N$ will be any Lie group in class ${\cal N}$, $\xi_{0}$ will be one commensurable class of lattices in $N$ and $\mathfrak{n}:=\mathfrak{n}(\xi_{0})$ will be the corresponding corresponding $\mathbb{Q}$ form of $\mathfrak{n}^{R}$. As before, set $\mathfrak{n}^{K}=K\otimes\mathfrak{n}$ for any field $K\subset\mathbb{C}$. Let ${\bf G}={\bf Aut}\,\mathfrak{n}$ the algebraic automorphism group of $\mathfrak{n}$, let ${\bf U}$ be its unipotent radical and set ${\bf\overline{G}}={\bf G}/{\bf U}$. By hypothesis, ${\bf\overline{G}}$ is the algebraic group $O(2)$. 10.1 The $\mathbb{Z}$-grading of $\mathfrak{n}^{\mathbb{C}}$ Since ${\bf\overline{G}}(\mathbb{C})=O(2,\mathbb{C})$, a maximal torus ${\bf H}$ of ${\bf G}(\mathbb{C})$ has dimension $1$. Therefore $\mathfrak{n}^{\mathbb{C}}$ has a $\mathbb{Z}$-grading $\mathfrak{n}^{\mathbb{C}}=\oplus_{k\in\mathbb{Z}}\,\mathfrak{n}^{\mathbb{C}}_{k}$, satisfying the following properties (i) the grading is essentially unique, namely any other grading is a multiple of the given grading, (ii) $\dim\mathfrak{n}^{\mathbb{C}}_{k}=\dim\mathfrak{n}^{\mathbb{C}}_{-k}$ for any $k$. In particular $\mathfrak{n}^{\mathbb{C}}$ does not admit a (non- trivial) non-negative grading, and (iii) the grading is not defined over $\mathbb{R}$. Indeed since ${\bf\overline{G}}(\mathbb{C})=O(2,\mathbb{C})$, the normalizer ${\bf K}(\mathbb{C})$ of ${\bf H}(\mathbb{C})$ has two connected components, and any $\sigma\in{\bf K}(\mathbb{C})\setminus{\bf K}(\mathbb{C})^{0}$ exchanges $\mathfrak{n}^{\mathbb{C}}_{k}$ and $\mathfrak{n}^{\mathbb{C}}_{-k}$, what shows Assertion (ii). Since ${\bf\overline{G}}(\mathbb{R})=O(2,\mathbb{R})$, no torus of ${\bf G}(\mathbb{R})$ is split, what implies Assertion (iii). Moreover, the grading is not very special, so $\mathrm{fcp}(\xi)=\infty$ for any commensurable class $\xi$. For the forthcoming computation of $\mathrm{cp}(\xi)$, the following quantity will be involved $e(N)=\sum_{k>0}\,k\dim\mathfrak{n}^{\mathbb{C}}_{k}$. For example, for the Lie group $N_{112}$ of [22], we have $e(N_{112})=126$. 10.2 Classification of commensurable lattices in $N$ ###### Lemma 27. Let $N\in{\cal N}$. Up to conjugacy, there is a bijection between (i) the commensurable class of cocompact lattices in $N$, and (ii) the positive definite quadratic form on $\mathbb{Q}^{2}$. ###### Proof. Let $q_{0}$ be a given definite quadratic form on $\mathbb{Q}^{2}$. It determines a $\mathbb{Q}$-form of the algebraic group $O(2)$, and $H^{1}({\mathrm{Gal}}(\mathbb{Q}),O(2,\overline{\mathbb{Q}}))$ classifies the quadratic forms on $\mathbb{Q}^{2}$, while the kernel of $H^{1}({\mathrm{Gal}}(\mathbb{Q}),O(2,\overline{\mathbb{Q}}))\to H^{1}({\mathrm{Gal}}(\overline{\mathbb{Q}}_{re}),O(2,\overline{\mathbb{Q}}))$ classifies the positive definite quadratic forms on $\mathbb{Q}^{2}$. Thus the lemma follows from Corollary 10. ∎ The classification of positive definite quadratic forms $q$ on $\mathbb{Q}^{2}$ is well known. Up to conjugacy, $q$ can be written as $q(x,y)=ax^{2}+ady^{2}$, where $a,\,d$ are positive and $d$ is a square-free integer. Then $q$ is determined by the following two invariants (i) its discriminant $-d$, viewed as an element of $\mathbb{Q}^{*}/\mathbb{Q}^{*2}$, (ii) the value $a$, viewed as an element in $\mathbb{Q}^{*}/N_{K/\mathbb{Q}}(K^{*})$, where $K=\mathbb{Q}(\sqrt{-d})$. Equivalently, this means that $q(\mathbb{Q}^{2}\setminus 0)=aN_{K/\mathbb{Q}}(K^{*})$. For any positive definite quadratic forms $q$ on $\mathbb{Q}^{2}$, let $\xi(q)$ be the corresponding commensurable class (or more precisely, the conjugacy class of the commensurable class). By Theorem 9, $\mathrm{cp}\,\xi(q)$ only depends on $O(q)$, therefore it only depends on the discriminant $-d$. 10.3 The function $F(d)$ Let $d$ be a positive square-free integer. Set $K=\mathbb{Q}(\sqrt{-d})$, let ${\cal O}$ be its ring of integers, let $R$ be te set of roots of unity in $K$ and set $K_{1}=\\{z\in K|z\overline{z}=1\\}$. For $z\in K^{*}$, recall that the integer $d(z)$ is defined by $d(z)=N_{K/\mathbb{Q}}(\pi_{z})=\mathrm{Card\,}{\cal O}/\pi_{z}$, where $\pi_{z}$ is the ideal $\pi_{z}=\\{a\in{\cal O}|az\in{\cal O}\\}$. Set $F(d)=\mathrm{Min}_{z\in K_{1}\setminus R}\,d(z)$. We will now show two formulas for $F(d)$. Indeed $F(d)$ is the norm of some specific ideal in $K=\mathbb{Q}(\sqrt{-d})$, and it is also the minimal solution of some diohantine equation. Let ${\cal J}$ be the set of all ideals $\pi$ of ${\cal O}$ such that $\pi$ and $\overline{\pi}$ are coprime and $\pi^{2}$ is principal. ###### Lemma 28. We have $F(d)=\mathrm{Min}_{\pi\in{\cal J}}\,N_{k/\mathbb{Q}}(\pi)$. In particular, we have $F(1)=5$ and $F(3)=7$. ###### Proof. The map $z\mapsto\pi_{z}$ induces a bijection $(K_{1}\setminus R)/R\simeq{\cal J}$, from which the first assertion follows. Moreover, if $\mathrm{Cl}(K)=\\{0\\}$, then $F(d)$ is the smallest split prime number. Therefore $F(1)=5$ and $F(3)=7$. ∎ Let’s consider the following diophantine equation $({\cal E})$ $4n^{2}=a^{2}+db^{2}$, with $n>0$, $a>0$ and $b\neq 0$. A solution $(n,a,b)$ of $({\cal E})$ is called primitive if $\gcd(n,a)=1$. Let ${\mathrm{Sol}}\,({\cal E})$ (respectively ${\mathrm{Sol}}\,_{prim}({\cal E})$) be the set of solutions (respectively of primitive solutions) of $({\cal E})$. Let $\pi\in{\cal J}$. Since $\pi^{2}$ is principal, there are integers $a(\pi)>0$ and $b(\pi)$ such that $a(\pi)+b(\pi)\sqrt{-d}$ is a generator of $4\pi^{2}$. Moreover, let’s assume that $d\neq 1$ or $3$. Then $R=\\{\pm 1\\}$ and the integers $a(\pi)$ and $b(\pi)$ are uniquely determined. Thus there is a map $\theta:{\cal J}\to{\mathrm{Sol}}\,({\cal E})$ defined by $\theta(\pi)=(N_{K/\mathbb{Q}}(\pi),a(\pi),b(\pi))$. ###### Lemma 29. Under the hypothesis that $d\neq 1$ or $3$, the map $\theta$ induces a bijection from ${\cal J}$ to ${\mathrm{Sol}}\,_{prim}({\cal E})$. In particular $F(d)=\mathrm{Min}_{(n,a,b)\in{\mathrm{Sol}}\,({\cal E})}\,n$. ###### Proof. Step 1: proof that $\theta({\cal J})\subset{\mathrm{Sol}}\,_{prim}({\cal E})$. An algebraic integer $z\in{\cal O}$ is called primitive if there are no integer $d>1$ such that $z/d$ is an algebraic integer. Equivalently, there are no integer $d>1$ such that $d\mid z+\overline{z}$ and $d^{2}\mid z\overline{z}$. Let $\pi\in{\cal J}$ and set $z=1/2(a(\pi)+b(\pi)\sqrt{-d})$. Since $z+{\overline{z}}=a(\pi)$ and $z.{\overline{z}}=N_{K/\mathbb{Q}}(\pi^{2})$, $z$ is an algebraic integer which is a generator of $\pi^{2}$. Since $\pi^{2}$ and $\overline{\pi}^{2}$ are coprime, $z$ is primitive. Since $z.{\overline{z}}=N_{K/\mathbb{Q}}(\pi)^{2}$, it follows that $N_{K/\mathbb{Q}}(\pi)$ and $a(\pi)$ are coprime. Hence $\theta(\pi)\in{\mathrm{Sol}}\,_{prim}({\cal E})$ and the claim is proved. Step 2: proof that $\theta({\cal J})={\mathrm{Sol}}\,_{prim}({\cal E})$. Let $(n,a,b)\in{\mathrm{Sol}}\,_{prim}({\cal E})$ and $z=1/2(a(\pi)+b(\pi)\sqrt{-d})$. Since $z\neq\overline{z}$, $z+{\overline{z}}=a$ and $z{\overline{z}}=n$, the number $z$ is an algebraic integer. Set $\tau=z{\cal O}$ and let $\tau=\pi_{1}^{m_{1}}\dots\pi_{k}^{m_{k}}$ be the factorization of $\tau$ into a product of prime ideals of ${\cal O}$, where, as usual we assume that $\pi_{i}\neq\pi_{j}$ for $i\neq j$ and all $m_{i}$ are positive. For $1\leq i\leq k$, let $p_{i}$ be the characteristic of the field ${\cal O}/\pi_{i}$. Since $n$ and $a$ are coprime, $\tau$ and $\overline{\tau}$ are coprime. It follows that $\overline{\pi_{i}}$ does not divide $\tau$. In particular $\pi_{i}\neq\overline{\pi_{i}}$ and $N_{K/\mathbb{Q}}(\pi_{i})=p_{i}$. Since $\pi_{i}$ and $\overline{\pi_{i}}$ are the only two ideals over $p_{i}$, we have $m_{i}=v_{p_{i}}(N_{K/\mathbb{Q}}(\tau))=v_{p_{i}}(n^{2})$. Since each $m_{i}$ is even, we have $\tau=\pi^{2}$ for some ideal $\pi\in{\cal J}$. Therefore $\theta(\pi)=(n,a,b)$, and the claim is proved. Step 3. It follows easily that $\theta$ is a bijection from ${\cal J}$ to ${\mathrm{Sol}}\,_{prim}({\cal E})$. In particular $F(d)=\mathrm{Min}_{(n,a,b)\in{\mathrm{Sol}}\,_{prim}({\cal E})}\,n$, from which the lemma follows. ∎ 10.4 Complexity computation ###### Theorem 11. Let $q$ be a positive definite quadratic form on $\mathbb{Q}^{2}$ of discriminant $-d$. Then we have $\mathrm{cp}\,\xi(q)=F(d)^{e(\mathfrak{n}^{\mathbb{C}})}$ ###### Proof. Step 1. Let $G\subset\mathrm{End}_{\mathbb{Q}}(K)$ be the group generated by the multiplication by elements in $K_{1}$ and by the complex conjugation. We have $G\simeq O(2)$ and $SO(2)\simeq K_{1}$. As a $O(q)$-module, there is an isomorphism $V\simeq\mathbb{Q}(\sqrt{-d})$, where $V=\mathfrak{n}(\xi(q))/[\mathfrak{n}(\xi(q)),\mathfrak{n}(\xi(q))]$. Step 2. Let ${\overline{\cal S}}(\mathfrak{n}(\xi(q))$ be the image of ${\cal S}(\mathfrak{n}(\xi(q))$ in $O(q)$. We claim that ${\overline{\cal S}}(\mathfrak{n}(\xi(q))=K_{1}\setminus R$. Indeed $O(q)$ can be identified with a Levi factor of ${\bf G}(\mathbb{Q})$ and let $\rho:O(q)\to{\bf G}(\mathbb{Q})$ a corresponding lift. Any element in $R\cup O(q)\setminus SO(q)$ has finite order, hence we have ${\overline{\cal S}}(\mathfrak{n}(\xi(q))\subset K_{1}\setminus R$. Let $z\in K_{1}\setminus R$. It is clear that $z$ is not an algebraic integer. Since the grading is special, we have $\mathfrak{z}^{\mathbb{C}}=\oplus_{k\neq 0}\mathfrak{z}^{\mathbb{C}}_{k}$. Since the eigenvalues of $\rho(z)$ on $\mathfrak{z}_{k}$ is $z^{k}$, it follows that $z$ belongs to ${\overline{\cal S}}(\mathfrak{n}(\xi(q))$, what proves the point. Step 3. Let $z\in K_{1}\setminus R$. We have $\overline{z}=\overline{z}^{-1}$. Therefore by Lemma 15 we have ${\mathrm{ht}}\,\rho(z)=\prod_{k\geq 1}d(z^{k})^{\dim\,\mathfrak{n}^{\mathbb{C}}_{k}}=d(z)^{e(N)}$. Therefore Theorem 4 implies Theorem 5. ∎ Since $F(d)\geq{\sqrt{1+d}\over 2}$, it follows that ###### Corollary 12. The group $N$ has infinite complexity. Since $F(7)=F(15)=2$ and $F(d)\geq 3$ otherwise, it follows that ###### Corollary 13. If the positive definite quadratic form $q$ has discriminant $-7$ or $-15$ we have $\mathrm{cp}\,\xi(q)=2^{e(N)}$, and $\mathrm{cp}\,\xi(q)\geq 3^{e(N)}$ otherwise. ## References * [1] A. Bansfield and D. Kan, Homotopy limits, completions and localizations. Springer Lecture Notes in Mathematics 304 (1972). * [2] L. Bartholdi and S. Sidki, Self-similar products of groups. Groups Geom. Dyn. 14 (2020) 107-115. * [3] Y. Benoist, Une nilvariete non-affine. J. Differential Geometry 41 (1995) 21-52. * [4] A. Berlatto and S. Sidki, Virtual endomorphisms of nilpotent groups. Groups Geom. Dyn. 1 (2007) 21-46. * [5] N. Bourbaki, Groupe et algèbres de Lie, ch. 2-3. Masson (1972). * [6] A. Borel Linear Algebraic Groups. Springer Graduate Text in Math. 126 (1991). * [7] C. Chevalley, Theorie des groupes de Lie, Tome III: Groupes algèbriques. Hermann (1955). * [8] Y. Cornulier, Gradings on Lie algebras, systolic growth, and cohopfian properties of nilpotent groups. Bull. de la SMF 144 (2016) 693-744. * [9] V. Futorny, D. Kochloukova and S.Sidki, On self-similar Lie algebras and virtual endomorphisms. Math. Z. 292 (2019) 1123–1156. * [10] R. I. Grigorchuk, On Burnside’s problem on periodic groups. Functional Anal. Appl. 14 (1980) 41-43. * [11] R. I. Grigorchuk, On the Milnor problem of group growth. Dokl. Akad. Nauk SSSR 271 (1983) 30-33. * [12] R. Grigorchuk, V. Nekrashevych and Z. Šunić Z, From Self-Similar Groups to Self-Similar Sets and Spectra, Fractal Geometry and Stochastics V. Progress in Probability 70 (2015) 175-207. * [13] N. D. Gupta and S. N. Sidki, On the Burnside problem for periodic groups. Math. Z. 182 (1983) 385-388. * [14] N. Gupta and S. Sidki, Extension of groups by tree automorphisms, in Contributions to group theory. Contemp. Math. 33 (1984) 232-246. * [15] M. Hall, The Theory of Groups. Macmillan compagny (1959). * [16] D. Kochloukova and S. Sidki, Self-similar groups of type FPn. Geom. Dedicata 204 (2020) 241-264. * [17] J.L. Loday and B. Valette, Algebraic Operads. Grundlehren der mathematischen Wissenschaften Book 346 (2012). * [18] A. I. Malcev, On a class of homogeneous spaces. Izvestiya Akad. Nauk. SSSR. Ser. Mat.13 (1949) 9-22. * [19] A. I. Malcev, Nilpotent groups without torsion. Izv. Akad. Nauk. SSSR, Math. 13 (1949) 201-212. * [20] A. Manning, Anosov diffeomorphisms on nilmanifolds. Proc. A.M.S. 38 (1973) 423-426. * [21] A. Manning, There are No New Anosov Diffeomorphisms on Tori. Am. J. of Math. 96 (1974) 422-429. * [22] O. Mathieu, Automorphisms Groups of nilpotent Lie algebras, and applications. In preparation. * [23] J. Milnor, On fundamental groups of complete affinely flat manifolds. Advances in Math., 25(1977) 178-187. * [24] V. Nekrashevych, Self-Similar Groups. Math. Survey and Monographs 117 (2005). * [25] V. Nekrashevych and S.Sidki, Automorphisms of the binary tree: state-closed subgroups and dynamics of 1/2-endomorphisms, in Groups: topological, combinatorial and arithmetic aspects. London Math. Soc. Lecture Note 311 (2004) 375–404. * [26] M.S. Raghunathan, Discrete subgroups in Lie groups. Springer-Verlag, Ergebnisse der Math. 68 (1972). * [27] M. Rosenlicht, Some rationality questions on algebraic groups. Annali di Matematica 43 (1957) 25-50. * [28] J. Scheuneman, Examples of locally affine spaces. Bull. Amer. Math. Soc. 77 (1971) 589-592. * [29] J. P. Serre, Cohomologie galoisienne. Springer Lecture Notes in Math. 5 (1965). * [30] S. Sidki: Regular trees and their automorphisms. Monografías de Matemática 56. Instituto de Matemática Pura e Aplicada (1998).
# Distributed Learning over Markovian Fading Channels for Stable Spectrum Access Tomer Gafni and Kobi Cohen Tomer Gafni and Kobi Cohen are with the School of Electrical and Computer Engineering, Ben-Gurion University of the Negev, Beer Sheva 8410501 Israel. Email<EMAIL_ADDRESS><EMAIL_ADDRESS>work has been submitted to the IEEE for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible.A short version of this paper was presented at the 57th Annual Allerton Conference on Communication, Control, and Computing, 2019 [1]. ###### Abstract We consider the problem of multi-user spectrum access in wireless networks. The bandwidth is divided into $K$ orthogonal channels, and $M$ users aim to access the spectrum. Each user chooses a single channel for transmission at each time slot. The state of each channel is modeled by a restless unknown Markovian process. Previous studies have analyzed a special case of this setting, in which each channel yields the same expected rate for all users. By contrast, we consider a more general and practical model, where each channel yields a different expected rate for each user. This model adds a significant challenge of how to efficiently learn a channel allocation in a distributed manner to yield a global system-wide objective. We adopt the stable matching utility as the system objective, which is known to yield strong performance in multichannel wireless networks, and develop a novel Distributed Stable Strategy Learning (DSSL) algorithm to achieve the objective. We prove theoretically that DSSL converges to the stable matching allocation, and the regret, defined as the loss in total rate with respect to the stable matching solution, has a logarithmic order with time. Finally, simulation results demonstrate the strong performance of the DSSL algorithm. ## I Introduction We consider the spectrum access problem, where a shared bandwidth is divided into $K$ orthogonal channels (i.e., sub-bands), and $M$ users want to access the spectrum, where $K\geq M$. Each channel is modeled by a Finite-State Markovian Channel (FSMC), which is independent and non-identically distributed across channels. The FSMC is a tractable model widely used to capture the time-varying behavior of a radio communication channel [2, 3]. It is often employed to model radio channel dynamics due to primary user occupancy effects in hierarchical cognitive radio networks (where the $M$ secondary (unlicensed) users are cognitive in terms of learning and adapting good access strategies), or the external interference effects in the open sharing model among $M$ users in the wireless network (e.g., ISM band) [4, 5]. At each time step, each user experiences a different transmission rate over each channel depending on its FSMC distribution, where the FSMC parameters (i.e., the transition probabilities that govern the Markov chain) are unknown. At each time step, each user is allowed to choose one channel to access, and observe the instantaneous channel state. If two users or more access the same channel at the same time, a collision occurs and the achievable rate is zero. We adopt the stable matching utility (see Section II for details) as the system objective, which is known to yield strong performance in multichannel wireless networks [6]. We define the regret as the loss in total rate with respect to the stable matching solution with known FSMCs. The objective is to develop a distributed learning algorithm for channel allocation and access under unknown FSMCs that minimizes the growth rate of the regret with time $t$. ### I-A Main Results The stable matching problem for multi-user spectrum access was first introduced in [6] under the assumption that the expected rates are known, and a distributed opportunistic CSMA algorithm that solves the problem was proposed. The model with an unknown expected rate matrix and rested setting (i.e., the states of the Markovian process do not change if not observed by the user) was studied in [7, 8]. A regret (with respect to the optimal allocation) of near-$O(\log t)$ was achieved. However, these algorithms require intensive communication between users in order to apply the auction algorithm [9]. In [10], the authors reduced the communication burden, but without guarantees on the achievable regret. Recently, it was shown in [11, 12] that achieving a sum-regret of near- $O(\log t)$ is possible without communication between users, but only for the case of i.i.d channels. In this paper we focus on the general case where the channel states may change whether or not they are being observed (i.e., the restless Markovian setting), and improve the regret scaling with the system parameters by a simple distributed implementation. The main contributions are summarized below. ##### A general model for spectrum access using a restless Markovian channel model As explained above, by contrast to [6, 7, 8, 10, 11, 12], in this paper we first solve the channel allocation and access problem under general unknown restless Markovian channel model. Handling this model adds significant challenges in algorithm design and regret analysis. Due to the restless nature of the channels and potential reward loss due to transient effects as compared to steady state when switching channels, learning the Markovian channel characteristics requires that the channels be accessed in a judicious consecutive manner for a period of time. This is reflected in a novel algorithm design that guarantees efficient learning, as detailed next. ##### Algorithm Development We are facing an online learning problem constituted by the well-known exploration versus exploitation dilemma. To remedy this, we propose a novel Distributed Stable Strategy Learning (DSSL) algorithm for solving the problem. Since the FSMCs are unknown, the rate means must be learned by accessing all channels via exploration phases. This results in increasing the regret, since the stable allocation is not performed. Thus, the exploration time must be minimized, while guaranteeing efficient learning. Roughly speaking, each channel can be learned by different exploration times, depending on its unknown parameters (see more details in Section III-D). The algorithm design in this paper contributes to both tackling the more general model, as well as improving the learning efficiency in a fully-distributed manner. Specifically, in existing algorithms [7, 8, 10, 11, 12], the exploration phase of all channels is determined by the channel that requires the largest exploration time. This results in oversampling the channels and significantly increases the regret. By contrast, the DSSL algorithm estimates online the desired (unknown) exploration rate of each channel. Thus, by sampling the channels according to the desired exploration rate, it avoids oversampling the channels, and thus reduces the regret scaling significantly as compared to existing algorithms. ##### Performance analysis In terms of theoretical performance analysis, we prove that the DSSL algorithm converges to the stable matching allocation, and the regret has a logarithmic order with time. When comparing to existing approaches, DSSL achieves this under the more general restless Markovian model, and also has significantly better scaling with the system parameters. Specifically, under a common benchmark setting of equal rates among users (but still vary among channels), and $K>M$, which allows a theoretical comparison of learning efficiency between different algorithms, in [8] and [13] the regret scales as $O(\frac{MK}{(\Delta_{\min})^{2}}\log(t))$ ,in [12] as $O(\frac{M^{3}K}{(\Delta_{\min})^{2}}\log(t))$ and in [11] the regret scales as $O(\frac{MK^{2}}{(\Delta_{\min})^{2}}\log(t))$, where $\Delta_{\min}$ is the difference in rates between the $M$th and $(M+1)$th best channels. In contrast, under DSSL, the regret scales as $O((\frac{1}{(\Delta_{\min})^{2}}+MK)\log(t))$. In addition, extensive numerical experiments were performed to demonstrate the efficiency of the proposed DSSL algorithm. ### I-B Related Work A number of studies have developed distributed learning algorithms for a special case of the restless Markovian channel model considered in this paper, where each channel yields the same expected rate for all users [14, 15, 16]. This special case significantly simplifies the channel allocation problem and the analysis (for instance, switching between assigned users does not affect the resulting regret in this special case). In this paper, we consider the general model where each channel yields a different expected rate for each user. This models the situation of different channel fading states across users and channels in actual wireless networks, and adds a significant challenge of how to learn the desired channel allocation in a distributed manner to achieve a global system-wide objective. Another set of related work on multi-user channel allocation has approached it from the angle of game theoretic and congestion control ( [17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27] and references therein), hidden channel states[28], and graph coloring ([29, 30, 31, 32] and references therein). The game theoretic aspects of the problem have been investigated from both non- cooperative (i.e., each user aims at maximizing an individual utility) [18, 19, 24, 25, 33], and cooperative (i.e., each user aims at maximizing a system- wide global utility) [17, 34, 26, 35] settings. Model-free learning strategies were developed in [36, 37] for orthogonal channels, compact models [38], and multiple access channel strategies were developed in [39, 40]. Graph coloring formulations have dealt with modeling the spectrum access problem as a graph coloring problem, in which users and channels are represented by vertices and colors, respectively (see [29, 30, 32, 31] and references therein for related studies). Finally, none of these studies have considered the problem of achieving provable stable strategies in the learning context under unknown restless Markovian dynamics, as considered in this paper. ## II System Model and Problem Formulation We consider a wireless network consisting of $K$ orthogonal channels indexed by the set $\mathcal{K}=\\{1,2,...,K\\}$ and $M$ cognitive users (referred to as users) indexed by the set $\mathcal{M}=\\{1,2,...,M\\}$, where $K\geq M$. The users aim at accessing the spectrum to send their data. Each user is allowed to choose a single channel for transmission at each time slot, and transmit if the channel is not occupied by a primary user. The users operate in a synchronous time-slotted fashion. Due to spatial geographic dispersion, each user can potentially experience different achievable rates over the channels. When a user $i$ transmits on channel $k$ (when the channel is free) at time slot $t$, its data rate is given by $r_{i,k}(t)$. This information is concisely represented by an $M\times K$ rate matrix $V(t)=\\{r_{ik}(t)\\}$, $i=1,...,M,k=1,...,K$. We consider the case where the rate process $r_{i,k}(t)$ is Markovian and has a well-defined steady state distribution. The transition probabilities associated with the Markov chain are unknown to the users. The process $r_{i,k}(t)$ evolves independently of the user’s actions (i.e., external process). Furthermore, the channel states may change depending on whether or not they are observed (i.e., restless setting). Specifically, the rate of user $i$ on channel $k$, $r_{i,k}(t)$, is modeled as a discrete time, irreducible and aperiodic Markov chain on a finite-state space $\mathcal{X}^{i,k}$ and is represented by a transition probability matrix $P^{i,k}\triangleq(p^{i,k}_{x,x^{\prime}}:x,x^{\prime}\in\mathcal{X}^{i,k})$. The process mean (i.e., the expected rate) is denoted by $\mu_{i,k}$ and is unknown to the users. We define the $M\times K$ expected rate matrix by $U=\\{\mu_{ik}\\}$, $i=1,...,M,k=1,...,K$. Let $X_{i,k}(t)$ be the actual achievable rate for user $i$ on channel $k$ at time $t$. If two or more users choose to access the same channel at the same time slot, a collision occurs. In this case, $X_{i,k}(t)=0$. Otherwise, if user $i$ has accessed channel $k$ without colliding with other users, then $X_{i,k}(t)=r_{i,k}(t)$. The users implement carrier sensing to observe the current channel state at each time slot as is typically done in cognitive radio networks [14, 22]. Hence, the channel states are observed regardless of collisions. The transmission scheme for the multi-user spectrum access model is detailed in Section III. ### II-A Notations We present the other notations that are used throughout the paper. Let $\vec{\pi}_{i,k}\triangleq(\pi^{x}_{i,k},x\in\mathcal{X}^{i,k})$ be the stationary distribution of the Markov chain $P^{i,k}$, and let: $\displaystyle\pi_{\min}\triangleq\min_{i\in\mathcal{M},k\in\mathcal{K},x\in\mathcal{X}^{i,k}}\ \pi^{x}_{i,k},\quad\hat{\pi}_{i,k}^{x}\triangleq\max\\{\pi_{i,k}^{x},1-\pi_{i,k}^{x}\\},\quad\hat{\pi}_{\max}\triangleq\max_{i\in\mathcal{M},k\in\mathcal{K},x\in\mathcal{X}^{i,k}}\\{\pi_{i,k}^{x},1-\pi_{i,k}^{x}\\}$. We define $X_{\max}\triangleq\max_{i\in\mathcal{M},k\in\mathcal{K}}\\{|\mathcal{X}^{i,k}|\\}$ as the maximal cardinality among the state spaces, and $\displaystyle x_{\max}\triangleq\max_{i\in\mathcal{M},k\in\mathcal{K},x\in\mathcal{X}^{i,k}}x,\quad r_{\max}\triangleq\max_{i\in\mathcal{M},k\in\mathcal{K}}\sum_{x\in{\mathcal{X}^{i,k}}}x$. Let $\lambda_{i,k}$ be the second largest eigenvalue of $P^{i,k}$, and $\displaystyle\lambda_{\max}\triangleq\max_{i\in\mathcal{M},k\in\mathcal{K}}\ \lambda_{i,k}$ be the maximal one among all channels and users. Also, $\displaystyle\overline{\lambda}_{\min}\triangleq 1-\lambda_{\max},\displaystyle\overline{\lambda}_{i,k}\triangleq 1-\lambda_{i,k}$ is the eigenvalue gap. Let $M^{i,k}_{x,y}$ be the mean hitting time of state $y$ starting at initial state $x$ for channel $k$ used by user $i$, and $\displaystyle M^{i,k}_{\max}\triangleq\max_{x,y\in\mathcal{X}^{i,k},x\neq y}M^{i,k}_{x,y}$. We also define: $A_{\max}\triangleq\max_{i,k}\;(\min_{x\in\mathcal{X}^{i,k}}\ \pi_{i,k}^{x})^{-1}\sum\limits_{x\in{\mathcal{X}^{i,k}}}x,$ and $\begin{array}[]{l}\displaystyle L\triangleq\frac{28x_{\max}^{2}r_{\max}^{2}\hat{\pi}_{\max}^{2}}{\bar{\lambda}_{\min}}.\end{array}$ (1) The expectations $\mu_{i,k}$ are given by: $\displaystyle\mu_{i,k}=\sum\limits_{x\in{\mathcal{X}^{i,k}}}x\cdot\pi_{i,k}^{x}$, and we define $\sigma_{i}$, for $i=1,...,M$, as a permutation of $\\{1,\ldots,K\\}$ such that $\displaystyle\mu_{i,\sigma_{i}(1)}>\mu_{i,\sigma_{i}(2)}>\ldots>\mu_{i,\sigma_{i}(K)}$. ### II-B A Stable Channel Allocation Let $a_{i}(t)\in\mathcal{K}$ be a selection rule, indicating which channel is chosen by user $i$ at time $t$, which is a mapping from the observed history of the process (i.e., all past actions and observations up to time $t-1$) to $\left\\{1,...,K\right\\}$. The expected aggregated data rate for all users up to time $t$ is given by: $\begin{array}[]{l}\displaystyle R(t)=\mathbb{E}[\sum\limits_{n=1}^{t}\sum\limits_{i=1}^{M}X_{i,a_{i}(n)}(n)].\end{array}$ (2) A policy $\phi_{i}$ is a time series vector of selection rules: $\phi_{i}=(a_{i}(t),t=1,2,...)$ for user $i$. Definition 1 ([6]): A bipartite matching between channels and users is a permutation $P:\mathcal{M}\rightarrow\mathcal{K}$. The optimal centralized allocation problem is to find a bipartite matching: $\displaystyle\mathbf{k}^{**}=\arg\max_{\mathbf{k}\in P}\sum\limits_{i=1}^{M}\mu_{i,k(i)}$. Definition 2 ([6]): A matching $S:\mathcal{M}\rightarrow\mathcal{K}$ is stable if for every $i\in\mathcal{M}$ and $k\in\mathcal{K}$ satisfying $S(i)\neq k$, if $\mu_{i,S(i)}<\mu_{i,k}$ then there exists some user $i^{\prime}\in\mathcal{M}$ such that $S(i^{\prime})=k$ and $\mu_{i^{\prime},k}>\mu_{i,k}$. Achieving the optimal allocation in Definition 1 requires implementing a centralized solution, or a distributed solution with heavy complexity and slow convergence [41]. Therefore, we are interested in developing a distributed algorithm with low complexity that converges to the stable matching solution in Definition 2 which is known to yield strong performance and very fast convergence (when the expected rates are known) by using distributed opportunistic CSMA (see Section III-B and [6] for more details on opportunistic CSMA for stable channel allocation). We assume that the entries in the matrix $U$ are all different, as in [6], which holds in wireless networks due to continuous-valued Shannon rates111Otherwise, we can add noise to the matrix.. Thus, there is a unique stable matching solution under our assumptions, and the expected aggregated rate under the stable matching solution $S$ is given by: $\sum\limits_{i=1}^{M}\mu_{i,S(i)}$. The channel $S(i)$ (i.e., the channel that user $i$ selects under the stable matching configuration) is referred to as the stable channel selection of user $i$. ###### Remark 1 We point out that under an i.i.d. or rested222In the rested model the Markov chain $P^{i,k}$ makes a state transition only when user $i$ accesses channel $k$. Markovian channel model, the optimal policy is to transmit on the same channels that achieves the optimal centralized allocation in terms of the sum expected rate. However, the optimal policy in the restless Markovian setting has been shown to be P-SPACE hard even under known Markoivan dynamics [42]. Therefore, a commonly adopted approach in this setting is to use a weaker definition of the regret, first introduced in [43] and used later; e.g., in [14, 15, 44, 45], where the policy is compared to a ”partially informed” genie who knows the expected rates of the channels, instead of the complete system dynamics. In this paper we adopt this approach as well. ### II-C The Objective Since the expected rates $\mu_{i,k}$ are unknown in our setting, the users must learn this information online effectively so as to converge to the stable matching solution. A widely used performance measure of online learning algorithms is the regret, which is defined as the reward loss with respect to an algorithm with a side information on the model. In our setting, we define the regret for policy $\phi=(\phi_{i},1\leq i\leq M)$ as the loss in the expected aggregated data rate with respect to the stable matching solution that uses the true expected rates: $\begin{array}[]{l}\displaystyle r_{\phi}(t)\triangleq t\cdot\sum\limits_{i=1}^{M}\mu_{i,S(i)}-\mathbb{E}_{\phi}[\sum\limits_{n=1}^{t}\sum\limits_{i=1}^{M}X_{i,\phi_{i}(n)}(n)].\end{array}$ (3) A policy $\phi$ that achieves a sublinear scaling rate of the regret with time (and consequently the time averaged regret tends to zero) approaches the required stable matching solution. The essence of the problem is thus to design an algorithm that learns the unknown expected rates efficiently to achieve the best sublinear scaling of the regret with time. ## III The Distributed Stable Strategy Learning (DSSL) Algorithm To achieve the objective, as detailed in Section II-C, we divide the time horizon into three phases, we term exploration, allocation, and exploitation. These three phases are performed repeatedly during the algorithm according to judiciously designed policy rules, as detailed later. The purpose of the exploration phase is to allow each user to explore all the channels to identify its $M$ best channels (i.e., the $M$ channels that yield the highest expected rates for the user). The users use the sample means as estimators for the expected rates of the channels to achieve this goal. This phase results in a regret loss, since users access sub-optimal channels to explore them, and the stable allocation is not performed. However, this phase is essential to identifying the $M$ best channels and consequently minimizing the regret scaling with time. The purpose of the exploitation phase is to use the currently learned information to execute the stable matching solution. The allocation phase allows users to allocate the channels among themselves properly in a distributed manner using opportunistic carrier sensing [46]. Since the rate process $r_{i,k}(t)$ can evolve even when channel $k$ is not selected by user $i$, learning the Markovian rate statistics requires using the channels in a consecutive manner for a period of time [14, 15]. Moreover, frequent switching between channels can cause a loss due to the transient effect. The high-level structure of the DSSL algorithm works as follows. Each user $i$ computes its sufficient number of samples in the exploration phases (condition (13) defined in III-E) for each channel $k$ at the end of every exploitation phase $t$. If the number of samples is greater than the required number for all $k$, user $i$ performs another exploitation phase. Otherwise, if the number of samples is smaller than the sufficient number for one or more channels, user $i$ carries out an exploration phase for those channels. When no exploration phase is needed, an allocation phase is performed. At the end of the allocation phase, each user identifies its stable channel selection, and an exploitation phase is carried out. We now discuss the structure of the DSSL algorithm in details. ### III-A The structure of the exploration phase: Let $n_{O}^{i,k}(t)$ be the number of exploration phases in which channel $k$ was selected by user $i$ up to time $t$. Each exploration phase is divided into two sub epochs: a Random size Epoch (RE), and a Deterministic size Epoch (DE). Let $\gamma^{i,k}(n_{O}^{i,k}(t)-1)$ be the last channel state observed at the $(n_{O}^{i,k}(t)-1)^{th}$ exploration phase. RE starts at the beginning of the exploration phase until state $\gamma^{i,k}(n_{O}^{i,k}(t)-1)$ is observed. This epoch ensures that the generated sample path (after removing the samples observed in the RE epochs) is equivalent to a sample path generated by continuously sensing the Markovian channel without switching. This step guarantees a consistent estimation of the expected rates. Then, DE starts by sensing the channel for a deterministic period of time $4^{n_{O}^{i,k}(t)}$. The deterministic period of time grows geometrically with time to ensure a relatively small number of channel switching. ### III-B The structure of the allocation phase: The allocation phase applies opportunistic CSMA among users. In opportunistic CSMA, the backoff function maps from an index (i.e., expected rate) to a backoff time [46]. The backoff function decreases monotonically with the rates, so that the user with the highest rate on a certain channel waits the minimal time before transmission. All other users sense that the channel is occupied and do not transmit on that channel. To obtain the stable matching allocation, this procedure continues until all $M$ users occupy $M$ channels. For more details on opportunistic CSMA for stable matching see [6]. The allocation phase has two goals in our setting. The first is to assign channels to users to yield a stable matching solution as in [6]. However, since the expected rates are unknown in our setting, the allocation phase is executed by using the sample means. The second goal is to use the backoff function to identify the differences in sample means among users and channels, which is needed for setting efficient learning rates. This requires a new mechanism that performs opportunistic CSMA, as detailed below. Let $\mathcal{T}_{k}$ be the set of all users that attempt to transmit on channel $k$ at a certain stage of the allocation phase. We initialize the phase by declaring each user to be unassigned. We divide the time horizon of the allocation phase into two sub-phases. In the first sub-phase, referred to as $S_{1}$, we perform opportunistic CSMA for stable matching as in [6], while replacing the expected rates by the sample means. Specifically, each unassigned user attempts to transmit on its best channel, out of those it has not yet attempted using opportunistic CSMA. On each channel $k$, the best user out of $\mathcal{T}_{k}$ in this sub-phase ($S_{1}$) is declared to be assigned. All the other users in $\mathcal{T}_{k}$ store the sample mean of the assigned user (by mapping from the sensed backoff time to the sample mean). This sub-phase continues until all $M$ users are assigned to $M$ channels. The second sub-phase, referred to as $S_{2}$, is used to obtain the side information required for efficient learning. Specifically, the opportunistic CSMA is executed again, but the assigned users of each channel do not transmit. All other users that attempted to transmit in $S_{1}$ transmit again on the same channel $k$. The sample mean of the best user in $S_{2}$ (i.e., the second best user in $\mathcal{T}_{k}$ for each channel $k$) is stored by the assigned user. This sub-phase continues until all $M$ users in $S_{2}$ were observed, and the phase ends. An example for $M=K=3$ is given next. The expected rate matrix is shown in Table I. Table II shows the transmission attempts made by the users in the allocation phase before the stable matching was achieved (the assigned users are shown in bold). At time $t=1$, each user transmits on its best channel (sub-phase $S_{1}$). Users $1$ and $2$ aim to access the same channel (channel $2$), and the channel is assigned to user $2$ since it has a higher expected rate on this channel (i.e., smaller backoff time). At time $t=2$, sub-phase $S_{2}$ is performed, in which user $1$ transmits again on channel $2$. At time $t=3$, user $1$ (the only unassigned user) tries to access its second best channel; i.e., channel $1$. However, the channel is assigned to user $3$ since it has a higher expected rate. The algorithm continues until the three users are assigned to the three channels. TABLE I: expected rate matrix U | channel 1 | channel 2 | channel 3 ---|---|---|--- user 1 | 45 | 70 | 35 user 2 | 30 | 90 | 60 user 3 | 65 | 10 | 50 TABLE II: allocation phase Sub-phase | Time | channel 1 | channel 2 | channel 3 ---|---|---|---|--- $S_{1}$ | t=1 | 3 | 1,2 | $S_{2}$ | t=2 | | 1 | $S_{1}$ | t=3 | 1,3 | 2 | $S_{2}$ | t=4 | 1 | | $S_{1}$ | t=5 | 3 | 2 | 1 ### III-C The structure of the exploitation phase: Let $n_{I}(t)$ be the number of exploitation phases up to time $t$. In the exploitation phase, each user transmits on the channel it was assigned according to the last allocation phase (during $S_{1}$) for a deterministic period of time $2\cdot 4^{n_{I}(t)-1}$ (for the $n_{I}^{th}$ exploitation phase). There are no channel switching and no sample mean updating during the exploitation phase. ### III-D Parameter setting for efficient learning: As discussed earlier, exploring the channels increases the regret since the stable matching allocation is not used. On the other hand, it is essential to reduce the estimation error and hence reduce the regret scaling order with time. In this section, we establish the sufficient exploration rate of each channel for each user to achieve efficient learning of the stable matching allocation. We next establish two parameters used in the learning strategy. #### III-D1 Identifying $M$ best channels We show in the analysis that a user (say user $i$) who is interested in distinguishing with a sufficiently high accuracy between two channels $k,l$ that yield expected rates $\mu_{i,k},\mu_{i,\ell}$, respectively, must explore them at least $\displaystyle\frac{4L}{(\mu_{i,k}-\mu_{i,\ell})^{2}}\cdot\log(t)$ times. Let $\mathcal{M}_{i}$ be the set of the $M$ best channels of user $i$. For each channel $k\in\mathcal{M}_{i}$ we define the deterministic row333This definition is consistent with the definition of the $M\times K$ expected rate matrix by $U=\\{\mu_{ik}\\}$, $i=1,...,M,k=1,...,K$. exploration coefficient as $\begin{array}[]{l}\displaystyle D_{i,k}^{(R)}\triangleq\frac{4L}{\displaystyle\min_{\ell\neq k}\\{(\mu_{i,k}-\mu_{i,\ell})^{2}\\}},\end{array}$ (4) and for channel $k{\not\in}\mathcal{M}_{i}$, $\begin{array}[]{l}\displaystyle D_{i,k}^{(R)}\triangleq\frac{4L}{(\mu_{i,k}-\mu_{i,\sigma_{i}(M)})^{2}}.\end{array}$ (5) Since the expected rates are unknown, the users need to estimate $D_{i,k}^{(R)}$ for each channel $k\in\mathcal{K}$. This estimator is denoted by $\widehat{D}_{i,k}^{(R)}(t)$. Let $\bar{s}_{i,k}(t)$ be the mean transmission rate of user $i$ on channel $k$. Thus, the adaptive row exploration coefficient for channels $k\in\mathcal{M}_{i}$ is defined by $\begin{array}[]{l}\displaystyle\widehat{D}_{i,k}^{(R)}(t)\triangleq\frac{4L}{\max\big{\\{}\Delta_{\min}^{2},\displaystyle\min_{\ell\neq k}\\{(\bar{s}_{i,k}(t)-\bar{s}_{i,\ell}(t))^{2}\\}-\epsilon\big{\\}}},\end{array}$ (6) and similarly for $k{\not\in}\mathcal{M}_{i}$ we have: $\begin{array}[]{l}\displaystyle\widehat{D}_{i,k}^{(R)}(t)\triangleq\frac{4L}{\max\\{\Delta_{\min}^{2},(\bar{s}_{i,k}(t)-\bar{s}_{i,\sigma_{i}(M)}(t))^{2}-\epsilon\\}},\end{array}$ (7) where $\Delta_{\min}$ is the smallest difference between two entries in the expected rate matrix $U$; i.e., $\hskip 85.35826pt\Delta_{\min}\triangleq\displaystyle\min_{i\in\mathcal{M}}\Delta_{i}$, $\vspace{0.2cm}\hskip 71.13188pt\Delta_{i}\triangleq\displaystyle\min_{k,\ell\in\mathcal{K},k\neq\ell}|\mu_{i,k}-\mu_{i,\ell}|.$ #### III-D2 CSMA protocol identification Consistent with the opportunistic CSMA protocol described above, each user $i$ needs to distinguish between a channel $k\in\mathcal{T}_{k}$ (this channel is in $\mathcal{M}_{i}$ as well), and the best channel in $\mathcal{T}_{k}$ (and the second best channel in $\mathcal{T}_{k}$ if $k$ is the best channel in $\mathcal{T}_{k}$), for all $k$. Hence, we define the deterministic column exploration coefficient for user $i$ for channel $k\in\mathcal{T}_{k}$ by: $\begin{array}[]{l}\displaystyle D_{i,k}^{(C)}\triangleq\frac{4L}{(\mu_{i,k}-\displaystyle\max_{j\neq i,j\in\mathcal{T}_{k}}\mu_{j,k})^{2}},\end{array}$ (8) and the adaptive column exploration coefficient by: $\begin{array}[]{l}\displaystyle\widehat{D}_{i,k}^{(C)}(t)\triangleq\frac{4L}{\displaystyle\max\\{\Delta_{\min}^{2},(\bar{s}_{i,k}(t)-\max_{j\neq i}\bar{s}_{j,k}(t))^{2}-\epsilon\\}}.\end{array}$ (9) Note that $\max_{j\neq i,j\in\mathcal{T}_{k}}\bar{s}_{j,k}(t)$ is known to user $i$ by the design of the opportunistic CSMA (by sub-phase $S_{2}$). By combining (4) and (8), the deterministic exploration-rate coefficient of user $i$ for channels $k\in\mathcal{M}_{i}\cap\mathcal{T}_{k}$ is given by: $\begin{array}[]{l}\displaystyle D_{i,k}\triangleq\max\\{D_{i,k}^{(R)},D_{i,k}^{(C)}\\},\end{array}$ (10) and by combining (6) and (9), the adaptive exploration-rate coefficient of user $i$ for channels $k\in\mathcal{M}_{i}\cap\mathcal{T}_{k}$ is given by: $\begin{array}[]{l}\displaystyle\widehat{D}_{i,k}(t)=\max\\{\widehat{D}_{i,k}^{(R)}(t),\widehat{D}_{i,k}^{(C)}(t)\\}.\end{array}$ (11) ###### Remark 2 The design of the adaptive exploration-rate coefficients under DSSL significantly reduces the regret as compared to existing algorithms that use deterministic exploration-rate coefficients determined by the channel that requires the largest exploration time [8, 10, 11, 12]. For example, consider the expected rate matrix $U$ given in Table I, where parameter $L$ in (1) equals $10^{4}$. In Table III, we present the deterministic exploration-rate coefficients $D_{i,k}$ defined in (10) for each channel-user pair under DSSL, where $D_{i,k}\cdot\log(t)$ is the number of samples required to achieve consistent estimates of the expected rates. By contrast, in other existing algorithms [8, 10, 11, 12], all channels are explored with the same exploration-rate coefficient, which is inversely proportional to the squared difference between the mean rate of the optimal allocation and the second best one. When applying this to our example, each channel should be explored for $1600\cdot\log(t)$ time steps (as seen in Table IV), which significantly increases the exploration times unnecessarily, and consequently increases the regret. TABLE III: Exploration coefficients under the DSSL algorithm $D_{i,k}$ | channel 1 | channel 2 | channel 3 ---|---|---|--- user 1 | 400 | 100 | 400 user 2 | 45 | 100 | 45 user 3 | 178 | 25 | 178 TABLE IV: Exploration coefficients under other existing algorithms [8, 10, 11, 12] $D_{i,k}$ | channel 1 | channel 2 | channel 3 ---|---|---|--- user 1 | 1600 | 1600 | 1600 user 2 | 1600 | 1600 | 1600 user 3 | 1600 | 1600 | 1600 ### III-E Choosing between phases types: Since $D_{i,k}$ is unknown, the algorithm replaces $D_{i,k}$ by its estimate $\widehat{D}_{i,k}(t)$. Furthermore, to ensure that $\widehat{D}_{i,k}(t)$ overestimates $D_{i,k}$, the users need to sense at least $I\cdot\log(t)$ times each of their channels in exploration phases, where $\begin{array}[]{l}\displaystyle I\triangleq\frac{7\epsilon^{2}}{48(r_{\max}+2)^{2}\cdot L},\end{array}$ (12) which can be viewed as the rate function of the estimators among all channels. At the end of the exploitation phases, the users check the condition: $\begin{array}[]{l}\displaystyle T_{i,k}^{(O)}(t)>\max\left\\{\widehat{D}_{i,k}(t),\frac{2}{I}\right\\}\cdot\log(t),\end{array}$ (13) where $T_{i,k}^{(O)}(t)$ is the number of samples in the exploration phases accessed in sub epochs DE for user $i$ on channel $k$ up to time $t$. If the condition holds for user $i$, the user enters another exploitation phase by transmitting on the same channel in which it transmitted during the last exploitation phase. Otherwise, if the condition does not hold, the user enters an exploration phase by sensing channel $k$. At the end of the phase, the user signals the other users that it has finished the exploration phase. If such an interruption occurred, all the users again check condition (13). If it holds for all users, they start an allocation phase. At the end of the allocation phase, an exploitation phase starts. A pseudocode of the DSSL algorithm is provided in Algorithm 1. Algorithm 1 DSSL Algorithm for user $i$ Initialization: For all $K$ channels, execute an exploration phase where a single observation is taken from each channel; while $t\leq T$ do if Condition (13) does not hold for channel $k$ then Enter an exploration phase with length $4^{n_{O}^{i,k}(t)}$; Update $\bar{s}_{i,k}(t)$ and increment $n_{O}^{i,k}(t)=n_{O}^{i,k}(t)+1$; goto step 3 end if Send an interrupt signal; Start an allocation phase; Start an exploitation phase with length $2\cdot 4^{n_{I}(t)}$. If an interruption occurs, go to step $3$; $n_{I}(t)=n_{I}(t)+1$; end while ## IV Regret Analysis Success in obtaining a logarithmic regret order depends on how fast $\widehat{D}_{i,k}(t)$ converges to a value which is no smaller than $D_{i,k}$ (so that user $i$ senses channel $k$ at least $D_{i,k}\cdot\log t$ time slots in most of the times). The analysis in the Appendix shows that exploring channels as in (13) guarantees the desired convergence speed. Specifically, in the following theorem we establish a finite-sample bound on the regret with time, which results in a logarithmic scaling of the regret. ###### Theorem 1 Assume that the proposed DSSL algorithm is implemented and that the assumptions on the system model described in Section II hold. Then, the regret at time $t$ is upper bounded by: $\begin{array}[]{l}\vspace{0.3cm}\hskip 8.5359pt\displaystyle r(t)\leq A_{\max}\cdot\bigg{(}\sum\limits_{i=1}^{M}\sum\limits_{k=1}^{K}(\lfloor\log_{4}(3A_{i,k}\log(t)+1)\rfloor+1)\bigg{)}\\\ \vspace{0.3cm}\hskip 19.91684pt\displaystyle+\sum\limits_{i=1}^{M}\sum\limits_{k=1}^{K}\bigg{[}\bigg{(}4A_{i,k}\cdot\log(t)+1\\\ \vspace{0.3cm}\hskip 42.67912pt\displaystyle+M_{\max}^{i,k}\big{(}\lfloor\log_{4}(3A_{i,k}\log(t)+1)\rfloor+1\big{)}\bigg{)}\\\ \vspace{0.3cm}\hskip 99.58464pt\displaystyle\cdot\bigg{(}\mu_{i,S(i)}+\mu_{S^{-1}(k),k}-\mu_{i,k}\bigg{)}\bigg{]}\end{array}$ $\begin{array}[]{l}\vspace{0.3cm}\hskip 19.91684pt\displaystyle+M^{2}\cdot A_{\max}\cdot\bigg{(}\sum\limits_{i=1}^{M}\sum\limits_{k=1}^{K}(\lfloor\log_{4}(3A_{i,k}\log(t)+1)\rfloor+1)\bigg{)}\\\ \vspace{0.3cm}\hskip 19.91684pt\displaystyle+\bigg{[}\bigg{(}2e\log(M+1)\bigg{)}\\\ \vspace{0.3cm}\hskip 34.14322pt\cdot\bigg{(}\sum\limits_{i=1}^{M}\sum\limits_{k=1}^{K}(\lfloor\log_{4}(3A_{i,k}\log(t)+1)\rfloor+1)\bigg{)}\bigg{]}\\\ \vspace{0.3cm}\hskip 22.76228pt\displaystyle\cdot\bigg{[}\sum\limits_{j=1}^{M}\mu_{j,S(j)}\bigg{]}\\\ \vspace{0.3cm}\hskip 19.91684pt\displaystyle+\bigg{(}A_{\max}+(M^{2}K+MK)\frac{6X_{\max}}{\pi_{\min}}\big{(}\sum\limits_{j=1}^{M}\mu_{j,S(j)}\big{)}\bigg{)}\\\ \vspace{0.3cm}\hskip 22.76228pt\displaystyle\cdot\bigg{(}\lceil\log_{4}(\frac{3}{2}t+1)\rceil\bigg{)}+O(1),\end{array}$ (14) where $A_{i,k}$ is given by: $\displaystyle\vspace{0.0cm}A_{i,k}\triangleq\left\\{\begin{matrix}\max\\{2/I\;,\;D_{i,k}^{(\max)}\\}\;,&\mbox{if $k\in\mathcal{G}_{i}$}\vspace{0.0cm}\\\ \max\\{2/I\;,\;4L/\Delta_{\min}^{2}\\}\;,&\mbox{if $k{\not\in}\mathcal{G}_{i}$}\end{matrix}\right.\;,$ (15) $\mathcal{G}_{i}$ is defined as the set of all indices $k\in\mathcal{K}$ of user $i$ that satisfy: $\displaystyle\min\\{(\displaystyle\min_{\ell\neq k}\\{\mu_{i,k}-\mu_{i,\ell}\\})^{2},(\mu_{i,k}-\max_{j\neq i}\mu_{j,k})^{2}\\}-2\epsilon>\Delta_{\min}^{2},$ for $k\in\mathcal{T}_{k}$, and $\displaystyle(\displaystyle\min_{\ell\neq k}\\{\mu_{i,k}-\mu_{i,\ell}\\})^{2}-2\epsilon>\Delta_{\min}^{2},$ for $k{\not\in}\mathcal{T}_{k}$, where $D_{i,k}^{(\max)}$ is defined as: $\begin{array}[]{l}\displaystyle D_{i,k}^{(\max)}\triangleq{\frac{4L}{\min\big{\\{}(\displaystyle\min_{\ell\neq k}\\{\mu_{i,k}-\mu_{i,\ell}\\})^{2},(\mu_{i,k}-\max_{j\neq i}\mu_{j,k})^{2}\big{\\}}-2\epsilon}}.\vspace{0.0cm}\end{array}$ (16) The proof is given in the Appendix. Note that Theorem 1 shows that similar to [13, 8, 11, 12], the regret under DSSL has a logarithmic order with time. DSSL, however, achieves this under the more general restless Markovian model, and also has significantly better scaling with $M,K$ and $\Delta_{\min}$. Specifically, under a common benchmark setting of equal rates among users (but still vary among channels), and $K>M$, which allows a theoretical comparison of learning efficiency between different algorithms, in [8] and [13] the regret scales as $O(\frac{MK}{(\Delta_{\min})^{2}}\log(t))$ ,in [12] as $O(\frac{M^{3}K}{(\Delta_{\min})^{2}}\log(t))$ and in [11] the regret scales as $O(\frac{MK^{2}}{(\Delta_{\min})^{2}}\log(t))$. In contrast, under DSSL, the regret scales as $O((\frac{1}{(\Delta_{\min})^{2}}+MK)\log(t))$ due to the novel algorithm design that explores every channel according to its unique adaptive exploration rate, while guaranteeing efficient learning. ## V Simulation Results In this section we present simulation results to evaluate the performance of DSSL numerically. In Subsection V-A we start by evaluating the convergence of DSSL under unknown restless fading FSMCs with respect to the stable matching solution solved under known restless fading FSMCs. We also evaluate the performance as compared to random allocation and the optimal centralized allocation schemes. Then, in Section V-B we examine the learning efficiency of DSSL as compared to other online learning algorithms under unknown restless FSMC, and verify our theoretical logarithmic regret. We performed $1,000$ Monte-Carlo experiments and averaged the performance over the experiments. ### V-A Convergence of DSSL to stable matching We start by describing the wireless channel model used in the simulations. Each user experiences a block fading channel which remains constant during each time slot, and varies between time slots. The channel response experienced by user $i$ at time slot $t$ is given by $h(i,t)=r(i,t)e^{j\rho(i,t)}$, where $r(i,t)=|h(i,t)|$ denotes the channel rate, and $\rho(i,t)$ denotes the channel phase experienced by user $i$ at time $t$. Let $f(i,r)$ denote the Probability Density Function (PDF) of the fading channel rate $r(i)$ experienced by user $i$ (e.g., Rayleigh fading distribution in the simulations). We consider independent but non-identically distributed channels across users, and Markovian correlated channels across time slots. The FSMC model [2, 3] partitions the range of the channel gain values into a finite number of intervals and represents each interval as a state of a Markov chain. The thresholds of the intervals at user $i$ are denoted by $\tau_{n}(i),n=0,\ldots N$, where $0=\tau_{0}(i)<\tau_{1}(i)<\ldots<\tau_{N-1}(i)<\tau_{N}(i)=\infty$. The channel rate $r(i,t)$ experienced by user $i$ is said to be in state $g_{n}(i),1<n<N$, if it lies in the interval: $t_{n-1}(i)\leq r(i,t)<\tau_{n}(i)$. The states are partitioned to yield an equal initial state probability for all states: $\displaystyle\int_{\tau_{n-1}(i)}^{\tau_{n}(i)}f(i,r)dr=\frac{1}{N},n=1,\ldots,N$. The transition probability to transition from state $g_{n}(i)$ to state $g_{\ell}(i)$ is defined by: $\displaystyle p_{n,\ell}(i)\triangleq Pr(\tau_{\ell-1}(i)\leq r(i,t+1)<\tau_{\ell}(i)$ $|\tau_{n-1}(i)\leq r(i,t)<\tau_{n}(i))$ where $r(i,t)$ and $r(i,t+1)$ are the current channel gain and the channel gain in the next time slot experienced by user $i$, respectively. In the simulations, we quantized the channel gain to $6$ states; i.e., $N=6$, and we simulated a case of $3$ users and $5$ channels. The transition probability matrix $P$ and the expected rate matrix $U$ are given by: $P=\left(\begin{array}[]{cccccc}3/6&2/6&1/6&0&0&0\\\ 2/8&3/8&2/8&1/8&0&0\\\ 1/9&2/9&3/9&2/9&1/9&0\\\ 0&1/9&2/9&3/9&2/9&1/9\\\ 0&0&1/8&2/8&3/8&2/8\\\ 0&0&0&1/6&2/6&3/6\end{array}\right)$, $U=\left(\begin{array}[]{ccccc}45&70&35&17.5&12.5\\\ 27.5&90&60&15&20\\\ 65&10&50&16.5&30\end{array}\right)$. We compared the expected rate evolution of DSSL under unknown FSMCs against stable matching, random allocation and the optimal centralized allocation solved under known FSMCs. The optimal centralized algorithm served as an upper bound benchmark for all algorithms, and the stable matching served as an upper bound for DSSL. In the random allocation scheme users access an arbitrary channel with equal probability. As shown in Fig. 1 the average rate under DSSL converged to that of the stable matching, as desired. The stable matching allocation allocates user 1 to channel 3, user 2 to channel 2, and user 3 to channel 1. Fig. 2 shows that the average achievable rate of each user in the DSSL algorithm converged to the stable allocation. Figure 1: Comparison of the system average rate of various schemes Figure 2: Comparison of users’ average rate for the proposed DSSL algorithm ### V-B Learning efficiency of DSSL We next evaluated the learning efficiency of DSSL as compared to other online learning algorithms under unknown restless FSMCs. We considered the hierarchical access channel model in spectrum access networks. This models the situation of primary and secondary users that share the spectrum. Primary users (licensed) occupy the spectrum occasionally, and a secondary user is allowed to transmit over a single channel when the channel is free. Thus, each channel has two states, good (free) and bad (occupied). The good state results in a positive expected rate, whereas bad state result in a zero rate. The occupancies of the channels by the primary users are modeled as Markov processes (i.e., Gilbert-Elliott channel). First, we simulated a special case of our model where each channel yielded the same expected rate for all users. In [14, 15], the RCA and DSEE algorithms were proposed to solve this special case. The RCA algorithm performs random regenerative cycles until catching predefined states in each phase, which results in oversampling the channels, and therefore is expected to increase the regret as compared to DSSL. The DSEE algorithm overcomes this issue by performing deterministic sequencing for both the exploration and exploitation phases. However, the deterministic sequencing requires the algorithm to explore all channels using the maximal exploration rate among all channels, which is expected to increase the regret as compared to DSSL (that learns the desired exploration rate for each channel) as well. We simulated the case of $2$ users, $6$ channels, each with two states: 0, 1. The transition probabilities for all channels to transition from 0 to 1 and from 1 to 0, respectively, were $p_{01}=[0.1,0.1,0.5,0.1,0.1,0.7]$, $p_{10}=[0.2,0.3,0.1,0.4,0.5,0.08]$, the expected rates for all channels at states 1, 0, respectively, are $r_{1}=[1,1,1,1,1,1]$, $r_{0}=[0.1,0.1,0.1,0.1,0.1,0.1]$. As can be seen in Fig. 3, the DSSL algorithm outperformed both RCA and DSEE and achieved the logarithmic regret order with time. Finally, we simulated the scenario where the stable matching allocation was also the optimal centralized allocation, and the channels were i.i.d. across time slots (and not Markovian). We compared DSSL to the $dE^{3}$ algorithm which was designed for this setting. However, $dE^{3}$ requires communication between users since it implements a distributed auction that requires users to observe the bids of other users [8]. We used the same parameters as selected and tuned by the authors in [8]. Similar to the DSEE algorithm, in $dE^{3}$ the exploration-rate coefficient was determined by the channel with the largest exploration time. Thus, we expected that DSSL would yield a faster convergence rate due to the adaptive design of the exploration epochs. As shown in Fig. 4, DSSL indeed outperformed the $dE^{3}$ algorithm. Figure 3: The regret (normalized by $\log t$) under DSSL, DSEE, and RCA as a function of time. Parameter setting: 2 users, 6 channels, each with two states: 0, 1. Transition probabilities for all channels to transition from 0 to 1 and from 1 to 0, respectively: $p_{01}=[0.1,0.1,0.5,0.1,0.1,0.7]$, $p_{10}=[0.2,0.3,0.1,0.4,0.5,0.08]$, expected rates for all channels at states 1, 0, respectively: $r_{1}=[1,1,1,1,1,1]$, $r_{0}=[0.1,0.1,0.1,0.1,0.1,0.1]$. Figure 4: The regret under DSSL and $dE^{3}$ as a function of time. Parameter setting: 3 users, 3 channels, with mean transmission rates: $[0.2,0.25,0.3;0.4,0.6,0.5;0.7,0.9,0.8]$. ## VI Conclusion We developed a novel algorithm for the multi-user spectrum access problem in wireless networks, dubbed the Distributed Stable Strategy Learning (DSSL) algorithm. In contrast to existing models, for the first time we considered the case of restless Markov channels, which requires a different algorithm structure to accurately learn the channel statistics. Moreover, the channels selection rules are adaptive in order to reduce the exploration time required for efficient learning. We showed theoretically that DSSL achieves a logarithmic regret with time, and better regret scaling with the system parameters as compared to existing approaches that have studied special cases of the model. Extensive simulation results supported the theoretical study and demonstrated the strong performance of DSSL. ## VII Appendix In this appendix we prove Theorem 1. ###### Definition 1 Let $T_{1}$ be the smallest integer, such that for all $t\geq T_{1}$ the following holds: $D_{i,k}\leq\widehat{D}_{i,k}(t)$ for all $i\in\mathcal{M},k\in\mathcal{K}$, and also $\widehat{D}_{i,k}(t)\leq D_{i,k}^{(\max)}$ for all $i\in\mathcal{M},k\in\mathcal{G}_{i}$. ###### Lemma 1 Assume that the DSSL algorithm is implemented as described in Section III. Then, $E(T_{1})<\infty$ is bounded independent of $t$. Proof: $E(T_{1})$ can be written as follows: $E[T_{1}]=\sum\limits_{n=1}^{\infty}n\cdot Pr\left(T_{1}=n\right)=\sum\limits_{n=1}^{\infty}\Pr\left(T_{1}\geq n\right)\\\ =\vspace{0.0cm}\hskip 8.5359pt\sum\limits_{n=1}^{\infty}\Pr\big{(}\bigcup\limits_{i\in\mathcal{M}}\bigcup\limits_{k\in\mathcal{G}_{i}}\bigcup\limits_{l=n}^{\infty}(\widehat{D}_{i,k}(l)<D_{i,k}\mbox{\;or\;}\\\ \vspace{0.0cm}\hskip 110.96556pt\widehat{D}_{i,k}(l)>D_{i,k}^{(\max)})\mbox{\;or\;}\\\ \vspace{0.0cm}\hskip 51.21504pt\bigcup\limits_{i\in\mathcal{M}}\bigcup\limits_{k{\not\in}\mathcal{G}_{i}}\bigcup\limits_{l=n}^{\infty}(\widehat{D}_{i,k}(l)<D_{i,k})\big{)}\\\ \leq\vspace{0.0cm}\sum\limits_{i\in\mathcal{M}}\sum\limits_{k\in\mathcal{G}_{i}}\sum\limits_{n=1}^{\infty}\sum\limits_{l=n}^{\infty}\Pr\big{(}\widehat{D}_{i,k}(l)<D_{i,k}\mbox{\;or\;}\widehat{D}_{i,k}(l)>D_{i,k}^{(\max)}\big{)}\\\ \vspace{0.0cm}\hskip 8.5359pt+\sum\limits_{i\in\mathcal{M}}\sum\limits_{k{\not\in}\mathcal{G}_{i}}\sum\limits_{n=1}^{\infty}\sum\limits_{l=n}^{\infty}\Pr\big{(}\widehat{D}_{i,k}(l)<D_{i,k}\big{)}$ Note that if we show that $\begin{array}[]{l}\Pr\big{(}\widehat{D}_{i,k}(l)<D_{i,k}\mbox{\;or\;}\widehat{D}_{i,k}(l)>D_{i,k}^{(\max)}\big{)}\leq C\cdot l^{-(2+\delta)}\end{array}$ (17) for some constants $C>0,\delta>0$ for all $i\in\mathcal{M},k\in\mathcal{G}_{i}$ for all $l\geq n$, then we get: $\displaystyle\sum\limits_{i\in\mathcal{M}}\sum\limits_{k\in\mathcal{G}_{i}}\sum\limits_{n=1}^{\infty}\sum\limits_{l=n}^{\infty}\Pr\big{(}\widehat{D}_{i,k}(l)<D_{i,k}\mbox{\;or\;}\widehat{D}_{i,k}(l)>D_{i,k}^{(\max)}\big{)}\\\ \leq\vspace{0.1cm}\hskip 8.5359ptMK\cdot C\left[\sum\limits_{l=1}^{\infty}l^{-(2+\delta)}+\sum\limits_{n=2}^{\infty}\sum\limits_{l=n}^{\infty}l^{-(2+\delta)}\right]\\\ \leq\vspace{0.1cm}\hskip 8.5359ptMK\cdot C\left[\sum\limits_{l=1}^{\infty}l^{-(2+\delta)}+\sum\limits_{n=2}^{\infty}\int\limits_{n-1}^{\infty}l^{-(2+\delta)}dl\right]\\\ =\vspace{0.1cm}\hskip 8.5359ptMK\cdot C\left[\sum\limits_{l=1}^{\infty}l^{-(2+\delta)}+\frac{1}{1+\delta}\sum\limits_{n=2}^{\infty}(n-1)^{-(1+\delta)}\right]<\infty$, which is bounded independent of $t$. Similarly, showing that $\Pr\big{(}\widehat{D}_{i,k}(l)<D_{i,k}\big{)}\leq C\cdot l^{-(2+\delta)}$ for some constants $C,\delta>0$ for all $i\in\mathcal{M},k{\not\in}\mathcal{G}_{i}$ for all $j\geq n$ completes the statement. We start bounding (17). We look at the first inequality of (17) for user $i$ with channel $k\in\mathcal{M}_{i}\cap\mathcal{T}_{k}$. The event $\widehat{D}_{i,k}(t)<D_{i,k}$ implies: $\vspace{0.3cm}\max\bigg{\\{}\Delta_{\min}^{2},\min\big{\\{}\displaystyle\min_{\ell\neq k}\\{(\bar{s}_{i,k}(t)-\bar{s}_{i,\ell}(t))^{2}\\}-\epsilon,\\\ \vspace{0.3cm}\hskip 34.14322pt(\bar{s}_{i,k}(t)-\max_{j\neq i}\bar{s}_{j,k}(t))^{2}-\epsilon\big{\\}}\bigg{\\}}\\\ \vspace{0.3cm}\hskip 5.69046pt>\min\big{\\{}\displaystyle\min_{\ell\neq k}\\{(\mu_{i,k}-\mu_{i,\ell})^{2}\\},(\mu_{i,k}-\max_{j\neq i}\mu_{j,k})^{2}\big{\\}},\\\ $ which after algebraic manipulations implies that at least one of the following holds: $\vspace{0.3cm}\hskip 14.22636pt\displaystyle\min_{\ell\neq k}\\{(\bar{s}_{i,k}(t)-\bar{s}_{i,\ell}(t))^{2}\\}-\epsilon>\displaystyle\min_{\ell\neq k}\\{(\mu_{i,k}-\mu_{i,\ell})^{2}\\}\\\ \vspace{0.3cm}\hskip 14.22636pt(\bar{s}_{i,k}(t)-\max_{j\neq i}\bar{s}_{j,k}(t))^{2}-\epsilon>(\mu_{i,k}-\max_{j\neq i}\mu_{j,k})^{2}.$ Similarly, the second inequality of (17) implies one of the following: $\vspace{0.3cm}\hskip 8.5359pt\displaystyle\min_{\ell\neq k}\\{(\bar{s}_{i,k}(t)-\bar{s}_{i,\ell}(t))^{2}\\}-\epsilon<\displaystyle\min_{\ell\neq k}\\{(\mu_{i,k}-\mu_{i,\ell})^{2}\\}-2\epsilon\\\ \vspace{0.3cm}\hskip 8.5359pt(\bar{s}_{i,k}(t)-\max_{j\neq i}\bar{s}_{j,k}(t))^{2}-\epsilon<(\mu_{i,k}-\max_{j\neq i}\mu_{j,k})^{2}-2\epsilon.$ Let $k^{*}=\displaystyle\text{arg}\min_{\ell\neq k}(\mu_{i,k}-\mu_{i,\ell}\\})^{2}$ (i.e., $(\mu_{i,k}-\mu_{i,k^{*}}\\})^{2}=\displaystyle\min_{\ell\neq k}\\{(\mu_{i,k}-\mu_{i,\ell})^{2}\\}$). Cascading the events written above we get : $\vspace{0.0cm}\hskip 8.5359pt\Pr\big{(}\widehat{D}_{i,k}(t)<D_{i,k}\mbox{\;or\;}\widehat{D}_{i,k}(t)>D_{i,k}^{(\max)}\ \big{)}$ $\displaystyle\vspace{0.3cm}\leq\Pr\big{(}|(\bar{s}_{i,k}(t)-\bar{s}_{i,k^{*}}(t))^{2}-(\mu_{i,k}-\mu_{i,k^{*}})^{2}|>\epsilon\big{)}$ $\displaystyle\vspace{0.3cm}+\Pr\big{(}|(\bar{s}_{i,k}(t)-\max_{j\neq i}\bar{s}_{j,k}(t))^{2}-(\mu_{i,k}-\max_{j\neq i}\mu_{j,k})^{2}|>\epsilon\big{)}.$ (18) Each of the terms in (18) is the probability of a deviation of the squared difference for two Markov chains’ sample means from the squared difference of their expected means by an $\epsilon$. We look at the first term of (18). Using conventional steps from set theory, it can be shown that: $\vspace{0.3cm}\Pr\big{(}|(\bar{s}_{i,k}(t)-\bar{s}_{i,k^{*}}(t))^{2}-(\mu_{i,k}-\mu_{i,k^{*}})^{2}|>\epsilon\big{)}\\\ \vspace{0.2cm}\leq\big{[}\Pr\big{(}|(\bar{s}_{i,k}(t)-\bar{s}_{i,k^{*}}(t))[(\bar{s}_{i,k}(t)-\bar{s}_{i,k^{*}}(t))\\\ \vspace{0.3cm}\hskip 133.72786pt-(\mu_{i,k}-\mu_{i,k^{*}})]|>\frac{\epsilon}{2}\big{)}\big{]}\\\ \vspace{0.2cm}+\big{[}\Pr\big{(}|(\mu_{i,k}-\mu_{i,k^{*}})[(\bar{s}_{i,k}(t)-\bar{s}_{i,k^{*}}(t))\\\ \vspace{0.3cm}\hskip 133.72786pt-(\mu_{i,k}-\mu_{i,k^{*}})]|>\frac{\epsilon}{2}\big{)}\big{]}\\\ \vspace{0.3cm}\leq\big{[}\Pr\big{(}|(\bar{s}_{i,k}(t)-\bar{s}_{i,k^{*}}(t))-(\mu_{i,k}-\mu_{i,k^{*}})|>1\big{)}\\\ \vspace{0.3cm}+\Pr\big{(}|(\bar{s}_{i,k}(t)-\bar{s}_{i,k^{*}}(t))-(\mu_{i,k}-\mu_{i,k^{*}})|>\frac{\epsilon}{2(R+1)}\big{)}\\\ \vspace{0.3cm}+\Pr\big{(}|(\mu_{i,k}-\mu_{i,k^{*}})+1|>R\big{)}\big{]}\\\ \vspace{0.3cm}+\big{[}\vspace{0.3cm}\Pr\big{(}\mu_{i,k}>R^{\prime}\big{)}\\\ \vspace{0.3cm}+\Pr\big{(}|(\bar{s}_{i,k}(t)-\bar{s}_{i,k^{*}}(t))-(\mu_{i,k}-\mu_{i,k^{*}})|>\frac{\epsilon}{2(R^{\prime}+1)}\big{)}\big{]},\\\ $ for every $R,R^{\prime}>0$. We choose $R=R^{\prime}=r_{\max}+1,$ hence the third and fourth terms are equal to $0$, and we get the concentration inequalities: $\vspace{0.0cm}\Pr\big{(}|(\bar{s}_{i,k}(t)-\bar{s}_{i,k^{*}}(t))^{2}-(\mu_{i,k}-\mu_{i,k^{*}})^{2}|>\epsilon\big{)}$ $\displaystyle<6\cdot\max\bigg{\\{}\Pr\big{(}|\bar{s}_{i,k}(t)-\mu_{i,k}|>\frac{\epsilon}{4(r_{\max}+2)}\big{)},$ (19) $\displaystyle\Pr\big{(}|\bar{s}_{i,k^{*}}(t)-\mu_{i,k^{*}}|>\frac{\epsilon}{4(r_{\max}+2)}\big{)}\bigg{\\}}.$ (20) Similar bounds can be obtained for the second term in (18). To bound (19) and (20) we use Lezaud’s results [47]: ###### Lemma 2 ([47]) Consider a finite-state, irreducible Markov chain $\\{X_{t}\\}_{t\geq 1}$ with state space $S$, matrix of transition probabilities $P$, an initial distribution $q$, and stationary distribution $\pi$. Let $N_{\textbf{q}}=\left\|(\frac{q^{(x)}}{\pi^{(x)}},x\in S)\right\|_{2}$. Let $\widehat{P}=P^{\prime}P$ be the multiplicative symmetrization of $P$ where $P^{\prime}$ is the adjoint of $P$ on $l_{2}(\pi)$. Let $\epsilon=1-\lambda_{2}$, where $\lambda_{2}$ is the second largest eigenvalue of the matrix $P^{\prime}$. $\epsilon$ will be referred to as the eigenvalue gap of $P^{\prime}$. Let $f:S\rightarrow\mathcal{R}$ be such that $\sum\limits_{y\in S}\pi_{y}f(y)=0,\quad\|f\|_{2}\leq 1$ and $0\leq\|f\|_{2}^{2}\leq 1$ if $P^{\prime}$ is irreducible. Then, for any positive integer $n$ and all $0<\lambda\leq 1$, we have: $P\displaystyle\left(\frac{\sum\limits_{t=1}^{n}f(X_{t})}{n}\geq\lambda\right)\leq N_{\textbf{q}}$ exp $[-\frac{n\lambda^{2}\epsilon}{12}].$ Consider an initial distribution $\textbf{q}^{i,k}$ for channel $k$ of user $i$. We have: $\displaystyle N_{\textbf{q}}^{(i,k)}=\left\|(\frac{q_{i,k}^{x}}{\pi_{i,k}^{x}},x\in X^{i,k})\right\|_{2}\leq\sum\limits_{x\in X^{i,k}}\left\|\frac{q_{i,k}^{x}}{\pi_{i,k}^{x}}\right\|_{2}\leq\frac{1}{\pi_{min}}.$ We point out that the sample rate mean $\bar{s}_{i,k}(t)$ is computed by $T^{(O)}_{i,k}(t)$ observation taken only from sub epochs DE in the exploration phases, thus the sample path that generated $\bar{s}_{i,k}(t)$ can be viewed as a sample path generated by a Markov chain with a transition matrix identical to the original channel $\\{i,k\\}$, so we can apply Lezaud’s result to bound (19) and (20). For equation (19): we define $n_{x}^{i,k}(t)$ to be the number of occurrences of state $x$ on channel $k$ sensed by user $i$ up to time t. $\vspace{0.3cm}\Pr\big{(}\bar{s}_{i,k}(t)-\mu_{i,k}>\frac{\epsilon}{4(r_{\max}+2)}\big{)}\\\ =\vspace{0.3cm}\Pr\big{(}\sum\limits_{x\in\mathcal{X}^{i,k}}x\cdot n_{x}^{i,k}(t)-T^{(O)}_{i,k}(t)\sum\limits_{x\in\mathcal{X}^{i,k}}x\cdot\pi_{i,k}^{x}>\frac{T^{(O)}_{i,k}(t)\cdot\epsilon}{4(r_{\max}+2)}\big{)}\\\ =\vspace{0.3cm}\Pr\big{(}\sum\limits_{x\in\mathcal{X}^{i,k}}(x\cdot n_{x}^{i,k}(t)-T^{(O)}_{i,k}(t)x\cdot\pi_{i,k}^{x})>\frac{T^{(O)}_{i,k}(t)\cdot\epsilon}{4(r_{\max}+2)}\big{)}\\\ \leq\vspace{0.3cm}\sum\limits_{x\in\mathcal{X}^{i,k}}\Pr\big{(}x\cdot n_{x}^{i,k}(t)-T^{(O)}_{i,k}(t)x\cdot\pi_{i,k}^{x}>\frac{T^{(O)}_{i,k}(t)\cdot\epsilon}{4(r_{\max}+2)|\mathcal{X}^{i,k}|}\big{)}\\\ =\vspace{0.3cm}\sum\limits_{x\in\mathcal{X}^{i,k}}\Pr\big{(}n_{x}^{i,k}(t)-T^{(O)}_{i,k}(t)\cdot\pi_{i,k}^{x}>\frac{T^{(O)}_{i,k}(t)\cdot\epsilon}{4(r_{\max}+2)|\mathcal{X}^{i,k}|\cdot x}\big{)}\\\ =\vspace{0.3cm}\sum\limits_{x\in\mathcal{X}^{i,k}}\Pr\bigg{(}\frac{\sum\limits_{n=1}^{t}\textbf{1}(x_{i,k}(n)=x)-T^{(O)}_{i,k}(t)\pi_{i,k}^{x}}{\hat{\pi}_{i,k}^{x}\cdot T^{(O)}_{i,k}(t)}\\\ \vspace{0.3cm}\hskip 128.0374pt>\frac{T^{(O)}_{i,k}(t)\cdot\epsilon}{4(r_{\max}+2)|\mathcal{X}^{i,k}|\cdot x\hat{\pi}_{i,k}^{x}}\bigg{)}\\\ \leq\vspace{0.3cm}|\mathcal{X}^{i,k}|\cdot N_{\textbf{q}}^{(i,k)}$ exp $\bigg{(}-T^{(O)}_{i,k}(t)\cdot\frac{\epsilon^{2}}{16(r_{\max}+2)^{2}\cdot x^{2}\cdot|\mathcal{X}^{i,k}|^{2}\cdot(\hat{\pi}_{i,k}^{x})^{2}}\\\ \vspace{0.3cm}\hskip 184.9429pt\cdot\frac{(1-\lambda_{i,k})}{12}\bigg{)},$ and from (13), we have: $T^{(O)}_{i,k}(t)>\frac{2}{I}\log(t)$ with $I$ defined in (12). Thus, $\displaystyle\displaystyle\Pr\big{(}|\bar{s}_{i,k}(t)-\mu_{i,k}|>\frac{\epsilon}{4(r_{\max}+2)}\big{)}\leq\frac{|X_{\max}|}{\pi_{\min}}\cdot t^{-2+\delta}.$ (21) The same bound can be obtained for (20), and with the same steps, for all terms in (18). The proof for all $i\in\mathcal{M},k{\not\in}\mathcal{G}_{i}$ is similar, and thus Lemma 1 follows. $\square$ We now bound the expected regret defined in (3). We divide the time horizon for $t<T_{1}$ and $t>T_{1}$. Since $T_{1}$ is finite (due to Lemma 1), the regret for all $t<T_{1}$ results in a constant term $O(1)$ which is independent of $t$. For $t>T_{1}$, we know that the adaptive exploration coefficient is no smaller than the deterministic exploration coefficient, and no larger than $D_{i,k}^{(\max)}$ defined in (16); i.e., $D_{i,k}\leq\widehat{D}_{i,k}(t)\leq D_{i,k}^{(\max)},$ (22) for all $i\in\mathcal{M},k\in\mathcal{G}_{i}$ , and the LHS of the inequality for $i\in\mathcal{M},k\in\mathcal{K}$. Thus, the exploration phases provides sufficient learning for the channel statistics (and the upper bound ensures that the channels are judiciously oversampled in the exploration phases). We continue bounding the regret for $t>T_{1}$: $\displaystyle\displaystyle r(t)\leq(t-T_{1})\cdot\sum\limits_{i=1}^{M}\mu_{i,S(i)}-\mathbb{E}[\sum\limits_{n=T_{1}+1}^{t}\sum\limits_{i=1}^{M}X_{i,a_{i}(n)}(n)].$ (23) For convenience, we will develop (23) between $n=1$ and $t$ with (22) (and the LHS for $k{\not\in}\mathcal{G}_{i}$) holds for all $1\leq n\leq t$, which upper bounds (23): $\vspace{0.0cm}r(t)\leq(t-T_{1})\cdot\sum\limits_{i=1}^{M}\mu_{i,S(i)}-\mathbb{E}[\sum\limits_{n=T_{1}+1}^{t}\sum\limits_{i=1}^{M}X_{i,a_{i}(n)}(n)]$ $\displaystyle\leq t\cdot\sum\limits_{i=1}^{M}\mu_{i,S(i)}-\mathbb{E}[\sum\limits_{n=1}^{t}\sum\limits_{i=1}^{M}X_{i,a_{i}(n)}(n)].$ (24) We can rewrite (24) as: $\displaystyle r(t)$ $\displaystyle\leq\sum\limits_{i=1}^{M}\sum\limits_{k=1}^{K}\big{(}\mu_{i,k}\cdot E[T_{i,k}(t)]-E[\sum\limits_{n=1}^{t}X_{i,k}(n)]\big{)}$ (25) $\displaystyle+\big{(}t\cdot\sum\limits_{i=1}^{M}\mu_{i,S(i)}-\sum\limits_{i=1}^{M}\sum\limits_{k=1}^{K}\mu_{i,k}\cdot E[T_{i,k}(t)]\big{)},$ (26) where $T_{i,k}(t)$ is the total number of transmission for user $i$ on channel $k$ up to time $t$ (and $X_{i,k}(n)=0$ if user $i$ did not try to access channel $k$ at time $n$). Equation (25) can be considered as the regret due to the transient effect (the initial state of the channel may not be given by the stationary distribution), and (26) is the regret caused by not playing the stable matching allocation. Both (25) and (26) can be thought of as the sum of three different regret terms, corresponding to the three phases described in Section III. We denote by $r^{O}(t),r^{A}(t),r^{I}(t)$ the regret caused in the exploration, allocation and exploitation phases respectively; i.e., the regret can be written as: $\displaystyle r(t)=r^{O}(t)+r^{A}(t)+r^{I}(t).$ (27) We next bound the regret in each of the three phases. Regret in the exploration phases: To bound the regret in the exploration phases, we first bound the number of exploration phases $n_{O}^{i,k}(t)$ for each user $i\in\mathcal{M}$ on each channel $k\in\mathcal{K}$ by time $t$. As described in Section (III-A), the total number of samples from the exploration phases in sub epochs DE for user $i$ on channel $k$ up to time $t$ is: $\displaystyle T_{i,k}^{(O)}(t)=\sum\limits_{n=1}^{n_{O}^{i,k}(t)}4^{n-1}=\frac{1}{3}(4^{n_{O}^{i,k}(t)}-1)$. Since we are in an exploration phase, from (13) together with (22), we have $T_{i,k}^{(O)}(t)<A_{i,k}\cdot\log(t)$ ($A_{i,k}$ is defined in (15). Hence, $\begin{array}[]{l}n_{O}^{i,k}(t)\leq\lfloor\log_{4}(3A_{i,k}\log(t)+1)\rfloor+1.\end{array}$ (28) We use the following lemma to show that the regret caused by channel switching is upper bounded by a constant independent of the number of transmissions on the channel in each phase. ###### Lemma 3 ([48]) Consider an irreducible, aperiodic Markov chain with state space $S$, a matrix of transition probabilities $P$, an initial distribution $\overrightarrow{q}$ which is positive in all states, and stationary distribution $\overrightarrow{\pi}(\pi_{s}$ is the stationary probability of state s). The state (reward) at time $t$ is denoted by $s(t)$. Let $\mu$ denote the mean reward. If we play the chain for an arbitrary time $T$, then there exists a value $A_{p}\leq(\min_{s\in S}\pi_{s})^{-}1\sum\limits_{s\in S}s$, such that: $E[\sum\limits_{t=1}^{T}s(t)-\mu T]\leq A_{p}$. Lemma 3 bounds the probability of a large deviation from the stationary distribution of a Markov chain (which we refer to as the transient effect). By the construction of the exploration phases described in Section (III-A), in each exploration phase there is no channel switching (each channel has its own unique exploration phases), therefore (25) in the exploration phases is bounded by: $\begin{array}[]{l}A_{\max}\cdot\big{(}\sum\limits_{i=1}^{M}\sum\limits_{k=1}^{K}(\lfloor\log_{4}(3A_{i,k}\log(t)+1)\rfloor+1)\big{)}.\end{array}$ (29) We next bound (26) in the exploration phases. Note that each user has its own exploration time, independent of the other users; i.e., when user $i$ explores, the other users (for which condition (13) holds) continue to exploit. However, user’s $i$ exploration may affect other users exploring during that time due to collision. Specifically, when user $i$ explores channel $k$ it affects the regret in two ways. First, user $i$ does not transmit in its stable channel; hence, the regret is increased by $\mu_{i,S(i)}-\mu_{i,k}$. Second, if $k$ is a stable channel of another user, then because of the collision, the regret will increase by $\mu_{S^{-1}(k),k}$ ($S^{-1}(k)$ is the user for which channel $k$ is its stable channel ). Combining these two terms, we bound (26) in exploration phases by: $\begin{array}[]{l}\sum\limits_{i=1}^{M}\sum\limits_{k=1}^{K}\bigg{(}E[N_{i,k}^{(O)}(t)]\cdot(\mu_{i,S(i)}+\mu_{S^{-1}(k),k}-\mu_{i,k})\bigg{)},\end{array}$ (30) where $N_{i,k}^{(O)}(t)$ consists of the time indices from RE and DE, and depends on the mean hitting time of the channel due to the regenerative cycles. With (28) we have: $\vspace{0.0cm}E[N_{i,k}^{(O)}(t)]\leq\sum\limits_{n=0}^{n_{O}^{i,k}-1}(4^{n}+M^{i,k}_{\max})\\\ \vspace{0.0cm}=\frac{1}{3}(4^{n_{O}^{i,k}(t)}-1)+M^{i,k}_{max}\cdot n_{O}^{i,k}(t)$ $\begin{array}[]{l}\vspace{0.0cm}\leq\frac{1}{3}[4(3A_{i,k}\cdot\log(t)+1)-1]\\\ \vspace{0.0cm}+M^{i,k}_{\max}\cdot\log_{4}(3A_{i,k}\log(t)+1).\end{array}$ (31) Combining (29) and (30) we can bound the first term in (27): $\begin{array}[]{l}\vspace{0.0cm}r^{O}(t)\leq A_{\max}\cdot\big{(}\sum\limits_{i=1}^{M}\sum\limits_{k=1}^{K}(\lfloor\log_{4}(3A_{i,k}\log(t)+1)\rfloor+1)\big{)}\\\ +\vspace{0.0cm}\sum\limits_{i=1}^{M}\sum\limits_{k=1}^{K}\bigg{(}E[N_{i,k}^{(O)}(t)]\cdot(\mu_{i,S(i)}+\mu_{S^{-1}(k),k}-\mu_{i,k})\bigg{)},\end{array}$ (32) which coincides with the first and second terms on the RHS of (14). Regret in the allocation phases: Since an allocation phase will only come after an exploration phase, the number of allocation phases by time $t$, $n_{A}(t)$ is bounded by the total number of exploration phases by time $t$; i.e., $\vspace{0.0cm}n_{A}(t)\leq\sum\limits_{i=1}^{M}\sum\limits_{k=1}^{K}n_{O}^{i,k}(t),$ and by using (28) we have: $\begin{array}[]{l}n_{A}(t)\leq\sum\limits_{i=1}^{M}\sum\limits_{k=1}^{K}\lfloor\log_{4}(3A_{i,k}\log(t)+1)\rfloor+1.\end{array}$ (33) Since the expected rates are unknown in our setting, the allocation phase is executed using the sample means. To bound the expected time required for each allocation phase, we use proposition VI.4. in [6]: ###### Lemma 4 ([6]) Denote the expected delay to reach a stable matching configuration by $T_{M}$. There is some constant $C$ s.t. for every $M$ we have: $T_{M}\leq C\log(M+1).$ Specifically, it was shown in [6] that it is sufficient to choose $C=2e$ for the bound to hold. Lemma 4 states that each allocation phase is finite with respect to $t$, and only depends on the number of users. The total time in allocation phases by time $t$, denoted by $T_{A}(t)$, can be bounded by combining (33) with lemma 4: $\begin{array}[]{l}\vspace{0.3cm}E[T_{A}(t)]\leq\big{(}2C\log(M+1)\big{)}\\\ \vspace{0.0cm}\hskip 8.5359pt\cdot\big{(}\sum\limits_{i=1}^{M}\sum\limits_{k=1}^{K}\lfloor\log_{4}(3A_{i,k}\log(t)+1)\rfloor+1\big{)},\end{array}$ (34) with $\displaystyle C=2e$. We now bound (25) and (26) for the allocation phases. In each allocation phase, the maximum number of channel switchings is $M\cdot M$; thus, the regret caused by the transient effect is bounded by: $\begin{array}[]{l}A_{\max}\cdot M^{2}\cdot\big{(}\sum\limits_{i=1}^{M}\sum\limits_{k=1}^{K}(\lfloor\log_{4}(3A_{i,k}\log(t)+1)\rfloor+1)\big{)}.\end{array}$ (35) and the regret due to sub-optimal allocation can be bounded by: $\begin{array}[]{l}E[T_{A}(t)]\cdot\big{(}\sum\limits_{i=1}^{M}\mu_{i,S(i)}\big{)}.\end{array}$ (36) Combining (35), (36) we have: $\begin{array}[]{l}\vspace{0.3cm}r^{A}(t)\leq A_{\max}\cdot M^{2}\cdot\big{(}\sum\limits_{i=1}^{M}\sum\limits_{k=1}^{K}(\lfloor\log_{4}(3A_{i,k}\log(t)+1)\rfloor+1)\big{)}\\\ \vspace{0.3cm}+\big{[}\big{(}C\log(M+1)\big{)}\cdot\big{(}\sum\limits_{i=1}^{M}\sum\limits_{k=1}^{K}\lfloor\log_{4}(3A_{i,k}\log(t)+1)\rfloor+1\big{)}\big{]}\\\ \vspace{0.0cm}\hskip 56.9055pt\cdot\big{(}\sum\limits_{i=1}^{M}\mu_{i,S(i)}\big{)},\end{array}$ (37) which coincides with the third and fourth terms in the RHS of (14). Regret in the exploitation phases: We first bound the number of exploitation phases up to time $t$. As described in Section III-C, the number of time slots in the $n^{th}$ exploitation phase is $2\cdot 4^{(n-1)}$. Thus we have: $\sum\limits_{n=1}^{n_{I}(t)}2\cdot 4^{n-1}=\frac{2}{3}(4^{n_{I}}-1)\leq t,$ which implies $\begin{array}[]{l}n_{I}\leq\lceil\log_{4}(\frac{3}{2}t+1)\rceil.\end{array}$ (38) During the exploitation phases, there are no channel switchings (each user exploits its stable channel). As a result, the regret caused by the transient effect in the exploitation phases is upper bounded by: $\begin{array}[]{l}A_{\max}\cdot\lceil\log_{4}(\frac{3}{2}t+1)\rceil.\end{array}$ (39) It remains to bound the regret as a result of not playing the stable matching allocation (which we refer to as a sub-optimal allocation) in the exploitation phases. The event of playing a sub-optimal allocation in an exploitation phase occurs if the previous allocation phase results in a sub-optimal allocation, which occurs if one of the following takes place. The first is that user $i$ did not correctly identify the order of its $M$ best channels entering the allocation phase. This event would be denoted by $Y_{i}$. The second eventuality is when the user with the highest expected rate in channel $k$ was not identified correctly in the allocation phase. This event is denoted by $Z_{k}$. We write these events explicitly: $\vspace{0.0cm}\displaystyle Y_{i}(t_{n})=\bigcup\limits_{k\in\mathcal{M}_{i}}\bigcup\limits_{l\in\mathcal{K}}\big{\\{}\bar{s}_{i,k}(t_{n})<\bar{s}_{i,l}(t_{n})|\mu_{i,k}>\mu_{i,l}\big{\\}}$ $\displaystyle Z_{k}(t_{n})=\bigcup\limits_{j\in\mathcal{T}_{k}}\big{\\{}\bar{s}_{i,k}(t_{n})<\bar{s}_{j,k}(t_{n})|\mu_{i,k}=\max_{l\in\mathcal{T}_{k}}\mu_{l,k}\big{\\}},$ where $t_{n}$ denotes the starting time of the $n^{th}$ exploitation phase. Based on the above notations, the probability for a sub-optimal allocation ($P_{S}(n)$) in an exploitation phase at time $t_{n}$ is given by: $\vspace{0.0cm}\hskip 8.5359pt\displaystyle P_{S}(n)\triangleq\Pr\big{(}\bigcup\limits_{i\in\mathcal{M}}Y_{i}(t_{n})\mbox{\;or\;}\bigcup\limits_{k\in\mathcal{K}}Z_{k}(t_{n})\big{)}.$ The number of time slots in a sub-optimal allocation in the exploitation phases can be written as: $\\\ \vspace{0.3cm}\hskip 8.5359pt\displaystyle E[\tilde{T}(t)]=\sum\limits_{n=1}^{n_{I}(t)}2\cdot 4^{n-1}\cdot P_{S}(n)\leq\sum\limits_{n=1}^{\lceil\log_{4}(\frac{3}{2}t+1)\rceil}2\cdot 4^{n-1}\cdot P_{S}(n)$ $\displaystyle\vspace{0.3cm}\hskip 8.5359pt\displaystyle\leq\sum\limits_{n=1}^{\lceil\log_{4}(\frac{3}{2}t+1)\rceil}3t_{n}\cdot P_{S}(n).$ (40) To complete Theorem 1, we need to show that: $\displaystyle\displaystyle P_{S}(n)=\Pr\big{(}\bigcup\limits_{i\in\mathcal{M}}Y_{i}(t_{n})\mbox{\;or\;}\bigcup\limits_{k\in\mathcal{K}}Z_{k}(t_{n})\big{)}\leq B\cdot t_{n}^{-1},$ (41) for some $B>0$ (there is only a logarithmic number of terms in (40)). Using union bounds we have: $\displaystyle\vspace{0.0cm}\Pr\big{(}\bigcup\limits_{i\in\mathcal{M}}Y_{i}(t_{n})\mbox{\;or\;}\bigcup\limits_{k\in\mathcal{K}}Z_{k}(t_{n})\big{)}\\\ $ $\displaystyle\leq$ $\displaystyle M^{2}K\cdot\Pr\big{(}\bar{s}_{i,k}(t_{n})<\bar{s}_{i,l}(t_{n})|\mu_{i,k}>\mu_{i,l}\big{)}$ (42) $\displaystyle+$ $\displaystyle MK\cdot\Pr\big{(}\bar{s}_{i,k}(t_{n})<\bar{s}_{j,k}(t_{n})|\mu_{i,k}=\max_{l\in\mathcal{T}_{k}}\mu_{l,k}\big{)}$ (43) To bound (42) and (43), we define $C_{t,v}=\sqrt{L\log(t)/v}$. Equation (42) implies that at least one of the following must hold $\displaystyle\bar{s}_{i,k}(t_{n})\leq\mu_{i,k}-C_{t_{n},T_{i,k}^{(O)}}$ (44) $\displaystyle\bar{s}_{i,l}(t_{n})\geq\mu_{i,l}+C_{t_{n},T_{i,l}^{(O)}}$ (45) $\displaystyle\mu_{i,k}<\mu_{i,l}+C_{t_{n},T_{i,l}^{(O)}}+C_{t_{n},T_{i,k}^{(O)}}.$ (46) First we show that the probability for event (46) is zero. $\vspace{0.3cm}\hskip 11.38092pt\Pr\big{(}\mu_{i,k}<\mu_{i,l}+C_{t_{n},T_{i,l}^{(O)}}+C_{t_{n},T_{i,k}^{(O)}}\big{)}\\\ \vspace{0.3cm}\displaystyle=\Pr\bigg{(}\mu_{i,k}-\mu_{i,l}<\sqrt{\frac{L\log t_{n}}{T_{i,l}^{(O)}(t_{n})}}+\sqrt{\frac{L\log t_{n}}{T_{i,k}^{(O)}(t_{n})}}\bigg{)}\\\ \vspace{0.3cm}\displaystyle\leq\Pr\bigg{(}\mu_{i,k}-\mu_{i,l}<2\sqrt{\frac{L\log t_{n}}{\min\left\\{T_{i,k}^{(O)}(t_{n}),T_{i,l}^{(O)}(t_{n})\right\\}}}\bigg{)}\\\ \vspace{0.3cm}\displaystyle\leq\Pr\bigg{(}\min\left\\{T_{i,k}^{(O)}(t_{n}),T_{i,l}^{(O)}(t_{n})\right\\}<\frac{4L}{(\mu_{i,k}-\mu_{i,l})^{2}}\log(t_{n})\bigg{)}.\\\ $ Combining (22) with (13) (which holds since we started an allocation phase), we have: $\vspace{0.3cm}\displaystyle T_{i,k}^{(O)}(t_{n})>\frac{4L}{\displaystyle\min_{\ell\neq k}\\{(\mu_{i,k}-\mu_{i,\ell})^{2}\\}}\log(t_{n})\\\ \vspace{0.3cm}\hskip 113.81102pt\geq\frac{4L}{(\mu_{i,k}-\mu_{i,l})^{2}}\log(t_{n})\\\ \vspace{0.3cm}\displaystyle T_{i,l}^{(O)}(t_{n})>\frac{4L}{\displaystyle\min_{j\neq\ell}\\{(\mu_{i,l}-\mu_{i,j})^{2}\\}}\log(t_{n})\\\ \vspace{0.3cm}\hskip 113.81102pt\geq\frac{4L}{(\mu_{i,k}-\mu_{i,l})^{2}}\log(t_{n}),$ which ensures that the probability of (46) is zero. Note that here we used the fact that $D_{i,k}\geq D_{i,k}^{(R)}.$ We now bound (44) and (45) using Lezaud’s result (Lemma 3). With similar steps as used above to bound (19), we can show: $\displaystyle\Pr\big{(}\bar{s}_{i,k}(t_{n})\leq\mu_{i,k}-C_{t_{n},v_{i,k}}\big{)}\leq\frac{|\mathcal{X}^{i,k}|}{\pi_{\min}}t^{-\frac{L\bar{\lambda}_{\min}}{28X_{\max}^{2}r_{\max}^{2}\hat{\pi}_{\max}^{2}}}$ (47) $\displaystyle\Pr\big{(}\bar{s}_{i,l}(t_{n})\geq\mu_{i,l}+C_{t_{n},v_{i,l}}\big{)}\leq\frac{|\mathcal{X}^{i,l}|}{\pi_{\min}}t^{-\frac{L\bar{\lambda}_{\min}}{28X_{\max}^{2}r_{\max}^{2}\hat{\pi}_{\max}^{2}}}.$ (48) Using (1), (42) is bounded by: $\displaystyle M^{2}K\cdot\Pr\big{(}\bar{s}_{i,k}(t_{n})<\bar{s}_{i,l}(t_{n})|\mu_{i,k}>\mu_{i,l}\big{)}$ $\displaystyle\vspace{0.0cm}\leq$ $\displaystyle M^{2}K\cdot\frac{2X_{\max}}{\pi_{\min}}\cdot t^{-1}.$ (49) Equation (43) can be bounded using similar techniques, this time using the fact that $D_{i,k}\geq D_{i,k}^{(C)}$, and we can bound (41): $\displaystyle\displaystyle\Pr\big{(}\bigcup\limits_{i\in\mathcal{M}}Y_{i}(t_{n})\mbox{\;or\;}\bigcup\limits_{k\in\mathcal{K}}Z_{k}(t_{n})\big{)}\leq(M^{2}K+MK)\frac{2X_{\max}}{\pi_{\min}}\cdot t^{-1}.$ (50) With (50) we can bound (40), and therefore the regret due to sub-optimal allocation in the exploitation phases is bounded by: $\begin{array}[]{l}\displaystyle 3\big{(}\sum\limits_{i=1}^{M}\mu_{i,S(i)}\big{)}(M^{2}K+MK)\frac{2X_{\max}}{\pi_{\min}}\cdot\lceil\log_{4}(\frac{3}{2}t+1)\rceil.\end{array}$ (51) By combining (51) with (39), the total regret in the exploitation phases is: $\begin{array}[]{l}\vspace{0.0cm}\displaystyle r^{I}(t)\leq A_{\max}\cdot\lceil\log_{4}(\frac{3}{2}t+1)\rceil\\\ \vspace{0.0cm}\hskip 22.76228pt\displaystyle+3\big{(}\sum\limits_{i=1}^{M}\mu_{i,S(i)}\big{)}(M^{2}K+MK)\frac{2X_{\max}}{\pi_{\min}}\cdot\lceil\log_{4}(\frac{3}{2}t+1)\rceil,\end{array}$ (52) which coincides with the two last terms on the RHS of (14). ## References * [1] T. Gafni and K. Cohen, “A distributed stable strategy learning algorithm for multi-user dynamic spectrum access,” in 57th Annual Allerton Conference on Communication, Control, and Computing (Allerton), pp. 347–351, 2019. * [2] H. S. Wang and N. Moayeri, “Finite-state markov channel-a useful model for radio communication channels,” IEEE transactions on vehicular technology, vol. 44, no. 1, pp. 163–171, 1995. * [3] P. Sadeghi, R. A. Kennedy, P. B. Rapajic, and R. Shams, “Finite-state markov modeling of fading channels-a survey of principles and applications,” IEEE Signal Processing Magazine, vol. 25, no. 5, pp. 57–80, 2008. * [4] Q. Zhao and B. Sadler, “A survey of dynamic spectrum access,” IEEE Signal Processing Magazine, vol. 24, no. 3, pp. 79–89, 2007. * [5] N. Slamnik-Kriještorac, H. Kremo, M. Ruffini, and J. M. Marquez-Barja, “Sharing distributed and heterogeneous resources toward end-to-end 5g networks: A comprehensive survey and a taxonomy,” IEEE Communications Surveys & Tutorials, vol. 22, no. 3, pp. 1592–1628, 2020. * [6] A. Leshem, E. Zehavi, and Y. Yaffe, “Multichannel opportunistic carrier sensing for stable channel access control in cognitive radio systems,” IEEE Journal on Selected Areas in Communications, vol. 30, no. 1, pp. 82–95, 2012. * [7] D. Kalathil, N. Nayyar, and R. Jain, “Decentralized learning for multiplayer multiarmed bandits,” IEEE Transactions on Information Theory, vol. 60, no. 4, pp. 2331–2345, 2014. * [8] N. Nayyar, D. Kalathil, and R. Jain, “On regret-optimal learning in decentralized multiplayer multiarmed bandits,” IEEE Transactions on Control of Network Systems, vol. 5, no. 1, pp. 597–606, 2016. * [9] D. P. Bertsekas, “The auction algorithm: A distributed relaxation method for the assignment problem,” Annals of operations research, vol. 14, no. 1, pp. 105–123, 1988. * [10] O. Avner and S. Mannor, “Multi-user lax communications: a multi-armed bandit approach,” in IEEE INFOCOM 2016-The 35th Annual IEEE International Conference on Computer Communications, pp. 1–9, IEEE, 2016. * [11] I. Bistritz and A. Leshem, “Distributed multi-player bandits-a game of thrones approach,” in Advances in Neural Information Processing Systems, pp. 7222–7232, 2018. * [12] E. Boursier, V. Perchet, E. Kaufmann, and A. Mehrabian, “A Practical Algorithm for Multiplayer Bandits when Arm Means Vary Among Players,” arXiv e-prints, p. arXiv:1902.01239, Feb 2019. * [13] H. Liu, K. Liu, and Q. Zhao, “Learning in a changing world: Restless multiarmed bandit with unknown dynamics,” IEEE Transactions on Information Theory, vol. 59, no. 3, pp. 1902–1916, 2013. * [14] C. Tekin and M. Liu, “Online learning of rested and restless bandits,” IEEE Transactions on Information Theory, vol. 58, no. 8, pp. 5588–5611, 2012\. * [15] H. Liu, K. Liu, and Q. Zhao, “Learning in a changing world: Restless multiarmed bandit with unknown dynamics,” IEEE Transactions on Information Theory, vol. 59, no. 3, pp. 1902–1916, 2012. * [16] T. Gafni and K. Cohen, “Learning in restless multi-armed bandits using adaptive arm sequencing rules,” in Proc. of the IEEE International Symposium on Information Theory (ISIT), pp. 1206–1210, Jun. 2018. * [17] Z. Han, Z. Ji, and K. R. Liu, “Fair multiuser channel allocation for OFDMA networks using Nash bargaining solutions and coalitions,” IEEE Transactions on Communications, vol. 53, no. 8, pp. 1366–1376, 2005. * [18] I. Menache and N. Shimkin, “Rate-based equilibria in collision channels with fading,” IEEE Journal on Selected Areas in Communications, vol. 26, no. 7, pp. 1070–1077, 2008. * [19] U. O. Candogan, I. Menache, A. Ozdaglar, and P. A. Parrilo, “Competitive scheduling in wireless collision channels with correlated channel state,” in Game Theory for Networks, 2009. GameNets’ 09. International Conference on, pp. 621–630, 2009. * [20] I. Menache and A. Ozdaglar, “Network games: Theory, models, and dynamics,” Synthesis Lectures on Communication Networks, vol. 4, no. 1, pp. 1–159, 2011. * [21] L. M. Law, J. Huang, and M. Liu, “Price of anarchy for congestion games in cognitive radio networks,” IEEE Transactions on Wireless Communications, vol. 11, no. 10, pp. 3778–3787, 2012. * [22] K. Cohen, A. Leshem, and E. Zehavi, “Game theoretic aspects of the multi-channel ALOHA protocol in cognitive radio networks,” IEEE Journal on Selected Areas in Communications, vol. 31, pp. 2276–2288, 2013. * [23] H. Wu, C. Zhu, R. J. La, X. Liu, and Y. Zhang, “Fasa: Accelerated S-ALOHA using access history for event-driven M2M communications,” IEEE/ACM Transactions on Networking (TON), vol. 21, no. 6, pp. 1904–1917, 2013. * [24] C. Singh, A. Kumar, and R. Sundaresan, “Combined base station association and power control in multichannel cellular networks,” IEEE/ACM Transactions on Networking, vol. 24, no. 2, pp. 1065–1080, 2016. * [25] K. Cohen and A. Leshem, “Distributed game-theoretic optimization and management of multichannel aloha networks,” IEEE/ACM Transactions on Networking, vol. 24, no. 3, pp. 1718–1731, 2016. * [26] K. Cohen, A. Nedić, and R. Srikant, “Distributed learning algorithms for spectrum sharing in spatial random access wireless networks,” IEEE Transactions on Automatic Control, vol. 62, no. 6, pp. 2854–2869, 2017. * [27] D. Malachi and K. Cohen, “Queue and channel-based aloha algorithm in multichannel wireless networks,” IEEE Wireless Communications Letters, vol. 9, no. 8, pp. 1309–1313, 2020. * [28] M. Yemini, A. Leshem, and A. Somekh-Baruch, “Restless hidden markov bandits with linear rewards,” arXiv preprint arXiv:1910.10271, 2019. * [29] W. Wang and X. Liu, “List-coloring based channel allocation for open-spectrum wireless network,” In proc. of IEEE Vehic. Tech. Conf., 2005. * [30] J. Wang, Y. Huang, and H. Jiang, “Improved algorithm of spectrum allocation based on graph coloring model in cognitive radio,” in WRI International Conference on Communications and Mobile Computing, vol. 3, pp. 353–357, 2009\. * [31] A. Checco and D. Leith, “Learning-based constraint satisfaction with sensing restrictions,” IEEE Journal of Selected Topics in Signal Processing, vol. 7, pp. 811–820, Oct 2013. * [32] A. Checco and D. J. Leith, “Fast, responsive decentralised graph colouring,” arXiv preprint arXiv:1405.6987, 2014. * [33] H. Cao and J. Cai, “Distributed opportunistic spectrum access in an unknown and dynamic environment: A stochastic learning approach,” IEEE Transactions on Vehicular Technology, vol. 67, no. 5, pp. 4454–4465, 2018. * [34] A. Leshem and E. Zehavi, “Bargaining over the interference channel,” in IEEE International Symposium on Information Theory, pp. 2225–2229, 2006. * [35] I. Bistritz and A. Leshem, “Approximate best-response dynamics in random interference games,” IEEE Transactions on Automatic Control, vol. 63, no. 6, pp. 1549–1562, 2018. * [36] O. Naparstek and K. Cohen, “Deep multi-user reinforcement learning for dynamic spectrum access in multichannel wireless networks,” in IEEE Global Communications Conference (GLOBECOM), pp. 1–7, 2017. * [37] O. Naparstek and K. Cohen, “Deep multi-user reinforcement learning for distributed dynamic spectrum access,” IEEE Transactions on Wireless Communications, vol. 18, no. 1, pp. 310–323, 2019. * [38] D. Livne and K. Cohen, “PoPS: Policy Pruning and Shrinking for deep reinforcement learning,” IEEE Journal of Selected Topics in Signal Processing, vol. 14, no. 4, pp. 789–801, 2020. * [39] T. Sery and K. Cohen, “On analog gradient descent learning over multiple access fading channels,” IEEE Transactions on Signal Processing, vol. 68, pp. 2897–2911, 2020. * [40] K. Cohen and D. Malachi, “A time-varying opportunistic multiple access for delay-sensitive inference in wireless sensor networks,” IEEE Access, vol. 7, pp. 170475–170487, 2019. * [41] O. Naparstek and A. Leshem, “Fully distributed optimal channel assignment for open spectrum access,” IEEE Transactions on Signal Processing, vol. 62, no. 2, pp. 283–294, 2013. * [42] C. H. Papadimitriou and J. N. Tsitsiklis, “The complexity of optimal queuing network control,” Mathematics of Operations Research, vol. 24, no. 2, pp. 293–305, 1999. * [43] P. Auer, N. Cesa-Bianchi, Y. Freund, and R. E. Schapire, “The nonstochastic multiarmed bandit problem,” SIAM journal on computing, vol. 32, no. 1, pp. 48–77, 2002. * [44] A. Lesage-Landry and J. A. Taylor, “The multi-armed bandit with stochastic plays,” IEEE Transactions on Automatic Control, vol. 63, no. 7, pp. 2280–2286, 2017. * [45] P. Reverdy, V. Srivastava, and N. E. Leonard, “Satisficing in multi-armed bandit problems,” IEEE Transactions on Automatic Control, vol. 62, no. 8, pp. 3788–3803, 2016. * [46] Q. Zhao and L. Tong, “Opportunistic carrier sensing for energy-efficient information retrieval in sensor networks,” EURASIP Journal on Wireless Communications and Networking, vol. 2005, no. 2, pp. 231–241, 2005. * [47] P. Lezaud, “Chernoff-type bound for finite markov chains,” Annals of Applied Probability, pp. 849–867, 1998. * [48] V. Anantharam, P. Varaiya, and J. Walrand, “Asymptotically efficient allocation rules for the multiarmed bandit problem with multiple plays-part ii: Markovian rewards,” IEEE Transactions on Automatic Control, vol. 32, no. 11, pp. 977–982, 1987.
# Dense Suspension Flow in a Penny-Shaped Crack Part I : Theory George R. Wyatt1 & Herbert E. Huppert2 1Emmanuel College, St. Andrew’s Street, Cambridge, CB2 3AP<EMAIL_ADDRESS> 2King’s College, King’s Parade, Cambridge, CB2 1ST<EMAIL_ADDRESS> ###### Abstract. We study the dynamics of proppants carried by fluid driven into an evolving penny-shaped fracture. The behaviour of the slurry flow is investigated in two phases: pressurised injection and elastic closure. During injection the slurry is modelled using a frictional rheology that takes into account the shear- induced migration and jamming of the proppants. Making pragmatic assumptions of negligible toughness and cross-fracture fluid slip, we find self-similar solutions supporting a range of proppant concentration profiles. In particular, we define an effective viscosity, which equates the fracture evolution of a slurry flow with a given proppant volume fraction, to a Newtonian flow with a particular viscosity. Using this framework, we are able to make predictions about the geometry of the growing fracture and the significance of tip screen-out. In the closure phase, proppants are modelled as incompressible and radially immobile within the narrowing fracture. The effects of proppant concentration on the geometry of the residual propped fracture are explored in full. The results have important applications to industrial fracking and geological dike formation by hot, intruding magma. ###### Key words and phrases: Hydraulic fracture, suspension flow, rheology, proppant transport, elastic, tip screen-out, penny-shaped, cavity flow ## 1\. Introduction Receiving a patent for his ‘exploding torpedo’ in 1865, US Civil War veteran Col. Edward Roberts established the practice of fracturing bedrock to stimulate oil wells [1]. A technique, known as hydraulic fracturing, which uses pressurised fluid rather than explosives to develop fracture networks, only came into practice much later, in 1947 [2], and is the topic of this paper. In particular, we will concentrate on the convective transport of proppants within an evolving cavity. These are small particles added to the fracturing fluid in order to prop open the developed fracture, which closes under far-field stress once the fluid pressure is released. Aside from its use in hydrocarbon recovery, hydraulic fracturing, or fracking, has uses including the measurement of in-situ stresses in rocks [3], generation of electricity in enhanced geothermal systems [4] and improvement of injection rates in CO2 sequestration [5]. Hydraulic fracturing processes are also ubiquitous in geology: dikes and sills arise from cracks whose growth is driven by magma, with magmatic crystals taking the place of synthetic proppants. Phenomena such as crystallisation and gas exsolution in the cooling magma mean models of dike propagation vary widely, as is summarised in [6]. Notably, Petford & Koenders [7] utilise granular flow theory to model the ascent of a granitic melt containing solids. This paper combines two significant, but often disconnected, fields of fracking study, cavity flow and suspension flow: * • The study of (elastohydrodynamic) cavity flow focusses on the interplay between hydrodynamic properties of the fracturing fluid and material properties of the medium being fractured. In the zero-proppant case, the problem of a fluid-driven, penny-shaped crack requires the joint solution of a nonlinear Reynold’s equation, which governs flow within the crack, and a singular integral boundary condition, which takes into account the elastic properties of the surrounding medium. The general strategy used in this paper takes inspiration from the work of Spence & Sharp [8], who in 1985, restricting to the two-dimensional case, were the first to solve these integro-differential equations. In particular, we will focus on cavities that keep the same shape in some evolving coordinate system, using series expansions to represent both the width and pressure profiles within the fracture. More recently, in 2002, Savitski & Detournay [9] solved similar three-dimensional versions of these equations, allowing them to find fracture evolutions with simple time dependence in both the viscous and toughness dominated regimes. In the former, the principal energy dissipation is by viscous flow, and in the latter, energy dissipation is mostly by creating new fracture surfaces. Notably, the same paper [9] verifies that industrial fracking occurs in the viscous regime; this assumption makes the problem considered in this paper tractable to a semi-analytical approach. * • The mathematical study of suspension flow dates back to 1906, when Einstein used properties of suspensions to estimate the size of a water molecule [10]. In particular, he showed that very dilute particle-laden flows are Newtonian, with a viscosity which increases with the concentration of particles. However, during hydraulic fracturing it is necessary to model a full range of proppant volume fractions, which we denote by $\phi$. It is typical to have both dilute flow near the crack walls, as well as plug flow at the centre of the cavity, where the slurry behaves as a porous granular medium. More recent experiments by Boyer et al. in 2011 [11] investigate dense suspension rheology. They show that particles in suspension, subject to a constant normal particle pressure that is applied by a porous plate, expand when a shear is applied to the mixture. As a result, it is possible to write $\phi=\phi(I)$, where the dimensionless parameter, $I$, is the ratio between the fluid shear stress, which is proportional to the shear rate, and the particle normal stress. Likewise, fixing the solid volume fraction, they showed that the normal particle pressure is proportional to the mixture shear stress. It is also shown that the constant of proportionality, $\mu$, can be expressed as a decreasing function of $\phi$. In the same paper [11], forms of the rheological functions $I$ and $\mu$ are suggested, showing good agreement with experimental data. Since then, several papers have suggested slightly different rheological models and are reviewed by Donstov et al. in [12]. These all feature a jamming limit, $\phi_{m}$, which is the volume fraction at which the flowing slurry transitions into a granular solid. We will utilise the frictional rheology given by Lecampion & Garagash [13], which is unique in allowing packings with $\phi>\phi_{m}$. These denser packings form due to ‘in- cage’ particle rearrangements caused by velocity and pressure fluctuations in the surrounding flow. The endeavours of this paper may be condensed into three main objectives. The first is to establish a mathematical framework that captures the behaviour of the proppant suspension as it interacts with the growing cavity. Here we will utilise a lubrication model, along with the assumption that the proppant flow is fully developed; equivalently, that the transverse fluid slip is negligible. Crucially, we will try to justify these assumptions using typical parameters from industrial fracking. We will also make a zero-toughness assumption, which is validated in [9]. Once we have developed this framework, an important step will be to compare its features to those derived in the zero-proppant, viscosity dominated case by Savitski & Detournay [9], particularly because we utilise a frictional rheology fitted to the dense regime. The second objective is to find and examine accurate numerical solutions modelling the developing cavity, given a range of proppant concentrations. We will explore the empirical effects of changing proppant concentration on the geometry of the developing fracture, as well as the distribution of proppants. Where possible, we will evaluate the consistency of our model and forecast potential shortfalls such as proppant screen-out near the crack tip. The third, and final, objective is to leverage our results to make predictions about the geometry of the fracture after the fluid pressure is released. By assuming the remaining proppants are immobile and incompressible, we aim to establish simple formulae predicting the width and radius of the developed fracture. Since these relate directly to the conductivity of the formation, this third objective is potentially the most significant. Aside from the availability of semi-analytical solutions, the problem of proppant flow in a penny-shaped crack is particularly appealing because of the potential of practical verification. Recent experiments by O’Keeffe, Huppert & Linden [14] have explored fluid-driven, penny-shaped fractures in transparent, brittle hydrogels, making use of small particle concentrations to measure in- crack velocities. This paper is the first of two; the second of which will be a practical treatise on slurry driven-fractures in hydrogels, aiming to verify the predictions made here by repeating the experiments of [14] including proppant concentrations. ## 2\. Injection: Problem Formulation Figure 1. Schematic of the penny-shaped crack. ### 2.1. Fracture Mechanics We model the propagation of a penny-shaped crack similar to that shown in Figure 1, using the framework of Detournay & Savitski [9]. We will make the following assumptions: * • The crack is axisymmetric and has reflectional symmetry in $z=0$, with half width $w(r,t)$ and total radius $R(t)$, so $w(R,t)=0$. * • The fluid is injected from a point source, with the wellbore radius negligible compared to the fracture radius. * • The lag between the fracture tip and the fluid front is negligible compared to the fracture radius. * • The fracture propagates in continuous mobile equilibrium. * • The normal stress on the fracture walls due to proppants is negligible compared to the fluid pressure. The third assumption is validated by Garagash & Detournay [15] and introduces a negative pressure singularity at the tip of the crack ($r=R$). The fourth and fifth assumptions lead to the following integral equations from linear elastic fracture mechanics. These relate the net fluid pressure, $p(r,t)$, to the opening of the fracture and the toughness of the surrounding rock. (1) $\displaystyle w(r,t)$ $\displaystyle=\frac{4R}{\pi E^{\prime}}\int_{r/R}^{1}\frac{y}{\sqrt{y^{2}-(r/R)^{2}}}\int_{0}^{1}\frac{xp(xyR,t)}{\sqrt{1-x^{2}}}dxdy,$ (2) $\displaystyle K_{Ic}$ $\displaystyle=\frac{2}{\sqrt{\pi R}}\int_{0}^{R}\frac{p(r,t)r}{\sqrt{R^{2}-r^{2}}}dr,$ where $E^{\prime}$ is the plane strain modulus, given by the Young modulus, $E$, and the Poisson ratio, $\nu$, as $E^{\prime}=E/(1-\nu^{2})$. $K_{Ic}$ is the material toughness. These equations can be attributed to Sneddon [16] and Rice [17] respectively. We note that $p$ represents the fluid pressure minus the in-situ stress of the surrounding rock, which is assumed to be isotropic. We write $p$ with radial spatial dependence only; this will be validated later, along with the fifth assumption, using a lubrication argument. ### 2.2. Frictional Rheology We model the injected flow as a Newtonian fluid containing identical spherical particles. Recent approaches in modelling dense slurry flow are characterised by empirical relations originally proposed by Boyer et al. [11]. The first of these relates the fluid shear stress to the normal stress required to confine the particles; the second gives the ratio of the mixture shear stress to the particle confining stress, (3) $\displaystyle I(\phi)$ $\displaystyle=\eta_{f}\dot{\gamma}/\sigma_{n}^{s},$ $\displaystyle\mu(\phi)$ $\displaystyle=\tau/\sigma_{n}^{s}.$ Here $\eta_{f}$ is the carrying fluid’s dynamic viscosity, $\phi$ is the volume fraction of the proppants, $\dot{\gamma}$ is the solid shear rate and $\sigma_{n}^{s}$ is the normal particle stress, which we will sometimes refer to as the particle pressure. The second ratio is given the symbol $\mu$, not to be confused with dynamic viscosity, because it resembles a friction coefficient. These relations are given a clear experimental grounding in [11], which is discussed in the introduction. Various forms of the dimensionless functions $I(\phi)$ and $\mu(\phi)$ have been compared to experimental results in [12] using the equivalent formulation: $\tau=\eta_{s}(\phi)\eta_{f}\dot{\gamma}$ and $\sigma_{n}=\eta_{n}(\phi)\eta_{f}\dot{\gamma}$, where $\eta_{s}=\mu(\phi)/I(\phi)$ and $\eta_{n}=1/{I(\phi)}$. In our calculations we will utilise the frictional rheology provided by B. Lecampion & D. I. Garagash [13], which is unique in allowing packings with volume concentrations greater than $\phi_{m}$. Here $I(\phi)=0$, meaning the proppants have zero shear rate and effectively resemble a permeable solid. Explicitly, we use the expressions (4) $\displaystyle\mu=\mu_{1}+\frac{\phi_{m}}{\delta}\left(1-\frac{\phi}{\phi_{m}}\right)$ $\displaystyle+\left(I(\phi)+\left[\frac{5}{2}\phi_{m}+2\right]I(\phi)^{0.5}\right)\left(1-\frac{\phi}{\phi_{m}}\right)^{2},$ (7) $\displaystyle I(\phi)$ $\displaystyle=\left\\{\begin{array}[]{rl}\left(\phi_{m}/\phi-1\right)^{2}&\textrm{ if }\phi<\phi_{m}\\\ 0&\textrm{ if }\phi\geq\phi_{m},\end{array}\right.$ where $\phi_{m}=0.585$, $\mu_{1}=0.3$ and $\delta=0.158$; these are plotted in Figure 2. We might have used a different rheology, but this model shows good agreement with the data of Boyer et al. [11] and Dagois-Bohy et al. [18] for $0.4<\phi<\phi_{m}$. Furthermore, owing to its linear extension beyond $\phi_{m}$, $\mu$ is a simple monotonic function, meaning we can invert it easily to find $\phi$. In other models $\phi(\mu)$ is constant for $\mu<\mu(\phi_{m})$; this means that $\phi_{m}$ is the maximum volume fraction, regardless of how small shear stresses in the jammed slurry become. An important observation is that $\mu=0$ implies $\phi=\phi_{m}+\delta\mu_{1}\approx 0.63\approx\phi_{rcp}$. Here $\phi_{rcp}$ is the random close packing limit, the maximal observed volume fraction due to random packing. This reflects the fact that, for a given confining stress, as the shear stress tends to zero, the particles pack to this maximal density. This rheology uses a continuum model that requires particles to be small compared to the size of the fracture. This is in order to well-define the proppant volume concentration, $\phi$. In our model the relevant ratio is that of the particle diameter to the typical crack width, the smallest cavity length scale. In [13], good results are obtained using the same rheological model, with this ratio taking values as large as $1/10$. However, as the ratio approaches unity we have to consider non-local effects, such as proppant bridging across the crack width. This is particularly important near the fracture tip, where $w$ approaches zero. These effects will be discussed in greater detail in Section 7, once we have formed a model of the evolving fracture. We must also be cautious applying these rheological models to dilute flows, since they are fitted to experimental data from the dense regime, where $\phi>0.4$. This difficulty is somewhat inevitable, since the determination of $I$ and $\mu$ requires measurement of the particle normal stress, or particle pressure, which becomes very small in the dilute regime. (a) $I$ (b) $\mu$ (c) $I/\mu$ (d) $I/\mu$ data Figure 2. Plots of the rheological functions $I$, $\mu$ and $I/\mu$ given by Lecampion & Garagash [13]. Also plotted is the experimental data of Boyer et al. [11] using polystyrene spheres of diameter 580$\mu$m in $2.15$Pa s fluid (red), as well as poly(methyl methacrylate) spheres of diameter 1100$\mu$m suspended in $3.10$Pa s fluid (orange); and of Dagois-Bohy et al. [18] using polystyrene spheres of diameter 580$\mu$m suspended in $2.27$Pa s fluid (purple). All experiments are carried out with a fixed particle pressure, applied by a porous plate. ### 2.3. Fluid Slip We define $\mathbf{u}$ as the slurry velocity, $\mathbf{v}$ as the particle velocity and $\mathbf{q}=\mathbf{u}-\mathbf{v}$ as the slip velocity. We then employ the slip relation (8) $\displaystyle\mathbf{q}$ $\displaystyle=\frac{a^{2}\kappa(\phi)}{\eta_{f}}\nabla\cdot\sigma^{f},$ (9) $\displaystyle\kappa(\phi)$ $\displaystyle=\frac{2(1-\phi)^{5.1}}{9\phi},$ where $a$ is the particle radius and $\sigma_{f}$ is the fluid stress tensor. Since fluid and particle shear rates are often similar, we ignore fluid shear stresses and take $\sigma^{f}=-pI$; this is typical in the analysis of porous media flow. This simplifies (8) to Darcy’s law. However, the effect of fluid shear stress is taken into account in the frictional rheology, where it is included as part of the solid shear stress. $\kappa$ is a normalised form of the permeability of the solid particles; we use the function suggested by Garside & Al-Dibouni [19], which is based on the phenomenology first described by Richardson & Zaki [20]. This choice of permeability function shows excellent agreement with the experimental results of Bacri et al. [21]. ### 2.4. Conservation Equations We consider the effective Reynolds number, (10) $\displaystyle\textrm{Re}_{\textrm{eff}}=\frac{\rho u_{r}w^{2}}{\eta_{f}R},$ to be negligible. We also neglect the effect of gravity, since we are mainly concerned with small or neutrally buoyant proppants, which settle slowly. Hence, our momentum balance becomes (11) $\displaystyle\nabla\cdot\sigma=0,$ where $\sigma=\sigma^{s}+\sigma^{f}$ is the mixture stress tensor, composed of the particle and fluid stresses respectively. We also note that, subtracting the hydrostatic pressure term, we write $\sigma=\tau-pI$. Since we assumed $\sigma^{f}=-pI$ in deriving the fluid slip equation, we deduce $\sigma_{s}=\tau$. This is a notational quirk arising from the frictional rheology because $\tau$ does include shear stress originating from the viscous carrier fluid. Herein we will refer to $\sigma^{s}_{zz}$ and $\tau_{rz}$, since the former generally arises from the proppants and the latter stems from both the proppants and the carrier fluid. The assumption of axisymmetry gives (12) $\displaystyle\frac{1}{r}\frac{\partial(r\tau_{rr})}{\partial r}+\frac{\partial\tau_{rz}}{\partial z}-\frac{\partial p}{\partial r}$ $\displaystyle=0,$ $\displaystyle\frac{1}{r}\frac{\partial(r\tau_{rz})}{\partial r}+\frac{\partial\sigma^{s}_{zz}}{\partial z}-\frac{\partial p}{\partial z}$ $\displaystyle=0.$ We also have the continuity equations (13) $\displaystyle\nabla\cdot(\mathbf{v}+\mathbf{q})$ $\displaystyle=0,$ $\displaystyle\frac{\partial\phi}{\partial t}+\nabla\cdot(\phi\mathbf{v})$ $\displaystyle=0.$ The first of these can be integrated over the fracture volume to give $Qt=4\pi\int_{0}^{R}rw(r,t)dr.$ Here, $Q$ is the rate at which the slurry is pumped into the crack, which we will assume is constant. We will also assume that the proppants are injected at a constant rate, meaning the average concentration at the wellbore is constant. ## 3\. Injection: Scalings To help implement the assumptions of a lubrication model, where the crack width is far smaller than the crack radius, we introduce the scaled coordinates, $\displaystyle T$ $\displaystyle=T(t),$ $\displaystyle r$ $\displaystyle=L(t)\Gamma(T)\xi,$ $\displaystyle z$ $\displaystyle=\epsilon(t)L(t)\eta.$ Here $T(t)$ is the internal time scale, a monotonic function to be specified later; $\epsilon(t)$ is a small number; and $\Gamma(T)$ is the crack radius, measured in the scaled coordinates, so $\xi=1$ implies $r=R$. We multiply the variables accordingly, $\displaystyle w(r,t)$ $\displaystyle\to\epsilon Lw(\xi,T),$ $\displaystyle p(r,z,t)$ $\displaystyle\to\epsilon E^{\prime}p(\xi,\eta,T),$ $\displaystyle R(t)$ $\displaystyle\to L\Gamma(T),$ $\displaystyle v_{z}(r,z,t)$ $\displaystyle\to-\dot{\epsilon}Lv_{z}(\xi,\eta,T),$ $\displaystyle v_{r}(r,z,t)$ $\displaystyle\to\frac{-\dot{\epsilon}L}{\epsilon}v_{r}(\xi,\eta,T),$ $\displaystyle q_{r}(r,z,t)$ $\displaystyle\to\frac{\epsilon}{L}\frac{a^{2}E^{\prime}}{\eta_{f}\Gamma}q_{r}(\xi,\eta,T),$ $\displaystyle q_{z}(r,z,t)$ $\displaystyle\to\frac{1}{L}\frac{a^{2}E^{\prime}}{\eta_{f}}q_{z}(\xi,\eta,T),$ $\displaystyle\tau(r,z,t)$ $\displaystyle\to-\frac{\dot{\epsilon}}{\epsilon^{2}}\eta_{f}\tau(\xi,\eta,T),$ $\displaystyle\sigma^{s}(r,z,t)$ $\displaystyle\to-\frac{\dot{\epsilon}}{\epsilon^{2}}\eta_{f}\sigma^{s}(\xi,\eta,T).$ The appearance of minus signs reflects the fact that $\epsilon$, the ratio of the characteristic radius to the characteristic width of the fracture, is decreasing. We also assume the scaling is suitable so that all the scaled variables are $\mathcal{O}(1)$. Herein, we will use $(\dot{})$ for derivatives with respect to $t$ and $(^{\prime})$ for those with respect to $T$. In the new, rescaled coordinates the equations describing the frictional rheology become $I(\phi)=\dot{\gamma}/\sigma_{n}^{s}$ and $\mu(\phi)=\tau/\sigma_{n}^{s}$. The slip equation becomes $\mathbf{q}=-\kappa(\phi)\nabla p,$ where $\nabla$ is now with respect to $(\xi,\eta)$. The integral equations become (14) $\displaystyle w(\xi,T)$ $\displaystyle=\frac{4\Gamma}{\pi}\int_{\xi}^{1}\frac{y}{\sqrt{y^{2}-\xi^{2}}}\int_{0}^{1}\frac{xp(xy,T)}{\sqrt{1-x^{2}}}dxdy,$ $\displaystyle\aleph\equiv\frac{K_{Ic}}{\epsilon E^{\prime}\sqrt{L}}$ $\displaystyle=2\sqrt{\frac{\Gamma}{\pi}}\int_{0}^{1}\frac{p(\xi,T)\xi}{\sqrt{1-\xi^{2}}}d\xi.$ The momentum equations are (15) $\displaystyle\frac{\epsilon}{\Gamma\xi}\frac{\partial(\xi\tau_{rr})}{\partial\xi}+\frac{\partial\tau_{rz}}{\partial\eta}+\frac{\epsilon^{3}E^{\prime}t}{\eta_{f}}\frac{\epsilon}{\dot{\epsilon}t\Gamma}\frac{\partial p}{\partial\xi}$ $\displaystyle=0,$ $\displaystyle\frac{\epsilon^{2}}{\Gamma\xi}\partialderivative{(\xi\tau_{rz})}{\xi}+\epsilon\partialderivative{\sigma^{s}_{zz}}{\eta}+\frac{\epsilon}{\dot{\epsilon}t}\frac{\epsilon^{3}E^{\prime}t}{\eta_{f}}\partialderivative{p}{\eta}$ $\displaystyle=0.$ Since we expect the radial pressure gradient to be comparable to the shear stress, $\tau_{rz}$, we choose $\epsilon$ so that the dimensionless quantity $\epsilon^{3}E^{\prime}t/\eta_{f}=1$. Finally, the global volume conservation equation then becomes $Qt/(\epsilon L^{3})=4\pi\Gamma^{2}\int_{0}^{1}\xi w(\xi,T)d\xi,$ so in a similar manner we choose the dimensionless quantity $Qt/\epsilon L^{3}=1.$ These choices mean (16) $\displaystyle\epsilon(t)$ $\displaystyle=(\eta_{f}/E^{\prime})^{\frac{1}{3}}t^{-1/3},$ $\displaystyle L(t)$ $\displaystyle=(E^{\prime}Q^{3}/\eta_{f})^{\frac{1}{9}}t^{4/9}.$ We will repeatedly use the relations $\dot{\epsilon}t/\epsilon=-1/3$ and $\dot{L}t/L=4/9$. Using this choice of $\epsilon$ we note that, before scaling, $\sigma^{s}/p=\mathcal{O}(\epsilon)$; this validates the assumption that particle pressure is negligible compared to hydrostatic pressure at the crack walls. Also, by the scaled momentum equations, (17) $\displaystyle\frac{\partial\tau_{rz}}{\partial\eta}$ $\displaystyle=\frac{3}{\Gamma}\frac{\partial p}{\partial\xi}+\mathcal{O}(\epsilon),$ $\displaystyle\frac{\partial p}{\partial\eta}$ $\displaystyle=\frac{\epsilon}{3}\frac{\partial\sigma^{s}_{zz}}{\partial\eta}+\mathcal{O}(\epsilon^{2}),$ the second of which verifies the assumption that $p$ has spatial dependence in the radial direction only. Because of the $\eta=0$ reflectional symmetry, we note that $\tau_{rz}(\xi,0)=0$. So, ignoring $\mathcal{O}(\epsilon)$ terms and integrating (17.1), we see that (18) $\displaystyle\tau_{rz}=\frac{3\eta}{\Gamma}\frac{\partial p}{\partial\xi},$ and, using the scaled equations from the frictional rheology, (19) $\displaystyle\sigma_{zz}^{s}$ $\displaystyle=\frac{3|\eta|}{\Gamma}\frac{1}{\mu(\phi)}\frac{\partial p}{\partial\xi},$ $\displaystyle\frac{\partial v_{r}}{\partial\eta}$ $\displaystyle=\frac{3\eta}{\Gamma}\frac{I(\phi)}{\mu(\phi)}\frac{\partial p}{\partial\xi}.$ Then, using the condition $v_{r}(\xi,\pm w)=0$, we deduce that (20) $\displaystyle v_{r}(\xi,\eta)=-\frac{3}{\Gamma}\partialderivative{p}{\xi}\int_{\eta}^{w}\frac{I(\phi)\eta}{\mu(\phi)}d\eta.$ ## 4\. Injection: Time Regimes In this choice of scaling, the slurry conservation equation becomes (21) $\displaystyle\frac{1}{3\Gamma\xi}\frac{\partial(\xi v_{r})}{\partial\xi}+\frac{1}{3}\frac{\partial v_{z}}{\partial\eta}+\left(\frac{a}{L\Gamma}\right)^{2}\frac{1}{\epsilon^{2}\xi}\frac{\partial(\xi q_{r})}{\partial\xi}+\left(\frac{a}{L}\right)^{2}\frac{1}{\epsilon^{4}}\frac{\partial q_{z}}{\partial\eta}=0.$ Combining this with the scaled slip equation, noting (17), we obtain (22) $\displaystyle\frac{1}{3\Gamma\xi}\frac{\partial(\xi v_{r})}{\partial\xi}+\frac{1}{3}\frac{\partial v_{z}}{\partial\eta}-\frac{\epsilon\lambda}{\Gamma^{2}\xi}\partialderivative{\xi}\left[\xi\kappa(\phi)\partialderivative{p}{\xi}\right]-\frac{\lambda}{3}\partialderivative{\eta}\left[\kappa(\phi)\partialderivative{\sigma^{s}_{zz}}{\eta}\right]=0.$ Here $\lambda=a^{2}/(L^{2}\epsilon^{3})$ is a constant; we will later identify it as the ratio of the fracture length scale to the development length scale, over which we expect proppant flow to stabilise. According to Shiozawa & McClure [22], Chen Zhixi et al. [23] and Liang et al. [24], we utilise the following constants, relevant to hydraulic fracturing, as given in Table 1. Constant | Typical Value ---|--- $Q$ | $0.04\textrm{m}^{3}\textrm{ s}^{-1}$ $E^{\prime}$ | $40\textrm{ GPa}$ $\eta_{f}$ | $0.01\textrm{ Pa s}$ $\rho_{f}$ | $1000\textrm{ kg m}^{-3}$ $K_{Ic}$ | $0.5\textrm{ MPa m}^{0.5}$ $a$ | $5\times 10^{-5}\textrm{m}$ Table 1. Typical values of constants, given by Shiozawa & McClure [22], Chen Zhixi et al. [23] and Liang et al. [24]. The choice of $a$ represents a typical diameter for the finer proppants commonly used at the initiation of fracturing [24]. This gives us the following estimates $\displaystyle\epsilon$ $\displaystyle\approx 6\times 10^{-5}\cdot t^{-1/3},$ $\displaystyle L$ $\displaystyle\approx 9\times 10^{0}\cdot t^{4/9},$ $\displaystyle\textrm{Re}_{\textrm{eff}}$ $\displaystyle\approx 1\times 10^{-2}\cdot t^{-7/9},$ $\displaystyle\aleph$ $\displaystyle\approx 4\times 10^{-2}\cdot t^{1/9},$ $\displaystyle\lambda$ $\displaystyle\approx 1\times 10^{2}\cdot t^{1/9},$ $\displaystyle a/(\epsilon L)$ $\displaystyle\approx 1\times 10^{-1}\cdot t^{-1/9}.$ The value of $\textrm{Re}_{\textrm{eff}}$ is calculated using formula (10), substituting each term with its typical scaling. Considering the same problem in the zero-proppant case, Detournay & Savitski [9] show that when $1.6\aleph<1$, the fracture evolution is well approximated by taking the dimensionless toughness $\aleph=0$. Also, the choice $T=\aleph$ is taken, reflecting the dependence of the scaled solution on this monotonically increasing parameter; assuming $\aleph$ is negligible it is possible to neglect any $T$ dependence. We will also use these assumptions, since toughness plays its greatest role near the fracture tip, where the crack is typically too narrow for proppants to interfere. Given our estimate for $\aleph$, this means we must take $t<1.5\times 10^{7}$. In general we will assume $t>250$, so we may ignore $\epsilon$ and $\textrm{Re}_{\textrm{eff}}$ terms. This also means $2a/(\epsilon L)<1/10$, so the fracture is typically more than 10 particles wide. Lecampion & Garagash [13], conclude that non-local phenomena such as proppant-bridging aren’t important in such cases; however we can still expect to see these effects near the narrow crack tip. The significance of this behaviour will be discussed in greater detail in Section 7. We also note that $\lambda$ is large; so in an effort to remove time dependence from our equations, we may neglect the first three terms in the continuity equation (22), (23) $\displaystyle\partialderivative{\eta}\left[\kappa(\phi)\partialderivative{\sigma^{s}_{zz}}{\eta}\right]=0.$ By the assumption of reflectional symmetry, the particle pressure gradient must vanish at $\eta=0$. Because $\kappa$ is generally non-zero, we deduce that the particle pressure is constant with $\eta$; and, by (19), so is $|\eta|/\mu(\phi)$. Hence, (24) $\displaystyle\phi(\xi,\eta)=\mu^{-1}\left(\mu_{w}(\xi)\frac{|\eta|}{w(\xi)}\right),$ where $\mu_{w}$ is an undetermined function of $\xi$, which we recognise as the value of $\mu$ at the crack wall. Noting that $\mu$ is a decreasing function, we see that $\mu_{w}$ also describes the rate at which the concentration drops from the centre to the wall of the cavity. We also notice that, in accordance to Donstov et al. [25], we have plug flow in the centre of the channel, where concentrations are greater than $\phi_{m}$. Because the slurry flows away from the wellbore, the distribution of proppants, which is described by $\mu_{w}$, depends on the concentration of proppants in the injected mixture and how that changes with time. Hence, an important step in the determination of $\mu_{w}$ will be implementing the assumption that the average concentration at the wellbore is constant. This will be discussed in greater detail in Section 7. It is interesting to note that [13] verifies a length scale of $\epsilon^{3}L^{3}/a^{2}$ for proppant flow in a channel, or pipe, to become fully established. This means the particle pressure gradient becomes negligible, and the cross fracture concentration profile becomes independent of the distance from the channel, or pipe, entrance. As a result, the constant $\lambda=a^{2}/(L^{2}\epsilon^{3})$ can be interpreted as the ratio of the fracture length to the development length. Because this is large, an alternative route to (24) would have been to assume the transverse particle pressure is constant, reflecting the full development of the flow. ## 5\. Injection: Governing Equation for fracture width In scaled coordinates, the governing equation for the conservation of proppant mass becomes (25) $\displaystyle\frac{\xi\dot{L}t}{L}\frac{\partial\phi}{\partial\xi}+\left[\frac{\dot{\epsilon}t}{\epsilon}+\frac{\dot{L}t}{L}\right]\eta\frac{\partial\phi}{\partial\eta}=-\frac{\dot{\epsilon}t}{\epsilon\Gamma\xi}\partialderivative{(\xi\phi v_{r})}{\xi}-\frac{\dot{\epsilon}t}{\epsilon}\partialderivative{(\phi v_{z})}{\eta}.$ Then, implementing our choices of $\epsilon$ and $L$, we obtain (26) $\displaystyle\frac{4\xi}{3}\frac{\partial\phi}{\partial\xi}+\frac{\eta}{3}\frac{\partial\phi}{\partial\eta}=\frac{1}{\Gamma\xi}\partialderivative{(\xi\phi v_{r})}{\xi}+\partialderivative{(\phi v_{z})}{\eta}.$ Integrating from $-w$ to $w$ with respect to $\eta$, leaving details to Appendix A for brevity, we obtain (27) $\displaystyle 4\xi\partialderivative{\xi}\left[w\Pi\circ\mu_{w}(\xi)\right]-w\Pi\circ\mu_{w}(\xi)=-\frac{9}{\Gamma^{2}\xi}\partialderivative{\xi}\left[\frac{\xi w^{3}}{\mu_{w}(\xi)^{2}}\partialderivative{p}{\xi}\Omega\circ\mu_{w}(\xi)\right].$ Here we have defined the rheological functions (28) $\displaystyle\Pi(x)$ $\displaystyle=\frac{1}{x}\int_{0}^{x}\mu^{-1}(u)du,$ $\displaystyle\Omega(x)$ $\displaystyle=\frac{1}{x}\int_{0}^{x}[\Pi(u)I\circ\mu^{-1}(u)u]du,$ which we plot in Figure 3. (a) $\Pi(x)$ as a function of $x$ (b) $\Omega(x)$ as a function of $x$ (c) $x^{2}\Pi/\Omega$ as a function of $\Pi(x)$ Figure 3. Plots of the rheological functions $\Omega$, $\Pi$ and $x^{2}\Pi/\Omega$. Multiplying by $\xi$ and integrating from $\rho$ to $1$, we obtain (29) $\displaystyle\int_{\rho}^{1}\xi w\Pi\circ\mu_{w}(\xi)d\xi+\frac{4}{9}\rho^{2}w\Pi\circ\mu_{w}(\rho)=-\frac{\rho w^{3}}{\Gamma^{2}\mu_{w}^{2}}\partialderivative{p}{\rho}\Omega\circ\mu_{w}(\rho),$ which lends itself more easily to computation. Here we have taken $w^{3}\partial p/\partial\xi\to 0$ as $\xi\to 1$; this is physically motivated by the fact that this term is proportional to the radial flux, which vanishes at the crack tip. Moreover, Spence & Sharp [8] show that, in the zero- proppant, zero-toughness regime, near the crack tip, $p\propto(1-\xi)^{-1/3}$ and $w\propto(1-\xi)^{2/3}$. In order to compare this equation to the zero-proppant case, we assume $\mu_{w}$ is independent of $\xi$ and take $\mu_{w}\to\infty$, to obtain (30) $\displaystyle\int_{\rho}^{1}\xi w(\xi)d\xi+\frac{4}{9}\rho^{2}w=-\frac{\rho w^{3}}{\Gamma^{2}}\partialderivative{p}{\rho}\lim_{\mu_{w}\to\infty}\left[\frac{\Omega(\mu_{w})}{\mu_{w}^{2}\Pi(\mu_{w})}\right].$ From Figure 3(c) we deduce the right hand limit is approximately $2/5$, which is confirmed exactly in Appendix B. Modelling the fluid as Newtonian, also leaving the details to Appendix B, we obtain the same equation, with a factor of $1/3$ instead. We conclude that the equations governing Newtonian flow are not the same as those in the zero-proppant slurry flow limit. This is clearly a limitation of our approach, which arises from using a dense-fitted rheology in the dilute regime. However, the fact that the equations share a nearly identical form is promising, as we expect the qualitative behaviour of slurry flow to be similar to that of Newtonian flow. ## 6\. Injection: Numerical Solution We implement the numerical method first used by Spence & Sharp [8], with the adaptions of Detournay & Savitski [9], to solve the equations we have derived so far. It will be useful to introduce $h(\xi)=w(\xi)/\Gamma$. The lubrication equation derived above, the elasticity equations and the global volume conservation equation become (31) $\displaystyle\int_{\rho}^{1}(\xi h\Pi\circ\mu_{w})d\xi$ $\displaystyle+\frac{4}{9}\rho^{2}h\Pi\circ\mu_{w}=-\rho h^{3}\partialderivative{p}{\rho}\frac{\Omega\circ\mu_{w}}{\mu_{w}^{2}},$ (32) $\displaystyle h(\xi)$ $\displaystyle=\frac{4}{\pi}\int_{\xi}^{1}\frac{y}{\sqrt{y^{2}-\xi^{2}}}\int_{0}^{1}\frac{xp(xy)}{\sqrt{1-x^{2}}}dxdy,$ (33) $\displaystyle 0$ $\displaystyle=\int_{0}^{1}\frac{p(\xi)\xi}{\sqrt{1-\xi^{2}}}d\xi,$ (34) $\displaystyle 1$ $\displaystyle=4\pi\Gamma^{3}\int_{0}^{1}(\xi h)d\xi.$ These equations alone do not give unique solutions for $\\{p,h,\mu_{w}\\}$, so we will prescribe $\mu_{w}$ as part of the problem data. This allows us to uniquely determine a solution for $\\{p,h\\}$. We seek series approximations of the form (35) $\displaystyle p(\xi)$ $\displaystyle=\sum_{i=-1}^{N-1}A_{i}p_{i}(\xi),$ $\displaystyle h(\xi)$ $\displaystyle=\sum_{i=-1}^{N}B_{i}h_{i}(\xi),$ where we define (42) $\displaystyle p_{i}(\xi)$ $\displaystyle=\left\\{\begin{array}[]{ll}-\ln\xi+\ln 2-1&(i=-1)\\\ &\\\ (1-\xi)^{-1/3}J_{i}(\frac{4}{3},2,\xi)+\omega_{i}&(i\geq 0)\end{array}\right\\},$ $\displaystyle h_{i}(\xi)=\left\\{\begin{array}[]{ll}\frac{4}{\pi}\left[(1-\xi^{2})^{1/2}-\xi\cos^{-1}(\xi)\right]&(i=-1)\\\ &\\\ \\\ (1-\xi)^{2/3}J_{i}(\frac{10}{3},2,\xi)&(i\geq 0)\end{array}\right\\}.$ Here the $i=-1$ terms are used to account for the logarithmic singularity in pressure at the inlet, expected as a result of the point source injection; the other terms allow for a general solution of (32). Importantly, we note that the $p_{i}$ terms have a $(1-\xi)^{-1/3}$ singularity near the crack tip and the $h_{i}$ terms are proportional to $(1-\xi)^{2/3}$ (for $i\geq 0$). This deliberately matches the asymptotic calculations from Spence & Sharp [8], which arise from the assumptions of zero-lag and zero-toughness in an expanding hydraulic fracture. This allows the numerical method to converge accurately with few terms. The $J_{i}(p,q,\xi)$ are Jacobi Polynomials of order $i$ defined on the interval $[0,1]$, in the sense defined by Abramowitz & Stegun [26], normalised to satisfy the orthonormality condition, (43) $\displaystyle\int_{0}^{1}(1-\xi)^{p-q}\xi^{q-1}J_{i}(p,q,\xi)J_{j}(p,q,\xi)d\xi=\delta_{ij}.$ This means that the $h_{i}$ ($i\geq 0$) are orthonormal with respect to an inner product weighted by $\xi$. The $\omega_{i}$ are simply constants to ensure each of the $p_{i}$ obey the zero-toughness equation; adding these constants means that the $p_{i}$ lose their orthonormality properties, however this doesn’t affect the solution finding process. Because of its linearity, these series approximations reduce (32) to a linear equation, (44) $\displaystyle B_{i}=\sum_{j=-1}^{N-1}P_{ij}A_{j}.$ Here $(P)_{ij}$ is an $(N+2)\times(N+1)$ matrix whose entries we only have to calculate once by using the orthogonality relation given above, along with the fact that $\\{p_{-1},\theta_{-1}\\}$ are a solution pair to (32). The entries of $M$, which can be found in [9], are listed in Appendix C for $N=4$. The subtleties of calculating elements of $P_{ij}$, in the face of strong singular behaviour, are important and described in depth in [9]. Finally, using the values of $B_{i}$ given above, we assign a cost to each choice of $A$ given by (45) $\displaystyle\Delta(A)=\sum_{\xi\in\\{0,1/M,...,1\\}}\left(\frac{\textrm{RHS}(\xi;A)}{\textrm{LHS}(\xi;A)}-1\right)^{2}.$ This is calculated by considering the discrepancies between the left and right hand sides of (31), calculated at M+1 equally spaced control points. We then minimise $\Delta$ with respect to $A$ using the Nelder-Mead Simplex method [27]. ## 7\. Injection: Solutions for a constant $\mu_{w}$ For most monotonic choices of $\mu_{w}$, the numerical method above shows good convergence. We see that the coefficients $A_{i}$ and $B_{i}$ drop off quickly with $i$, and the final value of $\Delta$ tends to zero rapidly as we increase $N$. If $\mu_{w}$ is a more complicated function, like in the case of Figure 4, we may need to use a larger value of $N$, but good convergence is still possible. Figure 4. Plot of cavity width profile and proppant distribution in the case where $\mu_{w}$ is sinusoidal. Here $N=8$ is used. This leads us to consider which choices of $\mu_{w}$ are most likely to appear in reality. We note that by (24), (46) $\displaystyle\Pi\circ\mu_{w}(\xi)=\frac{1}{2w}\int_{-w}^{w}\phi(\xi,\eta)d\eta,$ so we may view $\Pi\circ\mu_{w}(\xi)$ as the average proppant concentration at a given value of $\xi$. Since $\Pi\circ\mu_{w}$ is independent of time, we automatically satisfy the condition that the injection rates of the proppants and the fluid are constant. However this condition also means that the average concentration at the wellbore, $\Pi\circ\mu_{w}(0)$, must equal the average concentration taken by integrating over the entire crack volume. For a monotonic choice of $\mu_{w}$ this implies that $\mu_{w}$ must be independent of $\xi$. Herein we will make the assumption that $\mu_{w}$ is a constant and, as a result, so is $\Pi=\Pi(\mu_{w})$. This is a natural assumption: at early times we don’t expect significant concentration differences along the crack because radial length scales are small. A great advantage of a constant $\Pi$ is that we can define an ‘effective viscosity’, which we can absorb into our scaled variables the same way as we did with fluid viscosity. Under the assumption that $\mu_{w}$ is constant, (31) becomes (47) $\displaystyle\int_{\rho}^{1}\xi h(\xi)d\xi+\frac{4}{9}\rho^{2}h=-\frac{\rho h^{3}}{\eta_{e}}\partialderivative{p}{\rho},$ where $\eta_{e}=\mu_{w}^{2}\Pi/\Omega$ is what we call the effective viscosity. It is plotted in Figure 3(c), and is best thought of as a function of the average concentration, $\Pi$. Making the transformations (48) $\displaystyle h$ $\displaystyle=\eta_{e}^{1/3}\tilde{h},$ $\displaystyle p$ $\displaystyle=\eta_{e}^{1/3}\tilde{p},$ $\displaystyle\Gamma$ $\displaystyle=\eta_{e}^{-1/9}\tilde{\Gamma},$ our governing equations become (49) $\displaystyle\int_{\rho}^{1}\xi\tilde{h}d\xi$ $\displaystyle+\frac{4}{9}\rho^{2}\tilde{h}=-\rho\tilde{h}^{3}\partialderivative{p}{\rho},$ $\displaystyle\tilde{h}(\xi)$ $\displaystyle=\frac{4}{\pi}\int_{\xi}^{1}\frac{y}{\sqrt{y^{2}-\xi^{2}}}\int_{0}^{1}\frac{x\tilde{p}(xy)}{\sqrt{1-x^{2}}}dxdy,$ $\displaystyle 0$ $\displaystyle=\int_{0}^{1}\frac{\tilde{p}(\xi)\xi}{\sqrt{1-\xi^{2}}}d\xi,$ $\displaystyle 1$ $\displaystyle=4\pi\tilde{\Gamma}^{3}\int_{0}^{1}(\xi\tilde{h})d\xi.$ We will solve them using the numerical method described before, except with (49) in the place of (31-34). Figure 5 plots $\tilde{h}$ and $\tilde{p}$, calculated using $N=4$ and $M+1=501$ control points. Promisingly, we note that $\tilde{h}>0$ and $p$ shows the expected asymptotic behaviour. The value $\tilde{h}(0)=1.36$ will be important in later discussion. The first column of table 3 shows the coefficients $A_{i}$ and $B_{i}$, as well as the calculated value of $\tilde{\Gamma}=0.598$. Significantly, we see that $A_{i}$ and $B_{i}$ decrease rapidly with $i$, suggesting that a solution with higher order terms is unnecessary. This is supported by the small value of $\Delta\approx 5\times 10^{-5}$, with evenly spread contributions from control points along the radius of the crack. This suggests that we have found a genuine solution, and that the tip asymptotics are indeed suitable. Figure 5. $(\xi,\eta)$ plots of $\tilde{h}$ and $\tilde{p}$, the scaled width and pressure solutions to the absorbed effective viscosity system. We now focus on finding numerical solutions for different concentrations in order to consider features such as the velocity profile and proppant distribution within the cavity. We consider the case of four different values of the average concentration, $\Pi$. These are given in table 2, along with the corresponding values of $\mu_{w}$ and $\eta_{e}$. $\Pi$ | $\mu_{w}$ | $\eta_{e}$ ---|---|--- 0.05 | 487.3 | 2.74 0.20 | 23.35 | 3.92 0.40 | 3.93 | 10.37 0.55 | 1.06 | 96.60 Table 2. Test values of $\Pi$, $\mu_{w}$ and $\eta_{e}$. The latter columns of table 3 show the values of $A$, $B$ and $\Gamma$ calculated using the exact method suggested in Section 6. Again we use $M+1=501$ control points and $N=4$. Happily, the same values are observed by using the values of $A$, $B$ and $\Gamma$ listed in the first column, calculated after absorbing the effective viscosity, and using the relations (48) to return to the concentration-specific values. We calculate the same value of $\Delta\approx 5\times 10^{-5}$ each time; this is to be expected as the equations are equivalent once the solutions have been scaled. | $\Pi$ | | Absorbed | 0.05 | 0.20 | 0.40 | 0.55 ---|---|---|---|---|---|---|--- | $A_{-1}$ | | 0.14786 | 0.20710 | 0.23326 | 0.32238 | 0.67830 | $A_{0}$ | | 0.53529 | 0.74974 | 0.84444 | 1.16709 | 2.45559 | $A_{1}$ | | 0.01929 | 0.02702 | 0.03043 | 0.04206 | 0.08849 | $A_{2}$ | | 0.00402 | 0.00563 | 0.00634 | 0.00877 | 0.01844 | $A_{3}$ | | 0.00035 | 0.00049 | 0.00055 | 0.00076 | 0.00159 | $B_{-1}$ | | 0.14786 | 0.20710 | 0.23326 | 0.32238 | 0.67830 | $B_{0}$ | | 0.53805 | 0.75361 | 0.84879 | 1.17311 | 2.46825 | $B_{1}$ | | 0.05435 | 0.07612 | 0.08573 | 0.11849 | 0.24931 | $B_{2}$ | | 0.00012 | 0.00016 | 0.00019 | 0.00026 | 0.00054 | $B_{3}$ | | 0.00081 | 0.00114 | 0.00128 | 0.00177 | 0.00373 | $B_{4}$ | | 0.00029 | 0.00041 | 0.00046 | 0.00064 | 0.00134 | $\Gamma$ | | 0.59812 | 0.534579 | 0.513799 | 0.461261 | 0.359968 Table 3. Values of $A_{i}$, $B_{i}$ and $\Gamma$ obtained using (49) with effective viscosity absorbed into the scaling and (31-34) with $\Pi\in\\{0.05,0.20,0.40,0.55\\}$. We use $M=500$ and $N=4$ throughout. Figure 6 shows the distribution of proppants within the fracture for each value of $\Pi$. They are overlaid with an arrow plot of the proppant velocity profile, $\mathbf{v}$, scaled by $\xi$ to show the equivalent two-dimensional flux. The calculation of $\mathbf{v}$ is omitted since it is lengthy and similar to the derivation of (27) in Appendix A. As $\Pi$ increases we see a growing disk of plug flow where $\phi>\phi_{m}$, marked with a magenta contour. We also see a tendency towards proppant velocity across the crack, rather than along it; this is because the shape of the crack becomes shorter and wider as the effective viscosity increases. (a) $\Pi=0.05$ (b) $\Pi=0.20$ (c) $\Pi=0.40$ (d) $\Pi=0.55$ Figure 6. Concentration-specific $(\Gamma\xi,\eta)$ plots of developing fractures with total solid volume fraction, $\Pi$, taking the values $0.05$, $0.20$, $0.40$ and $0.55$. These are presented with filled contours displaying proppant concentration; arrows showing $\xi$-scaled velocity; and magenta contours indicating the transition into plug flow at the centre of each cavity. Drawing on calculations we have made so far, we are now in a position to assess the significance of tip screen-out in our model, something we have neglected so far by adopting a continuum model of proppant transport. This is where, near the crack tip, the narrowing crack aperture causes proppants to jam and block the fracture, significantly affecting the development of the evolving formation and the convective transport of proppants. In [28] this problem is addressed using a ‘blocking function’ which reduces proppant flux to zero in apertures smaller than three times the average particle’s diameter. We will use this threshold to weigh the significance of ignoring screen-out in our model. Figure 7(a) shows the volume-proportion of proppants predicted in fracture regions of width less than this threshold, dependant on the time, $t$, and the average proppant concentration, $\Pi$. We see that for early times and low concentrations, our model predicts a significant proportion of proppants in these regions, where the fracturing fluid is clear in reality. However, in concentrations greater than $0.3$ this proportion is relatively small; this means our model, which ignores tip screen-out, is self-consistent. This difference arises from the effective viscosity, which increases with $\Pi$ and causes the ratio of fracture width to length to decrease. Lecampion & Garagash [13] conclude that their rheology, which is employed throughout this paper, agrees very well with experimental results when the predicted width of plug flow is greater than a particle’s width. In figure 7(b), we see this condition holds for moderate times when $\phi>0.4$. It does not for $\phi<0.4$. Therefore, in this regime we can expect slight mismatches between predicted and practical concentration profiles; this arises from a breakdown of the continuum model in the jammed part of the flow [13]. (a) $w<6a$ (b) Plug width $<2a$ Figure 7. Proportion of proppants by volume, predicted in fracture regions where $w<6a$, or plug width $<2a$, given average concentration, $\Pi$, and time, $t$. ## 8\. Crack Closure: Problem Formulation In the zero-proppant case, Lai et al [29] have confirmed experimentally that for late times after the fluid pressure is released, the crack radius is constant and volume scales as $t^{-1/3}$. It is tempting to repeat our previous work in order to find an asymptotic solution with a generalised total fracture volume $Qt^{\alpha}$. We would then let $\alpha=-1/3$ to model the case of closure. This approach leads us to (50) $\displaystyle\alpha\int_{\rho}^{1}\xi h(\xi)d\xi+\beta\rho^{2}h=-\frac{\rho h^{3}}{\eta_{e}}\partialderivative{p}{\rho},$ in the place of (47). Here $\beta=(3\alpha+1)/9$ is the exponent for $L$, giving the radial growth of the fracture. However, we see that attempts to solve (50) using the previous numerical method fail as $(\alpha,\beta)\to(-1/3,0)$, corresponding to the case in [29]. This is because the tip asymptotes $w\propto(1-\xi)^{2/3}$ and $p\propto(1-\xi)^{-1/3}$ are a result of an advancing fracture in a zero- toughness medium. Spence & Sharp [8] note that $h\sim C(1-\xi)^{\tau}$ implies $p\sim C\tau(\cot\pi\tau)(1-\xi)^{\tau-1}$. Balancing terms in (50), we are forced with $C\leq 0$ if $\beta\leq 0$ which clearly can’t lead to physical solutions, given the constraint $h\geq 0$. In the same paper, solutions for $\beta=0$ are shown to exist without the assumption of zero-toughness; these have $h\sim(1-\xi^{2})^{1/2}$. However, this causes difficulties in the case of an evolving fracture, since a non-zero toughness parameter, $\aleph$, brings time dependence to the scaled equations we have derived. An alternative solution would be the addition of a non-zero fluid lag, providing a region of negative pressure between the fluid front and the crack tip. Such a region exists in reality, containing either vapour from the fracturing fluid or, if the surrounding medium is permeable, pore fluid [30, 31]. Zero-toughness solutions using this formulation are explored in [32]. Schematics of each possible solution type are shown in Figure 8. Figure 8. Possibilities for modelling the crack tip. Any model utilising a time independent concentration profile is likely to fail in describing fracture closure at late times. This is because the width of the crack is decreasing as $t^{-1/3}$, so it is bound to become comparable to the proppant diameter. At the point where $\epsilon L/a\approx 6$, the proppants begin to bridge across the fracture, effectively fixing them in position [28]; therein, concentrations will increase as the carrier fluid is forced from the cavity. For this reason, we will instead address the problem of finding the residual crack shape, given some axisymmetric initial distribution of proppants; we will assume these are radially immobile from the moment pressure is released. This method has been used with success to model the closure of a bi-wing fracture by Wang et al. [33, 34]. ## 9\. Crack Closure: Residual Width Profiles We model the residual shape of the fracture using $w_{p}(r)$, defined as the close packed width of proppants. That is to say, after packing the proppants as tightly as possible in the z direction, so $\phi=\phi_{rcp}$, this is the residual width. Given some radial distribution of proppants described by the average concentration, $\Pi$, and un-scaled width profile, $w$, we deduce that $w_{p}=w\Pi/\phi_{rcp}$. This description is compatible with the frictional rheology of Lecampion & Garagash [13], used previously, which asserts that a non-zero normal force on the proppants, along with vanishing shear stress, causes compression up to the random close packing limit. We then assume that the surrounding fracture simply collapses around the proppant pack. Our primary interest will be in using proppant distributions, arising from the injection phase described previously, to predict the geometry of the residual formation. In [34] a more complicated model is offered; this considers stress from the contact of opposing crack asperities, proppant embedment into the fracture walls, and compression of proppants. Since we will be concerned with cases where $w_{p}$ is non-zero along the entire crack radius; the contact term arising from the crack asperities, which is significant in the un-propped case, will not be necessary. Furthermore, in the same paper [34] the depth of proppant embedment is shown to be of the order $K_{e}=a(3/4E^{\prime})^{2}(16mE^{\prime 2}/9c_{p})^{2/3}$. Here, $m\approx 2\sqrt{3}$ is a constant which depends on the packing of proppants. Using the value of $c_{p}=3.9\times 10^{-8}\textrm{Pa}^{-1}$ [34], as well as the typical values of $a=50\mu\textrm{m}$ and $E^{\prime}=40\textrm{GPa}$ mentioned earlier, we note that $K_{e}\approx 1\mu\textrm{m}$, around 100 times smaller than the given proppant diameter. Since we will generally model proppant packs which are several times the size of the proppant diameter in width, we will ignore this phenomenon. Finally, we note that, according to our previous estimates, more than $10\textrm{s}$ into the injection phase we should expect pressures of less than $1\textrm{MPa}$. In [34] the compressive stress required to reduce the width of the closely packed proppant bed from $w_{p}$ to $w$ is given by $1/c_{p}\ln(w_{p}/w)$; using this, the same stress would only cause a $4\%$ reduction in width. Since typical stresses involved in the closure phase are much smaller than this, we will model the proppants as incompressible. This model of crack closure leads to a simple description of the residual crack profile. We have two parameters: one for average concentration, $\Pi$, and another for the time that injection ceases, $t_{0}$. Herein we will denote $\\{\tilde{h},\tilde{p},\tilde{\Gamma}\\}$ as the solution to the system of equations given in (49); $\tilde{h}$ and $\tilde{p}$ are plotted in Figure 5 and we use the value $\tilde{\Gamma}=0.598$. Then, using (48) and the original scaling arguments, we deduce that (51) $\displaystyle w_{p}(\xi;t_{0},\Pi)$ $\displaystyle=\frac{\Pi}{\phi_{rcp}}\epsilon(t_{0})L(t_{0})\eta_{e}(\Pi)^{2/9}\tilde{\Gamma}\tilde{h}(\xi),$ (52) $\displaystyle R(t_{0},\Pi)$ $\displaystyle=L(t_{0})\eta_{e}(\Pi)^{-1/9}\tilde{\Gamma}.$ From Figure 5 we notice that $\max(\tilde{h}_{1})\approx 1.35$. Using this, we may plot Figure 9(a), which shows the effect of average concentration on the maximum residual width of the formation. It is interesting to note that the propped width doesn’t grow proportional to the proppant concentration, as one may expect from the close packing of the suspended proppants. Instead, the dependance is superlinear, because greater proppant concentrations lead to a higher effective viscosity; this causes the fracture to take a wider shape before the release of injection pressure. We can also see that $t_{0}$ has relatively little effect on the maximum crack width. This is because the $t_{0}$ dependent term, $\epsilon L$, grows with $t_{0}^{1/9}$. By contrast, in Figure 9(b) we see a greater time dependence in the final radius, which grows with $L\propto t^{4/9}$. As the proppant concentration increases, with $t_{0}$ fixed, we see a decrease in the final radius of fracture achieved, arising from an increase in the effective viscosity. (a) Maximum fracture width. (b) Fracture radius. Figure 9. Plots showing the effect of average concentration on the maximum residual fracture width and radius for $t_{0}\in\\{100,500,1000\\}$. ## 10\. Conclusions We have established a mathematical framework that captures the behaviour of a slurry within a pressure driven cavity. Using typical parameters from industrial fracking, we predict that the development length, required to establish stable proppant flow away from the wellbore, is negligible compared to the typical radius of the penny-shaped fracture generated. As a result, we may assume the flow is fully developed, reducing the in-fracture distribution of proppants to a function of the radial distance from the wellbore. A further assumption of constant proppant injection rate allows us to describe the proppant distribution with one parameter, the total solid volume fraction. In the zero-concentration limit, our model becomes similar to one derived using Newtonian flow, with some disagreement arising from our choice of a dense frictional rheology. Within this framework, we are able to define an effective viscosity, which we may absorb into our equations using a suitable choice of scaling. This is a particularly striking result because it establishes an equivalence between slurry flow of a given solid fraction and simple Newtonian flow with some particular viscosity, at least in the sense of fracture development. Solving the resulting set of equations numerically, we may then return to our original scaling to investigate concentration-specific solutions. Unsurprisingly, we predict width and pressure profiles with the tip-asymptotic behaviour described in [9]. As the proppant concentration increases we expect shorter and wider fractures with steeper fluid pressure gradients. In the centre of the fracture, where shear rate vanishes, we predict the formation of a disk of plug flow with width, in relation to the crack, increasing with the average proppant concentration. Evaluating our model, we see that the unaccounted effect of tip screen-out is likely to be significant in the low concentration, low effective viscosity case, particularly at early times. Here, the cavity formed is narrow, so near its tip, particle bridging is likely. Moreover, we observe that for typical fracturing timescales, if $\Pi<0.4$, our model predicts plug flow thinner than one particle width: suggesting that our use of a continuum model may not be appropriate. Otherwise, the plug flow is broader than a particle’s width, meaning it is physically realisable and the results of [13] suggest we should have good experimental agreement. Lastly, we have adopted a simple model of crack closure which regards the remaining proppants to be immobile and incompressible. This allows us to predict the shape of the residual crack, based on two parameters: the average proppant concentration within the injected fluid and the length of time between the initiation of fracking and the release of pressure. Simple formulae show that the residual fracture width increases significantly with proppant concentration, and grows very slowly with time; fracture radius however, decreases with proppant concentration and increases with time. The results established here have important applications in both contexts of industrial fracking and geological dike formation. Diagnostics of tip screen- out and forecasts of residual fracture geometry are relevant to the formation of conductive fractures, whilst predictions about the shape and particle distribution of a slurry driven crack relate more to a cooling magma. The discovery of an effective viscosity may also provide a foothold in understanding slurry driven fractures, particularly given the bounty of literature surrounding cracks generated by Newtonian fluid. In spite of all this, experimental investigation is necessary to bolster the predictions we have made. We hope this will form the basis of a second article, with tentative title: ‘Proppant flow in a penny-shaped crack. Part II : Experimental Investigation’. ## 11\. Acknowledgements The authors would like to thank Derek Elsworth (Pennsylvania State University), Elisabeth Guazzelli (Centre National de la Recherche Scientifique) and Emmanuel Detournay (University of Minnesota) for their support and guidance in the drafting of this paper; with special gratitude to Elisabeth for providing the data used in Figure 2. We would also like to thank John Willis (University of Cambridge) for his support in the publication of the paper. ## Appendix A Integrating the $\phi$,phionservation equation over the crack width In this Appendix we integrate equation (25) over $(-w,w)$ to yield (27); we will take a term-by-term approach. First, we note that by (24), (53) $\displaystyle\int_{-z}^{z}\phi(\xi,\eta)d\eta$ $\displaystyle=2\int_{0}^{z}\mu^{-1}\left(\mu_{w}(\xi)\frac{\eta}{w}\right)d\eta,$ (54) $\displaystyle=2z\Pi\left(\mu_{w}(\xi)\frac{z}{w}\right).$ Hence, we see that (55) $\displaystyle\int_{-w}^{w}\partialderivative{\phi}{\xi}d\eta$ $\displaystyle=\partialderivative{\xi}\int_{-w}^{w}\phi d\eta-2\phi(\xi,w)\partialderivative{w}{\xi},$ (56) $\displaystyle=2\partialderivative{\xi}\left[w\Pi\circ\mu_{w}(\xi)\right]-2\phi(\xi,w)\partialderivative{w}{\xi}.$ Then, integrating by parts, we find (57) $\displaystyle\int_{-w}^{w}\eta\partialderivative{\phi}{\eta}d\eta=2\left[w\phi(\xi,w)-w\Pi\circ\mu_{w}(\xi)\right].$ Furthermore, utilising the expression of $v_{r}$ given in (20) and the condition $v_{r}(\xi,\pm w)=0$ we determine (58) $\displaystyle\int_{-w}^{w}\partialderivative{(\xi\phi v_{r})}{\xi}d\eta$ $\displaystyle=\partialderivative{\xi}\left[\xi\int_{-w}^{w}\phi v_{r}d\eta\right],$ (59) $\displaystyle=-\frac{6}{\Gamma}\partialderivative{\xi}\left[\xi\partialderivative{p}{\xi}\int_{0}^{w}\phi(\xi,\eta)\int_{\eta}^{w}\frac{I(\phi(\xi,z))z}{\mu(\phi(\xi,z))}dzd\eta\right],$ (60) $\displaystyle=-\frac{6}{\Gamma}\partialderivative{\xi}\left[\xi\partialderivative{p}{\xi}\int_{0}^{w}\int_{0}^{z}\phi(\xi,\eta)\frac{I(\phi(\xi,z))z}{\mu(\phi(\xi,z))}d\eta dz\right],$ (61) $\displaystyle=-\frac{6}{\Gamma}\partialderivative{\xi}\left[\xi\partialderivative{p}{\xi}\int_{0}^{w}z^{2}\Pi\left(\frac{\mu_{w}z}{w}\right)\frac{I(\phi(\xi,z))}{\mu(\phi(\xi,z))}dz\right].$ However, by (24), $\mu(\phi(\xi,z))=\mu_{w}z/w$, so (62) $\displaystyle\int_{-w}^{w}\partialderivative{(\xi\phi v_{r})}{\xi}d\eta$ $\displaystyle=-\frac{6}{\Gamma}\partialderivative{\xi}\left[\frac{w\xi}{\mu_{w}}\partialderivative{p}{\xi}\int_{0}^{w}z\Pi\left(\frac{\mu_{w}z}{w}\right)I\circ\mu^{-1}\left(\frac{\mu_{w}z}{w}\right)dz\right],$ (63) $\displaystyle=-\frac{6}{\Gamma}\partialderivative{\xi}\left[\frac{\xi w^{3}}{\mu_{w}(\xi)^{2}}\partialderivative{p}{\xi}\Omega\circ\mu_{w}(\xi)\right].$ Finally, we know that (64) $\displaystyle\int_{-w}^{w}\partialderivative{(\phi v_{z})}{\eta}d\eta=2\phi(\xi,w)v_{z}(\xi,w).$ In the original scaling we have the boundary condition $v_{z}(x,w)=\partialderivative{w}{t}(x,t)$; in the lubrication scaling this becomes (65) $\displaystyle-\dot{\epsilon}Lv_{z}(\xi,w)$ $\displaystyle=\left[\dot{\epsilon}L+\epsilon\dot{L}\right]w(\xi,T)-\epsilon L\xi\left[\frac{\dot{L}}{L}+\frac{\Gamma^{\prime}\dot{T}}{\Gamma}\right]\partialderivative{w}{\xi}+\dot{T}\partialderivative{w}{T}.$ Hence, (66) $\displaystyle v_{z}(\xi,w)=\frac{w}{3}-\frac{4\xi}{3}\partialderivative{w}{\xi},$ and so (67) $\displaystyle\int_{-w}^{w}\partialderivative{(\phi v_{z})}{\eta}d\eta=2\phi(\xi,w)\left[\frac{w}{3}-\frac{4\xi}{3}\partialderivative{w}{\xi}\right].$ Adding these terms together and making various cancellations, we derive equation (27). ## Appendix B Zero-Concentration Limit In this Appendix, we will compare the properties of equation (29) to the equivalent zero-proppant equation. Modelling the flow as Newtonian instead, we would have used the relation $\tau=\eta_{f}\dot{\gamma}$. In our choice of scaling this becomes $\tau=\dot{\gamma}$. Hence (19.2) is replaced by (68) $\displaystyle\partialderivative{v_{r}}{\eta}=\frac{3\eta}{\Gamma}\partialderivative{p}{\xi},$ where $\mathbf{v}$ is the fluid velocity. With the assumption that $\nabla\cdot v=0$, our scaled continuity equation is simply (69) $\displaystyle\frac{1}{\Gamma\xi}\partialderivative{(\xi v_{r})}{\xi}+\partialderivative{v_{z}}{\eta}=0.$ Integrating first over $(-w,w)$ as in Appendix A, making use of (66), (68) and $\tau=\dot{\gamma}$, we obtain (70) $\displaystyle\frac{w}{3}-\frac{4\xi}{3}\partialderivative{w}{\xi}=\frac{1}{\xi\Gamma^{2}}\partialderivative{\xi}\left[\partialderivative{p}{\xi}\xi w^{3}\right].$ Then, multiplying by $\xi$ and integrating from $\rho$ to 1, we use the $w^{3}\partial p/\partial\xi\to 0$ limit employed to derive (29), (71) $\displaystyle\int_{\rho}^{1}\xi wd\xi+\frac{4}{9}\rho^{2}w=-\frac{\rho w^{3}}{3\Gamma^{2}}\partialderivative{p}{\rho}.$ In order to compare (29) and (71), we are required to find the limit of $\Omega/(x^{2}\Pi)$ as $x\to\infty$. Explicitly we see that (72) $\displaystyle\lim_{x\to\infty}\frac{\Omega(x)}{x^{2}\Pi(x)}$ $\displaystyle=\lim_{x\to\infty}\frac{1}{x^{3}\Pi(x)}\int_{0}^{x}\Pi(u)I\circ\mu^{-1}(u)udu,$ (73) $\displaystyle=\lim_{x\to\infty}\int_{0}^{1}\frac{\Pi(vx)}{\Pi(x)}\cdot\frac{I\circ\mu^{-1}(vx)}{vx}\cdot v^{2}dv,$ (74) $\displaystyle=\int_{0}^{1}v^{2}\lim_{x\to\infty}\left[\frac{\Pi(vx)}{\Pi(x)}\right]dv,$ (75) $\displaystyle=\int_{0}^{1}v\lim_{x\to\infty}\left[\frac{\int_{0}^{vx}\mu^{-1}(u)du}{\int_{0}^{x}\mu^{-1}(u)du}\right]dv,$ (76) $\displaystyle=\int_{0}^{1}v^{2}\lim_{x\to\infty}\left[\frac{\mu^{-1}(vx)}{\mu^{-1}(x)}\right]dv,$ (77) $\displaystyle=\int_{0}^{1}v^{2}\lim_{x\to\infty}\left[\frac{I^{-1}(vx)}{I^{-1}(x)}\right]dv,$ (78) $\displaystyle=\int_{0}^{1}v^{2}\lim_{x\to\infty}\left[\frac{1+\sqrt{x}}{1+\sqrt{vx}}\right]dv,$ (79) $\displaystyle=\int_{0}^{1}v^{3/2}dv,$ (80) $\displaystyle=2/5.$ Here (74) and (77) arise from the fact $I(\phi)\sim\mu(\phi)$ as $\phi\to 0$, because the fluid shear stress approaches the slurry shear stress. (76) comes from L’Hôpital’s rule. We conclude that the equations governing Newtonian flow are not the same as those in the zero-proppant slurry flow limit. ## Appendix C Matrix $(P)_{ij}$, when $N=4$ The matrix $(P)_{ij}$ for $N=4$, as provided in [9], is given in table 4. | | | | | j | | ---|---|---|---|---|---|---|--- | | | -1 | 0 | 1 | 2 | 3 | -1 | | 1.0000 | 0.0000 | 0.0000 | 0.0000 | 0.0000 | 0 | | 0.0000 | 0.9560 | 1.2730 | 0.4101 | 0.3145 i | 1 | | 0.0000 | 0.0991 | -0.0185 | 0.4068 | 0.0610 | 2 | | 0.0000 | 0.0018 | -0.0429 | -0.0244 | 0.2293 | 3 | | 0.0000 | 0.0017 | 0.0039 | -0.0416 | -0.0141 | 4 | | 0.0000 | 0.0005 | 0.0026 | -0.0032 | -0.0372 Table 4. Matrix $(P)_{ij}$, for N=4. ## References * [1] Wells, Bruce A., ed. (2007). ”Shooters”. The Petroleum Age. American Oil and Gas Historical Society. 4 (3): 8–9. ISSN 1930-5915 * [2] Charlez, Philippe A. (1997). Rock Mechanics: Petroleum Applications. Paris: Editions Technip. p. 239. ISBN 9782710805861. * [3] National Earthquake Hazards Reduction Program (U.S.), Geological Survey (U.S.), Office of Earthquakes, Volcanoes, and Engineering, U.S. National Committee for Rock Mechanics (1983). Hydraulic Fracturing Stress Measurements. Volume 26 of International journal of rock mechanics and mining sciences and geomechanics abstracts. * [4] Pierce, Brenda (2010). Geothermal Energy Resources. National Association of Regulatory Utility Commissioners (NARUC). * [5] Miller, Bruce G. (2005). Coal Energy Systems. Sustainable World Series. Academic Press. p. 380. ISBN 9780124974517. * [6] E. Rivalta, B. Taisne, A.P. Bunger, R.F. Katz (2015). A review of mechanical models of dike propagation: Schools of thought, results and future directions. Tectonophysics. Volume 638,2015. Pages 1-42. ISSN 0040-1951. * [7] Petford, N., Koenders, M.A. (1998). Granular flow and viscous fluctuations in low Bagnold number granitic magmas. Journal of the Geological Society, 155 (5), pp. 873-881. 10.1144/gsjgs.155.5.0873 * [8] Spence, D.A., Sharp, P.W. (1985). Self-similar solution for elastohydrodynamic cavity flow. Proc. Roy. Soc. London, Ser. A (400),289–313. * [9] A.A. Savitski, E. Detournay (2002). Propagation of a penny-shaped fluid-driven fracture in an impermeable rock: asymptotic solutions, International Journal of Solids and Structures, Volume 39, Issue 26, Pages 6311-6337. * [10] Einstein, A. (1906). A new determination of molecular dimensions. Ann. Phys. 4 (19), 289–306. * [11] Boyer F., Guazzelli É., Pouliquen O. (2011). Unifying suspension and granular rheology. Phys. Rev. Lett. 107 (18), 188301. * [12] Dontsov EV, Boronin SA, Osiptsov AA, Derbyshev DY. (2019). Lubrication model of suspension flow in a hydraulic fracture with frictional rheology for shear-induced migration and jamming. Proc. R. Soc. A 475: 20190039. * [13] Lecampion, Garagash (2014). Confined flow of suspensions modelled by a frictional rheology. J. Fluid Mech. (2014), vol. 759, pp. 197–235. Cambridge University Press 2014. doi:10.1017/jfm.2014.557 * [14] Niall J. O’Keeffe, Herbert E. Huppert & P. F. Linden (2018). Experimental exploration of fluid-driven cracks in brittle hydrogels. J. Fluid Mech., vol. 844, pp. 435–458. * [15] Garagash, D.I., Detournay, E. (2000). The tip region of a fluid-driven fracture in an elastic medium. ASME J. Appl. Mech. 67, 183–192. * [16] Sneddon, I.N., (1951). Fourier Transforms. McGraw-Hill, New York, NY * [17] Rice, J.R., (1968). Mathematical analysis in the mechanics of fracture. In: Liebowitz, H. (Ed.), Fracture, an Advanced Treatise. Vol. II.Academic Press, New York, NY, pp. 191–311 (Chapter 3). * [18] Dagois-Bohy S., Hormozi S., Guazzelli É, Pouliquen O. (2015). Rheology of dense suspensions of non-colloidal spheres in yield-stress fluids. Journal of Fluid Mechanics, 776, R2. doi:10.1017/jfm.2015.329 * [19] Garside, J., Al-Dibouni, M. R. (1977). Velocity-voidage relationships for fluidization and sedimentation in solid–liquid systems. Ind. Eng. Chem. Process Des. Dev. 16 (2), 206–214. * [20] Richardson, J., Zaki, W. (1954) Sedimentation and fluidization: Part I. Trans. Inst. Chem. Engrs 32, 35–47. * [21] Bacri, J.-C., Frenois, C., Hoyos, M., Perzynski, R., Rakotomalala, N. & Salin, D. (1986). Acoustic study of suspension sedimentation. Europhys. Lett. 2 (2), 123–128. * [22] Shiozawa, S., Mcclure, M. (2016). Simulation of proppant transport with gravitational settling and fracture closure in a three-dimensional hydraulic fracturing simulator. J. Petrol. Sci. Engng, 138, 298–314. * [23] Chen Zhixi, Chen Mian, Jin Yan, Huang Rongzun (1997). Determination of rock fracture toughness and its relationship with acoustic velocity, International Journal of Rock Mechanics and Mining Sciences, Volume 34, Issues 3–4, 1997, Pages 49.e1-49.e11, ISSN 1365-1609 * [24] Feng Liang, Mohammed Sayed, Ghaithan A. Al-Muntasheri, Frank F. Chang, Leiming Li (2016). A comprehensive review on proppant technologies. Petroleum, Volume 2, Issue 1, March 2016, Pages 26-39. * [25] Dontsov EV, Boronin SA, Osiptsov AA, Derbyshev DY. (2019). Lubrication model of suspension flow in a hydraulic fracture with frictional rheology for shear-induced migration and jamming. Proc. R. Soc. A 475: 20190039. * [26] Abramowitz, M., Stegun, I.A. (Eds.), (1964). Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables.Applied Mathematics Series, 55. US Govt. Print. Off, Washington, DC. * [27] Lagarias, J. C., J. A. Reeds, M. H. Wright, & P. E. Wright (1998). Convergence Properties of the Nelder-Mead Simplex Method in Low Dimensions. SIAM Journal of Optimization. Vol. 9, Number 1, 1998, pp. 112–147. * [28] Dontsov, E. V., Peirce, A. P. (2014). Slurry flow, gravitational settling and a proppant transport model for hydraulic fractures. J. Fluid Mech. 760, 567–590. * [29] Ching-Yao Lai, Zhong Zheng, Emilie Dressaire, Guy Z. Ramon, Herbert E. Huppert, & Howard A. Stone (2016). Elastic Relaxation of Fluid-Driven Cracks and the Resulting Backflow. Physical Review Letters 117, 268001. * [30] A.M. Rubin. (1993). Tensile fracture of rock at high confining pressure: implications for dike propagation. J. Geophys. Res., 98 (B9) (1993), pp. 15,919-15,935. * [31] E. Detournay & D. Garagash (2003). The tip region of a fluid-driven fracture in a permeable elastic solid. J. Fluid Mech., 494, pp. 1-32. * [32] D. Garagash (2006). Propagation of a plane-strain hydraulic fracture with a fluid lag: Early-time solution, International Journal of Solids and Structures 43, 5811–5835. * [33] Jiehao Wang, Derek Elsworth & Martin K. Denison (2018). Propagation, proppant transport and the evolution of transport properties of hydraulic fractures. J. Fluid Mech., vol. 855, pp. 503–534. * [34] Wang, J. & Elsworth, D. (2018). Role of proppant distribution on the evolution of hydraulic fracture conductivity. J. Petrol. Sci. Engng 166, 249–262.
# Multilingual and cross-lingual document classification: A meta-learning approach Niels van der Heijden♣ Helen Yannakoudakis♠ Pushkar Mishra♢ Ekaterina Shutova♣ ♣ILLC, University of Amsterdam, the Netherlands ♠Dept. of Informatics, King’s College London, United Kingdom ♢Facebook AI, London, United Kingdom <EMAIL_ADDRESS><EMAIL_ADDRESS> <EMAIL_ADDRESS><EMAIL_ADDRESS> ###### Abstract The great majority of languages in the world are considered under-resourced for the successful application of deep learning methods. In this work, we propose a meta-learning approach to document classification in a limited- resource setting and demonstrate its effectiveness in two different settings: few-shot, cross-lingual adaptation to previously unseen languages; and multilingual joint training when limited target-language data is available during training. We conduct a systematic comparison of several meta-learning methods, investigate multiple settings in terms of data availability and show that meta-learning thrives in settings with a heterogeneous task distribution. We propose a simple, yet effective adjustment to existing meta-learning methods which allows for better and more stable learning, and set a new state of the art on several languages while performing on-par on others, using only a small amount of labeled data. ## 1 Introduction There are more than 7000 languages around the world and, of them, around 6% account for 94% of the population.111https://www.ethnologue.com/statistics Even for the 6% most spoken languages, very few of them possess adequate resources for natural language research and, when they do, resources in different domains are highly imbalanced. Additionally, human language is dynamic in nature: new words and domains emerge continuously and hence no model learned in a particular time will remain valid forever. With the aim of extending the global reach of Natural Language Processing (NLP) technology, much recent research has focused on the development of multilingual models and methods to efficiently transfer knowledge across languages. Among these advances are multilingual word vectors which aim to give word-translation pairs a similar encoding in some embedding space Mikolov et al. (2013a); Lample et al. (2017). There has also been a lot of work on multilingual sentence and word encoders that either explicitly utilizes corpora of bi-texts Artetxe and Schwenk (2019); Lample and Conneau (2019) or jointly trains language models for many languages in one encoder Devlin et al. (2018); Conneau et al. (2019). Although great progress has been made in cross- lingual transfer learning, these methods either do not close the gap with performance in a single high-resource language Artetxe and Schwenk (2019); Conneau et al. (2019); van der Heijden et al. (2019), e.g., because of cultural differences in languages which are not accounted for, or are impractically expensive Lai et al. (2019). Meta-learning, or learning to learn Schmidhuber (1987); Bengio et al. (1990); Thrun and Pratt (1998), is a learning paradigm which focuses on the quick adaption of a learner to new tasks. The idea is that by training a learner to adapt quickly and from a few examples on a diverse set of training tasks, the learner can also generalize to unseen tasks at test time. Meta-learning has recently emerged as a promising technique for few-shot learning for a wide array of tasks Finn et al. (2017); Koch et al. (2015); Ravi and Larochelle (2017) including NLP Dou et al. (2019); Gu et al. (2018). To our best knowledge, no previous work has been done in investigating meta-learning as a framework for multilingual and cross-lingual few-shot learning. We propose such a framework and demonstrate its effectiveness in document classification tasks. The only current study on meta-learning for cross-lingual few-shot learning is the one by Nooralahzadeh et al. (2020), focusing on natural language inference and multilingual question answering. In their work, the authors focus on applying meta-learning to learn to adapt a monolingually trained classifier to new languages. In contrast to this work, we instead show that, in many cases, it is more favourable to not initialize the meta-learning process from a monolingually trained classifier, but rather reserve its respective training data for meta-learning instead. Our contributions are as follows: 1) We propose a meta-learning approach to few-shot cross-lingual and multilingual adaptation and demonstrate its effectiveness on document classification tasks over traditional supervised learning; 2) We provide an extensive comparison of meta-learning methods on multilingual and cross-lingual few-shot learning and release our code to facilitate further research in the field;222https://github.com/mrvoh/meta_learning_multilingual_doc_classification 3) We analyse the effectiveness of meta-learning under a number of different parameter initializations and multiple settings in terms of data availability, and show that meta-learning can effectively learn from few examples and diverse data distributions; 4) We introduce a simple yet effective modification to existing methods and empirically show that it stabilizes training and converges faster to better local optima; 5) We set a new state of the art on several languages and achieve on-par results on others using only a small amount of data. ## 2 Meta-learning methods Algorithm 1 Meta-training procedure. 0: $p(\mathcal{D})$: distribution over tasks. 0: $\alpha,\beta$: step size hyper-parameters Initialize $\theta$ while not done do Sample batch of tasks $\\{D^{l}\\}=\\{(S^{l},Q^{l})\\}\sim p(\mathcal{D})$ for all $(S^{l},Q^{l})$ do Initialize $\theta_{l}^{(0)}=\theta$ for all steps k do Compute: $\theta_{l}^{(k+1)}=\theta_{l}^{(k)}-\alpha(\nabla_{\theta_{l}^{(k)}}\mathcal{L}_{S_{l}}(f_{\theta_{l}^{(k)}}))$ end for end for Update $\theta=\theta-\beta($MetaUpdate$(f_{\theta_{l}^{(K)}},Q^{l}))$ end while Meta-learning, or learning to learn, aims to create models that can learn new skills or adapt to new tasks rapidly from few training examples. Unlike traditional machine learning, datasets for either training or testing, which are referred to as meta-train and meta-test datasets, comprise of many tasks sampled from a distribution of tasks $p(\mathcal{D})$ rather than individual data points. Each task is associated with a dataset $\mathcal{D}$ which contains both feature vectors and ground truth labels and is split into a support set and a query set, $\mathcal{D}=\\{S,Q\\}$. The support set is used for fast adaptation and the query set is used to evaluate performance and compute a loss with respect to model parameter initialization. Generally, some model $f_{\theta}$ parameterized by $\theta$, often referred to as the base- learner, is considered. A cycle of fast-adaptation on a support-set followed by updating the parameter initialization of the base-learner based on the loss on the query-set is called an episode. In the case of classification, the optimal parameters maximize the probability of the true labels across multiple batches $Q\subset\mathcal{D}$ $\displaystyle\displaystyle\theta^{*}:=arg\underset{\theta}{max}\mathbb{E}_{Q\subset\mathcal{D}}[\sum_{(x,y)\in Q}P_{\theta}(y|x)]$ (1) In few-shot classification/fast learning, the goal is to minimize the prediction error on data samples with unknown labels given a small support set for learning. Meta-training (Algorithm 1) consists of updating the parameters of the base-learner by performing many of the formerly described episodes, until some stop criterion is reached. Following this procedure, the extended definition of optimal parameters is given in Eq. 2 to include fast adaptation based on the support set. The underlined parts mark the difference between traditional supervised-learning and meta-learning. The optimal parameters $\theta^{*}$ are obtained by solving $\displaystyle\scriptstyle arg\underset{\theta}{max}\underline{\mathbb{E}_{l\subset L}[}\mathbb{E}_{\underline{S^{l}\subset\mathcal{D}},Q^{l}\subset\mathcal{D}}[\sum_{(x,y)\in Q^{l}}P_{\theta}(y|x,\underline{S^{l}})]\underline{]}$ (2) In this work, we focus on metric- and optimization-based meta-learning algorithms. In the following sections, their respective characteristics and the update methods in Algorithm 1 are introduced. ### 2.1 Prototypical Networks Prototypical Networks Snell et al. (2017) belong to the metric-based family of meta-learning algorithms. Typically they consist of an embedding network $f_{\theta}$ and a distance function $d(x_{1},x_{2})$ such as Euclidean distance. The embedding network is used to encode all samples in the support set $S_{c}$ and compute prototypes $\mu_{c}$ per class $c\in C$ by computing the mean of the sample encodings of that respective class $\displaystyle\mu_{c}:=\frac{1}{|S_{c}|}\sum_{(x_{i},y_{i})\in S_{c}}f_{\theta}(x_{i})$ (3) Using the computed prototypes, Prototypical Networks classify a new sample as $\displaystyle p(y=c|x)=\frac{exp(-d(f_{\theta}(x),\mu_{c})}{\sum_{c^{^{\prime}}\in C}exp(-d(f_{\theta}(x),\mu_{c^{^{\prime}}})}$ (4) Wang et al. (2019) show that despite their simplicity, Prototypical Networks can perform on par or better than other state-of-the-art meta-learning methods when all sample encodings are centered around the overall mean of all classes and consecutively L2-normalized. We also adopt this strategy. ### 2.2 MAML Model-Agnostic Meta-Learning (MAML) Finn et al. (2017) is an optimization- based method that uses the following objective function $\displaystyle\theta^{*}:=arg\underset{\theta}{min}\sum_{D_{l}\sim p(D)}\mathcal{L}_{l}(f_{\theta_{l}^{(k)}})$ (5) $\mathcal{L}_{l}(f_{\theta_{l}^{(k)}})$ is the loss on the query set after updating the base-learner for $k$ steps on the support set. Hence, MAML directly optimizes the base-learner such that fast-adaptation of $\theta$, often referred to as inner-loop optimization, results in task-specific parameters $\theta_{l}^{(k)}$ which generalize well on the task. Setting $B$ as the batch size, MAML implements its MetaUpdate, which is also referred to as outer-loop optimization, as $\displaystyle\theta=\theta-\beta\frac{1}{B}\sum_{D_{l}\sim p(\mathcal{D})}(\nabla_{\theta}\mathcal{L}_{l}(f_{\theta_{l}^{(k)}}))$ (6) Such a MetaUpdate requires computing second order derivatives and, in turn, holding $\theta_{l}^{(j)}\forall j=1,\dots,k$ in memory. A first-order approximation of MAML (foMAML), which ignores second order derivatives, can be used to bypass this problem: $\displaystyle\theta=\theta-\beta\frac{1}{B}\sum_{D_{l}\sim p(\mathcal{D})}(\nabla_{\theta_{l}^{(k)}}\mathcal{L}_{l}(f_{\theta_{l}^{(k)}}))$ (7) Following previous work Antoniou et al. (2018), we also adopt the following improvements in our framework for all MAML-based methods: #### Per-step Layer Normalization weights Layer normalization weights and biases are not updated in the inner-loop. Sharing one set of weights and biases across inner-loop steps implicitly assumes that the feature distribution between layers stays the same at every step of the inner optimization. #### Per-layer per-step learnable inner-loop learning rate Instead of using a shared learning rate for all parameters, the authors propose to initialize a learning rate per layer and per step and jointly learn their values in the MetaUpdate steps. #### Cosine annealing of outer-loop learning rate It has shown to be crucial to model performance to anneal the learning rate using some annealing function Loshchilov and Hutter (2016). ### 2.3 Reptile Reptile Nichol et al. (2018) is a first-order optimization-based meta-learning algorithm which is designed to move the weights towards a manifold of the weighted averages of task-specific parameters $\theta_{l}^{(k)}$: $\displaystyle\theta=\theta-\beta\frac{1}{B}\sum_{D^{l}\sim p(\mathcal{D})}(\theta_{l}^{(k)}-\theta)$ (8) Despite its simplicity, it has shown competitive or superior performance against MAML, e.g., on Natural Language Understanding Dou et al. (2019). ### 2.4 ProtoMAML Triantafillou et al. (2020) introduce ProtoMAML as a meta-learning method which combines the complementary strengths of Prototypical Networks and MAML by leveraging the inductive bias of the use of prototypes instead of random initialization of the final linear layer of the network. Snell et al. (2017) show that Prototypical Networks are equivalent to a linear model when Euclidean distance is used. Using the definition of prototypes $\mu_{c}$ as per Eq. 3, the weights $w_{c}$ and bias $b_{c}$ corresponding to class $c$ can be computed as follows $\displaystyle\mathbf{w}_{c}:=2\mu_{c}\qquad b_{c}:=-\mu_{c}^{T}\mu_{c}$ (9) ProtoMAML is defined as the adaptation of MAML where the final linear layer is parameterized as per Eq. 9 at the start of each episode using the support set. Due to this initialization, it allows modeling a varying number of classes per episode. #### ProtoMAMLn Inspired by Wang et al. (2019), we propose a simple, yet effective adaptation to ProtoMAML by applying $L_{2}$ normalization to the prototypes themselves, referred to as ProtoMAMLn, and, again, use a first-order approximation (foProtoMAMLn). We demonstrate that doing so leads to a more stable, faster and effective learning algorithm at only constant extra computational cost ($\mathcal{O}(1))$. We hypothesize the normalization to be particularly beneficial in case of a relatively high-dimensional final feature space – in case of BERT-like models typically 768 dimensions. Let $x$ be a sample and $\hat{x}=f_{\theta}(x)$ be the encoding of the sample in the final feature space. Since the final activation function is the tanh activation, all entries of both $\hat{x}$ and $\mu_{c}$ have values between -1 and 1. The pre-softmax activation for class $c$ is computed as $\hat{x}^{T}\mu_{c}$. Due to the size of the vectors and the scale of their respective entries, this in-product can yield a wide range of values, which in turn results in relatively high loss values, making the inner-loop optimization unstable. ## 3 Related work ### 3.1 Multilingual NLP Just as the deep learning era for monolingual NLP started with the invention of dense, low-dimensional vector representations for words Mikolov et al. (2013b) so did cross-lingual NLP with works like those of Mikolov et al. (2013a); Faruqui et al. (2014). More recently, multilingual and/or cross- lingual NLP is approached by training one shared encoder for multiple languages at once, either by explicitly aligning representations with the use of parallel corpora Artetxe and Schwenk (2019); Lample and Conneau (2019) or by jointly training on some monolingual language model objective, such as the Masked Language Model (MLM) Devlin et al. (2018), in multiple languages Devlin et al. (2018); Conneau et al. (2019). The formerly described language models aim to create a shared embedding space for multiple languages with the hope that fine-tuning in one language does not degrade performance in others. Lai et al. (2019) argue that just aligning languages is not sufficient to generalize performance to new languages due to the phenomenon they describe as domain drift. Domain drift accounts for all differences for the same tasks in different languages which cannot be captured by a perfect translation system, such as differences in culture. They instead propose a multi-step approach which utilizes a multilingual teacher trained with Unsupervised Data Augmentation (UDA) Xie et al. (2019) to create labels for a student model that is pretrained on large amounts of unlabeled data in the target language and domain using the MLM objective. With their method, the authors obtain state-of-the-art results on the MLDoc document classification task Schwenk and Li (2018) and the Amazon Sentiment Polarity Review task Prettenhofer and Stein (2010). A downside, however, is the high computational cost involved. For every language and domain combination: 1) a machine translation system has to be inferred on a large amount of unlabeled samples; 2) the UDA method needs to be applied to obtain a teacher model to generate pseudo-labels on the unlabeled in-domain data; 3) a language model must be finetuned, which involves forwards and backwards computation of a softmax function over a large output space (e.g., 50k tokens for mBERT and 250k tokens for XLM-RoBERTa). The final classifier is then obtained by 4) training the finetuned language model on the pseudo-labels generated by the teacher. ### 3.2 Meta-learning in NLP #### Monolingual Bansal et al. (2019) apply meta-learning to a wide range of NLP tasks within a monolingual setting and show superior performance for parameter initialization over self-supervised pretraining and multi-task learning. Their method is an adaptation of MAML where a combination of a text-encoder, BERT Devlin et al. (2018), is coupled with a parameter generator that learns to generate task- dependent initializations of the classification head such that meta-learning can be performed across tasks with disjoint label spaces. Obamuyide and Vlachos (2019b) apply meta-learning on the task of relation extraction; Obamuyide and Vlachos (2019a) apply lifelong meta-learning for relation extraction; Chen et al. (2019) apply meta-learning for few-shot learning on missing link prediction in knowledge graphs. #### Multilingual Gu et al. (2018) apply meta-learning to Neural Machine Translation (NMT) and show its advantage over strong baselines such as cross-lingual transfer learning. By viewing each language pair as a task, the authors apply MAML to obtain competitive NMT systems with as little as 600 parallel sentences. To our best knowledge, the only application of meta-learning for cross-lingual few-shot learning is the one by Nooralahzadeh et al. (2020). The authors study the application of X-MAML, a MAML-based variant, to cross-lingual Natural Language Inference (XNLI) Conneau et al. (2018) and Multilingual Question Answering (MLQA) Lewis et al. (2019) in both a cross-domain and cross-language setting. X-MAML works by pretraining some model $M$ on a high-resource task $h$ to obtain initial model parameters $\theta_{mono}$. Consecutively, a set $L$ of one or more auxiliary languages is taken, and MAML is applied to achieve fast adaptation of $\theta_{mono}$ for $l\in L$. In their experiments, the authors use either one or two auxiliary languages and evaluate their method in both a zero- and few-shot setting. It should be noted that, in the few-shot setting, the full development set (2.5k instances) is used to finetune the model, which is not in line with other work on few-shot learning, such as Bansal et al. (2019). Also, there is a discrepancy in the training set used for the baselines and their proposed method. All reported baselines are either zero-shot evaluations of $\theta_{mono}$ or of $\theta_{mono}$ finetuned on the development set of the target language, whereas their proposed method additionally uses the development set in either one or two auxiliary languages during meta-training. MetaUpdate Method | Num inner-loop steps | Inner-loop lr | Class-head lr multiplier | Inner-optimizer lr ---|---|---|---|--- Reptile | 2,3,5 | 1e-5, 5e-5, 1e-4 | 1, 10 | - foMAML | 2,3,5 | 1e-5, 1e-4, 1e-3 | 1, 10 | 3e-5, 6e-5, 1e-4 foProtoMAMLn | 2,3,5 | 1e-5, 1e-4, 1e-3 | 1, 10 | 3e-5, 6e-5, 1e-4 Table 1: Search range per hyper-parameter. We consider the number of update steps in the inner-loop, Num inner-loop steps, the (initial) learning rate of the inner-loop, Inner-loop lr, the factor by which the learning rate of the classification head is multiplied, Class-head lr multiplier, and, if applicable, the learning rate with which the inner-loop optimizer is updated, Inner-optimizer lr. The chosen value is underlined. ## 4 Data In this section, we give an overview of the datasets we use and the respective classification tasks. #### MLDoc Schwenk and Li (2018) published an improved version of the Reuters Corpus Volume 2 Lewis et al. (2004) with balanced class priors for all languages. MLDoc consists of news stories in 8 languages: English, Spanish, French, Italian, Russian, Japanese and Chinese. Each news story is manually classified into one of four groups: Corporate/Industrial, Economics, Government/Social and Markets. The train datasets contain 10k samples whereas the test sets contain 4k samples. #### Amazon Sentiment Polarity Another widely used dataset for cross-lingual text classification is the Amazon Sentiment Analysis dataset Prettenhofer and Stein (2010). The dataset is a collection of product reviews in English, French, German and Japanese in three categories: books dvds and music. Each sample consists of the original review accompanied by meta-data such as the rating of the reviewed product expressed as an integer on a scale from one to five. In this work, we consider the sentiment polarity task where we distinguish between positive (rating $>$ 3) and negative (rating $<$ 3) reviews. When all product categories are concatenated, the dataset consists of 6K samples per language per dataset (train, test). We extend this with Chinese product reviews in the cosmetics domain from JD.com Zhang et al. (2015), a large e-commerce website in China. The train and test sets contain 2k and 20k samples respectively. ## 5 Experiments We use XLM-RoBERTa Conneau et al. (2019), a strong multilingual model, as the base-learner in all models. We quantify the strengths and weaknesses of meta- learning as opposed to traditional supervised learning in both a cross- and a multilingual joint-training setting with limited resources. #### Cross-lingual adaptation Here, the available data is split into multiple subsets: the auxiliary languages $l_{aux}$ which are used in meta-training, the validation language $l_{dev}$ which is used to monitor performance, and the target languages $l_{tgt}$ which are kept unseen until meta-testing. Two scenarios in terms of amounts of available data are considered. A small sample of the available training data of $l_{aux}$ is taken to create a limited-resource setting, whereas all available training data of $l_{aux}$ is used in a high-resource setting. The chosen training data per language is split evenly and stratified over two disjoint sets from which the meta-training support and query samples are sampled, respectively. For meta-testing, one batch (16 samples) is taken from the training data of each target language as support set, while we test on the whole test set per target language (i.e., the query set). #### Multilingual joint training We also investigate meta-learning as an approach to multilingual joint- training in the same limited-resource setting as previously described for the cross-lingual experiments. The difference is that instead of learning to generalize to $l_{tgt}\neq l_{aux}$ from few examples, here $l_{tgt}=l_{aux}$. If we can show that one can learn many similar tasks across languages from few examples per language, using a total number of examples in the same order of magnitude as in “traditional” supervised learning for training a monolingual classifier, this might be an incentive to change data collection processes in practice. For both experimental settings above, we examine the influence of additionally using all training data from a high-resource language $l_{src}$ during meta- training, English. $\mathbf{l_{src}}$ = en | Method | Limited-resource setting | High-resource setting | | ---|---|---|---|---|--- de | fr | it | ja | ru | zh | $\Delta$ | de | fr | it | ja | ru | zh | $\Delta$ Excluded | Non-episodic | 82.0 | 86.7 | 68.3 | 71.9 | 70.9 | 81.0 | 76.8 | 95.3 | 90.9 | 80.9 | 82.9 | 74.5 | 89.6 | 85.7 ProtoNet | 90.5 | 85.0 | 76.6 | 75.0 | 69.6 | 82.0 | 79.8 | 95.5 | 91.7 | 82.0 | 82.2 | 76.6 | 87.4 | 85.9 foMAML | 89.7 | 85.5 | 74.1 | 74.1 | 74.0 | 83.2 | 80.1 | 95.0 | 91.4 | 81.4 | 82.7 | 76.9 | 87.8 | 86.1 foProtoMAMLn | 90.6 | 86.2 | 77.8 | 75.6 | 73.6 | 83.8 | 80.7 | 95.6 | 92.1 | 82.6 | 83.1 | 77.9 | 88.9 | 86.7 Reptile | 87.9 | 81.8 | 72.7 | 74.4 | 73.9 | 80.9 | 78.6 | 95.0 | 90.1 | 81.1 | 82.7 | 72.5 | 88.7 | 85.0 Included | Zero-shot | 92.4 | 92.1 | 80.3 | 81.0 | 71.7 | 89.1 | 84.4 | 92.4 | 92.1 | 80.3 | 81.0 | 71.7 | 89.1 | 84.4 Non-episodic | 93.7 | 91.3 | 81.5 | 80.6 | 71.1 | 88.4 | 84.4 | 93.7 | 92.9 | 82.4 | 82.3 | 72.1 | 90.1 | 85.6 ProtoNet | 93.4 | 91.9 | 79.1 | 81.3 | 72.2 | 87.8 | 84.5 | 95.0 | 91.7 | 81.1 | 82.7 | 72.0 | 88.0 | 85.9 foMAML | 95.1 | 91.2 | 79.5 | 79.6 | 73.3 | 89.7 | 84.6 | 94.8 | 93.2 | 79.9 | 82.4 | 75.7 | 90.6 | 86.1 foProtoMAMLn | 94.9 | 91.7 | 81.5 | 81.4 | 75.2 | 89.9 | 85.5 | 95.8 | 94.1 | 82.7 | 83.0 | 81.2 | 90.4 | 87.9 | Reptile | 92.3 | 91.4 | 79.7 | 79.5 | 71.8 | 88.1 | 83.8 | 94.8 | 91.0 | 80.2 | 82.0 | 72.7 | 89.9 | 85.1 Table 2: Average accuracy of 5 different seeds on the unseen target languages for MLDoc. $\Delta$ corresponds to the average accuracy across test languages. ### 5.1 Specifics per dataset #### MLDoc As MLDoc has sufficient languages, we set $l_{src}=$ English and $l_{dev}=$ Spanish. The remaining languages are split in two groups: $l_{aux}=\\{\textrm{German, Italian, Japanese}\\}$; and $l_{tgt}=\\{\textrm{French, Russian, Chinese}\\}$. In the limited-resource setting, we randomly sample 64 samples per language in $l_{aux}$ for training. Apart from comparing low- and high-resource settings, we also quantify the influence of augmenting the training set $l_{aux}$ with a high-resource source language $l_{src}$, English. #### Amazon Sentiment Polarity The fact that the Amazon dataset (augmented with Chinese) comprises of only five languages has some implications for our experimental design. In the cross-lingual experiments, where $l_{aux}$, $l_{dev}$ and $l_{tgt}$ should be disjoint, only three languages, including English, remain for meta-training. As we consider two languages too little data for meta-training, we do not experiment with leaving out the English data. Hence, for meta-training, the data consists of $l_{src}=$ English, as well as two languages in $l_{aux}$. We always keep one language unseen until meta-testing, and alter $l_{aux}$ such that we can meta-test on every language. We set $l_{dev}=$ French in all cases except when French is used as the target language; then, $l_{dev}=$ Chinese. In the limited-resource setting, a total of 128 samples per language in $l_{aux}$ is used. For the multilingual joint-training experiments there are enough languages available to quantify the influence of English during meta-training. When English is excluded, it is used for meta-validation. When included, we average results over two sets of experiments: one where $l_{dev}=$ French and one where $l_{dev}=$ Chinese. Method | Limited-resource setting | | High-resource setting | ---|---|---|---|--- de | fr | ja | zh | $\Delta$ | de | fr | ja | zh | $\Delta$ Zero-shot | 91.2 | 90.7 | 87.0 | 84.6 | 88.4 | 91.2 | 90.7 | 87.0 | 84.6 | 88.4 Non-episodic | 90.9 | 90.6 | 86.1 | 86.9 | 88.6 | 91.6 | 91.0 | 85.5 | 87.9 | 89.0 ProtoNet | 89.7 | 90.2 | 86.6 | 85.2 | 87.9 | 90.7 | 92.0 | 86.7 | 84.0 | 88.4 foMAML | 88.3 | 90.5 | 86.8 | 88.1 | 88.4 | 91.4 | 92.5 | 88.0 | 90.4 | 90.6 foProtoMAMLn | 89.0 | 91.1 | 87.3 | 88.8 | 89.1 | 92.0 | 93.1 | 88.6 | 89.8 | 90.9 Reptile | 88.1 | 87.9 | 86.8 | 87.5 | 87.6 | 90.6 | 91.7 | 87.3 | 86.2 | 89.0 Table 3: Average accuracy of 5 different seeds on the unseen target languages for Amazon. $\Delta$ corresponds to the average accuracy across test languages. ### 5.2 Baselines We introduce baselines trained in a standard supervised, non-episodic fashion. Again, we use XLM-RoBERTa-base as the base-learner in all models. #### Zero-shot This baseline assumes sufficient training data for the task to be available in one language $l_{src}$ (English). The base-learner is trained in a non- episodic manner using mini-batch gradient descent with cross-entropy loss. Performance is monitored during training on a held-out validation set in $l_{src}$, the model with the lowest loss is selected, and then evaluated on the same task in the target languages. #### Non-episodic The second baseline aims to quantify the exact impact of learning a model through the meta-learning paradigm versus standard supervised learning. The model learns from exactly the same data as the meta-learning algorithms, but in a non-episodic manner: i.e., merging support and query sets in $l_{aux}$ (and $l_{src}$ when included) and training using mini-batch gradient descent with cross-entropy loss. During testing, the trained model is independently finetuned for 5 steps on the support set (one mini-batch) of each target language $l_{tgt}$. ### 5.3 Training setup and hyper-parameters We use the Ranger optimizer, an adapted version of Adam Kingma and Ba (2014) with improved stability at the beginning of training – by accounting for the variance in adaptive learning rates Liu et al. (2019) – and improved robustness and convergence speed Zhang et al. (2019); Yong et al. (2020). We use a batch size of 16 and a learning rate of 3e-5 to which we apply cosine annealing. For meta-training, we perform 100 epochs of 100 episodes and perform evaluation with 5 different seeds on the meta-validation set after each epoch. One epoch consists of 100 update steps where each update step consists of a batch of 4 episodes. Early-stopping with a patience of 3 epochs is performed to avoid overfitting. For the non-episodic baselines, we train for 10 epochs on the auxiliary languages while validating after each epoch. All models are created using the PyTorch library Paszke et al. (2017) and trained on a single 24Gb NVIDIA Titan RTX GPU. We perform grid search on MLDoc in order to determine optimal hyperparameters for the MetaUpdate methods. The hyper-parameters resulting in the lowest loss on $l_{dev}=$ Spanish are used in all experiments. The number of update steps in the inner-loop is 5; the (initial) learning rate of the inner-loop is 1e-5 for MAML and ProtoMAML and 5e-5 for Reptile; the factor by which the learning rate of the classification head is multiplied is 10 for MAML and ProtoMAML and 1 for Reptile; when applicable, the learning rate with which the inner-loop optimizer is updated is 6e-5. See Table 1 for the considered grid. $\mathbf{l_{src}}$ = en | Method | Amazon | MLDoc ---|---|---|--- de | fr | ja | zh | $\Delta$ | de | fr | it | ja | ru | zh | $\Delta$ Excluded | Non-episodic | 88.4 | 88.6 | 85.7 | 88.2 | 87.7 | 92.8 | 89.1 | 81.2 | 83.2 | 84.0 | 87.4 | 86.3 ProtoNet | 86.7 | 88.0 | 86.2 | 87.3 | 87.1 | 89.7 | 87.6 | 80.5 | 82.2 | 80.6 | 85.2 | 84.3 foMAML | 88.3 | 87.5 | 84.6 | 89.1 | 86.3 | 94.1 | 89.7 | 81.5 | 84.2 | 77.6 | 87.5 | 85.8 foProtoMAMLn | 88.9 | 89.5 | 86.5 | 89.0 | 88.5 | 94.8 | 89.5 | 81.5 | 84.8 | 81.0 | 88.7 | 86.6 Reptile | 86.1 | 86.3 | 82.9 | 87.0 | 85.6 | 92.4 | 88.2 | 80.5 | 82.5 | 79.5 | 87.8 | 85.3 Included | Non-episodic | 91.0 | 91.0 | 87.3 | 89.4 | 89.8 | 94.9 | 92.1 | 84.7 | 84.8 | 83.7 | 91.4 | 88.6 ProtoNet | 90.3 | 91.3 | 87.5 | 88.7 | 89.5 | 95.5 | 91.7 | 83.4 | 85.1 | 82.8 | 88.3 | 87.8 foMAML | 90.1 | 90.7 | 87.2 | 89.5 | 89.4 | 95.1 | 92.5 | 83.1 | 84.9 | 84.3 | 90.6 | 88.4 foProtoMAMLn | 90.7 | 91.5 | 88.0 | 90.4 | 90.2 | 96.0 | 93.6 | 85.0 | 85.7 | 84.8 | 90.8 | 89.3 Reptile | 90.0 | 89.5 | 86.5 | 87.6 | 88.4 | 94.4 | 93.1 | 83.8 | 85.2 | 83.6 | 90.4 | 88.4 Table 4: Average accuracy of 5 different seeds on the target languages in the joint-training setting for MLDoc and Amazon. $\Delta$ corresponds to the average accuracy across test languages. ## 6 Results #### Cross-lingual adaptation Tables 2 and 3 show the accuracy scores on the target languages on MLDoc and Amazon respectively. We start by noting the strong multilingual capabilities of XLM-RoBERTa as our base-learner: Adding the full training datasets in three extra languages (i.e., comparing the zero-shot with the non-episodic baseline in the high-resource, ‘Included’ setting) results in a mere 1.2% points increase in accuracy on average for MLDoc and 0.6% points for Amazon. Although the zero-shot333The zero-shot baseline is only applicable in the ‘Included’ setting, as the English data is not available under ‘Excluded’. and non- episodic baselines are strong, in the majority of cases, a meta-learning approach improves performance. This holds especially for our version of ProtoMAML (ProtoMAMLn), which achieves the highest average accuracy in all considered settings. The substantial improvements for Russian on MLDoc and Chinese on Amazon indicate that meta-learning is most advantageous when the considered task distribution is somewhat heterogeneous or, in other words, when domain drift Lai et al. (2019) is present. For the Chinese data used for the sentiment polarity task, the presence of domain drift is obvious as the data is collected from a different website and concerns different products than the other languages. For Russian in the MLDoc dataset, it holds that the non- episodic baseline has the smallest gain in performance when adding English data ($l_{src}$) in the limited-resource setting (0.2% absolute gain as opposed to 5.7% on average for the remaining languages) and even a decrease of 2.4% points when adding English data in the high-resource setting. Especially for these languages with domain drift, our version of ProtoMAML (foProtoMAMLn) outperforms the non-episodic baselines with a relatively large margin. For instance, in Table 2 in the high-resource setting with English included during training, foProtoMAMLn improves over the non-episodic baseline with 9.1% points whereas the average gain over the remaining languages is 0.9% points. A similar trend can be seen in Table 3 where, in the limited-resource setting, foProtoMAMLn outperforms the non-episodic baseline with 1.9% points on Chinese, with comparatively smaller gains on average for the remaining languages. #### Joint training In this setting, we achieve a new state of the art on MLDoc for German, Italian, Japanese and Russian using our method, foProtoMAMLn (Table 4).444The zero-shot baselines are the same as in Tables 2 and 3. The previous state of the art for German and Russian is held by Lai et al. (2019) (95.73% and 84.65% respectively). For Japanese and Italian, it is held by Eisenschlos et al. (2019) (80.55% and 80.12% respectively). The state of the art for French and Chinese is also held by Lai et al. (2019) (96.05% and 93.32% respectively). On the Amazon dataset, foProtoMAMLn also outperforms all other methods on average. The state of the art is held by Lai et al. (2019) with 93.3%, 94.2% and 90.6% for French, German and Chinese respectively and, although we do not outperform it, the differences are rather small – between 0.2% (Chinese) and 3.4% points (German) – even when grid search is based on MLDoc, while we use a much less computationally expensive approach. Figure 1: Validation accuracy for 3 seeds for original foProtoMAML and our new method, foProtoMAMLn. Again, we use Russian in MLDoc to exemplify the difference between meta- learning and standard supervised learning. When comparing the difference in performance between excluding and including English meta-training episodes ($l_{src}$), opposite trends are noticeable: for standard supervised, non- episodic learning, performance drops slightly by 0.3%, whereas all meta- learning algorithms gain between 2.2% and 6.7% in absolute accuracy. This confirms our earlier finding that meta-learning benefits from, and usefully exploits heterogeneity in data distributions; in contrast, this harms performance in the standard supervised-learning case. Dataset | de | fr | it | ja | ru | zh | Diff ---|---|---|---|---|---|---|--- Amazon | 90.4 | 90.9 | - | 87.3 | - | 88.3 | -1.7 MLDoc | 92.8 | 92.4 | 78.6 | 79.3 | 69.3 | 88.9 | -4.3 Table 5: Average accuracy of 5 different seeds on unseen target languages using the original/unnormalized foProtoMAML model. Diff is the difference in average accuracy $\Delta$ across languages against foProtoMAMLn. Method | Limited-resource setting | | High-resource setting | ---|---|---|---|--- de | fr | ja | zh | Diff | de | fr | ja | zh | Diff ProtoNet | 91.1 | 90.9 | 87.1 | 85.5 | +0.75 | 91.3 | 91.1 | 87.4 | 88.7 | +1.44 foMAML | 90.8 | 87.4 | 87.3 | 85.2 | -0.75 | 91.7 | 91.2 | 87.2 | 88.1 | -1.13 foProtoMAMLn | 87.7 | 87.8 | 83.9 | 84.4 | -3.1 | 90.8 | 89.8 | 86.2 | 82.3 | -3.96 Reptile | 89.3 | 90.2 | 86.7 | 85.5 | +0.35 | 90.0 | 89.3 | 87.1 | 85.7 | -1.04 Table 6: Average accuracy of 5 different seeds on unseen target languages for Amazon when initializing from monolingual classifier in $l_{src}$. Diff: difference in average accuracy $\Delta$ across languages compared to initializing from the XLM-RoBERTa language model. ## 7 Ablations #### foProtoMAMLn Figure 1 shows the development of the validation accuracy during training for 25 epochs for the original foProtoMAML and our model, foProtoMAMLn. By applying $L_{2}$ normalization to the prototypes, we obtain a more stable version of foProtoMAML which empirically converges faster. We furthermore re- run the high-resource experiments with English for both MLDoc and Amazon using the original foProtoMAML (Table 5) and find it performs 4.3% and 1.7% accuracy points worse on average, respectively, further demonstrating the effectiveness of our approach. #### Initializing from a monolingual classifier In our experiments, we often assume the presence of a source language (English). We now investigate (in the $l_{src}$ = en ‘Excluded’ setting) whether it is beneficial to pre-train the base-learner in a standard supervised way on this source language and use the obtained checkpoint $\theta_{mono}$ as an initialization for meta-training (Table 6) rather than initializing from the transformer checkpoint. We observe that only ProtoNet consistently improves performance, whereas foProtoMAMLn suffers the most with a decrease of 3.1% and 3.96% in accuracy in the low- and high-resource setting respectively. We surmise this difference is attributable to two factors. Intuitively, the monolingual classifier aims to learn a transformation from the input space to the final feature space, from which the prototypes for ProtoNet and ProtoMAML are created, in which the learned classes are encoded in their own disjoint sub-spaces such that a linear combination of these features can be used to correctly classify instances. ProtoNet aims to learn a similar transformation, but uses a Nearest Neighbours approach to classify instances instead. ProtoMAML on the other hand benefits the most from prototypes which can be used to classify instances after the inner-loop updates have been performed. This, in combination with the fact that the first-order approximation of ProtoMAML cannot differentiate through the creation of the prototypes, could explain the difference in performance gain with respect to ProtoNet. ## 8 Conclusion We proposed a meta-learning framework for few-shot cross- and multilingual joint-learning for document classification tasks in different domains. We demonstrated that it leads to consistent gains over traditional supervised learning on a wide array of data availability and diversity settings, and showed that it thrives in settings with a heterogenous task distribution. We presented an effective adaptation to ProtoMAML and, among others, obtained a new state of the art on German, Italian, Japanese and Russian in the few-shot setting on MLDoc. ## 9 Acknowledgements This work was supported by Deloitte Risk Advisory B.V., the Netherlands. ## References * Antoniou et al. (2018) Antreas Antoniou, Harrison Edwards, and Amos Storkey. 2018. How to train your maml. _arXiv preprint arXiv:1810.09502_. * Artetxe and Schwenk (2019) Mikel Artetxe and Holger Schwenk. 2019. Massively multilingual sentence embeddings for zero-shot cross-lingual transfer and beyond. _Transactions of the Association for Computational Linguistics_ , 7:597–610. * Bansal et al. (2019) Trapit Bansal, Rishikesh Jha, and Andrew McCallum. 2019. Learning to few-shot learn across diverse natural language classification tasks. _arXiv preprint arXiv:1911.03863_. * Bengio et al. (1990) Yoshua Bengio, Samy Bengio, and Jocelyn Cloutier. 1990. _Learning a synaptic learning rule_. Citeseer. * Chen et al. (2019) Mingyang Chen, Wen Zhang, Wei Zhang, Qiang Chen, and Huajun Chen. 2019. Meta relational learning for few-shot link prediction in knowledge graphs. In _Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)_ , pages 4217–4226, Hong Kong, China. Association for Computational Linguistics. * Conneau et al. (2019) Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Unsupervised cross-lingual representation learning at scale. _arXiv preprint arXiv:1911.02116_. * Conneau et al. (2018) Alexis Conneau, Guillaume Lample, Ruty Rinott, Adina Williams, Samuel R Bowman, Holger Schwenk, and Veselin Stoyanov. 2018. Xnli: Evaluating cross-lingual sentence representations. _arXiv preprint arXiv:1809.05053_. * Devlin et al. (2018) Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. _arXiv preprint arXiv:1810.04805_. * Dou et al. (2019) Zi-Yi Dou, Keyi Yu, and Antonios Anastasopoulos. 2019. Investigating meta-learning algorithms for low-resource natural language understanding tasks. _arXiv preprint arXiv:1908.10423_. * Eisenschlos et al. (2019) Julian Eisenschlos, Sebastian Ruder, Piotr Czapla, Marcin Kardas, Sylvain Gugger, and Jeremy Howard. 2019. Multifit: Efficient multi-lingual language model fine-tuning. _arXiv preprint arXiv:1909.04761_. * Faruqui et al. (2014) Manaal Faruqui, Jesse Dodge, Sujay K Jauhar, Chris Dyer, Eduard Hovy, and Noah A Smith. 2014. Retrofitting word vectors to semantic lexicons. _arXiv preprint arXiv:1411.4166_. * Finn et al. (2017) Chelsea Finn, Pieter Abbeel, and Sergey Levine. 2017. Model-agnostic meta-learning for fast adaptation of deep networks. In _Proceedings of the 34th International Conference on Machine Learning-Volume 70_ , pages 1126–1135. JMLR. org. * Gu et al. (2018) Jiatao Gu, Yong Wang, Yun Chen, Kyunghyun Cho, and Victor OK Li. 2018. Meta-learning for low-resource neural machine translation. _arXiv preprint arXiv:1808.08437_. * van der Heijden et al. (2019) Niels van der Heijden, Samira Abnar, and Ekaterina Shutova. 2019. A comparison of architectures and pretraining methods for contextualized multilingual word embeddings. _arXiv preprint arXiv:1912.10169_. * Kingma and Ba (2014) Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. _arXiv preprint arXiv:1412.6980_. * Koch et al. (2015) Gregory Koch, Richard Zemel, and Ruslan Salakhutdinov. 2015. Siamese neural networks for one-shot image recognition. In _ICML deep learning workshop_ , volume 2. Lille. * Lai et al. (2019) Guokun Lai, Barlas Oguz, and Veselin Stoyanov. 2019. Bridging the domain gap in cross-lingual document classification. _arXiv preprint arXiv:1909.07009_. * Lample and Conneau (2019) Guillaume Lample and Alexis Conneau. 2019. Cross-lingual language model pretraining. _arXiv preprint arXiv:1901.07291_. * Lample et al. (2017) Guillaume Lample, Alexis Conneau, Ludovic Denoyer, and Marc’Aurelio Ranzato. 2017\. Unsupervised machine translation using monolingual corpora only. _arXiv preprint arXiv:1711.00043_. * Lewis et al. (2004) David D Lewis, Yiming Yang, Tony G Rose, and Fan Li. 2004. Rcv1: A new benchmark collection for text categorization research. _Journal of machine learning research_ , 5(Apr):361–397. * Lewis et al. (2019) Patrick Lewis, Barlas Oğuz, Ruty Rinott, Sebastian Riedel, and Holger Schwenk. 2019. Mlqa: Evaluating cross-lingual extractive question answering. _arXiv preprint arXiv:1910.07475_. * Liu et al. (2019) Liyuan Liu, Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Jiawei Han. 2019. On the variance of the adaptive learning rate and beyond. _arXiv preprint arXiv:1908.03265_. * Loshchilov and Hutter (2016) Ilya Loshchilov and Frank Hutter. 2016. Sgdr: Stochastic gradient descent with warm restarts. _arXiv preprint arXiv:1608.03983_. * Mikolov et al. (2013a) Tomas Mikolov, Quoc V Le, and Ilya Sutskever. 2013a. Exploiting similarities among languages for machine translation. _arXiv preprint arXiv:1309.4168_. * Mikolov et al. (2013b) Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013b. Distributed representations of words and phrases and their compositionality. In _Advances in neural information processing systems_ , pages 3111–3119. * Nichol et al. (2018) Alex Nichol, Joshua Achiam, and John Schulman. 2018. On first-order meta-learning algorithms. _arXiv preprint arXiv:1803.02999_. * Nooralahzadeh et al. (2020) Farhad Nooralahzadeh, Giannis Bekoulis, Johannes Bjerva, and Isabelle Augenstein. 2020. Zero-shot cross-lingual transfer with meta learning. _arXiv preprint arXiv:2003.02739_. * Obamuyide and Vlachos (2019a) Abiola Obamuyide and Andreas Vlachos. 2019a. Meta-learning improves lifelong relation extraction. In _Proceedings of the 4th Workshop on Representation Learning for NLP (RepL4NLP-2019)_ , pages 224–229, Florence, Italy. Association for Computational Linguistics. * Obamuyide and Vlachos (2019b) Abiola Obamuyide and Andreas Vlachos. 2019b. Model-agnostic meta-learning for relation classification with limited supervision. In _Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics_ , pages 5873–5879, Florence, Italy. Association for Computational Linguistics. * Paszke et al. (2017) Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. 2017. Automatic differentiation in pytorch. In _NIPS 2017 Workshop Autodiff Submission_. * Prettenhofer and Stein (2010) Peter Prettenhofer and Benno Stein. 2010. Cross-language text classification using structural correspondence learning. In _Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics_ , pages 1118–1127, Uppsala, Sweden. Association for Computationprettenhoferal Linguistics. * Ravi and Larochelle (2017) Sachin Ravi and Hugo Larochelle. 2017. Optimization as a model for few-shot learning. In _International Conference on Learning Representations_. * Schmidhuber (1987) Jurgen Schmidhuber. 1987. Evolutionary principles in self-referential learning. _On learning how to learn: The meta-meta-… hook.) Diploma thesis, Institut f. Informatik, Tech. Univ. Munich_ , 1(2). * Schwenk and Li (2018) Holger Schwenk and Xian Li. 2018. A corpus for multilingual document classification in eight languages. In _Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018)_ , Paris, France. European Language Resources Association (ELRA). * Snell et al. (2017) Jake Snell, Kevin Swersky, and Richard Zemel. 2017. Prototypical networks for few-shot learning. In _Advances in neural information processing systems_ , pages 4077–4087. * Thrun and Pratt (1998) Sebastian Thrun and Lorien Pratt. 1998. Learning to learn: Introduction and overview. In _Learning to learn_ , pages 3–17. Springer. * Triantafillou et al. (2020) Eleni Triantafillou, Tyler Zhu, Vincent Dumoulin, Pascal Lamblin, Utku Evci, Kelvin Xu, Ross Goroshin, Carles Gelada, Kevin Swersky, Pierre-Antoine Manzagol, and Hugo Larochelle. 2020. Meta-dataset: A dataset of datasets for learning to learn from few examples. In _International Conference on Learning Representations_. * Wang et al. (2019) Yan Wang, Wei-Lun Chao, Kilian Q Weinberger, and Laurens van der Maaten. 2019. Simpleshot: Revisiting nearest-neighbor classification for few-shot learning. _arXiv preprint arXiv:1911.04623_. * Xie et al. (2019) Qizhe Xie, Zihang Dai, Eduard Hovy, Minh-Thang Luong, and Quoc V Le. 2019. Unsupervised data augmentation for consistency training. _arXiv preprint arXiv:1904.12848_. * Yong et al. (2020) Hongwei Yong, Jianqiang Huang, Xiansheng Hua, and Lei Zhang. 2020. Gradient centralization: A new optimization technique for deep neural networks. _arXiv preprint arXiv:2004.01461_. * Zhang et al. (2019) Michael Zhang, James Lucas, Jimmy Ba, and Geoffrey E Hinton. 2019. Lookahead optimizer: k steps forward, 1 step back. In _Advances in Neural Information Processing Systems_ , pages 9597–9608. * Zhang et al. (2015) Yongfeng Zhang, Min Zhang, Yi Zhang, Guokun Lai, Yiqun Liu, Honghui Zhang, and Shaoping Ma. 2015. Daily-aware personalized recommendation based on feature-level time series analysis. In _Proceedings of the 24th international conference on world wide web_ , pages 1373–1383.
# Compactness within the space of complete, constant $Q$-curvature metrics on the sphere with isolated singularities João Henrique Andrade , João Marcos do Ó and Jesse Ratzkin Institute of Mathematics and Statistics, University of São Paulo 05508-090, São Paulo-SP, Brazil and Department of Mathematics, Federal University of Paraíba 58051-900, João Pessoa-PB, Brazil<EMAIL_ADDRESS><EMAIL_ADDRESS>Department of Mathematics, Federal University of Paraíba 58051-900, João Pessoa-PB, Brazil<EMAIL_ADDRESS>Department of Mathematics, Universität Würzburg 97070, Würzburg-BA, Germany<EMAIL_ADDRESS> ###### Abstract. In this paper we consider the moduli space of complete, conformally flat metrics on a sphere with $k$ punctures having constant positive $Q$-curvature and positive scalar curvature. Previous work has shown that such metrics admit an asymptotic expansion near each puncture, allowing one to define an asymptotic necksize of each singular point. We prove that any set in the moduli space such that the distances between distinct punctures and the asymptotic necksizes all remain bounded away from zero is sequentially compact, mirroring a theorem of D. Pollack about singular Yamabe metrics. Along the way we define a radial Pohozaev invariant at each puncture and refine some a priori bounds of the conformal factor, which may be of independent interest. ###### Key words and phrases: Paneitz operator, $Q$-curvature, Critical exponent, Isolated singularities, Compactness, Pohozaev invariant ###### 2000 Mathematics Subject Classification: 35J60, 35B09, 35J30, 35B40 Research supported in part by Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq): grant 305726/2017-0, Coordenação de Aperfeiçoamento de Pessoal de Nível Superior (CAPES): grant 88882.440505/2019-01, and Fundação de Apoio à Pesquisa do Estado de São Paulo (FAPESP): grant 2020/07566-3 ## 1\. Introduction In 1960 H. Yamabe [28] proposed a program to find optimal metrics in a conformal class on a manifold of dimension at least three by minimizing the total scalar curvature functional, obtaining a constant scalar curvature representative in each conformal class. This program, which now bears his name, led to many advancements by N. Trudinger [26], T. Aubin [3], R. Schoen [25], and many others in the understanding of how geometry, topology and analysis interact with each other in compact Riemannian manifolds. The reader can find an excellent survey of the resolution of the original Yamabe problem in [17]. The lack of compactness of the group of conformal transformations of the sphere presents one of the many complications in carrying out Yamabe’s program. This same lack of compactness forces one to examine singular solutions which blow up along a closed subset. Many people continue to study both the regular and the singular Yamabe problems, and many open questions in both programs remain. More recent results include [19], in which the authors prove the set of solutions to the Yamabe problem within a conformal class is compact provided $3\leq n\leq 24$ and the conformal class is not the one of the round sphere. In recent years many people have pursued parts of Yamabe’s program for other notions of curvature. In the present note, we explore a part of the singular Yamabe program as applied to the fourth order $Q$-curvature, which is a higher order analog of scalar curvature. On a Riemannian manifold $(M,g)$ of dimension $n\geq 5$, the $Q$-curvature is $Q_{g}=-\frac{1}{2(n-1)}\Delta_{g}R_{g}-\frac{2}{(n-2)^{2}}|\operatorname{Ric}_{g}|^{2}+\frac{n^{3}-4n^{2}+16n-16}{8(n-1)^{2}(n-2)^{2}}R_{g}^{2},$ (1) where $R_{g}$ is the scalar curvature of $g$, $\operatorname{Ric}_{g}$ is the Ricci curvature of $g$, and $\Delta_{g}$ is the Laplace–Beltrami operator of $g$. After a conformal change, the $Q$-curvature transforms as $\widetilde{g}=u^{\frac{4}{n-4}}g\Rightarrow Q_{\widetilde{g}}=\frac{2}{n-4}u^{-\frac{n+4}{n-4}}P_{g}u,$ (2) where $P_{g}$ is the Paneitz operator $\displaystyle P_{g}u=\Delta_{g}^{2}u+\operatorname{div}\left(\frac{4}{n-2}\operatorname{Ric}_{g}(\nabla u,\cdot)-\frac{(n-2)^{2}+4}{2(n-1)(n-2)}R_{g}\langle\nabla u,\cdot\rangle\right)+\frac{n-4}{2}Q_{g}u.$ (3) Paneitz [22] first discovered the operator $P_{g}$ and investigated its conformal invariance. Thereafter Branson [5, 6] began a thorough investigation of $Q_{g}$ and its variants. The reader can find excellent summaries of the fourth order $Q$-curvature in [7, 9, 14]. The $Q$-curvature of the round metric $\overset{\circ}{g}$ is $\frac{n(n^{2}-4)}{8}$, and setting $Q_{g}$ to be this value gives the equation $P_{g}u=\frac{n(n-4)(n^{2}-4)}{16}u^{\frac{n+4}{n-4}}.$ (4) Just as in the scalar curvature setting, one can search for constant $Q$-curvature metrics in a conformal class by minimizing the total $Q$-curvature. However, because of the conformal invariance one encounters the same lack of compactness and presence of singular solutions. Hang and Yang [15] carry out part of this program in the regular case, assuming that the background metric also has positive Yamabe invariant. In any event, a complete understanding of the fourth order analog of the Yamabe problem would require an understanding of the following singular problem: let $(M,g)$ be a compact Riemannian manifold and let $\Lambda\subset M$ be a closed subset. A conformal metric $\widetilde{g}=u^{\frac{4}{n-4}}g$ is a singular constant $Q$-curvature metric if $Q_{\widetilde{g}}$ is constant and $\widetilde{g}$ is complete on $M\backslash\Lambda$. According to (2) we can write this geometric problem as $\displaystyle P_{g}u$ $\displaystyle=$ $\displaystyle\frac{n(n-4)(n^{2}-4)}{16}u^{\frac{n+4}{n-4}}\textrm{ on }M\backslash\Lambda,$ (5) $\displaystyle\liminf_{x\rightarrow x_{0}}u(x)$ $\displaystyle=$ $\displaystyle\infty\textrm{ for each }x_{0}\in\Lambda.$ For the remainder of our work we concentrate on the case that $(M,g)=(\mathbf{S}^{n},\overset{\circ}{g})$ is the round metric on the sphere and $\Lambda=\\{p_{1},\dots,p_{k}\\}$ is a finite set of distinct points. Thus we examine, given a singular set $\Lambda$ with $\\#(\Lambda)=k$, the set of functions $u:\mathbf{S}^{n}\backslash\Lambda=\mathbf{S}^{n}\backslash\\{p_{1},\dots,p_{k}\\}\rightarrow(0,\infty)$ that satisfy $\displaystyle\overset{\circ}{P}u$ $\displaystyle=$ $\displaystyle P_{\overset{\circ}{g}}u=\frac{n(n-4)(n^{2}-4)}{16}u^{\frac{n+4}{n-4}}$ (6) $\displaystyle\liminf_{x\rightarrow p_{j}}u(x)$ $\displaystyle=$ $\displaystyle\infty\textrm{ for each }j=1,2,\dots,k.$ For technical reasons we will also require $R_{g}\geq 0$. Following [21] we define the marked moduli space $\mathcal{M}_{\Lambda}=\left\\{g\in[\overset{\circ}{g}]:Q_{g}=\frac{n(n^{2}-4)}{8},\ R_{g}\geq 0,\ g\textrm{ is complete on }\ \mathbf{S}^{n}\backslash\Lambda\right\\}$ (7) and the unmarked moduli space $\displaystyle\mathcal{M}_{k}=\left\\{g\in[\overset{\circ}{g}]:Q_{g}=\frac{n(n^{2}-4)}{8},\ R_{g}\geq 0,\ g\textrm{ is complete on }\ \mathbf{S}^{n}\backslash\Lambda,\ \\#(\Lambda)=k\right\\}.$ (8) We equip each moduli space with the Gromov–Hausdorff topology. C. S. Lin [20] proved that $\mathcal{M}_{1}$ is the empty set, and recently Frank and König [11] classified all metrics in $\mathcal{M}_{2}$, proving $\mathcal{M}_{p,q}=(0,\epsilon_{n}]\textrm{ for each pair }p\neq q\in\mathbf{S}^{n},$ where $\epsilon_{n}=\left(\frac{n(n-4)}{n^{2}-4}\right)^{\frac{n-4}{8}}\in(0,1).$ (9) It follows that $\mathcal{M}_{2}=(0,\epsilon_{n}]\times((\mathbf{S}^{n}\times\mathbf{S}^{n}\backslash\textrm{diag})/SO(n+1,1)),$ where the group $SO(n+1,1)$ of conformal transformations acts on each $\mathbf{S}^{n}$ factor simultaneously. These metrics corresponding to a doubly punctured sphere are all rotationally invariant, and are called the Delaunay metrics. We describe them in detail in Section 2.1. In the present work we explore some of the structure of $\mathcal{M}_{k}$ when $k\geq 3$. Let $\Lambda=\\{p_{1},\dots,p_{k}\\}$ with $k\geq 3$ and let $g=u^{\frac{4}{n-4}}\overset{\circ}{g}\in\mathcal{M}_{\Lambda}$. As it happens, the metric $g$ is asymptotic to a Delaunay metric near each puncture $p_{j}$, and so one can associate a Delaunay parameter $\epsilon_{j}(g)\in(0,\epsilon_{n}]$ to each $p_{j}$ and $g\in\mathcal{M}_{\Lambda}$. (See Section 2.2.) Our main compactness theorem is the following. ###### Theorem 1. Let $k\geq 3$ and let $\delta_{1}>0,\delta_{2}>0$ be positive numbers. Then the set $\Omega_{\delta_{1},\delta_{2}}=\\{g\in\mathcal{M}_{k}:\operatorname{dist}_{\overset{\circ}{g}}(p_{j},p_{l})\geq\delta_{1}\textrm{ for each }j\neq l,\epsilon_{j}(g)\geq\delta_{2}\\}$ is sequentially compact in the Gromov–Hausdorff topology. We model this result on a compactness theorem of Pollack [23], which states that the similarly defined set in the moduli space of singular Yamabe metrics is sequentially compact. Very recently Wei [27] proved a similar theorem in the context of constant $\sigma_{k}$-curvature. Pollack’s theorem was an important first step in understanding the structure of the moduli space of singular Yamabe metrics on a finitely puctured sphere, a program that is still not complete. We hope our theorem above can play a similar role in advancing the general theory of constant $Q$-curvature metrics with isolated singularities. ## 2\. Preliminaries In this section we present some prerequisite analysis proven elsewhere which we will use below. We first rewrite (6). Pulling back by (the inverse of) stereographic projection, we can write $\overset{\circ}{g}=\left(\frac{2}{1+|x|^{2}}\right)^{2}\delta=U_{\textrm{sph}}^{\frac{4}{n-4}}\delta,\qquad U_{\textrm{sph}}=\left(\frac{1+|x|^{2}}{2}\right)^{\frac{4-n}{2}}.$ (10) In these coordinates (6) takes the form $u:\mathbf{R}^{n}\backslash\\{q_{1},\dots,q_{k}\\}\rightarrow(0,\infty),\qquad\Delta_{0}^{2}(U_{\textrm{sph}}u)=\frac{n(n-4)(n^{2}-4)}{16}(U_{\textrm{sph}}u)^{\frac{n+4}{n-4}},$ (11) where $\Delta_{0}$ is the usual flat Laplacian and $q_{j}$ is the image of $p_{j}$ under the stereographic map. Also, the condition $R_{g}\geq 0$ is equivalent to the differential inequality $-\Delta_{0}(U_{\textrm{sph}}u)^{\frac{n-2}{n-4}}\geq 0\Leftrightarrow-\Delta_{0}(U_{\textrm{sph}}u)\geq\frac{2}{n-4}\frac{|\nabla(U_{\textrm{sph}}u)|^{2}}{U_{\textrm{sph}}u}.$ (12) In this Euclidean setting the transformation rule (2) reads $\Delta_{0}^{2}u=Au^{\frac{n+4}{n-4}}\Rightarrow u_{\lambda}(x)=\lambda^{\frac{n-4}{2}}u(\lambda x)\textrm{ satsfies }\Delta_{0}^{2}u_{\lambda}=Au_{\lambda}^{\frac{n+4}{n-4}}\textrm{ for each }\lambda>0$ (13) for any constant $A$. ### 2.1. Delaunay metrics Let $p\neq q\in\mathbf{S}^{n}$ and let $g\in\mathcal{M}_{\\{p,q\\}}$. We may precompose by an appropriate dilation and assume $p=-q$, and then rotate $\mathbf{S}^{n}$ so that $p$ is the north pole and $q$ is the south pole. After reframing as in the previous paragraph we obtain a function $u:\mathbf{R}^{n}\backslash\\{0\\}\rightarrow(0,\infty)$ satisfying the PDE (11). Lin [20] proved that this solution must be rotationally invariant about $0$, and later Frank and König [11] classified all the ODE solutions. Their classification is easiest to see after changing to cylindrical coordinates. We let $\displaystyle t=-\log|x|,$ $\displaystyle\quad\theta=\frac{x}{|x|},$ (14) $\displaystyle v(t,\theta)$ $\displaystyle=$ $\displaystyle e^{\left(\frac{4-n}{2}\right)t}U_{\textrm{sph}}(e^{-t}\theta)u(e^{-t}\theta)=(e^{t}\cosh t)^{\frac{4-n}{2}}u(e^{-t}\theta)$ This transforms the Paneitz operator into $\displaystyle P_{\textrm{cyl}}$ $\displaystyle=$ $\displaystyle\frac{\partial^{4}}{\partial t^{4}}+\Delta_{\theta}^{2}+2\Delta_{\theta}\frac{\partial^{2}}{\partial t^{2}}-\left(\frac{n(n-4)+8}{2}\right)\frac{\partial^{2}}{\partial t^{2}}$ $\displaystyle-\frac{n(n-4)}{2}\Delta_{\theta}+\frac{n^{2}(n-4)^{2}}{16},$ so that (11) becomes $v:\mathbf{R}\times\mathbf{S}^{n-1}\rightarrow(0,\infty),\qquad P_{\textrm{cyl}}v=\frac{n(n-4)(n^{2}-4)}{16}v^{\frac{n+4}{n-4}}.$ (16) The fact that the orginal function $u$ is radial implies $v$ is a function of $t$ alone, and so (16) reduces to the ODE $\ddddot{v}-\left(\frac{n(n-4)+8}{2}\right)\ddot{v}+\frac{n^{2}(n-4)^{2}}{16}v=\frac{n(n-4)(n^{2}-4)}{16}v^{\frac{n+4}{n-4}}.$ (17) We find two solutions explicitly. The cylindrical solution is the only constant solution, namely $v_{\textrm{cyl}}=\epsilon_{n}$, given in (9). Also, the spherical solution $U_{\textrm{sph}}$ given in (10) transforms under the change of variables (14) into $v_{\textrm{sph}}=(\cosh t)^{\frac{4-n}{2}}$. The Delaunay solutions found by Frank and König in [11] interpolate between the cylindrical and spherical solutions. Indeed, for each $\epsilon\in(0,\epsilon_{n})$ there exists a unique solution $v_{\epsilon}$ of the ODE (17) realizing its minimal value of $\epsilon$ at $t=0$. Each $v_{\epsilon}$ is periodic with minimal period $T_{\epsilon}$, and these Delaunay solutions account for all global solutions of the ODE (17). Transforming back to Euclidean coordinates, we of course obtain the solutions $u_{\epsilon}:\mathbf{R}^{n}\backslash\\{0\\}\rightarrow(0,\infty),\qquad u_{\epsilon}(x)=|x|^{\frac{4-n}{2}}v_{\epsilon}(-\log|x|).$ (18) We may then apply global conformal transformations to construct the translated Delaunay solutions. The first such family is $\widetilde{u}_{\epsilon,a}(x)=u_{\epsilon}(x-a)$ for some fixed vector $a\in\mathbf{R}^{n}$. The second family is more important to our later analysis, and is given by translating the point at infinity. More precisely, we define $\displaystyle u_{\epsilon,a}(x)$ $\displaystyle=$ $\displaystyle\widehat{\mathbb{K}}_{0}(\widehat{\mathbb{K}}_{0}(u_{\epsilon}(\cdot-a))(x)$ $\displaystyle=$ $\displaystyle|x|^{\frac{n-4}{2}}\left|\frac{x}{|x|}-|x|a\right|^{\frac{4-n}{2}}v_{\epsilon}\left(-\log|x|-\log\left|\frac{x}{|x|}-|x|a\right|\right),$ where $\widehat{\mathbb{K}}_{0}(u)(x)=|x|^{4-n}u\left(\frac{x}{|x|^{2}}\right)$ is the Kelvin transform of $u$. In cylindrical coordinates we can write this expression for $u_{\epsilon,a}$ as $\displaystyle v_{\epsilon,a}(t,\theta)$ $\displaystyle=$ $\displaystyle|\theta-e^{-t}a|^{\frac{4-n}{2}}v_{\epsilon}(t+\log|\theta-e^{-t}a|)$ $\displaystyle=$ $\displaystyle v_{\epsilon}(t)+e^{-t}\langle\theta,a\rangle\left(-\dot{v}_{\epsilon}(t)+\frac{n-4}{2}v_{\epsilon}(t)\right)+\mathcal{O}(e^{-2t})$ ### 2.2. Asymptotics In [16] Jin and Xiong proved that any positive, superharmonic solution of (11) in a punctured ball is asymptotically symmetric. In other words, they show there exists $\alpha>0$ such that $u(x)=\overline{u}(|x|)(1+\mathcal{O}(|x|^{\alpha})),\qquad\overline{u}(r)=\frac{1}{r^{n-1}|\mathbf{S}^{n-1}|}\int_{|x|=r}u(\theta)d\theta.$ (20) Later the first two authors [2] and the third author [24] independently derived refined asymptotics for positive, singular solutions of (11). Roughly speaking, the translated Delaunay solutions of (2.1) give the next order term in the expansion of $u$. They show there exists $\beta>1$, $\epsilon\in(0,\epsilon_{n}]$, $T\in[0,T_{\epsilon})$, and $a\in\mathbf{R}^{n}$ such that $\displaystyle u(x)$ $\displaystyle=$ $\displaystyle|x|^{\frac{4-n}{2}}\left(v_{\epsilon}(-\log|x|+T)\right.$ $\displaystyle+$ $\displaystyle\left.|x|\left\langle\frac{x}{|x|},a\right\rangle\left(-\dot{v}_{\epsilon}(-\log|x|+T)+\frac{n-4}{2}v_{\epsilon}(-\log|x|+T)\right)+\mathcal{O}(|x|^{\beta})\right).$ In cylindrical coordinates this estimate has the form $v(t,\theta)=v_{\epsilon}(t+T)+e^{-t}\langle\theta,a\rangle\left(-\dot{v}_{\epsilon}(t+T)+\frac{n-4}{2}v_{\epsilon}(t+T)\right)+\mathcal{O}(e^{-\beta t}).$ (22) ### 2.3. Some other useful theorems For the sake of completeness, we state some background results which will be required later in the proof of our main result. We first quote the following theorem of Chang, Han and Yang [10, Theorem 1.1]. ###### Theorem 2 (Chang, Han and Yang). Let $n\geq 5$, let $\Lambda\subset\mathbf{S}^{n}$ be a proper closed set, and let $g=u^{\frac{4}{n-4}}\overset{\circ}{g}$ be a complete metric on $\mathbf{S}^{n}\backslash\Lambda$ such that $Q_{g}=\frac{n(n^{2}-4)}{8},\qquad R_{g}\geq 0.$ Then $\partial\overset{\circ}{\mathbf{B}}_{\rho}(x_{0})$ is has positive mean curvature with respect to $g$, computed with the inward pointing normal, where $\overset{\circ}{\mathbf{B}}_{\rho}(x_{0})$ is any ball with respect to the round metric contained in $\mathbf{S}^{n}\backslash\Lambda$. We will also need a version of Harnack’s inequality, which was proven by Caristi and Mitidieri [8, Theorem 3.6]. ###### Theorem 3 (Caristi and Mitidieri). Let $u$ be a superharmonic function defined in a domain $\Omega\subset\mathbf{R}^{n}$ such that $\Delta_{0}^{2}u=f(u)$, where $f$ is either linear or superlinear and $f(0)=0$. Then there exists $\rho_{0}>0$ such that for $\rho\leq\rho_{0}$ we have $\sup_{\overset{\circ}{\mathbf{B}}_{\rho}(p)}u\leq C\inf_{\overset{\circ}{\mathbf{B}}_{\rho}(p)}u,$ (23) where the constant $c$ depends only on the domain $\Omega$, the function $f$, and $\rho$. Gursky and Malchiodi [12, Proposition 2.5] prove the existence of a positive Greens function for the Paneitz operator of the round sphere. ###### Theorem 4 (Gursky and Malchiodi). Let $(M,g)$ be a compact Riemannian manifold such that $R_{g}\geq 0$ and $Q_{g}>0$. Then for each $p\in M$ there exists a Greens function $G_{p}$ satisfying $G_{p}:M\backslash\\{p\\}\rightarrow(0,\infty),\qquad P_{g}G_{p}=\delta_{p},$ where $\delta_{p}$ is the Dirac $\delta$-function with a singularity at $p$. Furthermore, if either $n=5,6,7$ or $g$ is conformally flat then there exists $c>0$ depending only on $n$ and $\alpha$ such that $G_{p}(x)=\frac{1}{2n(n-2)(n-4)\omega_{n}}\left(\operatorname{dist}_{g}(x,p)\right)^{4-n}+\mathcal{O}(1)$ (24) in conformal normal coordinates, where $\omega_{n}$ is the volume of a unit ball in $\mathbf{R}^{n}$. ## 3\. Pohozaev invariants One often finds integral invariants in geometric variational problems. The reader can find a general abstract framework for constructing these invariants in the paper by Gover and Ørsted [13]. In future work we will explicitly write out the full Pohozaev invariant using the first variation tensor defined in [18]. We consider a function $v:(a,b)\times\mathbf{S}^{n-1}\rightarrow\mathbf{R}$ satisfying (16). Given such a function $v$ we define the Hamiltonian functional $\displaystyle\mathcal{H}(v)$ $\displaystyle=$ $\displaystyle-\frac{\partial v}{\partial t}\frac{\partial^{3}v}{\partial t^{3}}+\frac{1}{2}\left(\frac{\partial^{2}v}{\partial t^{2}}\right)^{2}-\frac{1}{2}(\Delta_{\theta}v)^{2}-\left|\nabla_{\theta}\frac{\partial v}{\partial t}\right|^{2}+\frac{n(n-4)}{4}|\nabla_{\theta}v|^{2}$ $\displaystyle+\left(\frac{n(n-4)+8}{4}\right)\left(\frac{\partial v}{\partial t}\right)^{2}-\frac{n^{2}(n-4)^{2}}{32}v^{2}+\frac{(n-4)^{2}(n^{2}-4)}{32}v^{\frac{2n}{n-4}}.$ Integrating by parts we find $\frac{d}{dt}\int_{\\{t\\}\times\mathbf{S}^{n-1}}\mathcal{H}(v)d\theta=0,$ (26) which allows us to define our first integral invariant as $\widetilde{\mathcal{P}}_{\textrm{rad}}(v)=\int_{\\{t\\}\times\mathbf{S}^{n-1}}\mathcal{H}(v)d\theta.$ (27) Now we can define the radial (or dilational) Pohozaev invariants associated to a metric $g\in\mathcal{M}_{k}$ at each puncture point $p_{j}$. Recall that $g=u^{\frac{4}{n-4}}\overset{\circ}{g}$ is a complete, conformally flat metric on $\mathbf{S}^{n}\backslash\\{p_{1},\dots,p_{k}\\}$ with $Q_{g}=\frac{n(n^{2}-4)}{8}$ and $R_{g}\geq 0$. Completeness forces $\liminf_{\operatorname{dist}_{\overset{\circ}{g}}(x,p_{j})\rightarrow 0}u(x)=\infty$ for each $j$, while $Q_{g}=\frac{n(n^{2}-4)}{8}$ is equivalent to the PDE (11), after stereographic projection down to $\mathbf{R}^{n}\backslash\\{p_{1},\dots,p_{k}\\}$. Choose coordinates centered at one of the punctures $p_{j}$, and the perform the cylindrical change of variables (14), which gives us a function $v:(A,\infty)\times\mathbf{S}^{n-1}\rightarrow(0,\infty)$ satisfying (16). We define $\mathcal{P}_{\textrm{rad}}(g,p_{j})=\widetilde{\mathcal{P}}_{\textrm{rad}}(v)=\int_{\\{t\\}\times\mathbf{S}^{n-1}}\mathcal{H}(v)d\theta,$ which is well-defined by (26). In the special case that $k=2$ we can evaluate the dilational Pohozaev invariant more explicitly. In this situation we may as well let the puncture points be the north and south poles, and thus we obtain a function $v=v_{\epsilon}:\mathbf{R}\times\mathbf{S}^{n-1}\rightarrow(0,\infty)$ satisfying (17). Thus the Hamiltonian (3) reduces to $\overline{\mathcal{H}}(v)=-\dot{v}\dddot{v}+\frac{1}{2}\ddot{v}^{2}+\left(\frac{n(n-4)+8}{4}\right)\dot{v}^{2}-\frac{n^{2}(n-4)^{2}}{32}v^{2}-\frac{(n-4)^{2}(n^{2}-4)}{32}v^{\frac{2n}{n-4}}.$ (28) Moreover, because $\overline{\mathcal{H}}$ does not depend on $\theta$ and its integral over a sphere does not depend on $t$, this reduced Hamiltonian must be constant on solutions of (17). (One can, of course, explicitly verify this constancy by taking a derivative.) A computation reveals $\overline{\mathcal{H}}(v_{\textrm{sph}})=0,\qquad\overline{\mathcal{H}}(v_{\epsilon_{n}})=-\frac{(n-4)(n^{2}-4)}{8}\left(\frac{n(n-4)}{n^{2}-4}\right)^{n/4}<0.$ Furthermore, Proposition 6 of [4] implies the Delaunay solutions are ordered (in fact, uniquely determined!) by their energy (28). Combining our analysis above with (2.2) and (22) we immediately see the following ###### Lemma 5. Let $g=u^{\frac{4}{n-4}}\overset{\circ}{g}$ be a complete, conformally flat metric on $\mathbf{S}^{n}\backslash\\{p_{1},\dots,p_{k}\\}$ with $Q_{g}=\frac{n(n^{2}-4)}{8}$ and $R_{g}\geq 0$. For each puncture $p_{j}$ define $\mathcal{P}_{\textrm{rad}}(g,p_{j})$ as above. Then $\mathcal{P}_{\textrm{rad}}(g,p_{j})<0$ and depends only on the necksize $\epsilon_{j}$ of the Delaunay asymptote at $p_{j}$. Moreover, decreasing $\epsilon_{j}$ will increase $\mathcal{P}_{\textrm{rad}}(g,p_{j})$ towards $0$. In particular, bounding the radial Pohozaev invariants $\mathcal{P}_{\textrm{rad}}(g,p_{j})$ away from zero is equivalent to bounding the necksizes $\epsilon_{j}$ away from zero. ###### Proof. We have shown that $\mathcal{P}_{\textrm{rad}}(g,p_{j})=\int_{\\{t\\}\times\mathbf{S}^{n-1}}\mathcal{H}(v)d\theta$ is well-defined, because the integral does not depend on our choice of $t$. Now let $t\rightarrow\infty$, and observe that $v\rightarrow v_{\epsilon_{j}}$, the Delaunay asymptote of $g$ at the puncture $p_{j}$. In particular, $\mathcal{H}(v)\rightarrow\overline{\mathcal{H}}(v_{\epsilon_{j}})$, We conclude that $\mathcal{P}_{\textrm{rad}}(v)=\mathcal{P}_{\textrm{rad}}(v_{\epsilon_{j}})=|\mathbf{S}^{n-1}|\overline{\mathcal{H}}(v_{\epsilon_{j}})<0.$ The remainder of the lemma follows from the energy ordering theorem of van den Berg [4] applied to the Delaunay solutions, as described in the paragraph above. ∎ ###### Remark 1. Our radial Pohozaev invariant is basically the same as the one defined in Proposition 4.1 of [1]. Jin and Xiong [16] write out the same invariant for higher order equations. It will actually be useful for later computations to decompose the Hamiltonian energy $\mathcal{H}$ given in (3) as $\mathcal{H}(v)=\mathcal{H}_{\textrm{cyl}}(v)+\frac{(n-4)^{2}(n^{2}-4)}{32}v^{\frac{2n}{n-4}}.$ (29) The same computation as in (27) shows the following lemma. ###### Lemma 6. Let $v$ satisfy $v:(a,b)\times\mathbf{S}^{n-1}\rightarrow\mathbf{R},\qquad P_{\textrm{cyl}}v=Av^{\frac{n+4}{n-4}}$ for some constant $A$. Then the integral $\int_{\\{t\\}\times\mathbf{S}^{n-1}}\mathcal{H}_{\textrm{cyl}}(v)+\frac{(n-4)}{2n}Av^{\frac{2n}{n-4}}d\theta$ does not depend on $t$. ## 4\. Proof of the compactness theorem In this section we prove Theorem 1. We first use standard blow-up techniques to prove a priori bounds on the $\mathcal{C}^{4}$-norm of solutions of (5). Once we obtain these bounds, we use them to extract a convergent subsequence. Finally we prove that our limit is non-trivial, using the fact that the radial Pohozaev invariants of our original sequence of metrics remain bounded away from zero. ### 4.1. A priori bounds We prove some a priori bounds for solutions of $\overset{\circ}{P}u=\frac{n(n-4)(n^{2}-4)}{16}u^{\frac{n+4}{n-4}}.$ ###### Theorem 7. Let $n\geq 5$, let $\Lambda\subset\mathbf{S}^{n}$ be a proper closed set, and let $g=u^{\frac{4}{n-4}}\overset{\circ}{g}$ be a complete metric on $\mathbf{S}^{n}\backslash\Lambda$ such that $Q_{g}=\frac{n(n^{2}-4)}{8},\qquad R_{g}\geq 0.$ Then there exists $C>0$ depending only on the dimension $n$ such that $u(x)\leq C\left(\operatorname{dist}_{\overset{\circ}{g}}(x,\Lambda)\right)^{\frac{4-n}{2}}.$ (30) ###### Remark 2. In the context of $g\in\mathcal{M}_{k}$ with $k\geq 2$, our upper bound (30) is very similar to, but slightly stronger than, Proposition 3.2 of [16], because our constant $C$ depends only on the dimension $n$. ###### Proof. Our proof borrows from Pollack’s proof of the corresponding upper bound in the scalar curvature case. Given any $g$ satisfying the hypotheses of Theorem 7, $x_{0}\not\in\Lambda$, and $\rho>0$ such that $\overset{\circ}{\mathbf{B}}_{\rho}\subset\mathbf{S}^{n}\backslash\Lambda$ we define the auxiliary function $f:\overset{\circ}{\mathbf{B}}_{\rho}\rightarrow\mathbf{R},\qquad f(x)=(\rho-\operatorname{dist}_{\overset{\circ}{g}}(x,x_{0}))^{\frac{n-4}{2}}u(x).$ (31) Observe that choosing $\rho=\frac{1}{2}\operatorname{dist}_{\overset{\circ}{g}}(x_{0},\Lambda)$ yields $f(x_{0})=\rho^{\frac{n-4}{2}}u(x_{0})=\left(\frac{1}{2}\operatorname{dist}_{\overset{\circ}{g}}(x_{0},\Lambda)\right)^{\frac{n-4}{2}}u(x_{0}),$ (32) so it will suffice to find $C$ depending only on $n$ such that $f(x)\leq C$ for all admissible choices of $\Lambda$, $u$, $x_{0}$, and $\rho$. We suppose the contrary and derive a contradiction. To this end, let $\Lambda_{i}$, $u_{i}$, $x_{0,i}$ and $\rho_{i}$ be admissible as described above and suppose $M_{i}=f(x_{1,i})=\sup_{x\in\overset{\circ}{\mathbf{B}}_{\rho_{i}}(x_{0,i})}f(x)\rightarrow\infty.$ (33) Observe that $\left.f\right|_{\partial\overset{\circ}{\mathbf{B}}_{\rho_{i}}(x_{0,i})}=0$, so $x_{1,i}$ must lie in the interior of the ball $\overset{\circ}{\mathbf{B}}_{\rho_{i}}(x_{0,i})$. Next let $r_{i}=\rho_{i}-\operatorname{dist}_{\overset{\circ}{g}}(x_{1,i},x_{0,i}),$ let $y$ be geodesic normal coordinates centered at $x_{1,i}$, and define $\lambda_{i}=2(u_{i}(x_{1,i}))^{\frac{4-n}{2}},\qquad R_{i}=\frac{r_{i}}{\lambda_{i}}=\frac{r_{i}}{2}(u_{i}(x_{1,i}))^{\frac{n-4}{2}}=\frac{1}{2}M_{i}^{\frac{2}{n-4}}$ (34) and $w_{i}:\mathbf{B}_{R_{i}}(0)\rightarrow(0,\infty),\qquad w_{i}(y)=\lambda_{i}^{\frac{n-4}{2}}u_{i}(\lambda y).$ (35) By (2) (or, equivalently, (13)) the function $w_{i}$ solves $P_{\lambda\overset{\circ}{g}}w_{i}=\frac{n(n-4)(n^{2}-4)}{16}w_{i}^{\frac{n+4}{n-4}}.$ Moreover, by construction $2^{\frac{n-4}{2}}=w_{i}(0)=\sup_{\mathbf{B}_{R_{i}}(0)}w_{i}(x).$ Using the Arzela-Ascoli theorem we extract a subsequence, which we still denote by $w_{i}$, that converges uniformly on compact subsets of $\mathbf{R}^{n}$. Furthermore, as $i\rightarrow\infty$ the rescaled metrics $\lambda\overset{\circ}{g}$ converge to the Euclidean metric. Therefore, in the limit we obtain a function $\overline{w}:\mathbf{R}^{n}\rightarrow[0,\infty),\quad\Delta_{0}^{2}\overline{w}=\frac{n(n-4)(n^{2}-4)}{16}\overline{w}^{\frac{n+4}{n-4}},\quad\overline{w}(0)=\sup\overline{w}=2^{\frac{n-4}{2}}.$ (36) By the classification theorem in [20, Theorem 1.3] we must have $\overline{w}(x)=\left(\frac{1+|x|^{2}}{2}\right)^{\frac{4-n}{2}}.$ Thus each solution $u_{i}$ has a “bubble” when $i$ is sufficiently large, that is for $i$ sufficiently large a small neighborhood of $x_{1,i}$ is close (in $\mathcal{C}^{4}$-norm) to the round metric, and hence has a concave boundary. We verify this by computing the mean curvature of a geodesic sphere explicitly. The round metric has the form $g_{lm}=\frac{4}{(1+|x|^{2})^{2}}\delta_{lm}$ in stereographic coordinates, and in general the mean curvature of a hypersurface $\Sigma$ in a Riemannian manifold with unit normal $\eta$ is given by $H_{\Sigma}=-\operatorname{tr}_{g}\langle\nabla_{\partial l}\eta,\partial_{m}\rangle=-\partial_{l}\eta^{l}-\eta^{p}\Gamma_{lp}^{l}.$ (37) A geodesic sphere centered at $0$ coincides with a Euclidean round sphere centered at the origin (with a different radius, of course), and so the inward normal vector is $\eta=-\left(\frac{1+|x|^{2}}{2|x|}\right)x^{l}\partial_{l}.$ A computation shows $H=-2n|x|(1+|x|^{2})+\frac{n-1+n|x|^{2}}{|x|},$ which is negative, in particular, when $|x|>3$. Additionally, since $\|w_{i}-\overline{w}\|_{\mathcal{C}^{4}(\mathbf{B}_{\frac{3R_{i}}{4}}(0))}$ is arbitrarily small when $i$ is sufficiently large, we see that $\partial\mathbf{B}_{\frac{3R_{i}}{4}}(0)$ is also mean concave with respect to the metric $w_{i}^{\frac{4}{n-4}}\delta_{lm}$, which in turn implies $\partial\mathbf{B}_{\frac{3|x_{1,i}|}{8}}(x_{1,i})$ is mean concave with respect to the metric induced by $u_{i}^{\frac{4}{n-4}}\delta_{lm}$. This contradicts Theorem 2. ∎ We immediately obtain the following Corollary. ###### Corollary 8. For each compact subset $\Omega\subset\mathbf{S}^{n}\backslash\Lambda$, $l\in\mathbb{N}$ and $\alpha\in(0,1)$ there exists $C_{1}$ depending only on $n$, $l$, $\Omega$, and $\alpha$ such that $\|u_{i}\|_{\mathcal{C}^{l,\alpha}(\Omega)}\leq C_{1}.$ (38) We also record here a lower bound due to Jin and Xiong [16, Theorem 1.3]. ###### Theorem 9 (Jin and Xiong). Let $v:[A,\infty)\times\mathbf{S}^{n-1}\rightarrow(0,\infty)$ solve (16). Then $\mathcal{P}_{\textrm{rad}}(v)\leq 0$ with equality if and only if $\liminf_{t\rightarrow\infty}v(t,\theta)=\limsup_{t\rightarrow\infty}v(t,\theta)=\lim_{t\rightarrow\infty}v(t,\theta)=0.$ Otherwise, if $\mathcal{P}_{\textrm{rad}}(v)<0$, there exists $C_{2}>0$ (which depends on the solution $v$!) such that $v(t,\theta)\geq C_{2}$. ###### Corollary 10. Let $g=u^{\frac{4}{n-4}}\overset{\circ}{g}\in\mathcal{M}_{k}$ have the singular set $\Lambda=\\{p_{1},\dots,p_{k}\\}$. Then there exists $C_{2}>0$ (depending on the solution $u$!) such that $u(x)\geq C_{2}\left(\min_{1\leq j\leq k}\operatorname{dist}_{\overset{\circ}{g}}(x,p_{j})\right)^{\frac{4-n}{2}}.$ ### 4.2. Sequential compactness In this section we complete our proof of sequential compactness. To this end, let $\\{g_{i}=u_{i}^{\frac{4}{n-4}}\overset{\circ}{g}\\}\subset\Omega_{\delta_{1},\delta_{2}}\subset\mathcal{M}_{k}$ and denote the singular set of the conformal factor $u_{i}$ by $\Lambda_{i}=\\{p_{1}^{i},\dots,p_{k}^{i}\\}$. The following lemma will simplify our later analysis since it allows us to assume the singular points are fixed. ###### Lemma 11. Let $g_{i}=u_{i}^{\frac{4}{n-4}}\overset{\circ}{g}$ be a sequence in $\mathcal{M}_{k}$ as described above. After passing to a subsequence, we may assume that when $i$ is sufficiently large both $g_{i}$ and $u_{i}$ are regular on the $\mathbf{S}^{n}\backslash\left(\cup_{j=1}^{k}\overset{\circ}{\mathbf{B}}_{\delta_{1}/2}(p_{j}^{i})\right),$ where $\overset{\circ}{\mathbf{B}}_{r}(p)$ is the geodesic ball centered at $p$ with radius $r$, with respect to the round metric $\overset{\circ}{g}$. ###### Proof. The set $(\mathbf{S}^{n})^{k}\backslash\left\\{(q_{1},\dots,q_{k})\in(\mathbf{S}^{n})^{k}:\operatorname{dist}_{\overset{\circ}{g}}(q_{j},q_{l})\geq\delta_{1}\textrm{ for each }j\neq l\right\\}$ is compact and contains each singular set $\Lambda_{i}=\\{p_{1}^{i},\dots,p_{k}^{i}\\}$. Thus we may extract a convergent subsequence, which we still denote as $\Lambda_{i}=\\{p_{1}^{i},\dots,p_{k}^{i}\\}$, with $p_{j}^{i}\rightarrow\bar{p}_{j}$. The lemma now follows from $p_{j}^{i}\rightarrow\bar{p}_{j}$ for each $j$. ∎ To set notation, we define the compact sets $K_{m}=\mathbf{S}^{n}\backslash\left(\cup_{j=1}^{k}\overset{\circ}{\mathbf{B}}_{2^{-m}\delta_{1}}(\bar{p}_{j})\right)$ (39) for each natural number $m\in\mathbf{N}$. By construction the family $\\{K_{m}\\}$ is a compact exhaustion of $\mathbf{S}^{n}\backslash\\{\bar{p}_{1},\dots,\bar{p}_{k}\\}$. Furthermore, by the convergence $p_{j}^{i}\rightarrow\bar{p}_{j}$, for each fixed $m$ there exists $i_{0}$ such that $i\geq i_{0}$ implies $u_{i}$ is smooth in $K_{m}$. Therefore, combining Corollary 8 and the Arzela-Ascoli theorem we obtain a convergent subsequence, which we again denote by $u_{i}$, that converges uniformly on compact subsets of $\mathbf{S}^{n}\backslash\overline{\Lambda}$ to a limit $\overline{u}$, where $\overline{\Lambda}=\\{\bar{p}_{1},\dots,\bar{p}_{k}\\}$. Furthermore, combining our a priori bounds and elliptic regularity, we see that the limit function satisfies $\overline{u}:\mathbf{S}^{n}\backslash\overline{\Lambda}\rightarrow[0,\infty),\quad\overset{\circ}{P}\overline{u}=\frac{n(n-4)(n^{2}-4)}{16}\overline{u}^{\frac{n+4}{n-4}}.$ (40) ###### Proposition 12. The limit function $\overline{u}$ constructed in the paragraph above is positive on $\mathbf{S}^{n}\backslash\\{\bar{p}_{1},\dots,\bar{p}_{k}\\}$. ###### Proof. If the proposition does not hold then there exists $q\in\mathbf{S}^{n}\backslash\\{\bar{p}_{1},\dots,\bar{p}_{k}\\}$ such that $0=\overline{u}(q)=\lim_{i\rightarrow\infty}u_{i}(q).$ Let $\epsilon_{i}=u_{i}(q)$ and $w_{i}:\mathbf{S}^{n}\backslash\\{p_{1}^{i},\dots,p_{k}^{i}\\}\rightarrow(0,\infty),\qquad w_{i}(x)=\frac{1}{\epsilon_{i}}u_{i}(x).$ As a consequence of (13), we have $\overset{\circ}{P}w_{i}=\epsilon_{i}^{\frac{8}{n-4}}\frac{n(n-4)(n^{2}-4)}{16}w_{i}^{\frac{n+4}{n-4}}.$ (41) In addition, $w_{i}$ satisfies the normalization $w_{i}(q)=1$ (42) for each $i$ by construction. By (38), for each $m\in\mathbf{N}$ there exists $C_{1}$ depending on $m$ and the dimesnion $n$ such that $\sup_{K_{m}}u_{i}\leq C_{1}.$ (43) Next we find an upper bound for $1/\epsilon_{i}$. Using the fact that $0<U_{\textrm{sph}}\leq 2^{\frac{n-4}{2}}$ and the Harnack inequality in Theorem 3 we get $2^{\frac{n-4}{2}}\epsilon_{i}\geq\inf_{K_{m}}(U_{\textrm{sph}}u_{i})\geq\frac{1}{\widetilde{C}(m,n)}\sup_{K_{m}}(U_{\textrm{sph}}u_{i})=C_{2},$ and so $\frac{1}{\epsilon_{i}}\leq C_{3}.$ (44) Combining (43) and (44) we obtain a uniform upper bound $\sup_{K_{m}}w_{i}\leq C_{3},$ where $C_{3}$ depends only on $n$ and $m$. We conclude that $w_{i}$ converges uniformly on compact subsets of $\mathbf{S}^{n}\backslash\overline{\Lambda}$ to a function $\overline{w}:\mathbf{S}^{n}\backslash\overline{\Lambda}\rightarrow[0,\infty),\qquad\overset{\circ}{P}\overline{w}=0.$ By Theorem 4 we have $\overline{w}=\sum_{j=1}^{k}\alpha_{j}G_{\bar{p}_{j}},$ (45) for some coefficients $\alpha_{j}\geq 0$. By the normalization (42) at least one of the $\alpha_{j}$’s is positive, so (in particular) $\overline{w}$ is a smooth, positive function. Without loss of generality, we may assume $\alpha_{1}\neq 0$ and center our coordinate system at $\bar{p}_{1}$. We now use the cylindrical coordinates $t=-\log|x|$ and $\theta=\frac{x}{|x|}$ in a punctured ball centered on $\bar{p}_{1}=0$, and define $v_{i}(t,\theta)=e^{\left(\frac{4-n}{2}\right)t}u_{i}(e^{-t}\theta)U_{\textrm{sph}}(e^{-t}\theta),\quad z_{i}(t,\theta)=\frac{1}{\epsilon_{i}}v_{i}(t,\theta)=e^{\left(\frac{4-n}{2}\right)t}w_{i}(e^{-t}\theta)U_{\textrm{sph}}(e^{-t}\theta)$ and $\overline{v}(t,\theta)=e^{\left(\frac{4-n}{2}\right)t}\overline{u}(e^{-t}\theta)(\cosh t)^{\frac{4-n}{2}},\quad\overline{z}(t,\theta)=e^{\left(\frac{4-n}{2}\right)t}\overline{w}(e^{-t}\theta)(\cosh t)^{\frac{4-n}{2}}.$ By the expansion (24) we have $\displaystyle\overline{z}(t,\theta)$ $\displaystyle=$ $\displaystyle e^{\left(\frac{4-n}{2}\right)t}(\cosh t)^{\frac{4-n}{2}}\left(\frac{\alpha_{1}}{2n(n-2)(n-4)\omega_{n}}e^{\left(\frac{n-4}{2}\right)t}+\mathcal{O}(1)\right)$ $\displaystyle=$ $\displaystyle\frac{\alpha_{1}}{2n(n-2)(n-4)\omega_{n}}+\mathcal{O}(e^{(4-n)t}).$ Observe that $z_{i}$ satisfies the PDE $P_{\textrm{cyl}}z_{i}=\epsilon_{i}^{\frac{8}{n-4}}\frac{n(n-4)(n^{2}-4)}{16}z_{i}^{\frac{n+4}{n-4}},$ so, following Lemma 6, the integral $\int_{\\{t\\}\times\mathbf{S}^{n-1}}\mathcal{H}_{\textrm{cyl}}(z_{i})+\epsilon_{i}^{\frac{8}{n-4}}\frac{(n-4)^{2}(n^{2}-4)}{32}z_{i}^{\frac{2n}{n-4}}d\theta$ does not depend on $t$. Moreover, taking a limit as $i\rightarrow\infty$, we obtain $\displaystyle\lim_{i\rightarrow\infty}\int_{\\{t\\}\times\mathbf{S}^{n-1}}\mathcal{H}_{\textrm{cyl}}(z_{i})+\epsilon_{i}^{\frac{8}{n-4}}\frac{(n-4)^{2}(n^{2}-4)}{32}z_{i}^{\frac{2n}{n-4}}d\theta$ (47) $\displaystyle=$ $\displaystyle\int_{\\{t\\}\times\mathbf{S}^{n-1}}\mathcal{H}_{\textrm{cyl}}(\overline{z})d\theta$ $\displaystyle=$ $\displaystyle-\int_{\\{t\\}\times\mathbf{S}^{n-1}}\frac{n^{2}(n-4)^{2}}{32}\cdot\frac{\alpha_{1}^{2}}{4n^{2}(n-2)^{2}(n-4)^{2}\omega_{n}^{2}}+\mathcal{O}(e^{(4-n)t})$ $\displaystyle=$ $\displaystyle-\frac{n\alpha_{1}^{2}}{128(n-2)^{2}\omega_{n}}+\mathcal{O}(e^{(4-n)t}).$ On the other hand, by construction $\displaystyle\mathcal{P}_{\textrm{rad}}(v_{i})$ $\displaystyle=$ $\displaystyle\int_{\\{t\\}\times\mathbf{S}^{n-1}}\mathcal{H}_{\textrm{cyl}}(v_{i})+\frac{(n-4)^{2}(n^{2}-4)}{32}v_{i}^{\frac{2n}{n-4}}d\theta$ $\displaystyle=$ $\displaystyle\int_{\\{t\\}\times\mathbf{S}^{n-1}}\epsilon_{i}^{2}\mathcal{H}_{\textrm{cyl}}(z_{i})+\frac{(n-4)^{2}(n^{2}-4)}{32}\epsilon_{i}^{\frac{2n}{n-4}}z_{i}^{\frac{2n}{n-4}}d\theta\rightarrow 0,$ and so $\lim_{i\rightarrow\infty}\mathcal{P}_{\textrm{rad}}(g_{i},p_{1}^{i})=0.$ This contradicts the hypothesis that the asymptotic necksizes of $g_{i}=u_{i}^{\frac{4}{n-4}}\overset{\circ}{g}$ at the puncture points $p_{1}^{i}$ are all bounded away from $0$. ∎ We finally complete the proof of Theorem 1. ###### Proof. Given a sequence $\\{g_{i}\\}\in\Omega_{\delta_{1},\delta_{2}}$, we have already obtained a limit $\overline{g}=\overline{u}^{\frac{4}{n-4}}\overset{\circ}{g}$ as a limit of a subsequence. We know that $\overline{u}:\mathbf{S}^{n}\backslash\\{\bar{p}_{1},\dots,\bar{p}_{k}\\}\rightarrow(0,\infty),\quad\overset{\circ}{P}\overline{u}=\frac{n(n-4)(n^{2}-4)}{16}\overline{u}^{\frac{n+4}{n-4}},$ where $\bar{p}_{j}=\lim_{i\rightarrow\infty}p_{j}^{i}$. We have also shown that $\overline{u}>0$ on $\mathbf{S}^{n}\backslash\\{\bar{p}_{1},\dots,\bar{p}_{k}\\}$. It only remains to verify that $\overline{g}$ is complete. If $\overline{g}$ is incomplete then there exists $j\in\\{1,\dots,k\\}$ such that $\liminf_{x\rightarrow\bar{p}_{j}}\overline{u}(x)<\infty.$ In this case Theorem 9 implies $\mathcal{P}_{\textrm{rad}}(\overline{g},\bar{p}_{j})=0$. However, by construction $\mathcal{P}_{\textrm{rad}}(\overline{g},\bar{p}_{j})=\lim_{i\rightarrow\infty}\mathcal{P}_{\textrm{rad}}(g_{i},p_{j}^{i})\geq\delta_{2},$ giving a contradiction. We conclude that $\overline{g}$ is indeed in $\Omega_{\delta_{1},\delta_{2}}$. ∎ ## References * [1] M. O. Ahmedou, Z. Djadli, and A. Malchiodi. Prescribing a fourth-order conformal invariant on the standard sphere II: blow-up analysis and applications. Ann. Scuola Norm. Sup. Pisa 5 (2002), 387–434. * [2] J. H. Andrade and J. M. do Ó. Asymptotics for singular solutions of conformally invariant fourth order systems in the punctured ball. preprint, arxiv:2003.03487. * [3] T. Aubin. Équations différentielles non linéaires et problème de Yamabe concernant la courbure scalaire. J. Math. Pures Appl. 55 (1976), 269–296. * [4] J. van den Berg. The phase-plane picture for a class of fourth-order conservative differential equations. J. Differential Equations. 161 (2000), 110–153. * [5] T. Branson. Differential operators canonically associated to a conformal structure. Math. Scandinavia. 57 (1985), 293–345. * [6] T. Branson. Group representations arising from Lorentz conformal geometry. J. Funct. Anal. 74 (1987), 199–291. * [7] T. Branson and A. R. Gover. Origins, applications and generalisations of the $Q$-curvature. Acta Appl. Math. 102 (2008), 131–146. * [8] G. Caristi and E. Mitidieri. Harnack inequalities and applications to solutions of biharmonic equations. Operator Theory: Advances and Applications. 168 (2006), 1–26. * [9] S.-Y. A. Chang, M. Eastwood, B. Ørsted, and P. Yang. What is $Q$-curvature? Acta Appl. Math. 102 (2008), 119–125. * [10] S.-Y. A. Chang, Z.-C. Han, and P. Yang. Some remarks on the geometry of a class of locally conformally flat metrics. Progress in Math. 333 (2020), 37–56. * [11] R. Frank and T. König. Classification of positive solutions to a nonlinear biharmonic equation with critical exponent. Anal. PDE 12 (2019), 1101–1113. * [12] M. Gursky and A. Malchiodi. A strong maximum principle for the Paneitz operator and a non-local flow for the $Q$-curvature. J. Eur. Math. Soc. 17 (2015), 2137–2173. * [13] A. R. Gover and B. Ørsted. Universal principles for Kazdan-Warner and Pohozaev-Schoen type identities. Comm. Contemp. Math. 15 (2013), * [14] F. Hang and P. Yang. Lectures on the fourth order $Q$-curvature equation. Geometric analysis around scalar curvature, Lect. Notes Ser. Inst. Math. Sci. Natl. Univ. Singap. 31 (2016), 1–33. * [15] F. Hang and P. Yang. $Q$-curvature on a class of manifolds with dimension at least $5$. Comm. Pure Appl. Math. 69 (2016), 1452–1491. * [16] T. Jin and J. Xiong. Asymptotic symmetry and local behavior of solutions of higher order conformally invariant equations with isolated singularities. preprint, arxiv:1901.01678. * [17] J. Lee and T. Parker. The Yamabe problem. Bull. Amer. Math. Soc. 17 (1987), 37–91. * [18] Y.-J. Lin and W. Yuan. A symmetric $2$-tensor cannonically associated to $Q$-curvature and its applications. Pac. J. Math. 291 (2017) 425–438. * [19] M. Khuri, F. C. Marques, and R. Schoen. A compactness theorem for the Yamabe problem. J. Differential Geom. 81 (2009), 143–196. * [20] C. S. Lin. A classification of solutions of a conformally invariant fourth order equation in $\mathbf{R}^{n}$. Comment. Math. Helv. 73 (1998), 206–231. * [21] R. Mazzeo, D. Pollack, and K. Uhlenbeck. Moduli spaces of singular Yamabe metrics. J. Amer. Math. Soc. 9 (1996), 303–344. * [22] S. Paneitz. A quartic conformally covariant differential operator for arbitrary pseudo-Riemannian manifolds. SIGMA Symmetry Integrability Geom. Methods Appl. 4 (2008), 3 pages (preprint from 1983). * [23] D. Pollack. Compactness results for complete metrics of constant positive scalar curvature on subdomains of $\mathbf{S}^{n}$. Indiana Univ. Math. J. 42 (1993), 1441–1456. * [24] J. Ratzkin. On constant $Q$-curvature metrics with isolated singularities. preprint, arXiv:2001.07984. * [25] R. Schoen. Conformal deformation of a Riemannian metric to constant scalar curvature. J. Diff. Geom. 20 (1984), 479–495. * [26] N. Trudinger. Remarks concerning the conformal deformation of Riemannian structures on compact manifolds. Ann. Scuola Norm. Pisa. 22 (1968), 265–274. * [27] W. Wei. Compactness theorem of complete $k$-curvature manifolds with isolated singularities. preprint, arxiv.2008.08777. * [28] H. Yamabe. On the deformation of Riemannian structures on a compact manifold. Osaka Math. J. 12 (1960), 21–37.
# Doppler Estimation for High–Velocity Targets Using Subpulse Processing and the Classic Chinese Remainder Theorem Fernando Darío Almeida García, André Saito Guerreiro, Gustavo Rodrigues de Lima Tejerina, José Cândido S. Santos Filho, Gustavo Fraidenraich, and Michel Daoud Yacoub F. D. A. García, A. S. Guerreiro, G. R. de Lima Tejerina, J. C. S. Santos Filho, G. Fraidenraich, and M. D. Yacoub are with the Wireless Technology Laboratory, Department of Communications, School of Electrical and Computer Engineering, University of Campinas, 13083-852 Campinas, SP, Brazil, Tel:+55(19)3788-5106, e-mail:$\left\\{\right.$ferdaral, andsaito, tejerina, candido, gf, <EMAIL_ADDRESS> ###### Abstract In pulsed Doppler radars, the classic Chinese remainder theorem (CCRT) is a common method to resolve Doppler ambiguities caused by fast-moving targets. Another issue concerning high-velocity targets is related to the loss in the signal-to-noise ratio (SNR) after performing range compression. In particular, this loss can be partially mitigated by the use of subpulse processing (SP). Modern radars combine these techniques in order to reliably unfold the target velocity. However, the presence of background noise may compromise the Doppler estimates. Hence, a rigorous statistical analysis is imperative. In this work, we provide a comprehensive analysis on Doppler estimation. In particular, we derive novel closed-form expressions for the probability of detection (PD) and probability of false alarm (PFA). To this end, we consider the newly introduce SP along with the CCRT. A comparison analysis between SP and the classic pulse processing (PP) technique is also carried out. Numerical results and Monte- Carlo simulations corroborate the validity of our expressions and show that the SP–plus–CCRT technique helps to greatly reduce the PFA compared to previous studies, thereby improving radar detection. ###### Index Terms: Classic Chinese remainder theorem, robust Chinese remainder theorem, Doppler frequency estimation, subpulse processing, probability of detection. ## I Introduction One important concern in modern pulsed radars is related to the Doppler frequency estimation of fast-velocity targets. Due to the high target’s radial velocity, ambiguous estimates are more likely to occur. More specifically, ambiguous estimates appear whenever the target’s Doppler shift is greater than the pulse repetition frequency (PRF) [1]. It seems obvious to think that increasing the PRF will overcome this problem. However, if we are interested in detecting targets located at long distances, then the PRF will be restricted to a maximum value. Therefore, the choice of PRF is a trade-off between range and Doppler requirements [2]. Fortunately, there are some techniques that can resolve ambiguities, although at the cost of extra measurement time and processing load. These techniques make use of multiples PRFs [3, 4, 5, 6, 7]. The most known and used technique is the classic Chinese remainder theorem (CCRT). The CCRT is a fast and accurate method to resolve the unambiguous Doppler frequency. This is accomplished by solving a set of congruences, formed by the estimated measurements of each PRF [7, 8, 9]. Nevertheless, in this method, the number of PRFs will not be sufficient to resolve a certain quantity of targets. In general, $L$ PRFs are required to successfully disambiguate $L-1$ targets. If the number of targets exceeds $L-1$, then ghosts can appear.111Ghosts are false targets resulting from false coincidences of Doppler-ambiguous or range-ambiguous data [4]. Unless additional data (e.g., tracking information) is available, the radar has no way of recognizing possible false detections [4]. Care must be taken in the analysis and design since the number of PRFs and the number of targets to be detected have a direct relationship. Another issue concerning high-velocity targets is related to the signal-to- noise ratio (SNR) loss. This occurs because the Doppler shift of fast-moving targets will provoke a mismatch between the received signal and its replica [2]. Consequently, the SNR after range compression may be drastically reduced.222Range compression refers to the convolution operation between the received signal and the replica of the transmitted signal [10]. Some radar systems estimate and remove the Doppler shift prior to applying range compression. Nonetheless, some residual or uncompensated Doppler typically remains. This concern was partially alleviated in [11, 12]. Specifically, in [11], the authors proposed a subpulse processing (SP), which proved to have a higher Doppler tolerance,333Doppler tolerance refers to the degree of degradation in the compressed response due to uncompensated Doppler [13]. increasing the ability to detect fast-moving targets. The shortcomings of SP are computation time (critical for most radars), processing load, and poor velocity resolution. As stated before, the CCRT and SP have hardware and physical limitations when it comes to estimating high target velocities. In practice, modern pulsed radars take advantage of these two techniques so as to improve the system’s capability to accurately detect the target’s true Doppler frequency. Since SP the CCRT are affected by the presence of background noise, then a thorough statistical analysis involving these two estimation techniques must be carried out. Recently in [14], the authors proposed a novel expression for the probability to correctly estimate the unambiguous Doppler frequency considering the CCRT and the common pulse processing (PP) technique [2]. However, to the best of our knowledge, there is no performance analysis considering the SP–plus–CCRT technique. The main objective of this research is to combine the statistical analysis conducted in [14] along with the newly introduced SP and the CCRT. To do so, we adopt a stochastic model that suits our Doppler estimation techniques. Then, we derive novel and closed-form expressions for: i) the probability to correctly estimate the Doppler frequency, also called probability of detection (PD); and ii) the probability to erroneously estimate the Doppler frequency, also called probability of false alarm (PFA). The remainder of this paper is organized as follows. Section II introduces some key concepts to understand how the velocity estimation is performed. Section III describes the system model. Section IV analyzes the Doppler estimation using multiple PRFs; Section V discusses the representative numerical results. Finally, Section VI concludes this paper. In what follows, $(a)\text{mod}(b)$ denotes the remainder of the euclidean division of $a$ by $b$; $\left|\cdot\right|$, absolute value; $\lfloor\cdot\rfloor$, floor operation; $\text{round}(\cdot)$, rounding operation; $\text{Pr}\left[\cdot\right]$, probability; $\mathbb{E}(\cdot)$, expectation; $\text{Var}(\cdot)$, variance; $(\cdot)^{*}$, complex conjugate; $\bigcap$, intersection of events; $\bigcup$, union of events; $\mathcal{N}(\mu,\sigma^{2})$ denotes a Gaussian distribution with mean $\mu$ and variance $\sigma^{2}$; $\mathcal{N}_{c}(\mu,\sigma^{2})$ denotes a complex Gaussian distribution with mean $\mu$ and variance $\sigma^{2}$, and $j=\sqrt{-1}$ is the imaginary unit. ## II Preliminaries In this section, we present a brief introduction about the PP and SP techniques. Latter, we describe the basis to understand the CCRT algorithm. Finally, we show how the combined technique SP–plus–CCRT works in order to improve Doppler estimation. ### II-A Pulse Processing PP is the common technique employed by the radar to estimate the target velocity and improve the SNR. In this processing technique, the radar transmits a sequence of $M$ pulses during a coherent processing interval (CPI) [15]. Then, range compression is performed on each pulse to improve the radar’s range resolution. Finally, the discrete Fourier transform (DFT) is applied along the slow-time samples to increase the SNR and to estimate the target Doppler frequency [2]. These samples are collected at a rate equal to the PRF. The maximum Doppler frequency shift that the radar manages to detect using PP is $\Psi_{max}=\pm\text{PRF}/2$. If the target Doppler frequency, $\mathit{f}_{d}$, exceeds this value, then the radar will deliver ambiguous Doppler measurements. The Doppler frequency shift will be positive for closing targets and negative for receding targets. The target velocity, $\mathit{v}_{t}$, and its corresponding Doppler shift are related by the following equation [16]: $\displaystyle\mathit{f}_{d}=\frac{2\mathit{v}_{t}\mathit{f}_{R}}{\mathit{c}}=\frac{2\mathit{v}_{t}}{\lambda},$ (1) where $\mathit{f}_{R}$ is the radar’s operation frequency, $\mathit{c}$ is the speed of light, and $\lambda$ is the radar frequency. ### II-B Subpulse Processing SP improves Doppler tolerance by mitigating the loss in SNR caused by the uncompensated Doppler shift of fast-moving targets [2, 11]. Moreover, SP is used to overcome the problem of ambiguous Doppler measurements. The SP algorithm runs as follows: 1. 1. First, the replica of the transmitted signal is divided into $N$ subpulses – unlike PP that used the entire replica. 2. 2. Latter, range compression is carried out between each subpulse and the received signal (cf. [11, 12] for a detailed discussion on this). The use of shorter replicas will enhance the system’s Doppler tolerance [10], increasing the detection capability of fast-moving targets. Of course, this process leads to a reduction in the peak amplitude of the sub-compression response (by a factor of $1/N$). Here, the slow-time samples are collected at a rate of $\Phi=N/\tau$, where $\tau$ is the pulse width. It is important to emphasize that PP and SP are performed simultaneously, that is, for each of the $M$ compressions, the radar carried out $N$ sub-compressions [12]. 3. 3. Finally, the slow-time samples are coherently integrated to estimate the target Doppler frequency and to “restore” the peak amplitude of the sub- compression response. The number of subpulses can be chosen as high as needed, as long as it is taken into consideration that each additional subpulse requires an extra range compression operation, increasing the computational load and computation time. The maximum Doppler frequency shift that the radar can now manage to detect is $\Phi_{max}=\pm N/2\tau$ [11]. Since $\Phi_{max}>\Psi_{max}$, SP provides a higher frequency range of detection for fast-moving targets. Computation time is critical for most radars and depends strongly on the radar’s operation mode (e.g. tracking, searching or imaging), thereby limiting the number of subpulses. Commonly, the number of subpulses is set between 5 and 10. However, this small number yields to a poor discretization in the frequency domain and, consequently, producing inaccurate estimates. Figure 1: Block diagram for Doppler estimation. ### II-C Classic Chinese Remainder Theorem The use of multiples PRFs is a common approach to resolve range and Doppler ambiguities [3, 8, 17, 4]. In this work, we only focus on solving Doppler ambiguities. Consider for the moment a target with Doppler shift $\mathit{f}_{d}>\Psi_{max}$. In this scenario, the radar will detect the target with an apparent Doppler shift, $\mathit{f}_{d_{ap}}$, that satisfies $\displaystyle\mathit{f}_{d}=\mathit{f}_{d_{ap}}+n\text{PRF},$ (2) where $n$ is some integer. It is convenience to express the target’s Doppler shift $\mathit{f}_{d}$ in terms of its corresponding Doppler bin, $b_{d}$. Thus, (2) becomes $b_{d}=b_{ap}+nM,$ (3) in which $b_{{ap}}\in\left\\{0,1,2,\ldots,M-1\right\\}$ is the apparent Doppler bin, defined as $\displaystyle b_{ap}=$ $\displaystyle\left\lfloor\absolutevalue{\frac{\mathit{f}_{d_{ap}}}{\Delta D}}\right\rfloor,\ \ \ \ \ \ \ \ \ \ \ \ \mathit{f}_{d_{ap}}\geq 0$ (4) $\displaystyle b_{ap}=$ $\displaystyle M-\left\lfloor\absolutevalue{\frac{\mathit{f}_{d_{ap}}}{\Delta D}}\right\rfloor,\ \ \ \ \ \ \mathit{f}_{d_{ap}}<0$ (5) with $\Delta D=\text{PRF}/M$ being the Doppler bin spacing. Under this scenario, the radar is incapable to detect the target’s true Doppler frequency. Now, suppose that we have $L$ PRFs. Then, the unambiguous target’s Doppler bin must satisfies the following congruences: $\displaystyle b_{d}\equiv b_{{ap}_{i}}+n_{i}M_{i},\ \ \ \ 1\leq i\leq L$ (6) The CCRT states that if all PRFs are pairwisely coprimes, then the set of congruences in (6) will have a unique solution given by [17, 4, 18] $\displaystyle b_{d}=\left(\sum_{i=1}^{L}b_{{ap}_{i}}\beta_{i}\right)\text{mod}\left(\Theta\right),$ (7) where $\Theta=\prod_{i=1}^{L}M_{i}$, $\beta_{i}=b_{i}\Theta/\text{PRF}_{i}$, and $b_{i}$ is the smaller integer which can be computed by solving the following expression: $\displaystyle\left(\frac{b_{i}\Theta}{M_{i}}\right)\text{mod}\left(M_{i}\right)=1.$ (8) ### II-D Doppler Estimation Fig. 1 depicts the entire block diagram for Doppler estimation. First, the received signal passes through two types of independent range compression blocks, one for PP and one for SP. This process is performed in sequence for each pulse repetition interval (PRI). The outputs of both blocks are combined and stored in memory to form a datacube [2]. (The datacube’s data is organized by range, number of pulses, and number of subpulses.) More datacubes are needed when using more than one PRF, as shown in Fig. 1. Next, a 2D-DFT block is applied to each datacube to perform coherent integration. (The 2D-DFT block is referred to as a two-dimensional DFT applied along with pulses and subpulses.) Latter, the output of the 2D-DFT block is a matrix with the same size containing the estimated Doppler shifts. This new matrix is referred to as Doppler datacube. Finally, the CCRT is applied over the Doppler datacubes. This process will be clarified in Section V by means of simulation. Noise, jammer, and clutter are major concerns in all radar systems. In this work, we consider the presence complex white Gaussian noise (CWGN). Thus, the Doppler spectrum of fast moving targets will be compromised due to the intrinsic characteristics of noise. For example, a high noise power could mask small target returns, degrading radar performance. Even if the target return is entirely deterministic, the combined signal (target–plus–noise) is a random process and must be treated as such. Therefore, we need to assess the statistics underlying Doppler analysis, but first, we need to come up with a specific stochastic model that suits the requirements and design of our radar’s estimation scheme. This is discussed in the next section. ## III System Model In this section, we propose a stochastic model that fits our signal processing schemes. In addition, we describe the premises (hypotheses) used for Doppler estimation. According to Sections II-A and II-B, the collected signals in the slow-time domain corresponding to PP and SP can be expressed, respectively, as $\displaystyle g_{1}\left[m\right]=$ $\displaystyle s_{1}\left[m\right]+w_{1}\left[m\right]$ $\displaystyle=$ $\displaystyle a_{1}\exp\left(j2\pi\mathit{f}_{d}m/\text{PRF}\right)+w_{1}\left[m\right],\ \ 0\leq m\leq M-1$ (9) $\displaystyle g_{2}\left[n\right]=$ $\displaystyle s_{2}\left[n\right]+w_{2}\left[n\right]$ $\displaystyle=$ $\displaystyle a_{2}\exp\left(j2\pi\mathit{f}_{d}n/\Phi\right)+w_{2}\left[n\right],\ \ \ \ \ \ \ 0\leq n\leq N-1$ (10) where $s_{1}\left[m\right]$ and $s_{2}\left[n\right]$ are discrete complex sine signals444In most systems, the radio frequency (RF) signal is mixed to baseband prior to compression, and a coherent detector is used in the downconversion process to form in-phase (I) and quadrature (Q) receive channels, thereby creating a complex baseband signal. originated by changes in the target position; $w_{1}\left[m\right]$ and $w_{2}\left[n\right]$ are discrete additive complex Gaussian noises; and finally, $a_{1}$ and $a_{2}$ are the amplitudes at the output of the matched filters. Depending on the target velocity, the output amplitudes $a_{1}$ and $a_{2}$ maybe be greatly attenuated. However, the attenuation in $a_{2}$ is partially mitigated by the use of SP. In particular, it follows that $a_{2}>a_{1}$ for high-velocity targets [11]. Additionally, we define $2\sigma_{t_{1}}^{2}$ and $2\sigma_{t_{2}}^{2}$ as the total mean powers – in the time domain – for $w_{1}\left[m\right]$ and $w_{2}\left[n\right]$, respectively. As seen in practice, and due to the stationary characteristic of noise, we have that $\sigma_{t_{1}}^{2}=\sigma_{t_{2}}^{2}$ [19]. However, we will remain using separate notations for $\sigma_{t_{1}}^{2}$ and $\sigma_{t_{2}}^{2}$ so as to distinguish the noise power from PP and SP. Of course, these separate notations will not alter, in any form, our performance analysis. The SNR measured in the time domain considering PP and SP, can be expressed, respectively, as $\displaystyle\text{SNR}_{t_{1}}=$ $\displaystyle\frac{\left|a_{1}\right|^{2}}{2\sigma_{t_{1}}^{2}}$ (11) $\displaystyle\text{SNR}_{t_{2}}=$ $\displaystyle\frac{\left|a_{2}/N\right|^{2}}{2\sigma_{t_{2}}^{2}}.$ (12) Observe in (12) that the fact of dividing the replica into $N$ subpulses causes a reduction in the SNR by a factor of $1/N^{2}$, as mentioned in Section II-B. The DFT is the primary operation to implement coherent integration. More precisely, the DFT provides a mechanism to test multiple candidate frequencies to maximize the integration gain [2]. The corresponding DFTs for (III) and (III) are given, respectively, by $\displaystyle G_{1}\left[k^{\prime}\right]\triangleq$ $\displaystyle\ \mathscr{F}\left\\{g_{1}\left[m\right]\right\\}$ $\displaystyle=$ $\displaystyle\sum_{m=0}^{M-1}g_{1}\left[m\right]\exp\left(-j2\pi k^{\prime}m/M\right)$ $\displaystyle=$ $\displaystyle S_{1}\left[k^{\prime}\right]+W_{1}\left[k^{\prime}\right],\ \ \ 0\leq k^{\prime}\leq M-1$ (13) $\displaystyle G_{2}\left[l^{\prime}\right]\triangleq$ $\displaystyle\ \mathscr{F}\left\\{g_{2}\left[n\right]\right\\}$ $\displaystyle=$ $\displaystyle\sum_{n=0}^{N-1}g_{2}\left[n\right]\exp\left(-j2\pi l^{\prime}n/N\right)$ $\displaystyle=$ $\displaystyle S_{2}\left[l^{\prime}\right]+W_{2}\left[l^{\prime}\right],\ \ \ \ \ \ 0\leq l^{\prime}\leq N-1$ (14) The SNR measured in the frequency domain considering PP and SP, are given, respectively, by [2, Eq. (17.37)] $\displaystyle\text{SNR}_{1}=$ $\displaystyle\frac{|Ma_{1}|^{2}}{2\sigma_{1}^{2}}$ (15) $\displaystyle\text{SNR}_{2}=$ $\displaystyle\frac{|a_{2}|^{2}}{2\sigma_{2}^{2}},$ (16) in which $\sigma_{1}^{2}=M\sigma_{t_{1}}^{2}$ and $\sigma_{2}^{2}=N\sigma_{t_{1}}^{2}$ are half of the noise powers – in the frequency domain – for $W_{1}\left[k^{\prime}\right]$ and $W_{2}\left[l^{\prime}\right]$, respectively. The Doppler estimates are based on the absolute values of $G_{1}\left[k^{\prime}\right]$ and $G_{2}\left[l^{\prime}\right]$. That is, (III) and (III) will provide estimates for $\mathit{f}_{d}$, say $\hat{\mathit{f}}_{1}$ and $\hat{\mathit{f}}_{2}$, by searching $k^{\prime}$ and $l^{\prime}$, in which the absolute values of $G_{1}\left[k^{\prime}\right]$ and $G_{2}\left[l^{\prime}\right]$ are maximum. It is worth mentioning that if $\Psi_{max}<\mathit{f}_{d}$ and $\Phi_{max}<\mathit{f}_{d}$, then $\hat{\mathit{f}}_{1}$ and $\hat{\mathit{f}}_{2}$ will display ambiguous Doppler estimates. Now, considering $L$ PRFs (say, $\text{PRF}_{1},\ldots,\text{PRF}_{L}$), we can define the absolute values for $G_{1}\left[k^{\prime}\right]$ and $G_{2}\left[l^{\prime}\right]$ at the $i$-th PRF, respectively, as $\displaystyle H_{1,i}\left[k^{\prime}\right]\triangleq$ $\displaystyle|G_{1,i}\left[k^{\prime}\right]|\ \ \ \ \ 0\leq k^{\prime}\leq M_{i}-1$ (17) $\displaystyle H_{2,i}\left[l^{\prime}\right]\triangleq$ $\displaystyle|G_{2,i}\left[l^{\prime}\right]|\ \ \ \ \ \ 0\leq l^{\prime}\leq N_{i}-1$ (18) where the subscript $i\in\\{1,\ldots,L\\}$ denotes the association to the $i$-th PRF. Herein, we assume that $G_{1,i}\left[k^{\prime}\right]$ is composed of $M_{i}-1$ independent and identically distributed noise samples and one target–plus–noise sample, denoted as $\mathcal{G}_{1,i}$. On the other hand, $G_{2,i}\left[l^{\prime}\right]$ is composed of $N_{i}-1$ independent and identically distributed noise samples and one combined sample, denoted as $\mathcal{G}_{2,i}$. The target–plus–noise samples $\mathcal{G}_{1,i}$ and $\mathcal{G}_{2,i}$ can be modeled, respectively, by [20, Eq. (1)] $\displaystyle\mathcal{G}_{1,i}=$ $\displaystyle\sigma_{1,i}\left(\sqrt{1-\lambda_{1,i}^{2}}A_{1,i}+\lambda_{1,i}A_{0,i}\right)$ $\displaystyle+j\sigma_{1,i}\left(\sqrt{1-\lambda_{1,i}^{2}}B_{1,i}+\lambda_{1,i}B_{0,i}\right)$ (19) $\displaystyle\mathcal{G}_{2,i}=$ $\displaystyle\sigma_{2,i}\left(\sqrt{1-\lambda_{2,i}^{2}}A_{2,i}+\lambda_{2,i}A_{0,i}\right)$ $\displaystyle+j\sigma_{2,i}\left(\sqrt{1-\lambda_{2,i}^{2}}B_{2,i}+\lambda_{2,i}B_{0,i}\right),$ (20) where $A_{p,i}$ and $B_{p,i}$ ($p=1,2$) are mutually independent random variables (RVs) distributed as $\mathcal{N}(0,\frac{1}{2})$, and $\lambda_{p,i},\in(0,1]$. Then, for any $p$ and $q$ ($q=1,2$), it follows that $\mathbb{E}(A_{p,i}B_{q,i})=0$ and $\mathbb{E}(A_{p,i}A_{q,i})=\mathbb{E}(B_{p,i}B_{q,i})=\frac{1}{2}\delta_{pq}$. ($\delta_{pq}=1$ if $p=q$, and $\delta_{pq}=0$ otherwise.) In addition, $A_{0,i}$ and $B_{0,i}$ are mutually independent RVs distributed as $\mathcal{N}(m_{\textbf{Re},i},\frac{1}{2})$ and $\mathcal{N}(m_{\textbf{Im},i},\frac{1}{2})$, respectively. Thus, $\mathcal{G}_{1,i}$ and $\mathcal{G}_{2,i}$ are non-zero mean complex Gaussian RVs with probability density functions (PDFs) given, respectively, by $\mathcal{N}_{c}(\lambda_{1,i}(m_{\textbf{Re},i}+jm_{\textbf{Im},i}),\sigma_{1,i}^{2})$ and $\mathcal{N}_{c}(\lambda_{2,i}(m_{\textbf{Re},i}+jm_{\textbf{Im},i}),\sigma_{2,i}^{2})$. The correlation coefficient between any pair of ($\mathcal{G}_{1,i}$, $\mathcal{G}_{2,i}$), can be calculated as [20, Eq. (2)] $\displaystyle\rho_{kl,i}\triangleq$ $\displaystyle\frac{\mathbb{E}(\mathcal{G}_{1,i}\mathcal{G}_{2,i}^{*})-\mathbb{E}(\mathcal{G}_{1,i})\mathbb{E}(\mathcal{G}_{2,i}^{*})}{\sqrt{\text{Var}(\mathcal{G}_{1,i})\text{Var}(\mathcal{G}_{2,i})}}$ $\displaystyle=$ $\displaystyle\lambda_{1,i}\lambda_{2,i}.$ (21) This correlation exists because both PP and SP use the same received signal when performing range compression [2]. Observe that the parameters $\lambda_{1,i}^{2}$, $\lambda_{2,i}^{2}$, $m_{\textbf{Re},i}$ and $m_{\textbf{Im},i}$ can be used to model the compressed responses $\left|M_{i}a_{1,i}\right|^{2}$ and $\left|a_{2,i}\right|^{2}$. This can be done by making the following substitutions: $|M_{i}a_{1,i}|^{2}=\lambda_{1,i}^{2}(m_{\textbf{Re},i}^{2}+m_{\textbf{Im}}^{2})$ and $|a_{2,i}|^{2}=\lambda_{2,i}^{2}(m_{\textbf{Re},i}^{2}+m_{\textbf{Im}}^{2})$. On the other hand, $\lambda_{1,i}$ and $\lambda_{2,i}$ can be chosen to meet a desire correlation coefficient. By the above, it follows that $H_{1,i}\left[k^{\prime}\right]$ is composed of $M_{i}-1$ Rayleigh distributed samples, denoted as $X_{k,i}$ $\left(k\in\left\\{1,2,\ldots,M_{i}-1\right\\}\right)$, and one Rice distributed sample, denoted as $R_{1,i}$. Similarly, $H_{2,i}\left[l^{\prime}\right]$ is composed of $N_{i}-1$ Rayleigh distributed samples, denoted as $Y_{l,i}$ $\left(l\in\left\\{1,2,\ldots,N_{i}-1\right\\}\right)$, and one Rice distributed sample, denoted as $R_{2,i}$. The PDFs of $X_{k,i}$ and $Y_{l,i}$ are given, respectively, by $\displaystyle f_{X_{k,i}}(x_{k,i})=$ $\displaystyle\frac{x_{k,i}\exp\left(-\frac{x_{k,i}^{2}}{2\sigma_{k,i}^{2}}\right)}{\sigma_{k,i}}$ (22) $\displaystyle f_{Y_{l,i}}(y_{l,i})=$ $\displaystyle\frac{y_{l,i}\exp\left(-\frac{y_{l,i}^{2}}{2\sigma_{l,i}^{2}}\right)}{\sigma_{l,i}}.$ (23) Moreover, since $R_{2,i}$ and $R_{2,i}$ bear a certain degree of correlation, they are governed by a bivariate Rician distribution, given by [20, 21] $\displaystyle\mathit{f}_{R_{1,i},R_{2,i}}$ $\displaystyle\left(r_{1,i},r_{2,i}|\mathcal{H}_{1}\right)=\int_{0}^{\infty}\exp\left(-t\xi_{i}\right)$ $\displaystyle\times\exp\left(-\textbf{m}_{i}\right)I_{0}\left(2\sqrt{\textbf{m}_{i}t}\right)\prod_{p=1}^{2}\frac{r_{p,i}}{\Omega_{p,i}^{2}}$ $\displaystyle\times\exp\left(-\frac{r_{p,i}^{2}}{2\Omega_{p,i}^{2}}\right)I_{0}\left(\frac{r_{p,i}\sqrt{t\sigma_{p,i}^{2}\lambda_{p,i}^{2}}}{\Omega_{p,i}^{2}}\right)\text{d}t,$ (24) where $I_{0}(\cdot)$ is the modified Bessel function of the first kind and order zero [22, Eq. (9.6.16)], $\textbf{m}_{i}=m_{\textbf{Re},i}^{2}+m_{\textbf{Im},i}^{2}$, and $\displaystyle\Omega_{p,i}^{2}$ $\displaystyle=\sigma_{p,i}^{2}\left(\frac{1-\lambda_{p,i}^{2}}{2}\right)$ (25a) $\displaystyle\xi_{i}$ $\displaystyle=1+\sum_{p=1}^{2}\frac{\sigma_{p,i}^{2}\lambda_{p,i}^{2}}{2\Omega_{p,i}^{2}}.$ (25b) ## IV Doppler Analysis In this section, we provide a comprehensive statistical analysis on Doppler estimation. To do so, we derive the performance metrics for both SP and SP–plus–CCRT. ### IV-A SP Analysis First, let us define the following events: $\displaystyle\mathcal{A}_{k,i}=$ $\displaystyle\left\\{R_{1,i}>X_{k,i}\right\\}$ (26) $\displaystyle\mathcal{B}_{l,i}=$ $\displaystyle\left\\{R_{2,i}>Y_{l,i}\right\\}$ (27) $\displaystyle\mathcal{C}_{k,i}=$ $\displaystyle\left\\{X_{k,i}>R_{1,i}\right\\}$ (28) $\displaystyle\mathcal{D}_{l,i}=$ $\displaystyle\left\\{Y_{l,i}>R_{2,i}\right\\}.$ (29) ###### Proposition I. Let $\text{PD}_{i}$ be the probability of detection at the $i$-th PRF. Specifically, $\text{PD}_{i}$ is defined as the probability that $R_{1,i}$ is greater than $X_{k,i}$ and, simultaneously, that $R_{2,i}$ is greater than $Y_{l,i}$, i.e., $\displaystyle\text{PD}_{i}\triangleq\text{Pr}$ $\displaystyle\left[\left(\bigcap_{k=1}^{M_{i}-1}\mathcal{A}_{k,i}\right)\bigcap\left(\bigcap_{l=1}^{N_{i}-1}\mathcal{B}_{l,i}\right)\right].$ (30) Then, from (22)–(III), (30) can be expressed in closed-form as $\displaystyle\text{PD}_{i}=$ $\displaystyle\sum_{k=0}^{M_{i}-1}\sum_{l=0}^{N_{i}-1}\left(\begin{array}[]{c}M_{i}-1\\\ k\\\ \end{array}\right)\left(\begin{array}[]{c}N_{i}-1\\\ l\\\ \end{array}\right)$ (35) $\displaystyle\times\frac{(-1)^{-k-l+M_{i}+N_{i}}\mathcal{V}_{i}(k,l)}{\mathcal{U}_{i}(k,l)}\exp\left(-\textbf{m}_{i}+\frac{\textbf{m}_{i}}{\mathcal{U}_{i}(k,l)}\right),$ (36) wherein $\mathcal{U}_{i}(k,l)$ and $\mathcal{V}_{i}(k,l)$ are auxiliary functions defined, respectively, as $\displaystyle\mathcal{U}_{i}(k,l)=$ $\displaystyle\ \xi_{i}-\frac{\xi_{i}\lambda_{1,i}^{2}\sigma_{1,i}^{4}}{2\Omega_{1,i}^{2}\left(\Omega_{1,i}^{2}(k-M_{i}+1)-\sigma_{1,i}^{2}\right)}$ $\displaystyle-\frac{\xi_{i}\lambda_{2,i}^{2}\sigma_{2,i}^{4}}{2\Omega_{2,i}^{2}\left(\Omega_{2,i}^{2}(l-N_{i}+1)-\sigma_{2,i}^{2}\right)}$ (37a) $\displaystyle\mathcal{V}_{i}(k,l)=$ $\displaystyle\ \frac{\sigma_{1,i}^{2}}{\left(\Omega_{1,i}^{2}(-k+M_{i}-1)+\sigma_{1,i}^{2}\right)}$ $\displaystyle\times\frac{\sigma_{2,i}^{2}}{\left(\Omega_{2,i}^{2}(-l+N_{i}-1)+\sigma_{2,i}^{2}\right)}.$ (37b) ###### Proof. See Appendix A. ∎ ###### Corollary I. Let $\text{PFA}_{i}$ be the probability of false alarm at the $i$-th PRF. More precisely, $\text{PFA}_{i}$ is defined as the probability that at least one of $X_{k,i}$ is greater than $R_{1,i}$ and, simultaneously, that at least one of $Y_{l,i}$ is greater than $R_{2,i}$, i.e., $\displaystyle\text{PFA}_{i}\triangleq\text{Pr}$ $\displaystyle\left[\underset{k=1}{\overset{M_{i}-1}{\bigcup}}\underset{l=1}{\overset{N_{i}-1}{\bigcup}}\left(\mathcal{C}_{k,i}\bigcap\mathcal{D}_{l,i}\right)\right].$ (38) Then, from (22)–(III), (38) can be written in closed-form as in (Corollary I), shown at the top of the next page, where $\mathcal{P}_{i}\left(k,l\right)$ and $\mathcal{Q}_{i}\left(k,l\right)$ are auxiliary functions defined, respectively, by $\displaystyle\mathcal{P}_{i}\left(k,l\right)=$ $\displaystyle\ \xi_{i}-\frac{\lambda_{1,i}^{2}\sigma^{4}_{1,i}}{2\Omega_{1,i}^{2}\left(k\ \Omega_{1,i}^{2}+\sigma^{2}_{1,i}\right)}$ $\displaystyle-\frac{\lambda_{2,i}^{2}\sigma^{4}_{2,i}}{2\Omega_{2,i}^{2}\left(l\ \Omega_{2,i}^{2}+\sigma^{2}_{2,i}\right)}$ (39a) $\displaystyle\mathcal{Q}_{i}\left(k,l\right)=$ $\displaystyle\ \frac{\sigma^{2}_{1,i}\sigma^{2}_{2,i}}{\left(k\ \Omega_{1,i}^{2}+\sigma^{2}_{1,i}\right)\left(l\ \Omega_{2,i}^{2}+\sigma^{2}_{2,i}\right)}.$ (39b) $\displaystyle\textit{PFA}_{i}=$ $\displaystyle\frac{\left(M_{i}-1\right)\left(N_{i}-1\right)\mathcal{Q}_{i}\left(1,1\right)}{\mathcal{P}_{i}\left(1,1\right)}\exp\left(-\textbf{m}_{i}+\frac{\textbf{m}_{i}}{\mathcal{P}_{i}\left(1,1\right)}\right)-\binom{M_{i}-1}{2}\binom{N_{i}-1}{2}\frac{\mathcal{Q}_{i}\left(2,2\right)}{\mathcal{P}_{i}\left(2,2\right)}\exp\left(-\textbf{m}_{i}+\frac{\textbf{m}_{i}}{\mathcal{P}_{i}\left(2,2\right)}\right)+\ldots$ $\displaystyle+(-1)^{M_{i}-N_{i}-1}\frac{\mathcal{Q}_{i}\left(M_{i}-1,N_{i}-1\right)}{\mathcal{P}_{i}\left(M_{i}-1,N_{i}-1\right)}\exp\left(-\textbf{m}_{i}+\frac{\textbf{m}_{i}}{\mathcal{P}_{i}\left(M_{i}-1,N_{i}-1\right)}\right)$ (40) ###### Proof. See Appendix B. ∎ It is worth mentioning that (35) and (Corollary I) are novel and original contributions of this work, derived in closed-form even though (III) is given in integral form. ### IV-B SP–Plus–CCRT Analysis Similar to [14], we assume that each individual pulse on each sweep results in an independent random value for the target returns. Now, using (35) and taking into account the $\mathcal{M}$–of–$L$ detection criterion,555Instead of detecting a target on the basis of at least one detection in $L$ tries, system designers often require that some number $\mathcal{M}$ or more detections be required in $L$ tries before a target detection is accepted [2]. the probability of detection for the combined technique SP–plus–CCR can be calculated as follows [23] $\displaystyle\text{PD}_{\text{CCRT}}\triangleq$ $\displaystyle\sum_{l=\mathcal{M}}^{L}\sum_{\mathcal{E}\in\mathcal{F}_{l}}\left\\{\left(\prod_{i\in\mathcal{E}}\text{PD}_{i}\right)\left(\prod_{j\in\mathcal{E}^{c}}\left(1-\text{PD}_{j}\right)\right)\right\\},$ (41) where $\mathcal{F}_{l}$ is the set of all subsets of $l$ integers that can be selected from $\left\\{1,2,\ldots,L\right\\}$, and $\mathcal{E}^{c}$ is the complement of $\mathcal{E}$. For example, if $l=2$ and $L=3$, then $\mathcal{F}_{2}=\left\\{\left\\{1,2\right\\},\left\\{1,3\right\\},\left\\{2,3\right\\}\right\\}$, and $\mathcal{E}^{c}=\left\\{1,2,\ldots,L\right\\}\backslash\mathcal{E}$. On the other hand, the probability of false alarm for the combined technique SP–plus-CCRT can be calculated as [23] $\displaystyle\text{PFA}_{\text{CCRT}}\triangleq$ $\displaystyle\sum_{l=\mathcal{M}}^{L}\sum_{\mathcal{E}\in\mathcal{F}_{l}}\left\\{\left(\prod_{i\in\mathcal{E}}\text{PFA}_{i}\right)\left(\prod_{j\in\mathcal{E}^{c}}\left(1-\text{PFA}_{j}\right)\right)\right\\}.$ (42) For the case where $\mathcal{M}=L$, (41) and (42) reduce, respectively, to $\displaystyle\text{PD}_{\text{CCRT}}=$ $\displaystyle\prod_{i=1}^{L}\text{PD}_{i}$ (43) $\displaystyle\text{PFA}_{\text{CCRT}}=$ $\displaystyle\prod_{i=1}^{L}\text{PFA}_{i}.$ (44) ## V Numerical Results Figure 2: $\text{PD}_{i}$ vs $\text{SNR}_{1}$ using $N_{i}=8$, $\lambda_{1,i}=0.5$, $\lambda_{2,i}=0.99$, and different values of $M_{i}$ ($i\in\left\\{1,2,3,4\right\\}$). Figure 3: $\text{PMD}_{i}$ vs $\text{SNR}_{1}$ using $N_{i}=8$, $\lambda_{1,i}=0.5$, $\lambda_{2,i}=0.99$, and different values of $M_{i}$ ($i\in\left\\{1,2,3,4\right\\}$). --- (a–1) $\text{PRF}_{1}=1700$ [Hz], $M_{1}=11$ --- (a–2) $\text{PRF}_{2}=1900$ [Hz], $M_{2}=13$ --- (a–3) $\text{PRF}_{3}=2100$ [Hz], $M_{3}=17$ --- (a–4) $\text{PRF}_{4}=2300$ [Hz], $M_{4}=19$ --- (b–1) $\text{PRF}_{1}=1700$ [Hz], $N_{1}=8$ --- (b–2) $\text{PRF}_{2}=1900$ [Hz], $N_{2}=8$ --- (b–3) $\text{PRF}_{3}=2100$ [Hz], $N_{3}=8$ --- (b–4) $\text{PRF}_{4}=2300$ [Hz], $N_{4}=8$ --- (c–1) $\text{PRF}_{1}=1700$ [Hz] --- (c–2) $\text{PRF}_{2}=1900$ [Hz] --- (c–3) $\text{PRF}_{3}=2100$ [Hz] --- (c–4) $\text{PRF}_{4}=2300$ [Hz] Figure 4: Doppler estimation. In this section, we illustrate through Fig. 4 how the Doppler estimation process is carried out. Latter, we validate our derived expressions by means of Monte-Carlo simulations666The number of realizations in Monte-Carlo simulations was set to $10^{6}$.. To do so, we make use of the following radar setup: $\text{PRF}_{1}=700$ [Hz], $\text{PRF}_{2}=1100$ [Hz], $\text{PRF}_{3}=1300$ [Hz], $\text{PRF}_{4}=1700$ [Hz], $L=\mathcal{M}=4$, $\mathit{f}_{R}=6\ [\text{GHz}]$, $\tau=25\ [\mu s]$, $\lambda=0.05\ [\text{m}]$, $M_{1}=11$, $M_{2}=13$, $M_{3}=17$, $M_{4}=19$, and $N_{i}=8\ \forall i\in\left\\{1,2,3,4\right\\}$. In addition, we consider a linear frequency-modulated pulse with bandwidth $B=2\ [\text{MHz}]$. Fig. 4 illustrates the output data after the 2D-DFT blocks. In this simulation example, we placed a target at an initial range of 10 [Km], traveling with a constant velocity of $\mathit{v}_{t}=900$ [m/s] in the opposite direction of the radar (i.e., the target is receding). Fig. 4(a) shows the normalized output data – Velocity vs Range – using PP. Observe that in all 4 scenarios, the target at 10 [Km] is unlikely to be detected due to the high loss in SNR. On the other hand, Fig. 4(b) shows the normalized output data – Velocity vs Range – using SP. Observe that the loss in SNR is partially mitigated by means of SP. Therefore, the target located at 10 [Km] can now be easily be detect without further processing. At last, Fig. 4(c) shows the combined pulse and subpulse information. Note in Fig. 4(c) that SP provides a better intuition about the target location, but due to its poor discretization, it is not sufficient to determine the exact velocity. Conversely, PP provides a better discretization but, unfortunately, its velocity estimation is more likely to be ambiguous. Thus, by combining SP and the CCRT, we provide the system a high capability to unfold the target’s true velocity. Fig. 2 shows $\text{PD}_{i}$ versus $\text{SNR}_{1}$ using different values of $M_{i}$. Note how radar performance improves as $M_{i}$ increases, requiring a lower SNR for a given PD. This is because when increasing $M_{i}$, we are, in fact, increasing the compressed response of PP by means of coherent integration. In particular, for a fixed $\text{SNR}_{1}=10$ [dB], we obtain the following probabilities of detection: $\text{PD}_{1}=0.66$ for $M_{1}=7$; $\text{PD}_{2}=0.78$ for $M_{2}=11$; $\text{PD}_{3}=0.85$ for $M_{3}=13$; and $\text{PD}_{4}=0.93$ for $M_{4}=17$. Also, observe that for the high and medium SNR regime, our derived expression matches perfectly the PD of [14, Eq. (28)]. Nevertheless, there is a small difference in the PD for the low SNR regime. This occurs because if the compressed response of PP is less than the background noise, then the intersection probability in (30) will be less than the probability of $\bigcap_{k=1}^{M_{i}-1}\mathcal{A}_{k,i}$ . For example, given $\text{SNR}_{1}=4$ [dB] and $M_{1}=7$, we obtain $\text{PD}_{1}=0.15$ with our proposed SP–plus–CCRT technique, and $\text{PD}_{1}=0.18$ with [14, Eq. (28)]. However, this small reduction in the PD is compensated by a greater reduction in the PFA, as shall be seen next. Figure 5: $\text{PD}_{\text{CCRT}}$ vs $\text{SNR}_{1}$ using $N_{i}=8$, $\lambda_{1,i}=0.5$, $\lambda_{2,i}=0.99$, $\mathcal{M}=4$, and different values of $M_{i}$ ($i\in\left\\{1,2,3,4\right\\}$). Figure 6: $\text{PMD}_{\text{CCRT}}$ vs $\text{SNR}_{1}$ using $N_{i}=8$, $\lambda_{1,i}=0.5$, $\lambda_{2,i}=0.99$, $\mathcal{M}=4$, and different values of $M_{i}$ ($i\in\left\\{1,2,3,4\right\\}$). Fig. 3 shows $\text{PFA}_{i}$ versus $\text{SNR}_{1}$ using different values for $M_{i}$. Observe how $\text{PFA}_{i}$ decreases as $M_{i}$ increases. This occurs because as we increase $M_{i}$, the received target echo becomes stronger compared to the noise background. For example, for a fixed $\text{SNR}_{1}=5$ [dB], we obtain the following probabilities of false alarm: $\text{PFA}_{1}=0.83$ for $M_{1}=7$; $\text{PFA}_{2}=0.77$ for $M_{2}=11$; $\text{PFA}_{3}=0.73$ for $M_{3}=13$; and $\text{PFA}_{4}=0.60$ for $M_{4}=17$. More interesting, observe how $\text{PFA}_{i}$ decays rapidly compared to [14]. This difference in $\text{PFA}_{i}$ is because intuitively SP acts as a backup detection process. That is, since the compressed response of SP is greater of the PP response (for high-velocity targets), then the probability in (38) is lower than the probability of $\underset{k=1}{\overset{M_{i}-1}{\bigcup}}\mathcal{C}_{k,i}$ . For example, using the classic PP technique [14], we obtain the following probabilities of false alarm: $\text{PFA}_{1}=0.96$ for $M_{1}=7$; $\text{PFA}_{2}=0.97$ for $M_{2}=11$; $\text{PFA}_{3}=0.98$ for $M_{3}=13$; and $\text{PFA}_{4}=0.99$ for $M_{3}=17$. Finally, Figs. 5 and 6 show $\text{PD}_{\text{CCRT}}$ and $\text{PFA}_{\text{CCRT}}$ versus $\text{SNR}_{1}$, respectively. Observe in Fig. 5, the perfect agreement between (41) and [14, Eq. (29)]. Hence, in this case, we have no advantage when using SP–plus–CCRT. On the other hand, observe in Fig. 6, the high difference in the PFA between of (42) and that in [14]. In this case, the use of SP–plus–CCRT improves radar performance by considerably reducing the false alarms. For instance, for given $\text{SNR}_{1}=2$ [dB], we obtain probabilities of $\text{PFA}_{\text{CCRT}}=0.94$ using PP–plus–CCRT, and $\text{PFA}_{\text{CCRT}}=0.54$ using SP–plus–CCRT. ## VI Conclusion In this work, we provided a thorough statistical analysis on Doppler estimation when both SP and the CCRT were employed. To do so, we derived novel and closed-form expressions for the PD and PFA. Moreover, a comparison analysis between our proposed SP–plus–CCRT technique and the classic PP–plus–CCRT was carried out. Numerical results and Monte-Carlo simulations corroborated the validity of our expressions and showed that the PFA when using SP–plus–CCRT technique was greatly reduced compared to [14], thereby enhancing radar detection. ## Appendix A Proof of Proposition I Applying [24, Eq. (5.48)] and using the fact that $X_{k,i}$ and $Y_{l,i}$ are independent RVs, (30) can be rewritten as follows: $\displaystyle\text{PD}_{i}=$ $\displaystyle\int_{0}^{\infty}\int_{0}^{\infty}\left(\prod_{k=1}^{M_{i}-1}\text{Pr}\left[X_{k,i}<r_{1,i}|R_{1,i}=r_{1,i}\right]\right)$ $\displaystyle\times\left(\prod_{l=1}^{N_{i}-1}\text{Pr}\left[Y_{l,i}<r_{2,i}|R_{2,i}=r_{2,i}\right]\right)$ $\displaystyle\times\mathit{f}_{R_{1,i},R_{2,i}}(r_{1,i},r_{2,i})\ \text{d}r_{1,i}\ \text{d}r_{2,i}.$ (45) Now, with the aid of [24, Eq. (4.11)] and taking into account that $X_{k,i}$ and $Y_{l,i}$ are identically distributed RVs, yields $\displaystyle\text{PD}_{i}=$ $\displaystyle\int_{0}^{\infty}\int_{0}^{\infty}\left(\int_{0}^{r_{1,i}}f_{X_{1,i}}(x_{1,i})\ \text{d}x_{1,i}\right)^{M_{i}-1}$ $\displaystyle\times\left(\int_{0}^{r_{2,i}}f_{Y_{1,i}}(y_{1,i})\ \text{d}y_{1,i}\right)^{N_{i}-1}$ $\displaystyle\times\mathit{f}_{R_{1,i},R_{2,i}}(r_{1,i},r_{2,i})\ \text{d}r_{1,i}\ \text{d}r_{2,i}.$ (46) Replacing (22)–(III) in (A), we obtain $\displaystyle\text{PD}_{i}=$ $\displaystyle\int_{0}^{\infty}\int_{0}^{\infty}\underbrace{\left(\int_{0}^{r_{1,i}}\frac{x_{1,i}\exp\left(-\frac{x_{1,i}^{2}}{2\sigma_{1,i}^{2}}\right)}{\sigma_{1,i}}\text{d}x_{1,i}\right)^{M_{i}-1}}_{\triangleq\ \mathcal{I}_{1}}$ $\displaystyle\times\underbrace{\left(\int_{0}^{r_{2,i}}\frac{y_{1,i}\exp\left(-\frac{y_{1,i}^{2}}{2\sigma_{2,i}^{2}}\right)}{\sigma_{2,i}}\text{d}y_{1,i}\right)^{N_{i}-1}}_{\triangleq\ \mathcal{I}_{2}}$ $\displaystyle\times\int_{0}^{\infty}\exp(-\xi_{i}t)\exp\left(-\textbf{m}_{i}\right)I_{0}\left(2\sqrt{\textbf{m}_{i}t}\right)$ $\displaystyle\times\prod_{p=1}^{2}\frac{r_{p,i}}{\Omega_{p,i}^{2}}\exp\left(-\frac{r_{p,i}^{2}}{2\Omega_{p,i}^{2}}\right)$ $\displaystyle\times I_{0}\left(\frac{r_{p,i}\sqrt{t\sigma_{p,i}^{2}\lambda_{p,i}^{2}}}{\Omega_{p,i}^{2}}\right)\text{d}t\ \text{d}r_{1,i}\ \text{d}r_{2,i}.$ (47) In order to solve (A), we must first evaluate $\mathcal{I}_{1}$ and $\mathcal{I}_{2}$. In particular, $\mathcal{I}_{1}$ can be calculated as follows: $\displaystyle\mathcal{I}_{1}$ $\displaystyle\overset{(a)}{=}\left(1-\exp\left(-\frac{r_{1,i}^{2}}{2\sigma_{1,i}^{2}}\right)\right)^{M_{i}-1}$ $\displaystyle\overset{(b)}{=}\sum_{k=0}^{M_{i}-1}\left(\begin{array}[]{c}M_{i}-1\\\ k\\\ \end{array}\right)\left(-\exp\left(-\frac{r_{1,i}^{2}}{2\sigma_{1,i}^{2}}\right)\right)^{M_{i}-1-k},$ (50) where in step (a), we have developed the inner integral; and in step (b), we have used the binomial Theorem [24]. Using a similar approach to that used in (A), $\mathcal{I}_{2}$ can be calculated as $\displaystyle\mathcal{I}_{2}=\sum_{l=0}^{N_{i}-1}\left(\begin{array}[]{c}N_{i}-1\\\ l\\\ \end{array}\right)\left(-\exp\left(-\frac{r_{2,i}^{2}}{2\sigma_{2,i}^{2}}\right)\right)^{N_{i}-1-l}.$ (53) Inserting (A) and (53) in (A), followed by changing the order of integration777The change in the order of integration was performed without loss of generality since (22), (23) and (III) are non-negative real functions [25]. and along with minor manipulations, we obtain (58), displayed at the top of the next page. $\displaystyle\text{PD}_{i}=$ $\displaystyle\sum_{k=0}^{M_{i}-1}\sum_{l=0}^{N_{i}-1}1^{k+l}\left(\begin{array}[]{c}M_{i}-1\\\ k\\\ \end{array}\right)\left(\begin{array}[]{c}N_{i}-1\\\ l\\\ \end{array}\right)\int_{0}^{\infty}\exp(-\xi_{i}t)\exp\left(-\textbf{m}_{i}\right)I_{0}\left(2\sqrt{\textbf{m}_{i}t}\right)$ (58) $\displaystyle\times\underbrace{\int_{0}^{\infty}\left(-\exp\left(-\frac{r_{1,i}^{2}}{2\sigma_{1,i}^{2}}\right)\right)^{M_{i}-1-k}\frac{r_{1,i}}{\Omega_{1,i}^{2}}\exp\left(-\frac{r_{1,i}^{2}}{2\Omega_{1,i}^{2}}\right)I_{0}\left(\frac{r_{1,i}\sqrt{t\sigma_{1,i}^{2}\lambda_{1,i}^{2}}}{\Omega_{1,i}^{2}}\right)\text{d}r_{1,i}}_{\triangleq\ \mathcal{I}_{3}}$ $\displaystyle\times\underbrace{\int_{0}^{\infty}\left(-\exp\left(-\frac{r_{2,i}^{2}}{2\sigma_{2,i}^{2}}\right)\right)^{N_{i}-1-l}\frac{r_{2,i}}{\Omega_{2,i}^{2}}\exp\left(-\frac{r_{2,i}^{2}}{2\Omega_{2,i}^{2}}\right)I_{0}\left(\frac{r_{2,i}\sqrt{t\sigma_{2,i}^{2}\lambda_{2,i}^{2}}}{\Omega_{2,i}^{2}}\right)\text{d}r_{2,i}}_{\triangleq\ \mathcal{I}_{4}}\text{d}t.$ (59) Now, it remains to find $\mathcal{I}_{3}$ and $\mathcal{I}_{4}$. More precisely, $\mathcal{I}_{3}$ can be computed as $\displaystyle\mathcal{I}_{3}\overset{(a)}{=}$ $\displaystyle\int_{0}^{\infty}\left(-\exp\left(-\frac{r_{1,i}^{2}}{2\sigma_{1,i}^{2}}\right)\right)^{M_{i}-1-k}\frac{r_{1,i}}{\Omega_{1,i}^{2}}$ $\displaystyle\times\exp\left(-\frac{r_{1,i}^{2}}{2\Omega_{1,i}^{2}}\right)\sum_{q=0}^{\infty}\frac{\left(\frac{r_{1,i}\sqrt{t\lambda_{1,i}^{2}\sigma_{1,i}^{2}}}{2\Omega_{1,i}^{2}}\right)^{2q}}{q!\ \Gamma(q+1)}\text{d}r_{1,i}$ $\displaystyle\overset{(b)}{=}$ $\displaystyle\frac{(-1)^{-k+M_{i}+1}}{\Omega_{1,i}^{2}\left(\frac{-k+M_{i}-1}{\sigma_{1,i}^{2}}+\frac{1}{\Omega_{1,i}^{2}}\right)}$ $\displaystyle\times\sum_{q=0}^{\infty}\frac{\left(\frac{t\lambda_{1,i}^{2}\sigma_{1,i}^{4}}{2\Omega_{1,i}^{2}\left(\Omega_{1,i}^{2}(-k+M_{i}-1)+\sigma_{1,i}^{2}\right)}\right)^{q}}{q!}$ $\displaystyle\overset{(c)}{=}$ $\displaystyle\frac{(-1)^{-k+M_{i}+1}}{\Omega_{1,i}^{2}\left(\frac{-k+M_{i}-1}{\sigma_{1,i}^{2}}+\frac{1}{\Omega_{1,i}^{2}}\right)}$ $\displaystyle\times\exp\left(\frac{t\lambda_{1,i}^{2}\sigma_{1,i}^{4}}{2\Omega_{1,i}^{2}\left(\Omega_{1,i}^{2}(-k+M_{i}-1)+\sigma_{1,i}^{2}\right)}\right),$ (60) where in step (a), we have used the series representation of the modified Bessel function of the first kind and order zero [26, Eq. (03.02.02.0001.01)]; in step (b), we have solved the integral by first changing the order of integration; finally, in step (c), we have used [26, Eq. (01.03.06.0002.01)] and performed some algebraic manipulations. In like manner as in (A), $\mathcal{I}_{4}$ can be computed as $\displaystyle\mathcal{I}_{4}=$ $\displaystyle\frac{(-1)^{-l+N_{i}+1}}{\Omega_{2,i}^{2}\left(\frac{-l+N_{i}-1}{\sigma_{2,i}^{2}}+\frac{1}{\Omega_{2,i}^{2}}\right)}$ $\displaystyle\times\exp\left(\frac{t\lambda_{2,i}^{2}\sigma_{2,i}^{4}}{2\Omega_{2,i}^{2}\left(\Omega_{2,i}^{2}(-l+N_{i}-1)+\sigma_{2,i}^{2}\right)}\right).$ (61) Now, replacing (A) and (A) in (58), we obtain $\displaystyle\text{PD}_{i}=$ $\displaystyle\sum_{k=0}^{M_{i}-1}\sum_{l=0}^{N_{i}-1}\left(\begin{array}[]{c}M_{i}-1\\\ k\\\ \end{array}\right)\left(\begin{array}[]{c}N_{i}-1\\\ l\\\ \end{array}\right)$ (66) $\displaystyle\times\exp\left(-\textbf{m}_{i}\right)\left(\frac{\sigma_{1,i}^{2}(-1)^{-k+M_{i}+1}}{\Omega_{1,i}^{2}(-k+M_{i}-1)+\sigma_{1,i}^{2}}\right)$ $\displaystyle\times\left(\frac{\sigma_{2,i}^{2}(-1)^{-l+N_{i}+1}}{\Omega_{2,i}^{2}(-l+N_{i}-1)+\sigma_{2,i}^{2}}\right)$ $\displaystyle\times\int_{0}^{\infty}\exp(-\xi_{i}t)I_{0}\left(2\sqrt{\textbf{m}_{i}t}\right)$ $\displaystyle\times\exp\left(\frac{t\lambda_{1,i}^{2}\sigma_{1,i}^{4}}{2\Omega_{1,i}^{2}\left(\Omega_{1,i}^{2}(-k+M_{i}-1)+\sigma_{1,i}^{2}\right)}\right)$ $\displaystyle\times\exp\left(\frac{t\lambda_{2,i}^{2}\sigma_{2,i}^{4}}{2\Omega_{2,i}^{2}\left(\Omega_{2,i}^{2}(-l+N_{i}-1)+\sigma_{2,i}^{2}\right)}\right)\text{d}t.$ (67) Finally, using the following identity [27, Eq. (1.11.2.4)] $\int_{0}^{\infty}\exp(tb)I_{0}(\sqrt{t}a)\ \text{d}t=-\frac{\exp\left(-\frac{a^{2}}{4b}\right)}{b},$ (68) and after performing some minor simplifications, we can express (66) in closed-form as in (35), which completes the proof. ## Appendix B Proof of Corollary I By making use of [24, Coroll. 6], we can express (38) as $\displaystyle\text{PFA}_{i}=\sum_{k=1}^{M_{i}-1}\sum_{l=1}^{N_{i}-1}\text{Pr}\left[\mathcal{C}_{k,i}\bigcap\mathcal{D}_{l,i}\right]$ $\displaystyle\ -\underset{k<p,l<q}{\sum_{k=1}^{M_{i}-1}\sum_{l=1}^{N_{i}-1}\sum_{p=2}^{M_{i}-1}\sum_{q=2}^{N_{i}-1}}\text{Pr}\left[\mathcal{C}_{k,i}\bigcap\mathcal{D}_{l,i}\bigcap\mathcal{C}_{p,i}\bigcap\mathcal{D}_{q,i}\right]+\ldots$ $\displaystyle\ +(-1)^{M_{i}-N_{i}-1}\text{Pr}\left[\mathcal{C}_{1,i}\bigcap\mathcal{D}_{1,i}\bigcap\ldots\bigcap\mathcal{C}_{M_{i}-1,i}\bigcap\mathcal{D}_{N_{i}-1,i}\right].$ (69) Now, we need to find the event probabilities. First, let us derive the last event probability of (B), that is, Pr $\displaystyle\left[\mathcal{C}_{1,i}\bigcap\mathcal{D}_{1,i}\bigcap\ldots\bigcap\mathcal{C}_{M_{i}-1,i}\bigcap\mathcal{D}_{N_{i}-1,i}\right]$ $\displaystyle\overset{a}{=}$ $\displaystyle\int_{0}^{\infty}\int_{0}^{\infty}\left(\prod_{k=1}^{M_{i}-1}\text{Pr}\left[X_{k,i}>r_{1,i}|R_{1,i}=r_{1,i}\right]\right)$ $\displaystyle\times\left(\prod_{l=1}^{N_{i}-1}\text{Pr}\left[Y_{l,i}>r_{2,i}|R_{2,i}=r_{2,i}\right]\right)$ $\displaystyle\times\mathit{f}_{R_{1,i},R_{2,i}}(r_{1,i},r_{2,i})\ \text{d}r_{1,i}\ \text{d}r_{2,i}$ $\displaystyle\overset{b}{=}$ $\displaystyle\int_{0}^{\infty}\int_{0}^{\infty}\left(\int_{r_{1,i}}^{\infty}f_{X_{1,i}}(x_{1,i})\ \text{d}x_{1,i}\right)^{M_{i}-1}$ $\displaystyle\times\left(\int_{r_{2,i}}^{\infty}f_{Y_{1,i}}(y_{1,i})\ \text{d}y_{1,i}\right)^{N_{i}-1}$ $\displaystyle\times\mathit{f}_{R_{1,i},R_{2,i}}(r_{1,i},r_{2,i})\ \text{d}r_{1,i}\ \text{d}r_{2,i},$ (70) where in step (a) we have used [24, Eq. (5.48)]; and in step (b) we have used [24, Eq. (4.11)] along with the fact that $X_{k,i}$ and $Y_{l,i}$ are identically distributed RVs. Replacing (22)–(III) in (B), yields Pr $\displaystyle\left[\mathcal{C}_{1,i}\bigcap\mathcal{D}_{1,i}\bigcap\ldots\bigcap\mathcal{C}_{M_{i}-1,i}\bigcap\mathcal{D}_{N_{i}-1,i}\right]$ $\displaystyle=$ $\displaystyle\int_{0}^{\infty}\int_{0}^{\infty}\underbrace{\left(\int_{r_{1,i}}^{\infty}\frac{x_{1,i}\exp\left(-\frac{x_{1,i}^{2}}{2\sigma_{1,i}^{2}}\right)}{\sigma_{1,i}}\text{d}x_{1,i}\right)^{M_{i}-1}}_{\triangleq\ \mathcal{I}_{5}}$ $\displaystyle\times\underbrace{\left(\int_{r_{2,i}}^{\infty}\frac{y_{1,i}\exp\left(-\frac{y_{1,i}^{2}}{2\sigma_{2,i}^{2}}\right)}{\sigma_{2,i}}\text{d}y_{1,i}\right)^{N_{i}-1}}_{\triangleq\ \mathcal{I}_{6}}$ $\displaystyle\times\int_{0}^{\infty}\exp(-\xi_{i}t)\exp\left(-\textbf{m}_{i}\right)I_{0}\left(2\sqrt{\textbf{m}_{i}t}\right)$ $\displaystyle\times\prod_{p=1}^{2}\frac{r_{p,i}}{\Omega_{p,i}^{2}}\exp\left(-\frac{r_{p,i}^{2}}{2\Omega_{p,i}^{2}}\right)$ $\displaystyle\times I_{0}\left(\frac{r_{p,i}\sqrt{t\sigma_{p,i}^{2}\lambda_{p,i}^{2}}}{\Omega_{p,i}^{2}}\right)\text{d}t\ \text{d}r_{1,i}\ \text{d}r_{2,i}.$ (71) After some mathematical manipulations, $\mathcal{I}_{5}$ and $\mathcal{I}_{6}$ can be calculated, respectively, as $\displaystyle\mathcal{I}_{5}=$ $\displaystyle\exp\left(-\frac{r_{1,i}^{2}(M_{i}-1)}{2\sigma_{1,i}^{2}}\right)$ (72) $\displaystyle\mathcal{I}_{6}=$ $\displaystyle\exp\left(-\frac{r_{2,i}^{2}(N_{i}-1)}{2\sigma_{2,i}^{2}}\right).$ (73) $\displaystyle\text{PFA}_{i}=$ $\displaystyle\binom{M_{i}-1}{1}\binom{N_{i}-1}{1}\frac{\mathcal{Q}_{i}\left(1,1\right)}{\mathcal{P}_{i}\left(1,1\right)}\exp\left(-\textbf{m}_{i}+\frac{\textbf{m}_{i}}{\mathcal{P}_{i}\left(1,1\right)}\right)-\binom{M_{i}-1}{2}\binom{N_{i}-1}{2}\frac{\mathcal{Q}_{i}\left(2,2\right)}{\mathcal{P}_{i}\left(2,2\right)}\exp\left(-\textbf{m}_{i}+\frac{\textbf{m}_{i}}{\mathcal{P}_{i}\left(2,2\right)}\right)+\ldots$ $\displaystyle+(-1)^{M_{i}-N_{i}-1}\binom{M_{i}-1}{M_{i}-1}\binom{N_{i}-1}{N_{i}-1}\frac{\mathcal{Q}_{i}\left(M_{i}-1,N_{i}-1\right)}{\mathcal{P}_{i}\left(M_{i}-1,N_{i}-1\right)}\exp\left(-\textbf{m}_{i}+\frac{\textbf{m}_{i}}{\mathcal{P}_{i}\left(M_{i}-1,N_{i}-1\right)}\right)$ (74) Now, replacing (72) and (73) in (B), and after solving remaining three integrals by applying the same procedure as in (66), we obtain Pr $\displaystyle\left[\mathcal{C}_{1,i}\bigcap\mathcal{D}_{1,i}\bigcap\ldots\bigcap\mathcal{C}_{M_{i}-1,i}\bigcap\mathcal{D}_{N_{i}-1,i}\right]$ $\displaystyle=\frac{\mathcal{Q}_{i}\left(M_{i}-1,N_{i}-1\right)}{\mathcal{P}_{i}\left(M_{i}-1,N_{i}-1\right)}\exp\left(-\textbf{m}_{i}+\frac{\textbf{m}_{i}}{\mathcal{P}_{i}\left(M_{i}-1,N_{i}-1\right)}\right),$ (75) where $\mathcal{P}_{i}\left(k,l\right)$ and $\mathcal{Q}_{i}\left(k,l\right)$ are auxiliary functions defined in (39), and the parameters $k\in\left\\{1,2,\ldots,M_{i}-1\right\\}$ and $l\in\left\\{1,2,\ldots,N_{i}-1\right\\}$ denote the number of events for $\mathcal{C}_{k,i}$ and $\mathcal{D}_{l,i}$, respectively. Thus, the remaining event probabilities in (B) can be easily obtained by a proper choice of the parameters $k$ and $l$. For example, for $k=1$ and $l=3$, we obtain Pr $\displaystyle\left[\mathcal{C}_{1,i}\bigcap\mathcal{D}_{1,i}\bigcap\mathcal{D}_{2,i}\bigcap\mathcal{D}_{3,i}\right]$ $\displaystyle=\frac{\mathcal{Q}_{i}\left(1,3\right)}{\mathcal{P}_{i}\left(1,3\right)}\exp\left(-\textbf{m}_{i}+\frac{\textbf{m}_{i}}{\mathcal{P}_{i}\left(1,3\right)}\right),$ (76) whereas for $k=3$ and $l=2$, we have Pr $\displaystyle\left[\mathcal{C}_{1,i}\bigcap\mathcal{D}_{1,i}\bigcap\mathcal{C}_{2,i}\bigcap\mathcal{D}_{2,i}\bigcap\mathcal{C}_{3,i}\right]$ $\displaystyle=\frac{\mathcal{Q}_{i}\left(3,2\right)}{\mathcal{P}_{i}\left(3,2\right)}\exp\left(-\textbf{m}_{i}+\frac{\textbf{m}_{i}}{\mathcal{P}_{i}\left(3,2\right)}\right).$ (77) Later, with the aid of (B) and after some algebraic manipulations, we can rewrite (B) as in (B), displayed at the top of the next page. Finally, and after minor simplifications, (B) reduces to (Corollary I), which completes the proof. ## References * [1] G. Morris and L. Harkness, _Airborne Pulsed Doppler Radar_ , 2nd ed. Norwood, MA, USA: Artech House, 1996. * [2] M. A. Richards, J. Scheer, W. A. Holm, and W. L. Melvin, _Principles of Modern Radar: Basic Principles_ , 1st ed. West Perth, WA, Australia: SciTech, 2010. * [3] G. V. Trunk, “Range resolution of targets using automatic detectors,” _IEEE Trans. Aerosp. Electron. Syst._ , vol. AES-14, no. 5, pp. 750–755, Sept. 1978. * [4] S. A. Hovanessian, “An algorithm for calculation of range in a multiple PRF radar,” _IEEE Trans. Aerosp. Electron. Syst._ , vol. AES-12, no. 2, pp. 287–290, Mar. 1976. * [5] X.-G. Xia and G. Wang, “Phase unwrapping and a robust chinese remainder theorem,” _IEEE Signal Process. Lett._ , vol. 14, no. 4, pp. 247–250, Apr. 2007. * [6] X. Li, H. Liang, and X. Xia, “A robust chinese remainder theorem with its applications in frequency estimation from undersampled waveforms,” _IEEE Trans. Signal Process._ , vol. 57, no. 11, pp. 4314–4322, Nov. 2009\. * [7] W. Wang and X. Xia, “A closed-form robust chinese remainder theorem and its performance analysis,” _IEEE Trans. Signal Process._ , vol. 58, no. 11, pp. 5655–5666, Nov. 2010. * [8] G. V. Trunk and W. M. Kim, “Ambiguity resolution of multiple targets using pulse-Doppler waveforms,” _IEEE Trans. Aerosp. Electron. Syst._ , vol. 30, no. 4, pp. 1130–1137, Oct. 1994. * [9] F. D. A. García, A. S. Guerreiro, G. R. L. Tejerina, J. C. S. Santos Filho, G. Fraidenraich, M. D. Yacoub, M. A. M. Miranda, and H. Cioqueta, “Probability of detection for unambiguous doppler frequencies in pulsed radars using the chinese remainder theorem and subpulse processing,” in _Proc. 53rd Asilomar Conference on Signals, Systems, and Computers_ , Pacific Grove, CA, USA, Nov. 2019, pp. 138–142. * [10] M. I. Skolnik, _Introduction to Radar Systems_ , 3rd ed. Ney York, NY, USA: McGraw-Hill, 2001. * [11] G. Beltrao, L. Pralon, M. Menezes, P. Vyplavin, B. Pompeo, and M. Pralon, “Subpulse processing for long range surveillance noise radars,” in _Proc. International Conference on Radar Systems (Radar 2017)_ , Belfast, UK, Oct. 2017, pp. 1–4. * [12] A. Barreto, L. Pralon, B. Pompeo, G. Beltrao, and M. Pralon, “FPGA design and implementation of a real-time subpulse processing architecture for noise radars,” in _Proc. 2019 International Radar Conference (RADAR)_ , Toulon, France, Sept. 2019, pp. 1–6. * [13] D. S. Doviak and R. J. Zrnic, _Doppler Radar and Weather Observations_ , 2nd ed. San Diego, CA, USA: Academic Press, 2001. * [14] B. Silva and G. Fraidenraich, “Performance analysis of the classic and robust chinese remainder theorems in pulsed doppler radars,” _IEEE Trans. Signal Process._ , vol. 66, no. 18, pp. 4898–4903, Sept. 2018. * [15] M. A. Richards, _Fundamentals of Radar Signal Processing_ , 2nd ed. Ney York, NY, USA: McGraw-Hill, 2014. * [16] D. K. Barton, _Radar Equations for Modern Radar_ , 1st ed. Massachusetts, MA, USA: Artech House, 2013. * [17] G. Trunk and S. Brockett, “Range and velocity ambiguity resolution,” in _Proc. Record IEEE Nat. Radar Conf._ , Lynnfield, MA, USA, Apr. 1993, pp. 146–149. * [18] A. Ferrari, C. Berenguer, and G. Alengrin, “Doppler ambiguity resolution using multiple PRF,” _IEEE Trans. Aerosp. Electron. Syst._ , vol. 33, no. 3, pp. 738–751, Jul. 1997. * [19] A. Papoulis, _Probability, Random Variables, and Stochastic Processes_ , 4th ed. Ney York, NY, USA: McGraw-Hill, 2002. * [20] N. C. Beaulieu and K. T. Hemachandra, “Novel representations for the bivariate rician distribution,” _IEEE Trans. Commun._ , vol. 59, no. 11, pp. 2951–2954, Nov. 2011. * [21] A. Behnad, N. C. Beaulieu, and K. T. Hemachandra, “Correction to “Novel representations for the bivariate rician distribution”,” _IEEE Trans. Commun._ , vol. 60, no. 6, pp. 1486–1486, Jun. 2012. * [22] M. Abramowitz and I. A. Stegun, _Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables_. Washington, DC: US Dept. of Commerce: National Bureau of Standards, 1972. * [23] Y. H. Wang, “On the number of successes in independent trials,” _Statistica Sinica_ , vol. 3, no. 2, pp. 295–312, 1993. * [24] A. Leon-Garcia, _Probability and Random Processes for Electrical Engineering_ , 3rd ed. New Jersey, NJ, USA: Pearson Prentice Hall, 1994. * [25] H. Friedman, “A consistent Fubini-Tonelli theorem for nonmeasurable functions,” _Illinois J. Math._ , vol. 24, no. 3, pp. 390–395, 1980. * [26] Wolfram Research, Inc. (2018), _Wolfram Research_ , Accessed: Sept. 19, 2018\. [Online]. Available: http://functions.wolfram.com * [27] A. P. Prudnikov, Y. A. Bryčkov, and O. I. Maričev, _Integral and Series: Vol. 2_ , 2nd ed., Fizmatlit, Ed. Moscow, Russia: Fizmatlit, 1992.
11institutetext: University of Queensland, Brisbane QLD 4072, Australia 11email<EMAIL_ADDRESS>11email<EMAIL_ADDRESS>11email: <EMAIL_ADDRESS> # Towards a Standard Feature Set for Network Intrusion Detection System Datasets Mohanad Sarhan 11 Siamak Layeghy 11 Marius Portmann 11 ###### Abstract Network Intrusion Detection Systems (NIDSs) are important tools for the protection of computer networks against increasingly frequent and sophisticated cyber attacks. Recently, a lot of research effort has been dedicated to the development of Machine Learning (ML) based NIDSs. As in any ML-based application, the availability of high-quality datasets is critical for the training and evaluation of ML-based NIDS. One of the key problems with the currently available NIDS datasets is the lack of a standard feature set. The use of a unique and proprietary set of features for each of the publicly available datasets makes it virtually impossible to compare the performance of ML-based traffic classifiers on different datasets, and hence to evaluate the ability of these systems to generalise across different network scenarios. To address that limitation, this paper proposes and evaluates standard NIDS feature sets based on the NetFlow network meta-data collection protocol and system. We evaluate and compare two NetFlow-based feature set variants, a version with 12 features, and another one with 43 features. For our evaluation, we converted four widely used NIDS datasets (UNSW-NB15, BoT-IoT, ToN-IoT, CSE-CIC-IDS2018) into new variants with our proposed NetFlow based feature sets. Based on an Extra Tree classifier, we compared the classification performance of the NetFlow-based feature sets with the proprietary feature sets provided with the original datasets. While the smaller feature set cannot match the classification performance of the proprietary feature sets, the larger set with 43 NetFlow features, surprisingly achieves a consistently higher classification performance compared to the original feature set, which was tailored to each of the considered NIDS datasets. The proposed NetFlow-based standard NIDS feature set, together with four benchmark datasets, made available to the research community, allow a fair comparison of ML-based network traffic classifiers across different NIDS datasets. We believe that having a standard feature set is critical for allowing a more rigorous and thorough evaluation of ML-based NIDSs and that it can help bridge the gap between academic research and the practical deployment of such systems. ###### keywords: Machine Learning, NetFlow, Network Intrusion Detection System ## 1 Introduction Network Intrusion Detection Systems (NIDSs) aim to detect network attacks and to preserve the three principles of information security: confidentiality, integrity, and availability [1]. Signature-based NIDSs match attack signatures to observed traffic, giving a high detection accuracy to known attacks. However, these systems are unable to detect previously unseen (zero-day) attacks or new variants of known attacks. Therefore, researchers have investigated anomaly-based NIDSs that focus on matching attack behaviours and patterns [2]. Machine Learning (ML), a sub-field of artificial intelligence, is capable of learning and extracting complex network attack patterns that may threaten computer networks if undetected [3]. All network intrusions generate a unique set of security events, that would aid in their classification process. These identifying patterns can be extracted from network traffic in the form of data features. To generate a dataset, corresponding data features form network data flows that are ideally labelled with an attack or a benign class to allow for a supervised ML methodology. Real-world network flow datasets with labels that identify attack and benign flows are challenging to obtain, mainly due to security and privacy concerns. Therefore, researchers have designed network test-beds to generate synthetic datasets that consist of labelled network data flows [4]. The data flows are made of several network features that are often preselected based on the authors’ domain knowledge and available extraction tools. As a result, the currently available NIDS datasets are very distinct in terms of their feature sets and therefore the security events represented by the data flows. Due to the great impact of data features on the performance of ML models [5], the evaluation of the proposed ML-based NIDSs are often unreliable when tested on multiple datasets using their original feature sets. Finally, as certain network data features require a complex and deep packet inspection, the computational complexity of feature extraction and processing is not feasible The importance of having a standard feature set for all datasets is paramount. It will facilitate a fair and reliable evaluation of proposed ML models across various network environments and attack scenarios. This also enables an evaluation of the generalisability of the model, and hence its performance when deployed in practical network scenarios. Moreover, a standard feature set will ensure that the security events and network information presented by NIDS datasets are the same and in a controlled manner. NetFlow is an industry- standard protocol for network traffic collection [6].Its practical and scalable deployment properties are capable of enhancing the deployment feasibility of ML-based NIDSs. NetFlow features are capable of presenting key security events that are crucial in the identification of network attacks. Therefore, we believe that applying NetFlow-based features in the design of a universal feature set will facilitate the successful deployment of ML-based NIDS in practical network scenarios. Four widely used NIDS datasets, referred to as UNSW-NB15 [7], BoT-IoT [8], ToN-IoT [9], and CIC-CSE-IDS2018 [10] have been converted into a common basic NetFlow-based feature set [11]. The NetFlow datasets address some of the current research issues by applying a common feature set across multiple datasets. However, due to the insufficient security information represented by the basic NetFlow feature set, the ML models lead to limited detection accuracy, in particular when performing multi-class experiments. Therefore, this paper proposes an extended NetFlow feature set as the standard version to be used in future NIDS datasets. As part of its evaluation, the features have been extracted and labelled from four well-known datasets. The datasets generated are named NF-UNSW-NB15-v2, NF-BoT-IoT-v2, NF-ToN-IoT-v2, NF-CSE-CIC- IDS2018-v2 and NF-UQ-NIDS-v2, and have been made publicly available for research purposes [12]. This paper explores two variants of NetFlow-based feature sets along with their proprietary feature sets. The rest of the paper is organised as follows. Existing NIDS datasets and their limitations are discussed in Section 2. Section 3 motivates the case for having a standard and a common feature set in NIDS datasets. It also illustrates our methodology of extracting the proposed feature. Finally, in Section 4, we use an Extra Tree classifier to compare the predictive power of our proposed NetFlow based feature set, with the proprietary features sets provided with the original benchmark NIDS datasets. Finally, Section 5 concludes the paper. ## 2 Limitations of Existing Datasets Researchers have created engineered benchmark NIDS datasets because of the complexity in obtaining labelled realistic network traffic. A network testbed is designed to simulate the network behaviour of multiple end nodes. The artificial network environment overcomes the security and privacy issues faced by real-world networks. Besides, labelling the network flows generated by such controlled environments is more reliable than the open-world nature of realistic networks. During the experiments, benign network traffic and various attack scenarios are generated and conducted over the network testbed. In the meanwhile, the network packets are captured in their native packet capture (pcap) format and dumped onto storage devices. A set of network data features are extracted from the pcap files using appropriate tools and methods, forming network data flows. The result is a data source of labelled network flows reflecting benign and malicious network behaviour. The generated datasets are published and made publicly accessible for use in the design and evaluation phases of ML-based NIDS models [13]. The network data features that form these data flows are critical as they need to represent an adequate amount of security events that would aid in the ML model’s classification of benign and attack classes. They also need to be feasible in count and extraction’s complexity for scalable and practical deployments. A key task of designing an ML-based NIDS is the selection of the utilised data features. However, due to the lack of a standard feature set in generating NIDS datasets, the authors have applied their domain experience in the selection of these features. As a result, each available dataset is made up of its own unique set of features that their authors believe would lead to the best possible results in the classification stage. Each of the current feature sets is almost exclusive and completely different from other sets, sharing only a small number of features. The current evaluation method of ML models across multiple datasets requires the usage of the unique feature sets presented by each dataset. The differences in the security information represented by each dataset’s feature set have caused limitations and concerns regarding the reliability of the evaluation methods followed. The three main issues of not having a standard feature set are; 1. Complex extraction of several features from network traffic, some of which are irrelevant due to the lack of security events and 2. Limited ability to evaluate an ML model’s generalisation to a targeted feature set across multiple datasets and 3. Lack of a universal dataset containing network data flows collected over multiple network environments. It is believed that the lack of reliable evaluation methods has caused a gap between the extensive academic research produced and the practical deployment of ML-based NIDS models in production networks [14]. Four of the most recent and widely-used NIDS datasets are discussed, which represent modern behavioural network attacks due to their production time. * $\bullet$ UNSW-NB15 The Cyber Range Lab of the Australian Centre for Cyber Security (ACCS) released the widely used, UNSW-NB15, dataset in 2015. The IXIA PerfectStorm tool was utilised to generate a hybrid of testbed-based benign network activities as well as synthetic attack scenarios. The tcpdump tool was implemented to capture a total of 100 GB of pcap files. Argus and Bro-IDS, now called Zeek [15], and twelve additional SQL algorithms were used to extract the dataset’s original 49 features [7]. The dataset contains 2,218,761 (87.35%) benign flows and 321,283 (12.65%) attack ones, that is, 2,540,044 flows in total. * $\bullet$ BoT-IoT The Cyber Range Lab of the Australian Centre for Cyber Security (ACCS) designed a network environment in 2018 that consists of normal and botnet traffic [8]. The Ostinato and Node-red tools were utilised to generate the non-IoT and IoT traffic respectively. A total of 69.3GB of pcap files were captured and the Argus tool was used to extract the dataset’s original 42 features. The dataset contains 477 (0.01%) benign flows and 3,668,045 (99.99%) attack ones, that is, 3,668,522 flows in total. * $\bullet$ ToN-IoT A recent heterogeneous dataset released in 2019 [9] that includes telemetry data of Internet of Things (IoT) services, network traffic of IoT networks, and operating system logs. In this paper, the portion containing network traffic flows is utilised. The dataset is made up of a large number of attack scenarios conducted in a representation of a realistic large-scale network at the Cyber Range Lab by ACCS. Bro-IDS, now called Zeek [15], was used to extract the dataset’s original 44 features. The dataset is made up of 796,380 (3.56%) benign flows and 21,542,641 (96.44%) attack samples, that is, 22,339,021 flows in total. * $\bullet$ CSE-CIC-IDS2018 A dataset released by a collaborative project between the Communications Security Establishment (CSE) & Canadian Institute for Cybersecurity (CIC) in 2018 [10]. The victim network consisted of five different organisational departments and an additional server room. The benign packets were generated by network events using the abstract behaviour of human users. The attack scenarios were executed by one or more machines outside the target network. The CICFlowMeter-V3 tool was used to extract the original dataset’s 75 features. The full dataset has 13,484,708 (83.07%) benign flows and 2,748,235 (16.93%) attack flows, that is, 16,232,943 flows in total. Figure 1: Venn diagram of the shared and exclusive features of four NIDS datasets In Figure 1, the shared and unique features of the aforementioned datasets are displayed. The set of features available in all four datasets contains 3 features, and the pairwise shared feature numbers vary from 1 to 5. As most of the features are exclusive to individual datasets, the evaluation of proposed ML models using a targeted feature set across the four datasets is challenging. Moreover, the ratio of the classes, i.e., benign and attack flows, is extremely varied in each dataset. Where the UNSW-NB15 and CSE-CIC- IDS2018 datasets have very high benign-to-attack ratios, whereas the ToN-IoT and BoT-IoT datasets are mainly made up of attack samples, which do not represent a realistic network behaviour. Also, some of the features in the UNSW-NB15, BoT-IoT, and CSE-CIC-IDS2018 datasets are handcrafted features that are not originally found in network packets but are statistically calculated based on other features, such as the total number of bytes transferred over the last 100 seconds. All these differences in the security information presented by the datasets have led to the design of a standard feature set for NIDS datasets. ## 3 Benchmarking a Standard Feature Set Due to the aforementioned limitations faced by current NIDS datasets made up of unique feature sets, in this paper, a standard feature set is proposed. The feature set will be evaluated and benchmarked to be used in the releases of new NIDS datasets to efficiently design ML-based NIDS. The design of ML-based NIDS requires a feature set to be extracted and scanned for intrusions when implemented. The choice of these features significantly alters the performance of the NIDS as they need to contain an adequate amount of security events to aid the ML model classification. By having a standard feature set, researchers can evaluate their model’s classification ability based on their chosen features, across multiple datasets and hence different attack scenarios conducted over several network environments. This can be used to make sure their measured model performance generalises when deployed over different networks. Moreover, by having datasets sharing a common ground feature set, they can be merged to create a universal comprehensive source of data. Finally, having a standard feature set will grant control over the security information presented by NIDS datasets. We believe that a standard feature set will narrow the gap between the number of research experiments and the practical deployment of ML-based NIDS [14]. ### 3.1 NetFlow The collection and storage of network traffic are important for organisations to monitor, analyse, and audit their network environments. However, network traffic tends to overload in volume and therefore are aggregated in terms of flows. A network data flow is a sequence of packets, in either uni- or bi- direction, between two unique endpoints sharing some attributes such as source/destination IP address and L4 (transport layer) ports, and the L4 protocol, also known as the five-tuple [11]. A data flow can also be enhanced with additional features, each representing details of the respective network traffic. The information provided by these features contains security events that are essential in analysing network traffic in case of a threat [16]. Network flows can be represented in various formats where the NetFlow is the de-facto industry standard, developed and proposed by Darren and Barry Bruins from Cisco in 1996 [17]. NetFlow evolved over the years, where version 9 is the most common due to its larger variety of data features and bidirectional flow support [18]. Table 1: List of the proposed standard NetFlow features Feature | Description ---|--- IPV4_SRC_ADDR | IPv4 source address IPV4_DST_ADDR | IPv4 destination address L4_SRC_PORT | IPv4 source port number L4_DST_PORT | IPv4 destination port number PROTOCOL | IP protocol identifier byte L7_PROTO | Layer 7 protocol (numeric) IN_BYTES | Incoming number of bytes OUT_BYTES | Outgoing number of bytes IN_PKTS | Incoming number of packets OUT_PKTS | Outgoing number of packets FLOW_DURATION_MILLISECONDS | Flow duration in milliseconds TCP_FLAGS | Cumulative of all TCP flags CLIENT_TCP_FLAGS | Cumulative of all client TCP flags SERVER_TCP_FLAGS | Cumulative of all server TCP flags DURATION_IN | Client to Server stream duration (msec) DURATION_OUT | Client to Server stream duration (msec) MIN_TTL | Min flow TTL MAX_TTL | Max flow TTL LONGEST_FLOW_PKT | Longest packet (bytes) of the flow SHORTEST_FLOW_PKT | Shortest packet (bytes) of the flow MIN_IP_PKT_LEN | Len of the smallest flow IP packet observed MAX_IP_PKT_LEN | Len of the largest flow IP packet observed SRC_TO_DST_SECOND_BYTES | Src to dst Bytes/sec DST_TO_SRC_SECOND_BYTES | Dst to src Bytes/sec RETRANSMITTED_IN_BYTES | Number of retransmitted TCP flow bytes (src-$>$dst) RETRANSMITTED_IN_PKTS | Number of retransmitted TCP flow packets (src-$>$dst) RETRANSMITTED_OUT_BYTES | Number of retransmitted TCP flow bytes (dst-$>$src) RETRANSMITTED_OUT_PKTS | Number of retransmitted TCP flow packets (dst-$>$src) SRC_TO_DST_AVG_THROUGHPUT | Src to dst average thpt (bps) DST_TO_SRC_AVG_THROUGHPUT | Dst to src average thpt (bps) NUM_PKTS_UP_TO_128_BYTES | Packets whose IP size $<$= 128 NUM_PKTS_128_TO_256_BYTES | Packets whose IP size $>$ 128 and $<$= 256 NUM_PKTS_256_TO_512_BYTES | Packets whose IP size $>$ 256 and $<$= 512 NUM_PKTS_512_TO_1024_BYTES | Packets whose IP size $>$ 512 and $<$= 1024 NUM_PKTS_1024_TO_1514_BYTES | Packets whose IP size $>$ 1024 and $<$= 1514 TCP_WIN_MAX_IN | Max TCP Window (src-$>$dst) TCP_WIN_MAX_OUT | Max TCP Window (dst-$>$src) ICMP_TYPE | ICMP Type * 256 + ICMP code ICMP_IPV4_TYPE | ICMP Type DNS_QUERY_ID | DNS query transaction Id DNS_QUERY_TYPE | DNS query type (e.g., 1=A, 2=NS..) DNS_TTL_ANSWER | TTL of the first A record (if any) FTP_COMMAND_RET_CODE | FTP client command return code Most of the network devices such as routers and switches are capable of extracting NetFlow records. This is a great motivation for standardising NetFlow features for NIDS datasets, as the level of complexity and resources required to collect and store them is lower. In this paper, NetFlow v9 features have been utilised to form the proposed feature set, listed and described in Table 1. There are 43 features in total with some providing information on general flow statistics and others on specific protocol applications such as DNS and FTP. All features are flow-based, meaning they are extracted from packet headers and do not depend on the payload information which is often encrypted in secure communications due to privacy concerns. The chosen features are numerical in type for efficient ML experiments. These features contain useful security events to enhance the models’ intrusions detection capabilities. ### 3.2 Datasets Figure 2 shows the procedure of generating NIDS datasets using the proposed feature set. The nProbe tool by Ntop [19] is utilised to extract 43 NetFlow version 9 features from the publicly available pcap files. The output format is chosen as text flows, in which each feature is separated by a comma (,) to be utilised as CSV files. Two label features are created by matching the five flow identifiers; source/destination IPs and ports and protocol to the ground truth attack events published by the original dataset. If a data flow is located in the attack events it would be labelled as an attack (class 1) in the binary label and its respective attack’s type would be recorded in the attack label, otherwise, the sample is labelled as a benign flow (class 0). Figure 2: Feature set extraction and labelling procedure Table 2: Specifications of the datasets proposed in this paper, compared to the original and basic NetFlow datasets Dataset | | Release --- year Feature extraction tool | | Number --- of features | CSV size --- (GB) | Benign to attack --- samples ratio UNSW-NB15 | 2015 | Argus, Bro-IDS and MS SQL | 49 | 0.55 | 8.7 to 1.3 NF-UNSW-NB15 | 2020 | nProbe | 12 | 0.11 | 9.6 to 0.4 NF-UNSW-NB15-v2 | 2021 | nProbe | 43 | 0.41 | 9.6 to 0.4 BoT-IoT | 2018 | Argus | 42 | 0.95 | 0.0 to 10 NF-BoT-IoT | 2020 | nProbe | 12 | 0.05 | 0.2 to 9.8 NF-BoT-IoT-v2 | 2021 | nProbe | 43 | 5.60 | 0.0 to 10.0 ToN-IoT | 2020 | Bro-IDS | 44 | 3.02 | 0.4 to 9.6 NF-ToN-IoT | 2020 | nProbe | 12 | 0.09 | 2.0 to 8.0 NF-ToN-IoT-v2 | 2021 | nProbe | 43 | 2.47 | 3.6 to 6.4 CSE-CIC-IDS2018 | 2018 | CICFlowMeter-V3 | 75 | 6.41 | 8.3 to 1.7 NF-CSE-CIC-IDS2018 | 2020 | nProbe | 12 | 0.58 | 8.8 to 1.2 NF-CSE-CIC-IDS2018-v2 | 2021 | nProbe | 43 | 2.80 | 8.8 to 1.2 NF-UQ-NIDS | 2020 | nProbe | 12 | 1.0 | 7.7 to 2.3 NF-UQ-NIDS-v2 | 2021 | nProbe | 43 | 12.5 | 3.3 to 6.7 In this paper, the proposed feature set has been extracted from four well- known datasets; UNSW-NB15, BoT-IoT, ToN-IoT, and CSE-CIC-IDS2018. Their publicly available pcap files and ground truth events have been utilised in the features extraction and labelling processes respectively. The generated datasets have been named NF-UNSW-NB15-v2, NF-BoT-IoT-v2, NF-ToN-IoT-v2, NF- CSE-CIC-IDS2018-v2 and NF-UQ-NIDS-v2. The later dataset is a merge of all other datasets, which is a practical advantage of having a common feature set. Table 2 lists the NetFlow datasets and compares their properties to the original datasets in terms of the Feature Extraction (FE) tool utilised, the number of features, file size and the benign to attack samples ratio. As illustrated, two NetFlow datasets are corresponding to each original NIDS dataset, where v1 and v2 are the basic and extended versions respectively. The fifth NetFlow dataset is a comprehensive dataset that combines all four. * $\bullet$ NF-UNSW-NB15-v2 The NetFlow-based format of the UNSW-NB15 dataset, named NF- UNSW-NB15, has been extended with additional NetFlow features and labelled with its respective attack categories. The total number of data flows are 2,390,275 out of which 95,053 (3.98%) are attack samples and 2,295,222 (96.02%) are benign. The attack samples are further classified into nine subcategories, Table 3 represents the NF-UNSW-NB15-v2 dataset’s distribution of all flows. Table 3: NF-UNSW-NB15-v2 distribution Class | Count | Description ---|---|--- Benign | 2295222 | Normal unmalicious flows Fuzzers | 22310 | | An attack in which the attacker sends large amounts of random data which cause a system --- to crash and also aim to discover security vulnerabilities in a system. Analysis | 2299 | | A group that presents a variety of threats that target web applications through ports, --- emails and scripts. Backdoor | 2169 | | A technique that aims to bypass security mechanisms by replying to specific constructed --- client applications. DoS | 5794 | | Denial of Service is an attempt to overload a computer system’s resources with the aim --- of preventing access to or availability of its data. Exploits | 31551 | | Are sequences of commands controlling the behaviour of a host through a known --- vulnerability. Generic | 16560 | A method that targets cryptography and causes a collision with each block-cipher. Reconnaissance | 12779 | A technique for gathering information about a network host and is also known as a probe. Shellcode | 1427 | A malware that penetrates a code to control a victim’s host. Worms | 164 | Attacks that replicate themselves and spread to other computers. * $\bullet$ NF-BoT-IoT-v2 An IoT NetFlow-based dataset is generated by expanding the NF- BoT-IoT dataset. The features were extracted from the publicly available pcap files and the flows were labelled with their respective attack categories. The total number of data flows are 37,763,497 out of which 37,628,460 (99.64%) are attack samples and 135,037 (0.36%) are benign. There are four attack categories in the dataset, Table 4 represents the NF-BoT-IoT-v2 distribution of all flows. Table 4: NF-BoT-IoT-v2 distribution Class | Count | Description ---|---|--- Benign | 135037 | Normal unmalicious flows Reconnaissance | 2620999 | A technique for gathering information about a network host and is also known as a probe. DDoS | 18331847 | | Distributed Denial of Service is an attempt similar to DoS but has multiple --- different distributed sources. DoS | 16673183 | | An attempt to overload a computer system’s resources with the aim of preventing access --- to or availability of its data. Theft | 2431 | | A group of attacks that aims to obtain sensitive data such as data theft and keylogging --- * $\bullet$ NF-ToN-IoT-v2 The publicly available pcaps of the ToN-IoT dataset are utilised to generate its NetFlow records, leading to a NetFlow-based IoT network dataset called NF-ToN-IoT. The total number of data flows are 16,940,496 out of which 10,841,027 (63.99%) are attack samples and 6,099,469 (36.01%) are benign. Table 5 lists and defines the distribution of the NF-ToN-IoT-v2 dataset. Table 5: NF-ToN-IoT-v2 distribution Class | Count | Description ---|---|--- Benign | 6099469 | Normal unmalicious flows Backdoor | 16809 | | A technique that aims to attack remote-access computers by replying to specific constructed --- client applications DoS | 712609 | | An attempt to overload a computer system’s resources with the aim of preventing access to or --- availability of its data. DDoS | 2026234 | | An attempt similar to DoS but has multiple --- different distributed sources. Injection | 684465 | | A variety of attacks that supply untrusted inputs that aim to alter the course of --- execution, with SQL and Code injections two of the main ones. MITM | 7723 | | Man In The Middle is a method that places an attacker between a victim and host with which --- the victim is trying to communicate, with the aim of intercepting traffic and communications. Password | 1153323 | covers a variety of attacks aimed at retrieving passwords by either brute force or sniffing. Ransomware | 3425 | | An attack that encrypts the files stored on a host and asks for compensation in exchange for --- the decryption technique/key. Scanning | 3781419 | | A group that consists of a variety of techniques that aim to discover information about networks --- and hosts, and is also known as probing. XSS | 2455020 | | Cross-site Scripting is a type of injection in which an attacker uses web applications to send --- malicious scripts to end-users. * $\bullet$ NF-CSE-CIC-IDS2018-v2 The original pcap files of the CSE-CIC-IDS2018 dataset are utilised to generate a NetFlow-based dataset called NF-CSE-CIC-IDS2018-v2. The total number of flows are 18,893,708 out of which 2,258,141 (11.95%) are attack samples and 16,635,567 (88.05%) are benign ones, Table 6 represents the dataset’s distribution. Table 6: NF-CSE-CIC-IDS2018-v2 distribution Class | Count | Description ---|---|--- Benign | 16635567 | Normal unmalicious flows BruteForce | 120912 | | A technique that aims to obtain usernames and password credentials by accessing a list of --- predefined possibilities Bot | 143097 | | An attack that enables an attacker to remotely control several hijacked computers to perform --- malicious activities. DoS | 483999 | | An attempt to overload a computer system’s resources with the aim of preventing access to or --- availability of its data. DDoS | 1390270 | An attempt similar to DoS but has multiple different distributed sources. Infiltration | 116361 | | An inside attack that sends a malicious file via an email to exploit an application and is --- followed by a backdoor that scans the network for other vulnerabilities Web Attacks | 3502 | A group that includes SQL injections, command injections and unrestricted file uploads * $\bullet$ NF-UQ-NIDS-v2 A comprehensive dataset, merging all the aforementioned datasets. The newly published dataset represents the benefits of the shared dataset feature sets, where the merging of multiple smaller datasets is possible. This will eventually lead to a bigger and universal NIDS dataset containing flows from multiple network setups and different attack settings. It includes an additional label feature, identifying the original dataset of each flow. This can be used to compare the same attack scenarios conducted over two or more different testbed networks. The attack categories have been modified to combine all parent categories. Attacks named DoS attacks-Hulk, DoS attacks-SlowHTTPTest, DoS attacks-GoldenEye and DoS attacks-Slowloris have been renamed to the parent DoS category. Attacks named DDoS attack-LOIC-UDP, DDoS attack-HOIC and DDoS attacks-LOIC-HTTP have been renamed to DDoS. Attacks named FTP-BruteForce, SSH-Bruteforce, Brute Force -Web and Brute Force -XSS have been combined as a brute-force category. Finally, SQL Injection attacks have been included in the injection attacks category. The NF-UQ-NIDS dataset has a total of 75,987,976 records, out of which 25,165,295 (33.12%) are benign flows and 50,822,681 (66.88%) are attacks. Table 7 lists the distribution of the final attack categories. Table 7: NF-UQ-NIDS-v2 distribution Class | Count | Class | Count ---|---|---|--- Benign | 25165295 | Scanning | 3781419 DDoS | 21748351 | Fuzzers | 22310 Reconnaissance | 2633778 | Backdoor | 18978 Injection | 684897 | Bot | 143097 DoS | 17875585 | Generic | 16560 Brute Force | 123982 | Analysis | 2299 Password | 1153323 | Shellcode | 1427 XSS | 2455020 | MITM | 7723 Infilteration | 116361 | Worms | 164 Exploits | 31551 | Ransomware | 3425 Theft | 2431 | | ## 4 Evaluation In this section, the proposed NetFlow feature set is evaluated across five NIDS datasets; NF-UNSW-NB15-v2, NF-BoT-IoT-v2, NF-ToN-IoT-v2, NF-CSE-CIC- IDS2018-v2 and NF-UQ-NIDS-v2. An ensemble ML classifier, known as Extra Trees, that belongs to the trees family has been utilised for this purpose. The evaluation is conducted by comparing the classifier performance with the corresponding metrics of the basic NetFlow and original datasets. Various classification metrics are collected such as accuracy, Area Under the Curve (AUC), F1 Score, Detection Rate (DR), False Alarm Rate (FAR) and time required to predict a single test sample in microseconds (µs). As part of the data pre- processing, the flow identifiers such as IDs, source/destination IP and ports, timestamps, and start/end time are dropped to avoid learning bias towards attacking and victim end nodes. For the UNSW-NB15 and NF-UNSW-NB15-v2 datasets, The Time To Live (TTL)-based features are dropped due to their extreme correlation with the labels. Additionally, the min-max normalisation technique has been applied to scale all datasets’ values between 0 and 1. The datasets have been split into 70%-30% for training and testing purposes. For a fair evaluation, five cross-validation splits are conducted and the mean is measured. ### 4.1 Binary-class Classification In Table 8, the attack detection (binary classification) performance of the datasets has been measured and compared to the original and basic NetFlow datasets. Using the NF-UNSW-NB15-v2 dataset, the ML model’s performance has significantly increased with an AUC of 0.9845, compared to 0.9485 and 0.9545 when using the NF-UNSW-NB15 and UNSW-NB15 datasets respectively. The model achieved the highest F1 score of 0.97 in the shortest prediction time when using the extended NetFlow feature set. The NF-BoT-IoT-v2 dataset has enabled the ML model to achieve the highest possible detection accuracy and F1 score, the same as the BoT-IoT dataset. However, the model has a significantly lower FAR and prediction time, resulting in an increased AUC of 0.9987 and a lower prediction time of 3.90 µs compared. Using the extended NetFlow feature set, the ML model achieved significantly higher accuracy than NF-BoT-IoT of 100% compared to 93.82%. Table 8: Binary-class classification results Dataset | Accuracy | AUC | F1 Score | DR | FAR | Prediction Time (µs) ---|---|---|---|---|---|--- UNSW-NB15 | 99.25% | 0.9545 | 0.92 | 91.25% | 0.35% | 10.05 NF-UNSW-NB15 | 98.62% | 0.9485 | 0.85 | 90.70% | 1.01% | 7.79 NF-UNSW-NB15-v2 | 99.73% | 0.9845 | 0.97 | 97.07% | 0.16% | 5.92 BoT-IoT | 100.00% | 0.9948 | 1.00 | 100.00% | 1.05% | 7.62 NF-BoT-IoT | 93.82% | 0.9628 | 0.97 | 93.70% | 1.13% | 5.37 NF-BoT-IoT-v2 | 100.00% | 0.9987 | 1.00 | 100.00% | 0.26% | 3.90 ToN-IoT | 97.86% | 0.9788 | 0.99 | 97.86% | 2.10% | 8.93 NF-ToN-IoT | 99.66% | 0.9965 | 1.00 | 99.67% | 0.37% | 6.05 NF-ToN-IoT-v2 | 99.64% | 0.9959 | 1.00 | 99.76% | 0.58% | 8.47 CSE-CIC-IDS2018 | 98.31% | 0.9684 | 0.94 | 94.75% | 1.07% | 23.01 NF-CSE-CIC-IDS2018 | 95.33% | 0.9506 | 0.83 | 94.71% | 4.59% | 17.04 NF-CSE-CIC-IDS2018-v2 | 99.35% | 0.9829 | 0.97 | 96.89% | 0.31% | 21.75 NF-UQ-NIDS | 97.25% | 0.9669 | 0.94 | 95.66% | 2.27% | 14.35 NF-UQ-NIDS-v2 | 97.90% | 0.9830 | 0.98 | 97.12% | 0.52% | 14.18 The intrusion detection results of the ML model using the NF-ToN-IoT-v2 dataset are superior to its original ToN-IoT dataset. Compared to NF-ToN-IoT, it achieved a higher DR of but a slightly higher FAR. Overall, the accuracy achieved by the model using the NF-ToN-IoT-v2 is 99.64%, which is higher than ToN-IoT (97.86%) and similar to NF-ToN-IoT (99.66%). The model performance when using the NF-CSE-CIC-IDS2018-v2 dataset is notably more efficient than the CSE-CIC-IDS2018 and NF-CSE-CIC-IDS2018-v2 datasets. It achieved a high DR of 96.89% and a low FAR of 0.31% and required 21.75 µs per sample prediction. The overall accuracy achieved is 99.35%, which is higher than both the CSE- CIC-IDS2018 (98.31%) and NF-CSE-CIC-IDS2018 (95.33%) datasets. The merged NF- UQ-NIDS-v2 dataset enabled the model to achieve an accuracy of 97.90%, a DR of 97.12% and a FAR of 0.52%, outperforming the NF-UQ-NIDS dataset with a lower prediction time of 14.18 µs. Figure 3: Binary-class classification F1 score Figure 3 visually represents the F1 score obtained when applying an Extra Trees classifier on the three different feature sets of five NIDS datasets; the original as well as basic and proposed NetFlow feature sets. This fair comparison between the NetFlow feature sets demonstrates the benefit of having a common feature set across multiple datasets. It enables the evaluation of various attack detections using a common feature set. Overall, the proposed (extended) NetFlow feature set has outperformed the original and basic feature sets in terms of attack detection performance. All datasets have significantly achieved a higher or similar F1 score to their respective datasets. It is clear that using the proposed feature set achieves a reliable detection performance. Further feature selection experiments are required to identify its key features to enhance the extraction tasks. ### 4.2 Multi-class Classification Table 9: NF-UNSW-NB15-v2 multi-class classification results | UNSW-NB15 | NF-UNSW-NB15 | NF-UNSW-NB15-v2 ---|---|---|--- Class Name | DR | F1 Score | DR | F1 Score | DR | F1 Score Benign | 99.72% | 1.00 | 99.02% | 0.99 | 99.85% | 1.00 Analysis | 4.39% | 0.03 | 28.28% | 0.15 | 30.89% | 0.17 Backdoor | 13.96% | 0.08 | 39.17% | 0.17 | 40.30% | 0.18 DoS | 13.63% | 0.18 | 31.84% | 0.41 | 29.57% | 0.36 Exploits | 83.25% | 0.80 | 81.04% | 0.82 | 80.41% | 0.84 Fuzzers | 50.50% | 0.57 | 62.63% | 0.55 | 80.57% | 0.85 Generic | 86.08% | 0.91 | 57.13% | 0.66 | 85.15% | 0.90 Reconnaissance | 75.90% | 0.80 | 76.89% | 0.82 | 80.02% | 0.83 Shellcode | 53.61% | 0.59 | 87.91% | 0.75 | 87.67% | 0.69 Worms | 5.26% | 0.09 | 52.91% | 0.55 | 85.98% | 0.69 Weighted Average | 98.19% | 0.98 | 97.62% | 0.98 | 98.90% | 0.99 Prediction Time (µs) | 9.94 | 9.35 | 8.81 To further evaluate the proposed NetFlow feature set, multi-classification experiments are conducted to measure the weighted average of DR, F1 score and prediction time of each class present in the datasets. Tables 9, 10 11, 12 and 13 represent the performances of the NF-UNSW-NB15-v2, NF-BoT-IoT-v2, NF-ToN- IoT-v2, NF-CSE-CIC-IDS201-v2 and NF-UQ-NIDS-v2 datasets respectively. The datasets made up of the original and basic NetFlow feature sets are provided for comparison purposes. In Table 9, the benefits of using the NF-UNSW-NB15-v2 over the former datasets are realised by increasing the Ml model’s F1 score to 0.99 from 0.98 and decreasing the prediction time to 8.81 µs. The DR of certain attack types such as fuzzers, generic, and worms have significantly improved while the others have remained at slightly the same rate. The detection of the analysis, backdoor and DoS attacks are still unreliable when using the extended Netflow feature set, further analysis is required to identify the missing key features. However, due to their small number of samples, the overall accuracy of the NF-UNSW-NB15-v2 is higher (98.90%) than UNSW-NB15 (98.19%) and NF-UNSW-NB15 (97.62%). Table 10: NF-BoT-IoT-v2 multi-class classification results | BoT-IoT | NF-BoT-IoT | NF-BoT-IoT-v2 ---|---|---|--- Class Name | DR | F1 Score | DR | F1 Score | DR | F1 Score Benign | 99.58% | 0.99 | 98.65% | 0.43 | 99.76% | 1.00 DDoS | 100.00% | 1.00 | 30.37% | 0.28 | 99.99% | 1.00 DoS | 100.00% | 1.00 | 36.33% | 0.31 | 99.99% | 1.00 Reconnaissance | 100.00% | 1.00 | 89.95% | 0.90 | 99.93% | 1.00 Theft | 91.16% | 0.95 | 88.06% | 0.18 | 83.01% | 0.85 Weighted Average | 100.00% | 1.00 | 73.58% | 0.77 | 99.99% | 1.00 Prediction Time (µs) | 12.63 | 9.19 | 11.86 Table 10 shows that when using the NF-BoT-IoT-v2 dataset, the ML model is achieving the almost perfect multi-classification performance gained when using the BoT-IoT dataset of 100% accuracy and 1.00 F1 Score. The four attack categories are almost fully detected except for the theft attacks, where only 83.01% were successfully detected. The accuracy of the ML model is increased from 73.58% to 99.99% and the F1 score from 0.77 to 1.00 when applied to the extended NetFlow feature set compared to the basic set. Overall, it is a significant improvement that overcomes the performance limitations faced by the basic NetFlow datasets, despite the slight increase in prediction time. Table 11: NF-ToN-IoT-v2 multi-class classification results | ToN-IoT | NF-ToN-IoT | NF-ToN-IoT-v2 ---|---|---|--- Class Name | DR | F1 Score | DR | F1 Score | DR | F1 Score Benign | 89.97% | 0.94 | 98.97% | 0.99 | 99.44% | 0.99 Backdoor | 98.05% | 0.31 | 99.22% | 0.98 | 99.79% | 1.00 DDoS | 96.90% | 0.98 | 63.22% | 0.72 | 98.76% | 0.99 DoS | 53.89% | 0.57 | 95.91% | 0.48 | 89.41% | 0.91 Injection | 96.67% | 0.96 | 41.47% | 0.51 | 90.14% | 0.91 MITM | 66.25% | 0.16 | 52.81% | 0.38 | 37.45% | 0.45 Password | 86.99% | 0.92 | 27.36% | 0.24 | 97.16% | 0.97 Ransomware | 89.87% | 0.11 | 87.33% | 0.83 | 97.29% | 0.98 Scanning | 75.05% | 0.85 | 31.30% | 0.08 | 99.67% | 1.00 XSS | 98.83% | 0.99 | 24.49% | 0.19 | 96.83% | 0.96 Weighted Average | 84.61% | 0.87 | 56.34% | 0.60 | 98.05% | 0.98 Prediction Time (µs) | 12.02 | 21.21 | 12.15 In Table 11, the NF-ToN-IoT-v2 dataset has enabled the ML model to achieve outstanding results when conducting multi-classification experiments. The extended NetFlow feature set notably outperformed both the ToN-IoT and NF-ToN- IoT feature sets by increasing the model’s weighted F1 score to 0.98 from 0.87 and 0.60 respectively. The model also requires a lower prediction time compared to when applied to the basic NetFlow dataset. The extended Netflow feature set has increased the DR of all attack types except for DoS, MITM, and XSS attacks. Further analysis of features containing useful security events is essential to aid in their detection. Overall, the feature set of NF-ToN-IoT-v2 has aided the ML model in the detection of the attacks present in the dataset, with an enhanced accuracy of 98.05% confirming the reliability of the extended NetFlow feature set. Table 12: NF-CSE-CIC-IDS2018-v2 multi-class classification results | CSE-CIC-IDS2018 | NF-CSE-CIC-IDS2018 | NF-CSE-CIC-IDS2018-v2 ---|---|---|--- Class Name | DR | F1 Score | DR | F1 Score | DR | F1 Score Benign | 89.50% | 0.94 | 69.83% | 0.82 | 99.69% | 1.00 Bot | 99.92% | 0.99 | 100.00% | 1.00 | 100.00% | 1.00 Brute Force -Web | 71.36% | 0.01 | 50.21% | 0.52 | 28.05% | 0.01 Brute Force -XSS | 72.17% | 0.72 | 49.16% | 0.39 | 29.34% | 0.00 DDoS attack-HOIC | 100.00% | 1.00 | 45.66% | 0.39 | 57.33% | 0.73 DDoS attack-LOIC-UDP | 83.59% | 0.82 | 80.98% | 0.82 | 99.29% | 1.00 DDoS attacks-LOIC-HTTP | 99.93% | 1.00 | 99.93% | 0.71 | 100.00% | 1.00 DoS attacks-GoldenEye | 99.97% | 1.00 | 99.32% | 0.98 | 100.00% | 1.00 DoS attacks-Hulk | 100.00% | 1.00 | 99.65% | 0.99 | 100.00% | 1.00 DoS attacks-SlowHTTPTest | 69.80% | 0.60 | 0.00% | 0.00 | 100.00% | 1.00 DoS attacks-Slowloris | 99.44% | 0.62 | 99.95% | 1.00 | 99.99% | 1.00 FTP-BruteForce | 68.76% | 0.75 | 100.00% | 0.79 | 100.00% | 1.00 Infilteration | 36.15% | 0.08 | 62.66% | 0.04 | 39.58% | 0.43 SQL Injection | 49.34% | 0.30 | 25.00% | 0.22 | 41.44% | 0.00 SSH-Bruteforce | 99.99% | 1.00 | 99.93% | 1.00 | 100.00% | 1.00 Weighted Average | 90.28% | 0.94 | 71.92% | 0.80 | 96.90% | 0.98 Prediction Time (µs) | 24.17 | 17.29 | 27.28 Table 12 presents the detection results of the NF-CSE-CIC-IDS2018-v2 dataset. The ML model has improved the DR of most of the attacks present in the dataset, achieving an accuracy of 96.90% and an F1 score of 0.98. Most attacks were fully detected with a DR ranging between 99% to 100%. However, the detection of certain attack types such as Brute Force, DDoS attack-HOIC, infiltration, and SQL injection is still unreliable when using the extended Netflow feature set. Their respective F1 score is lower due to a high number of false positives. Overall, the performance of the ML model when applied to the NF-CSE-CIC-IDS2018-v2 dataset is superior compared to when using the CSE- CIC-IDS2018 and NF-CSE-CIC-IDS2018 datasets. However, there is an increased prediction time of 27.28 µs compared to 24.17 µs and 17.29 µs, respectively. Table 13: NF-UQ-NIDS-v2 multi-class classification results | NF-UQ-NIDS | NF-UQ-NIDS-v2 ---|---|--- Class Name | DR | F1 Score | DR | F1 Score Analysis | 69.63% | 0.21 | 78.43% | 0.24 Backdoor | 90.95% | 0.92 | 89.61% | 0.93 Benign | 71.70% | 0.83 | 93.45% | 0.96 Bot | 100.00% | 1.00 | 100.00% | 1.00 Brute Force | 99.94% | 0.85 | 98.16% | 0.74 DoS | 55.54% | 0.62 | 99.46% | 1.00 Exploits | 80.65% | 0.81 | 85.16% | 0.84 Fuzzers | 63.24% | 0.54 | 80.58% | 0.84 Generic | 58.90% | 0.61 | 85.41% | 0.88 Infilteration | 60.57% | 0.03 | 21.62% | 0.19 Reconnaissance | 88.96% | 0.88 | 98.24% | 0.76 Shellcode | 83.89% | 0.15 | 89.35% | 0.34 Theft | 87.22% | 0.15 | 81.66% | 0.22 Worms | 52.97% | 0.46 | 87.20% | 0.71 DDoS | 77.08% | 0.69 | 99.43% | 1.00 Injection | 40.58% | 0.50 | 90.03% | 0.90 MITM | 57.99% | 0.10 | 35.97% | 0.43 Password | 30.79% | 0.27 | 97.09% | 0.97 Ransomware | 90.85% | 0.85 | 96.82% | 0.87 Scanning | 39.67% | 0.08 | 97.36% | 0.98 XSS | 30.80% | 0.21 | 95.72% | 0.95 Weighted Average | 70.81% | 0.79 | 96.93% | 0.97 Prediction Time (µs) | 14.74 | 25.67 Table 13 compares the attack detection results of the merged NIDS dataset; NF- UQ-NIDS-v2 compared to its former (NF-UQ-NIDS) dataset. Most of the attacks DR has increased by using the extended NetFlow feature set. The detection of attacks such as DoS, Generic, Worms, DDoS, Injection, password, scanning and XSS has significantly improved. However, attacks such as infiltration and MITM have been detected less accurately. Moreover, the time consumed to predict a single test sample has increased from 14.74 µs to 25.67 µs. An increased accuracy from 70.81% to 96.96% and an F1 score from 0.79 to 0.97 confirms the enhanced ML model detection capabilities when applied to the extended NetFlow feature set across 20 attack types conducted over several network environments. Figure 4: Multi-class classification F1 score Overall, the proposed NetFlow feature set has significantly improved the multi-class classification performance of the datasets as displayed in Figure 4. The F1 score is plotted on the y-axis and the datasets in their three feature sets are on the x-axis. The detection performance is often comparable to the original feature set but remarkably superior to the basic NetFlow feature set. Hence, the generated datasets enjoy the benefits of adopting a standard common NetFlow feature set and with enhanced detection performance. This motivates the usage of the proposed feature set across future NIDS datasets and encourages researchers to generate their datasets in the proposed format for efficient and reliable ML experiments. ## 5 Conclusion This paper proposes a NetFlow based standard feature set for NIDS datasets, as listed in Table 2. The importance of having a standard feature set allows the reliable evaluation of ML-based NIDS across multiple datasets, network environments, and attack scenarios. Moreover, the use of a standard feature set allows multiple NIDS datasets to be merged, leading to a larger variety of labelled datasets. As part of the proposed feature set evaluation, five new NIDS datasets have been generated from existing NIDS benchmark datasets. These new dataset variants have been made publicly available to the research community. Our evaluation based on an Extra Tree classifier has shown that our NetFlow-based feature set with 43 features achieves a higher classification performance (F1-Score) than the proprietary feature sets, for all the considered benchmark NIDS datasets, for both binary and multi-class classification scenarios. The proposed NetFlow-based feature sets have the further advantage of being highly practical and scalable, due to the wide availability of efficient NetFlow exporter and collections. The key benefit of having a standard feature set for NIDS datasets, and the key contribution of this paper, is the ability to more rigorously and reliably evaluate ML-based traffic classifiers across a wide range of datasets, and hence a wider range of attack types, network topologies, etc. This allows the evaluation of how well these ML-based NIDSs can generalise from the dataset they have been trained on, to other network scenarios. We believe the inability to perform such thorough and rigorous evaluation is one of the reasons for the limited deployment of ML-based NIDSs in practical network settings. Therefore, we believe the contributions of this paper can provide a step towards bridging the gap between academic research on ML-based NIDSs and their practical deployment. ## References * [1] C. N. Modi, D. R. Patel, A. Patel, and R. Muttukrishnan, “Bayesian classifier and snort based network intrusion detection system in cloud computing,” in 2012 Third International Conference on Computing, Communication and Networking Technologies (ICCCNT’12), pp. 1–7, IEEE, 2012. * [2] P. Garcia-Teodoro, J. Diaz-Verdejo, G. Maciá-Fernández, and E. Vázquez, “Anomaly-based network intrusion detection: Techniques, systems and challenges,” computers & security, vol. 28, no. 1-2, pp. 18–28, 2009. * [3] S. K. Sahu, S. Sarangi, and S. K. Jena, “A detail analysis on intrusion detection datasets,” in 2014 IEEE International Advance Computing Conference (IACC), pp. 1348–1353, 2014. * [4] A. Shiravi, H. Shiravi, M. Tavallaee, and A. A. Ghorbani, “Toward developing a systematic approach to generate benchmark datasets for intrusion detection,” Computers & Security, vol. 31, no. 3, pp. 357 – 374, 2012. * [5] A. Binbusayyis and T. Vaiyapuri, “Identifying and benchmarking key features for cyber intrusion detection: An ensemble approach,” IEEE Access, vol. 7, pp. 106495–106513, 2019. * [6] B. Claise, G. Sadasivan, V. Valluri, and M. Djernaes, “Cisco systems netflow services export version 9,” 2004. * [7] N. Moustafa and J. Slay, “Unsw-nb15: a comprehensive data set for network intrusion detection systems (unsw-nb15 network data set),” 2015 Military Communications and Information Systems Conference (MilCIS), 2015. * [8] N. Koroniotis, N. Moustafa, E. Sitnikova, and B. Turnbull, “Towards the development of realistic botnet dataset in the internet of things for network forensic analytics: Bot-iot dataset,” CoRR, vol. abs/1811.00701, 2018. * [9] N. Moustafa, “Ton-iot datasets,” 2019. * [10] I. Sharafaldin, A. Habibi Lashkari, and A. A. Ghorbani, “Toward generating a new intrusion detection dataset and intrusion traffic characterization,” Proceedings of the 4th International Conference on Information Systems Security and Privacy, 2018. * [11] M. Sarhan, S. Layeghy, N. Moustafa, and M. Portmann, “Netflow datasets for machine learning-based network intrusion detection systems,” arXiv:2011.09144, 2020. * [12] “Netflow datasets.” http://staff.itee.uq.edu.au/marius/NIDS_datasets/, 2020\. * [13] M. Ring, S. Wunderlich, D. Scheuring, D. Landes, and A. Hotho, “A survey of network-based intrusion detection data sets,” Computers & Security, vol. 86, pp. 147 – 167, 2019. * [14] R. Sommer and V. Paxson, “Outside the closed world: On using machine learning for network intrusion detection,” 2010 IEEE Symposium on Security and Privacy, 2010. * [15] 2021\. * [16] B. Li, J. Springer, G. Bebis, and M. Hadi Gunes, “A survey of network flow applications,” Journal of Network and Computer Applications, vol. 36, no. 2, pp. 567–581, 2013. * [17] D. R. Kerr and B. L. Bruins, “Network Flow Switching and Flow Data Export,” 2001\. * [18] Cisco Systems, “Cisco IOS NetFlow Version 9 Flow-Record Format - White Paper.” https://www.cisco.com/en/US/technologies/tk648/tk362/technologies_white_paper09186a00800a3db9.pdf, 2011\. * [19] Ntop, “nProbe, An Extensible NetFlow v5/v9/IPFIX Probe for IPv4/v6.” https://www.ntop.org/guides/nprobe/cli_options.html, 2017.
# Optically induced Kondo effect in semiconductor quantum wells I. V. Iorsh1,2 O. V. Kibis2 Oleg.Kibis(c)nstu.ru 1Department of Physics and Engineering, ITMO University, Saint-Petersburg, 197101, Russia 2Department of Applied and Theoretical Physics, Novosibirsk State Technical University, Karl Marx Avenue 20, Novosibirsk 630073, Russia ###### Abstract It is demonstrated theoretically that the circularly polarized irradiation of two-dimensional electron systems can induce the localized electron states which antiferromagnetically interact with conduction electrons, resulting in the Kondo effect. Conditions of experimental observation of the effect are discussed for semiconductor quantum wells. ## I Introduction In 1964, Jun Kondo in his pioneering article kondo1964resistance suggested the physical mechanism responsible for the minimum of temperature dependence of the resistivity of noble divalent metals, which had remained a mystery for more than three decades de1934electrical . Within the developed theory, he showed that the antiferromagnetic interaction between the spins of conduction electrons and electrons localized on magnetic impurities leads to the $\log(T)$ corrections to the relaxation time of conduction electrons (the Kondo effect). The subsequent studies on the subject wilson1975renormalization ; Wiegmann_1981 ; Andrei_1983 demonstrated that physics of the Kondo effect is universal to describe the transformation of the ground state of various many-body systems in the broad range of energies. Particularly, the transformation is characterized by the single energy scale $T_{K}$ (the Kondo temperature) and can be effectively treated by the powerful methods of the renormalization group theory. Therefore, the Kondo problem is currently considered as an effective testing ground to solve many challenging many-body problems, including heavy-fermion materials, high-temperature superconductors, etc steglich1979superconductivity ; andres19754 ; tsunetsugu1993phase ; RevModPhys.56.755 . The Kondo temperature is defined by the Coulomb repulsion of the impurity atoms, hybridization of the conduction and impurity electrons, and other condensed-matter parameters which are fixed in bulk materials but can be effectively tuned in nanostructures. Since the first observation of the tunable Kondo effect in such nanostructures as quantum dots goldhaber1998kondo , it attracts the enormous attention of research community cronenwett1998tunable ; iftikhar2018tunable ; park2002coulomb ; Borzenets_2020 . While the tuning of Kondo temperature in nanostructures is usually achieved by stationary fields (produced, e.g., by the gate voltage iftikhar2018tunable ), it should be noted that all physical properties of them can be effectively controlled also by optical methods. Particularly, it has been demonstrated that the resonant laser driving of the impurity spins (such as quantum dot spins or single atom spins in the optical lattices) allows for the control over the onset and destruction of the Kondo resonance Latta_2011 ; Haupt_2013 ; Tureci_2011 ; Sbierski_2013 ; Nakagawa_2015 . An alternative optical way of controlling the Kondo effect could be based on the modification of electronic properties by an off-resonant high-frequency electromagnetic field (the Floquet engineering Basov_2017 ; Oka_2019 ), which became the established research area of modern physics and resulted in many fundamental effects in various nanostructures Goldman_2014 ; Bukov_2015 ; Lindner_2011 ; Savenko_2012 ; Iorsh_2017 ; Kibis_2016 ; Kibis_2017 ; Kozin_2018 ; Kozin_2018_1 ; Rechtsman_2013 ; Wang_2013 ; Glazov_2014 ; Torres_2014 ; Sentef_2015 ; Sie_2015 ; Cavalleri_2020 . Since the frequency of the off-resonant field lies far from characteristic resonant frequencies of the electron system, the field cannot be absorbed and only “dresses” electrons (dressing field), changing their physical characteristics. Particularly, it was demonstrated recently that a high-frequency circularly polarized dressing field crucially modifies the interaction of two-dimensional (2D) electron systems with repulsive scatterers, inducing the attractive area in the core of a repulsive potential Kibis_2019 . As a consequence, the light-induced electron states localized at repulsive scatterers appear Kibis_2020 ; Iorsh_2020 . Since such localized electron states are immersed into the continuum of conduction electrons and interact with them antiferromagnetically, the Kondo effect can exist. The present article is dedicated to theoretical analysis of this optically induced effect for 2D electron gas in semiconductor quantum wells (QWs). The article is organized as follows. In the second section, the model of Kondo effect based on the dressing field approach is developed. The third section is dedicated to the analysis of the Kondo resonance and the Kondo temperature in QWs. The last two sections contain conclusion and acknowledgements. ## II Model For definiteness, let us consider a semiconductor QW of the area $S$ in the plane $x,y$, which is filled by 2D gas of conduction electrons with the effective mass $m_{e}$ and irradiated by a circularly polarized electromagnetic wave incident normally to the $x,y$ plane. Then the behavior of a conduction electron near a scatterer with the repulsive potential $U(\mathbf{r})$ is described by the time-dependent Hamiltonian $\hat{\cal H}_{e}(t)=(\hat{\mathbf{p}}-e\mathbf{A}(t)/c)^{2}/2m_{e}+U(\mathbf{r})$, where $\hat{\mathbf{p}}$ is the plane momentum operator of conduction electron, $\mathbf{r}=(x,y)$ is the plane radius vector of the electron, $\mathbf{A}(t)=(A_{x},A_{y})=[cE_{0}/\omega_{0}](\sin\omega_{0}t,\,\cos\omega_{0}t)$ (1) is the vector potential of the wave, $E_{0}$ is the electric field amplitude of the wave, and $\omega_{0}$ is the wave frequency. If the field frequency $\omega_{0}$ is high enough and lies far from characteristic resonant frequencies of the QW, this time-dependent Hamiltonian can be reduced to the effective stationary Hamiltonian, $\hat{\cal H}_{0}=\hat{\mathbf{p}}^{2}/2m_{e}+U_{0}(\mathbf{r})$, where $U_{0}(\mathbf{r})=\frac{1}{2\pi}\int_{-\pi}^{\pi}U\big{(}\mathbf{r}-\mathbf{r}_{0}(t)\big{)}\,d(\omega_{0}t)$ (2) is the repulsive potential modified by the incident field (dressed potential), $\mathbf{r}_{0}(t)=(-r_{0}\cos\omega_{0}t,\,r_{0}\sin\omega_{0}t)$ is the radius-vector describing the classical circular trajectory of a free electron in the circularly polarized field (1), and $r_{0}={|e|E_{0}}/{m_{e}\omega^{2}_{0}}$ is the radius of the trajectory Kibis_2019 ; Kibis_2020 . In the case of the short-range scatterers conventionally modelled by the repulsive delta potential, $U(\mathbf{r})=u_{0}\delta(\mathbf{r}),$ (3) the corresponding dressed potential (4) reads Kibis_2020 $U_{0}(\mathbf{r})=\frac{u_{0}\,\delta({r}-{r}_{0})}{2\pi r_{0}}.$ (4) Thus, the circularly polarized dressing field (1) turns the repulsive delta potential (3) into the delta potential barrier of ring shape (4), which defines dynamics of an electron near the scatterer. As a consequence, the bound electron states which are localized inside the area fenced by the ring- shape barrier ($0<r<r_{0}$) appear. Certainly, such bound electron states are quasi-stationary since they can decay via the tunnel transition through the potential barrier (4) into the continuum of conduction electrons. As a consequence, the energy broadening of the localized states appears. In the following, we will restrict the analysis by the ground localized state with the energy $\varepsilon_{0}$, the energy broadening $\Gamma_{0}$ and the wave function $\psi_{0}(r)$ (see Fig. 1). Assuming the repulsive delta potential to be strong enough ($\alpha=2\hbar^{2}/m_{e}u_{0}\ll 1$), the solution of the Schrödinger problem with the stationary potential (4) can be written approximately as Kibis_2020 $\displaystyle\varepsilon_{0}=\frac{\hbar^{2}\xi^{2}_{0}}{2m_{e}r^{2}_{0}},\,\,\,{\Gamma}_{0}=\frac{2\varepsilon_{0}\alpha^{2}}{N^{3}_{0}(\xi_{0})J_{1}(\xi_{0})},$ $\displaystyle\psi_{0}(r)=\frac{J_{0}\left({\xi_{0}r}/{r_{0}}\right)}{\sqrt{\pi}r_{0}J_{1}(\xi_{0})}\theta(r_{0}-r),$ (5) where $J_{m}(\xi)$ and $N_{m}(\xi)$ are the Bessel functions of the first and second kind, respectively, $\xi_{0}$ is the first zero of the Bessel function $J_{0}(\xi)$, and $\theta(r)$ is the Heaviside function. It follows from the theory of the dressed potential approach Kibis_2019 that the discussed model based on the dressed potential (4) is applicable if the dressing field frequency, $\omega_{0}$, much exceeds the characteristic frequency of the bound electron state, $\varepsilon_{0}/\hbar$. Figure 1: Sketch of the system under consideration: (a) the optically induced potential (4) depicted by the vertical blue line, which confines the bound electron state (II) marked by the horizontal yellow strip and separates it from the states of conduction electrons with the wave vectors $\mathbf{k}$ marked by the wave arrow; (b) the energy structure of singly-occupied ($\uparrow$) and doubly occupied ($\uparrow\downarrow$) electron states (II) near the Fermi energy $\varepsilon_{F}$. Assuming the condition $\hbar\omega_{0}\gg\varepsilon_{0}$ to be satisfied, interaction between the localized electron state (II) and the conduction electrons can be described by the Hamiltonian $\displaystyle\hat{\cal H}$ $\displaystyle=$ $\displaystyle\sum_{\mathbf{k},\sigma}(\varepsilon_{\mathbf{k}}-\varepsilon_{F})\hat{c}_{\mathbf{k}\sigma}^{\dagger}\hat{c}_{\mathbf{k}\sigma}+\sum_{\sigma}(\varepsilon_{0}-\varepsilon_{F})\hat{d}_{\sigma}^{\dagger}\hat{d}_{\sigma}$ (6) $\displaystyle+$ $\displaystyle U_{C}\hat{d}_{\uparrow}^{\dagger}\hat{d}_{\uparrow}\hat{d}_{\downarrow}^{\dagger}\hat{d}_{\downarrow}+\sum_{\mathbf{k},\sigma}T_{\mathbf{k}}\left[\hat{c}_{\mathbf{k},\sigma}^{\dagger}\hat{d}_{\sigma}+\mathrm{H.c.}\right],$ where $\varepsilon_{\mathbf{k}}=\hbar^{2}k^{2}/2m_{e}$ is the energy spectrum of conduction electrons, $\mathbf{k}$ is the electron wave vector, $\varepsilon_{F}$ is the Fermi energy of conduction electrons, $\sigma=\uparrow,\downarrow$ is the spin quantum number, $\hat{c}_{\mathbf{k},\sigma}^{\dagger}$($\hat{c}_{\mathbf{k},\sigma}$) are the production (annihilation) operators for conduction electron states, $\hat{d}_{\sigma}^{\dagger}$($\hat{d}_{\sigma}$) are the production (annihilation) operators for the light-induced localized electron states (II), $\displaystyle U_{C}=e^{2}\int_{S}d^{2}\mathbf{r}\int_{S}d^{2}\mathbf{r}^{\prime}\frac{|\psi_{0}({r})|^{2}|\psi_{0}({r^{\prime}})|^{2}}{|\mathbf{r}-\mathbf{r}^{\prime}|}=\frac{\gamma e^{2}}{\epsilon r_{0}},$ (7) is the Coulomb interaction energy of two electrons with opposite spins $\sigma=\uparrow,\downarrow$ in the state (II), $\epsilon$ is the permittivity of QW, $\gamma\approx 0.8$ is the numerical constant, and $T_{\mathbf{k}}$ is the tunneling matrix element connecting the localized electron state (II) and the conduction electron state with the wave vector $\mathbf{k}$. Physically, the first term of the Hamiltonian (6) describes the energy of conduction electrons, the second term describes the energy of the localized electron in the state (II), the third therm describes the Coulomb energy shift of the double occupied state (II), and the fourth term describes the tunnel interaction between the conduction electrons and the localized electrons. Assuming the tunneling to be weak enough, one can replace the matrix element $T_{\mathbf{k}}$ with its resonant value, $|T_{\mathbf{k}}|^{2}=\Gamma_{0}/\pi N_{\varepsilon}$, corresponding to the energy $\varepsilon_{\mathbf{k}}=\varepsilon_{0}$, where $N_{\varepsilon}=Sm_{e}/\pi\hbar^{2}$ is the density of conduction electron states (see, e.g., Appendix B in Ref. Kibis_2020, ). If the localized and delocalized (conduction) electron states are decoupled from each other ($T_{\mathbf{k}}=0$), the localized eigenstates of the Hamiltonian (6) correspond to the singly occupied state (II) with the eigenenergy $\varepsilon_{0}-\varepsilon_{F}$ and the doubly occupied state (II) with the eigenenergy $2(\varepsilon_{0}-\varepsilon_{F})+U_{C}$, which are marked schematically in Fig. 1b by the symbols $\uparrow$ and $\uparrow\downarrow$, correspondingly. For completeness, it should be noted that the empty state (II) is also the eigenstate of the considered Hamiltonian with the zero eigenenergy corresponding to the Fermi level. Since the Kondo effect originates due to the emergence of magnetic moment (spin) of a localized electron, it appears only if the singly occupied state is filled by an electron but the doubly occupied state is empty. Assuming the temperature to be zero, this corresponds to the case of $\varepsilon_{0}-\varepsilon_{F}<0$ and $U_{C}>\varepsilon_{F}-\varepsilon_{0}$. Taking into account that the characteristic energies of the considered problem, $U_{C}$ and $\varepsilon_{0}$, depend differently on the irradiation amplitude $E_{0}$ and frequency $\omega_{0}$, the optically induced Kondo effect can exist only in the range of these irradiation parameters defined by the inequality $\sqrt{\frac{\hbar^{2}\xi_{0}^{2}}{2m_{e}\varepsilon_{F}}}<r_{0}<\frac{\gamma e^{2}}{2\epsilon\varepsilon_{F}}+\sqrt{\frac{\hbar^{2}\xi_{0}^{2}}{2m_{e}\varepsilon_{F}}+\left(\frac{\gamma e^{2}}{2\epsilon\varepsilon_{F}}\right)^{2}}.$ (8) Mathematically, the Hamiltonian (6) is identical to the famous Anderson Hamiltonian describing the microscopic mechanism for the magnetic moment formation in metals anderson1961localized . Therefore, one can apply the known Schrieffer-Wolff (SW) unitary transformation coleman2015introduction to turn the Hamiltonian (6) into the Hamiltonian of the Kondo problem Andrei_1983 . Assuming the condition (8) to be satisfied, we arrive at the Kondo Hamiltonian $\hat{\cal H}_{K}=\sum_{\mathbf{k}\sigma}(\varepsilon_{\mathbf{k}}-\varepsilon_{F})\hat{c}_{\mathbf{k}\sigma}^{\dagger}\hat{c}_{\mathbf{k}\sigma}+J\bm{\sigma}(0)\cdot\mathbf{S}_{0}-\frac{{V}}{2}\hat{\psi}_{\sigma}^{\dagger}(0)\hat{\psi}_{\sigma}(0),$ (9) where $\hat{\psi}_{\sigma}(0)=\sum_{\mathbf{k}}\hat{c}_{\mathbf{k}\sigma}$ is the $\hat{\psi}(\mathbf{r})$ operator of conduction electrons at the repulsive delta potential ($\mathbf{r}=0$), ${\hat{\bm{\sigma}}}(0)=\hat{\psi}^{\dagger}(0){\hat{\bm{\sigma}}}\hat{\psi}(0)$ is the spin density of conduction electrons at $\mathbf{r}=0$, $\mathbf{S}_{0}=\hat{d}^{\dagger}({{\hat{\bm{\sigma}}}}/{2})\hat{d}$ is the spin density of an electron in the localized state (II), $\hat{\bm{\sigma}}=(\sigma_{x},\sigma_{y},\sigma_{z})$ is the spin vector matrix, and the coupling coefficients $J$ and ${V}$ read $\displaystyle J$ $\displaystyle=$ $\displaystyle\frac{\Gamma_{0}}{\pi N_{\varepsilon}}\left[\frac{1}{\varepsilon_{0}-\varepsilon_{F}+U_{C}}+\frac{1}{\varepsilon_{F}-\varepsilon_{0}}\right],$ (10) $\displaystyle{V}$ $\displaystyle=$ $\displaystyle-\frac{\Gamma_{0}}{2\pi N_{\varepsilon}}\left[\frac{1}{\varepsilon_{0}-\varepsilon_{F}+U_{C}}-\frac{1}{\varepsilon_{F}-\varepsilon_{0}}\right].$ (11) It should be noted that the denominators in Eqs. (10)–(11) are the energy detunings between the singly occupied and empty states, $\varepsilon_{F}-\varepsilon_{0}$, and the singly and doubly occupied states, $\varepsilon_{0}-\varepsilon_{F}+U_{C}$. Since excitation of the empty (doubly occupied) state creates an electron (hole) in the Fermi sea, we will label the corresponding detunings as $D_{e}=\varepsilon_{F}-\varepsilon_{0}$ and $D_{h}=\varepsilon_{0}-\varepsilon_{F}+U_{c}$. It should be stressed also that the SW transformation is only applicable to the case of weak coupling between the localized electron state (II) and the conduction electrons as compared to the energy difference between the ground (singly occupied) and excited (empty and doubly occupied) localized states. As a consequence, the Hamiltonian (9) accurately describes the asymmetric Kondo problem under consideration for $\Gamma_{0}\ll[\varepsilon_{F}-\varepsilon_{0},\varepsilon_{0}-\varepsilon_{F}+U_{C}]$ and, therefore, the detunings $D_{e,h}$ are assumed to meet the condition $\Gamma_{0}\ll D_{e,h}$. Beyond the condition, the system enters the so-called mixed-valence regime, where the localized states (II) with different occupancy become effectively coupled riseborough2016mixed . Although the mixed valence regime hosts a rich class of the interesting phase transitions, in the following we will focus exclusively at the Kondo regime, where the singly- occupied ground localized state is well separated from the excited states. ## III Results and discussion Figure 2: Effect of the circularly polarized irradiation with the frequency $\omega_{0}/2\pi=200$ GHz and the intensity $I$ on (a) the hole and electron detunings $D_{h,e}$ and (b) the Kondo temperature in a GaAs-based QW filled by 2D electron gas with the Fermi energy $\varepsilon_{F}=5$ meV, energy broadening $\Gamma_{0}=0.1\varepsilon_{0}$ and the electron effective mass $m_{e}=0.067m_{0}$ ($m_{0}$ is the free electron mass) for the zero temperature. The green shadow areas mark the validity range of the model, where the applicability conditions are satisfied for both the Kondo Hamiltonian ($\Gamma_{0}\ll D_{e},D_{h}$) and the dressed potential approach ($\varepsilon_{0}\ll\hbar\omega_{0}$). For the particular case of 2D electrons in a GaAs-based QW, the dependence of the detunings $D_{e,h}$ on the irradiation is plotted in Fig. 2a. It should be noted that excitations of virtual electrons and holes should be considered within the whole conduction band of width $2D_{0}$, where $D_{0}\approx 1.5$ eV for GaAs. Since the typical Kondo temperature is essentially smaller than the bandwidth $D_{0}$, one needs to transform the initial high-energy Hamiltonian (9) to the low-energy range in order to find the Kondo temperature. Such a transformation can be performed within the poor man’s scaling renormalization approach anderson1970poor ; anderson1973kondo , which was originally proposed by Anderson. Following Anderson, the higher energy excitations corresponding to the first order processes pictured in Fig. 3a can be integrated out from the Hamiltonian with using the SW transformation. Then the highest energy excitations, which are remained in the renormalized Hamiltonian, correspond to the second order processes pictured in Fig. 3b-c. It should be stressed that the renormalized Hamiltonian has the same structure as the initial one with the coupling constants depending on the renormalized (decreased) bandwidth $D<D_{0}$, where the renormalized coupling constant $J(D)$ diverges at some critical bandwidth $D=D_{K}$. This critical bandwidth defines the sought Kondo temperature, $T_{K}=D_{K}$, which, particularly, indicates the applicability limit of the perturbation theory anderson1970poor ; anderson1973kondo . Figure 3: The diagrams illustrating all possible first order processes (a) and the second order processes for electrons (b) and holes (c). The solid lines depict propagators of conduction electrons and holes, the dashed lines depict the localized spin propagator, whereas the symbols $\sigma$ and $\alpha$($\beta$) mark the spins of localized electrons and conduction electrons, respectively. The only difference of the considered system from the original Kondo problem anderson1961localized is the strong electron-hole asymmetry since the typical Fermi energy in GaAs-based QWs is $\varepsilon_{F}\ll D_{0}$. As a consequence, the hole process shown in Fig. 3c cannot exists for $D>\varepsilon_{F}$ and only the second order process involving a virtual electron (see Fig. 3b) should be taken into account. On the contrary, the both processes contribute to the effective coupling rescaling for the case of $D<\varepsilon_{F}$. Applying the known general solution of the asymmetric electron-hole Kondo problem Zitko2016 to the considered system, the flow equations for the effective exchange constant $J^{\prime}(D)$ and the scalar potential $V^{\prime}(D)$ can be written as $\displaystyle\frac{1}{{\pi N_{\varepsilon}}}\frac{\partial{J}^{\prime}}{{\partial\ln(D_{0}/D)}}=[1+\theta(\varepsilon_{F}-D)]{{J^{\prime}}^{2}-\theta(D-\varepsilon_{F}){J^{\prime}}{V^{\prime}}},$ $\displaystyle\frac{2}{{\pi N_{\varepsilon}}}\frac{\partial{V}^{\prime}}{{\partial\ln(D_{0}/D)}}=-\theta(D-\varepsilon_{F})\left[{3{J^{\prime}}^{2}+{V^{\prime}}^{2}}\right],$ (12) with the boundary conditions $J^{\prime},V^{\prime}|_{D=D_{0}}=J,V$. The solving of Eqs. (12) should be performed in the two steps as follows. At the first step, we consider the interval $\varepsilon_{F}\leq D\leq D_{0}$. Within this interval, the two nonlinear differential equations (12) can be solved analytically (see, e.g., Ref. Zitko2016, for details) and result in the boundary condition $\displaystyle J^{\prime}(\varepsilon_{F})=$ (13) $\displaystyle=\frac{J}{\left[1+\frac{\pi N_{\varepsilon}}{2}\ln\frac{D_{0}}{\varepsilon_{F}}(J+V)\right]\left[1-\frac{\pi N_{\varepsilon}}{2}\ln\frac{D_{0}}{\varepsilon_{F}}(3J-V)\right]}.$ At the second step, we consider the interval $D\leq\varepsilon_{F}$. Within this interval, the scalar potential $V^{\prime}$ is constant and the differential equation defining the effective exchange constant $J^{\prime}$ does not depend on the scalar potential. Therefore, the system of two nonlinear differential equations (12) is reduced to the two independent differential equations. Solving them under the boundary condition (13), one can find the effective exchange coupling constant $J^{\prime}(D)$. Taking into account that the coupling constant diverges at the critical (Kondo) bandwidth $D=D_{K}$, we arrive the Kondo temperature $\displaystyle T_{K}=\varepsilon_{F}\exp\left[-\frac{1}{2\pi N_{\varepsilon}J^{\prime}(\varepsilon_{F})}\right].$ (14) Certainly, Eq. (14) turns into the known expression for the Kondo temperature $T_{K}=D_{0}\exp[-1/2J\pi N_{\varepsilon}]$ corresponding to the symmetric Kondo problem anderson1970poor if $\varepsilon_{F}=D_{0}$ (the particular case of the half-filled band). The dependence of the Kondo temperature on the irradiation is plotted in Fig. 2b, where the Kondo temperature is found to be of several Kelvin. In the theory developed above, the field frequency, $\omega_{0}$, was assumed to satisfy the high-frequency condition $\omega_{0}\tau\gg 1$, where $\tau$ is the mean free time of conduction electrons in the QW. To satisfy this condition for modern QWs and keep the irradiation intensity $I$ to be reasonable, we chose the field frequency for our calculations (see Fig. 2) near the upper limit of microwave range, $\omega_{0}/2\pi=200$ GHz, which can be easily realized in experiments. It should be noted also that the present theory is developed for the case of symmetric QW, although effects in asymmetric QWs are also studied actively (see, e.g., Refs. Stavrou_2001, ; Arulmozhi_2020, ). In such asymmetric QWs, particularly, there is the Rashba spin-orbit coupling which can lead to the exponential increase of the Kondo temperature Wong_2016 . To observe the discussed effect experimentally, the known contribution of the Kondo resonance to the electron mean free time kondo1964resistance , $\displaystyle 1/\tau\sim J^{4}\left[\frac{1}{\pi N_{\varepsilon}J}+2\ln\frac{D_{0}}{T}\right]^{2},$ (15) can be used. Indeed, the found Kondo temperature (14) corresponds to the minimum of the contribution (15) as a function of the temperature $T$. Since all electron transport phenomena depend on the electron mean free time $\tau$, this minimum can be detected in various transport experiments (e.g., conductivity measurements). In order to exclude effects arisen from the irradiation-induced heating of electron gas, the difference scheme based on using both a circularly polarized field and a linearly polarized one can be applied. Indeed, the heating does not depend on the field polarization, whereas the electron states bound at repulsive scatterers — and the related Kondo effect, respectively — can be induced only by a circularly polarized field Kibis_2019 . ## IV Conclusion We showed within the Floquet theory that a circularly polarized electromagnetic field irradiating a two-dimensional electron system can induce the localized electron states which antiferromagnetically interact with conduction electrons. As a consequence, the Kondo effect appears. For semiconductor quantum wells irradiated by a microwave electromagnetic wave of the intensity $\sim$kW/cm2, the Kondo temperature is found to be of several Kelvin and, therefore, the effect can be detected in state-of-the-art transport measurements. ## V Acknowledgements The reported study was funded by the Russian Science Foundation (project 20-12-00001). ## References * (1) Kondo G Resistance 1964 Minimum in Dilute Magnetic Alloys Progress of Theoretical Physics 32 37 * (2) De Haas W J, De Boer J and Van den Berg G J 1934 The electrical resistance of gold, copper and lead at low temperatures Physica 1 1115 * (3) Wilson K G 1973 The renormalization group: Critical phenomena and the Kondo problem Rev. Mod. Phys. 47 773 * (4) Fateev V A and Wiegmann P B 1981 The exact solution of the s-d exchange model with arbitrary impurity spin S (Kondo problem) Phys. Lett. A 81 179 * (5) Andrei N, Furuya K and Lowenstein J H 1983 Solution of the Kondo problem Rev. Mod. Phys. 55 331 * (6) Steglich F, Aarts J, Bredl C D, Lieke W, Meschede D, Franz W and Schäfer H 1979 Superconductivity in the Presence of Strong Pauli Paramagnetism: CeCu2Si2 Phys. Rev. Lett. 43 1892 * (7) Andres K, Graebner J E and Ott H R 1975 4f-Virtual-Bound-State Formation in CeAl3 at Low Temperatures Phys. Rev. Lett. 35 1779 * (8) Tsunetsugu H, Sigrist M and Ueda K 1993 Phase diagram of the one-dimensional Kondo-lattice model Phys. Rev. B 47 8345 * (9) Stewart G R 1984 Heavy-fermion systems Rev. Mod. Phys. 56 755 * (10) Kouwenhoven L and Glazman L 2001 Revival of the Kondo effect Phys. World 14(1) 33 * (11) Goldhaber-Gordon D, Göres J, Kastner M A, Shtrikman H, Mahalu D and Meirav U 1998 From the Kondo Regime to the Mixed-Valence Regime in a Single-Electron Transistor Phys. Rev. Lett. 81 5225 * (12) Cronenwett S M, Oosterkamp T H and Kouwenhoven L P 1998 A Tunable Kondo Effect in Quantum Dots Science 281 540 * (13) Iftikhar Z, Anthore A, Mitchell A, Parmentier F, Gennser U, Ouerghi A, Cavanna A, Mora C, Simon P and Pierre F 2018 Tunable quantum criticality and super-ballistic transport in a “charge” Kondo circuit Science 360 1315 * (14) Park J, Pasupathy A N, Goldsmith J I, Chang C, Yaish Y, Petta J R, Rinkoski M, Sethna J P, Abruna H D, McEuen P L and Ralph D C 2002 Coulomb blockade and the Kondo effect in single-atom transistors Nature 417 722 * (15) Borzenets I V, Shim J, Chen J C H, Ludwig A, Wieck A D, Tarucha S, Sim H-S and Yamamoto M 2020 Observation of the Kondo screening cloud Nature 579 210 * (16) Latta C, Haupt F, Hanl M, Weichselbaum A, Claassen M, Wuester W, Fallahi P, Faelt S, Glazman L, von Delft J, Türeci H E and Imamoglu A 2011 Quantum quench of Kondo correlations in optical absorption Nature 474 627 * (17) Haupt F, Smolka S, Hanl M, Wüster W, Miguel-Sanchez J, Weichselbaum A, von Delft J and Imamoglu A 2013 Nonequilibrium dynamics in an optical transition from a neutral quantum dot to a correlated many-body state Phys. Rev. B 88 161304 * (18) Türeci H E, Hanl M, Claassen M, Weichselbaum A, Hecht T, Braunecker B, Govorov A, Glazman L, Imamoglu A and von Delft J 2011 Many-Body Dynamics of Exciton Creation in a Quantum Dot by Optical Absorption: A Quantum Quench towards Kondo Correlations Phys. Rev. Lett. 106 107402 * (19) Sbierski B, Hanl M, Weichselbaum A, Türeci H E, Goldstein M, Glazman L I, von Delft J and Imamoglu A 2013 Proposed Rabi-Kondo Correlated State in a Laser-Driven Semiconductor Quantum Dot Phys. Rev. Lett. 111 157402 * (20) Nakagawa M and Kawakami N 2015 Laser-Induced Kondo Effect in Ultracold Alkaline-Earth Fermions Phys. Rev. Lett. 115 165303 * (21) Basov D N, Averitt R D and Hsieh D 2017 Towards properties on demand in quantum materials Nat. Mater. 16 1077 * (22) Oka T and Kitamura S 2019 Floquet Engineering of Quantum Materials Ann. Rev. Cond. Matt. Phys. 10 387 * (23) Goldman N and Dalibard J 2014 Periodically Driven Quantum Systems: Effective Hamiltonians and Engineered Gauge Fields Phys. Rev. X 4 031027 * (24) Bukov M, D’Alessio L and Polkovnikov A 2015 Universal high-frequency behavior of periodically driven systems: From dynamical stabilization to Floquet engineering Adv. Phys. 64 139 * (25) Lindner N H, Refael G and Galitski V 2011 Floquet topological insulator in semiconductor quantum wells Nat. Phys. 7 490 * (26) Savenko I G, Kibis O V and Shelykh I A 2012 Asymmetric quantum dot in a microcavity as a nonlinear optical element Phys. Rev. A 85 053818 * (27) Rechtsman M C, Zeuner J M, Plotnik Y, Lumer Y, Podolsky D, Dreisow F, Nolte S, Segev M and Szameit A 2013 Photonic Floquet topological insulator Nature 496 196 * (28) Wang Y H, Steinberg H, Jarillo-Herrero P and Gedik N 2013 Observation of Floquet-Bloch states on the surface of a topological insulator Science 342 453 * (29) Glazov M M and Ganichev S D 2014 High frequency electric field induced nonlinear effects in graphene Phys. Rep. 535 101 * (30) Usaj G, Perez-Piskunow P M, Foa Torres L E F and Balseiro C A 2014 Irradiated graphene as a tunable Floquet topological insulator Phys. Rev. B 90 115423 * (31) Sentef M A, Claassen M, Kemper A F, Moritz B, Oka T, Freericks J K and Devereaux T P 2015 Theory of Floquet band formation and local pseudospin textures in pump-probe photoemission of graphene Nat. Commun. 6 7047 * (32) Sie E J, McIver J W, Lee Y-H, Fu L, Kong J and Gedik N 2015 Valley-selective optical Stark effect in monolayer WS2 Nat. Mater. 14 290 * (33) Dini K, Kibis O V and Shelykh I A 2016 Magnetic properties of a two-dimensional electron gas strongly coupled to light Phys. Rev. B 93 235411 * (34) Kibis O V, Dini K, Iorsh I V and Shelykh I A 2017 All-optical band engineering of gapped Dirac materials Phys. Rev. B 95 125401 * (35) Iorsh I V, Dini K, Kibis O V and Shelykh I A 2017 Optically induced Lifshitz transition in bilayer graphene Phys. Rev. B 96 155432 * (36) Kozin V K, Iorsh I V, Kibis O V and Shelykh I A 2018 Quantum ring with the Rashba spin-orbit interaction in the regime of strong light-matter coupling Phys. Rev. B 97 155434 * (37) Kozin V K, Iorsh I V, Kibis O V and Shelykh I A 2018 Periodic array of quantum rings strongly coupled to circularly polarized light as a topological insulator Phys. Rev. B 97 035416 * (38) McIver J W, Schulte B, Stein F-U, Matsuyama T, Jotzu G, Meier G and Cavalleri A 2020 Light-induced anomalous Hall effect in graphene Nat. Phys. 16 38 * (39) Kibis O V 2019 Electron pairing in nanostructures driven by an oscillating field Phys. Rev. B 99 235416 * (40) Kibis O V, Boev M V and Kovalev V M 2020 Light-induced bound electron states in two-dimensional systems: Contribution to electron transport Phys. Rev. B 102 075412 * (41) Kibis O V, Kolodny S A and Iorsh I V 2021 Fano resonances in optical spectra of semiconductor quantum wells dressed by circularly polarized light Opt. Lett. 46 50 * (42) Anderson P W 1961 Localized Magnetic States in Metals Phys. Rev. 124 41 * (43) Coleman P 2015 Introduction to many-body physics (Cambridge: University Press) * (44) Riseborough P S and Lawrence J M 2016 Mixed Valnt Metals Rep. Prog. Phys. 79 084501 * (45) Anderson P W 1979 A poor man’s derivation of scaling laws for the Kondo problem J. Phys. C: Solid State Phys. 3 2436 * (46) Zitko R and Hovrat A 1994 Kondo effect at low electron density and high particle-hole asymmetry in 1D, 2D, and 3D Phys. Rev. B 94 125138 * (47) Anderson P W Kondo effect 1973 Comments on Solid State Phys. 5 73 * (48) Stavrou V N, Babiker M and Bennett C R 2001 Influences of asymmetric quantum wells on electron-phonon interactions J. Phys.: Condens. Matter 13 6489 * (49) Arulmozhi R, John Peterb A and Lee C W 2020 Optical absorption in a CdS/CdSe/CdS asymmetric quantum well Chem. Phys. Lett. 742 137129 * (50) Wong A, Ulloa S E, Sandler N and Ingersent K 2016 Influence of Rashba spin-orbit coupling on the Kondo effect Phys. Rev. B 93 075148
11institutetext: Institute of Theoretical Astrophysics, University of Oslo, P.O. Box 1029 Blindern, N-0315 Oslo, Norway 22institutetext: Rosseland Centre for Solar Physics, University of Oslo, P.O. Box 1029 Blindern, N-0315 Oslo, Norway # Signatures of ubiquitous magnetic reconnection in the deep atmosphere of sunspot penumbrae Luc H. M. Rouppe van der Voort 1122 Jayant Joshi 1122 Vasco M. J. Henriques 1122 Souvik Bose 1122 ( - submitted to A&A December 18, 2020 / accepted January 26, 2021.) ###### Abstract Context. Ellerman bombs are regions with enhanced Balmer line wing emission and mark magnetic reconnection in the deep solar atmosphere in active regions and quiet Sun. They are often found in regions where opposite magnetic polarities are in close proximity. Recent high resolution observations suggest that Ellerman bombs are more prevalent than thought before. Aims. We aim to determine the occurrence of Ellerman bombs in the penumbra of sunspots. Methods. We analyze high spatial resolution observations of sunspots in the Balmer H $\alpha$ and H $\beta$ lines as well as auxiliary continuum channels obtained with the Swedish 1-m Solar Telescope and apply the $k$-means clustering technique to systematically detect and characterize Ellerman Bombs. Results. Features with all the defining characteristics of Ellerman bombs are found in large numbers over the entire penumbra. The true prevalence of these events is only fully appreciated in the H $\beta$ line due to highest spatial resolution and lower chromospheric opacity. We find that the penumbra hosts some of the highest Ellerman bomb densities, only surpassed by the moat in the immediate surroundings of the sunspot. Some penumbral Ellerman bombs show flame morphology and rapid dynamical evolution. Many penumbral Ellerman bombs are fast moving with typical speed of 3.7 km s-1 and sometimes more than 10 km s-1. Many penumbral Ellerman bombs migrate from the inner to the outer penumbra over hundreds of km and some continue moving beyond the outer penumbral boundary into the moat. Many penumbral Ellerman bombs are found in the vicinity of regions with opposite magnetic polarity. Conclusions. We conclude that reconnection is a near continuous process in the low atmosphere of the penumbra of sunspots as manifest in the form of penumbral Ellerman bombs. These are so prevalent that they may be a major sink of sunspot magnetic energy. ###### Key Words.: Sun: activity – Sun: atmosphere – Sun: magnetic fields – sunspots – Magnetic reconnection ## 1 Introduction Magnetic reconnection is a fundamental process in magnetized astrophysical plasmas for which magnetic energy is dissipated and converted into heat. In the lower solar atmosphere, the hydrogen Balmer lines provide effective tracers of reconnection sites as they exhibit remarkable enhanced emission in their extended line wings as result of localised heating. This phenomenon of enhanced wing emission, referred to as Ellerman “bombs” (EBs, Ellerman, 1917), is particularly pronounced in emerging active regions with vigorous magnetic flux emergence. At locations where opposite polarities are in close proximity (i.e., at the polarity inversion line), EBs appear as subarcsecond sized brightenings in H $\alpha$ line wing (see, e.g., Georgoulis et al., 2002; Pariat et al., 2004; Fang et al., 2006; Pariat et al., 2007; Matsumoto et al., 2008; Watanabe et al., 2008) and H $\beta$ line wing (Libbrecht et al., 2017; Joshi et al., 2020) images. The fact that the enhancement is only in the wings and that the EBs are invisible in the H $\alpha$ line-core locate the height of the reconnection below the chromospheric canopy of fibrils (Watanabe et al., 2011; Vissers et al., 2013; Nelson et al., 2013b). When observed from an inclined observing angle, sufficiently away from the center of the solar disk, and at sufficient spatial resolution, H $\alpha$ wing images show EBs as tiny (1–2 Mm), bright, upright flames that flicker rapidly on a time scale of seconds (Watanabe et al., 2011; Rutten et al., 2013; Nelson et al., 2015). There is considerable spread in EB lifetimes but they rarely live longer than a few minutes. We refer to Rutten et al. (2013) and Vissers et al. (2019) for recent reviews of observational EB properties and their visibility in different spectral diagnostics. Traditionally, EBs have been associated with strong magnetic field environments and therefore regarded as a typical active region phenomenon. This view changed when Rouppe van der Voort et al. (2016) and later Shetye et al. (2018) reported the existence of tiny ($\lesssim 0\aas@@fstack{\prime\prime}5$) Ellerman-like brightenings in quiet Sun when observed at extremely high spatial resolution. Nelson et al. (2017) found cases of quiet Sun EBs (QSEB) that were also visible in UV channels, suggesting that at least some QSEBs are energetic enough to become detectable in higher energy diagnostics. New high spatial resolution quiet Sun H $\beta$ observations presented by Joshi et al. (2020) show that QSEBs are much more ubiquitous than the lower spatial resolution H $\alpha$ observations suggested. The shorter wavelength H $\beta$ line allows for higher spatial resolution and higher temperature sensitivity and the observations suggest that about half a million QSEBs are present in the solar atmosphere at any time. The interpretation of EBs as markers of small-scale magnetic reconnection in the lower solar atmosphere has been reinforced by the advanced numerical simulations of Hansteen et al. (2017, 2019) and Danilovic (2017). In these simulations, heating occurs along current sheets that extend over several scale heights from the photosphere into the chromosphere. In synthetic H $\alpha$ wing images, these current sheets are at the core of flame like structures that resemble the characteristic EB flames in observations. The sunspot penumbra is another environment in the lower solar atmosphere where magnetic reconnection is likely to occur. In the penumbra, harboring an “uncombed” magnetic field topology with strong magnetic fields at highly variable inclination angles and considerable dynamic forcing from convective flows, one may arguably expect ample occurrences of magnetic fields with differing angles at sufficient close proximity to effectively interact and reconnect (for reviews on the sunspot magnetic structure with strong-field vertical spines and and weaker-field horizontal inter-spines, see e.g., Borrero & Ichimoto, 2011; Tiwari, 2017). Scharmer et al. (2013) detected small regions of opposite polarity in a sunspot penumbra (also see Ruiz Cobo & Asensio Ramos, 2013; Franz & Schlichenmaier, 2013), and that these regions harbor convective downflows. Tiwari et al. (2015) found ample regions with polarity opposite to the dominant sunspot polarity in a high quality Hinode SOT/SP map. Based on the experience that EBs are often found at the interface between photospheric opposite polarity patches, we searched for EB signatures in high quality H $\alpha$ and H $\beta$ sunspot observations. In particular, we concentrated on the presence of flames in limbward observations as the telltale EB signature. After close inspection of 13 datasets acquired over more than a decade of observation campaigns, we conclude that EBs are prevalent in sunspot penumbrae. The signature of penumbral EBs (PEBs) however is often subtle and requires excellent observing quality. The H $\beta$ line offers more clear detection in comparison to H $\alpha$, where the EB spectral signature is often hidden by dense superpenumbral filaments. In this paper, we present results from analysis of the best datasets. ## 2 Observations The observations were obtained with the Swedish 1-m Solar Telescope (SST, Scharmer et al., 2003a) on the island of La Palma, Spain. We used the CRisp Imaging SpectroPolarimeter (CRISP, Scharmer et al., 2008) and the CHROMospheric Imaging Spectrometer (CHROMIS) to perform imaging spectrometry in the H $\alpha$ and H $\beta$ spectral lines. We used the standard SST data reduction pipelines (de la Cruz Rodríguez et al., 2015; Löfdahl et al., 2018) to process the data. This includes image restoration with the multi-object multi-frame blind deconvolution (MOMFBD, van Noort et al., 2005) method and the procedure for consistency across narrowband channels of Henriques (2012). High image quality was further aided with the SST adaptive optics system (Scharmer et al., 2003b) which has a 85-electrode deformable mirror operating at 2 kHz. The data recorded during the best seeing conditions was observed on 22 September 2017. During the best periods, the Fried’s parameter $r_{0}$ was above 50 cm, with a maximum of 79 cm (for a discussion of measurements of $r_{0}$ by the SST adaptive optics system, see Scharmer et al., 2019). Unfortunately, the seeing was not consistently of high quality and the data set is not optimal for temporal evolution studies. Most of the analysis and data presented in Figs. 1–5 is based on the CHROMIS and CRISP spectral scans recorded at 10:00:48 UT. The target area was the main sunspot in AR12681 at $(X,Y)=(-749\arcsec,-296\arcsec)$, $\mu=\cos\theta=0.54$ with $\theta$ the observing angle. With CHROMIS, we sampled the H $\beta$ line at 32 positions between $\pm$1.37 Å with equidistant steps of 0.074 Å around the line core and sparser in the wings to avoid line blends. The time to complete a full H $\beta$ scan was 11.1 s. The CHROMIS data has a pixel scale of 0$\aas@@fstack{\prime\prime}$038 and the telescope diffraction limit ($\lambda/D$) is 0$\aas@@fstack{\prime\prime}$1 at $\lambda=4861$ Å. The CHROMIS instrument has an auxiliary wide-band (WB) channel that is equipped with a continuum filter which is centered at 4846 Å and has a full-width at half-maximum (FWHM) of the transmission profile of 6.5 Å. This filter covers a spectral region that is dominated by continuum and has relatively weak spectral lines (see Löfdahl et al., 2018, for a plot of the transmission profile in comparison with an atlas spectrum). With CRISP, we sampled the H $\alpha$ line at 32 positions between $\pm$1.85 Å from the line core with equidistant steps of 0.1 Å between $-1.6$ and $+1.3$ Å. In addition, CRISP was sampling the Fe i 6301 and 6302 Å line pair in spectropolarimetric mode, with 9 positions in Fe i 6301 and 6 positions in Fe i 6302, avoiding the telluric blend in the red wing. Furthermore, a continuum position was sampled between the two lines. The time to complete full scans of the H $\alpha$ and Fe i spectral lines was 19.1 s. The pixel scale of the CRISP data is 0$\aas@@fstack{\prime\prime}$058. The other dataset that we analyzed in detail was observed on 29 April 2016 and was centered on the main sunspot in AR12533 at $(X,Y)=(623\arcsec,8\arcsec)$, $\mu=0.75$. The seeing conditions were very good for the whole 1 h 30 m duration of the time series which started at 09:43:09 UT. The $r_{0}$ values were averaging at about 20 cm with peaks up to 30 cm. The online material includes movies of the temporal evolution of the sunspot. For these movies we have applied frame selection by rejecting 32 low quality images, this corresponds to 12% of the total of 267 time steps. The CRISP instrument was running a program with Ca ii 8542 Å spectropolarimetry and H $\alpha$ imaging spectrometry at a cadence of 20 s. The H $\alpha$ line was sampled at 15 positions between $\pm$1.5 Å with 0.2 Å steps between $\pm$1.2 Å. We compare H $\alpha$ wing images with images from the CRISP 8542 Å WB channel. For CRISP, the WB channel branches off after the prefilter so that contrary to CHROMIS, one cannot have imaging in a clean continuum band. The prefilter has a FWHM of 9.3 Å and is centered on the Ca ii 8542 Å line. The Ca ii 8542 Å spectra were not included in our analysis. This data was earlier analyzed by Drews & Rouppe van der Voort (2020) to study penumbral micro jets and the co-aligned SST and IRIS data were publicly released as described by Rouppe van der Voort et al. (2020). For the exploration of all data, verification of detected events, and the study and measurement of the dynamical evolution of PEBs in the 29 April 2016 data, we made use of CRISPEX (Vissers & Rouppe van der Voort, 2012), a widget- based graphical user interface for exploration of multi-dimensional data sets written in the Interactive Data Language (IDL). ## 3 Methods #### Inversions. We have performed Milne-Eddington (ME) inversions of the Fe i line pair observed on 22 September 2017 to infer the magnetic field vector utilizing a parallel C++/Python implementation111https://github.com/jaimedelacruz/pyMilne (de la Cruz Rodríguez, 2019). The magnetic field vector retrieved from the ME inversions are an average over the formation height of the Fe i line pair. For these lines, the response of Stokes profiles to magnetic field is maximum around optical depth 0.1 at 5000 Å in sunspot penumbrae (e.g., see Fig. 9 of Joshi et al., 2017). We resolved the 180° ambiguity in our magnetic field vector measurements using the acute angle method (Sakurai et al., 1985; Cuperman et al., 1992). The inferred magnetic field vector in the line-of-sight frame of reference is projected to the disk center coordinates where $B_{z}$ represents the magnetic field component normal to the solar surface and $B_{x}$ and $B_{y}$ are the two orthogonal components projected onto the solar surface. To better resolve opposite polarity patches in the penumbra, we corrected for stray light prior to the inversions. We assumed a Gaussian point spread function (PSF) with FWHM of 1$\aas@@fstack{\prime\prime}$2 and 45% stray light, following similar stray light corrections that were considered for CRISP/SST observations in earlier studies. For example, Scharmer et al. (2011) and Scharmer & Henriques (2012) compensated for stray light using a PSF with FWHM of 1$\aas@@fstack{\prime\prime}$2 and 56% stray light contribution. Joshi et al. (2011) assumed 35% stray light and a Gaussian PSF with FWHM of 1$\aas@@fstack{\prime\prime}$6\. Moreover, from a detailed analysis of solar granulation contrast, Scharmer et al. (2019) concluded that stray light at the SST comes mainly from small-angle scattering and that the wings of the uncorrected PSF do not extend beyond 2″. #### k-means clustering. We used the $k$-means clustering technique (Everitt, 1972) to identify EB spectra in the H $\beta$ spectral line observed on 22 September 2017. The $k$-means method is widely used for the characterization of a variety of solar phenomena and observations. Examples include the classification of Mg ii h and k line profiles observed with IRIS (Sainz Dalda et al., 2019), the identification of Mg ii h and k spectra in flares (Panos et al., 2018), and Ca ii K observations of on-disk spicules (Bose et al., 2019, 2021). Our approach for clustering the H $\beta$ spectra is very similar to that employed by Joshi et al. (2020) and Joshi & Rouppe van der Voort (2020) to identify QSEBs in their H $\beta$ observations. With the $k$-means method we divided H $\beta$ spectra into 100 clusters and each cluster is represented by the mean of all profiles in that cluster. This mean profile is referred to as representative profile (RP). Out of 100 RPs we found that 29 RPs show line wing enhancement that is characteristic of EBs. Of these 29 selected RPs with enhanced wings, 25 RPs have essentially an unaffected line core, while the rest show an intensity enhancement in the line core. Inclusion of the four RPs, that have an enhanced line core as EB profiles, is motivated by Joshi et al. (2020) who found that unlike typical H $\alpha$ EB profiles, some EBs can show raised intensity level even in the H $\beta$ line core. A detailed description of selected RPs with EB-like H $\beta$ spectral profiles is provided in Sect. 4. Based on spatial locations of selected RPs, we created a binary mask which was then used to perform two dimensional (2D) connected component labeling (Fiorio & Gustedt, 1996) that assigns a unique label for each isolated patch in the binary mask. We then used the labels to estimate their area, brightness enhancement and radial distance from the geometric center of the sunspot for each individual EB. A detailed statistical analysis of these parameters is presented in Sect. 4. ## 4 Results ### 4.1 PEB morphology and general appearance Figure 1: Limb-side part of the sunspot in AR12681 observed on 22 September 2017 in H $\beta$ and H $\alpha$ blue wing and CHROMIS WB 4846 Å. PEBs are visible as small bright features all over the penumbra, some with clear flame morphology sticking straight up between filaments. These PEBs are invisible in the continuum WB image. The direction to the nearest limb is approximately upward along the $y$-axis. The top image includes six squares labelled A–F that mark ROIs that are shown in detail in Fig. 2. An animation of this figure is available as online material at https://www.mn.uio.no/astro/english/people/aca/rouppe/movies/. This animation shows a spectral scan through the H $\beta$ and H $\alpha$ lines. Figure 2: Details of EBs in and outside the penumbra of the sunspot shown in Fig. 1 in 6 ROIs. The spatial $X,Y$ coordinates are at the same scale as Fig. 1. The top row of panels for each ROI shows H $\beta$ and H $\alpha$ blue wing and CHROMIS WB 4846 Å images. The bottom left panels show $\lambda x$-diagrams of the spectral profiles along the red dotted line in the panels above. The bottom right panel shows spectral profiles for H $\beta$ (solid black line) and H $\alpha$ (dashed line) from the position of the red cross in the top left panels. The thin gray profiles are reference spectral profiles averaged over an area outside the sunspot. The intensity scaling is normalized to the level of the far red wing of the reference profile. The red tickmark in the bottom row panels indicates the line position of the wing images in the top left. ROI A is centered on a strong EB outside the sunspot. ROI B is centered on an EB at the outer edge of the penumbra. All other examples are PEBs inside the penumbra. Figure 1 shows the limb-side part of the 22 September 2017 sunspot in the blue wings of H $\beta$ and H $\alpha$ as well as in CHROMIS WB 4846 Å. The offset from line core was chosen to be close to the maximum of the typical EB profile as to show EBs at highest contrast. Some prominent EBs are visible as pronounced flames in the moat around the sunspot outside the penumbra. As expected, the EBs are not visible in the continuum dominated WB image. Inside the penumbra, there are a large number of small bright features present in the Balmer wing images but clearest in the H $\beta$ wing image and not visible in the WB image. Some of these appear as small linear features sticking straight up from between the penumbral filaments, resembling the larger EB flames in the surrounding sunspot moat. The animation associated with Fig. 1 shows a spectral line scan through the H $\beta$ and H $\alpha$ lines for comparison. From the animation it is evident that in the penumbra the EB-like brightenings in the H $\beta$ wings also persist in and close to the line core wavelength positions. However, these compact brightenings in the H $\beta$ line core are absent in the H $\alpha$ line core which predominantly show chromospheric superpenumbral fibril structures. Figure 2 zooms in on 6 regions of interest (ROI). In the upper left, ROI A is centered on the most prominent EB in the FOV, with the telltale flame towering about 600 km above the intergranular lane from where it appears to emanate. The CHROMIS WB image shows no trace of the EB, only some striations in the background faculae, which are unrelated to the EB phenomenon. The $\lambda x$-diagrams and spectral panel show the well-known characteristic EB Balmer profile with enhanced wings and unaffected line core. The peak wing enhancement is more than 2 times the level of the reference profile which is averaged over a quiet region. The higher contrast and higher spatial resolution in the H $\beta$ data compared to H $\alpha$ is clear, for example from the fine structure and spatial variation in the $\lambda x$-diagram. The EB in ROI A serves as reference for the EBs presented in the other ROIs. In ROI B, a clear EB flame is located at the outer edge of the penumbra. The vertical extension of this flame has a length of about 450 km. The other four ROIs are all inside the penumbra and are centered on PEBs. Of these, ROI F is centered on the tallest flame which has a length of about 350 km. The H $\beta$ wing image shows clear substructure in the PEB while it is more an extended fuzzy feature in the H $\alpha$ wing image. For this case, the wing enhancement in the H $\beta$ profile is only slightly larger than in H $\alpha$. For the PEB examples in ROIs C and D, the differences in wing enhancement is larger, in particular in ROI C where the peak in wing enhancement is almost as high as for the large EB in ROI A. Flame morphology in ROI C might be difficult to discern because the PEB is aligned along the penumbral filaments that, in this part of the penumbra, are aligned in the direction of the nearest limb (i.e. along the line-of-sight). ROI E is centered on a PEB that has hardly enhanced wings in the profile plot but clearly shows a little flame in the H $\beta$ wing image and is unmistakably present in the $\lambda x$-diagram. While this PEB might be weak, its absence in the WB image is striking. This weak event is detected as a PEB with the $k$-means method. In all of these ROIs the penumbra in WB appears smoother than in the Balmer wing images. Particularly in the H $\beta$ wing there are many small bright features, like bright “crumbs”, scattered over the penumbra. Some of these are very bright and show the characteristic EB wing enhancement and are clear PEBs. Many others show only subtle wing enhancement but are notably absent in the WB image. To the left of the PEB, in ROI F, the red dashed line crosses some of these “crumbs” and the $\lambda x$-diagram shows wing enhancement when compared to their surroundings, but clearly not so much as the central PEB. Figure 3 shows all H $\beta$ RPs that have been identified as showing EB spectral signatures. Representative profiles 0–24 have profiles similar to typical H $\alpha$ EB profiles with enhanced wings and essentially unaffected line core whereas RPs 25–28 display intensity enhancement in the line core along with enhancement in the wings. Each detected EB in our dataset displays a combination of RPs plotted in Fig. 3. For example, the PEB shown in the ROI E of Fig. 2 is identified as a line core brightening and represented by a combination of RPs 27 and 28. Similarly, a part of the PEB in ROI C is identified as a line core brightening by RP 25. The rest of the EBs and PEBs in Fig. 2 predominantly exhibit wing intensity enhancement in combination with unaffected line cores and are clustered following RP 0–24. Besides the RPs, Fig. 3 shows a density distribution of all H $\beta$ profiles that are included in each cluster. The density distributions are narrow and centered around the RPs. However, in some clusters the farthest profile shows some significant deviation from the corresponding RP. For example, in clusters represented by RP 7 and 23, the farthest profiles have quite different shape as compared to their respective RPs. Nevertheless, these farthest profiles also show characteristic EB-like spectral profiles. Figure 3: Twenty nine representative profiles (RPs) from the $k$-means clustering of the H $\beta$ line that are identified as signature of EB. The black lines show RPs whereas shaded colored areas represent density distribution of H $\beta$ spectra within a cluster; darker shades indicate higher density. Within a particular cluster, the H $\beta$ profile that is farthest (measured in euclidean distance) from the corresponding RPs is shown by the black dotted line. As reference, the average quiet Sun profile (gray line) is plotted in each panel. RPs 0–24 show the typical EB-like H $\beta$ profiles, i.e., enhanced wings and unaffected line core, while RPs 25–28 display both an enhancement in the wings as well as in the line core. The parameter $n$ represents the number of pixels in a cluster as percentage of the total of $\sim 1.73\times 10^{6}$ pixels. ### 4.2 Magnetic field environment Figure 4: The location of PEBs compared to the vertical magnetic field $B_{z}$. The top left panel shows a split image of the sunspot observed on 22 September 2017 with the left part in the H $\beta$ blue wing at $-0.2$ Å offset and the right part at $-0.6$ Å. The blue contours indicate the radial distance $r/\mathrm{R_{spot}}$ to the umbral center that is marked with the blue cross. The contour for $r/\mathrm{R_{spot}}=1.00$ is the outer penumbra boundary, defined from the associated WB image. The top right panel shows the $B_{z}$ map, derived from ME inversions of the Fe i lines, scaled between $-400$ and $+1600$ G. Regions with artifacts due to the de-projection method are marked in green. The sets of panels at the bottom show four ROIs in H $\beta$ $-0.2$ Å, $-0.6$ Å, and $B_{z}$ respectively. Red contours outline PEBs detected through the $k$-means method. Figure 5: Distribution of EBs and their properties with respect to radial distance from the sunspot center (observed on 22 September 2017). The outer sunspot boundary is at $r/\mathrm{R}_{\mathrm{spot}}=1$ and is marked with the yellow vertical line, also see the contours in Fig. 4. The statistics are based on $k$-means detections, the total number of EB detections is 372, of which 108 are PEBs. The top panel shows the EB occurrence. The red curve shows the ratio of negative magnetic polarity flux relative to total absolute flux (the sunspot is dominated by positive polarity). The blue curve shows the fraction of pixels with negative polarity relative to all pixels with significant magnetic signal ($|B_{z}|>50$ G). The middle panels show the area of the EB detections. The bottom panels show the H $\beta$ wing brightness enhancement of the brightest pixel in the EB detection relative to the local background in a 100$\times$100 pixel area and excluding EB detection pixels. The brightness enhancement is relative to the outer most wavelength positions on both sides of the line center of the background H $\beta$ profile and on a scaling set by the normalized quiet Sun reference profile. For the two bottom rows, the right panels show occurrence histograms with the black line outlining the histograms for PEBs. The histogram bin size is 0.003 Mm2 in area and 0.12 in brightness enhancement. The grey lines in the left panels mark the average values for each radial distance $r/\mathrm{R}_{\mathrm{spot}}$. In order to study the occurrence of PEBs with respect to the magnetic field in the vicinity, we compare EB detections from the $k$-means method with the $B_{z}$ map derived from the Fe i lines. This is illustrated in Fig. 4. The sunspot is dominated by positive magnetic polarity but the $B_{z}$ map also shows many small isolated patches with significant opposite (negative) polarity within the outer penumbra boundary. We find that many PEBs are located in the vicinity of these opposite polarity patches. This can be seen at close look in the four ROIs in the bottom of Fig. 4. Red contours outline EB detections and there are some clear examples of PEBs that are located at or close to the interface where opposite polarities meet. We note however, that we also find PEBs located in unipolar regions, for example the PEB in the center of the lower left ROI. As mentioned before, PEB brightenings can also be visible in and close to the H $\beta$ line core, see the left H $\beta$ $-0.2$ Å part of Fig. 4 (top left), that displays numerous compact brightenings in the penumbra. A number of PEBs that can be seen in the H $\beta$ line core are shown in more detail in the ROIs presented in the bottom of Fig. 4. For example, in the upper-left ROI, two big PEBs at the sunspot boundary are visible in the wing as well as close to the line core. The PEBs at $(X,Y)=(41\aas@@fstack{\prime\prime}5,23\aas@@fstack{\prime\prime}6)$ and $(X,Y)=(41\aas@@fstack{\prime\prime}0,22\aas@@fstack{\prime\prime}8)$ in the upper-right ROI are predominantly visible at $-0.2$ Å while they have only subtle brightenings in the outer H $\beta$ line wing. The statistics shown in Fig. 5 provide a quantified context of the observation that PEBs are often found in the vicinity of opposite polarities: the top diagram shows that both the number of PEBs and the contribution from opposite polarity patches increase towards the outer penumbra. Both the relative area (blue curve) and opposite polarity flux (red curve) increase to more than 10% at the outer penumbra boundary. With the $k$-means clustering method, we detected a total of 372 EBs of which 108 are in the penumbra. We found no EBs in the umbra. In the inner penumbra, $0.5\leq r/\mathrm{R}_{\mathrm{spot}}\leq 0.75$, the number density of detected PEBs is 0.29 Mm-2, and the fraction of the total area covered by PEBs is 0.007 (i.e., the area filling factor). In the outer penumbra, $0.75<r/\mathrm{R}_{\mathrm{spot}}\leq 1$, the number density is 0.76 Mm-2, and the area filling factor 0.032. In the immediate surroundings of the sunspot, in the moat, $1<r/\mathrm{R}_{\mathrm{spot}}\leq 1.25$, the EB number density is 1.72 Mm-2, and the area filling factor 0.037. The number density of all 372 EBs detected over the full CHROMIS FOV is 0.27 Mm-2. For two other H $\beta$ spectral scans of this sunspot, recorded under less optimal seeing conditions, we find fewer but comparable numbers of EB detections: for a scan with quiet Sun granulation contrast 15.0%, there are 304 EB detections of which 90 PEBs, and for a scan with 14.6% contrast, there are 252 EBs of which 75 PEB (the contrast for the best scan is 15.7%). So about 30% of the EBs detected in the FOV are inside the penumbra. Figure 5 further provides statistics on the area and brightness enhancement of the EB detections. The largest PEBs are found towards the outer penumbra and PEBs do not stand out as being smaller or larger than EBs outside the sunspot. The mean area for PEBs is 0.039 Mm2 (standard deviation $\sigma=0.055$ Mm2) and for EBs outside the sunspot 0.022 Mm2 ($\sigma=0.041$ Mm2). The trend in the area distribution has a sharp cut off at 0.0037 Mm2 (five pixels) which is set by the spatial resolution. This suggests that there exist smaller EBs that are not resolved. Many small bright features in the H $\beta$ wings described as “crumbs” in Sect. 4.1 were not detected by the $k$-means method. In some cases where these features were detected, only a few of the brightest pixels were identified as PEBs and not the whole morphological structure. Thus, these detections also contribute to the population of PEBs with smallest areas. We excluded all EB events with area less than five pixels in our statistical analysis. Also in terms of wing brightness enhancement, as shown in the bottom of Fig. 5, PEBs do not stand out as compared to EBs. The average wing brightness enhancement for PEBs is 0.72 ($\sigma=0.33$) and for EBs outside the sunspot 0.78 ($\sigma=0.35$). Here, the H $\beta$ wing enhancement was measured against the average intensity of the outermost wavelength positions in the local background (over an area of $100\times 100$ pixels). The majority of the EBs, within the penumbra as well as in the surroundings of the sunspot, have a brightness enhancement between 0.5 and 1. However, some PEBs in the outer penumbra and some EBs in close proximity of the sunspot are brighter, and for some, the intensity enhancement relative to the local surroundings is larger than 2. The brightest EBs were classified as RP 0–2 (see Fig. 3) for which the blue wing is raised more than three times the level of the reference quiet Sun. ### 4.3 Temporal evolution Figure 6: Temporal evolution of PEBs in the sunspot in AR12533 observed on 29 April 2016. The top left image shows an H $\alpha$ blue wing image at $-0.8$ Å with three regions of interest (ROI) marked with labels A, B, and C. The temporal evolution for these ROIs is shown in the rows with smaller H $\alpha$ wing images where the time is marked in the top left. The bottom row of images shows the temporal evolution in ROI C in WB 8542 Å for comparison. The spacing between large tick marks in the ROI images is 1″. The contrast in the top left overview image is enhanced by applying a gamma correction with $\Gamma$=2, all other images have linear scaling on a common scale for each ROI. Three animations associated to this figure are available as online material: a movie of the sunspot in the H $\alpha$ blue wing as in the upper left panel, the corresponding movie in WB 8542 Å, and a combined movie showing the left part of the sunspot. See https://www.mn.uio.no/astro/english/people/aca/rouppe/movies/. The temporal evolution of PEBs was studied in the 29 April 2016 observations of the sunspot in AR12533, see Fig. 6. The online material includes a movie of the full 90 min sequence of the entire sunspot at $-0.8$ Å offset from H $\alpha$ line center, equivalent to the FOV shown in the upper left panel. The movie shows many small bright PEBs in the penumbra that generally move radially outward, away from the umbra. These are not visible in the reference WB 8542 Å movie. This difference in visibility is best seen in the 3rd movie that zooms in on the left part of the penumbra and combines the H $\alpha$ blue wing and WB 8542 Å, as well as a panel that blinks between these diagnostics. There are numerous examples of PEBs that originate in the penumbra, migrate outwards and eventually cross the outer penumbra boundary where they continue their outward migration in the sunspot moat flow. From inspection of the H $\alpha$ blue wing movie, it is clear that most of the PEBs are found in the outer regions of the penumbra. This is similar to what is described above for the 22 September 2017 sunspot and what was found from the $k$-means detections. We tracked 32 events to measure lifetimes, trajectories and velocities. These PEBs were selected on the basis of clear visibility throughout their lifetime and regarded as a representative sample. We measured PEB lifetimes ranging between 1 and 9 min, and an average lifetime of about 3 min. During their lifetime, these PEBs traveled distances ranging between 100 and 1640 km, with an average of about 650 km. They traveled at an average speed of 3.7 km s-1 and the maximum speed measured is almost 13 km s-1. These velocities are apparent motions and from these observations we cannot determine whether these are real plasma flows or result from a moving front of for example reconnection moving along a magnetic interface. Figure 6 shows the evolution of selected PEBs in sequences of small images for three ROIs. The three images for ROI A cover 2:43 min during which the PEB migrates over a distance of 340 km with an average speed of 2 km s-1. The PEB flares up for a duration of 102 s, with its brightest moment at 10:13:16 UT in the middle panel. The PEB in ROI B migrates over 365 km with an average speed of 1.4 km s-1. During its lifetime, the PEB splits into a number of bright substructures, this is visible in the 3rd panel. The sequence for ROI C covers 1:41 min of a total lifetime of the PEB of 5:04 min. This PEB shows a clear flame structure which is strikingly absent in the reference row of WB 8542 Å images. This flame appears to eject a small bright blob that is visible in the three last panels. This ejection can be followed for 2:22 min during which it moves at a maximum speed of almost 11 km s-1. The PEB itself appears to move at a maximum speed of almost 4 km s-1 while it moves at an average speed of 2 km s-1 during a migration over 650 km. The rapid variability we see for these selected examples and other PEBs in the time series is clearly limited by the temporal resolution. There often are significant variations in brightness and morphology (e.g., in the form of splitting, merging, and ejections) between subsequent time steps. This suggests that PEBs are changing on a shorter time scale than 20 s. ## 5 Discussion and conclusion Using high spatial resolution observations in the Balmer H $\alpha$ and H $\beta$ lines, we find large numbers of EBs in the penumbrae of sunspots. The EB nature of these penumbral events is established by 1: characteristic spectral profiles with often strongly enhanced wings, 2: flame morphology under slanted viewing angle, 3: rapid temporal variability in brightness and morphology, and 4: absence in concurrent continuum passbands. We find many small patches in the penumbra with characteristic EB wing enhancement and note that there is considerable spread in the level of enhancement: some reach the level of strong EBs traditionally found in active region flux emergence regions with wing enhancement well above twice that of quiet Sun, others have weak wing enhancement that is only discernible in contrast to weak background penumbral profiles. In the H $\beta$ line, we find that PEBs do not stand out in terms of area or wing brightness as compared to EBs in the surroundings of the sunspot. We do note however, that PEBs are easier to discern in H $\beta$ than in H $\alpha$. The shorter wavelength of H $\beta$ offers the advantage of higher spatial resolution and higher contrast. Furthermore, we observe that the sunspot in the H $\alpha$ line is much more dominated by dense chromospheric fibrils. It appears that the sunspot chromosphere has much less opacity in H $\beta$. Recently, from non-LTE radiative transfer calculations, Zhang (2020) concluded that H $\beta$ Stokes signals originate from the sunspot umbra photosphere. The difference between the H $\alpha$ and H $\beta$ lines is well illustrated by the line scan animation associated with Fig. 1 in the online material. At wing offsets that have highest EB contrast, the H $\alpha$ line is much more dominated by the chromospheric superpenumbra fibrils than the H $\beta$ line. These combined reasons make it more difficult to detect PEBs in H $\alpha$ and appreciate the ubiquity of PEBs. Here we present detailed analysis of two different sunspots, but we note that we observe large numbers of PEBs in at least 11 other sunspot datasets that we have acquired over the past ten years. In the 22 September 2017 dataset, we find more than 100 PEBs in the highest quality H $\beta$ line scan which corresponds to almost 30% of all detected EBs. The number density of PEBs is higher than the average number density of EBs over the whole FOV. It is only in the sunspot moat, just outside the penumbra, that the number density of EBs is higher than in the outer penumbra. In the moat we detect on average about 3 EBs per typical granule, considering an average area of a granule of 1.75 Mm2 (see Rincon & Rieutord, 2018). For the outer penumbra, we detect about 1.3 PEBs per typical granule area. In the quiet Sun, Joshi et al. (2020) found a QSEB number density of 0.09 Mm-2, which is more than 8 times lower than the number densities of PEBs (0.76 Mm-2) and 19 times lower than EBs in the moat (1.72 Mm-2). Many events show clear flame morphology under inclined viewing angle which underlines the similarity with EBs in active regions and quiet Sun. The rapid variability and dynamics we see in the 29 April 2016 time series remind of the rapid variability found in EB flames in active regions (Watanabe et al., 2011) and quiet Sun (Rouppe van der Voort et al., 2016). We note however, that the temporal cadence of these studies ($\sim 1$ s) is much faster than for the time series presented here (20 s). Establishing the ubiquity of EBs in the penumbra is aided by the concurrent continuum observations that are available through the CHROMIS WB channel. Absence of a bright feature in the associated continuum image confirms the Balmer line wing enhancement. The EB features are as absent in the penumbra as they are outside the sunspot and this further confirms the EB nature of PEBs. The H $\beta$ wing images show many small bright features that are absent in the WB image but have too weak wing enhancement to be (fully) detected by the $k$-means method (in Sect. 4.1 described as “crumbs”). This suggests that PEBs are more prevalent than the detection numbers from the $k$-means method suggest. For CRISP H $\alpha$ observations, a clean continuum channel is not as readily available as the CRISP WB channel shares the same prefilter as the CRISP narrow band images. CRISP WB 6563 Å images show EBs because the CRISP prefilter transmission profile has relatively wide passband (FWHM=4.9 Å) and is centered on the H $\alpha$ line. For the 29 April 2016 time series (see Fig. 6) we compare H $\alpha$ blue wing images with concurrent CRISP WB 8542 Å images since EB signature in this channel is weaker due to wider passband and generally weaker EB emission in Ca ii 8542 Å. The presence of EBs in the sunspot penumbra has been reported before. For example, a number of small EBs inside the penumbra of a small sunspot can be seen in the H $\alpha$ wing detection maps of Nelson et al. (2013a), and Reardon et al. (2013) report the observation of EB profiles in the Ca ii 8542 Å line for two events in a study of penumbral transients. However, this is the first time that the presence of large numbers of EBs in the penumbra is reported. The significance of EBs lies in their capacity of being markers of magnetic reconnection in the low solar atmosphere. Numerical simulations demonstrated that enhanced Balmer wing emission and flame morphology stems from heating along current sheets at reconnection sites (Hansteen et al., 2017, 2019; Danilovic, 2017). The flames we observe for PEBs appear to be rooted deep down in the penumbra photosphere, in similar fashion as for EB flames in active regions and quiet Sun. Further support for PEBs being markers of magnetic reconnection in the deep penumbra photosphere comes from PEB detections being located in areas where opposite polarities are in close proximity. The sunspot of 22 September 2017 is of positive magnetic polarity. The $B_{z}$ map (Fig. 4) reveals the presence of many isolated patches of opposite (negative) polarity within the penumbra. Many PEBs are located in the vicinity of these opposite polarity patches and some are located right at the interface where the two magnetic polarities meet. We also observe that the number density of PEBs increases toward the outer penumbra, following the same trend of increasing opposite polarity flux with increasing distance from the sunspot umbra. We note however, that there are a few limitations that need to be kept in mind when combining the $B_{z}$ and EB detection maps for inferring that magnetic reconnection is taking place: spectral line inversions are sensitive to a limited range in height and simplifications assumed for the ME inversion method imply uncertainties. We estimate that our ME inversions of the Fe i lines are valid as $\@vec{B}$ field measurements over a height range over a few 100 km in the upper photosphere (see Grec et al., 2010). Joshi et al. (2017) have shown that opposite polarity magnetic flux found in the deeper penumbra could be more than four times larger than that in the the middle and upper photosphere. Therefore, there is solid ground to believe that our ME inversions which provide height-independent magnetic field vectors are not able to resolve all opposite polarity patches in the penumbra. Furthermore, stray light makes it difficult to detect weak signals and adds to the uncertainty in the interpretation. We have applied a correction for stray light that is consistent with previous studies but the full impact of stray light on our measurements remains unknown. Further uncertainties come from line-of-sight obscuration due to corrugation of the penumbral optical surface and it may be possible that regions with opposite polarity are hidden behind elevated foreground structures. Apart from these observational limitations that mitigate the detection of opposite polarity patches, it should be stressed that the condition of diametrically opposite direction fields is not strictly required for reconnection to take place. Even in areas that appear unipolar in observations, the complex magnetic topology of the penumbra can be expected to host gradients in the magnetic field that allow for small-angle magnetic reconnection. The large number of PEBs we observe suggests that magnetic reconnection is a very frequently occurring process in the low penumbra atmosphere. A significant amount of magnetic energy may be dissipated through reconnection in the highly abundant PEBs and as such PEBs may play an important role in sunspot decay. Outward moving magnetic elements that leave the penumbra and migrate through the sunspot moat, commonly referred to as moving magnetic features (MMF), carry net flux away from the sunspot and are traditionally regarded as main actors in sunspot decay (see, e.g. Solanki, 2003). The ubiquity of PEBs we find here may implicate that some fraction of magnetic energy is already dissipated and lost from the sunspot before MMFs cross the sunspot boundary. Moreover, high density of EBs in the immediate vicinity of the sunspot suggest that significant fraction of magnetic field in the moat flow regions might also dissipate through magnetic field reconnection occurring in the photosphere. What impact do PEBs have on the upper atmosphere? There exist several transient phenomena in sunspots that may be related to energy release in PEBs. Penumbral micro-jets (PMJ) are short-lived elongated brightenings that can be observed in the core of Ca ii lines (Katsukawa et al., 2007; Reardon et al., 2013). Magnetic reconnection has been suggested as their driver but the idea that they carry high-speed plasma flows as their name suggests has been contested (Esteban Pozuelo et al., 2019; Rouppe van der Voort & Drews, 2019). They can be observed in transition region diagnostics (Vissers et al., 2015; Drews & Rouppe van der Voort, 2020) and Tiwari et al. (2016) reported the existence of large PMJs originating from the outer penumbra in the regions with abundant mixed polarities. These large PMJs leave signatures in some of the transition region/coronal channels of the AIA instrument of NASA’s Solar Dynamics Observatory Drews & Rouppe van der Voort (2017) found that there exist on average 21 PMJs per time step in a time series of Ca ii 8542 Å observations. This is significantly fewer than the number of PEBs that we detect. Furthermore, clear PMJ detections are mostly found in the inner penumbra where we find fewer PEBs as compared to the outer penumbra. Recently, Buehler et al. (2019) and Drews & Rouppe van der Voort (2020) connected Ca ii 8542 Å PMJs with dark fibrilar structures close to the line core in H $\alpha$. The connection between PEBs and PMJs warrants further study and requires simultaneous observation of multiple spectral lines and extreme high temporal evolution to resolve the onset of PMJs (Rouppe van der Voort & Drews, 2019). Possibly there is also a connection with transition region bright dots observed above sunspots with IRIS (Tian et al., 2014; Samanta et al., 2017) and Hi-C (Alpert et al., 2016). Furthermore, magnetic reconnection in the deep atmosphere as marked by PEBs may play a role in the heating of bright coronal loops that are rooted in penumbrae (see, e.g., Tiwari et al., 2017) Finally, we conclude that EBs in the penumbra of sunspots are an excellent target for new telescopes such as the 4-m DKIST (Rimmele et al., 2020) and the planned EST (Schlichenmaier et al., 2019) since PEBs offer opportunities to study magnetic reconnection in kG magnetic field environments at the smallest resolvable scales in astrophysical plasmas. ###### Acknowledgements. The Swedish 1-m Solar Telescope is operated on the island of La Palma by the Institute for Solar Physics of Stockholm University in the Spanish Observatorio del Roque de los Muchachos of the Instituto de Astrofísica de Canarias. The Institute for Solar Physics is supported by a grant for research infrastructures of national importance from the Swedish Research Council (registration number 2017-00625). This research is supported by the Research Council of Norway, project number 250810, and through its Centres of Excellence scheme, project number 262622. VMJH is supported by the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (SolarALMA, grant agreement No. 682462). We thank Shahin Jafarzadeh, Ainar Drews, Tiago Pereira and Ada Ortiz for their help with the observations. We made much use of NASA’s Astrophysics Data System Bibliographic Services. ## References * Alpert et al. (2016) Alpert, S. E., Tiwari, S. K., Moore, R. L., Winebarger, A. R., & Savage, S. L. 2016, ApJ, 822, 35 * Borrero & Ichimoto (2011) Borrero, J. M. & Ichimoto, K. 2011, Living Reviews in Solar Physics, 8, 4 * Bose et al. (2019) Bose, S., Henriques, V. M. J., Joshi, J., & Rouppe van der Voort, L. 2019, A&A, 631, L5 * Bose et al. (2021) Bose, S., Joshi, J., Henriques, V. M. J., & Rouppe van der Voort, L. 2021, arXiv e-prints, arXiv:2101.07829 * Buehler et al. (2019) Buehler, D., Esteban Pozuelo, S., de la Cruz Rodriguez, J., & Scharmer, G. B. 2019, ApJ, 876, 47 * Cuperman et al. (1992) Cuperman, S., Li, J., & Semel, M. 1992, A&A, 265, 296 * Danilovic (2017) Danilovic, S. 2017, A&A, 601, A122 * de la Cruz Rodríguez (2019) de la Cruz Rodríguez, J. 2019, A&A, 631, A153 * de la Cruz Rodríguez et al. (2015) de la Cruz Rodríguez, J., Löfdahl, M. G., Sütterlin, P., Hillberg, T., & Rouppe van der Voort, L. 2015, A&A, 573, A40 * Drews & Rouppe van der Voort (2017) Drews, A. & Rouppe van der Voort, L. 2017, A&A, 602, A80 * Drews & Rouppe van der Voort (2020) Drews, A. & Rouppe van der Voort, L. 2020, A&A, 638, A63 * Ellerman (1917) Ellerman, F. 1917, ApJ, 46, 298 * Esteban Pozuelo et al. (2019) Esteban Pozuelo, S., de la Cruz Rodríguez, J., Drews, A., et al. 2019, ApJ, 870, 88 * Everitt (1972) Everitt, B. S. 1972, British Journal of Psychiatry, 120, 143–145 * Fang et al. (2006) Fang, C., Tang, Y. H., Xu, Z., Ding, M. D., & Chen, P. F. 2006, ApJ, 643, 1325 * Fiorio & Gustedt (1996) Fiorio, C. & Gustedt, J. 1996, Theoretical Computer Science, 154, 165 * Franz & Schlichenmaier (2013) Franz, M. & Schlichenmaier, R. 2013, A&A, 550, A97 * Georgoulis et al. (2002) Georgoulis, M. K., Rust, D. M., Bernasconi, P. N., & Schmieder, B. 2002, ApJ, 575, 506 * Grec et al. (2010) Grec, C., Uitenbroek, H., Faurobert, M., & Aime, C. 2010, A&A, 514, A91 * Hansteen et al. (2019) Hansteen, V., Ortiz, A., Archontis, V., et al. 2019, A&A, 626, A33 * Hansteen et al. (2017) Hansteen, V. H., Archontis, V., Pereira, T. M. D., et al. 2017, ApJ, 839, 22 * Henriques (2012) Henriques, V. M. J. 2012, A&A, 548, A114 * Joshi et al. (2017) Joshi, J., Lagg, A., Hirzberger, J., Solanki, S. K., & Tiwari, S. K. 2017, A&A, 599, A35 * Joshi et al. (2011) Joshi, J., Pietarila, A., Hirzberger, J., et al. 2011, ApJ, 734, L18 * Joshi & Rouppe van der Voort (2020) Joshi, J. & Rouppe van der Voort, L. H. M. 2020, (in preparation) * Joshi et al. (2020) Joshi, J., Rouppe van der Voort, L. H. M., & de la Cruz Rodríguez, J. 2020, A&A, 641, L5 * Katsukawa et al. (2007) Katsukawa, Y., Berger, T. E., Ichimoto, K., et al. 2007, Science, 318, 1594 * Libbrecht et al. (2017) Libbrecht, T., Joshi, J., Rodríguez, J. d. l. C., Leenaarts, J., & Ramos, A. A. 2017, A&A, 598, A33 * Löfdahl et al. (2018) Löfdahl, M. G., Hillberg, T., de la Cruz Rodriguez, J., et al. 2018, ArXiv e-prints 1804.03030 * Matsumoto et al. (2008) Matsumoto, T., Kitai, R., Shibata, K., et al. 2008, PASJ, 60, 577 * Nelson et al. (2013a) Nelson, C. J., Doyle, J. G., Erdélyi, R., et al. 2013a, Sol. Phys., 283, 307 * Nelson et al. (2017) Nelson, C. J., Freij, N., Reid, A., et al. 2017, ApJ, 845, 16 * Nelson et al. (2015) Nelson, C. J., Scullion, E. M., Doyle, J. G., Freij, N., & Erdélyi, R. 2015, ApJ, 798, 19 * Nelson et al. (2013b) Nelson, C. J., Shelyag, S., Mathioudakis, M., et al. 2013b, ApJ, 779, 125 * Panos et al. (2018) Panos, B., Kleint, L., Huwyler, C., et al. 2018, ApJ, 861, 62 * Pariat et al. (2004) Pariat, E., Aulanier, G., Schmieder, B., et al. 2004, ApJ, 614, 1099 * Pariat et al. (2007) Pariat, E., Schmieder, B., Berlicki, A., et al. 2007, A&A, 473, 279 * Reardon et al. (2013) Reardon, K., Tritschler, A., & Katsukawa, Y. 2013, ApJ, 779, 143 * Rimmele et al. (2020) Rimmele, T. R., Warner, M., Keil, S. L., et al. 2020, Sol. Phys., 295, 172 * Rincon & Rieutord (2018) Rincon, F. & Rieutord, M. 2018, Living Reviews in Solar Physics, 15, 6 * Rouppe van der Voort et al. (2020) Rouppe van der Voort, L. H. M., De Pontieu, B., Carlsson, M., et al. 2020, A&A, 641, A146 * Rouppe van der Voort & Drews (2019) Rouppe van der Voort, L. H. M. & Drews, A. 2019, A&A, 626, A62 * Rouppe van der Voort et al. (2016) Rouppe van der Voort, L. H. M., Rutten, R. J., & Vissers, G. J. M. 2016, A&A, 592, A100 * Ruiz Cobo & Asensio Ramos (2013) Ruiz Cobo, B. & Asensio Ramos, A. 2013, A&A, 549, L4 * Rutten et al. (2013) Rutten, R. J., Vissers, G. J. M., Rouppe van der Voort, L. H. M., Sütterlin, P., & Vitas, N. 2013, in Journal of Physics Conference Series, Vol. 440, Journal of Physics Conference Series, 012007 * Sainz Dalda et al. (2019) Sainz Dalda, A., de la Cruz Rodríguez, J., De Pontieu, B., & Gošić, M. 2019, ApJ, 875, L18 * Sakurai et al. (1985) Sakurai, T., Makita, M., & Shibasaki, K. 1985, in Theo. Prob. High Resolution Solar Physics, ed. H. U. Schmidt, 313 * Samanta et al. (2017) Samanta, T., Tian, H., Banerjee, D., & Schanche, N. 2017, ApJ, 835, L19 * Scharmer et al. (2003a) Scharmer, G. B., Bjelksjö, K., Korhonen, T. K., Lindberg, B., & Petterson, B. 2003a, in Proc. SPIE, Vol. 4853, Innovative Telescopes and Instrumentation for Solar Astrophysics, ed. S. L. Keil & S. V. Avakyan, 341–350 * Scharmer et al. (2013) Scharmer, G. B., de la Cruz Rodriguez, J., Sütterlin, P., & Henriques, V. M. J. 2013, A&A, 553, A63 * Scharmer et al. (2003b) Scharmer, G. B., Dettori, P. M., Löfdahl, M. G., & Shand, M. 2003b, in Proc. SPIE, Vol. 4853, Innovative Telescopes and Instrumentation for Solar Astrophysics, ed. S. L. Keil & S. V. Avakyan, 370–380 * Scharmer & Henriques (2012) Scharmer, G. B. & Henriques, V. M. J. 2012, A&A, 540, A19 * Scharmer et al. (2011) Scharmer, G. B., Henriques, V. M. J., Kiselman, D., & de la Cruz Rodríguez, J. 2011, Science, 333, 316 * Scharmer et al. (2019) Scharmer, G. B., Löfdahl, M. G., Sliepen, G., & de la Cruz Rodríguez, J. 2019, A&A, 626, A55 * Scharmer et al. (2008) Scharmer, G. B., Narayan, G., Hillberg, T., et al. 2008, ApJ, 689, L69 * Schlichenmaier et al. (2019) Schlichenmaier, R., Bellot Rubio, L. R., Collados, M., et al. 2019, arXiv e-prints, arXiv:1912.08650 * Shetye et al. (2018) Shetye, J., Shelyag, S., Reid, A. L., et al. 2018, MNRAS, 479, 3274 * Solanki (2003) Solanki, S. K. 2003, A&A Rev., 11, 153 * Tian et al. (2014) Tian, H., Kleint, L., Peter, H., et al. 2014, ApJ, 790, L29 * Tiwari (2017) Tiwari, S. K. 2017, ArXiv e-prints 1712.07174 * Tiwari et al. (2016) Tiwari, S. K., Moore, R. L., Winebarger, A. R., & Alpert, S. E. 2016, ApJ, 816, 92 * Tiwari et al. (2017) Tiwari, S. K., Thalmann, J. K., Panesar, N. K., Moore, R. L., & Winebarger, A. R. 2017, ApJ, 843, L20 * Tiwari et al. (2015) Tiwari, S. K., van Noort, M., Solanki, S. K., & Lagg, A. 2015, A&A, 583, A119 * van Noort et al. (2005) van Noort, M., Rouppe van der Voort, L., & Löfdahl, M. G. 2005, Sol. Phys., 228, 191 * Vissers & Rouppe van der Voort (2012) Vissers, G. & Rouppe van der Voort, L. 2012, ApJ, 750, 22 * Vissers et al. (2015) Vissers, G. J. M., Rouppe van der Voort, L. H. M., & Carlsson, M. 2015, ApJ, 811, L33 * Vissers et al. (2013) Vissers, G. J. M., Rouppe van der Voort, L. H. M., & Rutten, R. J. 2013, ApJ, 774, 32 * Vissers et al. (2019) Vissers, G. J. M., Rouppe van der Voort, L. H. M., & Rutten, R. J. 2019, A&A, 626, A4 * Watanabe et al. (2008) Watanabe, H., Kitai, R., Okamoto, K., et al. 2008, ApJ, 684, 736 * Watanabe et al. (2011) Watanabe, H., Vissers, G., Kitai, R., Rouppe van der Voort, L., & Rutten, R. J. 2011, ApJ, 736, 71 * Zhang (2020) Zhang, H. 2020, Science China Physics, Mechanics, and Astronomy, 63, 119611
# Minimum energy with infinite horizon: from stationary to non-stationary states P. Acquistapace111Dipartimento di Matematica, Università di Pisa, e-mail: <EMAIL_ADDRESS>F.Gozzi222Dipartimento di Economia e Finanza, Università _LUISS - Guido Carli_ Roma; e-mail<EMAIL_ADDRESS> ###### Abstract We study a non standard infinite horizon, infinite dimensional linear- quadratic control problem arising in the physics of non-stationary states (see e.g. [7, 9]): finding the minimum energy to drive a given stationary state $\bar{x}=0$ (at time $t=-\infty$) into an arbitrary non-stationary state $x$ (at time $t=0$). This is the opposite to what is commonly studied in the literature on null controllability (where one drives a generic state $x$ into the equilibrium state $\bar{x}=0$). Consequently, the Algebraic Riccati Equation (ARE) associated to this problem is non-standard since the sign of the linear part is opposite to the usual one and since it is intrinsically unbounded. Hence the standard theory of AREs does not apply. The analogous finite horizon problem has been studied in the companion paper [1]. Here, similarly to such paper, we prove that the linear selfadjoint operator associated to the value function is a solution of the above mentioned ARE. Moreover, differently to [1], we prove that such solution is the maximal one. The first main result (Theorem 4.7) is proved by approximating the problem with suitable auxiliary finite horizon problems (which are different from the one studied in [1]). Finally in the special case where the involved operators commute we characterize all solutions of the ARE (Theorem 5.5) and we apply this to the Landau-Ginzburg model. Keywords: Minimum energy; Null controllability; Landau-Ginzburg model; Optimal control with infinite horizon; Algebraic Riccati Equation in infinite dimension; Value function as maximal solution. ###### Contents 1. 1 Introduction 1. 1.1 Plan of the paper 2. 2 The problem and the main results 1. 2.1 The state equation 2. 2.2 Minimum energy problems with infinite horizon and associated Riccati equation 3. 2.3 The method and the main results 3. 3 The auxiliary problem 1. 3.1 A key comparison result 4. 4 Minimum energy with (negative) infinite horizon 1. 4.1 Optimal strategies 2. 4.2 Connection with the finite horizon case 3. 4.3 Algebraic Riccati Equation 5. 5 The selfadjoint commuting case 6. 6 A motivating example: from equilibrium to non-equilibrium states 7. A Minimum Energy with finite horizon 1. A.1 General formulation of the problem 2. A.2 The space $H$ and its properties ## 1 Introduction We study a non standard infinite dimensional, infinite horizon, linear- quadratic control problem: finding the minimum energy to drive a given stationary state $\bar{x}=0$ (at time $t=-\infty$) into an arbitrary non- stationary state $x$ (at time $t=0$). This kind of problems arises in the control representation of the rate function for a class of large deviation problems (see e.g. [13] and the references quoted therein; see also [18, Chapter 8] for an introduction to the subject). It is motivated by applications in the physics of non-equilibrium states and in this context it has been studied in various papers, see e.g. [4, 5, 6, 7, 8, 9] (see Section 6 for a description of a model case). The main goal here, as a departure point of the theory, is to apply the dynamic programming approach to characterize the value function as the unique (or maximal/minimal) solution of the associated Hamilton-Jacobi-Bellman (HJB) equation, a problem left open e.g. in [7, 9]. This problem is quite difficult since it deals with the opposite to what is commonly studied in the literature on null controllability (where one drives a generic state $x$ into the equilibrium state $\bar{x}=0$). For this reason we start studying here the simplest case, i.e. when the state equation is linear and the energy functional is purely quadratic: so the problem falls into the class of linear- quadratic optimal control problems, the value function is quadratic, and the associated HJB equation reduces to an Algebraic Riccati Equation (ARE). The above feature (i.e. the fact we bring $0$ to $x$ instead of the opposite) implies that the ARE associated to this problem is non-standard for two main reasons: first, the sign of the linear part is opposite to the usual one; second, since the set of reachable $x$ is strictly smaller than the whole state space $X$, the solution is intrinsically unbounded in $X$. The combination of these two difficulties does not allow to apply the standard theory of AREs. In the companion paper [1] we studied, as a first step, the associated finite horizon case. Here we partially exploit the results of such paper to deal with the more interesting infinite horizon case, which is the one that arises in the above mentioned papers in physics. Our main results (Theorems 4.7 and 5.5) show that, under a null controllability assumption (after a given time $T_{0}\geq 0$) and a coercivity assumption on the control operator, the linear selfadjoint operator $P$ associated to the value function is the maximal solution of the above mentioned ARE. The first result concerns the general case with some restrictions on the class of solutions, while the second one looks at the case where the state and the control operators commute, without any restriction on the class of solutions. This is only partially similar to what has been done in [1]. Indeed, the proof that $P$ is a solution of ARE is substantially similar to what is done in [1, Section 4.3]. On the other hand, while in [1, Section 4.4] we prove a partial uniqueness result (i.e. uniqueness in a suitable family of invertible operators), here we are able to prove, through a delicate comparison argument (based on a nontrivial approximation procedure), that $P$ is the maximal solution of the associated ARE. To prove the comparison argument (which is the content of the key Lemma 3.10) we need to introduce a family of auxiliary finite horizon problems, which are different from the one studied in [1]. Finally, in the special case where the involved operators commute, we are able, again differently from the finite horizon case, to characterize all solutions of the ARE. This allows to apply our result to the case of Landau- Ginzburg model. ### 1.1 Plan of the paper In Section 2 we illustrate the problem and the strategy to show the main results. It is divided in three subsections: in the first we present the state equation and the main Hypothesis; in the second one we describe our minimum energy problem; the third subsection briefly explains the method used to prove our main results. Section 3 concerns the study of the auxiliary problem. After devoting the first part of the section to some basic results on it, we show, in Subsection 3.1, the comparison Lemma 3.10 which will be used to prove the maximality result in the infinite horizon case. Section 4 is devoted to the main problem and the main maximality result. In Section 5 we analyze the case when the operators $A$ and $BB^{*}$ commute. In Section 6 we present, as an example, a special case of the motivating problem given in [7] (the case of the so-called Landau-Ginzburg model): we show that it falls into the class of problems treated in this paper. ## 2 The problem and the main results ### 2.1 The state equation ###### Notation 2.1. Given any two Banach spaces $Y$ and $Z$, we denote by ${\cal L}(Y,Z)$ the set of all linear bounded operators from $Y$ to $Z$, writing ${\cal L}(Y)$ when $Z=Y$. When $Y$ is a Hilbert space we denote by ${\cal L}_{+}(Y)$ the set of all elements of ${\cal L}(Y)$ which are selfadjoint and nonnegative. Let $-\infty<s<t<+\infty$. Consider the abstract linear equation $\left\\{\begin{array}[]{l}y^{\prime}(r)=Ay(r)+Bu(r),\quad r\in\,]s,t],\\\\[5.69054pt] y(s)=x\in X,\end{array}\right.$ (1) under the following assumption. ###### Hypothesis 2.2. (i) $X$, the state space, and $U$, the control space, are real separable Hilbert spaces; (ii) $A:{\cal D}(A)\subseteq X\rightarrow X$ is the generator of a $C_{0}$-semigroup on $X$ such that $\|e^{tA}\|_{{\cal L}(X)}\leq Me^{-\omega t},\qquad t\geq 0,$ (2) for given constants $M>0$ and $\omega>0$; (iii) $B:U\rightarrow X$ is a bounded linear operator; (iv) $u$, the control strategy, belongs to $L^{2}(s,t;U)$. We recall the following well known result, pointed out e.g. in [1, Proposition 2.2]. ###### Proposition 2.3. For $-\infty<s<t<+\infty$, $x\in X$ and $u\in L^{2}(s,t;U)$, the mild solution of (1), defined by $y(r;s,x,u)=e^{(r-s)A}x+\int_{s}^{r}e^{(r-\sigma)A}Bu(\sigma)\,\mathrm{d}\sigma,\quad r\in[s,t],$ (3) is in $C([s,t],X)$. We now consider the state equation in the half-line $\,]-\infty,t]$: $\left\\{\begin{array}[]{l}y^{\prime}(r)=Ay(r)+Bu(r),\quad r\in\,]-\infty,t],\\\\[5.69054pt] \displaystyle\lim_{s\rightarrow-\infty}y(s)=0.\end{array}\right.$ (4) Since (4) is not completely standard we introduce the following definition of solution. ###### Definition 2.4. Given $u\in L^{2}(-\infty,t;U)$, we say that $y\in C(\,]-\infty,t];X)$ is a solution of (4) if for every $-\infty<r_{1}\leq r_{2}\leq t$ we have $y(r_{2})=e^{(r_{2}-r_{1})A}y(r_{1})+\int_{r_{1}}^{r_{2}}e^{(r_{2}-\tau)A}Bu(\tau)\mathrm{d}\tau.$ (5) and $\lim_{s\rightarrow-\infty}y(s)=0.$ (6) ###### Lemma 2.5. Given any $u\in L^{2}(-\infty,t;U)$, there exists a unique solution of the Cauchy problem (4) and it is given by $y(r;-\infty,0,u):=\int_{-\infty}^{r}e^{(r-\tau)A}Bu(\tau)\,\mathrm{d}\tau,\qquad r\leq t.$ (7) ###### Proof. We prove first that the function $y(\cdot;-\infty,0,u)$ given by (7) is continuous. Fixed $r_{1}<r_{2}\leq t$, we have $\displaystyle y(r_{2};-\infty,0,u)-y(r_{1},-\infty,0,u)=$ $\displaystyle=\int_{-\infty}^{r_{2}}e^{(r_{2}-\tau)A}Bu(\tau)\,\mathrm{d}\tau-\int_{-\infty}^{r_{1}}e^{(r_{1}-\tau)A}Bu(\tau)\,\mathrm{d}\tau=$ $\displaystyle=\int_{-\infty}^{r_{1}}\left(e^{(r_{2}-r_{1})A}-I\right)e^{(r_{1}-\tau)A}Bu(\tau)\,\mathrm{d}\tau+\int_{r_{1}}^{r_{2}}e^{(r_{2}-\tau)A}Bu(\tau)\,\mathrm{d}\tau,$ and then continuity follows by standard arguments. We now prove that (5) holds. For $-\infty<r_{1}\leq r_{2}\leq t$, we have $\displaystyle y(r_{2};-\infty,0,u)$ $\displaystyle=$ $\displaystyle\int_{-\infty}^{r_{2}}e^{(r_{2}-\tau)A}Bu(\tau)\,\mathrm{d}\tau=$ $\displaystyle=$ $\displaystyle e^{(r_{2}-r_{1})A}\int_{-\infty}^{r_{1}}e^{(r_{1}-\tau)A}Bu(\tau)\,\mathrm{d}\tau+\int_{r_{1}}^{r_{2}}e^{(r_{2}-\tau)A}Bu(\tau)\,\mathrm{d}\tau=$ $\displaystyle=$ $\displaystyle e^{(r_{2}-r_{1})A}y(r_{1};-\infty,0,u)+\int_{r_{1}}^{r_{2}}e^{(r_{2}-\tau)A}Bu(\tau)\,\mathrm{d}\tau,$ so (5) is satisfied. Moreover letting $r\rightarrow-\infty$, since $u\in L^{2}(-\infty,t;U)$ and thanks to equation (2), we have $y(r;-\infty,x,u)\rightarrow 0$ as ${r\rightarrow-\infty}$. In order to prove uniqueness, consider two solutions $y_{1}(\cdot)$ and $y_{2}(\cdot)$ and a point $r\in(-\infty,t)$. Since $y_{1}(\cdot)$ and $y_{2}(\cdot)$ satisfy (5), for their difference we have, for $r_{0}<r<t$, $\|y_{1}(r)-y_{2}(r)\|_{X}=\|e^{(r-r_{0})A}(y_{1}(r_{0})-y_{2}(r_{0}))\|_{X}\leq M\,e^{-(r-r_{0})\omega}\|(y_{1}(r_{0})-y_{2}(r_{0}))\|_{X}\,.$ As $y_{1}(\cdot)$ and $y_{2}(\cdot)$ satisfy (6), letting $r_{0}\rightarrow-\infty$ above we get $y_{1}(r)=y_{2}(r)$ for every $r<t$. ∎ ###### Remark 2.6. Notice that, if the initial condition (6) is not zero, then the above equation cannot have any solution. Indeed any solution $y(\cdot;-\infty,x,u)$ of the state equation (4), with $0$ replaced by $x\in X\setminus\\{0\\}$ in (6), must satisfy (5) and $\lim_{s\rightarrow-\infty}y(s)=x$. But, as $r_{1}\rightarrow-\infty$, (5) implies, as in (7), that $y(r_{2};-\infty,x,u):=\int_{-\infty}^{r_{2}}e^{(r_{2}-\tau)A}Bu(\tau)\mathrm{d}\tau,\qquad r_{2}\leq t.$ (8) Taking the limit as $r_{2}\rightarrow-\infty$ we get $x=0$, a contradiction. ### 2.2 Minimum energy problems with infinite horizon and associated Riccati equation To better clarify our results we state, roughly and informally, the mathematical problem (see Section 4 for a precise description). The state space $X$ and the control space $U$ are both real separable Hilbert spaces. We take the linear controlled system in $X$ $\left\\{\begin{array}[]{ll}y^{\prime}(s)=Ay(s)+Bu(s),\quad s\in\,]-\infty,0],\\\\[5.69054pt] y(-\infty)=0,\end{array}\right.$ (9) where $A:{\cal D}(A)\subset X\rightarrow X$ generates a strongly continuous semigroup and $B:U\rightarrow X$ is a linear, possibly unbounded operator. Given a point $x\in X$ we consider the set ${\cal U}_{[-\infty,0]}(0,x)$ of all square integrable control strategies that drive the system from the equilibrium state $0$ (at time $t=-\infty$) into the generic non-equilibrium state $x$ (at time $t=0$). It is well known (see Proposition 4.2) that the set ${\cal U}_{[-\infty,0]}(0,x)$ is nonempty if and only if $x\in H$, where $H$ is a suitable subspace of $X$ that can be endowed with its own Hilbert structure (see next subsection for the precise definition of $H$ and Subsection A.2 for its properties). We want to minimize the “energy-like” functional $J_{[-\infty,0]}(u)=\frac{1}{2}\int_{-\infty}^{0}\|u(s)\|^{2}_{U}\,\mathrm{d}s.$ (10) As usual the value function $V_{\infty}$ is defined as $V_{\infty}(x)=\inf_{u\in{\cal U}_{[-\infty,0]}(0,x)}J_{[-\infty,0]}(u),$ (11) and it is finite only when $x\in H$. The peculiarity of the problem with respect to the most studied minimum energy problems in Hilbert spaces (see e.g. [10] [13], [14, 15], [20], [28], and the general surveys [3], [11], [22, 23], [32]) is that it gives rise to an Algebraic Riccati Equation with a ‘wrong’ sign in the linear term to which, to our knowledge, the standard theory developed in the current literature does not apply. Indeed, the associated ARE in $X$ (with unknown $R$), which can be found applying the dynamic programming principle, is, formally, $0=-\langle Ax,Ry\rangle_{X}-\langle Rx,Ay\rangle_{X}-\langle B^{*}Rx,B^{*}Ry\rangle_{U}\,,\quad x,y\in{\cal D}(A)\cap{\cal D}(R).$ (12) Since $R$ is unbounded (this comes from the fact that $V_{\infty}$ is defined only in $H$), it is convenient to rewrite (12) in $H$ (with unknown $P$, which is now a bounded operator on $H$). This way we get the equation $0=-\langle Ax,Py\rangle_{H}-\langle Px,Ay\rangle_{H}-\langle B^{*}Q_{\infty}^{-1}Px,B^{*}Q_{\infty}^{-1}Py\rangle_{U},$ (13) or, transforming the inner products in $H$ into inner products in $X$, $0=-\langle Ax,Q_{\infty}^{-1}Py\rangle_{X}-\langle Q_{\infty}^{-1}Px,Ay\rangle_{X}-\langle B^{*}Q_{\infty}^{-1}Px,B^{*}Q_{\infty}^{-1}Py\rangle_{U},$ (14) In the last two equations $Q_{\infty}$ is the so-called controllability operator (see (19)) and $Q_{\infty}^{-1}$ denotes its pseudoinverse, which is, in general, unbounded. Moreover the last two equations make sense for $x,y$ belonging to suitable sets to be specified later on. For more details on how these equations arise, the definitions of solution, and the relations among them, see the discussion at the beginning of Subsection 4.3. Here we just observe that the form of equation (14) turns out to be more suitable to prove our main results. The ‘wrong’ sign in the linear term333Evidently the two terms in equation (14) (or (12), or (13)) have the same sign, while in the standard case they do not. We infer that the ‘wrong’ sign is in the linear term looking at the corresponding finite horizon problem in [1]. of (14) (or (12), or (13)) does not allow us to approach it using the standard method (described e.g. in [3, pp. 390-394 and 479-486], see also [28, p.1018]), which consists in taking the associated evolutionary Riccati equation, proving that it has a solution $P(t)$ (using an a priori estimate, due to the fact that of both the linear and the quadratic terms have the same sign), and taking the limit of $P(t)$ as $t\rightarrow\infty$. On the other hand the ‘wrong’ sign comes from the nature of the motivating problem: to look at minimum energy paths from equilibrium to non-equilibrium states (see Section 6), which is the opposite direction of the standard one considered in the above quoted papers. This means that the value function depends on the final point, while in the above quoted problems it depends on the initial one (see Remark 3.3 to see what happens to our auxiliary problem using a time inversion). Therefore we are driven to use a different approach, that exploits the structure of the problem; we partly borrow some ideas from [28] and from the literature about model reduction444We thank prof. R. Vinter for providing us these references. (see e.g. [25] and [30]: indeed our results partly generalize Theorem 2.2 of [30], see Remark 4.4). ### 2.3 The method and the main results We now briefly explain our approach. First of all we consider the associated finite horizon problem (which has been studied in the companion paper [1] whose results, for the part needed here, are recalled in Appendix A), where the state equation is $\left\\{\begin{array}[]{ll}y^{\prime}(r)=Ay(r)+Bu(r),\quad r\in\,]-t,0],\\\\[5.69054pt] y(-t)=0,\end{array}\right.$ (15) and the energy to be minimized is $J_{[-t,0]}(u)=\frac{1}{2}\int_{-t}^{0}\|u(r)\|^{2}_{U}\,\mathrm{d}r.$ (16) The value function is $V(t,x)=\inf_{u\in{\cal U}_{[-t,0]}(0,x)}J_{[-t,0]}(u),$ (17) where ${\cal U}_{[-t,0]}(0,x)=\\{u\in L^{2}(-t,0;U):\;y(0)=x\\}.$ (18) We now recall the well known expression of the controllability operator $Q_{t}x=\int_{0}^{t}e^{rA}BB^{*}e^{rA^{*}}x\,\mathrm{d}r,\quad x\in X,\quad t\in[0,+\infty].$ (19) It is well known (see e.g [32, Part IV, Theorem 2.3]) that, for $t\geq 0$, the reachable set of the control systems (15) (finite horizon case) and (4) (infinite horizon case), is $\mathcal{\cal R}(Q^{1/2}_{t})$, i.e. the range of $Q_{t}^{1/2}$ ($t\in[0,+\infty]$). This is clearly the set where the value functions $V$ of (17) (finite horizon case) and $V_{\infty}$ of (11) (infinite horizon case) are well defined. Moreover, as pointed out, e.g., in [1, Proposition C.2-(i)], for $0\leq t_{1}\leq t_{2}$ we have $\mathcal{\cal R}(Q^{1/2}_{t_{1}})\subseteq\mathcal{\cal R}(Q^{1/2}_{t_{2}})$. It will be often useful to assume, beyond Hypothesis 2.2, also the following null controllability assumption. ###### Hypothesis 2.7. There exists $T_{0}\geq 0$ such that ${\cal R}(e^{T_{0}A})\subseteq{\cal R}(Q_{T_{0}}^{1/2}).$ (20) Under such assumption we get, for $t\geq T_{0}$, $\mathcal{\cal R}(Q^{1/2}_{t_{1}})=\mathcal{\cal R}(Q^{1/2}_{t_{2}}),\qquad T_{0}\leq t_{1}\leq t_{2}\leq+\infty.$ Consequently $\ker Q_{t_{1}}=\ker Q^{1/2}_{t_{1}}=\ker Q^{1/2}_{t_{2}}=\ker Q_{t_{1}}\,,\qquad T_{0}\leq t_{1}\leq t_{2}\leq+\infty.$ (21) We can now introduce the already announced space $H$. We define $H=\mathcal{\cal R}(Q_{\infty}^{1/2}).$ (22) Of course it holds $H\subseteq\overline{\mathcal{\cal R}(Q_{\infty}^{1/2})}=[\ker Q_{\infty}^{1/2}]^{\perp}=[\ker Q_{\infty}]^{\perp}.$ The inclusion is in general proper. Define in $H$ the inner product $\langle x,y\rangle_{H}=\langle Q_{\infty}^{-1/2}x,Q_{\infty}^{-1/2}y\rangle_{X}\,,\qquad x,y\in H.$ (23) Some useful results on the space $H$ which form the ground for our main results and are partly proved in [1], are recalled (and proved, when needed) in Appendix A.2. Using $H$ as the ground space we know (see [1, Proposition 4.8-(ii)]) that $V(t,x)=\frac{1}{2}\langle P(t)x,x\rangle_{H}$ , where $P(t)$ is a suitable extension of $Q_{\infty}Q_{t}^{-1}$ (here $Q_{t}^{-1}$ is the pseudoinverse of $Q_{t}$, see [1, Appendix A] or [32, Part IV, end of Section 2.1]). Moreover, using this explicit expression it is proved in [1, Theorem 4.12] that $P(t)$ solves the following Riccati equation in $H$: $\frac{d}{dt}\langle P(t)x,y\rangle_{H}=-\langle Ax,P(t)y\rangle_{H}-\langle P(t)x,Ay\rangle_{H}-\langle{B}^{*}Q_{\infty}^{-1}P(t)x,{B}^{*}Q_{\infty}^{-1}P(t)y\rangle_{U},\quad t>0,$ (24) whose natural condition at $t=0$ is, heuristically, $\lim_{t\rightarrow 0^{+}}P(t)=+\infty.$ It is not difficult to prove (see Proposition 4.3) that $V_{\infty}(x)=\lim_{t\rightarrow+\infty}V(t,x).$ (25) and that $V_{\infty}(x)=\frac{1}{2}\langle x,x\rangle_{H}$. This allows to prove that $P=I_{H}$ (the identity on $H$) solves the ARE (14) in $H$ (Theorem 4.7-(ii)). However, due to the infinite initial condition at $0$ of $P(t)$ (similarly to what happens in [28]), the above limit does not help to prove any comparison theorem for (14). Here comes the main difficulty since, even in very simple cases, it is not known in the literature whether the ARE characterizes $V_{\infty}$ or not (see e.g. [7]). To get a comparison result we proceed as follows. * • We first introduce a suitable auxiliary problem (beginning of Section 3). * • Next, we prove a comparison result for the auxiliary problem (Subsection 3.1, Lemma 3.10). * • Finally we use the relation among the auxiliary problem and the original problem to prove our main maximality result (Theorem 4.7). The idea of introducing an auxiliary problem is exploited in [28], too. However the method used there cannot work here, due to the different sign of the linear part of our equation. ## 3 The auxiliary problem In this section we introduce an auxiliary problem which can be considered a “time reversed” version of the auxiliary problem considered in [28] (see also Remark 3.3 about this). This problem will be a key tool to prove the main result, Theorem 4.7. Indeed, as we will see, any solution of our Algebraic Riccati Equation (14) is also, under appropriate assumptions, a solution of this auxiliary problem with itself as initial datum; a comparison argument will then allow to get the main result. Throughout this section Hypothesis 2.2 will be always assumed, while Hypothesis 2.7 will be used when necessary. Let us consider, for $x\in X$, the following set of controls: $\overline{{\cal U}}_{[-t,0]}(x)=\\{(z,u)\in H\times L^{2}(-t,0;U):y(0)=x\\},$ (26) where $y(\cdot):=y(\cdot;-t,z,u)$ is the solution of the Cauchy problem (similar to (15) but with generic initial datum $z$) $\left\\{\begin{array}[]{l}y^{\prime}(r)=Ay(r)+Bu(r),\quad r\in\,]-t,0],\\\ y(-t)=z.\end{array}\right.$ (27) Note that a control in $\overline{{\cal U}}_{[-t,0]}(x)$ is a couple: an initial point $z\in H$ and a control $u\in{\cal U}_{[-t,0]}(z,x)$, where ${\cal U}_{[-t,0]}(z,x)=\\{u\in L^{2}(-t,0;U):\;y(0;-t,z,u)=x\\}.$ (28) (this is similar to the set (18) but with a generic initial datum $z$). The following is true: ###### Proposition 3.1. Define the reachable set from the point $z$ as ${\mathbf{R}}_{[-t,0]}^{z}:=\left\\{x\in X:\ {\cal U}_{[-t,0]}(z,x)\neq\emptyset\right\\}.$ (29) and set $\bar{\mathbf{R}}_{[-t,0]}:=\bigcup_{z\in H}{\mathbf{R}}^{z}_{[-t,0]}.$ (30) Then the set $\overline{{\cal U}}_{[-t,0]}(x)$ introduced in (26) is nonempty if and only if $x\in\bar{\mathbf{R}}_{[-t,0]}$. Moreover we have $\bar{\mathbf{R}}_{[-t,0]}\subseteq H,$ (31) with equality for $t\geq T_{0}$, if Hypothesis 2.7 holds. ###### Proof. The first statement is an immediate consequence of the definition of reachable set in (29). The second one follows from (99), Lemma A.4-(i), the fact that ${\cal R}(Q_{t}^{1/2})\subseteq{\cal R}(Q_{\infty}^{1/2})$ (with equality, for $t\geq T_{0}$, when Hypothesis 2.7 holds), and the equality ${\cal R}\left({\cal L}_{-t,0}\right)={\mathbf{R}}^{0}_{[-t,0]}={\cal R}(Q_{t}^{1/2}),\qquad t\in[0,+\infty]$ (32) (here ${\cal L}_{-t,0}$ is the operator defined in (98)), which is proved in [32, Theorem 2.3] for $t<+\infty$. Such equality holds also when $t=+\infty$ with exactly the same proof. ∎ Given a bounded selfadjoint positive operator $N$ on $H$ we want to minimize, in the class $\overline{{\cal U}}_{[-t,0]}(x)$, the following functional with an initial cost: $J^{N}_{[-t,0]}(z,u)=\frac{1}{2}\langle Nz,z\rangle_{H}+\frac{1}{2}\int_{-t}^{0}\|u(s)\|_{U}^{2}\,\mathrm{d}s.$ (33) The presence of the operator $N\in{\cal L}_{+}(H)$ forces us to fix the starting point $z$ at time $-t$ in $H$, rather than in $X$. Define $V^{N}(t,x)=\inf_{(z,u)\in\overline{{\cal U}}_{[-t,0]}(x)}J^{N}_{[-t,0]}(z,u)=\inf_{z\in H}\left[\inf_{u\in{\cal U}_{[-t,0]}(z,x)}J^{N}_{[-t,0]}(z,u)\right],\ t>0,\ x\in X,$ (34) with the agreement that the infimum over the emptyset is $+\infty$, so that $V^{N}(t,x)$ is finite only when $x\in H$. Now we provide a relation between $V^{N}$ and the value function $V$ defined in (17). ###### Proposition 3.2. We have $V^{N}(t,x)=\inf_{z\in H}\left[V(t,x-e^{tA}z)+\frac{1}{2}\langle Nz,z\rangle_{H}\right],\quad t>0,\ x\in X$ (35) and, in particular, $V^{N}(t,x)\leq V(t,x)\qquad\forall x\in X,\quad\forall t>0.$ (36) ###### Proof. We use (96), (100) and (101) getting $\inf_{u\in{\cal U}_{[-t,0]}(z,x)}J^{N}_{[-t,0]}(z,u)=V_{1}(-t,0;z,x)+\frac{1}{2}\langle Nz,z\rangle_{H}=V(t,x-e^{tA}z)+\frac{1}{2}\langle Nz,z\rangle_{H}\,.$ This equality immediately implies (35). Taking $z=0$ we get (36). ∎ The following remark is crucial to understand what is the “natural” Riccati equation associated to this auxiliary problem. ###### Remark 3.3. If $A$ generates not just a $C_{0}$-semigroup but a $C_{0}$-group, the auxiliary problem can be shown, under appropriate assumptions, to be equivalent, reversing the time, to a standard optimization problem with final cost. Indeed, given $x\in H$, consider the problem of minimizing, over all $v(\cdot)\in L^{2}(0,t;U)$, the functional $\widehat{J}^{N}_{[0,t]}(x,v)=\frac{1}{2}\langle Nw(t),w(t)\rangle_{H}+\frac{1}{2}\int_{0}^{t}\|v(s)\|_{U}^{2}\,\mathrm{d}s,$ (37) where $w(\cdot):=w(\cdot;0,x,v)$ is the mild solution of the Cauchy problem $w^{\prime}(s)=-Aw(s)+Bv(s),\quad s\in\,]-t,0],\qquad w(0)=x.$ (38) Assume now that, for every $x\in H$, the mild solution $w(\cdot;0,x,v)$ belongs to $H$ for every $t>0$. Setting $\widehat{V}^{N}(t,x)=\inf_{v\in L^{2}(0,t;U)}\widehat{J}^{N}_{[0,t]}(x,v),$ it can be seen that $\widehat{V}^{N}(t,x)=V^{N}(t,x).$ To see this, fix $(t,x)\in[0,+\infty[\,\times H$ and recall that, for every $(z,u)\in\overline{\cal U}_{[-t,0]}(x)$, we have $e^{tA}z+\int_{-t}^{0}e^{-sA}Bu(s)ds=x\quad\Longleftrightarrow\quad z+\int_{-t}^{0}e^{(-t-s)A}Bu(s)ds=e^{-tA}x;$ hence, changing variable in the integral, $z=e^{t(-A)}x+\int_{0}^{t}e^{(t-s)(-A)}B(-u(-s))ds.$ This means that $\bar{\mathbf{R}}_{[-t,0]}=H$ (see (30)). Moreover, to any $(z,u)\in\overline{\cal U}_{[-t,0]}(x)$ we can associate a function $v\in L^{2}(0,t;U)$ such that $w(t)=z$, namely, $v(s)=-u(-s)$; consequently $J^{N}_{[-t,0]}(z,u)=\widehat{J}^{N}_{[0,t]}(x,v).$ (39) Conversely, given any $v\in L^{2}(0,t;U)$, set $z=w(t;0,x,v)$ and $u(s)=-v(-s)$: then, clearly, $(z,u)\in\overline{\cal U}_{[-t,0]}(x)$ and, again, (39) holds. In conclusion, there is a one-to-one correspondence between the control set of the two problems and, in particular, $\widehat{V}^{N}(t,x)=V^{N}(t,x)$. The equation for the “time-reversed” problem (37)-(38) turns out to be the following: $\left\\{\begin{array}[]{ll}\displaystyle\frac{d}{ds}\langle P^{N}(s)x,y\rangle_{H}=&-\langle Ax,P^{N}(s)y\rangle_{H}-\langle P^{N}(s)x,Ay\rangle_{H}-\\\\[5.69054pt] &-\langle{B}^{*}Q_{\infty}^{-1}P^{N}(s)x,{B}^{*}Q_{\infty}^{-1}P^{N}(s)y\rangle_{U}\,,\qquad s\in\,]0,t],\\\\[5.69054pt] P^{N}(0)=N.\end{array}\right.$ (40) To give sense to (40) we must take $x,y\in{\cal D}(A)\cap H$ with $Ax,Ay\in H$ and $P^{N}(t)x,P^{N}(t)y\in{\cal R}(Q_{\infty})$. When ${B}^{*}Q_{\infty}^{-1}$ can be extended to a bounded operator $H\rightarrow U$ and $A$ generates a group, then it is known that the value function $\widehat{V}^{N}$ is quadratic and $\widehat{V}^{N}(t,x)=\langle\widehat{P}^{N}(t)x,x\rangle_{H}$, where $\widehat{P}^{N}:[0,+\infty[\rightarrow{\cal L}_{+}(H)$ is the unique solution of (40). In our case this is not obvious, but it suggests anyway the right form of the Riccati equation for our auxiliary problem. Note, finally, that the right hand side of (40) is exactly one of the forms of the ARE we aim to study (see (13)). ###### Remark 3.4. As in the case $N=0$ treated in [1], in the above Riccati equations the sign of the linear part is opposite to the usual one. In fact the control problem (27)-(33) involves an “initial cost”, instead of a final cost like in the standard problems (see e.g. [28]). Our aim now is to prove that for every stationary solution $Q$ of the Riccati equation (40) (in a suitable class to be defined later) there exists an operator $N$, namely $Q$ itself, such that $\frac{1}{2}\langle Qx,x\rangle_{H}\leq V^{N}(t,x),\qquad\hbox{for sufficiently large $t$.}$ ###### Remark 3.5. It is possible to prove much more about the auxiliary problem, namely: * (i) that, for every $N\in{\cal L}_{+}(H)$ the value function $V^{N}$ is continuous and is a quadratic form in $H$; * (ii) that, when $N$ is coercive (i.e., for some $\nu>0$, $\langle Nx,x\rangle_{H}\geq\nu|x|^{2}_{H}$ for all $x\in H$), the linear operator $P^{N}$ associated to the value function solves the Riccati equation (40); * (iii) that the comparison result mentioned above translates in the inquality $P^{N}\geq Q^{N}$, in the preorder of positive operators, for every constant solution $Q^{N}$ of the Riccati equation (40) in a suitable class. This is the subject of a paper in progress. ### 3.1 A key comparison result Given any initial datum $N\in{\cal L}_{+}(H)$, we want to compare the “stationary” solutions of the Riccati equation (40) with the value function $V^{N}$ of the auxiliary problem. This fact will be used, in the next section, as a key tool to prove our main results. In order to do this we need first to give a precise meaning to the concept of stationary solution of (40). Roughly speaking, a stationary solution $P\in{\cal L}_{+}(H)$ of the Riccati Equation (40) should also be a solution of the following Algebraic Riccati Equation (ARE), which comes from the right hand side of (40): $0=-\langle Ax,Py\rangle_{H}-\langle Px,Ay\rangle_{H}-\langle B^{*}Q_{\infty}^{-1}Px,B^{*}Q_{\infty}^{-1}Py\rangle_{U}.$ (41) This equation is meaningful for every $x,y\in{\cal D}(A)\cap H$ with $Px,Py\in{\cal R}(Q_{\infty})$ and $Ax,Ay\in H$. Since the last requirement appears too restrictive, we rewrite (41) by taking the first two inner products in $X$, getting: $0=-\langle Ax,Q_{\infty}^{-1}Py\rangle_{X}-\langle Q_{\infty}^{-1}Ox,Ay\rangle_{X}-\langle B^{*}Q_{\infty}^{-1}Px,B^{*}Q_{\infty}^{-1}Py\rangle_{U}.$ (42) This makes sense in a larger set of vectors $x,y$, namely for every $x,y\in{\cal D}(A)\cap H$ with $Px,Py\in{\cal R}(Q_{\infty})$.555Note that (41) is the same as (13) while (42) is the same as (14). We can now provide the precise definition of solution of (42). ###### Definition 3.6. Let $P\in{\cal L}_{+}(H)$ and define the operator $\Lambda_{P}$ as follows: $\left\\{\begin{array}[]{l}{\cal D}(\Lambda_{P})=\\{x\in H:\ Px\in{\cal R}(Q_{\infty})\\}\\\\[5.69054pt] \Lambda_{P}x=Q_{\infty}^{-1}Px\qquad\forall x\in{\cal D}(\Lambda_{P}).\end{array}\right.$ (43) We say that $P$ is a solution of (42) (or, alternatively, a stationary solution of (40)) if ${\cal D}(A)\cap{\cal D}(\Lambda_{P})$ is dense in $[\ker Q_{\infty}]^{\perp}$ and $0=-\langle Ax,\Lambda_{P}y\rangle_{X}-\langle\Lambda_{P}x,Ay\rangle_{X}-\langle B^{*}\Lambda_{P}x,B^{*}\Lambda_{P}y\rangle_{U}\qquad\forall x,y\in{\cal D}(A)\cap{\cal D}(\Lambda_{P}).$ (44) We now define a subclass ${\cal Q}$ of the class of all stationary solutions of (40). First of all we recall that, by Lemma A.4-(i), $e^{tA}|_{H}$ is a strongly continuous semigroup in $H$. We then use the following notation. ###### Notation 3.7. We denote by $A_{0}:{\cal D}(A_{0})\subseteq H\rightarrow H$ the generator of $e^{tA}|_{H}$, and we write $e^{tA_{0}}$ in place of $e^{tA}|_{H}$. ###### Definition 3.8. Let $P\in{\cal L}_{+}(H)$. We say that $P\in{\cal Q}$ if there exists $D\subseteq{\cal D}(\Lambda_{P})$ such that $D$ is dense in ${\cal D}(A)\cap H$ with respect to the norm $\|\cdot\|_{H}+\|A\cdot\|_{X}$; ###### Lemma 3.9. The set ${\cal R}(Q_{\infty})\cap{\cal D}(A)$ is dense in ${\cal D}(A)\cap H$, equipped with the norm $\|\cdot\|_{H}+\|A\cdot\|_{X}$. Hence, choosing $D={\cal R}(Q_{\infty})\cap{\cal D}(A)$, we have $P=I_{H}\in{\cal Q}$. ###### Proof. Let $x\in H\cap{\cal D}(A)$ such that $\langle x,z\rangle_{H}+\langle Ax,Az\rangle_{X}=0,\qquad\forall z\in{\cal R}(Q_{\infty})\cap{\cal D}(A).$ It is enough to prove that $x=0$. Observe that, writing $z=Q_{\infty}y$, $\langle x,Q_{\infty}y\rangle_{H}+\langle Ax,AQ_{\infty}y\rangle_{X}=0,\qquad\forall y\in{\cal D}(AQ_{\infty}).$ Then $\langle Ax,AQ_{\infty}y\rangle_{X}=-\langle x,Q_{\infty}y\rangle_{H}=-\langle x,y\rangle_{X}\qquad\forall y\in{\cal D}(AQ_{\infty}).$ This means that $Ax\in{\cal D}((AQ_{\infty})^{*})$ and $(AQ_{\infty})^{*}Ax=-x$. Hence $\langle(AQ_{\infty})^{*}Ax,Ax\rangle_{X}=-\langle x,Ax\rangle_{X}=|(-A)^{1/2}x|_{X}^{2}\geq 0.$ On the other hand we know, from [1, Lemma 3.1-(ii)], that, for every $y\in{\cal D}((AQ_{\infty})^{*})\subseteq{\cal D}(AQ_{\infty})$ $2\langle(AQ_{\infty})^{*}y,y\rangle_{X}=-\|B^{*}y\|^{2}_{U}\,,$ so that $2\langle(AQ_{\infty})^{*}Ax,Ax\rangle_{X}=-\|B^{*}Ax\|^{2}_{U}\leq 0.$ This implies that $\|(-A)^{1/2}x\|_{X}^{2}=0$; hence $Ax=0$ and, since $A$ is invertible, $x=0$. ∎ ###### Lemma 3.10. Assume Hypothesis 2.7. Let $P\in{\cal L}_{+}(H)$ be a solution of (42) according to Definition 3.6. Assume also that $P\in{\cal Q}$ and that $BB^{*}$ is coercive, which is equivalent to require that, for some $\mu>0$, $\|B^{*}x\|_{U}\geq\mu\|x\|_{X}$ for all $x\in X$. Then, the following estimate holds: $\frac{1}{2}\langle Px,x\rangle_{H}\leq V^{P}(t-T_{0},x)\qquad\forall x\in H,\ \ \forall t>T_{0},$ where $V^{P}$ is the value function defined in (34) with $N=P$. ###### Proof. Step 1 We prove the estimate $\langle Px,x\rangle_{H}\leq\langle Py(T_{0}-t),y(T_{0}-t)\rangle_{H}+\int_{T_{0}-t}^{0}\|u(s)\|_{U}^{2}\,\mathrm{d}s,\qquad t>T_{0},$ (45) for every $(z,u)\in\overline{{\cal U}}_{[-t,0]}(x)$ with $x\in H$, where $y$ is the state corresponding to $(z,u)$, i.e. $y(s)=e^{(s+t)A}z+\int_{-t}^{s}e^{(s-\sigma)A}\,Bu(\sigma)\,d\sigma,\quad s\in[-t,0].$ (46) Such inequality would be easy to prove if we were able to compute $\frac{d}{ds}\langle Py(s),y(s)\rangle_{H}$ and prove that $\frac{d}{ds}\langle Py(s),y(s)\rangle_{H}\leq\|u(s)\|^{2}_{U},\qquad s\in[-t,0].$ Unfortunately we even do not know if such a derivative exists. Hence we need to build a delicate approximation procedure as follows. Fix $t>T_{0}$ and $x\in H$; consider any $(z,u)\in\overline{{\cal U}}_{[-t,0]}(x)$. It is not restrictive to assume in (46) that $u(\sigma)\in\overline{{\cal R}(B^{*})}$ for every $\sigma\in[-t,0]$: indeed, writing, for every such $\sigma$, $u(\sigma)=u_{1}(\sigma)+u_{2}(\sigma),\quad u_{1}(\sigma)\in\overline{{\cal R}(B^{*})},\quad u_{2}(\sigma)\in\overline{{\cal R}(B^{*})}^{\perp}=\ker B,$ it is clear that $e^{(s-\sigma)A}Bu_{2}(\sigma)=0$. Hence $y(s)=e^{(s+t)A}z+\int_{-t}^{s}e^{(s-\sigma)A}\,Bu_{1}(\sigma)\,d\sigma,\quad s\in[-t,0].$ Since, evidently, $J^{P}_{[-t,0]}(z,u)\geq J^{P}_{[-t,0]}(z,u_{1})$, we can always choose $u_{1}$ in place of $u$. Next, select a sequence $\\{(z_{n},u_{n})\\}\subseteq\big{[}{\cal D}(A_{0})\big{]}\times C^{1}_{0}([-t,0];U)$666$C^{1}_{0}([-t,0];U)$ is the set of $C^{1}$ $U$-valued functions which take the value $0$ at the boundary., such that $u_{n}$ is ${\cal R}(B^{*})$-valued and $(z_{n},u_{n})\rightarrow(z,u)$ in $H\times L^{2}(-t,0;U)$. Thus we can set $u_{n}=B^{*}v_{n}$, where $v_{n}\in C^{1}_{0}([-t,0],X)$ and, denoting by $y_{n}$ the corresponding state, we have $y_{n}\in C^{1}([-t,0];H)\cap C([-t,0];{\cal D}(A))$ (see e.g. [27, Chapter 4, Corollary 2.5]) and $y_{n}(s)=e^{(s+t)A}z_{n}+\int_{-t}^{s}e^{(s-\sigma)A}\,BB^{*}v_{n}(\sigma)\,d\sigma,\qquad s\in[-t,0].$ Thanks to the properties of the set $D$ of Hypothesis 3.8, we can now choose, for every $n\in\mathbb{N}$, another approximating sequence $\\{y_{nk}\\}_{h\in N}\subset C^{1}([-t,0],H)\cap C^{0}_{0}([-t,0],{\cal D}(A))$, such that $y_{nk}(s)\in D$ for every $s\in[-t,0]$ and satisfying, as $k\rightarrow+\infty$, $y_{nk}\rightarrow y_{n}\hbox{ in }C^{1}([-t,0];H),\qquad Ay_{nk}\rightarrow Ay_{n}\hbox{ in }C([-t,0];X)$ (47) (see e.g. [27, Chapter 4, Theorem 2.7]). Set now $w_{nk}=y^{\prime}_{nk}-Ay_{nk}$. By (47) we get, for every $n\in\mathbb{N}$, $w_{nk}\rightarrow y_{n}-Ay_{n}=BB^{*}v_{n}\hbox{ in }C^{0}([-t,0];X)\qquad\hbox{as $k\rightarrow+\infty$.}$ (48) We now can differentiate the quantity $\langle Py_{nk}(s),y_{nk}(s)\rangle_{H}$ for $s\in[-t,0]$. Indeed, taking into account the above definition of $w_{nk}$, we obtain, for $s\in[-t,0]$ and $n,k\in\mathbb{N}$: $\displaystyle\frac{d}{ds}\langle Py_{nk}(s),y_{nk}(s)\rangle_{H}=\langle y_{nk}^{\prime}(s),Py_{nk}(s)\rangle_{H}+\langle Py_{nk}(s),y_{nk}^{\prime}(s)\rangle_{H}=$ $\displaystyle=\langle y_{nk}^{\prime}(s),\Lambda_{P}y_{nk}(s)\rangle_{X}+\langle\Lambda_{P}y_{nk}(s),y_{nk}^{\prime}(s)\rangle_{X}=$ $\displaystyle=\langle Ay_{nk}(s)+w_{nk}(s),\Lambda_{P}y_{nk}(s)\rangle_{X}+\langle\Lambda_{P}y_{nk}(s),Ay_{nk}(s)+w_{nk}(s)\rangle_{X}.$ Since $P$ solves the ARE (44) we get, for every $s\in[-t,0]$, $\displaystyle\frac{d}{ds}\langle Py_{nk}(s),y_{nk}(s)\rangle_{H}=$ $\displaystyle=-\|B^{*}\Lambda_{P}y_{nk}(s)\|_{U}^{2}+\langle w_{nk}(s),\Lambda_{P}y_{nk}(s)\rangle_{X}+\langle\Lambda_{P}y_{nk}(s),w_{nk}(s)\rangle_{X}=$ $\displaystyle=-\|B^{*}\Lambda_{P}y_{nk}(s)\|_{U}^{2}+\langle B^{*}v_{n}(s),B^{*}\Lambda_{P}y_{nk}(s)\rangle_{U}+\langle B^{*}\Lambda_{P}y_{nk}(s),B^{*}v_{n}(s)\rangle_{U}=$ $\displaystyle\quad+\langle w_{nk}(s)-BB^{*}v_{n}(s),\Lambda_{P}y_{nk}(s)\rangle_{X}+\langle\Lambda_{P}y_{nk}(s),w_{nk}(s)-BB^{*}v_{n}(s)\rangle_{X}=$ $\displaystyle=-\|B^{*}\Lambda_{P}y_{nk}(s)-B^{*}v_{n}(s)\|_{U}^{2}+\|B^{*}v_{n}(s)\|_{U}^{2}+$ $\displaystyle\quad+\langle w_{nk}(s)-BB^{*}v_{n}(s),\Lambda_{P}y_{nk}(s)\rangle_{X}+\langle\Lambda_{P}y_{nk}(s),w_{nk}(s)-BB^{*}v_{n}(s)\rangle_{X}.$ Hence, recalling that $u_{n}=B^{*}v_{n}$, we may write for every $\varepsilon>0$, $\begin{array}[]{lcl}\displaystyle\frac{d}{ds}\langle Py_{nk}(s),y_{nk}(s)\rangle_{H}&\leq&-\|B^{*}\Lambda_{P}y_{nk}(s)-B^{*}v_{n}(s)\|_{U}^{2}+\|u_{n}(s)\|_{U}^{2}+\\\\[8.53581pt] &&+2\|w_{nk}(s)-Bu_{n}(s)\|_{X}\|\Lambda_{Q}Py_{nk}(s)\|_{X}\leq\\\\[5.69054pt] &\leq&-\|B^{*}\Lambda_{P}y_{nk}(s)-B^{*}v_{n}(s)\|_{U}^{2}+\|u_{n}(s)\|_{U}^{2}+\\\\[5.69054pt] &&\displaystyle+\frac{1}{\varepsilon}\|w_{nk}(s)-Bu_{n}(s)\|_{X}^{2}+\varepsilon\|\Lambda_{P}y_{nk}(s)\|_{X}^{2}.\end{array}$ (49) Now observe that $\varepsilon\|\Lambda_{P}y_{nk}(s)\|_{X}^{2}\leq\frac{\varepsilon}{\mu}\|B^{*}\Lambda_{P}y_{nk}(s)\|_{U}^{2}\leq 2\frac{\varepsilon}{\mu}\|B^{*}\Lambda_{P}y_{nk}(s)-B^{*}v_{n}(s)\|_{U}^{2}+2\frac{\varepsilon}{\mu}\|B^{*}v_{n}(s)\|_{U}^{2}.$ Inserting this inequality into (49) we get $\begin{array}[]{lcl}\displaystyle\frac{d}{ds}\langle Py_{nk}(s),y_{nk}(s)\rangle_{H}&\leq&\displaystyle-\left(1-2\frac{\varepsilon}{\mu}\right)\|B^{*}\Lambda_{P}y_{nk}(s)-B^{*}v_{n}(s)\|_{U}^{2}+\\\\[11.38109pt] &&\displaystyle+\left(1+2\frac{\varepsilon}{\mu}\right)\|u_{n}(s)\|_{U}^{2}+\frac{1}{\varepsilon}\|w_{nk}(s)-Bu_{n}(s)\|_{X}^{2}.\end{array}$ (50) Hence, for all positive $\varepsilon$ such that $2\frac{\varepsilon}{\mu}\leq\frac{1}{2}$ we get $\frac{d}{ds}\langle Py_{nk}(s),y_{nk}(s)\rangle_{H}\leq\left(1+2\frac{\varepsilon}{\mu}\right)\|u_{n}(s)\|_{U}^{2}+\frac{1}{\varepsilon}\|w_{nk}(s)-Bu_{n}(s)\|_{X}^{2}\,.$ (51) Now we have for every $s\in[-t,0]$, as $k\rightarrow\infty$, $\|y_{nk}(s)-y_{n}(s)\|_{H}\rightarrow 0,\quad\|y_{nk}^{\prime}(s)-y_{n}^{\prime}(s)\|_{H}\rightarrow 0,\quad\|w_{nk}(s)-Bu_{n}(s)\|_{X}\rightarrow 0;$ thus we get, for every $n\in\mathbb{N}^{+}$, $s\in[-t,0]$ and $0<\varepsilon\leq\mu/4$, $\frac{d}{ds}\langle Py_{n}(s),y_{n}(s)\rangle_{H}\leq\left(1+2\frac{\varepsilon}{\mu}\right)\|u_{n}(s)\|_{U}^{2}\,.$ Finally, letting $\varepsilon\rightarrow 0$, $\frac{d}{ds}\langle Py_{n}(s),y_{n}(s)\rangle_{H}\leq\|u_{n}(s)\|_{U}^{2}\quad\forall n\in\mathbb{N}^{+},\quad\forall s\in[-t,0].$ We now integrate in the smaller interval $[T_{0}-t,0]$: $\langle Py_{n}(0),y_{n}(0)\rangle_{H}\leq\langle Py_{n}(T_{0}-t),y_{n}(T_{0}-t)\rangle_{H}+\int_{T_{0}-t}^{0}\|u_{n}(s)\|_{U}^{2}\,\mathrm{d}s.$ Letting $n\rightarrow\infty$, since $y_{n}(s)\rightarrow y(s)$ for every $s\in[-t,0]$, $y(0)=x$, and $u_{n}\rightarrow u$ in $L^{2}(-t,0;U)$, we deduce for every $(z,u)\in\overline{{\cal U}}_{[-t,0]}(x)$ $\langle Px,x\rangle_{H}\leq\langle Py(T_{0}-t),y(T_{0}-t)\rangle_{H}+\int_{T_{0}-t}^{0}\|u(s)\|_{U}^{2}\,\mathrm{d}s,\qquad t>T_{0};$ this is equation (45). Step 2 We complete the proof of the Lemma. Consider a sequence $(\hat{z}_{n},\hat{u}_{n})\in\overline{{\cal U}}_{[T_{0}-t,0]}(x)$, such that, as $n\rightarrow\infty$, $J^{P}_{[T_{0}-t,0]}(\hat{z}_{n},\hat{u}_{n})\rightarrow\inf_{(z,u)\in\overline{{\cal U}}_{[T_{0}-t,0]}(x)}J^{P}_{[T_{0}-t,0]}(z,u)=V^{P}(t-T_{0},x).$ (52) Thus $\hat{z}_{n}\in H$, $\hat{u}_{n}\in L^{2}(T_{0}-t,0;U)$ and the corresponding state is $\hat{y}_{n}(s)=e^{(s+t-T_{0})A}\hat{z}_{n}+\int_{T_{0}-t}^{s}e^{(s-\sigma)A}B\hat{u}_{n}(\sigma)\,\mathrm{d}\sigma,\quad s\in[T_{0}-t,0];$ in particular $\hat{y}_{n}(0)=x$. Now choose $\hat{v}_{n}\in L^{2}(-t,T_{0}-t;U)$ such that $\int_{-t}^{T_{0}-t}e^{(T_{0}-t-\sigma)A}B\hat{v}_{n}(\sigma)\,\mathrm{d}\sigma=\hat{z}_{n};$ (53) this is possible since, due to Hypothesis 2.7, the range of the operator (defined in (98)) $v\mapsto{\cal L}_{-t,T_{0}-t}(v)={\cal L}_{-T_{0},0}(v(\cdot+t-T_{0}))$ is all of $H$ (see [32, Theorem 2.3]). Then, setting $\overline{u}_{n}=\left\\{\begin{array}[]{ll}\hat{v}_{n}&\textrm{in }[-t,T_{0}-t]\\\\[5.69054pt] \hat{u}_{n}&\textrm{in }[T_{0}-t,0],\end{array}\right.$ the state corresponding to $(0,\overline{u}_{n})$ in $[-t,0]$ is $\overline{y}_{n}(s)=\int_{-t}^{s}e^{(s-\sigma)A}B\overline{u}_{n}(\sigma)\,\mathrm{d}\sigma.$ By (53) we have $\overline{y}_{n}(T_{0}-t)=\int_{-t}^{T_{0}-t}e^{(T_{0}-t-\sigma)A}B\overline{u}_{n}(\sigma)\,\mathrm{d}\sigma=\hat{z}_{n};$ hence, by uniqueness, $\overline{y}_{n}(s)=e^{(s+t-T_{0})A}\hat{z}_{n}+\int_{T_{0}-t}^{s}e^{(s-\sigma)A}B\hat{u}_{n}(\sigma)\,\mathrm{d}\sigma=\hat{y}_{n}(s)\qquad\forall s\in[T_{0}-t,0],$ so that $\overline{y}_{n}(0)=\hat{y}_{n}(0)=x$. This shows that $(0,\overline{u}_{n})\in\overline{{\cal U}}_{[-t,0]}(x)$, and consequently, by (45), $\langle Px,x\rangle_{H}\leq\langle P\hat{z}_{n},\hat{z}_{n}\rangle_{H}+\int_{T_{0}-t}^{0}\|\hat{u}_{n}(s)\|^{2}_{U}\,\mathrm{d}s=2J^{P}_{[T_{0}-t,0]}(\hat{z}_{n},\hat{u}_{n}).$ Finally, by (52), as $n\rightarrow\infty$ we get $\frac{1}{2}\langle Px,x\rangle_{H}\leq V^{P}(t-T_{0},x)\qquad\forall t>T_{0},\quad\forall x\in H.$ ∎ ## 4 Minimum energy with (negative) infinite horizon We now give a precise formulation of our infinite horizon problem (see Subsection 2.2 and also [1, Remark 2.8]). We assume that Hypothesis 2.2 holds throughout this section without repeating it. For any given control $u\in L^{2}(-\infty,s;U)$ we take the state equation $\left\\{\begin{array}[]{l}y^{\prime}(r)=Ay(r)+Bu(r),\quad r\in\,]-\infty,s],\\\ y(-\infty)=0.\end{array}\right.$ (54) By Lemma 2.5 we know that the unique solution of (54) belongs to $C(\,]-\infty,s];X)$, is given by $y(r):=y(r;-\infty,0,u)=\int_{-\infty}^{r}e^{(r-\tau)A}Bu(\tau)\,\mathrm{d}\tau,\quad-\infty<r\leq s,$ and satisfies, for every $-\infty<r_{1}\leq r_{2}\leq s$, $y(r_{2})=e^{(r_{2}-r_{1})A}y(r_{1})+\int_{r_{1}}^{r_{2}}e^{(r_{2}-\tau)A}Bu(\tau)\,\mathrm{d}\tau,\quad\textrm{and}\quad\lim_{r\rightarrow-\infty}y(r)=0\quad\hbox{ in $X$}.$ As for the finite horizon case, we define: ${\cal U}_{[-\infty,s]}(0,x)\stackrel{{\scriptstyle\textrm{def}}}{{=}}\left\\{u\in L^{2}(-\infty,s;U)\;:\;y(s;-\infty,0,u)=x\right\\},$ (55) $J_{[-\infty,s]}(u)=\frac{1}{2}\int_{-\infty}^{s}\|u(r)\|^{2}_{U}\;\mathrm{d}r,$ $V_{1}(-\infty,s;0,x)\stackrel{{\scriptstyle\textrm{def}}}{{=}}\inf_{u\in{\cal U}_{[-\infty,s]}(0,x)}J_{[-\infty,s]}(u),$ with the agreement that the infimum over the empty set is $+\infty$. From (55) it is easy to see that $u(\cdot)\in{\cal U}_{[-\infty,s]}(0,x)\ \iff\ u(\cdot-s)\in{\cal U}_{[-\infty,0]}(0,x);$ (56) this implies that $V_{1}(-\infty,s;0,x)=V_{1}(-\infty,0;0,x).$ From now on we set, as in (101) $V_{\infty}(x)=V_{1}(-\infty,0;0,x)=\inf_{u\in{\cal U}_{[-\infty,0]}(0,x)}J_{[-\infty,0]}(u),\quad x\in X.$ (57) We collect now some results about the above problem and the function $V_{\infty}$. ### 4.1 Optimal strategies We start proving the existence of optimal strategies. ###### Proposition 4.1. The set ${\cal U}_{[-\infty,0]}(0,x)$ is nonempty if and only if $x\in H$. Moreover, for every $x\in H$ there exists a unique $\hat{u}_{x}\in{\cal U}_{[-\infty,0]}(0,x)$ such that $V_{\infty}(x)=J_{[-\infty,0]}(\hat{u}_{x}).$ ###### Proof. The first statement follows from (32) as in Proposition 3.1. Now take $x\in H$ and observe that any minimizing sequence $\\{u_{n}\\}_{n\in{\mathbb{N}}}$ must be bounded in $L^{2}(-\infty,0;U)$; so, passing to a subsequence, we have $u_{n}\rightharpoonup\hat{u}_{x}$ in $L^{2}(-\infty,0;U)$. As the functional $J_{[-\infty,0]}$ is weakly lower semicontinuous, we get $V_{\infty}(x)\leq J_{[-\infty,0]}(\hat{u}_{x})\leq\liminf_{n\rightarrow\infty}J_{[-\infty,0]}(u_{n})=V_{\infty}(x),$ i.e. $\hat{u}_{x}$ is optimal. Uniqueness is an easy consequence of the strict convexity of the functional $J_{[-\infty,0]}$. ∎ Moreover we have the following result about the optimal couples when $x\in{\cal R}(Q_{\infty})$ (see [1, Proposition C.3 and Remark C.4]). ###### Proposition 4.2. Let $x\in{\cal R}(Q_{\infty})$. Let $(\hat{y}_{x},\hat{u}_{x})$ be the optimal couple for our problem with target $x$. Then we have $\hat{u}_{x}(r)=B^{*}e^{-rA^{*}}{Q}^{-1}_{\infty}x,\quad r\in\,]-\infty,0].$ (58) Moreover the corresponding optimal state $\hat{y}_{x}$ satisfies $\hat{y}_{x}(r)=Q_{\infty}e^{-rA^{*}}Q_{\infty}^{-1}x,\quad r\in\,]-\infty,0];$ (59) hence the optimal couple satisfies the feedback formula $\hat{u}_{x}(r)=B^{*}Q_{\infty}^{-1}\hat{y}_{x}(r),\quad r\in\,]-\infty,0],$ (60) and, formally, $\hat{y}_{x}$ is a solution of the backward closed loop equation (BCLE) $y^{\prime}(r)=(A+BB^{*}Q_{\infty}^{-1})y(r),\quad r\in\,]-\infty,0[\,,\quad y(0)=x,$ (61) which, since $Q_{\infty}$ solves the Lyapunov equation (see [1, Proposition 3.3] rewrites as $y^{\prime}(r)=-Q_{\infty}A^{*}Q_{\infty}^{-1}y(r),\quad r\in\,]-\infty,0[\,.$ (62) If $A^{*}$ commutes with $Q_{\infty}$ (e.g. when $A$ is selfadjoint and invertible, and $A$ and $BB^{*}$ commute), then (62) becomes $y^{\prime}(r)=-A^{*}y(r),\quad r\in\,]-\infty,0[\,.$ (63) This means that, in such case, the optimal trajectory arriving at $x$ is given by $y(r)=e^{-rA^{*}}x,\quad r\in\,]-\infty,0].$ ### 4.2 Connection with the finite horizon case We now prove the connection between $V_{\infty}$ and the value function $V$ of the corresponding finite horizon problem which is studied in [1] (see also Appendix A). ###### Proposition 4.3. Under Hypothesis 2.7, for every $x\in H$ we have $V_{\infty}(x)=\lim_{t\rightarrow+\infty}V(t,x)=\inf_{t>0}V(t,x).$ Moreover $V_{\infty}(x)=\frac{1}{2}\|x\|_{H}^{2}$. ###### Proof. First of all, by [1, Proposition 4.8-(i)], the function $V(\cdot,x)$ is decreasing for every $x\in H$; hence, for every such $x$ $\exists\,\lim_{t\rightarrow+\infty}V(t,x)=\inf_{t>0}V(t,x).$ We now prove that $V_{\infty}(x)\leq\inf_{t>0}V(t,x)$. With an abuse of notation we can write ${\cal U}_{[-t,0]}(0,x)\subseteq{\cal U}_{[-\infty,0]}(0,x)\qquad\forall t>0:$ indeed, given a control bringing $0$ to $x$ in the interval $[-t,0]$, we can extend it to a control bringing $0$ to $x$ in the interval $[-\infty,0]$ just taking the null control on $\,]-\infty,-t]$. So, if the set ${\cal U}_{[-t,0]}(0,x)$ is not empty, a fortiori the set ${\cal U}_{[-\infty,0]}(0,x)$ will be not empty. This fact, together with the monotonicity of $V(\cdot,x)$ implies that $V_{\infty}(x)\leq\inf_{t>0}V(t,x)$. We prove now that $V_{\infty}(x)=\inf_{t>0}V(t,x)$. Assume by contradiction that $V_{\infty}(x)<\inf_{t>0}V(t,x)$, and let $\varepsilon>0$ be such that $V_{\infty}(x)+2\varepsilon<\inf_{t>0}V(t,x)$. Take $u_{\varepsilon}\in{\cal U}_{[-\infty,0]}(0,x)$ such that $J_{[-\infty,0]}(u_{\varepsilon})<V_{\infty}(x)+\varepsilon$. By (5) we get $x=\int_{-\infty}^{0}e^{-\tau A}Bu_{\varepsilon}(\tau)\,\mathrm{d}\tau=e^{tA}y(-t)+\int_{-t}^{0}e^{-\tau A}Bu_{\varepsilon}(\tau)\,\mathrm{d}\tau\qquad\forall t>0;$ hence we have $u_{\varepsilon}|_{[-t,0]}\in{\cal U}_{[-t,0]}(y(-t),x)$, which in turn implies that $V(t,x-e^{tA}y(-t))\leq\frac{1}{2}\int_{-t}^{0}\|u_{\varepsilon}(s)\|_{U}^{2}\,\mathrm{d}s.$ (64) Now we observe that for every $\delta\in\,]0,1[\,$ we may choose $t_{\delta}>T_{0}+1$ such that $\|e^{tA}y(-t)\|_{H}\leq\delta$ for every $t>t_{\delta}$: indeed, by Hypothesis 2.7 and Lemma A.1-(v) we have $\displaystyle\|e^{tA}y(-t)\|_{H}$ $\displaystyle=$ $\displaystyle\|Q_{\infty}^{-1/2}e^{tA}y(-t)\|_{X}\leq\|Q_{\infty}^{-1/2}e^{A}\|_{{\cal L}(X)}\|e^{(t-1)A}y(-t)\|_{X}\leq$ $\displaystyle\leq$ $\displaystyle\|Q_{\infty}^{-1/2}e^{A}\|_{{\cal L}(X)}Me^{-\omega(t-1)}\|y(-t)\|_{X}\,.$ Since $y(-t)$ is uniformly bounded in $X$ for $t>0$, we have the claim. Going ahead with the proof, we recall that, by [1, Proposition 4.8-(iii)-(b)], we have uniform continuity of $V$ on $[T_{0},+\infty]\times B_{H}(0,R)$ for every $R>0$, where $B_{H}(0,R)$ is the ball of center $0$ and radius $R$ in $H$. So, setting $R=\|x\|_{H}+1$, and denoting by $\rho_{R}$ the continuity modulus of $V$ on $[T_{0},+\infty]\times B_{H}(0,R)$, we have for $t>t_{\delta}$ $V\left(t,x-e^{tA}y(-t)\right)>V(t,x)-\rho_{R}(\delta).$ The above, together with (64), implies that $V(t,x)-\rho_{R}(\delta)\leq V_{\infty}(x)+\varepsilon\qquad\forall t>t_{\delta}\,.$ Now it is enough to choose $\delta$ such that $\rho_{R}(\delta)<\varepsilon$ to get a contradiction. Finally the last statement follows from [1, Proposition 4.8-(iii)-(d)]. ∎ ### 4.3 Algebraic Riccati Equation We deal with the Algebraic Riccati Equation (ARE from now on) associated to our infinite horizon problem. As well known, when the value function is a quadratic form in the state space $X$, the ARE is an equation whose unknown is an operator $R$. A typical goal in studying such ARE is to prove that the operator representing the quadratic form in $X$ given by the value function is a solution (possibly unique) of the associated ARE. Formally our ARE is given as follows: $0=-\langle Ax,Ry\rangle_{X}-\langle Rx,Ay\rangle_{X}-\langle{B}^{*}Rx,{B}^{*}Ry\rangle_{U},\qquad x,y\in{\cal D}(A).$ (65) In our case (see Proposition 4.1) the value function $V_{\infty}$ is finite only in $H$ so that the operator $R$ above must be unbounded and the above equation makes sense only for $x,y\in{\cal D}(A)\cap{\cal D}(R)$. Moreover, by Proposition 4.3, $V_{\infty}$ is a quadratic form on the space $H$, represented by the identity operator $I_{H}\in{\cal L}(H)$, i.e. $V_{\infty}(x)=\frac{1}{2}\|x\|^{2}_{H}$. Consequently, transforming such norm in $X$, it must be $V_{\infty}(x)=\frac{1}{2}\|Q_{\infty}^{-1/2}\|_{X}^{2}$, and, when $x\in R(Q_{\infty})$, $V_{\infty}(x)=\frac{1}{2}\langle Q_{\infty}^{-1}x,x\rangle_{X}$. Hence it is natural to deduce that the operator representing $V_{\infty}$ in the space $X$ is $Q_{\infty}^{-1}$. Due to the unboundedness of the candidate solution $R$ of the ARE (65), it seems better to study the corresponding ARE in the space $H$, with unknown $P\in{\cal L}(H)$ whose form (taking $R=Q_{\infty}^{-1}P$) must be (compare with (14)): $0=-\langle Ax,Q^{-1}_{\infty}Py\rangle_{X}-\langle Q^{-1}_{\infty}Px,Ay\rangle_{X}-\langle{B}^{*}Q_{\infty}^{-1}Px,{B}^{*}Q_{\infty}^{-1}Py\rangle_{U}.$ (66) Note that such expression makes sense only when $Px,Py\in{\cal R}(Q_{\infty})$ and $x,y\in{\cal D}(A)\cap H$. By Proposition 4.3, we expect that the positive selfadjoint operator $P=I_{H}$ associated with the value function $V_{\infty}$ is a solution of the above ARE (66). Similarly we expect that $R=Q_{\infty}^{-1}$ is a solution of the above ARE (65). As they cannot be unique (the zero operator is always a solution of both), we somehow expect such solutions to be maximal in some suitable sense. ###### Remark 4.4. In the finite-dimensional case, when the operator $Q_{\infty}$ is invertible, it is proved that the operator $R=Q_{\infty}^{-1}$ solves (65), using the fact that its inverse $W=Q_{\infty}$ is the unique solution of the Lyapunov equation $AW+WA^{*}=-BB^{*}$ (67) among all definite positive bounded operators $X\rightarrow X$. This is reported by Scherpen [30, Theorem 2.2], who quotes Moore [25] for the proof (see also [21, Chapters 5 and 7] for related results). In fact, as we will see, this procedure works in our infinite dimensional case, too, but with more difficulties. ###### Definition 4.5. * (i) An operator $P\in{\cal L}_{+}(H)$ is a solution of the ARE (66) if the set ${\cal D}(A)\cap{\cal D}(\Lambda_{P})$ (see (43)) is dense in $H$ and the equation (66) is satisfied for all $x,y\in{\cal D}(A)\cap{\cal D}(\Lambda_{P})$. * (ii) A positive, selfadjoint, possibly unbounded operator $R:{\cal D}(R)\subset X\rightarrow X$ is a solution of the ARE (65) if the set ${\cal D}(A)\cap{\cal D}(R)$ is dense in $[\ker Q_{\infty}]^{\perp}$ (in the topology inherited by $X$) and the equation (65) is satisfied for all $x,y\in{\cal D}(A)\cap{\cal D}(R)$. ###### Proposition 4.6. The following facts are equivalent. (i) $P\in{\cal L}_{+}(H)$ is a solution to (66); (ii) $R=Q_{\infty}^{-1}P$ is a solution to (65) and it satisfies, in addition, $Q_{\infty}^{1/2}RQ_{\infty}^{1/2}\in{\cal L}(X)$. ###### Proof. (i) Assume that $P\in{\cal L}_{+}(H)$ solves (66). Then, in particular the set ${\cal D}(A)\cap{\cal D}(\Lambda_{P})$ is dense in $H$. Setting $R=Q_{\infty}^{-1}P$ we see that its domain is exactly ${\cal D}(\Lambda_{P})$, which is dense in $[\ker Q_{\infty}]^{\perp}$. The fact that such $R$ satisfies (65) for every $x,y\in{\cal D}(A)\cap{\cal D}(\Lambda_{P})$ follows by simple substitution. Finally, for every $x\in X$ we have $\|Q_{\infty}^{1/2}RQ_{\infty}^{1/2}x\|_{X}=\|Q_{\infty}^{-1/2}PQ_{\infty}^{1/2}x\|_{X}=\|PQ_{\infty}^{1/2}x\|_{H}\leq\|P\|_{{\cal L}(H)}\|Q_{\infty}^{1/2}x\|_{H}=\|P\|_{{\cal L}(H)}\|x\|_{X}\,.$ (ii) Let $R:{\cal D}(R)\rightarrow X$ be a solution of (65), having the property $Q_{\infty}^{1/2}RQ_{\infty}^{1/2}\in{\cal L}(X)$: note that, in this case, ${\cal D}(R)$ must coincide with $H$. Thus ${\cal D}(A)\cap{\cal D}(R)$ is dense in H, since it contains ${\cal D}(A_{0})$. We set $P=Q_{\infty}R$: then $P\in{\cal L}_{+}(H)$ since, for every $x\in H$, $\displaystyle\|Px\|_{H}$ $\displaystyle=$ $\displaystyle\|Q_{\infty}Rx\|_{H}=\|Q_{\infty}^{1/2}[Q_{\infty}^{1/2}RQ_{\infty}^{1/2}]Q_{\infty}^{-1/2}x\|_{H}=\|[Q_{\infty}^{1/2}RQ_{\infty}^{1/2}]Q_{\infty}^{-1/2}x\|_{X}\leq$ $\displaystyle\leq$ $\displaystyle\|Q_{\infty}^{1/2}RQ_{\infty}^{1/2}\|_{{\cal L}(X)}\|Q_{\infty}^{-1/2}x\|_{X}=\|Q_{\infty}^{1/2}RQ_{\infty}^{1/2}\|_{{\cal L}(X)}\|x\|_{H}\,.$ Moreover, we see immediately that ${\cal D}(\Lambda_{P})=H$. In addition, (65) transforms into (66), and it holds for every $x,y\in{\cal D}(A)\cap{\cal D}(R)$, i.e. it holds for every $x,y\in{\cal D}(A)\cap H={\cal D}(A)\cap{\cal D}(\Lambda_{P})$, as required by Definition 4.5. ∎ Concerning the two AREs (66) and (65) we have the following result. ###### Theorem 4.7. Let Hypothesis 2.7 hold true. * (i) The operator $R=Q_{\infty}^{-1}$ is a solution of the Riccati equation (65) in the sense of Definition 4.5(ii). * (ii) The operator $P=I_{H}$ is a solution of the Riccati equation (66) in the sense of Definition 4.5(i). * (iii) Assume that $BB^{*}$ is coercive. Then the operator $I_{H}$ is the maximal solution of (66) in the following sense: if $\hat{P}$ is another solution of (66) in the sense of Definition 4.5-(i), belonging to the class ${\cal Q}$ introduced in Definition 3.8, then $\frac{1}{2}\langle\hat{P}x,x\rangle_{H}\leq\frac{1}{2}\langle x,x\rangle_{H}=V_{\infty}(x)\qquad\forall x\in H.$ ###### Proof. (i) By [1, Proposition 3.3], $Q_{\infty}$ solves the Lyapunov equation, i.e. we have for every $\xi\in{\cal D}(A^{*})$ $AQ_{\infty}\xi+Q_{\infty}A^{*}\xi+BB^{*}\xi=0.$ This implies that, for every $\xi\in{\cal D}(A^{*})$ and $\eta\in X$, $\langle AQ_{\infty}\xi,\eta\rangle_{X}+\langle Q_{\infty}A^{*}\xi,\eta\rangle_{X}+\langle B^{*}\xi,B^{*}\eta\rangle_{U}=0.$ When $\eta\in{\cal D}(AQ_{\infty})$ the second term above rewrites as $\langle\xi,AQ_{\infty}\eta\rangle_{X}$. Consequently, when $\eta\in{\cal D}(AQ_{\infty})$, the functional $\xi\rightarrow\langle AQ_{\infty}\xi,\eta\rangle_{X}$, well defined since $\xi\in{\cal D}(A^{*})$, can be extended to a bounded linear operator on $X$, since it is equal to $-\langle\xi,AQ_{\infty}\eta\rangle_{X}-\langle B^{*}\xi,B^{*}\eta\rangle_{U}$. Hence, choosing $\xi\in{\cal D}(AQ_{\infty})$, we get, for $\xi,\eta\in{\cal D}(AQ_{\infty})$, that $\langle AQ_{\infty}\xi,\eta\rangle_{X}+\langle\xi,AQ_{\infty}\eta\rangle_{X}+\langle B^{*}\xi,B^{*}\eta\rangle_{U}=0.$ (68) Now set $x=Q_{\infty}\xi$ and $y=Q_{\infty}\eta$. Then $x,y\in{\cal D}(A)$ and the above rewrites as $\langle Ax,\eta\rangle_{X}+\langle\xi,Ay\rangle_{X}+\langle B^{*}\xi,B^{*}\eta\rangle_{U}=0.$ (69) Observe that $\xi=Q_{\infty}^{-1}x+\xi_{0}$ and $\eta=Q_{\infty}^{-1}y+\eta_{0}$ for suitable $\xi_{0},\eta_{0}\in\ker Q_{\infty}\subseteq\ker B^{*}$. Hence, using the fact that $Q_{\infty}$ solves the Lyapunov equation in the form (68), we have, for $\xi\in{\cal D}(AQ_{\infty})$, $\langle Ax,\eta_{0}\rangle_{X}=\langle AQ_{\infty}\xi,\eta_{0}\rangle_{X}=-\langle\xi,AQ_{\infty}\eta_{0}\rangle_{X}-\langle B^{*}\xi,B^{*}\eta_{0}\rangle_{U}=0$ and, similarly, for $\eta\in{\cal D}(AQ_{\infty})$, $\langle\xi_{0},Ay\rangle_{X}=0$. We then get, substituting into (69) and observing that $B^{*}\xi_{0}=B^{*}\eta_{0}=0$, $\langle Ax,Q_{\infty}^{-1}y\rangle_{X}+\langle Q_{\infty}^{-1}x,Ay\rangle_{X}+\langle B^{*}Q_{\infty}^{-1}x,B^{*}Q_{\infty}^{-1}y\rangle_{U}=0,\quad+x,y\in Q_{\infty}({\cal D}(AQ_{\infty})).$ (70) The above is exactly equation (65) for $R=Q_{\infty}^{-1}$. To end the proof of (i), it is enough to observe that $Q_{\infty}({\cal D}(AQ_{\infty}))$ is dense in $[\ker Q_{\infty}]^{\perp}$ (using Remark A.3 and the fact that it contains $Q_{\infty}({\cal D}(A^{*}))$), and moreover that $Q_{\infty}({\cal D}(AQ_{\infty}))={\cal D}(A)\cap{\cal R}(Q_{\infty})={\cal D}(A)\cap{\cal D}(Q_{\infty}^{-1}).$ Indeed if $x\in Q_{\infty}({\cal D}(AQ_{\infty}))$ then it must be $x=Q_{\infty}\xi$ with $\xi\in({\cal D}(AQ_{\infty}))$, so that $AQ_{\infty}\xi$ is well defined and, clearly, it coincides with $Ax$, proving that $x\in{\cal D}(A)$. Obviously it must also be $x\in{\cal R}(Q_{\infty})$. The converse is similar. (ii) It is enough to observe that (70) coincides with (66) with $P=I_{H}$, and that ${\cal D}(\Lambda_{I_{H}})=R(Q_{\infty})$. (iii) Let $\hat{P}$ be a solution of (66) belonging to the class ${\cal Q}$ introduced in Definition 3.8. It is immediate to see that $\hat{P}$ is a stationary solution of (40) in the sense of Definition 3.6. Now we apply Lemma 3.10 and (36), getting $\frac{1}{2}\langle\hat{P}x,x\rangle_{H}\leq V^{\hat{P}}(t,x)\leq V(t,x)\quad x\in H,\quad t>T_{0}.$ Taking the limit as $t\rightarrow+\infty$, the result follows by Proposition 4.3. ∎ ###### Remark 4.8. The statement of Theorem 4.7 still holds if we consider the slightly more general problem where the energy functional has the integrand $\langle Cu,u\rangle_{U}$ instead of $\langle u,u\rangle_{U}\,$, where $C\in{\mathcal{L}}_{+}(U)$ is coercive and hence invertible. Indeed it is enough to define the new control variable $v=C^{1/2}u$ and, consequently, to replace the control operator $B$ in the state equation by $BC^{-1/2}$. ###### Remark 4.9. Theorem 4.7 can be applied to a variety of cases (e.g. delay equations treated in [1, Subsection 5.1] or wave equations). Here, according to our motivating example arising in physics, we develop more deeply the analysis when the operator $A$ is selfadjoint and commutes with $BB^{*}$ and, in particular, when both are diagonal. This will be done in the next section. ## 5 The selfadjoint commuting case We consider the case where $A$ is selfadjoint and invertible and commutes with $BB^{*}$. To apply Theorem 4.7 we need that $BB^{*}$ is coercive; hence we assume the following: ###### Hypothesis 5.1. $A$ is selfadjoint and invertible and commutes with $BB^{*}$ i.e. for every $x\in\mathcal{D}(A)$ we have $BB^{*}x\in\mathcal{D}(A)$ and $ABB^{*}x=BB^{*}Ax$. Moreover, $BB^{*}$ is coercive, i.e., for a suitable $\mu>0$, $\|B^{*}x\|_{U}\geq\mu\|x\|_{X}$ for all $x\in X$. From [1, Proposition C.1-(v)] we know that, for every $x\in X$, $Q_{\infty}x=-\frac{1}{2}A^{-1}BB^{*}x.$ This implies that ${\cal R}(Q_{\infty})={\cal D}(A)$, and, as $BB^{*}$ is invertible in $X$, we have $Q_{\infty}^{-1}x=-2(BB^{*})^{-1}Ax$ for every $x\in{\cal R}(Q_{\infty})$ (see again [1, Proposition C.1-(v)]). Hence the Riccati equation (66) in $H$ (with unknown $P\in{\cal L}(H)$), becomes $0=-\langle Ax,Q_{\infty}^{-1}Py\rangle_{X}-\langle Q_{\infty}^{-1}Px,Ay\rangle_{X}+2\langle APx,Q_{\infty}^{-1}Py\rangle_{X}.$ (71) This makes sense, as for (66), when $x,y\in{\cal D}(A)\cap{\cal D}(\Lambda_{P})$ (see Definition 3.6). We now want to rewrite this equation using the inner products in $H$. Observe first that in ${\cal R}(Q_{\infty})$ we have $Q_{\infty}^{-1}=Q_{\infty}^{-1/2}Q_{\infty}^{-1/2}$. Then, if $Ax$, $Ay$ and $APx$ belong to $H$, we rewrite (71) as $0=-\langle Ax,Py\rangle_{H}-\langle Px,Ay\rangle_{H}+2\langle APx,Py\rangle_{H}.$ (72) Now, recalling the definition of $A_{0}$ (see Notation 3.7 and Lemma A.4)-(ii)), equation (72) can be equivalently rewritten as $0=-\langle A_{0}x,Py\rangle_{H}-\langle Px,A_{0}y\rangle_{H}+2\langle A_{0}Px,Py\rangle_{H},$ (73) provided that $x,y,Px,Py$ belong to ${\cal D}(A_{0})$. We now clarify the relationship between (71) and (73). First we set $D^{P}:=\left\\{x\in{\cal D}(A_{0}):\;Px\in{\cal D}(A_{0})\right\\}.$ (74) Next, we provide the following definition of solution for (73) (compare with Definition 4.5): ###### Definition 5.2. An operator $P\in{\cal L}_{+}(H)$ is a solution of the ARE (73) if the set $D^{P}$ is dense in $H$ and the equation (73) is satisfied for every $x,y\in D^{P}$. Finally, we observe that every solution of (71) is also a solution of (73): indeed, if $P\in{\cal L}_{+}(H)$, then, by definition, we have $D^{P}\subseteq{\cal D}(A)\cap{\cal D}(\Lambda_{P})$. Hence, if $P\in{\cal L}_{+}(H)$ solves equation (71), then, choosing in particular $x,y\in D^{P}$ we can turn (71) into (73). The reverse procedure is also possible: we postpone the proof at the end of the Section, since some more informations on solutions $P$ of (73) are needed. We now give a preparatory result about the properties of such solutions. ###### Proposition 5.3. Assune Hypothesis 5.1. Then any solution $P$ of (73) satisfies $\langle A_{0}x,A_{0}Pz\rangle_{H}=\langle A_{0}Px,A_{0}z\rangle_{H}\qquad\forall x,z\in D^{P}.$ (75) ###### Proof. Let $P$ be a solution of (73). We observe that for all $x,y\in D^{P}$ we have, since $A_{0}$ is selfadjoint in $H$ (see Lemma (A.5)-(iii)), $\langle A_{0}Px,y\rangle_{H}+\langle PA_{0}x,y\rangle_{H}=2\langle PA_{0}Px,y\rangle_{H}\,.$ (76) By density, this equation holds for every $x\in D^{P}$ and $y\in H$. Symmetrically we have also $\langle x,PA_{0}y\rangle_{H}+\langle x,A_{0}Py\rangle_{H}=2\langle x,PA_{0}Py\rangle_{H}$ (77) for every $x\in H$ and $y\in D^{P}$. We choose in (76) $y=PA_{0}z-A_{0}Pz$, with $z\in D^{P}$, and we obtain: $\displaystyle\langle A_{0}Px,PA_{0}z\rangle_{H}-\langle A_{0}Px,A_{0}Pz\rangle_{H}+\langle PA_{0}x,PA_{0}z\rangle_{H}-\langle PA_{0}x,A_{0}Pz\rangle_{H}$ $\displaystyle\qquad\qquad\qquad\qquad\qquad\qquad=2\langle PA_{0}Px,PA_{0}z\rangle_{H}-2\langle PA_{0}Px,A_{0}Pz\rangle_{H}\,.$ We isolate on the left the symmetric terms: $\displaystyle 2\langle PA_{0}Px,A_{0}Pz\rangle_{H}-\langle A_{0}Px,A_{0}Pz\rangle_{H}+\langle PA_{0}x,PA_{0}z\rangle_{H}$ $\displaystyle\qquad\qquad\qquad=-\langle A_{0}Px,PA_{0}z\rangle_{H}+\langle PA_{0}x,A_{0}Pz\rangle_{H}+2\langle PA_{0}Px,PA_{0}z\rangle_{H}\,.$ Next, we apply (76) to the last term on the right: $\displaystyle 2\langle PA_{0}Px,A_{0}Pz\rangle_{H}-\langle A_{0}Px,A_{0}Pz\rangle_{H}+\langle PA_{0}x,PA_{0}z\rangle_{H}$ $\displaystyle=-\langle A_{0}Px,PA_{0}z\rangle_{H}+\langle PA_{0}x,A_{0}Pz\rangle_{H}+\langle A_{0}Px,PA_{0}z\rangle_{H}+\langle PA_{0}x,PA_{0}z\rangle_{H}\,,$ which simplifies to $2\langle PA_{0}Px,A_{0}Pz\rangle_{H}-\langle A_{0}Px,A_{0}Pz\rangle_{H}=\langle PA_{0}x,A_{0}Pz\rangle_{H}\,.$ Applying (77) to the term on the right, rewritten as $\langle A_{0}x,PA_{0}Px\rangle_{H}$, we obtain for every $x,z\in D^{P}$ $2\langle PA_{0}Px,A_{0}Pz\rangle_{H}-\langle A_{0}Px,A_{0}Pz\rangle_{H}-\frac{1}{2}\langle PA_{0}x,A_{0}z\rangle_{H}=\frac{1}{2}\langle A_{0}x,A_{0}Pz\rangle_{H}\,.$ (78) We now restart from (77), and choose $x=PA_{0}z-A_{0}Pz$, with $z\in D^{P}$: acting on the left variable of the inner product, and proceeding exactly in the same way as before, we get for every $z,y\in D^{P}$ $2\langle A_{0}Pz,PA_{0}Py\rangle_{H}-\langle A_{0}Pz,A_{0}Py\rangle_{H}-\frac{1}{2}\langle PA_{0}z,A_{0}y\rangle_{H}=\frac{1}{2}\langle A_{0}Pz,A_{0}y\rangle_{H}\,.$ (79) Comparing equations (78) and (79), both written with variables $x,y$, we immediately obtain $\frac{1}{2}\langle A_{0}x,A_{0}Py\rangle_{H}=\frac{1}{2}\langle A_{0}Px,A_{0}y\rangle_{H}\,,\quad x,y\in D^{P},$ which is (75). ∎ We can now prove: ###### Theorem 5.4. Assume Hypothesis 5.1. Then any solution $P$ of (73) commutes with $A_{0}$, i.e. $Px\in{\cal D}(A_{0})$ for every $x\in{\cal D}(A_{0})$ and $A_{0}Px=PA_{0}x\qquad\forall x\in{\cal D}(A_{0}).$ In particular $D^{P}={\cal D}(A_{0})$. ###### Proof. We start from (75) with $w=A_{0}x$ and $y=A_{0}z$, i.e. $\langle w,A_{0}PA_{0}^{-1}y\rangle_{H}=\langle A_{0}PA_{0}^{-1}w,y\rangle_{H}\qquad\forall w,y\in A_{0}(D^{P}).$ (80) Notice that $A_{0}(D^{P})$ is the natural domain of the operator $A_{0}PA_{0}^{-1}$; which might be (a priori) not dense in $H$. Let us denote by $Z$ the closure of ${\cal D}(A_{0}PA_{0}^{-1})$in $H$; so we have $Z:=\overline{A_{0}(D^{P})}=\overline{{\cal D}(A_{0}PA_{0}^{-1})}.$ Obvoiusly $Z$ is a Hilbert space with the inner product of $H$. Equation (80) then tells us that $A_{0}(D^{P})\subseteq{\cal D}((A_{0}PA_{0}^{-1})^{*})$ and $(A_{0}PA_{0}^{-1})^{*}w=A_{0}PA_{0}^{-1}w\qquad\forall w\in A_{0}(D^{P})={\cal D}(A_{0}PA_{0}^{-1}).$ (81) On the other hand, if $x\in{\cal D}(A_{0})$ and $y\in{\cal D}(A_{0}PA_{0}^{-1})$ we may write $\langle x,A_{0}PA_{0}^{-1}y\rangle_{H}=\langle A_{0}^{-1}PA_{0}x,y\rangle_{H}\,;$ consequently ${\cal D}(A_{0})\subseteq{\cal D}((A_{0}PA_{0}^{-1})^{*})$ (82) and $(A_{0}PA_{0}^{-1})^{*}x=A_{0}^{-1}PA_{0}x\qquad\forall x\in{\cal D}(A_{0}).$ (83) We now claim that $A_{0}PA_{0}^{-1}$ is selfadjoint in the space $H$, i.e. ${\cal D}((A_{0}PA_{0}^{-1})^{*})={\cal D}(A_{0}PA_{0}^{-1})=A_{0}(D^{P})$ (84) is dense in $H$ and (81) holds. Indeed, assume that $z\in{\cal D}((A_{0}PA_{0}^{-1})^{*})$: then there is $c>0$ such that $|\langle A_{0}PA_{0}^{-1}x,z\rangle_{H}|\leq c\|x\|_{H}\qquad\forall x\in{\cal D}(A_{0}PA_{0}^{-1}).$ In particular, by (81), $\langle x,(A_{0}PA_{0}^{-1})^{*}z\rangle_{H}=\langle A_{0}PA_{0}^{-1}x,z\rangle_{H}=\langle(A_{0}PA_{0}^{-1})^{*}x,z\rangle_{H}\quad\forall x\in{\cal D}(A_{0}PA_{0}^{-1}).$ This shows that $z\in{\cal D}(A_{0}PA_{0}^{-1})$ and $A_{0}PA_{0}^{-1}z=(A_{0}PA_{0}^{-1})^{*}z$. Hence ${\cal D}((A_{0}PA_{0}^{-1})^{*})\subseteq{\cal D}(A_{0}PA_{0}^{-1})\quad\textrm{and}\quad A_{0}PA_{0}^{-1}=(A_{0}PA_{0}^{-1})^{*}\ \textrm{on}\ {\cal D}((A_{0}PA_{0}^{-1})^{*}).$ Conversely, we know from (81) that ${\cal D}(A_{0}PA_{0}^{-1})=A_{0}(D^{P})\subseteq{\cal D}((A_{0}PA_{0}^{-1})^{*})\quad\textrm{and}\quad(A_{0}PA_{0}^{-1})^{*}=A_{0}PA_{0}^{-1}\ \textrm{on}\ {\cal D}(A_{0}PA_{0}^{-1}).$ In particular, by (82), $Z$ coincides with $H$, i.e. both domains in (84) are dense in $H$. This proves our claim. Take now $x\in{\cal D}(A_{0})$. As, by (82), $D(A^{0})\subseteq{\cal D}(A_{0}PA_{0}^{-1})$, we have $PA_{0}^{-1}\in{\cal D}(A_{0})\quad\forall x\in{\cal D}(A_{0}),\quad\textrm{i.e.}\quad D^{P}={\cal D}(A_{0}).$ (85) (see (74)). Moreover, by (83) and by the above claim we deduce $A_{0}^{-1}PA_{0}x=A_{0}PA_{0}^{-1}x\qquad\forall x\in{\cal D}(A_{0}).$ Applying $A_{0}^{-1}$ we have $A_{0}^{-2}PA_{0}x=PA_{0}^{-1}x$ for every $x\in{\cal D}(A_{0})$, or, equivalently, $A_{0}^{-2}PA_{0}^{2}z=z\qquad\forall z\in{\cal D}(A_{0}^{2}),\qquad\textrm{i.e.}\qquad A_{0}^{-2}Pw=PA_{0}^{-2}w\qquad\forall w\in H.$ This means that the bounded operators $A_{0}^{-2}$ and $P$ commute. Now, since $A_{0}^{-1}$ is a non-negative operator such that $(A_{0}^{-1})^{2}=A_{0}^{-2}$, by a well known result (see [29, Theorem VI.9]), $A_{0}^{-1}$ must commute with every bounded operator $B$ which commutes with $A_{0}^{-2}$, for instance $B=P$. So $A_{0}^{-1}Pw=PA_{0}^{-1}w\qquad\forall w\in H,\qquad\textrm{i.e.}\qquad Pz=A_{0}^{-1}PA_{0}z\qquad\forall z\in{\cal D}(A_{0});$ this implies that $P({\cal D}(A_{0}))\subseteq{\cal D}(A_{0})$ and $A_{0}Pz=PA_{0}z$ for every $z\in{\cal D}(A_{0})$. Thus $P$ commutes with $A_{0}$, as required. Moreover $P({\cal D}(A_{0}))\subseteq{\cal D}(A_{0})$ implies ${\cal D}(A_{0})\subseteq D^{P}$. The reverse inclusion immediately follows from the definition of $D^{P}$. ∎ We are now able to characterize all solutions of the ARE (73). ###### Theorem 5.5. Assume Hypothesis 5.1 and let $P\in{\cal L}_{+}(H)$. Then $P$ is a solution of (73) if and only if $P$ is an orthogonal projection in $H$ and it commutes with $A_{0}$. In particular the identity $I_{H}$ is the maximal solution among all solutions of (73). ###### Proof. Let $P$ be a solution of (73): by Theorem 5.4 we have $Px\in{\cal D}(A_{0})$ for every $x\in{\cal D}(A_{0})$ and $A_{0}Px=PA_{0}x$. Hence the ARE (76), equivalent to (73), becomes $0=-2\langle PA_{0}x,y\rangle_{H}+2\langle PA_{0}Px,y\rangle_{H},\quad x\in D^{P},\ y\in H.$ Since $y$ is arbitrary, using (85) we get $2PA_{0}x=2PA_{0}Px$ for every $x\in{\cal D}(A_{0})$, and successively, for all $x\in D(A_{0})$, $PA_{0}x-PA_{0}Px=0$ , $PA_{0}(I_{H}-P)x=0$, $A_{0}P(I_{H}-P)x=0$, $P(I_{H}-P)x=0$, $Px=P^{2}x$; finally, by density, $P=P^{2}$. Assume, conversely, that $P$ is an orthogonal projection in $H$ and it commutes with $A_{0}$. Then $PA_{0}Pz=P^{2}A_{0}z=PA_{0}z=A_{0}Pz\quad\forall z\in{\cal D}(A_{0}),$ and consequently $P$ solves (76). Finally, since $I_{H}$ solves (73), the last statement is immediate. ∎ We conclude this Section proving the equivalence of the two forms (71) and (73) of the ARE. ###### Proposition 5.6. Every solution of (71) is also a solution of (73) and vice versa. ###### Proof. We have already seen that every solution of (71) is also a solution of (73). Consider now a solution $P$ of (73). First of all, if $x,y\in D^{P}={\cal D}(A_{0})$, equation (73) transforms into (71), so that (71) holds true for $x,y\in D^{P}$. We claim that $D^{P}$ is dense in ${\cal D}(A)\cap{\cal D}(\Lambda_{P})$ (see (74)) with respect to the norm $\|\cdot\|_{H}+\|A\cdot\|_{X}+\|AP\cdot\|_{X}$. Indeed, for $z\in{\cal D}(A)\cap{\cal D}(\Lambda_{P})$, recalling Lemma A.4, we set $z_{n}=nR(n,A)z=nR(n,A)|_{H}\,z=nR(n,A_{0})z.$ Then $z_{n}\in{\cal D}(A_{0})=D^{P}$ and, as $n\rightarrow\infty$, $\begin{array}[]{c}z_{n}\rightarrow z\quad\textrm{in }H,\\\\[5.69054pt] A_{0}z_{n}=nA_{0}R(n,A_{0})z=nAR(n,A)z\rightarrow Az\quad\textrm{in }X,\\\\[5.69054pt] A_{0}Pz_{n}=nA_{0}PR(n,A_{0})z=nAR(n,A)Pz\rightarrow APz\quad\textrm{in }X;\end{array}$ this proves our claim. Let now $x,y\in{\cal D}(A)\cap{\cal D}(\Lambda_{P})$; select $\\{x_{n}\\},\\{y_{n}\\}\subseteq{\cal D}(A_{0})$ such that, as $n\rightarrow\infty$, $\begin{array}[]{l}x_{n}\rightarrow x\ \textrm{in }H,\quad Ax_{n}\rightarrow Ax\ \textrm{in }X,\quad APx_{n}\rightarrow APx\ \textrm{in }X,\\\\[5.69054pt] y_{n}\rightarrow y\ \textrm{in }H,\quad Ay_{n}\rightarrow Ay\ \textrm{in }X,\quad APy_{n}\rightarrow APy\ \textrm{in }X.\end{array}$ As a consequence, $Q_{\infty}^{-1}Px_{n}=-2BB^{*}APx_{n}\rightarrow-2BB^{*}APx=Q_{\infty}^{-1}Px\ \textrm{in }X\ \textrm{as }n\rightarrow\infty,$ and similarly $Q_{\infty}^{-1}Px_{n}\rightarrow Q_{\infty}^{-1}Py$ in $X$ as $n\rightarrow\infty$. For $x_{n}$ and $y_{n}$, (71) holds: $0=-\langle Ax_{n},Q_{\infty}^{-1}Py_{n}\rangle_{X}-\langle Q_{\infty}^{-1}Px_{n},Ay_{n}\rangle_{X}+2\langle APx_{n},Q_{\infty}^{-1}Py_{n}\rangle_{X}.$ In all terms, by what established above, we can pass to the limit as $n\rightarrow\infty$, obtaining $0=-\langle Ax,Q_{\infty}^{-1}Py\rangle_{X}-\langle Q_{\infty}^{-1}Px,Ay\rangle_{X}+2\langle APx,Q_{\infty}^{-1}Py\rangle_{X}\qquad\forall x,y\in{\cal D}(A)\cap{\cal D}(\Lambda_{P}),$ i.e. $P$ solves (71). ∎ ###### Remark 5.7. It is easy to verify that for every solution $P$ of (73) the space $D^{P}={\cal D}(A_{0})$ is dense in ${\cal D}(A)\cap H$ with respect to the norm $\|\cdot\|_{H}+\|A\cdot\|_{X}$: it suffices to repeat the argument above, i.e. to consider, for fixed $x\in{\cal D}(A)\cap H$, the approximation $x_{n}=nR(n,A_{0})x$, observing that $x_{n}\rightarrow x$ in $H$ and $A_{0}x_{n}=Ax_{n}\rightarrow Ax$ in $X$. Thus, $P$ belongs to the class ${\cal Q}$ introduced in Definition 3.8, and consequently, by Theorem 4.7, we have $P\leq I_{H}$. Of course, this follows as well by Theorem 5.5. ###### Corollary 5.8. Assume that $A_{0}$ is a diagonal operator with respect to an orthonormal complete system $\\{e_{n}\\}$ in $H$ with sequence of eigenvalues $\\{\lambda_{n}\\}\subset\,]-\infty,0[\,$. Let $P$ be a solution of the ARE (73). Then * (i) every eigenspace of $A_{0}$ is invariant for $P$; * (ii) if all eigenvalues are simple, then $P$ is diagonal with respect to the system $\\{e_{n}\\}$, too; * (iii) if at least one eigenspace $M$ has dimension $m\geq 2$, then the restriction of $P$ to $M$ needs not be diagonal: for instance, if $m=2$ a non-diagonal $P$ on $M$ must have the following explicit form: $\left(\begin{array}[]{cc}a&\pm\sqrt{a(1-a)}\\\ \pm\sqrt{a(1-a)}&1-a\\\ \end{array}\right)\qquad\textrm{for some }\ a\in\,]0,1[\,.$ (86) ###### Proof. To prove (i) it is enough to show that, for every eigenvalue $\lambda$ of $A_{0}$ and $x$ eigenvector of $A_{0}$ associated to $\lambda$, we have $\lambda Px=A_{0}Px$. This is immediate since $A_{0}$ and $P$ commute. Concerning (ii) we observe that, for every $n\in\mathbb{N}$ we have $A_{0}e_{n}=\lambda_{n}e_{n}$, so that $\lambda_{n}Pe_{n}=A_{0}Pe_{n}$. Since $\lambda_{n}$ is simple, it is $Pe_{n}=ke_{n}$ for some $k\in\mathbb{R}$. Since $P$ is a projection, it must be $k=0$ or $k=1$. Finally (iii) can be proved with straightforward algebraic calculations, using the fact that $M$ is invariant under $P$ and that $P$ is a projection. ∎ ###### Remark 5.9. Let $A$ be a diagonal operator with respect to an orthonormal complete system $\\{e_{n}\\}$ in $H$ with sequence of eigenvalues $\\{\lambda_{n}\\}\subset\,]-\infty,0[\,$, where all $\lambda_{n}$ are distinct and simple. Then $BB^{*}$ must be diagonal, too. Indeed we have, for every $n\in\mathbb{N}$, $\sum_{k=0}^{+\infty}\langle BB^{*}e_{n},e_{k}\rangle_{H}\,e_{k}=BB^{*}e_{n}=\frac{1}{\lambda_{n}}BB^{*}Ae_{n}=\frac{1}{\lambda_{n}}ABB^{*}e_{n}=\frac{1}{\lambda_{n}}\sum_{k=0}^{+\infty}\lambda_{k}\langle BB^{*}e_{n},e_{k}\rangle_{H}\,e_{k}\,,$ which implies $\langle BB^{*}e_{n},e_{k}\rangle_{H}\left(1-\frac{\lambda_{k}}{\lambda_{n}}\right)=0\quad\forall k,n\in\mathbb{N}.$ Since all eigenvalues are distinct, it must be $BB^{*}e_{n}=b_{n}e_{n}$ for all $n\in\mathbb{N}$ for a suitable sequence $\\{b_{n}\\}\in\ell^{\infty}$. This implies that $Q_{\infty}$ and $Q_{\infty}^{1/2}$ are diagonal with respect to $\\{e_{n}\\}$, too. Following [1, Subsection 5.2] we may also consider the case when $BB^{*}$ is unbounded and characterize the space $H$, for specific choices of $BB^{*}$, in terms of the domain of suitable powers of $(-A)$. In Section 6 we will consider a specific diagonal case arising in mathematical physics. ## 6 A motivating example: from equilibrium to non-equilibrium states In this section we describe, in a simple one-dimensional case, the optimal control problem outlined in the papers [4, 5, 6, 7, 8, 9]. Such special case fits into the application studied e.g. in [7, 9], in the case of the Landau- Ginzburg model. We consider a controlled dynamical system whose state variable is described by a function $\rho:\left]-\infty,0\right]$ (the choice of the letter $\rho$ comes from the fact that in many physical models $\rho$ is a density). The control variable is a function $F:\left]-\infty,0\right]\times\left[0,1\right]\rightarrow\mathbb{R}$ which we assume to belong to $L^{2}\left(-\infty,0;L^{2}(0,1)\right)$. The state equation is formally given by $\begin{cases}\frac{\partial\rho}{\partial t}\left(t,x\right)=\frac{1}{2}\frac{\partial^{2}\rho}{\partial x^{2}}\left(t,x\right)+\nabla F\left(t,x\right),&t\in\,]-\infty,0[\,,\ x\in\,]0,1[\,,\\\\[2.84526pt] \rho\left(-\infty,x\right)=\bar{\rho}(x),&x\in[0,1],\\\ \rho\left(t,0\right)=\rho_{-},\qquad\rho\left(t,1\right)=\rho_{+},&t\in\,]-\infty,0[\,,\\\ \rho\left(0,x\right)=\rho_{0}\left(x\right),&x\in[0,1],\end{cases}$ (87) where $\rho_{+},\rho_{-}\in\left(0,1\right)$, and $\bar{\rho}$ is an equilibrium state for the uncontrolled problem. Hence $\bar{\rho}$ is the unique solution of the following system $\left\\{\begin{array}[]{l}v^{\prime\prime}\left(x\right)=0,\\\ v\left(0\right)=\rho_{-},\\\ v\left(1\right)=\rho_{+};\end{array}\right.$ so we have $\bar{\rho}(x)=(\rho_{+}-\rho_{-})x+\rho_{-}$. For any datum $\rho_{0}\in L^{2}(0,1)$ we consider any control driving (in equation (87)) the equilibrium state $\bar{\rho}$ (at time $t=-\infty$) to $\rho_{0}$ at time $t=0$. Then we consider the problem of minimizing, over the set of such controls, the energy functional $J^{0}_{\infty}\left(F\right)=\frac{1}{2}\int_{-\infty}^{0}\|F(s)\|^{2}_{L^{2}(0,1)}\,\mathrm{d}s.$ Given the above structure it is natural to consider the new control $\nu=\nabla F\in L^{2}\left(-\infty,0;H^{-1}(0,1)\right)$ and take both the state space $X$ and the control space $U$ equal to $H^{-1}\left(0,1\right)$. We now rewrite (87) in our abstract setting as follows. First we denote by $A$ the Laplace operator in the space $H^{-1}(0,1)$ with Dirichlet boundary conditions, i.e. ${\cal D}\left(A\right)=H^{1}_{0}\left(0,1\right),\qquad A\eta=\eta^{\prime\prime}\quad\forall\eta\in H^{1}_{0}\left(0,1\right).$ Hence, formally, the state equation (87) becomes $\left\\{\begin{array}[]{l}\rho^{\prime}(t)=A[\rho(t)-\bar{\rho}]+\nu(t),\quad t<0,\\\\[2.84526pt] \rho(-\infty)=\bar{\rho}.\end{array}\right.$ (88) Using a standard argument (see e.g. [17, Appendix C]), the state equation (87) can be rewritten in the space $X$ and in the new variable $y(t):=\rho(t)-\bar{\rho}$ as $\left\\{\begin{array}[]{l}y^{\prime}(t)=Ay(t)+\nu(t),\quad t<0,\\\\[2.84526pt] y(-\infty)=0.\end{array}\right.$ (89) The function $y(t;-\infty,0,\nu)=\int_{-\infty}^{t}e^{(t-s)A}\nu(s)\,\mathrm{d}s,\qquad t\leq 0,$ (90) corresponding to $\rho(t;\nu)=\bar{\rho}+\int_{-\infty}^{t}e^{(t-s)A}\nu(s)\,\mathrm{d}s$, is the unique solution of (89), adopting Definition 2.4 and applying Lemma 2.5. The energy functional, in the new control variable $\nu$, becomes $\bar{J}^{0}_{\infty}\left(\nu\right)=\frac{1}{2}\int_{-\infty}^{0}\|A^{-1/2}\nu(s)\|^{2}_{L^{2}(0,1)}\,\mathrm{d}s=\frac{1}{2}\int_{-\infty}^{0}\|\nu(s)\|^{2}_{H^{-1}(0,1)}\,\mathrm{d}s.$ The set of admissible controls here is exactly $\mathcal{U}_{[-\infty,0]}(0,y_{0})$ (see Subsection 2.2), which is nonempty if and only if $y_{0}\in H:=R(Q_{\infty}^{1/2})=D(A^{1/2})=L^{2}(0,1)$ (see e.g. [1, Section 5.2]). The value function $V_{\infty}$ is defined as $V_{\infty}\left(y_{0}\right):=\inf_{\nu\in\mathcal{U}_{[-\infty,0]}(0,y_{0})}\bar{J}^{0}_{\infty}\left(\nu\right).$ (91) Now, recalling that $X=U=H^{-1}(0,1)$ and setting $B=I_{H^{-1}(0,1)}\in{\cal L}(U,X)$, this problem belongs to the class of the minimum energy problems studied in this paper. We know, from Proposition 4.3, that the value function is given by $V_{\infty}(y_{0})=\frac{1}{2}\|y_{0}\|^{2}_{L^{2}(0,1)}.$ We can now apply Theorem 4.7, obtaining that: * • the identity in $L^{2}$, $I_{L^{2}(0,1)}$, solves the ARE (66) where we replace $B$ and $B^{*}$ by $I_{H^{-1}(0,1)}$; * • the operator $Q_{\infty}^{-1}=2A$ solves the ARE (65) where we replace $B^{*}$ by $I_{H^{-1}(0,1)}$; * • $I_{L^{2}(0,1)}$ is the maximal solution of the ARE (66) among those in the class $\mathcal{Q}$ introduced in Definition 3.8. Moreover, here Hypothesis 5.1 holds; hence we can apply Theorem 5.5. Then, noting that $A_{0}$ is the Laplace operator with Dirichlet boundary conditions in the space $H=L^{2}(0,1)$, whose domain is $H^{2}(0,1)\cap H^{1}_{0}(0,1)$, we obtaing that: * • the identity in $L^{2}$, $I_{L^{2}(0,1)}$, is a solution of the two (equivalent) AREs (71) and (73); * • the set of all solutions of (71) and (73) consists of all orthogonal projections $P$ which commute with $A_{0}$, i.e. all projections whose image is generated by a subset of the eigenvectors of $A_{0}$; * • $I_{L^{2}(0,1)}$ is the maximal solution among all solutions of (71) and (73). ## Appendix ## Appendix A Minimum Energy with finite horizon This part of the Appendix is devoted to recall the formulation of the finite horizon minimum energy problem studied in [1] (briefly described at the beginning of Subsection 2.3) and to provide some related results which are useful in treating the infinite horizon problem (9)–(10). Throughout this section we will assume that Hypothesis 2.2 holds without repeating it. ### A.1 General formulation of the problem We take the Hilbert spaces $X$ (state space) and $U$ (control space), as well as the operators $A$ and $B$, as in Hypothesis 2.2. Given a time interval $[s,t]\subset\mathbb{R}$, an initial state $z\in X$ and a control $u\in L^{2}(s,t;U)$ we consider the state equation (1), which we rewrite here: $\left\\{\begin{array}[]{l}y^{\prime}(r)=Ay(r)+Bu(r),\quad r\in\,]s,t],\\\\[2.84526pt] y(s)=z.\end{array}\right.$ (92) Denote by $y(\cdot;s,z,u)$ the mild solution of (92) (see Proposition 2.3): $y(r;s,z,u):=e^{(r-s)A}z+\int_{s}^{r}e^{(r-\tau)A}Bu(\tau)\,\mathrm{d}\tau,\qquad r\in[s,t].$ (93) We define the class of controls $u(\cdot)$ bringing the state $y(\cdot)$ from a fixed $z\in X$ at time $s$ to a given target $x\in X$ at time $t$: ${\cal U}_{[s,t]}(z,x)\stackrel{{\scriptstyle\textrm{def}}}{{=}}\left\\{u\in L^{2}(s,t;U)\;:\;y(t;s,z,u)=x\right\\}.$ (94) Consider the quadratic functional (the energy) $J_{[s,t]}(u)=\frac{1}{2}\int_{s}^{t}\|u(r)\|_{U}^{2}\,\mathrm{d}r.$ (95) The minimum energy problem at $(s,t;z,x)$ is the problem of minimizing the functional $J_{[s,t]}(u)$ over all $u\in{\cal U}_{[s,t]}(z,x)$. The value function of this control problem (the minimum energy) is $V_{1}(s,t;z,x)\stackrel{{\scriptstyle\textrm{def}}}{{=}}\inf_{u\in{\cal U}_{[s,t]}(z,x)}J_{[s,t]}(u).$ (96) with the agreement that the infimum over the emptyset is $+\infty$. Similarly to what we did in Proposition 3.1, given any $z\in X$ we define the reachable set in the interval $[s,t]$, starting from $z$, as ${\mathbf{R}}_{[s,t]}^{z}:=\left\\{x\in X:\ {\cal U}_{[s,t]}(z,x)\neq\emptyset\right\\}.$ (97) Defining the operator ${\cal L}_{s,t}:L^{2}(s,t;U)\rightarrow X,\qquad{\cal L}_{s,t}u=\int_{s}^{t}e^{(t-\tau)A}Bu(\tau)\,\mathrm{d}\tau,$ (98) it is clear that ${\mathbf{R}}_{[s,t]}^{z}:=e^{(t-s)A}z+{\cal L}_{s,t}\left(L^{2}(s,t;U)\right).$ (99) The use of [1, Proposition 2.6] allows to reduce the number of variables from 4 to 2. In particular $V_{1}(s,t;z,x)=V_{1}(s-t,0;0,x-e^{(t-s)A}z)=V_{1}(0,t-s;0,x-e^{(t-s)A}z);$ (100) Hence from now on we set, for simplicity of notation, $V(t,x):=V_{1}(-t,0;0,x)=\inf_{u\in{\cal U}_{[-t,0]}(0,x)}J_{[-t,0]}(u)\qquad t\in\,]0,+\infty[,\quad x\in X.$ (101) ### A.2 The space $H$ and its properties In this subsection we provide some useful properties of the space $H$ introduced in Subsection 2.3 (see (22)-(23)). First recall that $H={\cal R}(Q_{\infty}^{1/2})\qquad\hbox{and}\qquad\langle x,y\rangle_{H}=\langle Q_{\infty}^{-1/2}x,Q_{\infty}^{-1/2}y\rangle_{X}\,,\quad x,y\in H,$ (102) and that (with, in general, proper inclusion) $H\subseteq\overline{{\cal R}(Q_{\infty}^{1/2})}=[\ker Q_{\infty}^{1/2}]^{\perp}=[\ker Q_{\infty}]^{\perp}.$ Next Lemmas A.1 and A.2 are exactly Lemmas 4.2 and 4.3 of [1]. ###### Lemma A.1. * * (i) The space $H$ is a Hilbert space continuously embedded into $X$. * (ii) The space ${\cal R}(Q_{\infty})$ is dense in $H$. * (iii) The operator $Q_{\infty}^{-1/2}$ is an isometric isomorphism from $H$ to $[\ker Q_{\infty}^{1/2}]^{\perp}$, and in particular $\|Q_{\infty}^{-1/2}x\|_{X}=\|x\|_{H}\qquad\forall x\in H.$ (103) * (iv) We have $Q_{\infty}^{1/2}\in{\cal L}(H)$ and $\|Q_{\infty}^{1/2}\|_{{\cal L}(X)}=\|Q_{\infty}^{1/2}\|_{{\cal L}(H)}.$ * (v) For every $F\in{\cal L}(X)$ such that ${\cal R}(F)\subseteq H$ we have $Q_{\infty}^{-1/2}F\in{\cal L}(X)$, so that $F\in{\cal L}(X,H)$. ###### Lemma A.2. For $0<t\leq+\infty$ let $Q_{t}$ be the operator defined by (19). Then, for every $t\in[T_{0},+\infty]$, the space $Q_{t}({\cal D}(A^{*}))$ is dense in $H$ and contained in ${\cal D}(A)$. In particular ${\cal D}(A)\cap H$ is dense in $H$. ###### Remark A.3. The above Lemma immediately implies that, for every $t\in[T_{0},+\infty]$, $Q_{t}({\cal D}(A^{*}))$ is dense in $[\ker Q_{\infty}]^{\perp}$ with the topology inherited by $X$, since the inclusion of $H$ into $[\ker Q_{\infty}]^{\perp}$ is continuous. Now we state and prove three very useful lemmas. ###### Lemma A.4. Assume Hypothesis 2.7. Then we have the following: * (i) For every $z\in H$ and $r\geq 0$ we have $e^{rA}z\in H$; moreover the semigroup $e^{tA}|_{H}$ is strongly continuous in $H$. In particular, for each $T>0$ there exists $c_{T}>0$ such that $\|e^{rA}z\|_{H}\leq c_{T}\|z\|_{H}\qquad\forall z\in H,\quad\forall r\in[0,T].$ * (ii) For every $\lambda\in\rho(A)$ we have $\lambda\in\rho(A_{0})$ and $R(\lambda,A_{0})=R(\lambda,A)|_{H}$. * (iii) The generator $A_{0}$ of the semigroup $e^{tA}|_{H}$ is given by $\left\\{\begin{array}[]{l}{\cal D}(A_{0})=\left\\{x\in{\cal D}(A)\cap H\,:\;Ax\in H\right\\}\\\\[5.69054pt] A_{0}x=Ax\quad\forall x\in{\cal D}(A_{0}).\end{array}\right.$ (104) We denote the semigroup $e^{tA}|_{H}$ by $e^{tA_{0}}$ (see Notation 3.7): thus $e^{tA_{0}}$ is a strongly continuous semigroup in $H$. ###### Proof. (i) Fix any $z\in H$ and $t>T_{0}$: then, by Hypothesis 2.7, $z\in{\cal R}(Q_{\infty}^{1/2})={\cal R}(Q_{t}^{1/2})={\cal R}({\cal L}_{-t,0})$ (see (32)), i.e. there exists $u\in L^{2}(0,r;U)$ such that $z={\cal L}_{-t,0}(u)=\int_{-t}^{0}e^{-\sigma A}\,Bu(\sigma)\,\mathrm{d}\sigma.$ Hence, for every $r>0$, $e^{rA}z=\int_{-t}^{0}e^{(r-\sigma)A}\,Bu(\sigma)\,d\sigma=\int_{-t}^{r}e^{(r-\sigma)A}\,B\overline{u}(\sigma)\,\mathrm{d}\sigma,$ where $\overline{u}(s)=\left\\{\begin{array}[]{ll}u(s)&\textrm{if }s\in[-t,0]\\\\[5.69054pt] 0&\textrm{if }s\in[0,r].\end{array}\right.$ Setting $r-\sigma=-s$ and $v(s)=\overline{u}(s+r)$, it follows that $e^{rA}z=\int_{-t-r}^{0}e^{-sA}\,B\overline{u}(r+s)\,ds={\cal L}_{-t-r,0}(v)\in{\cal R}({\cal L}_{-t-r,0})={\cal R}(Q_{t+r}^{1/2})={\cal R}(Q_{\infty}^{1/2})=H.$ Let us now prove that the restriction of $e^{rA}$ to $H$ has closed graph in $H$: if $z,\\{z_{n}\\}\subset H$ and $z_{n}\rightarrow z$ in $H$, $e^{rA}z_{n}\rightarrow w\in H$ in $H$, then, since $H$ is continuously embedded into $X$, $z_{n}\rightarrow z\textrm{ in }X,\qquad e^{rA}z_{n}\rightarrow w\textrm{ in }X;$ but $e^{rA}\in{\cal L}(X)$, so that $w=e^{rA}z$. Thus $e^{rA}z_{n}\rightarrow e^{rA}z$ in $H$, and it follows that $e^{rA}\in{\cal L}(H)$. Now fix $x\in H$ and consider for $t>0$ the quantity $e^{tA}x-x$. We have $\|e^{tA}x-x\|_{H}=\sup_{\|y\|_{H}=1}\langle e^{tA}x-x,y\rangle_{H}\,;$ thus, for every $\varepsilon\in\,]0,1[\,$ there exists $y_{\varepsilon}\in H$ with $\|y_{\varepsilon}\|_{H}=1$ such that $\|e^{tA}x-x\|_{H}<\varepsilon+\langle e^{tA}x-x,y_{\varepsilon}\rangle_{H};$ then, using Lemma A.2 and choosing $z_{\varepsilon}\in{\cal R}(Q_{\infty})$ such that $\|z_{\varepsilon}-y_{\varepsilon}\|_{H}<\varepsilon$, we obtain $\displaystyle\|e^{tA}x-x\|_{H}$ $\displaystyle<$ $\displaystyle\varepsilon+\langle e^{tA}x-x,y_{\varepsilon}-z_{\varepsilon}\rangle_{H}+\langle e^{tA}x-x,z_{\varepsilon}\rangle_{H}\leq$ $\displaystyle\leq$ $\displaystyle\varepsilon+\|e^{tA}x-x\|_{H}\,\|y_{\varepsilon}-z_{\varepsilon}\|_{H}+\langle e^{tA}x-x,Q_{\infty}^{-1}z_{\varepsilon}\rangle_{X}\leq$ $\displaystyle\leq$ $\displaystyle\varepsilon+\|e^{tA}x-x\|_{H}\,\varepsilon+\|e^{tA}x-x\|_{X}\,\|Q_{\infty}^{-1}z_{\varepsilon}\|_{X}\,.$ Hence $(1-\varepsilon)\|e^{tA}x-x\|_{H}<\varepsilon+\|e^{tA}x-x\|_{X}\,\|Q_{\infty}^{-1}z_{\varepsilon}\|_{X}\,,$ and letting $t\rightarrow 0^{+}$ we get $\limsup_{t\rightarrow 0^{+}}\|e^{tA}x-x\|_{H}\leq\frac{\varepsilon}{1-\varepsilon}+0.$ The arbitrariness of $\varepsilon$ leads to the conclusion. (ii) This is an immediate consequence of point (i) and of the well known resolvent formula $R(\lambda,A)=\int_{0}^{+\infty}e^{-\lambda t}e^{tA}\,\mathrm{d}t$. (iii) If $z\in{\cal D}(A_{0})$ then it must be, by definition, $z\in{\cal D}(A)$, $z\in H$, $Az\in H$; hence ${\cal D}(A_{0})\subseteq\left\\{x\in{\cal D}(A)\cap H\,:\;Ax\in H\right\\}$. To prove the converse we first observe that, using (ii), we get, for $n\in\mathbb{N}-\\{0\\}$ and $h\in H$, $nAR(n,A)x=nx-n^{2}R(n,A)x=nx-n^{2}R(n,A_{0})x=nA_{0}R(n,A_{0})x$ Now assume that $z\in{\cal D}(A)\cap H$ with $Az\in H$. To prove that $z\in{\cal D}(A_{0})$ it is enough to show that $nA_{0}R(n,A_{0})z$ converges to some element $y$ of $H$ when $n\rightarrow+\infty$. In this case such element is $A_{0}z$. To do this we observe that, by the above remarks for the resolvents and by the assumptions on $z$, we get $nA_{0}R(n,A_{0})z=nAR(n,A)z=nR(n,A)Az=nR(n,A_{0})Az$ The latter, by the properties of resolvents, converges in $H$ to $Az$ as $n\rightarrow+\infty$, since $Az\in H$. This shows that $z\in{\cal D}(A_{0})$ and $A_{0}z=Az$. ∎ ###### Lemma A.5. Assume Hypothesis 2.7. Then we have the following. * (i) $Q_{\infty}(H)$ is dense in $H$. * (ii) $Q_{\infty}({\cal D}(A_{0}^{*}))$ is dense in $H$. * (iii) Let $A$ be selfadjoint and commuting with $BB^{*}$. Then $Q_{\infty}({\cal D}(A_{0}^{*}))\subseteq{\cal D}(A_{0})$. Moreover $A_{0}^{*}$ is selfadjoint in $H$. ###### Proof. (i) Since $\ker Q_{\infty}^{1/2}=\ker Q_{\infty}$, we have $\overline{{\cal R}(Q_{\infty}^{1/2})}=\overline{{\cal R}(Q_{\infty})}$. Fix $x\in H$ and set $z:=Q_{\infty}^{-1/2}x\in\overline{{\cal R}(Q_{\infty}^{1/2})}$. Then there exists $\\{w_{n}\\}\subset X$ such that, defining $z_{n}:=Q_{\infty}w_{n}\in{\cal R}(Q_{\infty})$, we have $z_{n}\rightarrow z$ in $X$. Set $x_{n}:=Q_{\infty}^{1/2}z_{n}=Q_{\infty}^{1/2}Q_{\infty}w_{n}=Q_{\infty}Q_{\infty}^{1/2}w_{n}.$ Clearly $x_{n}\in Q_{\infty}(H)$. Moreover $\|x_{n}-x\|_{H}=\|Q_{\infty}^{1/2}z_{n}-x\|_{H}=\|z_{n}-z\|_{X}\rightarrow 0\quad\hbox{as $n\rightarrow+\infty$,}$ which proves the claim. [(ii) Fix $x\in H$. By part (i) there exists $\\{x_{n}\\}\subset Q_{\infty}(H)$ such that $x_{n}\rightarrow x$ in $H$. We must have $x_{n}:=Q_{\infty}z_{n}$, with $z_{n}\in H$. Since ${\cal D}(A^{*}_{0})$ is dense in $H$, then, for every $n\in\mathbb{N}_{+}$ there exists $w_{n}\in{\cal D}(A^{*}_{0})$ such that $\|z_{n}-w_{n}\|_{H}<1/n$. Consequently, setting $y_{n}:=Q_{\infty}w_{n}$, we have, using Lemma A.1-(iv), $\|y_{n}-x\|_{H}\leq\|Q_{\infty}(w_{n}-z_{n})\|_{H}+\|x_{n}-x\|_{H}\leq\|Q_{\infty}\|_{{\cal L}(H)}\frac{1}{n}+\|x_{n}-x\|_{H}\,.$ This proves the claim. (iii) Let $A$ be selfadjoint and commuting with $BB^{*}$. Observe first that ${\cal D}(A_{0}^{*})\subseteq{\cal D}(A^{*})={\cal D}(A)$. Indeed, when $x\in{\cal D}(A_{0}^{*})$, the linear map $y\rightarrow\langle x,A_{0}y\rangle_{H}$ is bounded in $H$. Using such boundedness and the fact that $A$ and $Q_{\infty}$ commute (see [1, Proposition C.1-(v)])), we get, for every $y\in{\cal D}(A)$, $\langle x,Ay\rangle_{X}=\langle x,Q_{\infty}Ay\rangle_{H}=\langle x,AQ_{\infty}y\rangle_{H}=\langle x,A_{0}Q_{\infty}y\rangle_{H}\leq C\|Q_{\infty}y\|_{H}\leq C^{\prime}\|y\|_{X}$ which implies $x\in{\cal D}(A^{*})={\cal D}(A)$. Now, let $x\in Q_{\infty}({\cal D}(A_{0}^{*}))$ (which is contained in ${\cal D}(A)$ by Lemma A.2, since ${\cal D}(A_{0}^{*})\subseteq{\cal D}(A^{*})$) and let $z\in{\cal D}(A_{0}^{*})$ be such that $x=Q_{\infty}z$. Using again the fact that $A$ and $Q_{\infty}$ commute, we get $Ax=AQ_{\infty}z=Q_{\infty}Az\in H$. Hence, by definition of $A_{0}$, we deduce that $x\in{\cal D}(A_{0})$ and $A_{0}x=Ax$. Now we prove that $A_{0}$ is selfadjoint in $H$. Let $x\in{\cal D}(A_{0})$ and $y\in Q_{\infty}({\cal D}(A_{0}^{*}))$. Then for some $z\in{\cal D}(A_{0}^{*})$ we have $y=Q_{\infty}z$ and $Q_{\infty}^{-1}y=z+z_{0}$, where $z_{0}\in\ker Q_{\infty}$. Hence it must be $\langle Ax,z_{0}\rangle_{X}=0$, since $Ax=A_{0}x\in H\subseteq[\ker Q_{\infty}]^{\perp}$. Using this fact, we get $\displaystyle\langle A_{0}x,y\rangle_{H}$ $\displaystyle=$ $\displaystyle\langle Ax,y\rangle_{H}=\langle Ax,Q_{\infty}^{-1}y\rangle_{X}=\langle Ax,z\rangle_{X}=\langle x,Az\rangle_{X}=$ $\displaystyle=$ $\displaystyle\langle x,Q_{\infty}Az\rangle_{H}=\langle x,AQ_{\infty}z\rangle_{H}=\langle x,Ay\rangle_{H}=\langle x,A_{0}y\rangle_{H}\,,$ where in the last step we used the inclusion ${\cal D}(A_{0}^{*})\subseteq{\cal D}(A)$, the fact that $Q_{\infty}$ and $A$ commute, and the inclusion $Q_{\infty}({\cal D}(A_{0}^{*}))\subseteq{\cal D}(A_{0})$. This implies that, for every $x\in{\cal D}(A_{0})$, the linear map $y\rightarrow\langle x,A_{0}y\rangle_{H}$ is defined on $Q_{\infty}({\cal D}(A_{0}^{*}))$ (which is dense in $H$) and is bounded in $H$. This implies that $x\in{\cal D}(A_{0}^{*})$ and $A_{0}^{*}x=A_{0}x$. Hence $A_{0}^{*}$ extends $A_{0}$. Since both $A_{0}$ and $A_{0}^{*}$ generate a semigroup, we can choose $\lambda>0$ such that $\lambda\in\rho(A_{0})\cap\rho(A_{0}^{*})$. For such $\lambda$ we now prove that $R(\lambda,A_{0}^{*})=R(\lambda,A_{0})$, which immediately implies that ${\cal D}(A_{0})={\cal D}(A_{0}^{*})$. Indeed for $z\in H$ we have $z=(\lambda-A_{0})R(\lambda,A_{0})z=(\lambda-A_{0}^{*})R(\lambda,A_{0})z,$ where in the last equality we used that ${\cal D}(A_{0})\subseteq{\cal D}(A_{0}^{*})$ and that $A_{0}^{*}x=A_{0}x$ for all $x\in{\cal D}(A_{0})$. Applying $R(\lambda,A_{0}^{*})$ to both sides we get the claim. ∎ ###### Lemma A.6. Assume Hypothesis 2.7. (i) For $0\leq s\leq T_{0}$ we have $Q_{\infty}^{-1/2}e^{sA}\in{\cal L}(H,X)$, and there exists $C_{1}(T_{0})>0$ such that $\|Q_{\infty}^{-1/2}e^{sA}x\|_{X}\leq C_{1}({T_{0}})\|x\|_{H}\qquad\forall x\in H,\quad\forall s\in[0,T_{0}],$ and $(Q_{\infty}^{-1/2}e^{sA})^{*}=e^{sA_{0}^{*}}Q_{\infty}^{-1/2}\in{\cal L}(X,H)$, with $\|(Q_{\infty}^{-1/2}e^{sA})^{*}\|_{{\cal L}(X,H)}\leq C_{1}({T_{0}})\qquad\forall s\in[0,T_{0}].$ (ii) For $s>T_{0}$ we have $Q_{\infty}^{-1/2}e^{sA}\in{\cal L}(X)$, with $\|Q_{\infty}^{-1/2}e^{sA}x\|_{X}\leq C_{1}({T_{0}})Me^{-\omega(s-T_{0})}\|x\|_{X}\qquad\forall x\in X,\quad\forall s>T_{0},$ and $\|(Q_{\infty}^{-1/2}e^{sA})^{*}\|_{{\cal L}(X)}\leq C_{1}({T_{0}})Me^{-\omega(s-T_{0})}\qquad\forall s>T_{0}.$ (iii) For $s\geq 0$, $x\in X$ we have $e^{sA_{0}^{*}}Q_{\infty}x=Q_{\infty}e^{sA^{*}}x.$ (iv) For $x\in{\cal D}(A^{*})$ we have $Q_{\infty}x\in{\cal D}(A_{0}^{*})$. Moreover, for every $s\geq 0$ we have $A_{0}^{*}e^{sA_{0}^{*}}Q_{\infty}x=Q_{\infty}A^{*}e^{tA^{*}}x.$ ###### Proof. (i) We have by Lemma A.4-(i) $\|Q_{\infty}^{-1/2}e^{sA}x\|_{X}=\|e^{sA}x\|_{H}\leq c_{T_{0}}\|x\|_{H}\qquad\forall x\in H,\quad\forall s\in[0,T_{0}];$ moreover (identifying here $X$ and $H$ with their duals), $(Q_{\infty}^{-1/2}e^{sA})^{*}\in{\cal L}(X,H)$ and, for all $z\in H$ and $x\in X$, we have $\langle(Q_{\infty}^{-1/2}e^{sA})^{*}x,z\rangle_{H}=\langle x,Q_{\infty}^{-1/2}e^{sA}z\rangle_{X}=\langle Q_{\infty}^{1/2}x,e^{sA_{0}}z\rangle_{H}=\langle e^{sA_{0}^{*}}Q_{\infty}^{1/2}x,z\rangle_{H}\,.$ This shows that $(Q_{\infty}^{-1/2}e^{sA})^{*}=e^{sA_{0}^{*}}Q_{\infty}^{1/2}\in{\cal L}(X,H),$ with $\|(Q_{\infty}^{-1/2}e^{sA})^{*}\|_{{\cal L}(X,H)}=\|e^{sA_{0}}Q_{\infty}^{1/2}\|_{{\cal L}(X,H)}=\|Q_{\infty}^{-1/2}e^{sA}\|_{{\cal L}(H,X)}\leq c_{T_{0}}\quad\forall s\in[0,T_{0}].$ (ii) By Hypothesis 2.7 we have $Q_{\infty}^{-1/2}e^{sA}\in{\cal L}(X)$, and by (i) we get $\displaystyle\|Q_{\infty}^{-1/2}e^{sA}\|_{{\cal L}(X)}$ $\displaystyle=$ $\displaystyle\|Q_{\infty}^{-1/2}e^{T_{0}A}e^{(s-T_{0})A}\|_{{\cal L}(X)}$ $\displaystyle\leq$ $\displaystyle\|Q_{\infty}^{-1/2}e^{T_{0}A}\|_{{\cal L}(X)}Me^{-\omega(s-T_{0})}\leq C_{1}({T_{0}})Me^{-\omega(s-T_{0})}\quad\forall s>T_{0}.$ The claim easily follows. (iii) For $s\geq 0$, $x\in X$, $z\in H$, we have $\langle e^{sA_{0}^{*}}Q_{\infty}x,z\rangle_{H}=\langle Q_{\infty}x,e^{sA_{0}}z\rangle_{H}=\langle x,e^{sA}z\rangle_{X}=\langle e^{sA^{*}}x,z\rangle_{X}=\langle Q_{\infty}e^{sA^{*}}x,z\rangle_{H}\,,$ which proves the claim. Let $x\in{\cal D}(A^{*})$ and $s\geq 0$. For $h>0$ we have, using the point (iii) above, $\frac{e^{(s+h)A_{0}^{*}}-e^{sA_{0}^{*}}}{h}Q_{\infty}x=Q_{\infty}\frac{e^{(s+h)A^{*}}-e^{sA^{*}}}{h}x$ Letting $h\rightarrow 0$, the claim follows. ∎ ## References * [1] P. Acquistapace, and F. Gozzi. Minimum energy for linear systems with finite horizon: a non-standard Riccati equation. Math. Control Signals Syst. 29, 19 (2017). * [2] E. Barucci, and F. Gozzi. Technology Adoption and Accumulation in a Vintage Capital Model. J. of Economics, 74, N.1, 1–38 (2001). * [3] A. Bensoussan, G. Da Prato, M.C. Delfour, and S.K. Mitter. Representation and control of Infinite dimensional system. Second edition, Birkhäuser, Boston 2007. * [4] L. Bertini, A. De Sole, D. Gabrielli, G. Jona–Lasinio, and C. Landim. Fluctuations in stationary nonequilibrium states of irreversible processes. Phys. Rev. Lett. 87, 040601 (2001). * [5] L. Bertini, A. De Sole, D. Gabrielli, G. Jona–Lasinio, and C. Landim. Macroscopic fluctuation theory for stationary non equilibrium state. J. Statist. Phys. 107, 635–675 (2002). * [6] L. Bertini, A. De Sole, D. Gabrielli, G. Jona–Lasinio, and C. Landim. Large deviations for the boundary driven simple exclusion process. Math. Phys. Anal. Geom. 6, 231–267 (2003). * [7] L. Bertini, A. De Sole, D. Gabrielli, G. Jona–Lasinio, and C. Landim. Minimum Dissipation Principle in Stationary Non-Equilibrium States. J. Statist. Phys. 116, 831–841 (2004). * [8] L. Bertini, A. De Sole, D. Gabrielli, G. Jona–Lasinio, and C. Landim. Action Functional and Quasi-Potential for the Burgers Equation in a Bounded Interval. Comm. Pure. Appl. Math. 64, 649–696 (2011). * [9] L. Bertini, D. Gabrielli, and J.L. Lebowitz. Large deviations for a stochastic model of heat flow. J. Statist. Phys. 121, 843–885 (2005). * [10] 0\. Carja. The minimal time function in infinite dimensions. SIAM J. Contr. Optimiz. 31 (5), 1103-1114 (1993). * [11] R. Curtain, and A. J. Pritchard. Infinite Dimensional Linear Systems Theory. Springer 1978. * [12] G. Da Prato, and J. Zabczyk. Stochastic Equations in Infinite Dimension. Springer 1992. * [13] G. Da Prato, A. J. Pritchard, and J. Zabczyk. Null controllability with vanishing energy. SIAM J. Control Optim. 29 (1), 209–221 (1991). * [14] Z. Emirsajlow. A feedback for an infinite-dimensional linear-quadratic control problem with a fixed terminal state. IMA J. Math. Control Inf. 6 (1), 97-117 (1989). * [15] Z. Emirsajlow, and SD. Townley. Uncertain systems and minimum energy control. J. Appl. Math. Comput. Sci. 5 (3), 533–545 (1995). * [16] K-J. Engel, and R. Nagel. One Parameter Semigroups for Linear Evolution Equations. Springer 1999. * [17] G. Fabbri, F. Gozzi, and A. Swiech. Stochastic optimal control in infinite dimension. Springer 2017. * [18] J. Feng, and T. Kurtz. Large Deviations for Stochastic Processes. Mathematical Surveys and Monographs, AMS 2006. * [19] J.A. Goldstein. Semigroups of linear operators and applications. Oxford Mathematical Monographs, 1985. * [20] F. Gozzi, and P. Loreti. Regularity of the minimum time function and minimum energy problems: The linear case. SIAM J. Control Optim. 37 (4), 1195–1221 (1999). * [21] P. Lancaster, and L. Rodman. Algebraic Riccati Equations. Oxford Science Publications, Clarendon Press, Oxford, 1995. * [22] I. Lasiecka, and R. Triggiani. Control Theory for Partial Differential Equations: Continuous and Approximation Theories. Part 1. Cambridge University Press 2000. * [23] I. Lasiecka, and R. Triggiani. Control Theory for Partial Differential Equations: Continuous and Approximation Theories. Part 2. Cambridge University Press 2000. * [24] A. Lunardi. Analytic semigroups and optimal regularity in parabolic problems. Birkhäuser Verlag, Basel, 1995. * [25] B. C. Moore. Principal component analysis in linear systems: controllability, observability, and model reduction. IEEE Trans. Automat. Control 26 (1), 17–32 (1981). * [26] A. W. Olbrot, and L. Pandolfi. Null controllability of a class of functional-differential systems. Internat. J. Control 47 (1), 193–208 (1988). * [27] A. Pazy. Semigroups of linear operators and applications to partial differential equations. Springer Verlag, New-York 1983. * [28] E. Priola, and J. Zabczyk. Null controllability with vanishing energy. SIAM J. Control and Optim. 42 (6), 1013-1032, 2003. * [29] M. Reed, B. Simon. Functional Analysis. Academic Press, London 1980. * [30] J.M.A. Scherpen. Balancing for Nonlinear Systems. Syst. Contr. Lett. 21 (2), 143-153 (1993). * [31] H. Triebel. Interpolation theory, function spaces, differential operators. North-Holland Publishing Co., Amsterdam 1978. * [32] J. Zabczyk. Mathematical control theory: an introduction. Birkhäuser Verlag, Boston 1995.
# Echo State Network for two-dimensional turbulent moist Rayleigh-Bénard convection Florian Heyder Institut für Thermo- und Fluiddynamik, Technische Universität Ilmenau, Postfach 100565, D-98684 Ilmenau, Germany Jörg Schumacher Institut für Thermo- und Fluiddynamik, Technische Universität Ilmenau, Postfach 100565, D-98684 Ilmenau, Germany Tandon School of Engineering, New York University, New York City, NY 11201, USA ###### Abstract Recurrent neural networks are machine learning algorithms which are suited well to predict time series. Echo state networks are one specific implementation of such neural networks that can describe the evolution of dynamical systems by supervised machine learning without solving the underlying nonlinear mathematical equations. In this work, we apply an echo state network to approximate the evolution of two-dimensional moist Rayleigh- Bénard convection and the resulting low-order turbulence statistics. We conduct long-term direct numerical simulations in order to obtain training and test data for the algorithm. Both sets are pre-processed by a Proper Orthogonal Decomposition (POD) using the snapshot method to reduce the amount of data. Training data comprise long time series of the first 150 most energetic POD coefficients. The reservoir is subsequently fed by these data and predicts of future flow states. The predictions are thoroughly validated by original simulations. Our results show good agreement of the low-order statistics. This incorporates also derived statistical moments such as the cloud cover close to the top of the convection layer and the flux of liquid water across the domain. We conclude that our model is capable of learning complex dynamics which is introduced here by the tight interaction of turbulence with the nonlinear thermodynamics of phase changes between vapor and liquid water. Our work opens new ways for the dynamic parametrization of subgrid-scale transport in larger-scale circulation models. ## I Introduction Machine learning (ML) algorithms have changed our way to model and analyse turbulent flows including thermally driven convection flows Brenner et al. (2019); Brunton et al. (2020); Pandey et al. (2020). The applications reach from subgrid-scale stress models Ling et al. (2016); Duraisamy et al. (2019), via the detection of large-scale patterns in mesoscale convection Fonda et al. (2019) to ML-based parametrizations of turbulent transport and clouds in climate and global circulation models Brenowitz and Bretherton (2018); O’Gorman and Dwyer (2018); Gentine et al. (2018). The implementations of such algorithms help to process and reduce increasing amounts of data coming from numerical simulations and laboratory experiments Moller et al. (2020) by detecting patterns and statistical correlations Goodfellow et al. (2016). Moreover, the computational cost that is in line with a direct numerical simulation (DNS) of the underlying highly nonlinear Navier-Stokes equations can often be reduced significantly by running a neural network instead that generates synthetic fields with the same low-order statistics as in a full simulation Schneider et al. (2017); Mohan et al. (2020). Turbulent convection, as all other turbulent flows, is inherently highly chaotic so that specific flow states in the future are hard to predict after exceeding a short horizon. The additional possibility of the fluid to change its thermodynamic phase, as it is the case in moist turbulent convection Stevens (2005); Mellado (2017), adds further nonlinearities and feedbacks to the turbulence dynamics. Learning low-order statistics of such a turbulent flow provides a challenge to an ML algorithm. For such a task, an internal memory is required since statistical correlations decay in a finite time. This necessitates the implementation of a short-term memory or internal cyclic feedbacks in the network architecture. That is why a particular class of supervised machine learning algorithms – recurrent neural networks (RNNs) – will be in the focus of the present work Hochreiter and Schmidhuber (1997); Lukoševičius and Jaeger (2009); Tanaka et al. (2019). In this paper we apply a specific implementation of an RNN, the echo state network (ESN) Jaeger and Haas (2004); Yildiz et al. (2012), to two-dimensional (2d) turbulent moist Rayleigh-Bénard convection (RBC) flow. We use this RNN to predict the low-order statistics, such as the buoyancy and liquid water fluxes or fluctuation profiles of these quantities across the layer. Our present work extends a recent application of ESNs to two-dimensional dry Rayleigh-Bénard convection Pandey and Schumacher (2020) in several points. (1) The physical complexity of turbulent convection is enhanced in the present moist case. This is due to the total water content, an additional active scalar field which comprises of vapor and liquid water contents. The total water content couples as an additional field to the original dry Boussinesq model for temperature and velocity. (2) The size of the convection domain is increased by a factor of 4 such that the necessary degree of data reduction is significantly higher. (3) Moist convection requires also the satisfying reproduction of new physical quantities which are derived from different turbulence fields, e.g. the cloud cover in or the liquid water flux across the layer. This can be considered as a firmer test of the capabilities of the ESN to model complex dynamics. Such quantities are of particular interest for larger-scale models of atmospheric circulation that include mesoscale convection processes in form of parameters or minimal models Grabowski and Smolarkiewicz (1999); Grabowski (2001). (4) Finally, the hyperparameters of the ESN, in particular the spectral radius $\rho(W^{r})$ of the reservoir matrix $W^{r}$, has been tested in more detail (exact definitions follow). The grid search in our work thus sums up to a total of more than 1800 different hyperparameter sets. Echo state networks have recently received renewed attention for their capability of equation-free modeling of several chaotic systems, such as of the Lorenz96 model Vlachas et al. (2020) or the one-dimensional partial differential Kuramoto-Sivashinsky equation Lu et al. (2017); Pathak et al. (2018). Furthermore, low-order flow statistics in 2d dry Rayleigh-Bénard convection with ${\rm Ra}=10^{7}$ have been successfully reproduced using an ESN that trains the most energetic time coefficients of a Karhunen-Loéve expansion (or proper orthogonal decomposition) of the convection fields Holmes et al. (2012). This latter step can be considered as an autoencoder that provides training and test data for the ESN which cannot take the data directly, even for the present 2d case. A second popular implementation of RNNs, which we want to mention here for completeness, are long short-term memory networks (LSTM) Hochreiter and Schmidhuber (1997) which have been also applied to fluid mechanical problems, such as the dynamics of Galerkin models of shear flows in Srinivasan et al. (2019). These models also demonstrated to capture the longer-term time dependency and low-order statistics of important turbulent transport properties well (see also Pandey et al. (2020) for a direct comparison). An acceleration of the training, which requires the backpropagation of the errors through the whole network in contrast to ESNs, were obtained recently with Momentum LSTMs that apply the momentum extension of gradient descent search of the cost function minimum to determine the weights at the network nodes Nguyen et al. (2020). In this work, a moist Rayleigh-Bénard convection model with moist Rayleigh number ${\rm Ra}_{\rm M}\simeq 10^{8}$ and Prandtl number $\text{Pr}=0.7$ is considered as an example case. We choose a 2d domain $\Omega=L\times H$ with aspect ratio $A=L/H=24$. Here $L$ is the horizontal length and $H$ the height of the simulation domain. The number of grid points was chosen as $N_{x}\times N_{y}=7200\times 384$. The data are obtained by direct numerical simulations (DNS) which apply a spectral element method Fischer (1997); Scheel et al. (2013); nek (2017). Comprehensive studies of further data sets with different parameters, such as Rayleigh numbers, to study the generalization properties will be presented elsewhere. Our intention is to demonstrate the capability of the ESN to deliver reliable low-order statistics for a turbulent convection flow with phase changes. The manuscript is structured as follows. The next section introduces the moist RBC model and the central ideas of ESNs. Then the generation and analysis of the DNS data, including a brief description of the numerical scheme and the proper orthogonal decomposition (POD), is specified. Finally the results of our machine learning approach to moist turbulent convection will be discussed in detail. In section V we summarize our results and provide a conclusion and an outlook. ## II Methods ### II.1 Moist Rayleigh-Bénard Convection Model We now briefly review our model for moist Rayleigh-Bénard convection in two spatial dimensions. A detailed derivation can be found in Pauluis and Schumacher (2010); Weidauer et al. (2010); Schumacher and Pauluis (2010). The framework is based on the mathematically equivalent formulation by Bretherton Bretherton (1987, 1988). Similar simplified models of moist convection with precipitation have been developed by Smith and Stechmann Smith and Stechmann (2017) and Vallis et al. Vallis et al. (2019). Evaporative cooling and buoyancy reversal effects were discussed for example by Abma et al. Abma et al. (2013). The buoyancy $B$ in atmospheric convection is given by Emmanuel (1994) $\displaystyle B=-g\frac{\rho(S,q_{v},q_{l},q_{i},p)-\overline{\rho}}{\overline{\rho}}$ (1) with the gravity acceleration $g$, a mean density $\overline{\rho}$, the pressure $p$, the entropy $S$ and the contents of water vapor $q_{v}$, liquid water $q_{l}$ and ice $q_{i}$. We consider warm clouds only, i.e. $q_{i}=0$ and assume local thermodynamic equilibrium. From the latter assumption, it follows that no precipitation is possible and the number of independent variables in eq. (1) reduces to three. By introducing the total water content $q_{T}=q_{v}+q_{l}$ the buoyancy can be expressed as $B(S,q_{T},p)$. In the Boussinesq approximation, pressure variations about a hydrostatic equilibrium profile are small, such that the buoyancy becomes $B(S,q_{T},y)$ with the vertical spatial coordinate $y$ for the present 2d case. Furthermore, the convection layer is near the vapor-liquid phase boundary. The buoyancy can then be expressed as a piecewise linear function of $S$ and $q_{T}$ on both sides of the saturation line. This step preserves the discontinuity of the first partial derivatives of $B$ and therefore the physics of a first-order phase transition. The advantage of this formulation is that locally the saturation state of the air can be determined. In the last step, we substitute the linear combinations of $S$ and $q_{T}$ on both sides of the phase boundary by a dry buoyancy $D$ and a moist buoyancy $M$. Consequently the buoyancy field $B$ is given by Pauluis and Schumacher (2010) $\displaystyle B(x,y,t)=\max\left(M(x,y,t),D(x,y,t)-N_{s}^{2}y\right)$ (2) where the fixed Brunt-Väisälä frequency $N_{s}=\sqrt{g(\Gamma_{u}-\Gamma_{s})/T_{\rm ref}}$ is determined by the lapse rates of the saturated and unsaturated moist air, $\Gamma_{s}$ and $\Gamma_{u}$, and a reference temperature $T_{\rm ref}$. An air parcel at height $y$ and time $t$ is unsaturated if $M(x,y,t)<D(x,y,t)-N_{s}^{2}y$ and saturated if $M(x,y,t)>D(x,y,t)-N_{s}^{2}y$. Note that the newly introduced dry buoyancy field $D$ is proportional to the liquid water static energy and the moist buoyancy field $M$ to the moist static energy. As in dry Rayleigh- Bénard convection with Dirichlet boundary conditions for the temperature, the static diffusive profiles $\overline{D}(y),\overline{M}(y)$ are vertically linear $\displaystyle\overline{D}(y)$ $\displaystyle=D_{0}+\frac{D_{H}-D_{0}}{H}y$ (3) $\displaystyle\overline{M}(y)$ $\displaystyle=M_{0}+\frac{M_{H}-M_{0}}{H}y$ (4) where $D_{0}$, $M_{0}$ and $D_{H}$, $M_{H}$ are the imposed values of $D$, $M$ at the bottom ($y=0$) and top ($y=H$) of the computational domain. Here, $D_{0}=M_{0}$. The governing equations of the moist Boussinesq system are given by $\displaystyle\frac{d\mathbf{v}}{dt}$ $\displaystyle=-\nabla\tilde{p}+\nu\nabla^{2}\mathbf{v}+B\left(D,M,y\right)\hat{\mathbf{e}}_{y}$ (5) $\displaystyle\nabla\cdot\mathbf{v}$ $\displaystyle=0$ (6) $\displaystyle\frac{dD}{dt}$ $\displaystyle=\kappa\nabla^{2}D$ (7) $\displaystyle\frac{dM}{dt}$ $\displaystyle=\kappa\nabla^{2}M.$ (8) Here $\mathbf{v}=(v_{x}(x,y,t),v_{y}(x,y,t))^{T}$ is the two-dimensional velocity field, $\tilde{p}=p/\rho_{\rm ref}$ the kinematic pressure, $\nu$ the kinematic viscosity and $\kappa$ the scalar diffusivity. The term $d/dt=\partial/\partial t+(\mathbf{v}\cdot\nabla)$ is the material derivative. This idealized model describes the formation of warm, non-precipitating low clouds in a shallow layer up to a depth of $\sim 1$km. The assumptions, which are made here, hold for example to a good approximation over the subtropical oceans. The problem is made dimensionless using length scale $[x,y]=H$, buoyancy scale $[B]=M_{0}-M_{H}$, and (free-fall) time scale $[t]=\sqrt{H/(M_{0}-M_{H})}$. The characteristic velocity scale follows by $[v_{x},v_{y}]=\sqrt{(M_{0}-M_{H})H}$. Four dimensionless numbers can be identified: the Prandtl number, dry Rayleigh number and moist Rayleigh number are given by Pr $\displaystyle=\frac{\nu}{\kappa}$ (9) $\displaystyle{\rm Ra}_{\rm D}$ $\displaystyle=\frac{\left(D_{0}-D_{H}\right)H^{3}}{\nu\kappa}$ (10) $\displaystyle{\rm Ra}_{\rm M}$ $\displaystyle=\frac{\left(M_{0}-M_{H}\right)H^{3}}{\nu\kappa}\,.$ (11) An additional parameter arises from the additional phase changes, the dimensionless form of (2) ${\rm CSA}=\frac{N_{s}^{2}H}{M_{0}-M_{H}}\,.$ (12) The condensation in saturated ascent (CSA) controls the amount of latent heat an ascending saturated parcel can release on its way to the top. The saturation condition (2) implies that liquid water is immediately formed at a point in space and time when $M>D-N_{s}^{2}y$. There is no supersaturation considered in this model and the liquid water content field $q_{l}$ and thus the clouds are given by $q_{l}(x,y,t)=M(x,y,t)-(D(x,y,t)-N_{s}^{2}y)\,.$ (13) Note that in this formulation, $q_{l}$ can become negative, as it is a measure of the degree of saturation. In a nutshell, $q_{l}<0$ stand thus for a liquid water deficit. When the atmosphere is saturated, $q_{l}\geq 0$ and the conventional liquid water content is retained. Here we study the case of $D_{0}>D_{H}$ and $M_{0}>M_{H}$. Both fields are linearly unstable. For the case of a conditionally unstable moist layer with $D_{0}\leq D_{H}$ we refer to refs. Bretherton (1987, 1988) or Pauluis and Schumacher (2011). ### II.2 Reservoir Computing Reservoir computing (RC) is a special type of RNN implementation. Contrary to standard feed forward networks, neurons in the hidden layers of RNN are recurrently connected to each other. In this way, RNNs have a similar architecture to biological brains and are said to posses an internal memory as cyclic connections allow information to stay inside the network for a certain amount of time before fading out, known as the echo state property. Yet the training of such RNNs is exceedingly difficult, as common training schemes like the back propagation through time struggle with fading error gradients, slow convergence Jaeger (2002) and bifurcations Doya (1992). An alternative training method was proposed by Jaeger Jaeger (2001) and in an independent work by Maass Maass et al. (2002). Their idea, which is now known as reservoir computing Lukoševičius and Jaeger (2009), was to train the weights of the output layer only, which connect the RNN, denoted to as the reservoir, to the output neurons. The weights of the input layer as well as the internal reservoir weights should be initialized at random and then kept constant. This training procedure reduces the computational costs for training significantly and shifts the focus to an adequate initialization of the input and reservoir weight matrix $W^{r}$ (which is an adjacency matrix in network theory). While Jaeger’s approach is known as ESN, Maass’ framework is called liquid state machine. They differ in their field of origin, as the ESN stems from the field of machine learning and the liquid state machine from computational neuroscience. We will stick to the ESN formulation, but note that RC refers to the concept mentioned above and summarizes both models. Despite their simple training scheme, ESNs have been said to tackle many tasks, from forecasting closing prices of stock markets Lin et al. (2008) to estimating the life span of a fuel cell Morando et al. (2013). Especially its application to dynamical systems shows great promise. For instance it has been demonstrated that the dynamics of two of the three degrees of freedom of the Rössler system can be inferred from the evolution of the third one Lu et al. (2017). Further, the Lyapunov exponents of the dynamical system that a trained ESN represents have been shown to match the exponents of the data generating system Pathak et al. (2017). Figure 1: Sketch of the echo state network in the training phase (a) for time steps $n<0$ and the prediction phase (b) for time steps $n\geq 0$. Figure 1 shows the concept and components of the ESN, for the training phase in panel (a) and for the prediction phase in panel (b). The input $\mathbf{u}(n)\in\mathbb{R}^{\rm N_{\rm in}}$ at a time instance $n$ as well as a constant scalar bias $b=1$ are passed to the reservoir via the input weight matrix $W^{\rm in}\in\mathbb{R}^{\rm N_{\rm r}\times(1+\rm N_{\rm in})}$. The weighted input contributes to the dynamics of the reservoir state $\mathbf{s}\in\left[-1,1\right]^{\rm N_{\rm r}}$ at time $n$ which is given by $\displaystyle\mathbf{s}(n)$ $\displaystyle=(1-\gamma)\mathbf{s}(n-1)$ $\displaystyle+\gamma\tanh\left[W^{\rm in}\left[b;\mathbf{u}(n)\right]+W^{\rm r}\mathbf{s}(n-1)\right].$ (14) Here $\left[b;\mathbf{u}(n)\right]$ stands for the vertical concatenation of the scalar bias and the input 111In some cases the bias $b=0$ and $\mathbf{u}(n-1)$ are used in eq. (II.2).. This update rule comprises external forcing by the inputs $\mathbf{u}(n)$ as well as a self-interaction with the past instance $\mathbf{s}(n-1)$. The reservoir weight matrix $W^{\rm r}\in\mathbb{R}^{\rm N_{\rm r}\times\rm N_{\rm r}}$ blends the state dimensions, while the nonlinear hyperbolic tangent $\tanh\left(\cdot\right)$, applied to each component of its argument vector, is the nonlinear activation function of the neurons in this model. The leaking rate $\gamma\in\left(0,1\right]$ moderates the linear and nonlinear contributions and assures that the state is confined to $\left[-1,1\right]^{N_{\rm r}}$. As mentioned above, the existence of echo states, i.e., states that are purely defined by the input history, is crucial. An ESN is said to include such echo states when two different states $\mathbf{s}(n-1)$, $\mathbf{s}^{\prime}(n-1)$ converge to the same state $\mathbf{s}(n)$, provided the same input $\mathbf{u}(n)$ is given and the system has been running for many iterations $n$ Jaeger (2001). Therefore, the first few state iterations are considered as a reservoir washout, even for a reservoir with echo state property. After this transient phase, the updated state is concatenated with the bias and the current input to form the extended reservoir state $\tilde{\mathbf{s}}(n)=\left[b;\mathbf{u}(n);\mathbf{s}(n)\right]$. Finally, $\tilde{\mathbf{s}}$ is mapped via the output matrix $W^{\rm out}\in\mathbb{R}^{\rm N_{\rm in}\times\left(1+\rm N_{\rm in}+\rm N_{\rm r}\right)}$ to form the reservoir output $\mathbf{y}(n)\in\mathbb{R}^{\rm N_{\rm in}}$ $\mathbf{y}(n)=\mathrm{W}^{\rm out}\tilde{\mathbf{s}}(n).$ (15) For our application the output dimension matches the input dimension $\rm N_{\rm in}$, which generally does not need to be the case. Before the ESN can be used in the prediction phase, as sketched in Fig. 1(b), the elements of $W^{\rm out}$ have to be computed first. This process is known as training phase of this supervised machine learning algorithm. Only when the reservoir is properly trained it will produce reasonable output. A set of $n_{\rm train}$ training data instances $\\{\mathbf{u}(n),\mathbf{y}^{\rm target}(n)\\}$ (where $n=-n_{\rm train},-(n_{\rm train}-1),...,-1$) needs to be prepared. The target output $\mathbf{y}^{\rm target}(n)$ represents the desired output the ESN should produce for the given input $\mathbf{u}(n)$. The reservoir state $\mathbf{s}$ is computed for all inputs in the training data set and assembled into a mean square cost function with an additional $L_{2}$ regularization $C\left(W^{\rm out}\right)$ which is given by $\displaystyle C\left(W^{\rm out}\right)$ $\displaystyle=\frac{1}{n_{\rm train}}\sum_{n=-n_{\rm train}}^{-1}\left(W^{\rm out}\tilde{\mathbf{s}}(n)-\mathbf{y}^{\rm target}(n)\right)^{2}$ $\displaystyle+\beta\sum\limits_{i=1}^{N_{\rm in}}\|w_{i}^{\rm out}\|_{2}^{2},$ (16) and has to be minimized corresponding to $\displaystyle W_{\ast}^{\rm out}$ $\displaystyle=\arg\min C\left(W^{\rm out}\right)\,.$ (17) Here $w_{i}^{\rm out}$ denotes the $i$-th row of $W^{\rm out}$ and $\|\cdot\|_{2}$ the $L_{2}$ norm. Equations (16) and (17) are known as ridge regression with the ridge regression parameter $\beta$. The last term in (16) suppresses large values of the rows of $W^{\rm out}$, which could inadvertently amplify small differences of the state dimensions in (15). This well known regression problem is solved by the fitted output matrix $\displaystyle W_{\ast}^{\rm out}=Y^{\rm target}S^{\rm T}\left(SS^{\rm T}+\beta{\rm Id}\right)^{-1}$ (18) where $\left(\cdot\right)^{\rm T}$ denotes the transposed and ${\rm Id}$ the identity matrix. $Y^{\rm target}$ and $S$ are matrices where the $n$-th column is the target output $\mathbf{y}^{\rm target}(n)$ and the extended reservoir state $\tilde{\mathbf{s}}(n)$, respectively. As the output weights are the only parameters that are trained, RC is computationally inexpensive. However, the algebraic properties of the initially randomly generated matrices $W^{\rm in}$ and $W^{\rm r}$ are hyperparameters which have to be tuned beforehand. In our approach we draw the elements of $W^{\rm in}$ from a uniform distribution in $\left[-0.5,0.5\right]$ and impose no further restrictions on this matrix. For the generation of the reservoir weights in $W^{\rm r}$ it has been reported that the proportion of non-zero elements, i.e., the reservoir density D and the spectral radius $\varrho\left(W^{\rm r}\right)$, i.e., the largest absolute eigenvalue of $W^{\rm r}$ are crucial parameters that determine whether the desired echo state property holds Lukoševičius (2012). We choose a sparse reservoir ($\text{D}<1$) with few internal node connections and draw the elements from a uniform distribution in $\left[-1,1\right]$. We then normalize $W^{\rm r}$ by its largest absolute eigenvalue and multiply it with the desired spectral radius $\varrho\left(W^{\rm r}\right)$. This scaling approach, initially proposed by Jaeger Jaeger (2002), has established itself as one of the standard ways Morando et al. (2013); Lu et al. (2017); Pathak et al. (2018); Pandey and Schumacher (2020) to control the spectral radius. Nevertheless, other procedures have been proposed Yildiz et al. (2012); Strauss et al. (2012). In addition, the size of the reservoir $\rm N_{\rm r}$ is a hyperparameter. Usually $\mathbf{s}$ should be a high-dimensional extension of the inputs $\mathbf{u}$, so that $\rm N_{\rm r}\gg\rm N_{\rm in}$ is satisfied. Moreover, we consider the leaking rate $\gamma$ and ridge regression parameter $\beta$ as further quantities that have to be adjusted to our data. ## III Echo state network for 2d Moist Convection ### III.1 Direct Numerical Simulations DNS using the spectral element solver Nek5000 Fischer (1997); Scheel et al. (2013); nek (2017) were conducted to solve the two-dimensional moist Rayleigh- Bénard system (5)-(8) in a domain $\Omega=L\times H$ with aspect ratio $A=L/H=24$. The Rayleigh numbers are ${\rm Ra}_{\rm D}=2\cdot 10^{8}$, ${\rm Ra}_{\rm M}=4\cdot 10^{8}$. The Prandtl number is ${\rm Pr}=0.7$ representing moist air. The additional parameter is set to ${\rm CSA=0.3}$. In the vertical direction $y$, Dirichlet boundary conditions are imposed for both buoyancy fields at top and bottom in combination with free-slip boundary conditions for the velocity field. Periodic boundaries are set for all fields in the horizontal direction. We chose a spatial resolution of $N_{x}\times N_{y}=7200\times 384$ grid points and a time step size of $5.0\cdot 10^{-4}$. This setup corresponds to an absolutely unstable atmosphere, i.e. where both unsaturated and saturated air are unstable w.r.t. vertical displacements. The initial conditions are small perturbations around the diffusive equilibrium state $\overline{M}(y)$ and $\overline{D}(y)$, which result in turbulent convection. The flow statistics relaxes into a statistically stationary state (see Fig. 2) which provides training and test data for further processing. Figure 2: Turbulent kinetic energy $E_{\rm kin}(t)=\langle v_{x}^{2}+v_{y}^{2}\rangle_{x,y}$ of the moist Rayleigh-Bénard flow versus time $t$. After an initial transient the values such as those of $E_{\rm kin}(t)$ become statistically stationary. In the statistically stationary regime, $2000$ snapshots, each separated by $\Delta t=0.25$, are analyzed by a POD (see Sec. III.2). ### III.2 Data reduction by POD We sample a total of $n_{s}=2000$ snapshots of $v_{x}$, $v_{y}$, $M$ and $D$ at a time interval $\Delta t=0.25$ in the statistically stationary regime. The original DNS data snapshots have been sampled at a constant time interval such that we stick to constant time steps throughout this work, including the subsequent RC model. Furthermore, the data are spectrally interpolated on a uniform grid with a resolution of $N_{x}^{\prime}\times N_{y}^{\prime}=640\times 80$ points from the originally unstructured element mesh. They are decomposed into temporal mean and fluctuations subsequently, $\displaystyle v_{x}(x,y,t)$ $\displaystyle=$ $\displaystyle\langle v_{x}\rangle_{t}(x,y)+v_{x}^{\prime}(x,y,t)$ (19) $\displaystyle v_{y}(x,y,t)$ $\displaystyle=$ $\displaystyle\langle v_{y}\rangle_{t}(x,y)+v_{y}^{\prime}(x,y,t)$ (20) $\displaystyle D(x,y,t)$ $\displaystyle=$ $\displaystyle\langle D\rangle_{t}(x,y)+D^{\prime}(x,y,t)$ (21) $\displaystyle M(x,y,t)$ $\displaystyle=$ $\displaystyle\langle M\rangle_{t}(x,y)+M^{\prime}(x,y,t)$ (22) $\displaystyle q_{l}(x,y,t)$ $\displaystyle=$ $\displaystyle\langle q_{l}\rangle_{t}(x,y)+q_{l}^{\prime}(x,y,t)\,.$ (23) Figure 3: Eigenvalue spectrum of the POD mode obtained from the analysis of 2000 snapshots. (a) Individual and cumulative contribution of each mode. The shaded region indicates the first $\rm N_{\rm POD}=150$ modes which capture $81\%$ of the total energy of the original snapshot data. (b) Time coefficients $a_{p}(n)$ for the 1st, 2nd, 10th, 50th, and 100th mode are shown. The first coefficients show a slow variation compared to higher coefficients. Time series are shifted with respect to each other for better visibility. A grid dimension of $640\times 80$ in terms of ESN input dimensions ${\rm N_{\rm in}}$ is still too big. We therefore follow the approach in Pandey and Schumacher (2020) and introduce an intermediate step of data reduction before handing the data to the ESN. We make use of the periodicity and expand the data in a Fourier series in the horizontal $x$-direction and take the Karhunen-Loéve expansion, also known as POD of our data. In particular we choose the snapshot method Sirovich (1987) which decomposes the $k$-th component of a vector field $\mathbf{g}(x,y,t)$ into $\displaystyle g_{k}(x,y,t)$ $\displaystyle=\sum\limits_{p=1}^{\rm n_{\rm s}}\sum_{n_{x}=-N_{x}^{\prime}/2}^{N_{x}^{\prime}/2}a_{p,n_{x}}(t)\Phi_{k,n_{x}}^{(p)}(y)\exp{\left(i\frac{2\pi n_{x}x}{L}\right)}$ $\displaystyle=\sum\limits_{p=1}^{\rm n_{\rm s}}a_{p}(t)\Phi_{k}^{(p)}(x,y).$ (24) Here $a_{p}(t)$ and $\Phi_{k}^{(p)}(x,y)$ are the $p$-th time coefficient and the corresponding spatial POD mode respectively. In our approach we take the POD of $\mathbf{g}=(v_{x}^{\prime},v_{y}^{\prime},D^{\prime},M^{\prime})^{T}$. The POD spectrum of the turbulent convection data can be seen in Fig. 3(a). The eigenvalues of the covariance matrix fall off quickly and we therefore truncate the POD expansion at $\rm N_{\rm POD}=150\ll\rm n_{\rm s}$ and include only the most energetic modes (green shaded region). These capture $81\%$ of the total energy of the original data. In Fig. 3(b) the time coefficients $a_{p}(n)$ are shown for $p=1,2,10,50$, and 100 for all POD time steps. The first time coefficients ($a_{1}$ to $a_{10}$) posses only few temporal features opposed to higher coefficients. This range of active scales is inherent to turbulence as kinetic energy of large-scale motion is transferred to small eddies down to the Kolmogorov scale. Moreover, the influence of the additional nonlinearity due to the phase changes impacts the dynamics, as the first coefficients varied more in the dry RBC case with aspect ratio $6$ at ${\rm Ra=10^{7}}$ Pandey and Schumacher (2020). This is one order of magnitude below our Rayleigh number values. Nevertheless our RC model will receive values for all $\rm N_{\rm POD}$ coefficients and therefore for a wide range of temporal frequencies. We note here that the design of the RC model could be adapted to the different frequencies in future approaches. Figure 4 shows the spatial modes $\Phi_{2}^{(1)}$, $\Phi_{2}^{(50)}$ and $\Phi_{4}^{(1)}$, $\Phi_{4}^{(50)}$. Figure 4: Spatial structure of two POD modes for $v_{y}$: (a) $\Phi_{2}^{(1)}(x,y)$, (b) $\Phi_{2}^{(50)}(x,y)$ and the moist buoyancy field $M$: (c) $\Phi_{4}^{(1)}(x,y)$, (d) $\Phi_{4}^{(50)}(x,y)$. For visualization purposes the aspect proportions do not match the actual aspect ratio of $A=24$. We limit ourselves to the last $1400={\rm n}_{\rm train}+{\rm n}_{\rm test}$ time instances of our data and use the first 150 time coefficients $\mathbf{a}(n)=\left(a_{1}(n),a_{2}(n),...,a_{150}(n)\right)^{T}$ as the input for the ESN. The first ${\rm n}_{\rm train}=700$ snapshots are assembled into a training data set $\displaystyle\\{\mathbf{u}(n),\mathbf{y}^{\rm target}(n)\\}=\\{\mathbf{a}(n),\mathbf{a}(n+1)\\}$ (25) with $-{\rm n}_{\rm train}\leq n\leq-1$. The training data span 175 free-fall time units $T_{f}$ that correspond to 61 eddy turnover times. This time scale is given by $\tau_{\rm eddy}=H/u_{\rm rms}\approx 2.9T_{f}$ with $u_{\rm rms}=\langle u_{x}^{2}+u_{y}^{2}\rangle^{1/2}_{x,y,t}$. The ESN is trained to predict the time instance $\mathbf{a}(n+1)$ when given the POD time coefficients $\mathbf{a}(n)$ at the last time step as input. We use the first $46$ time steps to initialize the reservoir. Using eq. (18) we compute $W^{\rm out}$, which can then be used for prediction. In the prediction phase, we give the initial input $\mathbf{u}(0)=\mathbf{a}(0)$ to the reservoir and redirect the output to the input layer (see Fig. 1(b)) such that $\displaystyle\mathbf{u}(n)$ $\displaystyle=\mathbf{y}(n-1)\qquad n=1,...,{\rm n}_{\rm test}-1$ (26) with ${\rm n}_{\rm test}=700$. This coupling creates an autonomous system that generates new outputs without providing external inputs. Contrary to the teacher forcing approach, where at each time step the input is given by the actual time coefficients, this method is more suited for real world application. Finally, the outputs at each time step are gathered and validated by the last ${\rm n}_{\rm test}$ snapshots $\\{\mathbf{y}^{\rm val}(n)\\}=\\{\mathbf{a}(n+1)\\}$. ### III.3 Training of ESN and choice of hyperparameters We quantify the quality of ESN predictions, with the set of hyperparameters $\rm h=\\{\gamma,\beta,\rm N_{\rm r},{\rm D},\varrho(W^{\rm r})\\}$, by two types of measures. The mean square error ${\rm MSE}_{\rm h}$ of ESN output to the validation data $\displaystyle{\rm MSE}_{\rm h}$ $\displaystyle=\frac{1}{{\rm n}_{\rm test}}\sum\limits_{n=0}^{{\rm n}_{\rm test}}{\rm mse}(n)$ (27) where $\displaystyle{\rm mse}(n)$ $\displaystyle=\frac{1}{\rm N_{\rm POD}}\sum\limits_{i=1}^{{\rm N}_{\rm POD}}\left(y_{i}(n)-y^{\rm val}_{i}(n)\right)^{2}$ (28) is the mean square error at time step $n$ averaged over all ${\rm N}_{\rm POD}$ modes. Additionally, we take the physically more relevant normalized average relative error (NARE) as defined in Srinivasan et al. (2019) into account. For the moist buoyancy field $M$, it is for example given by $\displaystyle E_{\rm h}\left[\langle M\rangle_{x,t}\right]$ $\displaystyle=\frac{1}{C_{\max}}\int\limits_{0}^{1}\Big{|}\langle M\rangle_{x,t}^{\rm ESN}(y)-\langle M\rangle_{x,t}^{\rm POD}(y)\Big{|}dy$ (29) with $\displaystyle C_{\max}$ $\displaystyle=\frac{1}{2\max_{y\in[0,1]}(|\langle M\rangle_{x,t}^{\rm POD}|)}.$ (30) The superscript defines whether the field was reconstructed with $a_{i}(n)$ (POD) or $y_{i}(n)$ (ESN). It measures the integral deviation of the reconstructed line-time average profile $\langle\cdot\rangle_{x,t}$ of a specific field. We consider the three NAREs: $E_{\rm h}\left[\langle M\rangle_{x,t}\right]$, $E_{\rm h}\left[\langle q_{l}^{\prime 2}\rangle_{x,t}\right]$ and $E_{\rm h}\left[\langle v_{y}^{\prime}M^{\prime}\rangle_{x,t}\right]$, that is the NARE of the total moist buoyancy field $M$, the liquid water content fluctuations $q_{l}^{\prime}$, and the moist buoyancy flux fluctuations $v_{y}^{\prime}M^{\prime}$. $\gamma$ | $\beta$ | $D$ | $\varrho(W^{\rm r})$ ---|---|---|--- $0.50$ | $5\cdot 10^{-4}$ | $0.1$ | $0.00$ $0.60$ | $5\cdot 10^{-3}$ | $0.2$ | $0.90$ $0.70$ | $5\cdot 10^{-2}$ | $0.3$ | $0.91$ $0.80$ | $5\cdot 10^{-1}$ | $0.4$ | $0.92$ $0.90$ | | $0.5$ | $0.93$ $0.95$ | | $0.6$ | $0.94$ | | $0.7$ | $0.95$ | | | $0.96$ | | | $0.97$ | | | $0.98$ | | | $0.99$ | | | $1.00$ Table 1: Range of the four hyperparameters upon which a grid search was conducted. For each of the $1848$ combinations, an ESN was trained and validated with the training and validation data set. The reservoir dimension $N_{\rm r}$ was set to $4000$ for all studies. The MSE and NARE measures were computed and evaluated to find the optimal parameter set $h^{*}$. $\gamma^{*}$ | $\beta^{*}$ | $N_{\rm r}^{*}$ | $D^{*}$ | $\varrho(W^{\rm r})^{*}$ ---|---|---|---|--- $0.9$ | $5\cdot 10^{-1}$ | $4000$ | $0.1$ | $1.0$ MSE($\rm h^{*}$) | $E_{\rm h}\left[\langle M\rangle_{x,t}\right]$ | $E_{\rm h}\left[\langle q_{l}^{\prime 2}\rangle_{x,t}\right]$ | $E_{\rm h}\left[\langle v_{y}^{\prime}M^{\prime}\rangle_{x,t}\right]$ ---|---|---|--- $8.18\cdot 10^{-4}$ | $0.032$% | $0.033$% | $4.5$% Table 2: Hyperparameter set $h^{*}$ that was chosen for the ESN setup and the associated errors, which this ESN run has produced. The results of the reservoir with these listed hyperparameters are presented in section IV. Figure 5: Representative profiles taken from the error landscape for the leaking rate $\gamma$ (a,b,c) and the spectral radius $\varrho(W^{\rm r})$ (d,e,f). The data are obtained by a grid search study (see Table 1). We find a systematic dependence for the two quantities in this parameter domain. Note the different magnitudes between single quantity-NARE in (b,e) and multiple quantity-NARE in panels (c,f). The legends in (c) and (f) hold also for panels (a,b) and (d,e), respectively. A grid search for the four quantities $\gamma$, $\beta$, $D$, $\varrho(W^{\rm r})$ was conducted in a suitable range (see Table 1), based on the results in Pandey and Schumacher (2020). The reservoir size was fixed to $N_{\rm r}=4000$ for all runs. The resulting ${\rm MSE}_{h}$ and NAREs were computed to find an adequate parameter setting. Figure 5 shows ${\rm MSE}_{h}$, $E_{\rm h}\left[\langle q_{l}^{\prime 2}\rangle_{x,t}\right]$, $E_{\rm h}\left[\langle v_{y}^{\prime}M^{\prime}\rangle_{x,t}\right]$ and their dependence on $\varrho\left(W^{\rm r}\right)$ and $\gamma$. We detect a systematic dependence of both, spectral radius and leaking rate, even when slightly changing a third parameter (see legends). Interestingly, as both parameters are increased, the mean square error increases as well, while the NARE either decrease or pass a local minimum. We emphasize that our grid search is only an excerpt of the much bigger error landscape. Further, we did not average over multiple random realizations which would be the basis for a more rigorous discussion of parameter dependencies. Nevertheless, a starting point of the discussion of the hyperparameter dependencies is as follows: as the spectral radius grows, the magnitude of the argument of the hyperbolic tangent builds up too. This will saturate the activation function, which in turn will act in an increasingly binary way since $\lim\limits_{x\rightarrow\pm\infty}\tanh(x)=\pm 1$. In this limit of a fully saturated activation function, eq. (II.2) simplifies to $\displaystyle\mathbf{s}(n)\simeq(1-\gamma)\mathbf{s}(n-1)\pm\gamma\mathbf{1},$ (31) where $\mathbf{1}=(1,1,...,1)^{T}\in\mathbb{R}^{N_{\rm r}}$. This corresponds to a linear dependence of each reservoir state on its last instance plus the constant leaking rate $\gamma$ with stochastically changing sign, depending on the randomly generated weight matrices. As the leaking rate is increased towards unity, the memory effect is lost as well and the reservoir state is basically updated by the constant last term in (31). The resulting reservoir output will lead to increased mean square deviations from the varying ground truth signal. We thus speculate that such activation saturation is already satisfied for several reservoir state components at $\rho(W^{r})\lesssim 1$ which in turn contribute to the increasing ${\rm MSE}_{\rm h}$ in panel (d) of Fig. 5. Finally, we chose the hyperparameter set $h^{*}$, listed in Table 2 as it results in a minimum of $E_{\rm h}\left[\langle v_{y}^{\prime}M^{\prime}\rangle_{x,t}\right]$. The reason for settling with this measure is that it is susceptible to two ESN estimates. Quantities like $E_{\rm h}\left[\langle q_{l}^{\prime 2}\rangle_{x,t}\right]$, which depend on only one ESN estimate, exhibit small values for many parameter settings (see Fig. 5 (b),(e)). ## IV Results for the moist RBC case Figure 6: Time evolution of the POD time coefficients $a_{i}(t)$. The gray shaded area marks the training phase (reservoir output not shown). At the end of the training phase the prediction phase starts. The curves labeled POD stand for the ground truth of the evolution of the coefficient, while those labeled ESN are the network predictions. Panels (f)–(j) show the initial part of the forecast and correspond to (a)–(e). Figure 7: Fourier spectrum $\|\mathcal{F}(a_{i})\|^{2}$ of the POD time coefficients and of the corresponding reservoir prediction. The first 200 frequencies are shown only. After the ESN receives the initial input $\mathbf{u}(0)=\mathbf{a}(0)$, the autonomous predictor (see eq. (26)) produces estimates for the POD time coefficients which can be seen in Fig. 6. From the predicted coefficients and the known POD modes we can reconstruct all fields and compare these with the ground truth which is the POD expansion of the test data. Deviations for the first ten coefficients are detected while predictions for subsequent POD coefficients associated with less energetic modes agree with the values of the validation data for the first few time steps. Nevertheless, the ESN accomplishes to produce a time series with matching temporal frequency as the actual data, but shows bigger deviations to compute the trend of the slowly varying first coefficients. The frequency spectra of the ESN predictions for the coefficients $a_{i}(t)$ in comparison to those of the test data are shown in Figure 7 for 5 different cases. While the spectral values of the first 100 frequencies are captured well by the ESN, the higher frequency part starts to deviate in most cases. As discussed already above, this might be due to a simple RC model architecture, which does not differentiate between significantly different time scales that are always present in a turbulent flow. Nevertheless, the result underlines that the ESN is able to learn the time scales of the most significant POD coefficients. Figure 8(a) shows the mean square error ${\rm mse}(n)$ over all modes, as defined in (28), as a function of time steps after training. The mean error initially rises and then saturates. The fact that errors increase stems from the coupling scheme of output to input; small errors will inadvertently be amplified by the nonlinearity of the activation function in (II.2). Figure 8(b) shows the deviations $(y_{i}^{\rm val}(n)-y_{i}(n))$ for $i=1,10,50,100$. Figure 8: ESN prediction error: (a) Mean square error over all modes ${\rm mse}(n)$, see. eq. (28), a function of time steps $n$ after the training phase. (b) The difference $(y_{i}^{\rm val}(n)-y_{i}(n))$, i.e. the deviation of the ESN prediction $y_{i}(n)$ of $a_{i}$ at time steps $n$ after training. Figures 9(a)–(c) show the three weight matrices of the RC model. As described in section II.2, the input and reservoir weights are initialized randomly and left unchanged. With a reservoir density of $D=0.2$ the reservoir matrix $W^{\rm r}$ is a sparse matrix containing many weights equal to zero. Note further, that the fitted output weights $W^{\rm out}$ have low magnitudes in comparison to the entries of the input matrix $W^{\rm in}$. This is adjusted according to the number of reservoir nodes $N_{\rm r}$ and the magnitude of the training data. Moreover, the magnitude of the first $1+N_{\rm in}$ columns are close to zero. This indicates that the contributions of the output bias $b$ and the current input $\mathbf{u}$ to the reservoir output (see eq. (15)) are small. Figure 10 (a,b) shows the training and prediction phase dynamics of three exemplary hidden reservoir state components $s_{1},s_{1000}$, and $s_{4000}$. As the first $46$ time steps of the training data were used to initialize the reservoir state, the last $n_{\rm Train}-46$ time steps are shown in panel (a) only. During both phases the individual time series $s_{i}$ are confined to a certain subrange of the whole range $\left[-1,1\right]$. They have comparable amplitudes. Nevertheless, the prediction phase time series differ from the their training phase counterparts by slightly smoother variations with respect to time, see also the corresponding Fourier frequency spectra in panels (c,d) of the same figure. This might be explained by the fact that the states for $n\geq 0$ experience the learned output matrix $W^{\rm out}_{\ast}$ via the closed feedback loop. The states in the training phase, $n<0$, on the other hand, do neither experience the fitted output matrix, nor is the last output fed back to the reservoir. We suspect that this has a significant impact on the evolution of the $s_{i}$. Figure 9: Reservoir weight matrices: (a) Input weight matrix $W^{\rm in}$ which is a $4000\times 151$ matrix in our case. (b) Reservoir weight matrix $W^{\rm r}$ which is a $4000\times 4000$ matrix in the present case. All weights that are unequal to zero are marked as black dots. (c) Optimized output weight matrix $W_{\ast}^{\rm out}$ which is a $150\times 4151$ matrix. The aspect ratios of $W^{\rm in}$ and $W_{\ast}^{\rm out}$ have been adjusted for illustration purposes. Figure 10: Reservoir state components $s_{1},s_{1000},s_{4000}$ versus time step $n$ during (a) training and (b) prediction phases. Panels (c) and (d) show their corresponding Fourier spectra $\|\mathcal{F}(s_{i})\|^{2}$. Note that in (a), the first $46$ time steps are not shown, as they were used for the initialization of the reservoir. We now take a look at the reconstruction of the three fields $v_{x}(x,y)$, $v_{y}(x,y)$ and $M(x,y)$ with the reservoir outputs as temporal coefficients to see whether large-scale features are captured correctly. An instantaneous snapshot at the half-time of the prediction phase at time step $n=350$ is depicted in Fig. 11. We apply a POD-mode composition (24) to obtain the fluctuation fields $v_{x}^{\prime}$, $v_{y}^{\prime}$ and $M^{\prime}$ from the reservoir outputs $a_{p}(t)$. The mean profiles $\langle v_{x}\rangle_{t}$ and $\langle v_{y}\rangle_{t}$ and $\langle M\rangle_{t}$ are subsequently added to obtain the full fields, see eqns. (20) and (23). The resulting ESN predictions are displayed in panels (b), (d), and (f). For reference, the validation (POD) fields are shown in panels (a), (c), and (e). The horizontal velocity field $v_{x}(x,y)$ in (a) and (b) shows some differences in the structure of the right- and left-going fluid patches, but the large-scale structure as a whole is in surprisingly good qualitative agreement. The structure of vertical velocity field $v_{y}(x,y)$ in panel (c) and (d) does not show a systematic distinction, even though slight differences in shape and maximum values of up- and downdrafts are detectable. Finally the moist buoyancy field $M(x,y)$ in (f) does not fully reproduce all moist plumes that detach from the bottom plate, see validation field in (e). Nevertheless the predicted time coefficients lead to reconstructed fields that contain the same features as the original fields. Figure 11: Instantaneous snapshot of the fields (a, b) $v_{x}(x,y)$, (c, d) $v_{y}(x,y)$ and (e, f) $M(x,y)$ at time step $n=350$ after the training phase. Panels (a, c, e) are validation data from the POD and (b, d, f) the ESN output data. The fields were reconstructed using the first 150 $a_{p}(n)$ (POD) and the predictions $y_{p}(n)$ (ESN). Here, $n$ is a discrete time step. The ESN predictions deviate locally from the POD fields, but capture large- scale features of the flow. The aspect ratio has been adjusted again for illustration purposes. The corresponding colorbars can be seen on the right. To get a better grasp on the time evolution of the error in the field reconstruction, we compute a normalized field deviation at constant height $y$ of the vertical velocity field component in Fig. 12 which is given by $\displaystyle{\rm Err}(x,n)=\frac{|v_{y}^{\prime\rm POD}(x,n)-v_{y}^{\prime\rm ESN}(x,n)|}{\max_{x,y,n}\left(v_{y}^{\prime\rm POD}\right)}\Bigg{|}_{y={\rm const}}$ (32) where the superscript defines whether the field was reconstructed with $a_{i}(n)$ (POD) or $y_{i}(n))$ (ESN). We find that the fast growing errors in the time coefficients lead to fast amplifications of local field errors. Furthermore, different horizontal and vertical positions in the domain show different error magnitudes. Figure 12: Time evolution of the prediction error ${\rm Err}(x,n)$, which is given by eq. (32), at specified height $y$ (see top). As time progresses, the errors start to grow in magnitude. Different positions $(x,y)$ in the domain give rise to different magnitudes of the deviation. We now discuss the ability of the ESN to reproduce the low-order statistical properties of the turbulent flow. This is done by comparison of vertical line- time average profiles $\langle\cdot\rangle_{x,t}(y)$. The averages are taken along $x$-direction in combination with respect to time $t$. Such profiles are for example of interest in larger-scale atmospheric models for the parameterization of sub-grid scale transport Grabowski (2001); Khairoutdinov and Randall (2001). Figure 13 depicts the profiles as a function of the domain height $y$. The actual profiles obtained by the original DNS are plotted as a dash-dotted, the POD reconstruction as a solid and the reconstruction from the ESN outputs as a dashed line. Figure 13(a) shows the moist buoyancy $M$. Here the time mean $\langle M\rangle_{t}$ was added to see whether the reconstructed POD and ESN fields would deviate from the full DNS data. We observe an excellent agreement and find that the ESN produces correct fluctuations which preserve this profile. The fluctuations of the vertical moist buoyancy flux $\langle v_{y}^{\prime}M^{\prime}\rangle$ are shown in Fig. 13(b). Again, an excellent agreement between the curves in the bulk of the domain and small deviations in the boundary layers only are observable, despite the fact that this quantity is much more susceptible to errors since it consists of two ESN estimates. Finally the fluctuations from the liquid water content and the liquid water flux, further derived fields, are shown in 13(c) and (d), respectively. Here we see that POD and ESN curves match throughout the whole of the domain. We thus conclude that the ESN is able to reproduce essential low-order statistics very well. For comparison, we add a comparison of test data with the output of an LSTM network. The network parameters are as follows: the number of hidden states is 300 and the number of stacked LSTM layers 3. The loss function is again the mean-squared error, the optimization applies the method of adaptive moments, and the learning rate is $10^{-3}$ Goodfellow et al. (2016). A training of the network proceeds over 1000 epochs. We can conclude that the LSTM performs similarly well as the ESN even though the reproduced profiles deviate a bit for both fluxes in the center of the convection layer. Figure 13: Line-time averaged vertical profiles $\langle\cdot\rangle_{x,t}$ of (a) the full moist buoyancy $M$, (b) the vertical moist buoyancy flux $v_{y}^{\prime}M^{\prime}$, (c) the fluctuations of the liquid water content $q_{l}$ and (d) the vertical liquid water flux $v_{y}^{\prime}q_{l}^{\prime}$. The time average for DNS (dash-dotted) and POD (solid) were computed over the whole range of 1400 time steps, while for the ESN (dashed) and LSTM (dotted) only the range of the prediction phase ($700$ time steps) was taken into account. $\langle{\rm CC}^{\rm DNS}\rangle_{t}$ | $\langle q_{l}^{\rm DNS}\geq 0\rangle_{x,y,t}\cdot 10^{3}$ | $\langle{\rm CC}^{\rm POD}\rangle_{t}$ | $\langle q_{l}^{\rm POD}\geq 0\rangle_{x,y,t}\cdot 10^{3}$ | $\langle{\rm CC}^{\rm ESN}\rangle_{t}$ | $\langle q_{l}^{\rm ESN}\geq 0\rangle_{x,y,t}\cdot 10^{3}$ ---|---|---|---|---|--- $89.43$% | $3.24$ | $83.25$% | $2.72$ | $82.49$% | $2.72$ Table 3: Time mean average $\langle\rm CC\rangle_{t}$ and $\langle q_{l}\geq 0\rangle_{x,y,t}$ of the cloud cover CC and the volume average of liquid water for the DNS, POD and ESN case. See also Fig. 14. Figure 14: Clouds, i.e. $q_{l}(x,y,n)\geq 0$ at time step $n=350$ of (a) the POD and the (b) ESN prediction. The liquid water content $q_{l}$ differs in the magnitude and shape of its envelope, the isosurface $q_{l}=0$. (c) shows the cloud cover as defined in eq. (33) computed for the DNS data (brown), POD data (green) and the ESN predictions (blue). The POD approximation does not capture all of the original cloud cover; the value of the DNS exceeds the one of the POD by a few per cent. The cloud cover prediction of the ESN itself deviates by a few per cent in comparison to that of the POD. (d) shows the change of the volume average of positive liquid water content $\langle q_{l}\geq 0\rangle_{x,y}$ with time. Motivated by these results, we now investigate whether quantities such as the cloud cover can be also modeled by the ESN. We define the cloud cover CC of the two-dimensional domain $N_{x}^{\prime}\times N_{y}^{\prime}$ as the ratio of the number of vertical grid lines $N_{q_{l}>0}$ that contain at least one mesh point with $q_{l}>0$ along their vertical line of sight and the total number of vertical grid lines, $N_{x}^{\prime}$. Thus follows ${\rm CC}=\frac{N_{q_{l}>0}}{N_{x}^{\prime}}\times 100\%.$ (33) The time average $\langle{\rm CC}\rangle_{t}$ and volume-time average of positive liquid water content $\langle q_{l}\geq 0\rangle_{x,y,t}$ are given in Table 3. The truncation to the first $\rm N_{\rm POD}$ POD modes leads to a loss of about $6.9\%$ of the original DNS CC and a $16.1\%$ loss of $\langle q_{l}\geq 0\rangle_{x,y,t}$. We find good agreement between ESN estimate and POD results. In Fig. 14, the POD and ESN results of the cloud distribution at time step $n=350$ are shown. Despite small discrepancies in the local distribution of the liquid water content and the shape of the cloud boundaries, i.e. the isosurfaces $q_{l}=0$, the overall distribution is comparable. In panels (c) and (d) of the same figure, the time evolution of cloud cover and volume-time average positive liquid water content are displayed. While the predicted cloud cover does not deviate too much from the reference case, the variations in the amount of liquid water are less well reproduced. ## V Summary and Conclusion In the present work, we have applied a machine learning algorithm to two- dimensional turbulent moist Rayleigh-Bénard convection, in order to infer the large-scale evolution and low-order statistics of convective flow undergoing phase changes. We apply a specific learning scheme for recurrent neural networks, called echo state network, which has been applied for learning the dynamics of nonlinear systems, such as turbulent shear flows. Here, we test its capabilities successfully by fitting a reservoir to complex convection flow dynamics which results by the interaction of turbulence with the nonlinear thermodynamics originating from the first order phase changes between vapor and liquid water as present for example in atmospheric clouds. We therefore generate comprehensive data by means of a simple moist convection model in the Boussinesq framework, the moist Rayleigh-Bénard convection model. We obtain moist convection data from direct numerical simulations. As the 2d set of data has still a large amount of degrees of freedom and therefore cannot be passed directly to the echo state network, we introduce the POD as an additional dimensionality reduction step. We therefore decompose the data into a data-driven, spatially dependent basis with temporal coefficients, where the latter are fed to the reservoir. We truncate the POD to its most energetic modes and coefficients, reducing the degrees of freedom of the dynamical system at hand considerably. This reduced data set serves as the ground truth for the training of the echo state network as well as validation of its outputs. The network setup is tuned by conducting a comprehensive grid search of important hyperparameters. By coupling the output of the trained network back to its input, the autonomous system estimates the evolution of the temporal coefficients. Reconstructing the velocity and thermodynamic fields from these estimates, allow us to check whether the dynamics have been learned. We find an excellent agreement of the vertical profiles of moist buoyancy, vertical moist buoyancy transport as well as liquid water content. Furthermore, we report a good agreement of essential parameters in moist convection such as the fraction of clouds covering the two-dimensional atmosphere, as well its content of liquid water. This first approach of our reservoir computing model to moist Rayleigh-Bénard convection shows, its potential to infer low-order statistics from a set of training data. Though the reservoir output quickly diverges from the actual system trajectory, time averaged quantities are robustly reproduced. This result might seem trivial at first glance, yet the reservoir produces velocity and thermodynamic fluctuation fields which do not deviate too strongly from those of the original flow, even for combined quantities such as the liquid water flux across the layer. This indicates that the present echo state network did not just learn the statistics, but the dynamical system itself. Our additional comparison with an LSTM network gives a similar outcome. A more detailed comparison of both RNN implementations has to be left however as a future work. Our approach can be considered as a first step of applying reservoir computing as a machine learning-based parameterization. General circulation models already use multi-scale modeling methods where small-scale resolving models interact with the large-scale motion by their low-order statistics, essentially relaxing one to each other, e.g. in superparametrizations Grabowski and Smolarkiewicz (1999); Grabowski (2001); Khairoutdinov and Randall (2001). An echo state network can serve as a simple dynamical substitute for the unresolved subgrid scale transport. Even though the present results are promising, the development is still in its infancy. We state that for example the mathematical foundations of reservoir computing, which could provide deeper insights on the role of the hyperparameters on the prediction quality, are still mostly unexplored. Moreover, we reckon that for an extension of the ESN approach to a three dimensional flow, the data reduction step via POD will not suffice to cope with the large amount of simulation data. For this scenario one might propose the usage of a convolutional autoencoder/-decoder neural network in combination with the RC model. Furthermore, we mention that the machine learning algorithm is supposed here to learn dynamics of a nonlinear system which incorporates processes on different spatial and temporal scales. This circumstance is so far not fully captured by the network architecture. Particularly for turbulence, this might imply that the different spatial scales which interact with each other and exchange their energy, could be trained separately allowing for a subsequent coupling. The exploration of such ideas is currently under way and will be reported elsewhere. ## Acknowledgments This work is supported by the project ”DeepTurb – Deep Learning in and of Turbulence” which is funded by the Carl Zeiss Foundation. The authors gratefully acknowledge the Gauss Centre for Supercomputing e.V. (www.gauss- centre.eu) for funding this project by providing computing time through the John von Neumann Institute for Computing (NIC) on the GCS Supercomputer JUWELS at Jülich Supercomputing Centre (JSC). We thank Martina Hentschel, Erich Runge, and Sandeep Pandey for helpful comments. ## References * Brenner et al. (2019) M. P. Brenner, J. D. Eldredge, and J. B. Freund, Phys. Rev. Fluids 4, 100501 (2019). * Brunton et al. (2020) S. L. Brunton, B. R. Noack, and P. Koumoutsakos, Annu. Rev. Fluid Mech. 52, 477 (2020). * Pandey et al. (2020) S. Pandey, J. Schumacher, and K. R. Sreenivasan, J. Turbul. 21, 567 (2020). * Ling et al. (2016) J. Ling, A. Kurzawski, and J. Templeton, J. Fluid Mech. 807, 155 (2016). * Duraisamy et al. (2019) K. Duraisamy, G. Iaccarino, and H. Xiao, Annu. Rev. Fluid Mech. 51, 357 (2019). * Fonda et al. (2019) E. Fonda, A. Pandey, J. Schumacher, and K. R. Sreenivasan, Proc. Natl. Acad. Sci. 116, 8667 (2019). * Brenowitz and Bretherton (2018) N. D. Brenowitz and C. S. Bretherton, Geophys. Res. Lett. 45, 6289 (2018). * O’Gorman and Dwyer (2018) P. A. O’Gorman and J. G. Dwyer, J. Adv. Model Earth Sy. 10, 2548 (2018). * Gentine et al. (2018) P. Gentine, M. Pritchard, S. Rasp, G. Reinaudi, and G. Yacalis, Geophys. Res. Lett. 45, 5742 (2018). * Moller et al. (2020) S. Moller, C. Resagk, and C. Cierpka, Exp. Fluids 61, 111 (2020). * Goodfellow et al. (2016) I. Goodfellow, Y. Bengio, and A. Courville, _Deep Learning_ (MIT Press, 2016). * Schneider et al. (2017) S. Schneider, T. ans Lan, A. Stuart, and J. Teixeira, Geophys. Res. Lett. 44, 12396 (2017). * Mohan et al. (2020) A. Mohan, D. Tretiak, M. Chertkov, and D. Livescu, J. Turbul. 21, 525 (2020). * Stevens (2005) B. Stevens, Annu. Rev. Earth Planet. Sci. 33, 605 (2005). * Mellado (2017) J. P. Mellado, Annu. Rev. Fluid Mech. 49, 145 (2017). * Hochreiter and Schmidhuber (1997) S. Hochreiter and J. Schmidhuber, Neural Comput. 9, 1735 (1997). * Lukoševičius and Jaeger (2009) M. Lukoševičius and H. Jaeger, Comp. Sci. Rev. 3, 127 (2009). * Tanaka et al. (2019) G. Tanaka, T. Yamane, J. B. Héroux, R. Nakane, N. Kanazawa, S. Takeda, H. Numata, D. Nakano, and A. Hirose, Neural Netw. 115, 100 (2019). * Jaeger and Haas (2004) H. Jaeger and H. Haas, Science 304, 78 (2004). * Yildiz et al. (2012) I. B. Yildiz, H. Jaeger, and S. J. Kiebel, Neural Netw. 35, 1 (2012). * Pandey and Schumacher (2020) S. Pandey and J. Schumacher, Phys. Rev. Fluids 5, 113506 (2020). * Grabowski and Smolarkiewicz (1999) W. W. Grabowski and P. K. Smolarkiewicz, 133, 171 (1999). * Grabowski (2001) W. W. Grabowski, J. Atm. Sci. 58, 978 (2001). * Vlachas et al. (2020) P. R. Vlachas, J. Pathak, B. R. Hunt, T. P. Sapsis, G. M., E. Ott, and P. Koumoutsakos, Neural Netw. 126, 191 (2020). * Lu et al. (2017) Z. Lu, J. Pathak, B. Hunt, M. Girvan, R. Brockett, and E. Ott, Chaos 27 (2017). * Pathak et al. (2018) J. Pathak, B. Hunt, M. Girvan, Z. Lu, and E. Ott, Phys. Rev. Lett. 120, 024102 (2018). * Holmes et al. (2012) P. Holmes, J. L. Lumley, G. Berkooz, and C. W. Rowley, _Turbulence, Coherent Structures, Dynamical Systems and Symmetry_ , Cambridge Monographs on Mechanics (Cambridge University Press, Cambridge, UK, 2012), 2nd ed. * Srinivasan et al. (2019) P. A. Srinivasan, L. Guastoni, H. Azizpour, P. Schlatter, and R. Vinuesa, Phys. Rev. Fluids 4, 054603 (2019). * Nguyen et al. (2020) T. Nguyen, R. Baraniuk, A. Bertozzi, S. Osher, and B. Wang, in _Advances in Neural Information Processing Systems_ , edited by H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin (Curran Associates, Inc., 2020), vol. 33, pp. 1924–1936. * Fischer (1997) P. F. Fischer, J. Comput. Phys. 133, 84 (1997). * Scheel et al. (2013) J. D. Scheel, M. S. Emran, and J. Schumacher, New J. Phys. 15, 113063 (2013). * nek (2017) _nek5000 version 17.0_ (2017), URL https://nek5000.mcs.anl.gov. * Pauluis and Schumacher (2010) O. Pauluis and J. Schumacher, Commun. Math. Sci. 8, 295 (2010). * Weidauer et al. (2010) T. Weidauer, O. Pauluis, and J. Schumacher, New J. of Phys. 12, 105002 (2010). * Schumacher and Pauluis (2010) J. Schumacher and O. Pauluis, J. Fluid Mech. 648, 509–519 (2010). * Bretherton (1987) C. S. Bretherton, J. Atmos. Sci. 44, 1809 (1987). * Bretherton (1988) C. S. Bretherton, J. Atmos. Sci. 45, 2391 (1988). * Smith and Stechmann (2017) L. M. Smith and S. N. Stechmann, J. Atmos. Sci. 74, 3285 (2017). * Vallis et al. (2019) G. K. Vallis, D. J. Parker, and S. M. Tobias, J. Fluid Mech. 862, 162–199 (2019). * Abma et al. (2013) D. Abma, T. Heus, and J. P. Mellado, J. Atmos. Sci. 70, 2088 (2013). * Emmanuel (1994) K. A. Emmanuel, _Atmospheric Convection_ (Oxford University Press, 1994). * Pauluis and Schumacher (2011) O. Pauluis and J. Schumacher, Proc. Natl. Acad. Sci. USA 108, 12623 (2011). * Jaeger (2002) H. Jaeger, GMD-Forschungszentrum Informationstechnik (2002). * Doya (1992) K. Doya, [Proceedings] 1992 IEEE International Symposium on Circuits and Systems 6, 2777 (1992). * Jaeger (2001) H. Jaeger, GMD-Forschungszentrum Informationstechnik Technical Report 148 (2001). * Maass et al. (2002) W. Maass, T. Natschläger, and H. Markram, Neural Computation 14, 2531 (2002). * Lin et al. (2008) X. Lin, Z. Yang, and Y. Song, in _Advances in Knowledge Discovery and Data Mining_ (Springer Berlin Heidelberg, Berlin, Heidelberg, 2008), pp. 932–937. * Morando et al. (2013) S. Morando, S. Jemei, R. Gouriveau, N. Zerhouni, and D. Hissel, in _IECON 2013 - 39th Annual Conference of the IEEE Industrial Electronics Society_ (2013), pp. 1632–1637. * Pathak et al. (2017) J. Pathak, Z. Lu, B. R. Hunt, M. Girvan, and E. Ott, Chaos 27, 121102 (2017). * Lukoševičius (2012) M. Lukoševičius, LNCS 7700, 659 (2012). * Strauss et al. (2012) T. Strauss, W. Wustlich, and R. Labahn, Neural Comput. 24, 3246 (2012). * Sirovich (1987) L. Sirovich, Q. Appl. Math. XLV, 561 (1987). * Khairoutdinov and Randall (2001) M. F. Khairoutdinov and D. A. Randall, Geophys. Res. Lett. 28, 3617 (2001).
LATTICE ISOMORPHISMS OF LEIBNIZ ALGEBRAS DAVID A. TOWERS Department of Mathematics and Statistics Lancaster University Lancaster LA1 4YF England <EMAIL_ADDRESS> ###### Abstract Leibniz algebras are a non-anticommutative version of Lie algebras. They play an important role in different areas of mathematics and physics and have attracted much attention over the last thirty years. In this paper we investigate whether conditions such as being a Lie algebra, cyclic, simple, semisimple, solvable, supersolvable or nilpotent in such an algebra are preserved by lattice isomorphisms. Mathematics Subject Classification 2000: 17B05, 17B20, 17B30, 17B50. Key Words and Phrases: Lie algebras, Leibniz algebras, cyclic, simple, semisimple, solvable, supersolvable, nilpotent, lattice isomorphism. ## 1 Introduction An algebra $L$ over a field $F$ is called a Leibniz algebra if, for every $x,y,z\in L$, we have $[x,[y,z]]=[[x,y],z]-[[x,z],y]$ In other words the right multiplication operator $R_{x}:L\rightarrow L:y\mapsto[y,x]$ is a derivation of $L$. As a result such algebras are sometimes called right Leibniz algebras, and there is a corresponding notion of left Leibniz algebras, which satisfy $[x,[y,z]]=[[x,y],z]+[y,[x,z]].$ Clearly the opposite of a right (left) Leibniz algebra is a left (right) Leibniz algebra, so, in most situations, it does not matter which definition we use. Leibniz algebras which satisfy both the right and left identities are sometimes called symmetric Leibniz algebras. Every Lie algebra is a Leibniz algebra and every Leibniz algebra satisfying $[x,x]=0$ for every element is a Lie algebra. They were introduced in 1965 by Bloh ([7]) who called them $D$-algebras, though they attracted more widespread interest, and acquired their current name, through work by Loday and Pirashvili ([18], [19]). They have natural connections to a variety of areas, including algebraic $K$-theory, classical algebraic topology, differential geometry, homological algebra, loop spaces, noncommutative geometry and physics. A number of structural results have been obtained as analogues of corresponding results in Lie algebras. The Leibniz kernel is the set $I=$ span$\\{x^{2}:x\in L\\}$. Then $I$ is the smallest ideal of $L$ such that $L/I$ is a Lie algebra. Also $[L,I]=0$. We define the following series: $L^{1}=L,L^{k+1}=[L^{k},L](k\geq 1)\hbox{ and }L^{(0)}=L,L^{(k+1)}=[L^{(k)},L^{(k)}](k\geq 0).$ Then $L$ is nilpotent of class n (resp. solvable of derived length n) if $L^{n+1}=0$ but $L^{n}\neq 0$ (resp.$L^{(n)}=0$ but $L^{(n-1)}\neq 0$) for some $n\in{\mathbb{N}}$. It is straightforward to check that $L$ is nilpotent of class n precisely when every product of $n+1$ elements of $L$ is zero, but some product of $n$ elements is non-zero.The nilradical, $N(L)$, (resp. radical, $R(L)$) is the largest nilpotent (resp. solvable) ideal of $L$. The set of subalgebras of a nonassociative algebra forms a lattice under the operations of union, $\cup$, where the union of two subalgebras is the subalgebra generated by their set-theoretic union, and the usual intersection, $\cap$. The relationship between the structure of a Lie algebra $L$ and that of the lattice ${\cal L}(L)$ of all subalgebras of $L$ has been studied by many authors. Much is known about modular subalgebras (modular elements in ${\cal L}(L)$) through a number of investigations including [1, 12, 14, 27, 28, 29]. Other lattice conditions, together with their duals, have also been studied. These include semimodular, upper semimodular, lower semimodular, upper modular, lower modular and their respective duals (see [8] for definitions). For a selection of results on these conditions see [9, 13, 15, 16, 17, 20, 24, 26, 30, 31]. The subalgebra lattice of a Leibniz algebra, however, is rather different; in a Lie algebra every element generates a one-dimensional subalgebra, whereas in a Leibniz algebra elements can generate subalgebras of any dimension. So, one could expect different results to hold for Leibniz algebras anf this has been shown to be the case in [21].. Of particular interest is the extent to which important classes of Leibniz algebras are determined by their subalgebra lattices. In order to investigate this question we introduce the notion of a lattice isomorphism. If we denote the subalgebra lattice of $L$ by ${\mathcal{L}}(L)$, then a lattice isomorphism from $L$ to $L^{*}$ is a bijective map $\theta:{\mathcal{L}}(L)\rightarrow{\mathcal{L}}(L^{*})$ such that $\theta(A\cup B)=\theta(A)\cup\theta(B)$ and $\theta(A\cap B)=\theta(A)\cap\theta(B)$ for all $A,B\in{\mathcal{L}}(L)$. If $L$ is a Lie algebra over a field of characteristic zero the following were proved in [23]. ###### Theorem 1.1 * (i) If $L$ is simple then either * (a) $L^{*}$ is simple, or * (b) $L$ is three-dimensional non-split simple and $L^{*}$ is two-dimensional. * (ii) If $L$ is semisimple then either * (a) $L^{*}$ is semisimple, or * (b) $L$ is three-dimensional non-split simple and $L^{*}$ is two-dimensional. * (iii) If $\dim L,L^{*}>2$ and $R$ is the radical of $L$, then $R^{*}$ is the radical of $L^{*}$. * (iv) If $L$ is supersolvable of dimension $>2$, then $L^{*}$ is supersolvable. In [15] the following was proved. ###### Theorem 1.2 If $L$ is a solvable Lie algebra over a perfect field of characteristic different from $2,3$, then either * (i) $L^{*}$ is solvable, or * (ii) $L^{*}$ is three-dimensional non-split simple. We say that a Lie algebra $L$ is almost abelian if it is a split extension $L=L^{2}\dot{+}Fa$ with ad $a$ acting as the identity map on the abelian ideal $L^{2}$; $L$ is quasi-abelian if it is abelian or almost abelian. The quasi- abelian Lie algebras are precisely the ones in which every subspace is a subalgebra. The following is well-known and easy to show. ###### Proposition 1.3 If $L$ is a quasi-abelian Lie algebra over a field of characteristic zero then $L^{*}$ is quasi-abelian unless $\dim L=2$ and $L^{*}$ is three-dimensional non-split simple. In this paper we consider corresponding results for Leibniz algebras. First, in section two, we show that cyclic Leibniz algebras are characterised by their subalgebra lattice, and that a non-Lie Leibniz algebra cannot be lattice isomorphic to a Lie algebra. In section three we see that if $L$ is a non-Lie simple or semisimple Leibniz algebra then so is $L^{*}$. In section four, it is shown that if $L$ is a non-Lie solvable or supersolvable Leibniz algebra then so is $L^{*}$. It is also proved that the radical of a non-Lie Leibniz algebra is preserved by lattice isomorphisms. The final section is devoted to showing that if $L$ is a non-Lie nilpotent Leibniz algebra then so is $L^{*}$. Most of the above results are over fields of characteristic zero. Throughout, $L$ will denote a finite-dimensional Leibniz algebra over a field $F$. Algebra direct sums will be denoted by $\oplus$, whereas vector space direct sums will be denoted by $\dot{+}$. The notation ‘$A\subseteq B$’ will indicate that $A$ is a subset of $B$, whereas ‘$A\subset B$’ will mean that $A$ is a proper subset of $B$. If $A$ and $B$ are subalgebras of $L$ we will write $\langle A,B\rangle$ for $A\cup B$. The centre of $L$ is $Z(L)=\\{z\in L\mid[z,x]=[x,z]=0$ for all $x\in L\\}$. The Frattini ideal of $L$, $\phi(L)$, is the largest ideal of $L$ contained in all maximal subalgebras of $L$ . ## 2 Cyclic Leibniz algebras The only previous paper that we are aware of on this topic is by Barnes ([5]). The following example shows that the Leibniz kernel of a non-Lie Leibniz algebra is not necessarily preserved by a lattice isomorphism. ###### Example 2.1 Let $L=Fb+Fa$ where the only non-zero products are $[b,b]=a$, $[a,b]=a$. Then the only subalgebras of $L$ are ${0}$, $Fa$, $F(b-a)$ and $L$, and $I=Fa$. Then we can define a lattice automorphism of $L$ which interchanges $Fa$ and $F(b-a)$, and the latter is not an ideal of $L$ as $[b,b-a]=a$. Barnes called the above example the diamond algebra because of the structure of its lattice of subalgebras as a Hasse diagram, but that name has since been used for a different Leibniz algebra. He further showed that this example is exceptional in the following result. ###### Theorem 2.1 ([5, Theorem 3.1]) Let $L,L^{*}$ be Leibniz algebras with Leibniz kernels $I,I^{*}$ respectively, and let $\theta:L\rightarrow L^{*}$ be a lattice isomorphism. Suppose that $\dim L\geq 3$. Then $\theta(I)=I^{*}$. However, this paper does not appear to have been followed by further investigations into the subalgebra structure of a Leibniz algebra. Theorem 2.1, of course, has an immediate corollary. ###### Corollary 2.2 Let $L$ be a non-Lie Leibniz algebra. Then $L$ cannot be lattice isomorphic to a Lie algebra $L^{*}$. Proof. If $\dim L\geq 3$ then $I\neq 0$ if and only if $I^{*}\neq 0$. If $\dim L=2$ there are only two possibilities for $L$, both of them cyclic with basis $x,x^{2}$. In the first, $[x^{2},x]=0$ and the only proper subalgebra is $Fx^{2}$, and in the second, $[x^{2},x]=x^{2}$ and the only proper subalgebras are $Fx^{2}$ and $F(x-x^{2})$. However, every Lie algebra of dimension greater than one has more than two proper subalgebras. There is no non-Lie Leibniz algebra of dimension one. $\Box$ A Leibniz algebra $L$ is called cyclic if it is generated by a single element. In this case, $L$ has a basis $x,x^{2},\ldots,x^{n}(n>1)$ and products $[x^{i},x]=x^{i+1}$ for $1\leq i\leq n-1$, $[x^{n},x]=\alpha_{2}x^{2}+\ldots+\alpha_{n}x^{n}$, all other products being zero. Then we have the following. ###### Theorem 2.3 If $L$ is a cyclic Leibniz algebra over an infinite field $F$, then $L^{*}$ is also a cyclic Leibniz algebra of the same dimension. Proof. Over an infinite field a Leibniz algebra is cyclic if and only if it has finitely many maximal subalgebras, by [21, Corollary 2.3]. Moreover, the length of a maximal chain of subalgebras of a cyclic algebra is equal to its dimension. $\Box$ ###### Corollary 2.4 If $L$ is a nilpotent cyclic Leibniz algebra, then $L^{*}\cong L$. Proof. A nilpotent cyclic Leibniz algebra has only one maximal subalgebra, namely $I$, its Leibniz kernel. It follows that $L^{*}$ is nilpotent of the same dimension. Note that the restriction on the field is unnecessary here, since, if $M$ is the only maximal subalgebra of $L$ and $x\in L\setminus M$, we must have $L=\langle x\rangle$. $\Box$ Note that, in both of the above results, if $L=\langle x\rangle$ then $L^{*}=\langle x^{*}\rangle$, since $x$ does not belong to any of the maximal subalgebras of $L$, and this is inherited by $x^{*}$ in $L^{*}$. ###### Proposition 2.5 Let $L=A\dot{+}Fx$ be a non-Lie Leibniz algebra in which $A$ is a minimal abelian ideal of $L$ and $x^{2}=0$. Then $L$ is cyclic and $A=I$. Proof. Since $L$ is not a Lie algebra, $A=I$, $[L,A]=0$ and $[A,L]\neq 0$, so $[A,x]=A$. Let $0\neq a\in A$. Then $(x+a)^{n}=R_{x}^{n-1}(a)$ for $n\geq 2$, which implies that $[(x+a)^{n},x]=R_{x}^{n}(a)\in\langle x+a\rangle$ for $n\geq 1$. Hence $\langle x+a\rangle\cap A$ is an ideal of $L$ and so equals $A$ or $0$. However, the latter implies that $[a,x]=0$, whence $A=Fa$ and $[A,x]=0$, a contradiction. It follows that $L=\langle x+a\rangle$. $\Box$ ## 3 Semisimple Leibniz algebras The following useful result was proved by Barnes in [4]. Note that we have modified the statement to take account of the fact that Barnes’ result is stated for left Leibniz algebras and we are dealing with right Leibniz algebras. ###### Lemma 3.1 Let $A$ be a minimal ideal of the Leibniz algebra $L$. Then $[L,A]=0$ or $[x,a]=-[a,x]$ for all $a\in A$, $x\in L$. A Leibniz algebra $L$ is called simple if its only ideals of $L$ are $0$, $I$ and $L$, and $L^{2}\neq I$. If $L/I$ is a simple Lie algebra then $L$ is not necessarily a simple Leibniz algebra. It is said to be semisimple if $R(L)=I$. This definition agrees with that of a semisimple Lie algebra, since, in this case, $I=0$. Semisimple Leibniz algebras are not necessarily direct sums of simple Leibniz algebras (see [10] or [2]). We have the following version of Levi’s Theorem. ###### Theorem 3.2 (Barnes [3]) Let $L$ be a finite-dimensional Leibniz algebra over a field of characteristic $0$. Then there is a semisimple Lie subalgebra $S$ of $L$ such that $L=S\dot{+}R(L)$. We shall need the following result which was proved by Gein in [11, p. 23]. ###### Lemma 3.3 Let $S$ be a three-dimensional non-split simple Lie algebra, and let $R$ be an irreducible $S$-module. Then, for any $s\in S$, $R$ has an ad $s$-invariant subspace of dimension less than or equal to two. If $U$ is a subalgebra of $L$ and $0=U_{0}<U_{1}<\ldots<U_{n}=U$ is a maximal chain of subalgebras of $U$ we will say that $U$ has length $n$. ###### Theorem 3.4 Let $L=S\dot{+}A$ be a Leibniz algebra over a field of characteristic zero, where $S$ is a three-dimensional non-split simple Lie algebra and $A$ is a minimal abelian ideal of $L$. Then $L^{*}$ has a simple Lie subalgebra. Proof. Suppose that $L^{*}$ does not have a simple Lie subalgebra. Then $R(L^{*})\neq 0$, by Theorem 3.2, and so $L^{*}$ has a minimal abelian ideal $B^{*}$. As $S^{*}$ is a maximal subalgebra of $L^{*}$ we must have that $L^{*}=S^{*}\dot{+}B^{*}$. If $\dim A=1$ we have that $L=S\oplus A$ is a Lie algebra and hence, so is $L^{*}$, giving that $S^{*}\cong S$, by [23, Lemma 3.3] and contradicting our supposition. Hence $\dim A\geq 2$. Now maximal subalgebras of $L$ are of two types: they are isomorphic to $S$, and so have length $2$, or they are of the form $Fs\dot{+}A$, where $s\in S$, and so are solvable of length at least $3$. Moreover, $A$ is the intersection of those of the second type. The same must be true of the maximal subalgebras of $L^{*}$ and so $B^{*}=A^{*}$ and $L^{*}=S^{*}\dot{+}A^{*}$. Also, $\dim S^{*}=2$, by Theorem 1.1. Now $\phi(L^{*})=(\phi(L))^{*}=0$, so $L^{*}=A^{*}\dot{+}C^{*}$, where $C^{*}$ is abelian, by [6, Corollary 2.9]. Since $S^{*}\cong C^{*}$, we have that $S^{*}$ is abelian. Let $0\neq s^{*}\in S^{*}$, $0\neq a^{*}\in A^{*}$ and let $f(\theta)$ be the polynomial of smallest degree for which $f(R_{s^{*}})(a^{*})=0$. It follows from the fact that $S^{*}$ is abelian that $\\{x^{*}\in A^{*}:f(R_{s^{*}})(x^{*})=0\\}$ is an ideal of $L^{*}$, and hence that it coincides with $A^{*}$. Clearly then $f(\theta)$ is the minimum polynomial of $R_{s^{*}}|_{A^{*}}$. Suppose that there is an $s_{1}^{*}\in S^{*}$ for which the minimum polynomial for $R_{s_{1}^{*}}$ has degree two. and let this polynomial be $f(\theta)=\theta^{2}-\lambda_{2}\theta-\lambda_{1}$. Pick $s_{2}^{*}\in S^{*}$ linearly independent of $s_{1}^{*}$. Then $\displaystyle[[a^{*},s_{1}^{*}],s_{1}^{*}]$ $\displaystyle=\lambda_{1}a^{*}+\lambda_{2}[a^{*},s_{1}^{*}]\hbox{ and }$ $\displaystyle[[a^{*},s_{2}^{*}],s_{2}^{*}]$ $\displaystyle=\alpha_{1}a^{*}+\alpha_{2}[a^{*},s_{2}^{*}]\hbox{ so }$ $\displaystyle[[a^{*},s_{1}^{*}],s_{2}^{*}]$ $\displaystyle=[a^{*},[s_{1}^{*},s_{2}^{*}]]+[[a^{*},s_{2}^{*}],s_{1}^{*}]=[[a^{*},s_{2}^{*}],s_{1}^{*}]$ $\displaystyle=\beta_{1}a^{*}+\beta_{2}[a^{*},s_{1}^{*}]+\beta_{3}[a^{*},s_{2}^{*}],$ since $[[a^{*},s_{1}^{*}+s_{2}^{*}],s_{1}^{*}+s_{2}^{*}]\in Fa^{*}+F[a^{*},s_{1}^{*}+s_{2}^{*}]$. Now $[[[a^{*},s_{2}^{*}],s_{1}^{*}],s_{1}^{*}]=\lambda_{1}[a^{*},s_{2}^{*}]+\lambda[[a^{*},s_{2}^{*}],s_{1}^{*}],$ so $\displaystyle(\beta_{2}\lambda_{1}+\beta_{3}\beta_{1})a^{*}+(\beta_{1}+\beta_{2}\lambda_{2})[a^{*},s_{1}^{*}]+\beta_{3}^{2}[a^{*},s_{2}^{*}]$ $\displaystyle=\lambda_{2}\beta_{1}a^{*}+\lambda_{2}\beta_{2}[a^{*},s_{1}^{*}]+(\lambda_{1}+\lambda_{2}\beta_{3})[a^{*},s_{2}^{*}].$ Since $f(\theta)$ is irreducible, $\beta_{3}^{2}\neq\lambda_{1}+\lambda_{2}\beta_{3}$ and so $[a^{*},s_{2}^{*}]=\gamma_{1}a^{*}+\gamma_{2}[a^{*},s_{1}^{*}]$. Hence $A^{*}$ is two dimensional. Put $A=Fa+F[a,s]$. Choose $s_{1},s_{2}$ to be elements of $S$ such that $s,s_{1},s_{2}$ are linearly independent. Then $[a,s_{1}]=\alpha a+\beta[a,s]$ and $[a,s_{2}]=\gamma a+\delta[a,s]$ for some $\alpha,\beta,\gamma,\delta\in F$. Thus $[a,s_{1}-\beta s]=\alpha a$ and $[a,s_{2}-\delta s]=\gamma a$. But $s_{1}-\beta s$ and $s_{2}-\delta s$ are linearly independent, so $[a,S]=[a,<s_{1}-\beta s,s_{2}-\delta s>]\subseteq Fa$ and $A$ is one dimensional, a contradiction. $\Box$ ###### Corollary 3.5 Let $L$ be a non-Lie semisimple Leibniz algebra over a field of characteristic zero. Then $L^{*}$ is a non-Lie semisimple Leibniz algebra. Proof. We have that $I\neq 0$, so $L=I\dot{+}S$ where $S$ is a semisimple Lie algebra, by Theorem 3.2. Then $L^{*}/I^{*}$ is a semisimple Lie algebra or $\dim L^{*}/I^{*}=2$ and $S$ is $3$-dimensional non-split simple, by Theorem 1.1(ii). Suppose that the latter holds, so $L^{*}$ is solvable. Let $A$ be a minimal ideal of $L$ inside $I$ and put $B=A\dot{+}S$. Then $B^{*}$ has a simple Lie subalgebra, by Theorem 3.4 and $L^{*}$ cannot be solvable. Hence the former holds and $L^{*}$ is a non-Lie semsimple Leibniz algebra. $\Box$ A subalgebra $U$ of $L$ is called upper semi-modular if $U$ is a maximal subalgebra of $\langle U,B\rangle$ for every subalgebra $B$ of $L$ such that $U\cap B$ is maximal in $B$. Using this concept we have a further corollary. ###### Corollary 3.6 Let $L$ be a non-Lie simple Leibniz algebra over a field of characteristic zero. Then $L^{*}$ is a non-Lie simple Leibniz algebra. Proof. We have that $L=I\dot{+}S$ where $S$ is a simple Lie subalgebra of $L$ and $I\neq 0$. If $L^{*}/I^{*}$ is not simple then $S$ must be three- dimensional non split simple, by Theorem 1.1(i), and we get a contradiction as in the previous corollary. Let $0\neq A^{*}$ be an ideal of $L^{*}$. Suppose first that $A^{*}\subseteq I^{*}$. Then $A$ is an upper semi-modular subalgebra of $L$ with $A\subseteq I$. Let $s\in S$. Then $A\cap Fs=0$ is a maximal subalgebra of $Fs$. Hence $A$ is a maximal subalgebra of $C=\langle A,s\rangle$. Now $A\subseteq C\cap I\subset C$, so $A=C\cap I$. Thus $[s,A],[A,s]\subseteq C\cap I=A$, so $A$ is an ideal of $L$, whence $A=I$. It follows that $A^{*}=I^{*}$. Next, suppose that $A^{*}\not\subseteq I^{*}$. Then $I^{*}+A^{*}=L^{*}$ and $I^{*}\cap A^{*}=I^{*}$ or $0$, by the previous paragraph. The former implies that $A^{*}=L^{*}$; the latter gives that $L^{*}=I^{*}\oplus A^{*}$ giving $I^{*}=0$ and $L^{*}=A^{*}$ again. Clearly $(L^{*})^{2}\neq I^{*}$, so $L^{*}$ is a non-Lie simple Leibniz algebra. $\Box$ ## 4 Solvable and supersolvable Leibniz algebras ###### Proposition 4.1 Let $L$ be a non-Lie solvable Leibniz algebra over a field of characteristic zero. Then $L^{*}$ is a non-Lie solvable Leibniz algebra. Proof. Let $L$ be a minimal counter-example. Then $L^{*}$ has a semisimple Lie subalgebra $S^{*}$, and so $S(\neq L)$ must be two dimensional and $S^{*}$ must be three-dimensional non-split simple. Moreover, $L^{*}=S^{*}\dot{+}A^{*}$, where $A^{*}$ is a minimal ideal of $L^{*}$, since, otherwise, this is a smaller counter-example. But then $L$ has a simple subalgebra, by Theorem 3.4, a contradiction. $\Box$ ###### Lemma 4.2 Let $L$ be a Leibniz algebra over a field of characteristic zero. Then the radical, $R$, of $L$ is the intersection of the maximal solvable subalgebras of $L$. Proof. Let $\Gamma$ be the intersection of the maximal solvable subalgebras of $L$. Then $R\subseteq\Gamma$. Furthermore, $\Gamma$ is invariant under all automorphisms of $L$, and hence is invariant under all derivations of $L$, by [22, Corollary 3.2]. It follows that $\Gamma$ is a right ideal of $L$. But $[x,y]+[y,x]\in I\subseteq\Gamma$ for all $x\in L$, $y\in\Gamma$, so $\Gamma$ is an ideal of $L$, whence $\Gamma\subseteq R$. $\Box$ Then we have the following corollaries to Proposition 4.1. ###### Corollary 4.3 Let $L$ be a non-Lie Leibniz algebra over a field of characteristic zero, and let $R$ be the radical of $L$. Then $R^{*}$ is the radical of $L^{*}$. Proof. Let $U$ be a maximal solvable subalgebra of $L$. If $U$ is non-Lie then $U$ is solvable, by Proposition 4.1. If $U$ is Lie, then $U^{*}$ is solvable, unless $\dim U=2$ and $U^{*}$ is three-dimensional non-split simple. If $R^{*}=0$ then $L^{*}$, and hence $L$, is a Lie algebra, a contradiction. Hence $\dim R^{*}\neq 0$. Moreover, $R\subseteq U$. If $R=U$ then $R$ is a maximal solvable subalgebra of $L$, which is impossible unless $R=L$. But then the result follows from Proposition 4.1. So suppose $\dim R=0,1$. The former implies that $L$ is a semisimple Lie algebra, which is impossible. The latter implies that $L=S\oplus Fa$, where $S$ is a semisimple Lie algebra. But this is also a Lie algebra and so is impossible. It follows that $U^{*}$ must be a maximal solvable subalgebra of $L^{*}$. The result now follows from Lemma 4.2. $\Box$ A subalgebra $U$ of $L$ is called lower semi-modular in $L$ if $U\cap B$ is maximal in $B$ for every subalgebra $B$ of $L$ such that $U$ is maximal in $\langle U,B\rangle$. We say that $L$ is lower semi-modular if every subalgebra of $L$ is lower semi-modular in $L$. ###### Corollary 4.4 Let $L$ be a non-Lie supersolvable Leibniz algebra over a field of characteristic zero. Then $L^{*}$ is supersolvable. Proof. We have that $L$ is solvable and lower semi-modular, by [21, Proposition 5.1]. It follows from Proposition 4.1 that the same is true of $L^{*}$. Hence $L^{*}$ is supersolvable, by [21, Proposition 5.1] again. $\Box$ ## 5 Nilpotent Leibniz algebras A Lie algebra $L$ is callled almost nilpotent of index $n$ if it has a basis $\\{x;e_{11},\ldots,e_{1r_{1}};\ldots;e_{n1},\ldots,e_{nr_{n}}\\}$ such that $\displaystyle-[e_{ij},x]=[x,e_{ij}]$ $\displaystyle=e_{ij}+e_{i+1,j}\hbox{ for }1\leq i\leq n-1,1\leq j\leq r_{i},$ $\displaystyle-[e_{nj},x]=[x,e_{nj}]$ $\displaystyle=e_{nj}\hbox{ and }r_{j}\leq r_{j+1}\hbox{ for }1\leq j\leq n-1$ all other products being zero. The following result was proved in [25] ###### Theorem 5.1 Let $L$ be a nilpotent Lie algebra of index $n$ and of dimension greater than two for which $L^{*}$ is not nilpotent, over a field of characteristic zero. Then $L^{*}$ is almost nilpotent of index $n$. Moreover, every almost nilpotent Lie algebra is lattice isomorphic to a nilpotent Lie algebra. For non-Lie Leibniz algebras we have the following result. ###### Theorem 5.2 Let $L$ be a nilpotent non-Lie Leibniz algebra over a field of characteristic zero. Then $L^{*}$ is a non-Lie nilpotent Leibniz algebra. First we need a lemma. ###### Lemma 5.3 Let $L$ be a nilpotent Leibniz algebra and let $W=Fw$ be a minimal ideal of $L$ contained in the Leibniz kernel, $I$, of $L$. Then $W^{*}$ is a minimal ideal of $L^{*}$ and $W^{*}\subseteq Z(L^{*})$. Proof. Suppose that $x\notin I$, where $x^{n}=0$ but $x^{n-1}\neq 0$. Then $S=\langle x,W\rangle=\langle x\rangle+W$ and $\langle x\rangle\cap W=0$ or $1$. The former implies that $\langle x\rangle$ is a maximal subalgebra of $S$, whence $\langle x^{*}\rangle$ is a maximal subalgebra, and hence an ideal, of $S^{*}=\langle x^{*},W^{*}\rangle$. The latter implies that $W\subseteq\langle x\rangle$, whence $W^{*}\subseteq\langle x^{*}\rangle$. In either case, $[w^{*},x^{*}]\in\langle x^{*}\rangle\cap I^{*}$. Hence $[w^{*},x^{*}]=\sum_{i=2}^{n}\lambda_{i}(x^{*})^{i}$. Suppose that $\lambda_{2}\not=0$ and consider $\langle\lambda_{2}x-w\rangle$. If $W\subseteq\langle x\rangle$ then $W=Fx^{n}$ and $\langle\lambda_{2}x-w\rangle=\langle x\rangle$. If $W\not\subseteq\langle x\rangle$ then $(\lambda_{2}x-w)^{k}=\lambda_{2}^{k}x^{k}-\lambda_{2}^{k-1}\mu^{k-1}w$, where $[w,x]=\mu w$. In either case, $\langle\lambda_{2}x-w\rangle$ is a cyclic subalgebra of dimension $n$. However, $(\lambda_{2}x^{*}-w^{*})^{2}=\lambda_{2}^{2}(x^{*})^{2}-\lambda_{2}\sum_{i=2}^{n}\lambda_{i}(x^{*})^{i}=\lambda_{2}\sum_{i=3}^{n}\lambda_{i}(x^{*})^{i},$ so $\langle\lambda_{2}x^{*}-w^{*}\rangle$ is a cyclic subalgebra of dimension $n-1$, contradicting Corollary 2.4. It follows that $\lambda_{2}=0$. A similar argument shows that $\lambda_{i}=0$ for all $2\leq i\leq n$, so $[w^{*},x^{*}]=0$. Also, $[x^{*},w^{*}]=0$, since $w^{*}\in I^{*}$, from which the result follows. $\Box$ Now we can prove Theorem 5.2. Proof. We have $L/L^{2}$ is abelian and $L^{2}=\phi(L)$, so $L^{*}/\phi(L^{*})$ is almost abelian or three-dimensional non-split simple. The latter is impossible, as it would imply that $L^{*}=\phi(L^{*})\dot{+}S^{*}=S^{*}$, where $S^{*}$ is three-dimensional non- split simple, by Theorem 3.2. But then $L$ is a two-dimensional Lie algebra, by Theorem 2.2, a contradiction. It follows that $L^{*}/\phi(L^{*})$, and hence $L^{*}$, is supersolvable (see [5, Theorems 3.9 and 5.2]) and has nilradical $N^{*}=\phi(L^{*})+Fe_{11}^{*}+\dots+Fe_{1r_{1}}^{*}.$ Let $L$ be a minimal counter-example, so $L$ is non-Lie and nilpotent, but $L^{*}$ is not nilpotent. Now $I$ is non-zero, so choose a minimal ideal $W=Fw$ of $L$ inside $I$. We have that $W^{*}$ is a minimal ideal of $L^{*}$ inside $Z(L^{*})$, by Lemma 5.3. Then $L^{*}/W^{*}$ is not nilpotent, so $L/W$ is a Lie algebra and $L^{*}/W^{*}$ is almost nilpotent. Hence there is a basis $\\{x^{*};e_{11}^{*},\ldots,e_{1r_{1}}^{*};\ldots;e_{n1}^{*},\ldots,e_{nr_{n}}^{*},w^{*}\\}\hbox{ for }L^{*}$ such that $\displaystyle[x^{*},e_{ij}^{*}]$ $\displaystyle=e_{ij}^{*}+e_{i+1,j}^{*}+\lambda_{ij}w^{*}\hbox{ for }1\leq i\leq n-1,1\leq j\leq r_{i},$ $\displaystyle[x^{*},e_{nj}^{*}]$ $\displaystyle=e_{nj}^{*}+\lambda_{nj}w^{*}\hbox{ and }r_{j}\leq r_{j+1}\hbox{ for }1\leq j\leq n-1,$ where $\lambda_{ij}\in F$, $I^{*}=Fw^{*}$ and $(N^{*})^{2}\subseteq W^{*}$. Let $M^{*}$ be spanned by all of the basis vectors for $L^{*}$ apart from $e_{11}^{*}$. Then $M^{*}$ is not nilpotent and has nilradical $F^{*}$ spanned by all of the basis vectors apart from $e_{11}^{*}$ and $x^{*}$. By the minimality, we must have that $M$ is Lie and $M^{*}$ is almost nilpotent, so $(F^{*})^{2}=0$, $(x^{*})^{2}=0$ and $[e_{ij}^{*},x^{*}]=-[x^{*},e_{ij}^{*}]$ for all of the $e_{ij}^{*}$’s apart from $e_{11}^{*}$. But also $[e_{11}^{*},x^{*}]=-e_{11}^{*}-e_{21}^{*}+\mu w^{*}$ for some $\mu\in F$, so $\displaystyle[e_{11}^{*},x^{*}]$ $\displaystyle=[[x^{*},e_{11}^{*}],x^{*}]-[e_{21}^{*},x^{*}]$ $\displaystyle=[x^{*},[e_{11}^{*},x^{*}]]+[(x^{*})^{2},e_{11}]+[x^{*},e_{21}^{*}]$ $\displaystyle=-[x^{*},e_{11}^{*}]-[x^{*},e_{21}^{*}]+[x^{*},e_{21}^{*}]=-[x^{*},e_{11}^{*}]$ We now claim that $(N^{*})^{2}=0$. It suffices to show that $[N^{*},e_{11}^{*}]=0$, which we do by a backwards induction argument. We have, for any $f^{*}\in F^{*}$, $\displaystyle[f^{*},e_{11}^{*}]$ $\displaystyle=[f^{*},[x^{*},e_{11}^{*}]-e_{21}^{*}-\lambda_{11}w^{*}]=[f^{*},[x^{*},e_{11}^{*}]]$ $\displaystyle=[[f^{*},x^{*}],e_{11}^{*}]-[[f^{*},e_{11}^{*}],x^{*}]=[[f^{*},x^{*}],e_{11}^{*}],$ (1) since $[f^{*},e_{11}^{*}]\in W\subseteq Z(L^{*})$. Now putting $f^{*}=e_{nj}^{*}$ gives $[e_{nj}^{*},e_{11}^{*}]=[[e_{nj}^{*},x^{*}],e_{11}^{*}]=-[e_{nj}^{*},e_{11}^{*}],$ whence $[e_{nj}^{*},e_{11}^{*}]=0$. So now suppose that $[e_{ij}^{*},e_{11}^{*}]=0$ for some $2\leq i\leq n$. Putting $f^{*}=e_{i-1,j}^{*}$ ($(i-1,j)\neq(2,1)$) in (1) gives $[e_{i-1,j}^{*},e_{11}^{*}]=[[e_{i-1,j}^{*},x^{*}].e_{11}^{*}]=-[e_{i-1,j}^{*},e_{11}^{*}]$ which, again, yields that $[e_{i-1,j}^{*},e_{11}^{*}]=0$. Finally, note that, if we now put $f^{*}=e_{11}^{*}$, then (1) remains valid, so $(e_{11}^{*})^{2}=0$ and $(N^{*})^{2}=0$. Now replace $e_{nj}^{*}$ by $e_{nj}^{*}+\lambda_{nj}w^{*}$, $e_{ij}^{*}$ by $e_{ij}^{*}+(-1)^{n-i}\lambda_{i+1,j}w^{*}$ to see that $L^{*}$ is almost nilpotent and $L$ is a Lie algebra, a contradiction. Hence the result holds. $\Box$ ## References * [1] R.K. Amayo and J. Schwarz, ‘Modularity in Lie Algebras’, Hiroshima Math. J. 10 (1980), 311-322. * [2] Sh. Ayupov, K. Kudaybergenov, B. Omirov and K. Zhao, ‘Semisimple Leibniz algebras, their derivations and automorphisms’, Linear and Multilinear Alg. (2019), https://doi.org/10.1080/03081087.2019.1567674. * [3] D.W. Barnes, ‘On Levi’s Theorem for Leibniz algebras’, Bull. Aust. Math. Soc. 86 (2012), 184-185. * [4] D.W. Barnes, ‘Some theorems on Leibniz algebras’, Comm. Alg. 39 (7) (2011), 2463-2472. * [5] D.W. Barnes, ‘Lattices of subalgebras of Leibniz algebras’, Comm. Alg. 40 (11) (2012), 4330-4335. * [6] C. Batten Ray, L. Bosko-Dunbar, A. Hedges, J.T. Hird, K. Stagg and E. Stitzinger, ‘A Frattini theory for Leibniz algebras’, Comm. Alg. 41(4) (2013), 1547–1557. * [7] A. Bloh. ‘On a generalization of the concept of Lie algebra’. Dokl. Akad. Nauk SSSR. 165 (1965), 471–473. * [8] K. Bowman and D.A.Towers, ‘Modularity conditions in Lie algebras’, Hiroshima Math. J. 19 (1989), 333-346. * [9] K. Bowman and V.R. Varea, ‘Modularity* in Lie algebras’, Proc. Edin. Math. Soc. 40(2) (1997), 99-110. * [10] I. Demir, K.C. Misra and E. Stitzinger, ‘On some structures of Leibniz algebras’, in Recent Advances in Representation Theory, Quantum Groups, Algebraic Geometry, and Related Topics, Contemporary Mathematics, 623. Amer. Math. Soc., Providence, RI, (2014), 41–54. * [11] A.G. Gein, ‘Projections of a Lie algebra of characteristic zero’, Izvestija vyss̃. ucebn. Zaved. Mat. 22(4) (1978), 26-31. * [12] A.G. Gein, ‘Modular rule and relative complements in the lattice of subalgebras of a Lie algebra’, Sov. Math. 31(3) (1987), 22-32; translated from Izv. Vyssh. Uchebn. Zaved. Mat. 83 (1987), 18-25. * [13] A.G. Gein, ‘Semimodular Lie algebras’, Siberian Math. J. 17 (1976), 243-248; translated from Sibirsk Mat. Z. 17 (1976), 243-248. * [14] A.G. Gein, ‘On modular subalgebras of Lie algebras’, Ural Gos. Univ. Mat. Zap. 14 (1987), 27-33. * [15] A.G. Gein and V.R. Varea, ‘Solvable Lie algebras and their subalgebra lattices’, Comm. Alg. 20(8) (1992), 2203-2217. Corrigenda: Comm. Alg. 23(1) (1995), 399-403. * [16] B. Kolman, ’Semi-modular Lie algebras’, J. Sci. Hiroshima Univ. Ser. A-I 29 (1965), 149-163. * [17] A.A. Lashi, ‘On Lie algebras with modular lattices of subalgebras’, J. Algebra 99 (1986), 80-88. * [18] J.-L. Loday, ‘Une version non commutative des algèbres de Lie: les algèbres de Leibniz’. Enseign. Math. (2) 39 (3–4) (1993), 269–293. * [19] J.-L. Loday and T. Pirashvili, ‘Universal enveloping algebras of Leibniz algebras and (co)homology’, Math. Annalen 296 (1) (1993) 139–158. * [20] C. Scheiderer, ‘Intersections of maximal subalgebras in Lie algebras’, J. Algebra 105 (1987), 268-270. * [21] S. Siciliano and D.A. Towers, ‘On the subalgebra lattice of a Leibniz algebra’, arXiv:2010.12254 (2020). * [22] D.A. Towers, ‘A Frattini theory for algebras’, Proc. London Math. Soc. (3) 27 (1973), 440–462. * [23] D.A. Towers, ‘Lattice isomorphisms of Lie algebras’, Math. Proc. Camb. Phil. Soc. 89 (1981), 285-292. * [24] D.A. Towers, ‘Semimodular subalgebras of a Lie algebra’, J. Algebra 103 (1986), 202-207. * [25] D.A. Towers, ‘Almost nilpotent Lie algebras’, Glasgow Math. J. 29 (1987), 7-11. * [26] D.A. Towers, ‘On modular* subalgebras of a Lie algebra’, J. Algebra 190 (1997), 461-473. * [27] V.R. Varea, ‘Modular subalgebras, Quasi-ideals and inner ideals in Lie Algebras of prime characteristic’, Comm. Alg. 21(11) (1993), 4195–4218. * [28] V.R. Varea, ‘The subalgebra lattice of a supersolvable Lie algebra’, In Lecture Notes in Mathematics. Springer-Verlag: New York, 1989; Vol. 1373, 81-92. * [29] V.R. Varea, ‘On modular subalgebras in Lie algebras of prime characteristic,, Contemporary Math. 110 (1990), 289-307. * [30] V.R. Varea, ‘Lower Semimodular Lie algebras’, Proc. Edin. Math. Soc. 42(1999), 521-540. * [31] V.R. Varea, ‘Lie algebras whose maximal subalgebras are modular’, Proc. Roy. Soc. Edinburgh Sect. A 94 (1983), 9-13.
# Charge affinity and solvent effects in numerical simulations of ionic microgels Giovanni Del Monte1,2,3,∗, Fabrizio Camerin1,4,∗, Andrea Ninarello1,2, Nicoletta Gnan1,2, Lorenzo Rovigatti2,1, Emanuela Zaccarelli1,2,∗ 1 CNR Institute of Complex Systems, Uos Sapienza, piazzale Aldo Moro 2, 00185, Roma, Italy 2 Department of Physics, Sapienza University of Rome, piazzale Aldo Moro 2, 00185 Roma, Italy 3 Center for Life NanoScience, Istituto Italiano di Tecnologia, viale Regina Elena 291, 00161 Rome, Italy 4 Department of Basic and Applied Sciences for Engineering, via Antonio Scarpa 14, 00161 Roma, Italy <EMAIL_ADDRESS><EMAIL_ADDRESS><EMAIL_ADDRESS> ###### Abstract Ionic microgel particles are intriguing systems in which the properties of thermo-responsive polymeric colloids are enriched by the presence of charged groups. In order to rationalize their properties and predict the behaviour of microgel suspensions, it is necessary to develop a coarse-graining strategy that starts from the accurate modelling of single particles. Here, we provide a numerical advancement of a recently-introduced model for charged co- polymerized microgels by improving the treatment of ionic groups in the polymer network. We investigate the thermoresponsive properties of the particles, in particular their swelling behaviour and structure, finding that, when charged groups are considered to be hydrophilic at all temperatures, highly charged microgels do not achieve a fully collapsed state, in favorable comparison to experiments. In addition, we explicitly include the solvent in the description and put forward a mapping between the solvophobic potential in the absence of the solvent and the monomer-solvent interactions in its presence, which is found to work very accurately for any charge fraction of the microgel. Our work paves the way for comparing single-particle properties and swelling behaviour of ionic microgels to experiments and to tackle the study of these charged soft particles at a liquid-liquid interface. ††: J. Phys.: Condens. Matter * August 27, 2024 Keywords: ionic microgels, charge affinity, solvophobic attraction, volume phase transition, form factors ## 1 Introduction Soft matter is a very active branch of condensed matter physics, which comprises, among other systems, colloidal suspensions, whose constituent particles can greatly vary in shape, softness and function. Soft matter encompasses not only synthetic particles, but also constituents of many biological systems, such as proteins, viruses and even cells, whose size ranges between the nano and the micrometer scale. A peculiar aspect of soft matter systems is the great variety of amorphous states they can form, including glasses [1, 2] and gels [3, 4, 5]. Indeed, a large amount of work in this field is devoted to the study of these non-ergodic states which may form due to different kind of interactions, such as steric, hydrophobic or electrostatic ones, both of attractive and repulsive nature. Sometimes, a single colloidal particle is already quite a complex object whose behaviour at the collective level is strongly connected to the microscopic features of the particle itself. This situation is typical of soft colloids, i.e. deformable particles with internal degrees of freedom strongly influencing their mutual interactions, which makes them already intrinsically multi-scale. For these systems a theoretical approach is quite challenging even at the single-particle level, thus it is convenient to rely on the development of suitable coarse-grained models [6] that allow to greatly reduce the system complexity, still capturing the important ingredients to be retained for a correct description of the collective behavior. This strategy is very profitable for the case of microgel particles [7] that, combining together the properties of colloids and polymers, can be viewed as a prototype example of soft particles [8, 9]. A microgel is a microscale gel whose internal polymeric network controls its peculiar properties. By varying the constituent monomers, microgels can be made responsive to temperature, pH or to external forces [7]. For their intriguing properties, they are employed in a wide variety of applications, ranging from biomedical purposes [10, 11] to paper restoration [12]. In order to be able to predict the behaviour of dense microgel suspensions and the formation of arrested states, it is important to properly take into account the internal degrees of freedom of the particles, by modelling their effective interactions in such a way that the resulting object can still shrink, deform and interpenetrate [13, 14, 15, 16]. Hence, an accurate modelling of a single microgel is a necessary pre-requisite for a correct description of bulk suspensions. To validate numerical models at the single- particle level, there are a number of different experiments we can refer to. One of the most straightforward is the measurement of the effective size of the microgels via dynamic light scattering experiments. Upon varying the controlling parameter of the dispersion, the so-called swelling curves can be determined. For instance, microgels synthesized by employing a thermoresponsive polymer, such as Poly(N-isopropyl-acrylamide) (PNIPAM), undergo a so-called Volume Phase Transition (VPT) [7] at a temperature $T_{\scriptscriptstyle\mathit{VPT}}\approx 32^{\circ}$C from a swollen to a collapsed state. In addition, form factors can be measured by small-angle scattering experiments of dilute microgel suspensions, either using neutrons [17], x-rays [18] or even visible light for large enough microgels [19]. This observable directly provides information on the inner structure of the microgels and shows that microgels prepared via precipitation polymerization [20] can be modelled as effective fuzzy spheres [17], where a rather homogeneous core is surrounded by a fluffy corona, giving rise to what is usually called a core- corona structure. A more complex situation arises when ionic groups are added to the synthesis to make microgels responsive also to external electric fields [21, 22] and to pH variations [23]. A case study of such these co-polymerized microgels is the one made of PNIPAM and polyacrylic acid (PAAc) [20, 24, 25, 26], that is pH-responsive due to the the weak acidic nature of AAc monomers. An increasing amount of work in the last years has focused on modelling single-particle behaviour both of neutral [27, 28, 29] and ionic microgels [30, 31, 32, 33]. For the latter case, we have recently shown [33] that it is important to take into account both the disordered nature of the network, as opposed to the diamond lattice structure [29], and an explicit treatment of charges and counterions. Indeed, mean-field approaches completely neglect the complex, heterogeneous nature of the charge distribution within the microgel. In this work, we extend our previous effort by going one step further towards a realistic numerical treatment of ionic co-polymerized microgels. In Ref. [33], we modelled a single microgel particle such that all of its monomers, including charged ones, experienced a solvophobic attraction on increasing temperature. Here, instead, charged monomers experience Coulomb and steric interactions only. This is expected to be more realistic, since charged or polar groups always remain hydrophilic irrespectively of temperature, thus having a distinct behaviour with respect to all other monomers. We examine the consequences of this difference on the microgel swelling behaviour as well as on its structure and charge distributions across the VPT. In the second part of the manuscript, we consider the presence of an explicit solvent, to examine its effects on the structural properties of the microgel. In this way, we aim to make our model suitable for situations where solvent effects become important. In particular, this will enable us to study the effect of charges for microgels adsorbed at liquid-liquid interfaces, similarly to what we recently put forward for neutral microgels [34, 35]. ## 2 Methods The coarse-grained microgels used in this work are prepared as in Refs. [27] starting from $N$ patchy particles of diameter $\sigma$, which sets the unit of length, confined in a spherical cavity. A fraction $c=0.05$ of these particles has four patches on their surface to mimic the typical crosslinker connectivity in a chemical synthesis, while all the others have two patches to represent monomers in a polymer chain. During the assembly, an additional force is employed on the crosslinking particles to increase their concentration in the core of the microgel in agreement with experimental features [18]. Once a fully-bonded configuration is reached (when the fraction of formed bonds is greater than $0.995$), a permanent topology is obtained by enforcing the Kremer-Grest bead-spring model [36]. In this way, all particles experience a steric repulsion via the Weeks-Chandler-Anderson (WCA) potential, $V_{\rm WCA}(r)=\begin{cases}4\epsilon\left[\left(\frac{\sigma}{r}\right)^{12}-\left(\frac{\sigma}{r}\right)^{6}\right]+\epsilon&\text{if $r\leq 2^{\frac{1}{6}}\sigma$}\\\ 0&\text{otherwise,}\end{cases}$ (1) where $\epsilon$ sets the energy scale and $r$ is the distance between the centers of two beads. Additionally, bonded particles interact via the Finitely Extensible Nonlinear Elastic potential (FENE), $V_{\rm FENE}(r)=-\epsilon k_{F}R_{0}^{2}\ln\left[1-\left(\frac{r}{R_{0}\sigma}\right)^{2}\right]\text{ if $r<R_{0}\sigma$,}$ (2) with $k_{F}=15$ which determines the stiffness of the bond and $R_{0}=1.5$ which determines the maximum bond distance. The resulting microgel is thus constituted by a disordered polymer network with a core-corona structure and form factors across the VPT that closely resemble experimental ones for microgels synthesized via the precipitation polymerization procedure [18]. Once the microgels are prepared, we add electrostatic interactions to mimic co-polymerized polymer networks with charged groups. To this aim, we randomly assign a negative charge $-e^{*}$ to a given fraction $f$ of microgel monomers, to mimic the acrylic acid dissociation in water, where $e^{*}=\sqrt{4\pi\varepsilon_{0}\varepsilon_{r}\sigma\epsilon}$ is the reduced unit charge (which roughly corresponds to the elementary charge $e$, considering $\epsilon\approx k_{B}T$ at room temperature and $\sigma$ as the polymer’s Kuhn length), and $\varepsilon_{0}$ and $\varepsilon_{r}$ are the vacuum and relative dielectric constants. Accordingly, we insert in the simulation box an equal number of positively charged counterions with charge $e^{*}$ to ensure the neutrality of the system. Interactions among charged beads are given by the reduced Coulomb potential $V_{\rm coul}(r)=\frac{q_{i}q_{j}\sigma}{e^{*2}r}\epsilon,$ (3) where $q_{i}$ and $q_{j}$ are the charges of counterions or charged monomers. We adopt the particle-particle-particle-mesh method [37] as a long-range solver for the Coulomb interactions. As discussed in a previous contribution [33], the size of the counterions is set to $0.1\sigma$ to facilitate their diffusion within the microgel network and to avoid spurious excluded volume effects. They interact with all other species simply via the WCA potential. The swelling behaviour of a thermoresponsive microgel can be reproduced in molecular dynamics simulations either by means of an implicit solvent, namely by adding a potential that mimics the effect of the temperature on the polymer, or by explicitly adding coarse-grained solvent particles within the box. In the first case, we employ a solvophobic potential $V_{\alpha}(r)=\begin{cases}-\epsilon\alpha&\text{if }r\leq 2^{1/6}\sigma\\\ \frac{1}{2}\alpha\epsilon\left\\{\cos\left[\gamma{\left(\frac{r}{\sigma}\right)}^{2}+\beta\right]-1\right\\}&\text{if }2^{1/6}\sigma<r\leq R_{0}\sigma\\\ 0&\text{if }r>R_{0}\sigma\end{cases}$ (4) with $\gamma=\pi\left(\frac{9}{4}-2^{1/3}\right)^{-1}$ and $\beta=2\pi-\frac{9}{4}\gamma$ [38]. This potential introduces an effective attraction among polymer beads, modulated by the parameter $\alpha$, whose increase corresponds to an increase in the temperature of the dispersion. For $\alpha=0$ no attraction is present, which corresponds to fully swollen, i.e. low-temperature, conditions. For neutral microgels, the VPT is encountered at $\alpha\approx 0.65$, while a full collapse is usually reached for $\alpha\gtrsim 1.2$. In the first part of this work, we analyze two different models, based on a different treatment of the interactions of charged ions on the microgels: in Model I all monomers experience a total interaction potential where $V_{\alpha}$ (Eq. 4) is added to the Kremer-Grest interactions, as previously done in Ref. [33]; in Model II only neutral monomers experience this total interaction potential, while the charged monomers do not interact with $V_{\alpha}$, i.e. $\alpha=0$ for them in all cases. This second situation is equivalent to leaving unaltered the behaviour of charged groups of the microgel as the solvent conditions change, so that they always retain a good affinity for the solvent ($\alpha=0$). A similar treatment is also adopted for the counterions, for which $\alpha=0$ for both Model I and Model II. In the second part of this work, we explicitly consider the presence of the solvent in driving the Volume Phase Transition. Solvent particles are modelled within the Dissipative Particle Dynamics (DPD) framework in order to avoid spurious effects which may arise from the use of a standard Lennard-Jones potential due to the excessive excluded volume of the solvent [39]. In the DPD scheme, two particles $i$ and $j$ experience a force $\vec{F}_{ij}=\vec{F}^{C}_{ij}+\vec{F}^{D}_{ij}+\vec{F}^{R}_{ij}$, where: $\displaystyle\vec{F}^{C}_{ij}$ $\displaystyle=$ $\displaystyle a_{ij}w(r_{ij})\hat{r}_{ij}$ (5) $\displaystyle\vec{F}^{D}_{ij}$ $\displaystyle=$ $\displaystyle-\gamma w^{2}(r_{ij})(\vec{v}_{ij}\cdot\vec{r}_{ij})\hat{r}_{ij}$ (6) $\displaystyle\vec{F}^{R}_{ij}$ $\displaystyle=$ $\displaystyle 2\gamma\frac{k_{B}T}{m}w(r_{ij})\frac{\theta}{\sqrt{\Delta t}}\hat{r}_{ij}$ (7) where $\vec{F}^{C}_{ij}$ is a conservative repulsive force, with $w(r_{ij})=1-r_{ij}/r_{c}$ for $r_{ij}<r_{c}$ and $0$ elsewhere, $\vec{F}^{D}_{ij}$ and $\vec{F}^{R}_{ij}$ are a dissipative and a random contribution of the DPD, respectively; $\gamma$ is a friction coefficient, $\theta$ is a Gaussian random variable with average $0$ and unit variance, and $\Delta t$ is the integration time-step. We set $r_{c}=1.75\sigma$ and $\gamma=2.0$ in all the simulations. Here $a_{i,j}$ quantifies the repulsion between two particles $i$ and $j$, which effectively allows the tuning of the monomer-solvent (m,s) and solvent-solvent (s,s) interactions. Following our previous work [39], we fix $a_{s,s}=25.0$ while we vary $a_{m,s}\equiv a$ between 5.0 and 16.0, that is the range where the collapse of a neutral microgel takes place. The reduced DPD density is set to $\rho_{s}r_{c}^{3}=3.9$ (with $\rho_{s}$ the actual number density of solvent beads). With this choice of parameters, we previously showed that this model reproduces the swelling behaviour and structural properties of a neutral microgel particle, in quantitative agreement with the implicit solvent model that was explicitly tested against experiments [18]. To compare the explicit solvent model with the implicit one, we only consider Model II, where charged monomers always retain a high affinity for the solvent. We will show later that, in the explicit treatment, this translates to having charged monomers- solvent interactions (ch,s) always set to $a_{ch,s}=0$. All other interactions are identical to the implicit solvent model. Simulations are performed with LAMMPS [40]. The equations of motion are integrated with the velocity-Verlet algorithm. All particles have unit mass $m$, the integration time-step is $\Delta t=0.002\sqrt{m\sigma^{2}/\epsilon}$ and the reduced temperature $T^{*}=k_{B}T/\epsilon$ is set to 1.0 by means of a Nosè-Hoover thermostat for implicit solvent simulations or via the DPD thermostat for explicit solvent ones. In the former case the number of monomers in the microgels is fixed to $N\approx 42000$, while for the latter case we limit the analysis to $N\approx 5000$ due to the very large computational times owing to the presence of the solvent. From equilibrated trajectories, we directly calculate the form factor of the microgel as, $P(q)=\left\langle\frac{1}{N}\sum_{i,j=1}^{N}\exp{(-i\vec{q}\cdot\vec{r}_{ij})}\right\rangle,$ (8) where $r_{ij}$ is the distance between monomers $i$ and $j$, while the angular brackets indicate an average over different configurations and over different orientations of the wavevector $\vec{q}$ (for each $q$ we consider $300$ distinct directions randomly chosen on a sphere of radius $q$). Also, we determine the radius of gyration $R_{g}$ of the microgels as a measure of their swelling degree. This is calculated as $R_{g}=\left\langle\left[\frac{1}{N}\sum_{i=1}^{N}(\vec{r}_{i}-\vec{r}_{CM})^{2}\right]^{\frac{1}{2}}\right\rangle,$ (9) where $\vec{r}_{CM}$ is the position of the center of mass of the microgel. For each swelling curve, representing $R_{g}$ as a function of the effective temperature $\alpha$ (implicit solvent) or $a$ (explicit solvent), we define an effective VPT temperature, either $\alpha_{\scriptscriptstyle\mathit{VPT}}$ or $a_{\scriptscriptstyle\mathit{VPT}}$, as the inflection point of a cubic spline interpolating the simulation points. Finally, the radial density profile for all monomers is defined as $\rho(r)=\left\langle\frac{1}{N}\sum_{i_{=1}}^{N}\delta(|\vec{r}_{i}-\vec{r}_{CM}|-r)\right\rangle.$ (10) By restricting the sum in Eq. 10 to only charged monomers or to counterions, we also calculate $\rho_{CH}(r)$ and $\rho_{CI}(r)$, that are the radial density profiles of charged microgel monomers and of counterions, respectively. In addition, we define the net charge density profile as $\rho_{Q}(r)=-\rho_{CH}(r)+\rho_{CI}(r).$ (11) ## 3 Results and Discussion Figure 1: Simulation snapshots. Ionic microgels with $f=0.2$ and $N\approx 42000$ obtained in implicit solvent for (top) Model I and (bottom) Model II for $\alpha=0,0.74$ and $1.4$ (from left to right panels). Blue (cyan) particles are neutral (charged) microgel monomers; smaller purple particles represent counterions. ### 3.1 On the role of the affinity of charged monomers for the solvent #### 3.1.1 Swelling behaviour Figure 2: Swelling curves. Radius of gyration $R_{g}$ as a function of the solvophobic parameter $\alpha$ and different values of $f$ for microgels with $N\approx 42000$ for the case where charged monomers (a) have a varying affinity for the solvent (Model I) and (b) have always a good affinity for the solvent (Model II). Figure 3: Form factors. Form factors for charged microgels with $N\approx 42000$ and (top) $f=0.032$ and (bottom) $f=0.2$, simulated in implicit solvent for Models I and II. The models are compared at the same $\alpha$: for $f=0.032$, $\alpha=0,0.48,0.64,0.8,1.4$; for $f=0.2$, $\alpha=0,0.6,0.74,1.0,1.4$ (from left to right). The corresponding neutral case ($f=0$) is also displayed for comparison. Straight lines in the central panel of the bottom row highlight the two power-law behaviours of the form factors at intermediate (full line) and high (dashed line) $q$ values, that are present for both models, extensively discussed in Ref. [33]. We start by discussing the influence of charges on the VPT for microgels with $N\approx 42000$ in implicit solvent. As explained in the Methods section, we directly compare the case where the affinity of charged beads for the solvent varies (Model I) to the case where it remains unchanged (Model II). Representative simulation snapshots of the two models for the highest value of charge fraction investigated in this work ($f=0.2$) are reported in Fig. 1. Here we focus on different swelling stages of the microgels upon increasing $\alpha$. We notice immediately that a large amount of inhomogeneities persists in Model II at large $\alpha$, in contrast with the behavior of Model I where a full collapse is achieved. This can be better quantified by the swelling curves, reporting the variation of the radius of gyration $R_{g}$ versus the effective temperature $\alpha$, that are shown in Fig. 2 for different values of the charge fraction $f$. For both models we observe that the increase of $f$ shifts the transition towards larger effective temperatures, but important differences arise at large $\alpha$, as displayed in the snapshots. In Model I, where charged beads experience Coulomb as well as solvophobic interactions, the VPT occurs at all studied $f$, as shown in Fig. 2(a). Using the $\alpha$-temperature mapping established in Ref. [18] through a comparison to experiments, the VPT temperature observed for $f=0.2$ microgels would correspond to $T\approx 38^{\circ}$C. However, experiments on ionic microgels, for which the amount of charges was systematically varied [41, 26], have shown that even for values of $f$ smaller than $0.2$, the microgel does not collapse below $40^{\circ}$C. As hypothesized in our earlier work [33], Model I neglects the interplay between the hydrophilic character of the co-polymer and its charge content. However, charged or polar groups, such as AAc groups, are known to remain hydrophilic even at high temperatures [42], which would increase the stability of the microgel in the swollen state with increasing $f$. We thus incorporate such a feature in Model II by removing solvophobic interactions for charged microgel beads. The resulting swelling curves, shown in Fig. 2(b), clearly demonstrate that for $f=0.20$ the VPT is not encountered within the investigated solvophobicity range, up to values of $\alpha$ that would correspond to temperatures above $50^{\circ}$C, in qualitative agreement with experimental observations [41, 26, 43]. #### 3.1.2 Structural properties It is now important to compare the two models from the structural point of view, to check whether major differences arise. We start the analysis by looking at the form factors which, in our previous work on Model I [33], were shown to exhibit novel features with respect to neutral microgels. In particular, we found evidence that for $\alpha<\alpha_{\scriptscriptstyle\mathit{VPT}}$, the standard fuzzy-sphere- like model was not able to describe the numerical form factors, Instead, the emergence of two distinct power-law behaviours was found immediately after the first peak, at intermediate and high $q$ values, respectively [33]. This was attributed to the presence of charges in the inhomogeneous structure of the microgel, which gives rise to different features for core and corona regions, each being characterized by a different domain size. It is now crucial to verify whether such distinctive behaviour also persists when the interactions among charged beads are modelled more realistically. Figure 4: Density profiles. Top panels show the monomers density profiles for an ionic microgel with $f=0.2$ and $N\approx 42000$ as a function of the distance from the microgel center of mass $r$ obtained in implicit solvent for Models I and II. Bottom panels report the ions and counterions (c-ions) density profiles for $f=0.2$ for both models. The models are compared at the same $\alpha$: $0,0.6,0.74,1.0,1.4$ (from left to right). The corresponding neutral case ($f=0$) is also displayed for comparison. Figure 5: Comparison of Models I and II at the same $R_{g}$. Radial density profiles for an ionic microgel with $f=0.2$ and $N\approx 42000$ at $R_{g}\approx 21$, where $\alpha=0.9$ and $\alpha=1.4$ for Models I and II, respectively. The inset shows the corresponding ions and counterions (c-ions) density profiles. Figure 6: Charge density profile. Net charge density profile $\rho_{Q}(r)$ as defined in Eq. 11 for ionic microgels with $N\approx 42000$ and (top) $f=0.032$ , (bottom) $f=0.2$, as a function of the distance from the microgel center of mass $r$, simulated in implicit solvent for Models I and II. The models are compared at the same $\alpha$: for $f=0.032$, $\alpha=0,0.48,0.64,0.8,1.4$; for $f=0.2$, $\alpha=0,0.6,0.74,1.0,1.4$ (from left to right panels). Fig. 3 reports the form factors for Models I and II with $f=0.032$ and $f=0.2$, in comparison to the neutral case ($f=0$), at different values of $\alpha$. For $f=0.032$, the amount of charges in the microgel is still too low to observe differences between the two models and the neutral microgel. Also, at large $\alpha$, the form factor is that of a collapsed microgel in all cases, as expected from the swelling curves in Fig. 2. For $f=0.2$ and low enough $\alpha$, the behaviour of the two models is again very similar, with the form factors of ionic microgels showing a first peak that is systematically smaller with respect to that of the neutral case. At intermediate $\alpha$, we find that two power-law-like behaviours are compatible with both sets of data for charged microgels, while the neutral case does not show such a feature. This finding, already elaborated in Ref. [33], appears to be a distinctive feature of our numerical model of ionic microgels and is the result of the combination of a random charge distribution within a disordered, heterogeneous network topology with the explicit treatment of ions and counterions. Such a distinctive feature was tentatively attributed to the different degree of swelling of the corona and of the core, but still awaits a direct experimental confirmation. However, hints of a similar two-step decay for $P(q)$ were reported in Ref. [44] and would certainly deserve further investigation in future experiments. On the other hand, major differences between the two charged models arise for large values of $\alpha$. Indeed, in Model I the microgel approaches and crosses the VPT leading to a fully collapsed state, while in Model II it remains in a quasi-swollen configuration for all studied $\alpha$. Consequently, for high $\alpha$ values, the form factor does not resemble that of a homogeneous sphere, with only a second peak becoming evident, as opposed to the neutral case where many sharp peaks emerge. We notice that Model I fully coincides with the neutral case for very large $\alpha$, even for $f=0.2$. In Fig. 4, we compare the monomers density profiles for the two models as a function of $\alpha$. These data further indicate that, for Model II, the microgel does not achieve a collapsed state, as also visible from the behaviour of the profiles of charged monomers and of counterions, respectively. These are reported in the bottom panels of Fig. 4, showing that, for both models, the counterions are always found to be very close to the charged monomers, in order to neutralize the overall charge of the microgel. However, all profiles remain much more extended for Model II as compared to Model I, for all $\alpha$. We stress that the comparison is performed for microgels with different affinity of the charged monomers for the solvent at the same $\alpha$, which corresponds to very different swelling conditions, as evident from Fig. 2. Additional information can be extracted by comparing the two cases for a similar value of $R_{g}$, as reported in Fig. 5. Also in this case, we find that Model II displays a more slowly decaying radial profile, albeit having a very similar mass distribution with respect to Model I, which is due to the presence of more stretched external dangling chains. Similar results also apply to ions and counterions profiles, that are shown in the inset of Fig. 5: even at the same $R_{g}$, there is a surplus of charges at the surface in the case where the affinity of charged monomers for the solvent does not change with the effective temperature (Model II). Overall, these findings confirm an enhanced stabilization of the swollen configuration operated by the charged groups of the microgel, hindering the tendency of the remaining (neutral) monomers to collapse. To complete the structural analysis of the two models, it is instructive to consider the net charge density profile inside the microgels, that is reported in Fig. 6 for both $f=0.032$ and $f=0.2$. We confirm that, for both models, the net charge of the core region is roughly zero. However, it was shown in Ref. [33] that in the collapsed configuration a charged double layer arises at the surface of microgels, signalling the onset of a charge imbalance that grows with $\alpha$. This feature, that is clearly visible in the behaviour of Model I at high $\alpha$ for all values of $f$, is also present for Model II for the low charge case ($f=0.032$). However, the double peak in the net charge distribution is smeared out for $f=0.2$, due to the fact that, up to the largest explored values of $\alpha$, the microgel does not fully collapse. In this way, it maintains a low concentration of charged beads, that is always roughly balanced by counterions, resulting in a rather uniform charge profile. Instead, the peaks at the surface appear when the microgel collapses: this is indeed the case for both models at low charge fraction and even for large $f$ when charged monomers are assigned a solvophobic behaviour (Model I). We conclude from this analysis that the hydrophilicity of the charged monomers at all effective temperatures enhances the tendency of the microgel to remain swollen, even when most of the monomers experience a very large solvophobic attraction. Thanks to the charge neutralization operated by counterions, the microgel remains very stable in a rather swollen configuration up to very large $\alpha$, avoiding collapse for large enough values of $f$. This scenario agrees well with experimental observations, where the suppression of the VPT [26, 43, 42] is found when the concentration of charged hydrophilic groups in the polymer network is large enough. These considerations imply that Model I should not be used to describe microgels with high charge content. Indeed, its identical treatment of the solvophilic character of both neutral and charged monomers leads the particle to collapse at extremely high $\alpha$. Incidentally, we report that this was observed also for unrealistic values of $f$ up to $0.4$ (not shown), in evident contrast with experiments. We will thus rely on Model II in the future to correctly incorporate charge effects in modelling microgels in a realistic fashion. ### 3.2 Solvent effects We now go one step further in modelling ionic microgels, by explicitly adding the solvent to the simulations. This is a necessary prerequisite to tackle phenomena that cannot be described with an implicit solvent, e.g. situations in which hydrodynamics or surface tension effects at a liquid-liquid interface [34] play a fundamental role. In this subsection, we compare results for swelling behaviour and structural properties of the microgels for implicit and explicit solvent simulations. In particular, we restrict our discussion to Model II, having established this to be more in line with experimental observations. Since simulations with an explicit solvent require a much higher computational effort, we limit the following discussion to microgels with $N\approx 5000$. #### 3.2.1 Swelling curves and explicit-implicit ($a$-$\alpha$) mapping We start by reporting the swelling curves of charged microgels, stressing the point that they have been obtained by fixing the value of $a_{ch,s}$, which tunes the solvophilic properties of charged beads and counterions. We find that setting $a_{ch,s}=0$, while $a_{m,s}\equiv a$ varies, the explicit model is essentially equivalent to the implicit one. This means that it is possible to find a relation that links every implicit system with a certain value of the solvophobic attraction $\alpha$ to an explicit one with solvophobic parameter $a$ that shows the same structure and swelling properties. Figure 7: Implicit-explicit solvent mapping and swelling curves. (a) Mapping between $\alpha$ and $a$ obtained by comparing neutral microgels with implicit (Model II) and explicit solvents: the linear mapping is expressed by Eq. 12 and the numerical mapping via Eq. 13; (b-e) Normalized radius of gyration $R_{g}/R_{g,max}$ as a function of the swelling parameter $a$ for microgels with different charge content: (b) neutral, (c) $f=0.032$, (d) $f=0.1$ and (e) $f=0.2$, for explicit (full lines and filled diamonds) and implicit solvent conditions (rescaled along the horizontal axis using the linear mapping $a_{\text{lin}}(\alpha)$, dashed lines and empty squares, and using the numerical mapping $a_{\text{num}}(\alpha)$, full lines and filled squares). The present figure and the following ones refer to the same microgel topology with $N\approx 5000$. In order to establish such a $a$-$\alpha$ mapping, we explored two different routes. The first one, referred to as linear mapping in the following, is based on the assumption that the dependence of $a$ on $\alpha$ is linear, as previously adopted for neutral microgels [39]. In this way, the mapping relation is obtained through a horizontal rescaling of the relative swelling curves $R_{g}^{\text{imp}}(\alpha)/R_{g}^{\text{imp}}(\alpha=0)$ and $R_{g}^{\text{exp}}(a)/R_{g}^{\text{exp}}(a=0)$ for the neutral implicit and explicit microgels onto each other. Specifically, given two points for each curve, $(a_{1},a_{2})$ and ($\alpha_{1},\alpha_{2}$), the rescaled $x$-coordinate is calculated using the following relationship: $a_{lin}(\alpha)=\left(\alpha-\langle\alpha\rangle\right)\Delta a/\Delta\alpha+\langle a\rangle$ (12) where $\langle x\rangle=0.5(x_{1}+x_{2})$ and $\Delta x=x_{1}-x_{2}$ with $x=a,\alpha$. The second mapping $a_{\text{num}}(\alpha)$, referred to as numerical mapping, has been obtained by numerically inverting the equation $R_{g}^{\text{imp}}(\alpha)/R_{g}^{\text{imp}}(\alpha=0)=R_{g}^{\text{exp}}(a)/R_{g}^{\text{exp}}(a=0),$ (13) after spline fitting the two swelling curves. We report both mapping relations in Fig. 7(a), finding that they fall onto each other for almost the entire range of investigated solvophobic parameters in the two models, confirming the overall correctness of the assumption of linearity. However, we find some differences in the region $\alpha>1.0$ ($a>15$). Having established the mapping for neutral microgels, we now use it to directly remap also the results for ionic microgels for all studied $f$ without any further adjustments. #### 3.2.2 Swelling behaviour The normalized swelling curves with varying charge fraction $f$, comparing implicit and explicit solvent, are reported in Fig. 7(b-e). Data from implicit simulations are mapped via both methods described above. For the neutral case, the presence of the solvent does not affect the swelling behaviour, as shown in Fig. 7(b), where no appreciable differences are found between linear and numerical mapping even at high $\alpha$. Using the same relations for comparing charged microgels in explicit and implicit solvent, we find that, remarkably, the same swelling behaviour works for all charge contents. The swelling curves are virtually identical, which ensures that the inclusion of the solvent does not alter the microgel behavior in temperature even in the presence of charges. Small deviations, as expected from Fig. 7(a), appear only at large $\alpha$ values, being more pronounced for high charge content. This confirms the robustness of the DPD model which, as already discussed in Ref. [39], does not induce spurious effects, e.g. due to excluded volume, even in the collapsed state. An important result of this work is that, even in the presence of an explicit solvent, the microgel at high $f$ does not fully collapse at large $\alpha$, being entirely equivalent to implicit Model II and compatible with experimental findings. Figure 8: Density profiles. Density profiles of monomers (top row), charged beads and counterions (middle row) and net charge (bottom row) for ionic microgels with $f=0.1$ as a function of the distance from the microgel center of mass $r$ obtained in explicit and implicit solvent conditions. Curves from the explicit case refer to values of $a=5,11,12.3,14,16$, from the (left) swollen to the (right) collapsed state. Implicit and explicit solvent cases are compared at values of $\alpha$ approximately corresponding to each $a$ value according to both the linear ($\alpha=0,0.56,0.74,1.0,1.1$) and the numerical ($\alpha=0,0.56,0.74,1.0,1.2$) mapping. #### 3.2.3 Structural properties In this subsection, we will show that the implicit and explicit solvent treatments with the newly established numerical mapping (Eq. 13) lead to identical structural features of the microgels. Small differences arise when using the linear mapping (Eq. 12) at high $f$ and large values of $\alpha$. We show in Fig. 8 the monomer (top panels), ion and counterion (middle panels) and charge (bottom panels) density profiles only for the $f=0.1$ case, since similar results are also found for the other studied charge fractions. Reported data for different values of monomer-solvent interactions show an overall similarity between implicit and explicit solvent descriptions at all swelling conditions. Small deviations arise only for $f=0.2$ for states with the highest values of $a$ or $\alpha$, when using the linear mapping: as we can observe from the rightmost panels of Fig. 8, the linear mapping fails to associate implicit and explicit states in the most collapsed state, where a visible difference arises between the profiles. The distribution of ions and counterions within the microgel is an observable that should be more sensitive to the presence of the solvent. However, quite remarkably, also in this case, we find excellent agreement between the two models, as shown in the middle panels of Fig. 8. In particular, the emergence of a clear double-peak structure in the ion distribution is found in both models for large $\alpha$ (implicit) and $a$ (explicit), signalling an accumulation of ions at the exterior surface of the microgels. This can be understood from the fact that ions, remaining always hydrophilic, never completely collapse onto the core of the particle. Thus, the appearance of a peak at distances corresponding to the outer region of the microgel is the result of an attempt of ions to maximize their contact with solvent. This is preceded by a minimum, which indicates a region where ions are depleted within the particle. This feature is the echo of the minimum that arises in the net charge density distributions, already anticipated for the large microgel treated with the implicit model in Fig. 6. Importantly, a minimum also occurs in $\rho_{Q}(r)$ for smaller microgels, as shown in the bottom panels of Fig. 8, for the most collapsed conditions. Here a charged double layer is clearly present, with an excess of positive charges inside the microgel corona due to the increased amount of counterions in this region. At the same time, a negative charge surplus is found at the surface of the microgel, since charged ions preferably remain in contact with solvent particles. The net charge distribution is also identical for explicit and implicit solvent when using the numerical mapping, with again very small differences arising for the linear mapping at large $\alpha$. Figure 9: Solvent density profiles for charged microgels of different $f$ values, as a function of the distance from the microgel center of mass $r$. The different panels refer to $a=5,11,12,14,16$ from (left) good to (right) bad solvent conditions. It is important to notice that, although a double layer was also observed with the implicit solvent in Ref. [33] (equivalent to Model I), the two distributions (the one in Fig. 8 of the current manuscript and that reported in Fig. 6 of Ref. [33]) have opposite signs. Indeed in Ref. [33] the superposition of electrostatic and solvophobic effects led to an accumulation of counterions at the microgel surface, with the onset of a seemingly Donnan equilibrium [45]. Notwithstanding the different origin of the double layer, both models demonstrate that an almost perfect neutrality is achieved within the core of the microgel, and it is only at the surface that inhomogeneous distributions appear. Besides, the reduced size of the microgels studied with the explicit solvent facilitates the onset of peaks due to the increased surface-to-volume ratio of the microgels. A more precise assessment of size effects and a careful comparison to experiments will be the subject of future works. Finally, the explicit solvent model allows us to quantify the amount of solvent that is located inside the microgel as temperature increases. This is illustrated in Fig. 9, where the solvent density profile $\rho_{s}(r)$ is reported for different values of $a$ and all investigated charge fractions. These plots confirm the reduced tendency to collapse of charged microgels which retain a large amount of solvent within the network structure. No inhomogeneities within the microgel are in general observed. At large $f$ and $\alpha$ some oscillations arise which may be due to reduced statistics. Finally, this study confirms that even at temperatures above the VPT there is quite a residual amount of solvent within the microgel, that is significantly enhanced by increasing the charge. These findings are in line with expectations [43, 46], that are thus confirmed by our simulations. ## 4 Conclusions In this work we report an extensive numerical study of single microgel particles, a prototype of soft colloids that is of great interest for the colloidal community, particularly for the formation of arrested states with tunable rheological properties [9], including glasses [47, 48] and gels [49]. The use of different polymers within the microgel network allows to exploit responsiveness to different control parameters, such as temperature and pH, giving rise also to unusual responses in the fragility of the system [47, 50, 16]. In order to be able to model dense suspensions of these soft particles, we can rely on two possible strategies. On one hand, we can exploit highly coarse- grained models, such as the Hertzian one, which completely neglect the polymeric degrees of freedom of the particles and thus cannot reproduce the complex phenomenology observed in experiments in the gel or glassy regimes, such as shrinking, faceting and interpenetration [51, 15]. On the other hand, we can try to model a single microgel in a realistic way, aiming to reproduce its structural properties and, from this, to build effective interactions which retain the polymeric features of the single particle. Adopting the second strategy, the aim of this work is to improve the current numerical modelling of single ionic microgels with randomly distributed charged groups, aiming to describe PNIPAM-co-PAAc microgels across the Volume Phase Transition. In particular, we assess two different ways to model the interactions of the charged monomers belonging to the polymer network, either considering or not a solvophobic attraction that mimics their hydrophilic/hydrophobic interactions. We find that, as long as the charged groups maintain the same affinity for the solvent, the tendency of the microgel to remain in swollen conditions is enhanced even at high effective temperatures. Thus, for a charge fraction of $f=0.2$ we find no evidence of the collapse of the microgel within the investigated range of our simulations, in agreement with experimental observations that are currently available [41, 26, 43, 42]. This result is different from the case where charged beads also attract each other like neutral monomers upon increasing temperature, which undergoes a Volume Phase Transition to a fully collapsed state [33]. Despite this fundamental difference, the structural properties of the microgels treated with both models are rather similar, especially at low and intermediate temperatures. For instance, we confirm that the peculiar power- law regimes observed in the form factors are independent of the chosen model. Having established the most appropriate modelling for charged monomers, we then performed another necessary step in the modelling of ionic microgels, namely to explicitly consider the presence of the solvent, which may affect the rearrangement of the charges during the swelling-deswelling transition. To this aim, we build on previous results showing that for neutral microgels a description with an explicit solvent can be directly and quantitatively superimposed to the implicit modelling by using a DPD representation of the solvent, leaving unchanged the treatment of the polymer network with a bead- spring model. In this way, the solvophobic potential in Eq. 4, modulated by the parameter $\alpha$, is replaced by the DPD repulsive interactions between monomers and solvent. The latter is varied through a change of the parameter $a$ controlling the repulsion between non-charged beads, while the interaction between charged monomers always retains a solvophilic nature. We have thus carried out a careful comparison between explicit and implicit solvent treatments, finding quantitative agreement between the two. Interestingly, the relation among $a$ and $\alpha$ established by the comparison of neutral microgels can be used also to compare charged microgels, even with large values of $f$ (some deviations occur only at $f=0.2$ and large $\alpha,a$ values), where the same correspondence between implicit and explicit solvent states is retrieved. We showed that a linear mapping between the two control parameters of the interactions in the implicit and explicit case is sufficient to obtain a very good agreement between the two descriptions. From our analysis of the internal structure of the microgels across the VPT, we found that counterions have a rather similar distribution within the microgel core, effectively neutralizing the internal charge at small distances, but being in excess close to the surface. This gives rise to a charged double layer for large values of $a$ and $\alpha$. Interestingly, such peaks in the charge density distributions are swapped with respect to the case of Model I, where ions do not experience a tendency to remain at the surface, since they are also treated as solvophobic. These detailed predictions will have to be compared to experiments on ionic microgels as a function of charge fraction, pH and $T$, in order to establish the limit of validity of our model and to further improve it, towards a more realistic description of experimental microgels. In perspective, this work paves the way to study realistic charged microgels in diffusing conditions, such as in electrophoresis and thermophoresis experiments [52], or at liquid-liquid interfaces and to calculate their effective interactions, similarly to what has been done for neutral microgels [34, 35]. In this way, we will be able to determine the conditions under which electrostatic effects play a dominant role over elastic ones. Another important line of research will be the assessment of the role of the network topology: examples of interesting cases whose properties could be investigated are microgels consisting of two interpenetrated networks [50, 53] or ultra-low crosslinked [54, 55] and hollow [56, 57] microgels. Finally, we hope that our theoretical efforts will stimulate further experimental activity on charged microgels to verify the predicted behaviour so that it will be possible to tackle the investigation of dense suspensions in the near future. ## Acknowledgments This research has been performed within the PhD program in ”Mathematical Models for Engineering, Electromagnetics and Nanosciences”. We acknowledge financial support from the European Research Council (ERC Consolidator Grant 681597, MIMIC). FC and EZ also acknowledge funding from Regione Lazio, through L.R. 13/08 (Progetto Gruppo di Ricerca GELARTE, n.prot.85-2017-15290). ## References ## References * [1] Sciortino F and Tartaglia P 2005 Advances in Physics 54 471–524 * [2] Pusey P 2008 Journal of Physics: Condensed Matter 20 494202 * [3] Zaccarelli E 2007 Journal of Physics: Condensed Matter 19 323101 * [4] Lu P J and Weitz D A 2013 Annu. Rev. Condens. Matter Phys. 4 217–233 * [5] Joshi Y M 2014 Annu. Rev. Chem. Biomol. Eng 5 181–202 * [6] Likos C N 2001 Physics Reports 348 267–439 * [7] Fernandez-Nieves A, Wyss H, Mattsson J and Weitz D A (eds) 2011 Microgel suspensions: fundamentals and applications (New York, New York, USA: John Wiley & Sons) * [8] Lyon L A and Fernandez-Nieves A 2012 Annual review of physical chemistry 63 25–43 * [9] Vlassopoulos D and Cloitre M 2014 Current opinion in colloid & interface science 19 561–574 * [10] Oh J K, Drumright R, Siegwart D J and Matyjaszewski K 2008 Progress in Polymer Science 33 448–477 * [11] Karg M, Pich A, Hellweg T, Hoare T, Lyon L A, Crassous J J, Suzuki D, Gumerov R A, Schneider S and Potemkin I I 2019 Langmuir 35 6231–6255 * [12] Di Napoli B, Franco S, Severini L, Tumiati M, Buratti E, Titubante M, Nigro V, Gnan N, Micheli L, Ruzicka B et al. 2020 ACS Applied Polymer Materials 2 2791–2801 * [13] Rovigatti L, Gnan N, Ninarello A and Zaccarelli E 2019 Macromolecules 52 4895–4906 * [14] Bergman M J, Gnan N, Obiols-Rabasa M, Meijer J M, Rovigatti L, Zaccarelli E and Schurtenberger P 2018 Nature communications 9 1–11 * [15] Conley G M, Zhang C, Aebischer P, Harden J L and Scheffold F 2019 Nature communications 10 1–8 * [16] Gnan N and Zaccarelli E 2019 Nature Physics 15 683–688 * [17] Stieger M, Richtering W, Pedersen J S and Lindner P 2004 The Journal of chemical physics 120 6197–6206 * [18] Ninarello A, Crassous J J, Paloli D, Camerin F, Gnan N, Rovigatti L, Schurtenberger P and Zaccarelli E 2019 Macromolecules 52 7584–7592 * [19] Bergman M J, Pedersen J S, Schurtenberger P and Boon N 2020 Soft Matter 16 2786–2794 * [20] Pelton R and Hoare T 2011 Microgels and their synthesis: An introduction pp 1–32 in Fernandez-Nieves et al. [7] * [21] Nöjd S, Mohanty P S, Bagheri P, Yethiraj A and Schurtenberger P 2013 Soft Matter 9 9199–9207 * [22] Colla T, Mohanty P S, Nojd S, Bialik E, Riede A, Schurtenberger P and Likos C N 2018 ACS nano 12 4321–4337 * [23] Nigro V, Angelini R, Bertoldo M, Castelvetro V, Ruocco G and Ruzicka B 2015 Journal of Non-Crystalline Solids 407 361–366 * [24] Nöjd S, Holmqvist P, Boon N, Obiols-Rabasa M, Mohanty P S, Schweins R and Schurtenberger P 2018 Soft Matter 14 4150–4159 * [25] Rochette D, Kent B, Habicht A and Seiffert S 2017 Colloid and Polymer Science 295 507–520 * [26] Capriles-González D, Sierra-Martín B, Fernández-Nieves A and Fernández-Barbero A 2008 The Journal of Physical Chemistry B 112 12195–12200 * [27] Gnan N, Rovigatti L, Bergman M and Zaccarelli E 2017 Macromolecules 50 8777–8786 * [28] Moreno A J and Verso F L 2018 Soft Matter 14 7083–7096 * [29] Rovigatti L, Gnan N, Tavagnacco L, Moreno A J and Zaccarelli E 2019 Soft matter 15 1108–1119 * [30] Quesada-Pérez M, Ramos J, Forcada J and Martín-Molina A 2012 The Journal of chemical physics 136 244903 * [31] Kobayashi H and Winkler R G 2014 Polymers 6 1602–1617 * [32] Martín-Molina A and Quesada-Pérez M 2019 Journal of Molecular Liquids 280 374–381 * [33] Del Monte G, Ninarello A, Camerin F, Rovigatti L, Gnan N and Zaccarelli E 2019 Soft matter 15 8113–8128 * [34] Camerin F, Fernandez-Rodriguez M A, Rovigatti L, Antonopoulou M N, Gnan N, Ninarello A, Isa L and Zaccarelli E 2019 ACS nano 13 4548–4559 * [35] Camerin F, Gnan N, Ruiz-Franco J, Ninarello A, Rovigatti L and Zaccarelli E 2020 Physical Review X 10 031012 * [36] Kremer K and Grest G S 1990 The Journal of Chemical Physics 92 5057–5086 * [37] Deserno M and Holm C 1998 The Journal of chemical physics 109 7678–7693 * [38] Soddemann T, Dünweg B and Kremer K 2001 The European Physical Journal E 6 409–419 * [39] Camerin F, Gnan N, Rovigatti L and Zaccarelli E 2018 Scientific reports 8 1–12 * [40] Plimpton S 1995 Journal of computational physics 117 1–19 * [41] Holmqvist P, Mohanty P, Nägele G, Schurtenberger P and Heinen M 2012 Physical review letters 109 048302 * [42] Wiehemeier L, Brändel T, Hannappel Y, Kottke T and Hellweg T 2019 Soft matter 15 5673–5684 * [43] Brändel T, Wiehemeier L, Kottke T and Hellweg T 2017 Polymer 125 110–116 * [44] Fernandez-Barbero A, Fernandez-Nieves A, Grillo I and Lopez-Cabarcos E 2002 Physical Review E 66 051803 * [45] Hunter R J 2001 Foundations of colloid science (Oxford, New York, USA: Oxford University Press) * [46] Bischofberger I and Trappe V 2015 Scientific reports 5 15520 * [47] Mattsson J, Wyss H M, Fernandez-Nieves A, Miyazaki K, Hu Z, Reichman D R and Weitz D A 2009 Nature 462 83–86 * [48] Philippe A M, Truzzolillo D, Galvan-Myoshi J, Dieudonné-George P, Trappe V, Berthier L and Cipelletti L 2018 Physical Review E 97 040601 * [49] Wu J, Huang G and Hu Z 2003 Macromolecules 36 440–448 * [50] Nigro V, Angelini R, Bertoldo M, Bruni F, Ricci M A and Ruzicka B 2017 Soft matter 13 5185–5193 * [51] Conley G M, Aebischer P, Nöjd S, Schurtenberger P and Scheffold F 2017 Science advances 3 e1700969 * [52] Wongsuwarn S, Vigolo D, Cerbino R, Howe A M, Vailati A, Piazza R and Cicuta P 2012 Soft Matter 8 5857–5863 * [53] Nigro V, Angelini R, Rosi B, Bertoldo M, Buratti E, Casciardi S, Sennato S and Ruzicka B 2019 Journal of colloid and interface science 545 210–219 * [54] Bachman H, Brown A C, Clarke K C, Dhada K S, Douglas A, Hansen C E, Herman E, Hyatt J S, Kodlekere P, Meng Z et al. 2015 Soft Matter 11 2018–2028 * [55] Scotti A, Brugnoni M, Lopez C G, Bochenek S, Crassous J J and Richtering W 2020 Soft Matter 16 668–678 * [56] Nayak S, Gan D, Serpe M J and Lyon L A 2005 Small 1 416–421 * [57] Nickel A C, Scotti A, Houston J E, Ito T, Crassous J, Pedersen J S and Richtering W 2019 Nano letters 19 8161–8170
# OffCon3: What is State-of-the-Art Anyway? Philip J. Ball Department of Engineering Science University of Oxford Oxford, UK <EMAIL_ADDRESS> Stephen J. Roberts Department of Engineering Science University of Oxford Oxford, UK <EMAIL_ADDRESS> ###### Abstract Two popular approaches to model-free continuous control tasks are SAC and TD3. At first glance these approaches seem rather different; SAC aims to solve the entropy-augmented MDP by minimising the KL-divergence between a stochastic proposal policy and a hypotheical energy-basd soft Q-function policy, whereas TD3 is derived from DPG, which uses a deterministic policy to perform policy gradient ascent along the value function. In reality, both approaches are remarkably similar, and belong to a family of approaches we call ‘Off-Policy Continuous Generalized Policy Iteration’. This illuminates their similar performance in most continuous control benchmarks, and indeed when hyperparameters are matched, their performance can be statistically indistinguishable. To further remove any difference due to implementation, we provide OffCon3 (_Off_ -Policy _Con_ tinuous _Con_ trol: _Con_ solidated), a code base featuring state-of-the-art versions of both algorithms. ## 1 Introduction State-of-the-art performance in model-free continuous control reinforcement learning (RL) has been dominated by off-policy maximum-entropy/soft-policy based methods, namely Soft Actor Critic [1, 2]. This is evidenced by the plethora of literature, both model-free and model-based, that chooses SAC as the standard [3, 4, 5, 6, 7], often showing it as the best performing model- free approach. ### 1.1 Deterministic Policy Gradients At this point we introduce the notion of deterministic policy gradients (DPGs) for off-policy reinforcement learning. This is in contrast to stochastic policy gradients (SPGs) that rely on a stochastic policy for gradient estimation. It can be shown that DPG is simply a limiting case of SPG [8], and intuitively the key difference between them focuses on how they each rely on samples for estimating gradients. For both approaches the policy gradient proof is required. For details see [8, 9], but through changing the order of integration and/or differentiation we can remove the reliance of the derivative on having access the underlying state distribution. As aforementioned, DPG is a limiting case of SPG, specifically when the variance parameter of the SPG policy tends to $0$ (i.e., $\sigma\rightarrow 0$). However the similarities and differences between these two methods are nuanced and merit further investigation. This is under-explored and often incorrect equivalences are drawn (i.e., DPG necessitates off-policy learning, SPG necessitates on-policy learning). We start by presenting a simple explanation as to why DPG facilitates off- policy learning: $\displaystyle Q(s,a)$ $\displaystyle=\mathbb{E}_{r,s^{\prime}\sim E}\left[r_{t}+\gamma Q(s^{\prime},\mu(s^{\prime}))\right].$ (1) Observing Eq 1, we note that the expectation is only dependent on the environment itself, and not the policy. Therefore, all we need to train the Q-function is environment samples (i.e., tuples $(s,a,r,s_{t+1})$ from a replay buffer), and the deterministic policy $\pi$. We are now in a position to write down the objectives we wish to maximize for both the critic and the actor. For the critic, we use a standard Bellman update, and for the actor, we maximize the expected return under a Q-function: $\displaystyle J_{Q}$ $\displaystyle=\mathbb{E}_{s,a,r,s^{\prime}\sim E}\left[\left(Q(s,a)-\left(r+\gamma Q(s^{\prime},a^{\prime})|_{a^{\prime}=\pi(s^{\prime})}\right)\right)^{2}\right]$ (2) For the actor $\pi$, we wish to maximize the expected return: $\displaystyle J_{\pi}$ $\displaystyle=\mathbb{E}_{s\sim E}\left[V(s)\right]$ (3) $\displaystyle=\mathbb{E}_{s\sim E}\left[Q(s,a)|_{a=\pi(s)}\right]$ (4) We can now write out the update steps required for DPG-style algorithms, or more specifically DDPG [10] considering the use of neural networks. This will facilitate the comparisons to SAC later on. We now denote the neural network weights of the Q-function and policy as $\theta$ and $\phi$ respectively, with $Q_{\theta}$ and $\pi(\cdot)=f_{\phi}(\cdot)$ respectively. We define the policy as a deterministic function for now, as it will make it clearer later when we start defining policy $\pi$ as a distribution that is dependent on a deterministic function. We now write the critic and actor update rules: Critic: $\displaystyle\nabla_{\theta}J_{Q}$ $\displaystyle\approx\nabla_{\theta}\mathbb{E}_{r,s,s^{\prime}\sim E}\left[\left(Q_{\theta}(s,a)-\left(r+\gamma Q_{\theta}(s^{\prime},a^{\prime})|_{a^{\prime}=f_{\phi}(s^{\prime})}\right)\right)^{2}\right]\text{}.$ (5) Actor: $\displaystyle\nabla_{\phi}J_{\pi}$ $\displaystyle\approx\mathbb{E}_{s\sim E}\left[\nabla_{a}Q_{\theta}(s,a)|_{a=f_{\phi}(s)}\nabla_{\phi}f_{\phi}(s)\right].$ (6) Note that, due to the chain rule, the gradient through the actor requires a Q-value estimator that is differentiable with respect to actions. Finally, we observe that in the generalized policy iteration analogues of Actor-Critic with dynamic programming ([9] Chapter 4); critic training is analogous to policy evaluation, and actor training is analogous to policy improvement. ### 1.2 Stochastic Value Gradients Here we discuss Stochastic Value Gradients [11], an algorithm that introduces the idea of taking a gradient through a learned model of the environment and associated stochastic policy. We specifically focus on the model-free limiting case of this approach, SVG(0). This is of particular interest as it represents a stepping stone between DPG and the maximum entropy methods introduced later. We observe that, unlike the other versions of SVG, SVG(0) specifically uses a Q-function to estimate expected return (as opposed to a state-value function). Therefore we must derive the stochastic Bellman equation in the form of the Q-function, following a similar approach to [11]: $\displaystyle Q^{\pi}(s_{t},a_{t})$ $\displaystyle=\mathbb{E}\left[\gamma^{\tau-t}r^{\tau}|s=s_{t},a=a_{t}\right]$ (7) $\displaystyle=\int r_{t}p(r_{t}|s_{t},a_{t})+\gamma\left[\int\int Q^{\pi}(s_{t+1},a_{t+1})\pi(a_{t+1}|s_{t+1})p(s_{t+1}|s_{t},a_{t})\differential{a_{t+1}}\differential{s_{t+1}}\right]\differential{r_{t}}$ (8) $\displaystyle=\mathbb{E}_{r_{t},s_{t+1}\sim E}\left[r_{t}+\gamma\mathbb{E}_{a_{t+1}\sim\pi}\left[Q^{\pi}(s_{t+1},a_{t+1})\right]\right].$ (9) Observe how Eq 9 is just Eq 1 except with a stochastic policy $a\sim\pi(\cdot)$. To make its derivative tractable, we treat the policy as a spherical Gaussian, and amortize its parameter ($\mu,\Sigma$) inference using a neural network with weights $\theta$. This allows the use of the reparameterization/pathwise derivative trick [12, 13]. This means $a\sim\pi(s,\eta;\theta)$ where $\eta\sim\mathcal{N}(0,I)$. As a result, we move policy action sampling outside (i.e., we sample from both the environment $E$ and a $\mathcal{N}(0,I)$), and can backpropagate through the policy weights: $\displaystyle Q^{\pi}(s_{t},a_{t})$ $\displaystyle=\mathbb{E}_{r_{t},s_{t+1}\sim E}\left[r_{t}+\gamma\mathbb{E}_{\eta\sim\mathcal{N}(0,I)}\left[Q^{\pi}(s_{t+1},\pi(s_{t+1},\eta;\theta))\right]\right].$ (10) $\displaystyle=\mathbb{E}_{r_{t},s_{t+1}\sim E,\eta\sim\mathcal{N}(0,I)}\left[r_{t}+\gamma\left[Q^{\pi}(s_{t+1},\pi(s_{t+1},\eta;\theta))\right]\right].$ (11) This means we can write the derivative of the Actor and Critic as follows: Critic: $\displaystyle\nabla_{\theta^{Q}}J$ $\displaystyle\approx\nabla_{\theta^{Q}}\mathbb{E}_{r,s,s^{\prime}\sim\rho^{\pi},\eta\sim\mathcal{N}(0,I)}\left[\left(Q(s,a;\theta^{Q})-\left(r+\gamma Q(s^{\prime},a^{\prime};\theta^{Q^{\prime}})|_{a^{\prime}=\pi(s^{\prime},\eta)}\right)\right)^{2}\right].$ (12) Actor: $\displaystyle\nabla_{\theta^{\pi}}J$ $\displaystyle\approx\mathbb{E}_{s\sim\rho^{\pi},\eta\sim\mathcal{N}(0,I)}\left[\nabla_{a}Q(s,a;\theta^{Q})|_{a=\pi(s,\eta)}\nabla_{\theta^{\pi}}\pi(s,\eta;{\theta^{\pi}})\right].$ (14) Observe how similar this is to the DPG-style gradients; note that when determining actions an additional sample from a Gaussian distribution is all that is necessary. Furthermore, we observe that this is still an off-policy algorithm, with no dependency on the policy that gave rise to the trajectory samples. ### 1.3 Soft Actor Critic SAC is an actor-critic method which aims to learn policy that maximizes both return and entropy over each visited state in a trajectory [14]: $\displaystyle\pi^{*}={\arg\max}_{\pi}\sum_{t}\mathbb{E}_{(\mathbf{s}_{t},\mathbf{a}_{t})\sim\rho_{\pi}}\left[r(\mathbf{s}_{t},\mathbf{a}_{t})\color[rgb]{0,0.88,0}+\alpha\mathcal{H}(\pi(\cdot|\mathbf{s}_{t}))\right]$ (15) where the part of the equation in green describes the additional entropy objective (N.B.: the conventional objective is therefore recovered as $\alpha\rightarrow 0$). This is done using soft-policy iteration, and involves repeatedly applying the following entropy Bellman operator [1]: $\displaystyle\mathcal{T}^{\pi}Q(s,a)=$ $\displaystyle r(s,a)+\gamma\mathbb{E}_{s^{\prime}\sim p}[V(s^{\prime})]$ (16) where $\displaystyle V(s)=$ $\displaystyle\mathbb{E}_{a\sim\pi}[Q(s,a)-\alpha\log\pi(a|s)].$ (17) For consistency of presentation, we present this as a (soft) Bellman update: $\displaystyle Q^{\pi}(s,a)$ $\displaystyle=\int rp(r|s,a)+\gamma\left[\iint(Q^{\pi}(s^{\prime},a^{\prime})-\alpha\log\pi(a^{\prime}|s^{\prime}))\pi(a^{\prime}|s^{\prime})p(s^{\prime}|s,a)\differential{a^{\prime}}\differential{s^{\prime}}\right]\differential{r}$ $\displaystyle=\mathbb{E}_{r,s^{\prime}\sim E}\left[r+\gamma\mathbb{E}_{a^{\prime}\sim\pi}\left[Q^{\pi}(s^{\prime},a^{\prime})-\alpha\log\pi(a^{\prime}|s^{\prime})\right]\right]$ $\displaystyle=\mathbb{E}_{r,s^{\prime}\sim E,\eta\sim\mathcal{N}(0,I)}\left[r+\gamma\left[Q^{\pi}(s^{\prime},\pi(s^{\prime},\eta;\theta))-\alpha\log\pi(s^{\prime},\eta;\theta)\right]\right]$ (18) where in the last line we make the same assumption about amortizing the policy distribution as in SVG. At this point we can directly write down the objective of the actor/policy, namely to maximize expected return _and_ entropy, i.e., Eq 17. This follows the method for determining the objective function for the policy gradient in DPG (Eq 4): $\displaystyle J_{\pi}$ $\displaystyle=\mathbb{E}_{s\sim\rho^{\mu}}\left[V(s)\right]$ (19) $\displaystyle=\mathbb{E}_{s\sim\rho^{\mu}}\left[\mathbb{E}_{a\sim\pi}\left[Q(s,a;\theta^{Q})-\alpha\log\pi(a|s)\right]\right]$ (20) Similarly for the critic $Q$, we have $J_{Q}$: $\displaystyle J_{Q}$ $\displaystyle=\mathbb{E}_{r,s,s^{\prime}\sim\rho^{\mu},a^{\prime}\sim\pi}\left[\left(Q(s,a;\theta^{Q})-\left(r+\gamma(Q(s^{\prime},a^{\prime};\theta^{Q^{\prime}})-\alpha\log\pi(a^{\prime}|s^{\prime}))\right)\right)^{2}\right]$ (21) The ‘soft’ critic gradient is similar to the DPG style update as the Q-value parameters don’t depend on the additional entropy term, however the actor gradient requires both the chain rule and the law of total derivatives. Here we write down the gradients directly: Critic: $\displaystyle\nabla_{\theta^{Q}}J$ $\displaystyle\approx\nabla_{\theta^{Q}}\mathbb{E}_{r,s,s^{\prime}\sim\rho^{\pi},\eta\sim\mathcal{N}(0,I)}\left[\left(Q(s,a;\theta^{Q})-\left(r+\gamma(Q(s^{\prime},a^{\prime};\theta^{Q^{\prime}})-\alpha\log\pi(a^{\prime}|s^{\prime}))|_{a^{\prime}=\pi(s^{\prime},\eta)}\right)\right)^{2}\right].$ (22) Actor: $\displaystyle\nabla_{\theta^{\pi}}J$ $\displaystyle\approx\mathbb{E}_{s\sim\rho^{\pi},\eta\sim\mathcal{N}(0,I)}\left[\left(-\nabla_{\theta^{\pi}}\log\pi(a|s)+\nabla_{a}\left(Q(s,a;\theta^{Q})-\alpha\log\pi(s,\eta;\theta^{\pi})\right)\right)|_{a=\pi(s,\eta)}\nabla_{\theta^{\pi}}\pi(s,\eta;{\theta^{\pi}})\right].$ (23) What remains to be optimized is the temperature $\alpha$, which balances the entropy/reward trade-off in Eq 15. In [2] the authors learn this during training using an approximation to constrained optimization, where the mean trajectory entropy $\mathcal{H}$ is the constraint. ### 1.4 DPG $\rightarrow$ SVG $\rightarrow$ SAC Having outlined DPG, SVG(0), and SAC we are now in a position to directly compare all three approaches. We do this by observing the Critic and Actor objectives, highlighting in different colors the components that are attributable to each: $\displaystyle J_{\pi}$ $\displaystyle=\mathbb{E}_{s\sim E,\color[rgb]{1,0,1}\eta\sim\mathcal{N}(0,I)}\left[Q_{\theta}(s,a)\color[rgb]{0,0.88,0}-\alpha\log\pi(a|s)|_{a=f_{\phi}(s,\color[rgb]{1,0,1}\eta)}\right]$ (24) $\displaystyle J_{Q}$ $\displaystyle=\mathbb{E}_{r,s,s^{\prime}\sim E,\color[rgb]{1,0,1}\eta\sim\mathcal{N}(0,I)}\left[\left(Q_{\theta}(s,a)-\left(r+\gamma(Q_{\theta}(s^{\prime},a^{\prime})\color[rgb]{0,0.88,0}-\alpha\log\pi(a^{\prime}|s^{\prime}))\right)\right)^{2}|_{a^{\prime}=f_{\phi}(s^{\prime},\color[rgb]{1,0,1}\eta)}\right]$ (25) where terms in pink are introduced by SVG(0), and terms in green are introduced by SAC. Here we describe the natural progression of DPG to SAC: 1. 1. DPG introduces the policy iteration framework, including the deterministic policy gradient, that allows the learning of policies through Q-learning over continuous action spaces. DDPG introduces heuristics that allow the use of neural network function approximators. 2. 2. SVG introduces the idea of stochastic policies, and its limiting model-free case SVG(0) allows the learning of stochastic policies in the Q-learning policy improvement framework proposed in DPG. This uses the pathwise derivative through the amortized Gaussian policy. 3. 3. SAC leverages the policy variance learning in amortized inference by ensuring a maximum-entropy action distribution for any given state through the addition of an entropy term into the traditional maximum return objective. We observe therefore that all three algorithms can be considered as belonging to the same family, namely ‘Off-Policy Continuous Generalized Policy Iteration’, where the policy evaluation step represents a gradient step along $J_{Q}$, and policy improvement a gradient step along $J_{\pi}$. All that distinguishes these approaches is whether the actor is deterministic, and whether there is an additional entropy objective. We note that the SAC policy has been derived using standard gradient ascent of the value function (as in [8]), and similarly the DPG policy gradient can be derived as a KL- minimization (as in [1]). ## 2 Practical Reinforcement Learning Two methods derived from the aforementioned approaches have emerged as being most popular, namely SAC with entropy adjustment [2] and the DDPG derived TD3 [15]. At first glance, it may appear coincidental that both approaches have achieved similar levels of success in continuous control tasks, such as OpenAI Gym [16], but the above analysis shows that they are closely related. We briefly explain the merits of TD3, and understand how this has influenced SAC. ### 2.1 TD3 TD3 [15] is based on DDPG, and introduces several heuristics to improve upon it. These include: * • Training two Q-functions, then taking their minimum when evaluating to address Q-function overestimation bias. * • Update target parameters and actor parameters less frequently than critic updates. * • Add noise to the target policy action during critic training, making it harder for the actor to directly exploit the critic. The original SAC paper [1] does not train two Q-functions, and instead trains a Q-function and a state-value (V) function. Furthermore the trade-off between entropy and reward is fixed. The ‘applied’ SAC paper [2] removes the state- value function, and instead trains two Q-functions similar to TD3, and automatically adjusts the temperature trade-off to ensure some expected policy entropy (a function of action dimension). Interestingly, in their original papers, TD3 and SAC claim to outperform each other, and it would appear that the incorporation of the TD3-style Q-learning and temperature adjustment results in the ultimately better performance in the ‘applied’ SAC paper. However there are still key differences between SAC and TD3 training, namely heuristics such as network architecture, learning rate, and batch size. For the purposes of fair comparison, we choose these to be the same across both SAC and TD3, as shown in Table 1222Note we include an additional hidden layer, see Appendix C for details. | Algorithm ---|--- Hyperparamater | TD3 | SAC Collection Steps | 1,000 Random Action Steps | 10,000 Network Hidden Layers | $256:256:256$ Learning Rate | $3\text{\times}{10}^{-4}$ Optimizer | $Adam$ Replay Buffer Size | $1\text{\times}{10}^{6}$ Action Limit | $[-1,1]$ Exponential Moving Averaging Parameter | $5\text{\times}{10}^{-3}$ (Critic Update:Environment Step) Ratio | 1 (Policy Update:Environment Step) Ratio | 2 | 1 Has Target Policy? | Yes | No Expected Entropy Target | N/A | $-\text{dim}(\mathcal{A})$ Policy Log-Variance Limits | N/A | [-20, 2] Target Policy $\sigma$ | 0.2 | N/A Target Policy Clip Range | [-0.5, 0.5] | N/A Rollout Policy $\sigma$ | 0.1 | N/A Table 1: Hyperparameters used in OffCon3 ### 2.2 What is the effect of Gaussian exploration? One difference between TD3 and DDPG is the noise injection applied during Q-value function training. This turns a previously deterministic mapping $a=\mu(s)$ into a stochastic one ($a\sim\mu(s)+\text{clip}(\mathcal{N}(0,I)\times 0.2,-0.5,0.5)$). This means that the policies used in both data collection and critic training are in fact stochastic, making TD3 closer to SAC. We may ask how this compares to a deterministic objective; evidently, veering from the mean action selected by the deterministic actor should reduce expected return, so what does this stochasticity provide? To explore this question, we split our analysis into two sections: the effect on the Critic, and the effect on the Actor. #### Effect on Critic: We simplify analysis by assuming all added noise is diagonal Gaussian333Action clipping technically makes this assumption untrue, but in reality policies are still very nearly Gaussian, and write the deterministic objective as $J_{D}$. We also assume 1-D actions without loss of generality. Performing a Taylor series expansion of the stochastic policy, we find that the objective maximized by this Gaussian actor ($J_{R}$) as (see Appendix A for proof): $\displaystyle J_{R}$ $\displaystyle\approx J_{D}+\frac{\sigma^{2}}{2}\mathbb{E}_{s_{t}\sim E}\left[\nabla^{2}_{a}Q(s_{t},a)|_{a=\mu(s_{t})}\right]$ (26) This is the deterministic objective with an additional term proportional to the fixed variance of the policy, as well as the 2nd derivative (Hessian for multi-dimensional actions) of the critic with respect to actions. Unpacking this latter term, noting that the following residual between the stochastic ($J_{R}$) and deterministic ($J_{D}$) objectives, leads to: $\displaystyle J_{R}-J_{D}\approx\frac{\sigma^{2}}{2}\mathbb{E}_{s_{t}\sim E}\left[\nabla^{2}_{a}Q(s_{t},a)|_{a=\mu(s_{t})}\right]$ (27) First, let us consider a well trained policy that is able to produce actions that maximize the critic $Q$ for all states. This means that the value of the 2nd order term must be negative (equivalently, the Hessian must be negative semi-definite). Evidently, any non-zero $\sigma^{2}$ will result in the stochastic return $J_{D}$ being lower than $J_{R}$. This implies that the stochastic policy objective $J_{R}$ can only ever realistically lower bound the deterministic objective $J_{D}$. However in Gaussian exploration we fix this $\sigma^{2}$ to be non-zero, therefore the only degree of freedom is in the second-order term itself. Evidently we want a policy that maximizes $Q$ (i.e., 0th order term), therefore making this term positive is not viable. However the magnitude of the second-order term can be reduced by making $Q$ ‘smoother’. Since $Q$ is twice differentiable w.r.t. $a$, we can invoke the identity $\nabla^{2}_{a}Q(s,a)\preceq\beta\text{I}$ [17], implying that the largest eigenvalue of the Hessian of $Q$ is smaller than $\beta$, where $\beta$ is the Lipschitz constant of $Q$. Therefore, to minimize the magnitude of $\nabla^{2}_{a}Q(s,a)$, we must learn a $Q$ that is smoother with respect to actions. This can be viewed as a spectral norm regularizer of the $Q$ function [18], and the mechanism that is used in [15] to ensure the stability of the critic can be viewed as approximating this approach. It must be noted that we get this smoothing behavior by default in SAC as the learned policy has non- zero variance due to its entropy objective. #### Effect on Actor: The forced non-zero $\sigma^{2}$ variance term also has implications to entropy based approaches. SAC learns a non-zero variance to ensure some minimum entropy per trajectory timestep: $\displaystyle\max_{\pi}\left[Q^{\pi}(s_{0},a_{0})\right]\quad\text{s.t.}\quad\mathbb{E}_{s_{t}\sim E}[\mathbb{E}_{a_{t}\sim\pi}[-\log\pi(a_{t}|s_{t})]]>\mathcal{H}\quad\forall t$ (28) where, for a Gaussian policy, we can write $\displaystyle\max_{\pi}\left[Q^{\pi}(s_{0},a_{0})\right]\quad\text{s.t.}\quad\mathbb{E}_{s_{t}\sim E}\left[\log(\sigma(s_{t})\sqrt{2\pi e})\right]>\mathcal{H}\quad\forall t.$ (29) This optimization is non-trivial for SAC (see Sec. 5 in [2]) as the amount of policy variance is learned by the policy (hence the $s_{t}$ dependency above). However for a policy with fixed variance $\sigma$, the optimization becomes trivial as we can drop the state dependency, and simply maximize policy over the critic444Consider that for a fixed $\sigma$ we can always find an $\mathcal{H}$ such that the constraint inequality evaluates to an exact equality for all policies, therefore the dual simply collapses into a maximization of the primal without a constraint. (which is done in standard actor training): $\displaystyle\max_{\pi}\left[Q^{\pi}(s_{0},a_{0})\right]\quad\text{s.t.}\quad\log(\sigma\sqrt{2\pi e})>\mathcal{H}\quad\forall t.$ (30) In the case of a deployed policy which has standard deviation $0.1$, such as in TD3, we can view this performing exploration with a maximum entropy policy that ensures a policy entropy of $\approx-0.884$ nats. ### 2.3 Why not SVG(0)? In principle, SVG(0) appears to be a strong candidate for policy training; it is stochastic, does not incorporate an entropy term, which adds computation and hyperparameters. In reality, since the environments evaluated are deterministic, the variance head of the SVG(0) policy tends to 0 very quickly. The reason for this is outlined in Sec. 2.2. As a consequence, the resultant algorithm is effectively DDPG. Indeed this is supported in the analysis performed in Appendix A, where only the 0th order Thompson term remains for relatively small $\sigma$. We illustrate this effect in Figure 1, and include this algorithm for illustrative purposes as ‘TDS’ in OffCon3. Figure 1: Variance of rollout policies on HalfCheetah ## 3 Experiments For these experiments, we run both algorithms for 5 seeds on 4 different MuJoCo environments: HalfCheetah, Hopper, Walker2d, and Ant. We then perform a two-tailed Welch’s $t$-test [19] to determine whether final performance is statically significantly different. Observing Table 2, Ant and Walker2d performance is statistically indistinguishable (although TD3 learns quicker in Ant, see Appendix B); in HalfCheetah SAC does convincingly outperform TD3, but in Hopper, the opposite is true. (a) HalfCheetah (b) Hopper (c) Ant (d) Walker2d (e) Humanoid Figure 2: SAC and TD3 Training Curves on MuJoCo Environments. ### 3.1 Authors’ Results For completeness we compare the results from each algorithm’s available code with ours to ensure our implementation does not unfairly penalize either approach. The results are shown in Tables 3 and 4555Authors’ SAC and TD3 code is here and here respectively. We tabulate the results provided in these repos. Results are the max (averaged over seeds) performance up to and including the Timesteps column.. We note that our implementation appears to generally match, or exceeds, the authors’ code. Note that unlike the original code in [15], we do not discard ‘failure’ seeds666See discussion here.; this may explain why our implementation doesn’t always outperform the author’s code, especially on less stable environments (such as Hopper and Walker2d). | $t$-Test Result ---|--- Environment | $t$ | $p$ HalfCheetah | 4.29 | 0.00927 Hopper | -2.92 | 0.0293 Ant | -0.481 | 0.653 Walker2d | 1.59 | 0.155 Humanoid | 1.29 | 0.29 Table 2: Two-tailed Welch’s $t$-test results | SAC Return | ---|---|--- Environment | Ours | Author | Timesteps HalfCheetah | $16,784\pm 292$ | $12,219\pm 4,899$ | $3\text{\times}{10}^{6}$ Hopper | $3,142\pm 654$ | $3,319\pm 175$ | $1\text{\times}{10}^{6}$ Ant | $4,987\pm 784$ | $3,845\pm 759$ | $1\text{\times}{10}^{6}$ Walker2d | $5,703\pm 408$ | $5,523\pm 466$ | $3\text{\times}{10}^{6}$ Humanoid | $5,871\pm 171$ | $6,268\pm 186$ | $3\text{\times}{10}^{6}$ Table 3: SAC Implementation Comparison to Author’s Code | TD3 Return | ---|---|--- Environment | Ours | Author | Timesteps HalfCheetah | $12,804\pm 493$ | $9,637\pm 859$ | $1\text{\times}{10}^{6}$ Hopper | $3,498\pm 99$ | $3,564\pm 115$ | $1\text{\times}{10}^{6}$ Ant | $5,700\pm 334$ | $4,372\pm 1,000$ | $1\text{\times}{10}^{6}$ Walker2d | $4,181\pm 607$ | $4,683\pm 540$ | $1\text{\times}{10}^{6}$ Humanoid | $5,085\pm 144$ | N/A | $1\text{\times}{10}^{6}$ Table 4: TD3 Implementation Comparison to Author’s Code ## 4 Conclusion In conclusion, we show that TD3 and SAC are closely related algorithms, and that it is possible to categorize them as belonging to the same general family of algorithms, namely ‘Off-Policy Continuous Generalized Policy Iteration’. We make this comparison complete by comparing against an oft-forgotten approach SVG(0). We then show that by matching hyperparameters, their performance is more similar than is often shown in the literature, and can be statistically indistinguishable; furthermore TD3 can in fact outperform SAC on certain environments whilst being more computationally efficient. To make this link from theory to practice explicit, we have implemented both in the open-source code base OffCon3, whereby many major elements of the code are shared for each algorithm. ## References * [1] Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel and Sergey Levine “Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor” In _ICML_ , 2018 * [2] Tuomas Haarnoja et al. “Soft Actor-Critic Algorithms and Applications”, 2018 arXiv:1812.05905 [cs.LG] * [3] Michael Janner, Justin Fu, Marvin Zhang and Sergey Levine “When to Trust Your Model: Model-Based Policy Optimization” In _Advances in Neural Information Processing Systems_ 32 Curran Associates, Inc., 2019, pp. 12519–12530 URL: https://proceedings.neurips.cc/paper/2019/file/5faf461eff3099671ad63c6f3f094f7f-Paper.pdf * [4] Ignasi Clavera, Yao Fu and Pieter Abbeel “Model-Augmented Actor-Critic: Backpropagating through Paths” In _International Conference on Learning Representations_ , 2020 URL: https://openreview.net/forum?id=Skln2A4YDB * [5] Yinlam Chow, Brandon Cui, MoonKyung Ryu and Mohammad Ghavamzadeh “Variational Model-based Policy Optimization”, 2020 arXiv:2006.05443 [cs.LG] * [6] Kate Rakelly et al. “Efficient Off-Policy Meta-Reinforcement Learning via Probabilistic Context Variables” 97, Proceedings of Machine Learning Research Long Beach, California, USA: PMLR, 2019, pp. 5331–5340 URL: http://proceedings.mlr.press/v97/rakelly19a.html * [7] Wenjie Shi, Shiji Song and Cheng Wu “Soft Policy Gradient Method for Maximum Entropy Deep Reinforcement Learning” In _Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, IJCAI-19_ International Joint Conferences on Artificial Intelligence Organization, 2019, pp. 3425–3431 DOI: 10.24963/ijcai.2019/475 * [8] David Silver et al. “Deterministic Policy Gradient Algorithms” 32.1, Proceedings of Machine Learning Research Bejing, China: PMLR, 2014, pp. 387–395 URL: http://proceedings.mlr.press/v32/silver14.html * [9] Richard S Sutton and Andrew G Barto “Reinforcement learning: An introduction” MIT press, 2018 * [10] Timothy P. Lillicrap et al. “Continuous control with deep reinforcement learning.” In _ICLR (Poster)_ , 2016 URL: http://arxiv.org/abs/1509.02971 * [11] Nicolas Heess et al. “Learning Continuous Control Policies by Stochastic Value Gradients” In _Advances in Neural Information Processing Systems 28_ Curran Associates, Inc., 2015, pp. 2944–2952 URL: http://papers.nips.cc/paper/5796-learning-continuous-control-policies-by-stochastic-value-gradients.pdf * [12] Diederik P. Kingma and Max Welling “Auto-Encoding Variational Bayes” In _2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings_ , 2014 * [13] Shakir Mohamed, Mihaela Rosca, Michael Figurnov and Andriy Mnih “Monte Carlo Gradient Estimation in Machine Learning” In _Journal of Machine Learning Research_ 21.132, 2020, pp. 1–62 * [14] Brian D. Ziebart “Modeling Purposeful Adaptive Behavior with the Principle of Maximum Causal Entropy” USA: Carnegie Mellon University, 2010 * [15] Scott Fujimoto, Herke Hoof and David Meger “Addressing Function Approximation Error in Actor-Critic Methods” In _ICML_ , 2018, pp. 1582–1591 URL: http://proceedings.mlr.press/v80/fujimoto18a.html * [16] Greg Brockman et al. “OpenAI Gym”, 2016 eprint: arXiv:1606.01540 * [17] Sébastien Bubeck “Convex Optimization: Algorithms and Complexity” In _Found. Trends Mach. Learn._ 8.3–4 Hanover, MA, USA: Now Publishers Inc., 2015, pp. 231–357 DOI: 10.1561/2200000050 * [18] Yuichi Yoshida and Takeru Miyato “Spectral Norm Regularization for Improving the Generalizability of Deep Learning”, 2017 arXiv:1705.10941 [stat.ML] * [19] B.. Welch “The Generalization of ’Student’s’ Problem when Several Different Population Variances are Involved” In _Biometrika_ 34.1-2, 1947, pp. 28–35 DOI: 10.1093/biomet/34.1-2.28 * [20] Aviral Kumar, Aurick Zhou, George Tucker and Sergey Levine “Conservative Q-Learning for Offline Reinforcement Learning” In _Advances in Neural Information Processing Systems_ 33 Curran Associates, Inc., 2020, pp. 1179–1191 URL: https://proceedings.neurips.cc/paper/2020/file/0d2b2061826a5df3221116a5085a6052-Paper.pdf ## Appendix A Objective of a Stochastic Gaussian Policy with fixed Variance Consider a Deterministic Policy: $\displaystyle J_{D}$ $\displaystyle=\mathbb{E}_{s_{t}\sim E}\left[Q(s_{t},a_{t})|_{a_{t}=\mu_{(s_{t})}}\right].$ The Random Policy is defined as $\pi(a_{t}|s_{t})=\mathcal{N}(a_{t}|\mu(s_{t}),\sigma^{2})$ where $\sigma$ is fixed: $\displaystyle J_{R}$ $\displaystyle=\mathbb{E}_{s_{t}\sim E}\left[\mathbb{E}_{a_{t}\sim\pi(a_{t}|s_{t})}\left[Q(s_{t},a_{t})\right]\right]$ $\displaystyle=\mathbb{E}_{s\sim E}\left[\int\mathcal{N}(a_{t}|\mu(s_{t}),\sigma^{2})Q(s_{t},a_{t})da_{t}\right].$ Performing a Taylor expansion of $Q(s_{t},a_{t})$ around $a_{t}=\mu(s)$ provides: $Q(s_{t},a_{t})=Q(s_{t},\mu(s_{t}))+\nabla_{a}Q(s_{t},a)|_{a=\mu(s_{t})}(a_{t}-\mu(s_{t}))\\\ +\frac{1}{2}\nabla^{2}_{a}Q(s_{t},a)|_{a=\mu(s_{t})}(a_{t}-\mu(s_{t}))^{2}+\dots.$ We address the different Taylor expansion orders separately (labeled 0, 1, 2, etc.). First 0th order: 0 $\displaystyle=\mathbb{E}_{s_{t}\sim E}\left[\int\mathcal{N}(a_{t}|\mu(s),\sigma^{2})Q(s_{t},\mu(s_{t}))da_{t}\right]$ $\displaystyle=\mathbb{E}_{s_{t}\sim E}\left[Q(s_{t},\mu(s_{t}))\int\mathcal{N}(a_{t}|\mu(s_{t}),\sigma^{2})da_{t}\right]$ $\displaystyle=\mathbb{E}_{s_{t}\sim E}\left[Q(s_{t},\mu(s_{t}))\right]$ $\displaystyle=J_{D}.$ Now 1st order: 1 $\displaystyle=\mathbb{E}_{s_{t}\sim E}\left[\int\mathcal{N}(a_{t}|\mu(s_{t}),\sigma^{2})\left(\nabla_{a}Q(s_{t},a)|_{a=\mu(s_{t})}(a_{t}-\mu(s))\right)da_{t}\right]$ $\displaystyle=\mathbb{E}_{s_{t}\sim E}\left[\nabla_{a}Q(s_{t},a)|_{a=\mu(s_{t})}\int\mathcal{N}(a_{t}|\mu(s_{t}),\sigma^{2})\left(a_{t}-\mu(s_{t})\right)da_{t}\right]$ $\displaystyle=\mathbb{E}_{s_{t}\sim E}\left[\nabla_{a}Q(s_{t},a)|_{a=\mu(s_{t})}\left(\int\mathcal{N}(a_{t}|\mu(s_{t}),\sigma^{2})a_{t}da_{t}-\mu(s_{t})\right)\right]$ $\displaystyle=\mathbb{E}_{s_{t}\sim E}\left[\nabla_{a}Q(s_{t},a)|_{a=\mu(s_{t})}\left(\mu(s_{t})-\mu(s_{t})\right)\right]$ $\displaystyle=0.$ Now 2nd order: 2 $\displaystyle=\mathbb{E}_{s_{t}\sim E}\left[\int\mathcal{N}(a_{t}|\mu(s_{t}),\sigma^{2})\left(\nabla^{2}_{a}Q(s_{t},a)|_{a=\mu(s_{t})}(a_{t}-\mu(s))^{2}\right)da_{t}\right]$ $\displaystyle=\mathbb{E}_{s_{t}\sim E}\left[\nabla^{2}_{a}Q(s_{t},a)|_{a=\mu(s_{t})}\int\mathcal{N}(a_{t}|\mu(s_{t}),\sigma^{2})\left(a_{t}-\mu(s))^{2}\right)da_{t}\right]$ $\displaystyle=\mathbb{E}_{s_{t}\sim E}\left[\nabla^{2}_{a}Q(s_{t},a)|_{a=\mu(s_{t})}\mathbb{E}_{a_{t}}\left[(a_{t}-\mu(s))^{2}\right]\right]$ $\displaystyle=\mathbb{E}_{s_{t}\sim E}\left[\nabla^{2}_{a}Q(s_{t},a)|_{a=\mu(s_{t})}\right]\sigma^{2}$ So putting it all (0, 1, 2) together: $\displaystyle J_{R}$ $\displaystyle=J_{D}+\frac{\sigma^{2}}{2}\mathbb{E}_{s_{t}\sim E}\left[\nabla^{2}_{a}Q(s_{t},a)|_{a=\mu(s_{t})}\right]$ ## Appendix B Efficiency Gain of TD3 in Ant Figure 3: Efficiency gain of TD3 over SAC ## Appendix C 2 v.s. 3 Layers Some recent Q-learning work for MuJoCo continuous control has used 3 hidden layers instead of the 2 hidden layers in the original author’s code, such as [20]. Following their lead, and noticing the particularly strong performance on HalfCheetah, we choose to implement 3 hidden layers in OffCon3. However, for fairness, we run a set of experiments with 2 hidden layers; the results are displayed in 4. We notice that apart from the significantly improved HalfCheetah performance, and SAC improving on Hopper, the convergent differences are marginal. Note that the small network plots are smoother as we evaluate performance at longer intervals. (a) HalfCheetah (b) Hopper (c) Ant (d) Walker2d (e) Humanoid Figure 4: SAC and TD3 Training Curves on MuJoCo Environments with different network depths (all 5 seeds). (Small) denotes a 2 hidden layer network for both actor and critic.
AI artificial intelligence ANN artificial neural network ASIC application-specific integrated circuit BD bounded delay BC breast cancer BNN binarized neural network CB connectionist bench CD completion detection CNN convolutional neural network CTM convolutional Tsetlin machine CPOG conditional partial order graph DI delay insensitive DR dual-rail DRAM dynamic RAM DSP digital signal processing DVFS dynamic voltage and frequency scaling FD-SOI fully-depleted silicon-on-insulator FPGA field-programmable gate array FSM finite state machine HDL hardware description language HVR house voting record INWE inverse-narrow-width effect ISA instruction set architecture IoT Internet of Things LA learning automaton LEC logical equivalence checking LFSR linear feedback shift register LU learning unit MAC multiply-accumulate MNIST Modified National Institute of Standards and Technology MEP minimum energy point ML machine learning MLP multi-layer perceptron MPP maximum power point MPPT maximum power point tracking MRAM NCL null convention logic NN neural network OCV on-chip variation PVT process, variation and temperature QDI quasi delay insensitive RAM random-access memory RDF random dopant fluctuation RTM recurrent Tsetlin machine SCM standard cell memory SI speed independent SR single-rail SRAM static RAM STG signal transition graph SVM support vector machine TA Tsetlin Automaton TAT Tsetlin automaton team TM Tsetlin machine ULV ultra-low voltage VLSI very-large-scale integration WSN wide sensor network IoT Internet of Things IC integrated circuit MFCC Mel-frequency cepstrum coefficients KWS keyword spotting DNN Deep Neural Network QCNN Quantized Convolutional Neural Network BCNN Binarized Convolutional Neural Network LSTM Long Short-term Memory SoC system on a chip GPU Graphics Processing Unit HPC High Performance Computing CUDA Compute Unified Device Architecture # Low-Power Audio Keyword Spotting using Tsetlin Machines Jie Lei Microsystems Research Group School of Engineering, Newcastle University, NE1 7RU, UK <EMAIL_ADDRESS> Tousif Rahman Microsystems Research Group School of Engineering, Newcastle University, NE1 7RU, UK <EMAIL_ADDRESS> Rishad Shafik Microsystems Research Group School of Engineering, Newcastle University, NE1 7RU, UK <EMAIL_ADDRESS> Adrian Wheeldon Microsystems Research Group School of Engineering, Newcastle University, NE1 7RU, UK <EMAIL_ADDRESS> Alex Yakovlev Microsystems Research Group School of Engineering, Newcastle University, NE1 7RU, UK <EMAIL_ADDRESS> Ole-Christoffer Granmo Centre for AI Research (CAIR), University of Agder, Kristiansand, Norway; <EMAIL_ADDRESS> Fahim Kawsar Pervasive Systems Centre, Nokia Bell Labs Cambridge, UK; <EMAIL_ADDRESS> Akhil Mathur Pervasive Systems Centre, Nokia Bell Labs Cambridge, UK; <EMAIL_ADDRESS> ###### Abstract The emergence of Artificial Intelligence (AI) driven Keyword Spotting (KWS) technologies has revolutionized human to machine interaction. Yet, the challenge of end-to-end energy efficiency, memory footprint and system complexity of current Neural Network (NN) powered AI-KWS pipelines has remained ever present. This paper evaluates KWS utilizing a learning automata powered machine learning algorithm called the Tsetlin Machine (TM). Through significant reduction in parameter requirements and choosing logic over arithmetic based processing, the TM offers new opportunities for low-power KWS while maintaining high learning efficacy. In this paper we explore a TM based keyword spotting (KWS) pipeline to demonstrate low complexity with faster rate of convergence compared to NNs. Further, we investigate the scalability with increasing keywords and explore the potential for enabling low-power on-chip KWS. _K_ eywords Speech Command $\cdot$ Keyword Spotting $\cdot$ MFCC $\cdot$ Tsetlin Machine $\cdot$ Learning Automata $\cdot$ Pervasive AI $\cdot$ Machine Learning $\cdot$ Artificial Neural Network $\cdot$ ## 1 Introduction Continued advances in Internet of Things (IoT) and embedded system design have allowed for accelerated progress in artificial intelligence (AI) based applications [1]. AI driven technologies utilizing sensory data have already had a profoundly beneficial impact to society, including those in personalized medical care [2], intelligent wearables [3] as well as disaster prevention and disease control [4]. A major aspect of widespread AI integration into modern living is underpinned by the ability to bridge the human-machine interface, viz. through sound recognition. Current advances in sound classification have allowed for AI to be incorporated into self-driving cars, home assistant devices and aiding those with vision and hearing impairments [5]. One of the core concepts that has allowed for these applications is through using KWS[6]. Selection of specifically chosen key words narrows the training data volume thereby allowing the AI to have a more focused functionality [7]. With the given keywords, modern keyword detection based applications are usually reliant on responsive real-time results [8] and as such, the practicality of transitioning keyword recognition based machine learning to wearable, and other smart devices is still dominated by the challenges of algorithmic complexity of the KWS pipeline, energy efficiency of the target device and the AI model’s learning efficacy. The algorithmic complexity of KWS stems from the pre-processing requirements of speech activity detection, noise reduction, and subsequent signal processing for audio feature extraction, gradually increasing application and system latency [7]. When considering on-chip processing, the issue of algorithmic complexity driving operational latency may still be inherently present in the AI model [7, 9]. AI based speech recognition often offload computation to a cloud service. However, ensuring real-time responses from such a service requires constant network availability and offers poor return on end-to-end energy efficiency [10]. Dependency on cloud services also leads to issues involving data reliability and more increasingly, user data privacy [11]. Currently most commonly used AI methods apply a neural network (NN) based architecture or some derivative of it in KWS [9, 12, 8, 13] (see Section Section 5.1 for a relevant review). The NN based models employ arithmetically intensive gradient descent computations for fine-tuning feature weights. The adjustment of these weights require a large number of system-wide parameters, called hyperparameters, to balance the dichotomy between performance and accuracy [14]. Given that these components, as well as their complex controls are intrinsic to the NN model, energy efficiency has remained challenging [15]. To enable alternative avenues toward real-time energy efficient KWS, low- complexity machine learning (ML) solutions should be explored. A different ML model will eliminate the need to focus on issues NN designers currently face such as optimizing arithmetic operations or automating hyper-parameter searches. In doing so, new methodologies can be evaluated against the essential application requirements of energy efficiency and learning efficacy, The challenge of energy efficiency is often tackled through intelligent hardware-software co-design techniques or a highly customized AI accelerator, the principal goal being to exploit the available resources as much as possible. To obtain adequate learning efficacy for keyword recognition the KWS-AI pipeline must be tuned to adapt to speech speed and irregularities, but most crucially it must be able to extract the significant features of the keyword from the time-domain to avoid redundancies that lead to increased latency. Overall, to effectively transition keyword detection to miniature form factor devices, there must be a conscious design effort in minimizing the latency of the KWS-AI pipeline through algorithmic optimizations and exploration of alternative AI models, development of dedicated hardware accelerators to minimize power consumption, and understanding the relationships between specific audio features with their associated keyword and how they impact learning accuracy. This paper establishes an analytical and experimental methodology for addressing the design challenges mentioned above. A new automata based learning method called the Tsetlin machine (TM) is evaluated in the KWS-AI design in place of the traditional perceptron based NNs. The TM operates through deriving propositional logic that describes the input features [16]. It has shown great potential over NN based models in delivering energy frugal AI application while maintaining faster convergence and high learning efficacy [17, 18, 19] Through exploring design optimizations utilizing the TM in the KWS-AI pipeline we address the following research questions: * • How effective is the TM at solving real-world KWS problems? * • Does the Tsetlin Machine scale well as the KWS problem size is increased? * • How robust is the Tsetlin Machine in the KWS-AI pipeline when dealing with dataset irregularities and overlapping features? This initial design exploration will uncover the relationships concerning how the Tsetlin Machine’s parameters affect the KWS performance, thus enabling further research into energy efficient KWS-TM methodology. ### 1.1 Contributions The contributions of this paper are as follows: * • Development of a pipeline for KWS using the TM. * • Using data encoding techniques to control feature granularity in the TM pipeline. * • Exploration of how the Tsetlin Machine’s parameters and architectural components can be adjusted to deliver better performance. ### 1.2 Paper Structure The rest of the paper is organized as follows: Section 2 offers an introduction to the core building blocks and hyper-parameters of the Tsetlin Machine. Through exploring the methods of feature extraction and encoding process blocks, the KWS-TM pipeline is proposed in Section 3.3. We then analyze the effects of manipulating the pipeline hyper-parameters in Section 4 showing the Experimental Results. We examine the effects of changing the number of Mel-frequency cepstrum coefficientss generated, the granularity of the encoding and the the robustness of the pipeline through the impact of acoustically similar keywords. We then apply our understanding of the Tsetlin Machines attributes to optimize performance and energy expenditure through Section 4.5. Through the related works presented in Section 5.2, we explore the current research progress on AI powered audio recognition applications and offer an in-depth look at the key component functions of the TM. We summarize the major findings in Section 6 and present the direction of future work in Section 7. ## 2 A Brief Introduction to Tsetlin Machine The Tsetlin Machine is a promising, new ML algorithm based on formulation of propositional logic [16]. This section offers a high level overview of the main functional blocks; a detailed review of relevant research progress can be found in Section 5.2. The core components of the Tsetlin Machine are: _a team of Tsetlin Automata (TA) in each clause_ , _conjunctive clauses_ , _summation and threshold_ module and the _feedback_ module, as seen in Figure 1. The TA are finite state machine (FSM)s that are used to form the propositional logic based relationships that describe an output class through the inclusion or exclusion of input features and their complements. The states of the TAs for each feature and its compliment are then aligned to a stochastically independent clause computation module. Through a voting mechanism built into the summation and threshold module the expected output class ${Y}$ is generated. During the training phase this class is compared against the target class $\hat{Y}$ and the TA states are incremented or decremented accordingly (this is also referred to as as issuing rewards or penalties). Figure 1: Block diagram of TM (dashed green arrow indicates penalties and rewards)[19] A fundamental difference between the TM and NNs is the requirement of a _Booleanizer_ module. The key premise is to convert the raw input features and their complements to Boolean features rather than binary encoded features as seen with NNs. These Boolean features are also referred to as literals: $\hat{X}$ and ${X}$. Current research has shown that significance-driven Booleanization of features for the Tsetlin Machine is vital in controlling the Tsetlin Machine size and processing requirements [18]. Increasing the number of features will increase the number of TA and increase computations for the clause module and subsequently the energy spent in incrementing and decrementing states in the feedback module. The choice of the number of clauses to represent the problem is also available as a design knob, which also directly affects energy/accuracy tradeoffs [19]. The Tsetlin Machine also has two hyper parameters, the _s_ value and the _Threshold (T)_. The Threshold parameter is used to determine the clause selection to be used in the voting mechanism, larger Thresholds will mean more clauses partake in the voting and influence the feedback to TA states. The _s_ value is used to control the fluidity with which the TAs can transition between states. Careful manipulation of these parameters can be used to determine the flexibility of the feedback module and therefore control the TMs learning stability [17]. As seen in Figure 2, increasing the Threshold and decreasing the _s_ value will lead to more events triggered as more states are transitioned. These parameters must be carefully tuned to balance energy efficiency through minimizing events triggered, and achieving good performance through finding the optimum _s_ -_T_ range for learning stability in the KWS application. Figure 2: The affect of $T$ and $s$ on reinforcements on TM[19] In order to optimize the TM for KWS, due diligence must be given to designing steps that minimize the Boolean feature set. This allows for finding a balance between performance and energy usage through varying the TM hyper parameters and the number of clause computation modules. Through exploitation of these relationships and properties of the TM, the KWS pipeline can be designed with particular emphasis on feature extraction and minimization of the number of the TMs clause computation modules. An extensive algorithmic description of Tsetlin Machine can be found in [16]. The following section will detail how these ideas can be implemented through audio pre-processing and Booleanization techniques for KWS. ## 3 Audio Pre-processing Techniques for KWS When dealing with audio data, the fundamental design efforts in pre-processing should be to find the correct balance between reducing data volume and preserving data veracity. That is, while removing the redundancies from the audio stream the data quality and completeness should be preserved. This is interpreted in the proposed KWS-TM pipeline through two methods: feature extraction through MFCCs, followed by discretization control through quantile based binning for Booleanization. These methods are expanded below. ### 3.1 Audio Feature Extraction using MFCC Audio data streams are always subject to redundancies in the channel that formalize as nonvocal noise, background noise and silence [20, 21]. Therefore the challenge becomes identification and extraction of the desired linguistic content (the keyword) and maximally discarding everything else. To achieve this we must consider transformation and filtering techniques that can amplify the characteristics of the speech signals against the background information. This is often done through the generation of MFCCs as seen in the signal processing flow in Figure 3. Figure 3: MFCC pipeline. The MFCC is a widely used audio file pre-processing method for speech related classification applications [22, 21, 23, 24, 25, 12]. The component blocks in the MFCC pipeline are specifically designed for extracting speech data taking into account the intricacies of the human voice. The _Pre-Emphasis step_ is used to compensate for the structure of the human vocal tract and provide initial noise filtration. When producing glottal sounds when speaking, higher frequencies are damped by the vocal tract which can be characterized as a step roll-off in the signals’ frequency spectrum [26]. The Pre-Emphasis step, as its name-sake suggests, amplifies (adds emphasis to) the energy in the high frequency regions, which leads to an overall normalization of the signal [27]. Speech signals hold a quasi-stationary quality when examined over a very short time period, which is to say that the statistical information it holds remains near constant [20]. This property is exploited through the _Framing and Windowing_ step. The signal is divided into around 20ms frames, then around 10-15ms long window functions are multiplied to these overlapping frames, in doing so we preserve the temporal changes of the signal between frames and minimize discontinuities (this is realized through the smoothed spectral edges and enhanced harmonics of the signal after the subsequent transformation to the frequency domain) [28]. The windowed signals are then transformed to the frequency domain through a Discrete Fourier Transform (DFT) process using the _Fast Fourier Transform (FFT)_ algorithm. FFT is chosen as it is able to find the redundancies in the DFT and reduce the amount of computations required offering quicker run-times. The human hearing system interprets frequencies linearly up to a certain range (around 1KHz) and logarithmically thereafter. Therefore, adjustments are required to translate the FFT frequencies to this non-linear function [29]. This is done through passing signal through the _Mel Filter Banks_ in order to transform it to the _Mel Spectrum_ [30]. The filter is realized by overlapping band-pass filters to create the required warped axis. Next, the logarithm of the signal is taken, this brings the data values closer and less sensitive to the slight variations in the input signal [30]. Finally we perform a _Discrete Cosine Transform (DCT)_ to take the resultant signal to the _Cepstrum_ domain [31]. The DCT function is used as energies present in the signal are very correlated as a result of the overlapping Mel filterbanks and the smoothness of the human vocal tract; the DCT finds the co-variance of the energies and is used to calculate the MFCC features vector [27, 32]. This vector can be passed to the Booleanizer module to produce the input Boolean features, as described next. ### 3.2 Feature Booleanization As described in Section 2, Booleanization is an essential step for logic based feature extraction in Tsetlin Machines. Minimizing the Boolean feature space is crucial to the Tsetlin Machine’s optimization. The size and processing volume of a TM is primarily dictated by the number of Booleans [18]. Therefore, a pre-processing stage for the audio features must be embedded into the pipeline before the TM to allow for granularity control of the raw MFCC data. The number of the Booleanized features should be kept as low as possible while capturing the critical features for classification [18]. The discretization method should be able to adapt to, and preserve the statistical distribution of the MFCC data. The most frequently used method in categorizing data is through binning. This is the process of dividing data into group, individual data-points are then represented by the group they belong to. Data points that are close to each other are put into the same group thereby reducing data granularity [16]. Fixed width binning methods are not effective in representing skewed distribution and often result in empty bins, they also require manual decision making for bin boundaries. Therefore, for adaptive and scalable Booleanization quantile based binning is preferred. Through binning the data using its own distribution, we maintain its statistical properties and do not need to provide bin boundaries, merely the number of bins the data should be discretized into. The control over the number of quantiles is an important parameter in obtaining the final Boolean feature set. Choosing two quantiles will result in each MFCC coefficient being represented using only one bit; however, choosing ten quantiles (or bins) will result in four bits per coefficient. Given the large number of coefficients present in the KWS problem, controlling the number of quantiles is an effective way to reduce the total TM size. ### 3.3 The KWS-TM pipeline The KWS-TM pipeline is composed of the the data encoding and classification blocks presented in Figure 4. The data encoding scheme encompasses the generation of MFCCs and the quantile binning based Booleanization method. The resulting Booleans are then fed to the Tsetlin Machine for classification. The figure highlights the core attributes of the pre-processing blocks: the ability to extract the audio features only associated with speech through MFCCs and the ability to control their Boolean granularity through quantile binning. To explore the functionality of the pipeline and the optimizations that can be made, we return to our primary intentions, i.e., to achieve energy efficiency and high learning efficacy in KWS applications. We can now use the design knobs offered in the pre-processing blocks, such as variable window size in the MFCC generation, and control over the number of quantiles to understand how these parameters can be used in presenting the Boolean data to the TM in a way to returns good performance utilizing the least number of Booleans. Through Section 2 we have also seen the design knobs available through variation of the hyperparameters _s_ and Threshold _T_ , as well as the number of clause computation modules used to represent the problem. Varying the parameters in both the encoding and classification stages through an experimental context will uncover the impact they have on the overall KWS performance and energy usage. Figure 4: The data encoding and classification stages in the KWS-TM pipeline ## 4 Experimental Results To evaluate the proposed KWS-TM pipeline, Tensorflow speech command dataset was used111Tensorflow speech command: https://tinyurl.com/TFSCDS. The dataset consists of many spoken keywords collected from a variety of speakers with different accents, as well as male and female gender. The datapoints are stored as 1 second long audio files where the background noise is negligible. This reduces the effect of added redundancies in the MFCC generation, given our main aim is predominantly to test functionality, we will explore the impact of noisy channels in our future work. This dataset is commonly used in testing the functionality of ML models and will therefore allow for fair comparisons to be drawn [33]. From the Tensorflow dataset, 10 keywords: "Yes", "No", "Stop", "Seven", "Zero", "Nine", "Five", "One", "Go" and "Two", have been chosen to explore the functionality of the pipeline using some basic command words. Considering other works comparing NN based pipelines, 10 keywords is the maximum used [34, 13]. Among the keywords chosen, there is an acoustic similarity between "No" and "Go", therefore, we explore the impact of 9 keywords together (without "Go") and then the effect of "No" and "Go" together. The approximate ratio of training data, testing data and validation data is given as 8:1:1 respectively with a total of 3340 datapoints per class. Using this setup, we will conduct a series of experiments to examine the impact of the various parameters of the KWS-TM pipeline discussed earlier. The experiments are as follows: * • Manipulating the window length and window steps to control the number of MFCCs generated. * • Exploring the effect of different quantile bins to change the number of Boolean features. * • Using a different number of the keywords ranging from the 2 to 9 to explore the scalability of the pipeline. * • Testing the effect on performance of acoustically different and similar keywords. * • Changing the size of the TM through manipulating the number of clause computation modules, optimizing performance through tuning the feedback control parameters _s_ and _T_. ### 4.1 MFCC Setup It is well defined that the number of input features to the TM is one of the major factors that affect its resource usage [17, 18, 19]. Increased raw input features means more Booleans are required to represent them and thus the number of Tsetlin Automaton (TA) in the TM will also increase leading to more energy required to provide feedback to them. Therefore, reducing the number of features at the earliest stage of the data encoding stage of the pipeline is crucial to implementing energy-frugal TM applications. The first set of parameters available in manipulating the number of features comes in the form of the _Window Step_ and the _Window Length_ (this takes place in the "Framing an Windowing" stage in Figure 4) in MFCC generation and can be seen through Figure 5(a). --- (a) The windowing process. (b) Effect of increasing window length. (c) Effect of increasing window step. Figure 5: The Hamming window function applied to audio pre-processing. The window function is effective in reducing the spectral distortion by tapering the sample signal at the beginning and ending of each frame (We use overlapping frames to ensure signal continuity is not lost). Smaller Window Steps lead to a more fine grained and descriptive representation of the audio features through more frames and therefore more MFCCs but this also increases computations and latency. Increasing the Window Length leads to a linear decrease in the total number of frames and therefore the MFCCs as seen in Figure 6(a). Given that the Window Steps are kept constant for this experiment, we have a linearly decreasing number of window overlaps resulting in a linearly decreasing total number of window functions, FFTs and subsequent computations. This leads to the linear decrease in the MFCCs across all frames. Increasing the Window Step leads to much sharper fall given the overlapping regions now no longer decrease linearly as seen in Figure 6(b). This results in a total number of non-linearly decreasing window functions and therefore much fewer FFTs and so on, leading to much fewer MFCCs across all frames. As a result of this, the smaller the increase in the Window Step the larger the decrease in the number of frames and therefore MFCCs. | ---|--- (a) The effect of increasing window length. | (b) The effect of increasing window step. Figure 6: Changing the number of MFCC coefficients through manipulating the Window parameters. To test the effectiveness of manipulating the Window Length and Window Step, the MFCC coefficients were produced for 4 keywords and the TMs classification performance was examined as seen in Figure 7(a) and Figure 7(b). Changing the Window Length results in much bigger falls in accuracy compared to Window Step. This is due to the diminished signal amplitudes at the window edges, longer windows mean more tapering of the edge amplitudes and fewer overlaps to preserve the signal continuities as seen through Figure 5(b). As a result, the fidelity of generated the MFCC features is reduced. | ---|--- (a) Effect of window length on accuracy. | (b) Effect of window step on accuracy. Figure 7: Effect of changing window parameters on classification accuracy. The effect of increasing the Window Step leads to a smaller drop in accuracy. We see the testing and validation accuracy remain roughly the same at around 90.5$\%$ between 0.03 and 0.10 second Window Steps and then experience a slight drop. Once again this is due to the tapering effect of the window function, given the window length remains the same for this experiment, we know that the increasing of window steps will mean far fewer total overlaps and a shrinking overlapping region as seen in Figure 5(c). The overlaps are used to preserve the continuity of the signal against the window function edge tapering, as the size of the overlapping regions decrease, the effect of edge tapering increases thereby leading to increased loss of information. The accuracy remains constant up to a Window Step of 0.1s as the Window Length is sufficiently long to capture enough of the signal information, once the overlapping regions start to shrink we experience the loss in accuracy. We can see that increasing the Window Step is very effective in reducing the number of frames and therefore the total number MFCC coefficients across all frames and providing the Window Length is long enough, the reduction in performance is minimal. To translate these findings toward energy efficient implementations, we must give increased design focus to finding the right balance between the size of the Window Step parameter and the achieved accuracy given the reduction in computations from the reduction in features produced. ### 4.2 Impact of Number of Quantiles Increased granularity through more bins will lead to improved performance but it is observed that this is not the case entirely. Table 1 shows the impact of the KWS-TM performance when increasing the number of bins. The testing and validation accuracy remain around the same with 1 Boolean per feature compared with 4 Booleans per feature. Figure 8 shows the large variance in some feature columns and no variance in others. The zero variance features are redundant in the subsequent Booleanization, they will be represented through the same Boolean sequence. The features with large variances are of main interest. We see that the mean for these features is relatively close to zero compared to their variance (as seen in Figure 9), therefore one Boolean per feature representation is sufficient, a 1 will represent values above the mean and 0 will represent below. The logical conclusion to be made from these explorations is that the MFCC alone is sufficient in both eliminating redundancies and extracting the keyword properties and does not require additional granularity beyond one Boolean per feature to distinguish classes. Figure 8: Variance between MFCC features. Figure 9: Mean of MFCC features. We have seen that the large variance of the MFCCs mean that they are easily represented by 1 Boolean per feature and that is sufficient to achieve high performance. This is an important initial result, for offline learning we can now also evaluate the effect of removing the no variance features in future work to further reduce the total number of Booleans. From the perspective of the Tsetlin Machine there is an additional explanation as to why the performance remains high even when additional Boolean granularity is allocated to the MFCC features. Given that there are a large number datapoints in each class (3340), if the MFCCs that describe these datapoints are very similar then the TM will have more than sufficient training data to settle on the best propositional logic descriptors. This is further seen by the high training accuracy compared to the testing and validation accuracy. Table 1: Impact of increasing quantiles with 4 classes Training | Testing | Validation | Num.Bins | Bools per Feature | Total Bools ---|---|---|---|---|--- 94.8% | 91.3% | 91.0% | 2 | 1 | 378 96.0% | 92.0% | 90.7% | 4 | 2 | 758 95.9% | 90.5% | 91.0% | 6 | 3 | 1132 95.6% | 91.8% | 92.0% | 8 | 3 | 1132 97.1% | 91.0% | 90.8% | 10 | 4 | 1512 ### 4.3 Impact of Increasing the Number of Keywords Figure 10(a) shows the linear nature with which the training, testing and validation accuracy decrease as the number of keywords are increased for a TM with 450 clauses with 200 epochs for training. We note that the testing and validation accuracy start to veer further away from the training accuracy with the increase of keywords. This performance drop is expected in ML methods as the problem scales [35]. Despite the large number of datapoints per keyword this is an indicator of overfitting, as confirmed through Figure 10(b) showing around a 4$\%$ increase. The implication of this is that increased number of keywords make it difficult for the TM to create distinct enough propositional logic to separate the classes. The performance drop is caused when the correlation of keywords outweighs the number of datapoints to distinguish each of them. This behavior is commonly observed in ML models for audio classification applications [23]. The explained variance ratio of the dataset with an increasing number of keywords was taken for the first 100 Principle Component Analysis eigenvalues, as seen in Figure 10(b). We observe that as the number of keywords is increased, the system variance decreases, i.e. the inter-class features start to become increasingly correlated. Correlated inter-class features will lead to class overlap and degrade TM performance [18]. Through examination of the two largest Linear Discriminant component values for the 9 keyword dataset, we clearly see in Figure 11 that there is very little class separability present. | ---|--- (a) The effect on accuracy. | (b) The amount of overfitting. Figure 10: The effect of increasing the number of keywords. To mitigate against the effect on performance of increasing keywords, there are two methods available: Firstly to adjust the Tsetlin Machines hyperparameters to enable more events triggered (see Figure 2). In doing so the this may allow the TM to create more differing logic to describe the classes. Then, by increasing the number of clause computation modules, the TM will have a larger voting group in the Summation and Threshold module and potential reach the correct classification more often. Secondly the quantity of the datapoints can be increased, however, for this to be effective the new dataset should hold more variance and completeness when describing each class. This method of data regularization is often used in audio ML applications to deliberately introduce small variance between datapoints [21]. Figure 11: LDA of 9 keywords. ### 4.4 Acoustically Similar Keywords In order to test the robustness of the KWS-TM pipeline functionality, we must emulate real-word conditions where a user will use commands that are acoustically similar to others. Table 2 shows the results of such circumstances. The _Baseline_ experiment is a KWS dataset consisting of 3 keywords: ’Yes’, ’No’ and ’Stop’. The second experiment then introduces the keyword ’Seven’ to the dataset and the third experiment introduces the keyword ’Go’. The addition of ’Seven’ causes a slight drop in accuracy adhering to our previously made arguments of increased correlation and the presence of overfitting. However the key result is the inclusion of ’Go’; ’Go’ is acoustically similar to ’No’ and this increases the difficulty in separating these two classes. We see from Figure 12(a), showing the first two LDA components that adding ’Seven’ does not lead to as much class overlap as adding ’Go’ as seen in Figure 12(b). As expected, the acoustic similarities of ’No’ and ’Go’ lead to significant overlap. We have seen from the previous result (Figure 11) that distinguishing class separability is increasingly difficult when class overlaps are present. Table 2: Impact of acoustically similar keywords. Experiments | Training | Testing | Validation ---|---|---|--- Baseline | 94.7% | 92.6% | 93.1% Baseline + ‘Seven’ | 92.5% | 90.1% | 90.2% Baseline + ‘Go’ | 85.6% | 82.6% | 80.9% | ---|--- (a) The Baseline with ‘Seven’. | (b) The Baseline with ‘Go’. Figure 12: The LDA of 4 keywords - the Baseline with one other. ### 4.5 Number of Clauses per Class So far we have considered the impact of Booleanization granularity, problem scalabilty and robustness when dealing with acoustically similar classes. Now, we turn our attention towards optimizing the KWS-TM pipeline to find the right functional balance between performance and energy efficiency. This is made possible through two streams of experimentation: manipulating the number of clauses for each keyword class in the TM and observing the energy expenditure and accuracy, and experimenting with the TMs hyperparameters to enable better performance using fewer clauses. | ---|--- (a) The effect on accuracy. | (b) The effect on overfitting. Figure 13: Effect of increasing the number of clauses on accuracy and overfitting. The influence of increasing the number of clauses was briefly discussed in Section 2, here we can see the experimental result in Figure 13(a) showing the impact of increasing clauses with 4 classes. Increased number of clauses leads to better performance. However, upon closer examination we can also see the impact of overfitting at the clause level, i.e., increasing the number of clauses has resulted in a larger difference in the training accuracy with the testing and validation. The datapoints for the 4 classes were sufficient to create largely different sub-patterns for the TAs during training, but not complete enough to describe new data in the testing and validation. As a result, when clauses are increased, more clauses reach incorrect decisions and sway the voting in the summation and threshold module toward incorrect classification, which is seen through Figure 14(a). The TM has two types of feedback, Type I, which introduces stochasticity to the system and Type II, which bases state transitions on the results of corresponding clause value. Type II feedback is predominantly used to diminish the effect of false positives. We see that as the clause value increases the TM uses more Type II feedback indicating increased false positive classifications. This result is for due to the incompleteness in the training data in describing all possible logic propositions for each class. We see this through 14(b); despite increasing the number of epochs we do not experience a boost in testing and validation accuracy and through Figure 13(b) we find the point where the overfitting outweighs the accuracy improvement at around 190-200 clauses. | ---|--- (a) The effect of clauses on feedback. | (b) The effect of epoch on accuracy. Figure 14: Effect of increasing the number of clauses on TM feedback (a), and the effect of increasing the number of epochs on accuracy (b). From the perspective of energy efficiency, these results offer two possible implications for the KWS-TM pipeline, if a small degradation of performance in the KWS application is acceptable, then operating at a lower clause range will be more beneficial for the TM. The performance can then be boosted through hyperparameters available to adjust feedback fluidity. This approach will reduce energy expenditure through fewer clause computations and reduce the effects of overfitting when the training data lacks enough completeness. Alternatively, if performance is the main goal, then the design focus should be on injecting training data with more diverse datapoints to increase the descriptiveness of each class. In that case, increased clauses will provide more robust functionality. Table 3: Impact of the number of clauses on energy/accuracy tradeoffs. | Clauses | Current | Time | Energy | Accuracy ---|---|---|---|---|--- Training | 100 | 0.50 A | 68 s | 426.40 J | - Training | 240 | 0.53 A | 96 s | 636.97 J | - Inference | 100 | 0.43 A | 12 s | 25.57 J | 80 % Inference | 240 | 0.47 A | 37 s | 87.23 J | 90 % The impacts of being resource efficient and energy frugal are most prevalent when implementing KWS applications into dedicated hardware and embedded systems. To explore this practically, the KWS-TM pipeline was implemented onto a Raspberry Pi. The same 4 keyword experiement was ran with 100 and 240 clauses. As expected, we see that increased clause computations lead to increased current, time and energy usage, but also delivers better performance. We can potentially boost the performance of the Tsetlin Machine at lower clauses through manipulating the hyperparameters as seen Table 4. Table 4: Impact of the T values on accuracy Clauses | T | Training | Testing | Validation | Better Classification ---|---|---|---|---|--- 30 | 2 | 83.5 % | 80.5 % | 83.8 % | ✓ 30 | 23 | 74.9 % | 71.1 % | 76.1 % | 450 | 2 | 89.7 % | 86.1 % | 84.9 % | 450 | 23 | 96.8 % | 92.5 % | 92.7 % | ✓ The major factor that has impacted the performance of the KWS is the capacity of the TM which is determined by the number of clauses per class. The higher the number clauses, the higher the overall classification accuracy [18]. Yet, the resource usage will increase linearly along with the energy consumption and memory footprint. Through Table 4 we see that at 30 clauses the accuracy can be boosted through reducing the Threshold hyperparameter. The table offers two design scenarios; firstly, very high accuracy is achievable through a large number of clauses (450 in this case) and a large Threshold value. With a large number of clauses an increased number of events must be triggered in terms of state transitions (see Figure 2) to encourage more feedback to clauses and increases the TMs decisiveness. While this offers a very good return on performance, the amount of computations are increased with more clauses and more events triggered and this leads to increased energy expenditure as seen through Table 3. In contrast, using 960 clauses and a lower Threshold still yields good accuracy but at a much lower energy expenditure through fewer clause computations and feedback events. A smaller number of clauses mean that the vote of each clause has more impact, even at a smaller Threshold the inbuilt stochasticity of the TM’s feedback module allows the TAs to reach the correct propositional logic. Through these attributes it is possible to create more energy frugal TMs requiring fewer computations and operating at a much lower latency. ### 4.6 Comparative Learning Convergence and Complexity Analysis of KWS-TM Both TMs and NNs have modular design components in their architecture; For the TM, this is in the form of clauses and for the NN it is the number of neurons. NNs require input weights for the learning mechanism which define the neurons’ output patterns. The number of weights and the number of neurons are variable, however more neurons will lead to better overall NN connectivity due to more refined arithmetic pathways to define a learning problem. For the TM the clauses are composed of TAs. The number of TAs are defined by the number of Boolean features which remains static throughout the course of the TMs learning. It is the number of clauses that is variable, increasing the clauses typically offers more propositional diversity to define a learning problem. Through Figure 15 and Table 5 we investigate the learning convergence rates of the TM against 4 ’vanilla’ NN implementations. The TM is able to converge to 90.5$\%$ after less than 10 epochs highlighting its quick learning rate compared to NNs which require around 100 epochs to converge to the isoaccuracy target ($\approx$90%). After further 100 epochs the NN implementations reach only marginally better accuracy than TM. The indirect implication of faster convergence is improved energy efficiency as fewer training epochs will result in fewer computations required for the TA states to settle. Table 5 shows one of the key advantages of the TM over all types of NNs, the significantly fewer parameters required, i.e. low-complexity. Large number of parameters needed for NNs are known to limit their practicality for on-chip KWS solutions [34, 36, 12, 13], where as the TM offers a more resource-frugal alternative. With only 960 clauses, which require only logic based processing, the TM outperforms even the most capable large and deep NNs. In our future work, we aim to exploit this to enable on-chip learning based KWS solutions. | ---|--- (a) Convergence of the TM against shallow NNs. | (b) Convergence of the TM against deep NNs. Figure 15: Training convergence of TM and NN implementations. Table 5: The required parameters for different NNs and the TM for a 4 keyword problem. KWS-ML Configuration | Num. neurons | Num. hyperparameters ---|---|--- NN Small & Shallow: 256+512X2 | 1,280 | 983,552 NN Small & Deep: 256+512X5 | 2,816 | 2,029,064 NN Large & Shallow: 256+1024X2 | 2,304 | 2,822,656 NN Large & Deep: 256+1024X5 | 5,376 | 7,010,824 TM with 240 Clauses per Class | 960 (clauses) | 2 hyperparameters with 725760 TAs ## 5 Related Work This section will provide a brief examination into current KWS research, industrial challenges with KWS, deeper look in the component blocks of the TM and provide insight into the current developments and the future research directions. ### 5.1 Current KWS developments The first KWS classification methods proposed in the late 1970s used MFCCs for their feature extraction ability and because the coefficients produced offered a very small dimensionality compared to the raw input data that was being considered then [37]. It was later shown that compared to other audio extraction methods such as near prediction coding coefficients (LPCC)s and perceptual linear production (PLP), MFCCs perform much better with increased background noise and low SNR [12]. For the classifier, Hidden Markov Models (HMMs) were favored after the MFCC stage due to their effectiveness in modelling sequences [37]. However they rely on many summation and Bayesian probability based arithmetic operations as well as the computationally intensive _Viterbi_ decoding to identify the final keyword [34, 38, 39]. Later it was shown that Recurrent Neural Networks (RNN)s outperform HMMs but suffer from operational latency as the problem scales, albeit RNNs still have faster run-times than HMM pipelines given they do not require a decoder algorithm [38]. To solve the latency issue, the Deep Neural Network (DNN) was used, it has smaller memory footprint and reduced run-times compared to HMMs [12, 39]. However, DNNs are unable to efficiently model the temporal correlations of the MFCCs and their transitional variance [36] [34]. In addition to this, commonly used optimization techniques used for DNNs such as pruning, encoding and quantization lead to great accuracy losses with KWS applications [12]. The MFCC features exist as a 2D array as seen in Figure 4, to preserve the temporal correlations and transitional variance, this array can be treated as an image and a convolutional neural network (CNN) can be used for classification [13, 36]. With the use of convolution comes the preservation of the spatial and temporal dependencies of the 2D data as well as the reduction of features and computations from the convolution and pooling stages [13]. However, once again both the CNN and DNN suffer from the large number of parameters (250K for the dataset used in [36] and 9 million Multiplies required for the CNN). Despite the gains in performance and reductions in latency, the computational complexity and large memory requirements from parameter storage are ever present with all NN based KWS solutions. The storage and memory requirements played a major part in transitioning to a micro-controller system for inference where memory is limited through the size of the SRAM [34]. In order to accommodate for the large throughput of running NN workloads, micro-controllers with integrated DSP instructions or integrated SIMD and MAC instructions can accelerate low-precision computations [34]. When testing for 10 keywords, it was shown experimentally in [34], that for systems with limited memory and compute abilities DNNs are favorable given they use the fewer operations despite having a lower accuracy (around 6$\%$ less) compared to CNNs. It is when transitioning to hardware that the limitations of memory and compute resources become more apparent. In these cases it is better to settle for energy efficiency through classifiers with lower memory requirements and operations per second even if there is a slight drop in performance. A 22nm CMOS based Quantized Convolutional Neural Network (QCNN) Always-ON KWS accelerator is implemented in [12], they explore the practicalities of CNN in hardware through quantized weights, activation values and approximate compute units. Their findings illustrate the effectiveness of hardware design techniques; the use of approximate compute units led to a significant decrease in energy expenditure, the hardware unit is able to classify 10 real-time keywords under different SNRs with a power consumption of 52$\mu$W. This impact of approximate computing is also argued in [13] with design focus given to adder design, they propose an adder with a critical path that is 49.28$\%$ shorter than standard 16-bit Ripple Carry Adders. Through their research work with earables Nokia Bell Labs Cambridge have brought an industrial perspective to the idea of functionality while maintaining energy frugality into design focus for AI powered KWS [40, 41], with particular emphasis on user oriented ergonomics and commercial form factor. They discovered that earable devices are not as influenced by background noise compared to smartphones and smartwatches and offer better signal-to-noise ratio for moving artefacts due to their largely fixed wearing position in daily activities (e.g. walking or descending stairs) [41]. This was confirmed when testing using Random Forest classifiers. ### 5.2 The Tsetlin Machine We briefly discussed the overall mechanism of the TM and the main building blocks in the Section 2. In this section, we will have a closer look to the fundamental learning element of the TM, namely the Tsetlin Automaton, as described in Figure 16. We will also present a more detailed look at the clause computing module as seen in Figure 17, and we will discuss the first application-specific integrated circuit (ASIC) implementation of the TM, the Mignon 222 Mignon AI: http://mignon.ai/, as seen in Figure 18. Figure 16: Mechanism of a TA. The TA is the most fundamental part of the TM forming the core learning element that drives classification (Figure 16). Developed by Mikhail Tsetlin in the 1950s, the TA is an FSM where the current state will transition towards or away from the middle state upon receiving Reward or Penalty reinforcements during the TMs training stage. The current state of the TA will decide the output of the automaton which will be either an Include (aA) or Exclude (aB). Figure 17: Mechanism of a Clause computing module (assuming TA1= 1 means _Include_ and TA1= 0 means _Exclude_). Figure 17 shows how the clause module create logic propositions that describe the literals based on the TA decisions through logic _OR_ operations between the negated TA decision and the literal. The TA decision is used to bit mask the literal and through this we can determine which literals are to be excluded. The proposition is then logic _AND_ ed and this forms the raw vote for this clause. Clauses can be of positive and negative polarity, as such, a sign will be added to the clause output before it partakes in the class voting. It is important to note the reliance purely on logic operations making the TM well suited to hardware implementations. Clauses are largely independent of each other, only coalescing for voting giving the TM good scalability potential. The feedback to the TM can be thought of on three levels, at the TM level, at the clause level and at the TA level. At the TM level, the type of feedback to issue is decided based on the target class and whether the TM is in learning or inference mode. For inference no feedback is given, we simply take the clause computes for each class and pass to the summation and threshold module to generate the predicted class. However, in training mode there is a choice of Type I feedback to combat false negatives or Type II feedback to combat false positives. This feedback choice is further considered at the clause level. At the clause level there are three main factors that will determine feedback type to the TAs, the feedback type decision from the TM level, the current clause value, and whether the magnitude of clause vote is above the magnitude of the Threshold. At the TA level, the feedback type from the clause level will be used in conjunction with the current TA state and the s parameter to determine whether there is inaction, penalty or reward given to the TA states. The simplicity of the TM shows its potential to be a promising NN alternative. Lei et al [19] comparatively analyzed the architecture, memory footprint and convergence of these two algorithms for different datasets. This research shows the fewer number of hyperparameter of the TM will reduce the complexity of the design. The convergence of the TM is higher than the NN in all experiments conducted. The most unique architectural advances of the TM is the propositional logic based learning mechanism which will be beneficial in achieving energy frugal hardware AI. Wheeldon et al. [18] presented the first ASIC implementation of the TM for Iris flower classifications (see Figure 18). Figure 18: The Mignon AI: ASIC microchip acclerating TM (left). This 65-nm technology based design is a breakthrough in achieving an energy efficiency of up to 63 Tera Operations per Joule (Tops/Joule) while maintaining high convergence rate and performance. The early results from this microchip has been extensively compared with Binarized Convolutional Neural Network (BCNN) and neuromorphic designs in [18]. In addition, Wheeldon et al. [18] also proposed a system-wide design space exploration pipeline in deploying TM into ASIC design. They introduced a detailed methodology from 1) dataset encoding building on the work seen in [4] to 2) software based design exploration and 3) an FPGA based hyperparameter search to 4) final ASIC synthesis. A follow-up work of this [42] also implemented a self-timed and event-driven hardware TM. This implementation showed power and timing elasticity properties suitable for low-end AI implementations at-the-microedge. Other works include mathematical lemma based analysis of clause convergence using the XOR dataset [43], natural language (text) processing [44], disease control [4], methods of automating the s parameter [45] as well as exploration of regression and convolutional TMs [46, 47]. The TM has so far, been implemented with many different programming languages such as, C, C++, C$\\#$, Python and Node.js, to name a few. It has also been optimized for High Performance Computing (HPC) through Compute Unified Device Architecture (CUDA) for accelerating Graphics Processing Unit (GPU) based solutions and currently through OpenCL for heterogeneous embedded systems [48]. Exploiting the natural logic underpinning there are currently ongoing efforts in establishing explainability evaluation and analysis of TMs[17]. Deterministic implementation of clause selection in TM, reported by [49], is a promising direction to this end. Besides published works, there are numerous talks, tutorials and multimedia resources currently available online to mobilize the hardware/software community around this emerging AI algorithm. Below are some key sources: Videos: https://tinyurl.com/TMVIDEOSCAIR. Publications: https://tinyurl.com/TMPAPERCAIR & www.async.org.uk. Software implementations: https://tinyurl.com/TMSWCAIR Hardware implementations, Mignon AI: http://www.mignon.ai/. A short video demonstrating KWS using TM can be found here: https://tinyurl.com/KWSTMDEMO. ## 6 Summary and Conclusions The paper presented the first ever TM based KWS application. Through experimenting with the hyperparameters of the proposed KWS-TM pipeline we established relationships between the different component blocks that can be exploited to bring about increased energy efficiency while maintaining high learning efficacy. From current research work we have already determined the best methods to optimize for the TM is through finding the right balance between reduction of the number of features, number of clauses and number of events triggered through the feedback hyper-parameters against the resulting performance from these changes. These insights were carried into our pipeline design exploration experiments. Firstly, we fine tuned the window function in the generation of MFCCs, we saw that increasing the window steps lead to much fewer MFCCs and if the window length is sufficient enough to reduce edge tapering then the performance degradation is minimal. Through quantile binning to manipulate the discretization of the Boolean MFCCs, it was seen that this did not yield change in performance. The MFCC features of interest have very large variances in each feature column and as such less precision can be afforded to them, even as low as one Boolean per feature. This was extremely useful in reducing the resulting TM size. Through manipulating the number of clause units to the TM on a Raspberry Pi, we confirmed the energy and latency savings possible by running the pipeline at a lower clause number and using the Threshold hyper-parameter the classification of the accuracy can also be boosted. Through these design considerations we are able to increase the energy frugality of the whole system and transition toward low-power hardware accelerators of the pipeline to tackle real time applications. The KWS-TM pipeline was then compared against some different NN implementations, we demonstrated the much faster convergence to the same accuracy during training. Through these comparisons we also highlighted the far fewer parameters required for the TM as well as a fewer number of clauses compared to neurons. The faster convergence, fewer parameters and logic over arithmetic processing makes the KWS-TM pipeline more energy efficient and enables future work into hardware accelerators to enable better performance and low power on-chip KWS. Acknowledgement: The authors gratefully acknowledge the funding from EPSRC IAA project “Whisperable” and EPSRC grant STRATA (EP/N023641/1). The research also received help from the computational powerhouse at CAIR333https://cair.uia.no/house-of-cair/. ## 7 Future Work Through testing the KWS-TM pipeline against the Tensorflow Speech data set we did not account for background noise effects. In-field IoT applications must be robust enough to minimize the effects of additional noise, therefore, future work in this direction should examine the effects of the pipeline with changing signal-to-noise ratios. The pipeline will also be deployed to a micro-controller in order to benefit from the effects of energy frugality by operating at a lower power level. ## References * [1] T. Rausch and S. Dustdar. Edge intelligence: The convergence of humans, things, and ai. In 2019 IEEE International Conference on Cloud Engineering (IC2E), pages 86–96, 2019. * [2] Itsuki Osawa, Tadahiro Goto, Yuji Yamamoto, and Yusuke Tsugawa. Machine-learning-based prediction models for high-need high-cost patients using nationwide clinical and claims data. * [3] Tiago M. Fernández-Caramés and Paula Fraga-Lamas. Towards the internet-of-smart-clothing: A review on iot wearables and garments for creating intelligent connected e-textiles. Electronics (Switzerland), 7, 12 2018. * [4] K. D. Abeyrathna, O. C. Granmo, X. Zhang, and M. Goodwin. Adaptive continuous feature binarization for tsetlin machines applied to forecasting dengue incidences in the philippines. In 2020 IEEE Symposium Series on Computational Intelligence (SSCI), pages 2084–2092, 2020. * [5] K. Hirata, T. Kato, and R. Oshima. Classification of environmental sounds using convolutional neural network with bispectral analysis. In 2019 International Symposium on Intelligent Signal Processing and Communication Systems (ISPACS), pages 1–2, 2019. * [6] Hadas Benisty, Itamar Katz, Koby Crammer, and David Malah. Discriminative keyword spotting for limited-data applications. Speech Communication, 99:1 – 11, 2018. * [7] J. S. P. Giraldo, C. O’Connor, and M. Verhelst. Efficient keyword spotting through hardware-aware conditional execution of deep neural networks. In 2019 IEEE/ACS 16th International Conference on Computer Systems and Applications (AICCSA), pages 1–8, 2019. * [8] J. S. P. Giraldo, S. Lauwereins, K. Badami, H. Van Hamme, and M. Verhelst. 18uw soc for near-microphone keyword spotting and speaker verification. In 2019 Symposium on VLSI Circuits, pages C52–C53, 2019. * [9] S. Leem, I. Yoo, and D. Yook. Multitask learning of deep neural network-based keyword spotting for iot devices. IEEE Transactions on Consumer Electronics, 65(2):188–194, 2019\. * [10] A depthwise separable convolutional neural network for keyword spotting on an embedded system. EURASIP Journal on Audio, 2020:10, 2020. * [11] Massimo Merenda, Carlo Porcaro, and Demetrio Iero. Edge machine learning for ai-enabled iot devices: A review. Sensors (Switzerland), 20, 5 2020. * [12] B. Liu, Z. Wang, W. Zhu, Y. Sun, Z. Shen, L. Huang, Y. Li, Y. Gong, and W. Ge. An ultra-low power always-on keyword spotting accelerator using quantized convolutional neural network and voltage-domain analog switching network-based approximate computing. IEEE Access, 7:186456–186469, 2019. * [13] S. Yin, P. Ouyang, S. Zheng, D. Song, X. Li, L. Liu, and S. Wei. A 141 uw, 2.46 pj/neuron binarized convolutional neural network based self-learning speech recognition processor in 28nm cmos. In 2018 IEEE Symposium on VLSI Circuits, pages 139–140, 2018. * [14] Nebojsa Bacanin, Timea Bezdan, Eva Tuba, Ivana Strumberger, and Milan Tuba. Optimizing convolutional neural network hyperparameters by enhanced swarm intelligence metaheuristics. 2020\. * [15] Rishad Shafik, Alex Yakovlev, and Shidhartha Das. Real-power computing. IEEE Transactions on Computers, 2018. * [16] Ole-Christoffer Granmo. The Tsetlin Machine - A Game Theoretic Bandit Driven Approach to Optimal Pattern Recognition with Propositional Logic. arXiv, April 2018. * [17] Rishad Shafik, Adrian Wheeldon, and Alex Yakovlev. Explainability and dependability analysis of learning automata based AI hardware. In IEEE IOLTS, 2020. * [18] Adrian Wheeldon, Rishad Shafik, Tousif Rahman, Jie Lei, Alex Yakovlev, and Ole-Christoffer Granmo. Learning automata based AI hardware design for IoT. Philosophical Trans. A of the Royal Society, 2020. * [19] J. Lei, A. Wheeldon, R. Shafik, A. Yakovlev, and O. C. Granmo. From arithmetic to logic based ai: A comparative analysis of neural networks and tsetlin machine. In 2020 27th IEEE International Conference on Electronics, Circuits and Systems (ICECS), pages 1–4, 2020. * [20] S. Chu, S. Narayanan, and C. . J. Kuo. Environmental sound recognition with time–frequency audio features. IEEE Transactions on Audio, Speech, and Language Processing, 17(6):1142–1158, 2009. * [21] Zohaib Mushtaq and Shun-Feng Su. Environmental sound classification using a regularized deep convolutional neural network with data augmentation. Applied Acoustics, 167:107389, 2020. * [22] W. Shan, M. Yang, J. Xu, Y. Lu, S. Zhang, T. Wang, J. Yang, L. Shi, and M. Seok. 14.1 a 510nw 0.41v low-memory low-computation keyword-spotting chip using serial fft-based mfcc and binarized depthwise separable convolutional neural network in 28nm cmos. In 2020 IEEE International Solid- State Circuits Conference - (ISSCC), pages 230–232, 2020. * [23] Muqing Deng, Tingting Meng, Jiuwen Cao, Shimin Wang, Jing Zhang, and Huijie Fan. Heart sound classification based on improved mfcc features and convolutional recurrent neural networks. Neural Networks, 130:22 – 32, 2020. * [24] L. Xiang, S. Lu, X. Wang, H. Liu, W. Pang, and H. Yu. Implementation of lstm accelerator for speech keywords recognition. In 2019 IEEE 4th International Conference on Integrated Circuits and Microsystems (ICICM), pages 195–198, 2019. * [25] Kirandeep Kaur and N. Jain. Feature extraction and classification for automatic speaker recognition system – a review. 2015\. * [26] Joseph W. Picone. Signal modeling techniques in speech recognition. In PROCEEDINGS OF THE IEEE, pages 1215–1247, 1993. * [27] Uday Kamath, John Liu, and James Whitaker. Automatic Speech Recognition, pages 369–404. Springer International Publishing, Cham, 2019. * [28] Automatic speech recognition. In Speech and Audio Signal Processing, pages 299–300. John Wiley & Sons, Inc., oct 2011. * [29] N.J. Nalini and S. Palanivel. Music emotion recognition: The combined evidence of mfcc and residual phase. Egyptian Informatics Journal, 17(1):1 – 10, 2016. * [30] Q. Li, Y. Yang, T. Lan, H. Zhu, Q. Wei, F. Qiao, X. Liu, and H. Yang. Msp-mfcc: Energy-efficient mfcc feature extraction method with mixed-signal processing architecture for wearable speech recognition applications. IEEE Access, 8:48720–48730, 2020. * [31] C. Paseddula and S. V. Gangashetty. Dnn based acoustic scene classification using score fusion of mfcc and inverse mfcc. In 2018 IEEE 13th International Conference on Industrial and Information Systems (ICIIS), pages 18–21, 2018. * [32] S. Jothilakshmi, V. Ramalingam, and S. Palanivel. Unsupervised speaker segmentation with residual phase and mfcc features. Expert Systems with Applications, 36(6):9799 – 9804, 2009. * [33] Pete Warden. Speech commands: A dataset for limited-vocabulary speech recognition, 2018\. * [34] Yundong Zhang, Naveen Suda, Liangzhen Lai, and Vikas Chandra. Hello edge: Keyword spotting on microcontrollers. CoRR, abs/1711.07128, 2017. * [35] Z. Zhang, S. Xu, S. Zhang, T. Qiao, and S. Cao. Learning attentive representations for environmental sound classification. IEEE Access, 7:130327–130339, 2019. * [36] Tara Sainath and Carolina Parada. Convolutional neural networks for small-footprint keyword spotting. In Interspeech, 2015. * [37] J. G. Wilpon, L. R. Rabiner, C. . Lee, and E. R. Goldman. Automatic recognition of keywords in unconstrained speech using hidden markov models. IEEE Transactions on Acoustics, Speech, and Signal Processing, 38(11):1870–1878, 1990. * [38] Santiago Fernández, Alex Graves, and Jürgen Schmidhuber. An application of recurrent neural networks to discriminative keyword spotting. In Proceedings of the 17th International Conference on Artificial Neural Networks, ICANN’07, page 220–229, Berlin, Heidelberg, 2007\. Springer-Verlag. * [39] G. Chen, C. Parada, and G. Heigold. Small-footprint keyword spotting using deep neural networks. In 2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4087–4091, 2014. * [40] Chulhong Min, Akhil Mathur, and Fahim Kawsar. Exploring audio and kinetic sensing on earable devices. In Proceedings of the 4th ACM Workshop on Wearable Systems and Applications, WearSys ’18, page 5–10, New York, NY, USA, 2018. Association for Computing Machinery. * [41] F. Kawsar, C. Min, A. Mathur, and A. Montanari. Earables for personal-scale behavior analytics. IEEE Pervasive Computing, 17(3):83–89, 2018. * [42] Adrian Wheeldon, Alex Yakovlev, Rishad Shafik, and Jordan Morris. Low-latency asynchronous logic design for inference at the edge. arXiv preprint arXiv:2012.03402, 2020. * [43] Lei Jiao, Xuan Zhang, Ole-Christoffer Granmo, and K. Darshana Abeyrathna. On the convergence of tsetlin machines for the xor operator, 2021. * [44] Bimal Bhattarai, Ole-Christoffer Granmo, and Lei Jiao. Measuring the novelty of natural language text using the conjunctive clauses of a tsetlin machine text classifier, 2020. * [45] Saeed Rahimi Gorji, Ole-Christoffer Granmo, Adrian Phoulady, and Morten Goodwin. A tsetlin machine with multigranular clauses, 2019. * [46] K Darshana Abeyrathna, Ole-Christoffer Granmo, Xuan Zhang, Lei Jiao, and Morten Goodwin. The regression tsetlin machine: a novel approach to interpretable nonlinear regression. Philosophical Trans. A of the Royal Society, 2019. * [47] Ole-Christoffer Granmo, Sondre Glimsdal, Lei Jiao, Morten Goodwin, Christian W. Omlin, and Geir Thore Berge. The convolutional tsetlin machine. CoRR, abs/1905.09688, 2019. * [48] K Darshana Abeyrathna, Bimal Bhattarai, Morten Goodwin, Saeed Gorji, Ole-Christoffer Granmo, Lei Jiao, Rupsa Saha, and Rohan K Yadav. Massively parallel and asynchronous tsetlin machine architecture supporting almost constant-time scaling. arXiv preprint arXiv:2009.04861, 2020. * [49] K Darshana Abeyrathna, Ole-Christoffer Granmo, Rishad Shafik, Alex Yakovlev, Adrian Wheeldon, Jie Lei, and Morten Goodwin. A novel multi-step finite-state automaton for arbitrarily deterministic tsetlin machine learning. In International Conference on Innovative Techniques and Applications of Artificial Intelligence, pages 108–122. Springer, 2020. *[KWS]: keyword spotting *[IoT]: Internet of Things *[AI]: artificial intelligence *[NN]: neural network *[ML]: machine learning *[TM]: Tsetlin machine *[FSM]: finite state machine *[MFCC]: Mel-frequency cepstrum coefficients *[MFCCs]: Mel-frequency cepstrum coefficients *[TA]: Tsetlin Automaton *[TMs]: Tsetlin machine *[NNs]: neural network *[DNN]: Deep Neural Network *[CNN]: convolutional neural network *[QCNN]: Quantized Convolutional Neural Network *[ASIC]: application-specific integrated circuit *[BCNN]: Binarized Convolutional Neural Network *[HPC]: High Performance Computing *[CUDA]: Compute Unified Device Architecture *[GPU]: Graphics Processing Unit
# Launchers and Targets in Social Networks Pedro Martins Polytechnic Institute of Coimbra, Coimbra Business School - ISCAC, Portugal, and Centro de Matemática, Aplicações Fundamentais e Investigação Operacional (CMAFcIO), Universidade de Lisboa, 1749-016 Lisboa, Portugal e-mail<EMAIL_ADDRESS>Filipa Alarcão Martins Digital Marketing Analyst, Portugal e-mail<EMAIL_ADDRESS> ###### Abstract Influence propagation in social networks is a subject of growing interest. A relevant issue in those networks involves the identification of key influencers. These players have an important role on viral marketing strategies and message propagation, including political propaganda and fake news. In effect, an important way to fight malicious usage on social networks is to understand their properties, their structure and the way messages propagate. This paper proposes two new indices for analysing message propagation in social networks, based on the network topological nature and the power of the message. The first index involves the strength of each node as a launcher of the message, dividing the nodes into launchers and non-launchers. The second index addresses the potential of each member as a receiver (target) of the message, dividing the nodes into targets and non-targets. Launcher individuals should indicate strong influencers and target individuals should identify the best target consumers. These indices can assist other known metrics when used to select efficient influencers in a social network. For instance, instead of choosing a strong and probably expensive member according to its degree in the network (number of followers), we may previously select those belonging to the launchers group and look for the lowest degree members, which are probably cheaper but still guarantying almost the same influence effectiveness as the largest degree members. On a different direction, using the second index, the strong target members should characterize relevant consumers of information in the network, which may include fake news’ regular collectors. We discuss these indices using small-world randomly generated graphs and a number of real-world social networks available in known datasets repositories. Keywords: influencers, influence propagation, social networks, launchers and targets in social networks ## 1 Introduction Social networks have long existed in society, but the fast growth of the web made these networks emerge at an unimaginable scale. These networks represent linkage among people and besides their relevancy in communication and in society, they also provide influence exertion, information dissemination (true or false) and, in some cases, gossip spread. Message or information propagation follows on a cascade setting in an online social network. As an example, suppose that John writes on a friend’s Facebook wall about a party he has been last night. This post is communicated to his friends who may typically comment, getting the message across their friends, and so on. This way, John’s initial message propagates transitively throughout the network. In another example, suppose that Mary posts on Twitter about a new mobile phone she bought. Some of her followers on Twitter reply to her tweet while others retweet it, producing again a cascade of information propagation. In these examples, the messages start on a single person as is usual in many posts and comments on social networks. Our study will focus on these cases, where a message is launched by a single user. In these cases, an interesting point to observe is how far the message can go throughout the network when sent by each of the users; and who are the most targeted users in the network. These are the aspects that we propose discussing in the present work. Along the text, we will use the word ”message” in a broad sense, representing a content launched by a member into the network. This paper considers a social network on a simple and oriented graph (or network) $G=(N,A)$, with $N=\\{1,\ldots,n\\}$ the set of nodes or individuals or members (homogeneous) and $A\subseteq\\{(i,j)\in N\times N:i\neq j\\}$ the set of arcs, representing the existing links among individuals. We denote by $\delta^{-}(j)$ and $\delta^{+}(j)$ the set of predecessors and the set of successor of $j$ in $G$, respectively. We also denote by $g^{-}(i)$ and $g^{+}(i)$ the in-degree and out-degree of node $i$, respectively, thus, $g^{-}(i)=|\delta^{-}(i)|$ and $g^{+}(i)=|\delta^{+}(i)|$. Arc $(i,j)\in A$ indicates that node $i$ influences node $j$. In the context of a social network, this means that individual $i$ is followed by individual $j$ (note that the orientation of the arc can seem contradictory, however, we are assuming that an individual influences its followers, as is usual). A node is said to be a _seed_ if it is the starting point of a message launched into the network, meaning that this node/individual is the launcher of a message. The _strength of influence_ that node $i$ exerts on $j$ is characterized by $d(i,j)$, for all $(i,j)\in A$; and the _hurdle_ of each node $j\in N$ to adopt the message is denoted by $h(j)$. Thus, following other authors notation, node $j$ is said to be _active_ (or _covered_) if it is a seed node or if $\sum_{i\in S\cap\delta^{-}(j)}d(i,j)\geq h(j)$, for $S\subseteq N$ the current subset of active nodes. This expression is denoted by _activating condition_. So, in an online marketing setting for promoting a product, a node/individual is active if it has adopted or promoted the product, otherwise it is inactive. In this study we assume that an active member is an adopter. A more detailed discussion about the roles of an influencer node is conducted in [Chen et al. 2013]. In addition, if the original graph is undirected, then each edge $\\{i,j\\}$ should be substituted by the two arcs $(i,j)$ and $(j,i)$. The activating condition previously described is based on the Linear Threshold (LT) model proposed in [Kempe et al. 2003], where a person will adopt a product/message if the influence received from its neighbors has reached a certain threshold (the hurdle value of the node). This condition follows similar expressions used on influence propagation and influence maximization problems in the literature, namely in [Chen 2009], and in very recent publications in [Fischetti et al. 2018], [Raghavan and Zhang 2019] and [Günneç et al. 2020]. All these works address propagation originated on a set of members and not on a single individual as in our case. Finding such a set of the smallest size that would lead the entire network to repass the message or adopt the product is known as the Target Set Selection (TSS) problem, discussed in [Chen 2009]. In the TSS, 100% adoption is required, the hurdle $h(i)$ of a node $i\in N$ is a value between 1 and its degree in $G$ and a node $i$ becomes active if it has at least $h(i)$ active neighbors, that is, equal influence is assumed. Later on, [Raghavan and Zhang 2019] considered a weighted version of the TSS, denoted by WTSS problem. In the WTSS, equal influence from neighbors and 100% adoption is still present, but the hurdle of a node is a value (a weight) that characterizes the amount of effort required by an individual (node) to be convinced to adopt the product. Equal influence and 100% adoption are no longer required on the Least-Cost Influence Problem (LCIP) discussed in [Günneç et al. 2020]. In the LCIP, the activating condition includes a tailored (i.e., partial) incentive (monetary in most cases) on each node, exerting influence on that node together with the usual influence employed by its predecessors to promote adoption. The LCIP seeks minimizing the sum of all tailored incentives provided to individuals in a social network while ensuring that a given fraction of the network is influenced. This problem was also addressed in [Fischetti et al. 2018], considering a nonlinear influence structure, besides the usual settings that characterize the LCIP. An extensive survey on influence propagation in social networks is described in [Peng et al. 2018]. It encompasses various types of social networks, their properties, social influence, state-of-the-art evaluation metrics and models, and an overview of known methods for influence maximization. The interpretation of the strength of influence $d(i,j)$ that individual $i$ exerts on $j$ is not a theme of consensus, as observed in most works previously reported on influence propagation and influence maximization. In fact, it is not easy to set a function for characterizing the influence parameters $d(i,j)$, neither the hurdle of a node as in most cases it is individual dependent, characterized by the message/product and it may also vary along the time. In our case, we set these parameters as functions of the nodes’ degrees, and thus defining the activating condition as a linear function of nodes’ degrees. This way, we do not have to assess influence strength and nodes’ hurdles for each particular instance. Instead, the activating conditions are automatically build, translating the inherent topological nature of the network for characterizing influence propagation, with an additional single parameter to assess the viral power of the message/product involved. Thus, in the present work we consider that influence strength of an individual can be related with the number of other individuals that he/she directly exerts influence on. In this case, we define this influence as its out-degree if working on an oriented graph, or degree if the graph is undirected. The same way, the hurdle can be associated with the influence strength of the individual (its out degree) and with the influence/viral power of the message (parameter $\alpha$, denoted by _hurdle coefficient_). It acts as a threshold for this individual to become active, that is, to adopt a product or to repass a message. Thus, we define: * • $d(i,j)=g^{+}(i)$, for all $(i,j)\in A$, that is, $d(i,j)$ is the out-degree of node $i$, indicating that the strength of influence that $i$ exerts on $j$ is defined by the number of individuals that $i$ can influence, that is, the number of followers of node $i$; and * • $h(j)=\alpha\cdot g^{+}(j)$, for all $j\in N$, where $\alpha$ is the hurdle coefficient, used to leverage the difficulty to activate node $j$, being related with the viral power of the message sent by the seed; and $g^{+}(j)$ is, again, node´s $j$ own strength. Hence, for any node $j\in N$ and $S\subseteq N$ the current subset of active nodes, the activating condition becomes: $\sum_{i\in S\cap\delta^{-}(j)}g^{+}(i)\geq\alpha\cdot g^{+}(j)$ (1) According to the previous definition of the influence that a node $j\in N$ receives from their direct neighbors (characterized by parameters $d(i,j)$, for all $(i,j)\in A$), equal influence from predecessor neighbors is not present, except if $G$ is regular, which is quite unlikely in a social network. Equal influence from predecessors was considered on the TSS and WTSS problems previously mentioned, being unavoidable when privacy concerns are present in the network. In addition, the hurdle is also node dependent as in the WTSS and LCIP problems, indicating that different nodes require different levels of effort to become active. In this paper, we assume that this hurdle is proportional to the node’s out-degree (its number of followers), as mentioned above. The relevancy of the number of followers is also stressed in the literature (see, e.g., [Huang et al. 2012, Bakshy et al. 2011]) and widely used on a number of public social activities, namely to select marketing influencers, or people to TV shows, or other public exhibitions, which motivates our option. This way, all the influential process is based on the topological nature of the graph, using just the degree information of the nodes in the entire activating condition. This excludes other external incident features, like monetary incentives, user specific characterizations of the influence among individuals or specific characterizations of the hurdle. This paper proposes two new indices for analysing message propagation in social networks. These indices are based on two new concepts: * • the _Individual Launching Power_ (ILP) of node $i\in N$, denoted by $ilp^{\prime}(i)$, representing the number of activated nodes in $G$ when $i$ is the launcher of a message; and * • the _Individual Target Potential_ (ITP) of node $i\in N$, denoted by $itp^{\prime}(i)$, representing the number of times that node $i$ is activated when each of the other nodes $j\in N\setminus\\{i\\}$ launch their own messages. Thus, we define the following two indices, respectively: * • the ILP index of $i$: $ilp(i)=\frac{ilp^{\prime}(i)}{(n-1)}$, for all $i\in N$; and * • the ITP index of $i$: $itp(i)=\frac{itp^{\prime}(i)}{(n-1)}$, for all $i\in N$. Index $ilp(i)$ represents the potential strength of individual $i$ as a launcher of a message, while index $itp(i)$ is the potential of individual $i$ as a receiver (target) of a message sent from the nodes of $G$. Launcher individuals should correspond to strong influencers and target individuals should identify the best target consumers or the most prominent message collectors. These indices are assuming that all nodes are receptive for the message/product and all nodes have equal chance to be the seeds of a message/product. There are three relevant issues involved on message propagation: i) the launcher strength, ii) the message power, and iii) the network topological structure. The launcher strength can be set by the ILP index, the message influence/viral power is characterized by the hurdle coefficient $\alpha$ and the network structure is the particular topology of graphs $G$. Note that the relevancy of classifying the viral power of the message was also stressed in [Bakshy et al. 2011] on the cascade size of influence propagation. To exemplify, consider the oriented graph $G1$ in Figure 1a, with 9 nodes and 19 arcs. graph $G1$ | | $\alpha=1.5$ | | $\alpha=2.0$ ---|---|---|---|--- | | | | (a) | | (b) | | (c) Figure 1: Oriented graph $G1$ in (a) and the active nodes when node 3 is the seed, for $\alpha=1.5$ and $\alpha=2.0$, in (b) and (c), respectively. Table 1 shows the ILP and ITP indices’ values for all the nodes in $G1$, for $\alpha=1.5,2.0\mbox{ and }2.5$. It also includes the out-degree of each node. | | $ilp(i)$ | $itp(i)$ ---|---|---|--- node $i$ | $g^{+}(i)$ | $\alpha=1.5$ | $\alpha=2.0$ | $\alpha=2.5$ | $\alpha=1.5$ | $\alpha=2.0$ | $\alpha=2.5$ 1 | 1 | 0.000 | 0.000 | 0.000 | 0.125 | 0.125 | 0.125 2 | 1 | 0.000 | 0.000 | 0.000 | 0.125 | 0.125 | 0.125 3 | 5 | 0.875 | 0.625 | 0.500 | 0.000 | 0.000 | 0.000 4 | 1 | 0.000 | 0.000 | 0.000 | 0.375 | 0.375 | 0.375 5 | 0 | 0.000 | 0.000 | 0.000 | 0.250 | 0.250 | 0.250 6 | 5 | 0.625 | 0.625 | 0.625 | 0.000 | 0.000 | 0.000 7 | 3 | 0.375 | 0.125 | 0.125 | 0.250 | 0.250 | 0.125 8 | 1 | 0.000 | 0.000 | 0.000 | 0.500 | 0.250 | 0.125 9 | 2 | 0.125 | 0.125 | 0.000 | 0.375 | 0.125 | 0.125 Table 1: ILP and ITP indices for graph $G1$, including the nodes’ out-degree ($g^{+}(i)$) In this example and considering a message launched in the network by node 3, if the hurdle coefficient is $\alpha=1.5$, representing a rather viral message, then it can reach all nodes but the 6, covering 87.5% of the entire network (excluding node 3), represented in Figure 1b. In this case, node 7 is made active just by influence of node 3. Then, in turn, node 7 can influence node 9, which will then influence node 8. Alternatively, if node 3 releases a less viral message, with hurdle coefficient 2.0, then it only reaches 62.5% of the other nodes, shown in Figure 1c. In this case, node 7 is activated through the joint influence of 3 and 4; and nodes 8 and 9 are no longer influenceable. However, if the hurdle coefficient is $\alpha=2.5$, the message sent by 3 will reach only nodes 1, 2, 4 and 5, covering just 0.5% of the nodes in $N\setminus\\{3\\}$. Besides, note that although nodes 3 and 6 have the same out-degree ($g^{+}(3)=g^{+}(6)=5$), their influence power in the entire graph is different due to the influential cascade each one generates. In effect, in this case, node 3 can go further than node 6. On the other hand, if observing the $itp(i)$ indices, representing the targeting potential of each node $i\in N$, when the message is very viral, with $\alpha=1.5$, node 8 can be reached by half of the other individuals. However, when the message is less viral, with $\alpha=2.0$ or $\alpha=2.5$, the easiest individual to reach is node 4, which can be targeted by 3 of the remaining individuals (37.5%). If graph $G1$ is undirected, by substituting each arc by an edge between the same nodes, the connectivity in the network may seem extended. However, considering again node 3 as the seed of a message, the ILP indices are $ilp(3)=1$ for $\alpha=1.5\mbox{ and }2.0$, and $ilp(3)=0.375$ when $\alpha=2.5$. While for $\alpha=1.5\mbox{ and }2.0$ the influence capability increases; it shrinks instead when the message becomes less viral ($\alpha=2.5$). The reason for the influence reduction in the last case is related with the hurdle increase in most nodes, due to the out-degree growth, making some nodes harder to cover when the message is less viral. The various social networks available these days vary quite a lot concerning oriented and undirected graph topologies. For example, Facebook and LinkedIn should be represented by undirected graphs, as each connection requires both users to accept the link. On the other hand, Instagram and ResearchGate, for instance, should be described using directed graphs, as user A can decide to follow B, but B may not be interested in following A. The contribution of the present paper involves the following three aspects: * • We propose two new indices especially devoted for identifying strong influencers and strong targeting nodes in a social network, through the ILP and the ITP indices, respectively. These indices incorporate a hurdle criteria on each node to model its hardness to forward the message, while most link topological measures in the literature are entirely build upon the relevancy of neighbors, ignoring the nodes’ ability to decide on a message propagation scheme. * • The construction of the new indices uses just the underlying topological nature of the graph and a hurdle coefficient ($\alpha$) to characterize the viral strength of the message/product. * • The ILP index results divide the nodes into launchers and non-launchers. The launchers class can be used to assist other metrics to find the best influencers, providing better decisions. The ITP results divide the nodes into targets and non-targets, allowing to identify strong consumers of information that may possibly magnetize malicious information in the network. The paper is organized as follows: an algorithm for constructing the ILP and ITP indices is described in the next section; computational tests are conducted in Section 3; and the paper ends with a conclusions section (Section 4). ## 2 Algorithm for calculating the ILP and ITP indices The construction of the two indices (ILP and ITP) is made in the same algorithm, on a given simple oriented graph $G=(N,A)$. It calculates the $ilp(i)$ index for each node $i\in N$ and updates the $itp(i)$ values along the calculation of the ILP indices. So, starting from a given node $i\in N$ and for a given input parameter value $\alpha$, the algorithm generates a sequential cascade of newly active nodes through influence propagation, using the activating condition set in (1). Its pseudocode is depicted in Figure 2. Before, we describe the sets and the variables involved. $S$ | is the current subset of activated nodes ($S\subset N$) ---|--- $L$ | is the current list of nodes to analyse (still non activated nodes) $\delta^{-}(i)$ | is the set of nodes converging to $i$ in $G$, for all $i\in N$ $\delta^{+}(i)$ | is the set of nodes diverging from $i$ in $G$, for all $i\in N$ $g^{+}(i)$ | is the out-degree of node $i$ in $G$, for all $i\in N$ (that is, $g^{+}(i)=|\delta^{+}(i)|$) $\alpha$ | is the hurdle coefficient | Algorithm ILP-ITP($\alpha$) | ---|---|--- | Input: $G=(N,A)$ and $\alpha$ | | Output: $ilp(i)$ and $itp(i)$ for all $i\in N$ | 1 | for all $i\in N$ do | 2 | $itp(i)\leftarrow 0$ | //_initialize vector itp_ // 3 | end_do | 4 | for all $i\in N$ do | 5 | $S\leftarrow\\{i\\}$, $L\leftarrow\\{j\in N\setminus S:(i,j)\in A\\}$ | //_initialize the sets_ // 6 | while ($L\neq\emptyset$) do | 7 | $v\leftarrow\arg\min_{j\in L}\\{g^{+}(j)\\}$ | //_select the next candidate to activate_ // 8 | if $\left(\sum_{j\in S\cap\delta^{-}(v)}g^{+}(j)\geq\alpha\cdot g^{+}(v)\right)$ then | //_if it becomes active, then_ // 9 | $S\leftarrow S\cup\\{v\\}$ | //_put the newly activated node in $S$_// 10 | $L\leftarrow L\cup\\{j\in N\setminus S:(v,j)\in A\\}$ | //_add to $L$ all inactive successors of $v$_// 11 | $itp(v)\leftarrow itp(v)+1$ | //_$i$ is also able to activate node $v$_// 12 | end_if | 13 | $L\leftarrow L\setminus\\{v\\}$ | //_remove node $v$ from $L$_// 14 | end_while | 15 | $ilp(i)\leftarrow|S|-1$ | //_$|S|$ is the number of nodes that node $i$ can_ | | _activate, excluding itself_ // 16 | end_do | 17 | $ilp(i)\leftarrow\frac{ilp(i)}{n-1}$ and $itp(i)\leftarrow\frac{itp(i)}{n-1}$ for all $i\in N$ | 18 | return $ilp(i)$ and $itp(i)$ for all $i\in N$ | Figure 2: Algorithm for computing the ILP and ITP indices. The algorithm starts initializing the $itp(i)$ variables, for all $i\in N$. Then, for each node $i\in N$, it generates a cascade of influences, trying to activate the nodes placed in the candidates list $L$. So, $i$ is the first node to enter the active nodes list $S$ and the activating candidates list $L$ is initialized with all the successors of $i$ in $G$. Then, while list $L$ is nonempty, we take from $L$ the node ($v$) with the lowest out-degree in $G$ and test it for activation. If $v$ passes the test, then we include it in $S$ and put in $L$ all the inactive successors of $v$. In addition, and if $v$ becomes active, we increase variable $itp(v)$ in one unit, meaning that one additional node ($i$) is able to activate $v$ through influence propagation. Then, node $v$ is removed from $L$, whether it entered $S$ or not. When $L$ becomes empty, the process terminates for the influential search started in node $i$, allowing to set the $ilp(i)$ result, which is equal to the number of nodes that were made active, excluding node $i$, divided by $n-1$. Each new candidate to activate taken from list $L$ is selected according to the minimum out-degree of the nodes in $L$. The advantage of this option is to promote a fast activation of the easiest nodes. The execution times increase when priority is given to the nodes with largest out-degree. ## 3 Computational tests and discussion In this section we use the ILP-ITP($\alpha$) algorithm for computing the ILP and ITP indices on a number of randomly generated and real-world graphs. We start describing the instances in the first subsection; then in Subsection 3.2, we use a small known example to compare the ILP and ITP indices with other metrics usually considered in the literature. The computational results and comments on larger sized graphs are conducted in Subsection 3.3. All tests were run on an Intel Core i7-2600 with 3.40 GHz and 8 GB RAM. The experiments were performed under Microsoft Windows 10 operating system. The algorithm described in Figure 2 was coded in Fortran and compiled on gfortran. Times are reported in seconds. ### 3.1 Instances In the tests conducted in this section, we consider two classes of instances: randomized and real-world. Randomized instances represent small-world networks, following the methodology described in [Watts and Strogatz 1998]. Real-world instances are taken from known online repositories and used in a number of publications addressing virtual social networks. The repositories are: the Stanford Large Network Dataset Collection (SNAP, [Leskovec and Krevl 2014]), the Koblenz Network Collection (KONECT, [Kunegis 2017]), and from the Social Networks Security Research Group from the Ben-Gurion University of the Negev (BGU; [Lesser et al. 2013]). All these instances are described next. Small-world randomly generated instances: The randomly generated instances are classified by nodes’ average degree and sparsity. All instances involve $n=10,000$ nodes. Their initial average degrees are 10, 20 and 50. We follow the methodology described in [Watts and Strogatz 1998] for generating small-world networks. Depending on the construction process, these graphs can have social network properties, as described in [Barabási 2016, Chap. 3] and [Günneç and Raghavan 2017]. The small-world construction of each network is initially made for the undirected version. Then, we use the undirected graph to build two additional oriented versions by substituting some of the edges by an arc in one of the two directions, while substituting the remaining edges by the two associated oriented arcs. The two oriented versions involve the substitution of $(2/3)m$ and $(1/3)m$ edges by a single directed arc, for $m$ the number of edges in the undirected counterpart. The edges selected to be oriented are taken at random. These networks are denoted by WS-_k_ -_o_ , with $k$ representing the initial average nodes’ degrees ($k=10,20\mbox{ and }50$) and $o$ indicating the proportion of bidirected links between pairs of nodes (edges in the original graph) that will remain, taking values $o=1.00,0.66\mbox{ and }0.33$). In version $o=1.00$, all links are bidirected, representing the undirected graph; for $o=0.66$, the graph is oriented, keeping 66% of the initial set of edges (as bidirected arcs) and 33% of unidirected arcs; and for $o=0.33$ the graph is also oriented, keeping 33% of the initial set of edges (as bidirected arcs) and 66% unidirected arcs. The Watts and Strogatz procedure starts with a regular graph with nodes’ degree _k_. Each edge of the graph is rewired, being reconnected, with probability _p_ , to another node chosen uniformly at random (duplicate edges are forbidden). The process is repeated for all original edges. Following a number of works in the literature, also addressing social networks (see, .e.g, [Günneç and Raghavan 2017, Fischetti et al. 2018, Raghavan and Zhang 2019]), we consider the most destructive rewiring probability in the recommended range ($0.1\leq p\leq 0.3$), setting $p=0.3$. This rewiring probability represents most closely the social networks studied by Watts and Strogatz ([Watts and Strogatz 1998]). Table 2 summarizes the main characteristics of the WS randomly generated graphs. The information concerning nodes’ degrees and density is adapted to the correspondent version of the graph, considering the degrees’ values for the undirected versions and the out-degrees ($g^{+}(i)$) for the oriented cases. | | | | degree/out-degree | ---|---|---|---|---|--- | nodes ($n$) | edges/arcs ($m$) | density | min | average | max | type WS-10-33 | 10,000 | 66,500 | 0.0007 | 0 | 6.65 | 15 | oriented WS-10-66 | 10,000 | 83,000 | 0.0008 | 2 | 8.30 | 16 | oriented WS-10-100 | 10,000 | 50,000 | 0.0010 | 5 | 10.00 | 18 | undirected WS-20-33 | 10,000 | 133,000 | 0.0013 | 3 | 13.30 | 25 | oriented WS-20-66 | 10,000 | 166,000 | 0.0017 | 6 | 16.60 | 28 | oriented WS-20-100 | 10,000 | 100,000 | 0.0020 | 12 | 20.00 | 30 | undirected WS-50-33 | 10,000 | 332,500 | 0.0033 | 17 | 33.25 | 50 | oriented WS-50-66 | 10,000 | 415,000 | 0.0042 | 20 | 41.50 | 59 | oriented WS-50-100 | 10,000 | 500,000 | 0.0100 | 38 | 50.00 | 66 | undirected Table 2: Main characteristics of the WS randomly generated graphs Real-world instances: $\bullet$ Zachary karate club [Zachary 1977]: Data source: http://vlado.fmf.uni-lj.si/pub/networks/data/ucinet/ucidata.htm This dataset was collected from the members of a university karate club by the sociologist Wayne Zachary in 1977 [Zachary 1977]. Each node represents a member of the club and an edge between two members indicates that they are connected, generating an undirected graph. Each club member knew all the others, but the network only represents links between members who regularly interact outside the club. This network is widely used in a number of papers in the literature. Most of these works try to find the two groups of people into which the karate club split after an argument between the president (John A.) and an instructor (Mr. Hi). We use this network to compare the ILP and ITP indices with other known metrics in the literature. $\bullet$ Konect - Advogato network [Massa et al. 2009]: Data source: http://konect.uni-koblenz.de/networks/advogato The Advogato trust network is build using the Advogato online community platform for developers of free software. The original dataset has 3992 loops that were removed. The resulting graph is oriented, with 5,155 nodes (Advogato users) and 47,135 arcs (trust relationships, called a ”certification” on Advogato). $\bullet$ Konect - Hamsterster network [Kunegis 2013]: Data source: http://konect.uni-koblenz.de/networks/petster-friendships-hamster This Network is based on friendships between users of the website hamsterster.com. The network is undirected, with 1,858 nodes (users) and 12,534 edges (friendships). $\bullet$ SNAP - ego-Facebook [Leskovec and McAuley 2012]: Data source: https://snap.stanford.edu/data/index.html This network represents social circles (circles of friends) from Facebook (anonymized), collected from survey participants using this online social network. Nodes are Facebook users and edges represent interactions between users. The dataset includes users’ profiles, circles and ego networks. The resulting graph is undirected, with 4,039 nodes (users) and 88,234 edges (interactions). The anonymized process permits relating users by their affiliations but does not allow to identify those affiliations. $\bullet$ SNAP - email-EU-core network [Yin et al. 2017, Leskovec et al. 2007]: Data source: https://snap.stanford.edu/data/email-Eu-core.html This network was generated using email data from a large European research institution. It used anonymized information about all incoming and outgoing email messages between members. The resulting graph is oriented, with 1,005 nodes (members of the research institution) and 25,571 arcs, where an arc $(i,j)$ in the graph indicates that member $i$ sent at least one email message to $j$, considering just email messages shared between members (the core). $\bullet$ SNAP - CollegeMsg temporal network [Panzarasa et al. 2009]: Data source: https://snap.stanford.edu/data/CollegeMsg.html This network involves an online social network at the University of California, Irvine. It is a temporal network based on private messages shared among members, derived from a dataset hosted by Tore Opsahl [Panzarasa et al. 2009]. Users could search the network for others and then initiate conversation based on profile information. The resulting graph is oriented, with 1,899 nodes (members) and 20,296 arcs, involving 59,835 messages shared along a given time frame. An arc $(i,j)$ means that user $i$ sent at least one message to $j$ within a given time frame. $\bullet$ BGU - Ning network [Lesser et al. 2013]: Data source: http://proj.ise.bgu.ac.il/sns/ning.html Ning is a very large online community building platform for people and organizations to create social networks. This particular network is a snapshot of the friendship and group affiliation networks from Ning, harvested during September 2012. The resulting graph is oriented, with 10,298 nodes (members) and 76,262 arcs. It has 5,512 pairs of nodes linked by a single arc. The remaining arcs represent bidirected links among pairs of nodes. We have removed one loop from the original dataset. We have switched all arcs’ orientation in the real-world networks because a link from node $i$ to node $j$ in the original dataset indicates that $i$ follows $j$, that is, $j$ is followed by $i$. Thus, according to the definition of the influence graph $G$ introduced in Section 1, this link should be represented by arc $(j,i)$, that is, node $j$ influences node $i$. Table 3 summarizes the main characteristics of the graphs, using the same column labels considered in Table 2, besides ”source”, representing the online repository dataset. | | | | | degree/out-degree | ---|---|---|---|---|---|--- | source | nodes ($n$) | edges/arcs ($m$) | density | min | average | max | type Zachary karate club | Konect | 34 | 78 | 0.1390 | 1 | 4.59 | 17 | undirected Advogato | Konect | 5,155 | 47,135 | 0.0018 | 0 | 9.14 | 721 | oriented Hamsterster | Konect | 1,858 | 12,534 | 0.0073 | 2 | 13.49 | 272 | undirected ego-Facebook | SNAP | 4,039 | 88,234 | 0.0108 | 1 | 43.69 | 1045 | undirected email-EU | SNAP | 1,005 | 25,571 | 0.0253 | 0 | 25.44 | 212 | oriented CollegeMsg | SNAP | 1,899 | 20,296 | 0.0056 | 0 | 10.69 | 137 | oriented Ning | BGU | 10,298 | 76,262 | 0.0007 | 0 | 7.41 | 633 | oriented Table 3: Main characteristics of the real world instances used in the computational tests An important difference observed among the WS randomly generated instances and the real world examples here considered, is the range of variation of nodes’ degrees. Although the average settings are comparable, the differences between minimum and maximum degrees in the real world examples are much larger, which is closer to the usual behavior of a social network. ### 3.2 Other metrics The most usual metrics in the literature and in real-life problems belong to two classes: centralized and link topological metrics (see, e.g., [Kiss and Bichler 2008, Peng et al. 2018]). Centralized metrics characterize the spread capabilities of the nodes and also describes nodes’ proximity to the other players in the network; while link topological metrics emphasize important neighbors, benefiting from their relevancy. These metrics are also included in the measures list considered in Gephi [Bastian et al. 2009]. The lists in these classes are vast, particularly in the centralized group. In the present paper, we consider the selection described next. Before, we must introduce the concept of _path length_ between two nodes, denoted by $p_{ij}$ for all pairs in $\left\\{\\{i,j\\}:i,j\in N\mbox{ and }i\neq j\right\\}$, where $p_{ij}$ is the length (number of arcs or edges) in the shortest path between $i$ and $j$ in $G$. Centralized metrics: $\bullet$ degree centrality ([Foulds 2012, Peng et al. 2018]): The degree of a node $i\in N$ is the number of links incident on $i$. If the graph is oriented, we consider just its out-degree, as defined in Section 1, that is, $g^{+}(i)=|\delta^{+}(i)|$, related with the strength of influence and the hurdle of a node, considered in (1). We use the notation $g(i)$ when addressing the undirected case. Nodes with higher degree (more friends or more followers) are considered to be more influential. $\bullet$ eccentricity ([Hage and Harary 1995, Foulds 2012]): The eccentricity of a node $i\in N$ is the maximum path length between $i$ and any other node in the graph, that is, $ec(i)=\max_{j\in N}p_{ij}$. The central nodes in the graph are those with the lowest eccentricity. $\bullet$ closeness centrality ([Okamoto et al. 2008]): The closeness centrality of node $i\in N$ is the inverse of the average path length between $i$ and all the other nodes in the graph, that is, $cc(i)=\frac{n-1}{\sum_{j\in N\setminus\\{i\\}}p_{ij}}$. Nodes with larger closeness centrality can reach sooner the other nodes in the graph, on average, being better positioned to spread the information. $\bullet$ betweenness centrality ([Freeman 1977, Brandes 2001, Boccaletti et al. 2006]): The betweenness centrality (or load) of node $i\in N$ represents a measure of the number of shortest paths passing through node $i$, being defined by $bc(i)=\sum_{s,t\in N\setminus\\{i\\},s\neq t}\frac{\sigma_{st}(i)}{\sigma_{st}}$, with $\sigma_{st}(i)$ representing the number of shortest paths between $s$ and $t$ passing through node $i$, and $\sigma_{st}$ representing the number of shortest paths between $s$ and $t$ in $G$. Nodes with larger betweenness centrality are in the ”preferable” path between many pairs of nodes, acting as bridges. Under the assumption that information flows through shortest paths, these nodes are better placed to control most communications’ traffic in the graph. In addition, the betweenness centrality value of node $i$ also indicates the disruption induced in the graph when $i$ is removed. Link topological metrics ([Peng et al. 2018]): $\bullet$ eigenvector centrality ([Bonacich 2007, Lü et al. 2016]): The eigenvector centrality of node $i\in N$ measures the level of influence of node $i$ based on the influential strength of their direct neighbors, being denoted by $ev(i)$. This means that the spreading ability of $i$ is enforced if it is connected to other influential nodes. In this case, a person $i$ with few connections can have a high $ev(i)$ if those connections are also well- connected with others in the network. Nodes with higher eigenvector value are expected to be more influential in the network. The eigenvector metric is the same as degree centrality if all neighbor nodes have equal degree, namely on regular graphs. The eigenvector centralities can be computed iteratively through the expression $ev(i)=\frac{1}{\lambda}\sum_{j\in\delta^{-}(i)}ev(j)$, that is, $Ax=\lambda x$, where $A$ is the adjacency matrix, $x$ is the $ev$ vector and $\lambda$ is the largest eigenvalue in absolute value of $A$ ($\lambda\neq 0$), related to the principal eigenvector of $A$. $\bullet$ PageRank ([Brin and Page 2012, Lü et al. 2016, Liu et al. 2017]): The PageRank index, proposed by Sergey Brin and Lawrence Page [Brin and Page 1998, Brin and Page 2012], and used by Google’s search engine, is a measure of the relevancy of a given web page in a webgraph (oriented, in general). It is also used in other network environments ([Gleich 2015]), including social networks. As the eigenvector metric, this index is based on the number and relevancy of the hiperlinks a website receives. Thus, higher PageRank values correspond to more meaningful websites in a webgraph. The PageRank considers the number of inward arcs (i.e., converging links to a website) and the relevancy of the linkers (i.e., the PageRank of the adjacent converging websites). The PageRank index of a node $i$, here denoted by $pr(i)$ for all $i\in N$, is calculated using a recursive procedure, defined by $pr_{t}(i)=\beta\sum_{j\in\delta^{-}(i)}\frac{pr_{t-1}(j)}{g^{+}(j)}+(1-\beta)\frac{1}{n}$, where $\beta$ is the probability that a user clicks on a hyperlink on the current page; while $1-\beta$ is the probability of teleportation by typing the name of a web page of choice in the address bar, moving elsewhere. The procedure starts with $pr_{1}(i)=1$ for all $i\in N$ and terminates when the PageRank values become stable. The final result is $pr(i)=pr_{tt}(i)$ for all $i\in N$, with $tt$ the total number of iterations of the recursive procedure. It has good convergence properties when $\beta$ is not too close to 1 ([Gleich 2015]). The usually recommended value for $\beta$ is 0.85 (e.g., [Brin and Page 2012, Gleich 2015, Lü et al. 2016]). $\bullet$ HITS (Hyperlink-Induced Topic Search) ([Kleinberg 1999, Lü et al. 2016]): HITS was also developed to produce metrics on webgraphs, classifying websites into authoritative and hub pages, here denoted by $au(i)$ and $hu(i)$, respectively, for all $i\in N$. Authoritative pages are those with many incoming links from hub websites, exposing the value of the content of the page; while hub pages are those incident on many relevant authoritative pages. These two types of nodes are intimately related to each other, being calculated through a recursive procedure that should converge to stable authoritative and hub scores. The recursive expressions are: $au_{t}(i)=\sum_{\delta^{-}(i)}hu_{t-1}^{\prime}(j)$ and $hu_{t}(i)=\sum_{\delta^{+}(i)}au_{t-1}^{\prime}(j)$, with $au_{t}^{\prime}(i)=\frac{au_{t}(i)}{\sqrt{\sum_{k\in N}\left(au_{t}(k)\right)^{2}}}$ and $hu_{t}^{\prime}(i)=\frac{hu_{t}(i)}{\sqrt{\sum_{k\in N}\left(hu_{t}(k)\right)^{2}}}$ the normalized scores. The procedure starts with $au_{1}(i)=hu_{1}(i)=1$ for all $i\in N$; and terminates when the scores of all nodes reach the steady state, controlled by a given threshold, denoted by $\varepsilon$ and set to 0.0001 in our computational tests. The final values are $au(i)=au_{tt}(i)$ and $hu(i)=hu_{tt}(i)$ for all $i\in N$, with $tt$ the total number of iterations of the recursive procedure. The two metrics’ results are similar on undirected graphs. A good authoritative page should have many incoming links from many good hubs, that is, being linked from pages known as hubs for information; and a good hub page should point to many good authoritative pages, that is, linked to pages that are considered to be authorities on the subject. The nodes with higher authoritative value are the most central in the graph. Unlike other authors, we chose to include the eigenvector metric within the link topological class because it is build upon neighborhood influence, as PageRank and HITS. All these centralized measures are focused on direct indistinct neighbors or on the path length distance to the other nodes in the graph. They do not reveal information spread based on nodes’ influence strength over their neighbors. This influence strength over the neighbors is essential to model influence spread cascades, used to show each node’s influence capability over the entire network, performed by the ILP measure here proposed. The same way, the selected link topological metrics described above are based on direct neighborhood influence, distinguishing the relevancy of the various neighbors. These neighbors also benefit from their neighbors influence and so on, producing a kind of influence propagation towards the nodes. However, in this cases, the relevancy of a member is entirely build upon the relevancy of its neighbors, ignoring its own ability to decide on a message propagation scheme. In effect, message propagation capability depends on the strength of inward neighbors, but it then must break the node’s own hurdle to further propagate the message. This point establishes the main difference between the link topological metrics described above and influence propagation considered on Threshold and Cascade models, including the ILP and ITP indices architecture. Instead of just reading neighbors’ direct influence, the ILP index uses the network topology and the message viral power to truly simulate message propagation. On a different perspective, the ITP index describes each node capability as a consumer of information flowing in the graph. It may resemble the betweenness centrality measure, but instead of receiving the flow due to its geodesic position in the graph, it receives the flow by true influence propagation cascades started on each node in the graph, being also sensitive to the message viral power. To detach these differences, we use the Zachary karate club network described in Subsection 3.2, represented by an undirected graph. Table 8 in the Appendix shows the ILP and ITP indices’ values for all the nodes in the Zachary’s club graph, for $\alpha=1.5\mbox{ and }3.0$. It also includes all the results produced by the metrics described above, calculated using Gephi [Bastian et al. 2009]. Starting with the ILP index results, when $\alpha=1.5$, representing a rather viral message, for instance, an opinion about an adversary, there are 5 members able to spread out this opinion throughout the entire network (members: 1 (Mr. Hi), 2, 3, 33 and 34 (John A.)). The other members have low or null influence capability in the graph. This is a typical behavior of this index that separates the members in two groups: strong and low/null message launchers. However, when $\alpha=3.0$, representing a less viral message, for instance, a non-consensual opinion about one of the leaders (Mr. Hi or John A.), the message propagation should be harder because they are all members of the same club. In this case, only node 1 (Mr. Hi) and node 34 (John A.) can spread the message, being able to cover only 20 other members, when starting in each one of them, represented in Figure 3 (a) and (b), for Mr. Hi and John A., respectively. There is another member (node 33) that can also do some damage, being able to reach 9 other members, but clearly with lower strength compared to the two leaders. | | ---|---|--- (a) | | (b) Figure 3: ILP index solution with $\alpha=3.0$, for node 1 (in (a)) and node 34 (in (b)) as the origins, for the Karate club Zachary’s graph. The two solutions in Figure 3 also show that there are some members being reached by the two potential influencers: Mr. Hi and John A., namely members 3, 9, 10, 14, 20, 29, 31 and 32. Some of these members stayed with Mr. Hi (3, 9, 14, 20) and the others stayed with John A. (10, 29, 31, 32) after the club’s fission. Concerning the ITP index results, when the message is volatile ($\alpha=1.5$) there are two nodes (17 and 26) that are more influenceable than the others, being reached by 7 members of the club. These two members have great potential for being consumers or collectors of information in the network, which may suggest that they should be observed closely for all sort of message propagation in the graph, namely gossips or fake news. However, for messages with less spread capability ($\alpha=3.0$), namely classified information, the members to observe are no longer the former, but 9, 10, 20, 29 and 31. Curiously, they all belong to the subset of those influenced by both sides, as observed above. Compared to the other metrics, the ILP and ITP indices incorporate the capability to classify the message being spread. In effect, the viral strength of the message is a key issue for its spreadability, while the previously mentioned metrics make no distinction on this matter. If we observe the degree information of each node, it provides close answers for the more volatile message (with $\alpha=1.5$), but it helps less when the message has more friction (with $\alpha=3.0$), namely on nodes 2, 3 and 33. However, nodes’ degree can complement the ILP index information, revealing, for instance, that some lower degree members (2 and 3) are also able to cover the entire network when the message is volatile. These members can be easier to convince to act as influencers in a marketing campaign, for instance. In addition, among the members with the lowest eccentricity there are a number of them with low or null ability to spread the message, namely members 4, 9, 14, 20 and 32. Instead, John A. (member 34) has higher eccentricity. Also, closeness centrality detaches members 9, 14, and 32 with low average path length (below $\frac{1}{0.52}$) and very low or null capability to be message spreaders. On the other hand, member 17 has high average path length ($\frac{1}{0.28}$) while being a good consumer/collector of volatile messages when $\alpha=1.5$, with $ipt(17)=0.21$. Then, the betweenness centrality metric detaches members 3 and 32 as interesting bridges for the flow passing among nodes in the graph, but they have low capability to be good influencers, specially for less viral messages ($\alpha=3.0$). This is curious because they can be highly disruptive if removed from the graph. In addition, the stronger ability of the members as consumers/collectors (index ITP) is not followed by higher values of betweenness centrality. Now, considering the link topological metrics, they all bring similar results for this example, detaching members 1 and 34, and also members 2, 3 and 33 for their relevancy in the graph, which is directly related with the number of neighbors and their weight. However, they lack indicating how far this relevancy can go considering message propagation. In effect, they also stress nodes 4, 9, 14 and 32 (as the eccentricity and closeness centrality metrics) with moderate relevancy, although these members have low or null ability for message spreading in the graph. To conclude, these observations stress that the new indices should not be used to substitute the known metrics, but used in conjunction with those metrics to complement the decision process on influencers selection in social networks. In the next subsection we provide a few more observations involving the metrics here described using a larger sized graph. ### 3.3 Computational results This section describes the computational tests on the ILP-ITP($\alpha$) algorithm described in Section 2 for generating the ILP (Individual Launching Power) and ITP (Individual Target Potential) indices on two given classes of instances. These classes involve the randomized and real-world instances selected in Subsection 3.1. Table 4 reports the ILP and ITP values on the WS randomly generated graphs, involving $n=10,000$ nodes. The table includes the execution times (in seconds) of the algorithm and the percentage of nodes that are able to cover (activate) at least 99% of the remaining nodes in the graph, representing strong candidates to be influencers, using the ILP index results. The tests were conducted considering the following values for the hurdle coefficient: $\alpha=$ 1.0, 1.5 and 2.0, indicating the virality level of a message generated in each of the nodes. | | | percentage of strong influencers | execution times (sec.) ---|---|---|---|--- instance | density | type | $\alpha=1.0$ | $\alpha=1.5$ | $\alpha=2.0$ | $\alpha=1.0$ | $\alpha=1.5$ | $\alpha=2.0$ WS-10-33 | 0.0007 | oriented | 80.99 | 4.23 | 0.00 | 1853 | 70 | $<1$ WS-10-66 | 0.0008 | oriented | 86.57 | 14.57 | 0.35 | 2443 | 371 | 7 WS-10-100 | 0.0010 | undirected | 90.31 | 16.87 | 0.44 | 3472 | 552 | 13 WS-20-33 | 0.0013 | oriented | 91.70 | 21.01 | 5.32 | 4227 | 1730 | 231 WS-20-66 | 0.0017 | oriented | 94.23 | 23.05 | 1.66 | 4923 | 1214 | 89 WS-20-100 | 0.0020 | undirected | 94.87 | 6.77 | 0.03 | 7183 | 502 | 3 WS-50-33 | 0.0033 | oriented | 97.18 | 21.01 | 0.37 | 9808 | 2142 | 38 WS-50-66 | 0.0042 | oriented | 98.24 | 8.53 | 0.09 | 11023 | 993 | 11 WS-50-100 | 0.0100 | undirected | 97.87 | 0.14 | 0.00 | 16792 | 24 | $<1$ Table 4: Percentage of strong influencers and execution times of the ILP- ITP($\alpha$) algorithm on the WS randomly generated graphs, considering $\alpha=$ 1.0, 1.5 and 2.0. As expected, all the ILP index results observed using the WS graphs divide the nodes in two classes: launchers and non-launchers. The same way, the results with index ITP also found two classes of nodes: targets and non-targets. The breaking point that separates launchers from non-launchers represents the minimum proportion of nodes that launchers can cover. This value was observed to be higher than 0.9959 in all experiments reported in Table 4, except for instances WS-10-33 and WS-50-100 for $\alpha=2.0$. Considering the results in this table, when the message is very viral ($\alpha=1.0$), most nodes are able to act as strong launchers in the graph, especially when it becomes denser, as expected. As an example, suppose a social network (connected) of football (soccer) supporters. If we think of a very viral message in this network, for instance, ”Ronaldo returns to Manchester”, it can easily reach most nodes of the network if launched by any member, no matter its strength, in most cases. A similar result holds for a less viral message ($\alpha=1.5$) when $k=10$, but it suddenly changes when the graph becomes denser. In effect, for $k=20$ the number of strong launchers is lower in the undirected graph (WS-20-100) compared with the lower density instances (WS-20-33 and WS-20-66); and this behavior is even clearer for the $k=50$ instances, where the number of stronger launchers decreases along with the density increase of the graphs, which is somehow unexpected. This performance is even more noticeable for the less viral message cases, with $\alpha=2.0$. Actually, although density growth increases the number of connections, it also makes the nodes stronger in their own hurdle, turning message dissemination harder to pass. These small-world networks are particularly sensitive to this aspect, due to the homogeneity of their nodes’ degrees. As observed in the forthcoming tests, this low variation on nodes’ degrees is not typical in a social network, which may cast doubt on the suitability of small-world artificial graphs to simulate social networks. Now, considering the larger sized real-world instances proposed in Subsection 3.1, we show in Tables 6 and 7 the results of the ILP and ITP indices returned by the ILP-ITP($\alpha$) algorithm, considering the following values for the hurdle coefficient: $\alpha=$ 1.0, 1.5, 2.0 and 3.0. The algorithm was run for the entire graphs, despite the fact that most of them are not connected (or strong connected in the oriented cases). The sizes of these connected (or strongly connected) components (in percentage of nodes over the entire graph) and the execution times of the algorithm for the various hurdle coefficients are reported in Table 5. | | | execution times (in seconds) ---|---|---|--- instance | type | largest component (in %) | $\alpha=1.0$ | $\alpha=1.5$ | $\alpha=2.0$ | $\alpha=3.0$ Advogato | oriented | 60.91 | 12 | 6 | 4 | 2 Hamsterster | undirected | 96.23 | 3 | 2 | 1 | $<1$ ego-Facebook | undirected | 100.00 | 189 | 125 | 88 | 30 email-EU | oriented | 79.90 | 2 | 1 | $<1$ | $<1$ CollegeMsg | oriented | 68.14 | 2 | 1 | 1 | $<1$ Ning | oriented | 87.37 | 69 | 32 | 18 | 7 Table 5: Sizes (in percentage) of the largest connected (strongly connected) components in the real-world instances; and the execution times of the ILP- ITP($\alpha$) algorithm for $\alpha=$ 1.0, 1.5, 2.0 and 3.0. The execution time of the algorithm is influenced by the size and density of the graph, but also by the hurdle coefficient. In effect, when the hurdle coefficient increases, the number of launcher nodes diminishes, being reflected on a lower execution effort by the algorithm. In addition, the largest connected (or strongly connected) component has almost the initial graph size on the undirected instances, being smaller on the oriented counterparts, especially on the Advogato graph. Despite these differences, we chose to run and report the tests on the original graphs because it is closer to reality. Here again, all the ILP index results on the RW graphs divided the nodes in two classes: launchers and non-launchers. The same way, the results with index ITP found two classes of nodes: targets and non-targets. We consider again the breaking point that separates launchers from non-launchers as the minimum proportion that launcher members can cover, here denoted as _minimum influential breaking point_ (mibp in short). We also consider the breaking point that separates targeting from non-targeting nodes as the minimum proportion of nodes covering target nodes, denoted by _minimum targeting breaking point_ (mtbp in short). Thus, Table 6 reports the percentage of launcher nodes and the mibp values found in each instance and for each hurdle coefficient. Table 7 represents the percentage of targeting nodes and the mtbp values for the same instances and the same hurdle coefficients. | percentage of launchers | min influential breaking point (mibp) ---|---|--- instance | $\alpha=1.0$ | $\alpha=1.5$ | $\alpha=2.0$ | $\alpha=3.0$ | $\alpha=1.0$ | $\alpha=1.5$ | $\alpha=2.0$ | $\alpha=3.0$ Advogato | 21.51 | 11.70 | 7.29 | 3.16 | 0.7241 | 0.7115 | 0.6995 | 0.6814 Hamsterster | 32.88 | 20.61 | 12.70 | 5.71 | 0.9553 | 0.9467 | 0.9435 | 0.9413 ego-Facebook | 73.31 | 52.34 | 36.37 | 15.23 | 1.0000 | 0.9861 | 0.9861 | 0.9854 email-EU | 52.64 | 36.42 | 26.27 | 12.14 | 0.8137 | 0.8137 | 0.8118 | 0.8078 CollegeMsg | 25.43 | 12.06 | 6.79 | 2.05 | 0.6965 | 0.6907 | 0.6886 | 0.6797 Ning | 14.58 | 7.44 | 4.47 | 1.93 | 0.8757 | 0.8642 | 0.8527 | 0.8196 Table 6: ILP index results on the real-world selected instances, considering $\alpha=$ 1.0, 1.5, 2.0 and 3.0. The mibp percentage on these instances is lower, in general, compared to the WS instances results, especially on the Advogato and CollegeMsg datasets, probably due to the smaller size of the largest strongly connected components. On the other instances, the launcher nodes can cover more than 80% of the nodes. If observing the largest oriented graph (Ning), when the message is very viral ($\alpha=1.0$), 14.58% of the nodes (1501 members) can be classified as launchers and the nodes in this group can reach at least 87.57% (9018 nodes) of the entire set of members. The remaining 85.42% (non-launcher nodes) can reach no more than 83 other nodes in the graph. However, if the message becomes less viral, with $\alpha=1.5$, the successful launcher nodes falls to 7.44% (766 members) and each of these nodes can cover 86.42% (8900 nodes) of the graph, or more. The remaining 9532 non-launcher nodes can reach less than 184 other members in the graph. Further, if the message has low virality (with $\alpha=3.0$), the percentage of launcher nodes is 1.93%, representing only 199 members that are able to cover at least 81.96% (8440 nodes) of the graph. The remaining 98.07% members (nonlaunchers) can only cover at most 509 other nodes. An additional observation considering, for instance, the $\alpha=1.0$ case, the mibp of the launchers (0.8757, representing 9018 nodes) is bigger than the largest strongly connected component in that graph (87.37%, that is, 8997 nodes). The reason for this is that the message can propagate across nodes in and out of the largest strongly connected component, namely among nodes in weakly connected components. Considering a different case, if we observe the largest undirected graph (ego- Facebook network), which is entirely connected, it has 73.31% launcher members that are able to cover the entire graph through influence propagation if the message is very viral, with $\alpha=1.0$. The remaining 26.69% members can reach no more than 60 other nodes in the graph, thus, being non-launchers. Once again, if the virality of the message decreases, considering $\alpha=1.5$, the percentage of launcher nodes decreases to 51.34%, each of which being able to cover 98.61% of the graph; and if the message becomes harder to pass, with $\alpha=2.0$, the percentage of launcher nodes decreases further, to 36.37%, falling even deeper (15.23%) if the virality of the message is further decreased. These 15.23% launchers are 615 Facebook members that are able to reach (individually) 98.54% of the other nodes in the graph, at least. The detached launcher members able to cover almost the entire graph are still too much if we are intended to choose some of them to propagate a message or initiate a marketing campaign. Therefore, we propose complementing the information with some of the metrics discussed in Subsection 3.2. To assist on this discussion we show in Figure 4 an image with the ego-Facebook instance, where the nodes’ color (from red to light yellow) and size are proportional to their degree in the graph. The network was build using Gephi [Bastian et al. 2009]. Observing this image, a natural choice for strong influencers (as launchers) would be those with larger degree, according to the criteria used in the construction of the activating condition (1) described in Section 1. In fact, the five nodes with largest degree are 108, 1685, 1913, 3438 and 1, with degrees 1045, 792, 755, 547 and 347, respectively. These 5 nodes are also on top of the list for closeness centrality, betweenness centrality and PageRank, indicating that they are central on communication and neighbors influence. These are probably the most relevant players in the graph, but they should also be the more expensive if we think about a financial incentive to pay these members to support a marketing campaign or to decide sending a message. In effect, they all belong to the launchers’ list determined by the ILP index. However, and still thinking about the cost to pay to these members, are there strong launcher nodes that may cost less? A possible answer to this question can arise from the lower degree nodes, or other metric, still belonging to the launchers’ list. In effect, among the nodes in this list, we can find a number of members with degree below 30 (nodes 679, 3081, 3232 and 991) which may represent good candidates to act as launchers (influencers) in practice. Probably due to their position in the graph, these apparently weak members are so effective on influential spread as node 108 that exhibits the strongest degree ($g(108)=1045$) in the graph. Note that the average degree in this graph is 43.69 and the standard-deviation is 52.41. A curious aspect to mention involves node 568 that has the lowest eccentricity ($ec(568)=4$), node degree slightly above average ($g(568)=63$) and very high betweenness centrality (above 750,000), suggesting that it could be placed in a privileged position as an influencer. Yet, it is classified in the non-launchers class by the ILP index for virality level $\alpha=3.0$, as it is able to reach only 21 other nodes in the graph. So, the other known metrics may not be tailored for assessing influence propagation on their own, but their performance can be improved if used together with the ILP index information. Figure 4: ego-Facebook network, build using Gephi. Besides instances ego-Facebook and email-EU, the launchers’ groups on all other networks are relatively small. The number of members in these groups should fall below 1% when $\alpha>3.0$ in most of the studied graphs. This percentage is close to the number of nodes that are able to launch an effective influence cascade over the network, detached in [Goel et al. 2016] and based on a very extensive variety of contents launched on Twitter. As shown in our experiments, these proportions are strongly influenced by the viral power of the message ($\alpha$) and the network topological nature. | percentage of targets | min targeting breaking point (mtbp) ---|---|--- instance | $\alpha=1.0$ | $\alpha=1.5$ | $\alpha=2.0$ | $\alpha=3.0$ | $\alpha=1.0$ | $\alpha=1.5$ | $\alpha=2.0$ | $\alpha=3.0$ Advogato | 72.42 | 71.15 | 69.95 | 68.15 | 0.2150 | 0.1168 | 0.0728 | 0.0314 Hamsterster | 95.53 | 94.67 | 94.35 | 94.13 | 0.3285 | 0.2057 | 0.1265 | 0.0565 ego-Facebook | 100.00 | 98.61 | 98.61 | 98.54 | 0.7330 | 0.5233 | 0.3777 | 0.1521 email-EU | 81.39 | 81.39 | 81.19 | 80.80 | 0.5259 | 0.3635 | 0.2620 | 0.1205 CollegeMsg | 69.67 | 69.09 | 68.88 | 67.98 | 0.2540 | 0.1201 | 0.0674 | 0.0200 Ning | 87.57 | 86.42 | 85.27 | 81.96 | 0.1457 | 0.0743 | 0.0446 | 0.0192 Table 7: ITP index results on the real-world selected instances, considering $\alpha=$ 1.0, 1.5, 2.0 and 3.0. Now, concerning the ITP index results reported in Table 7, most nodes in the graphs belong to the targets group, with a slight exception on instances CollegeMsg and Advogato, possibly due to their lower connectivity properties. However, most of these targeting nodes are reachable by a relatively small number of members, except on the ego-Facebook and email-EU graphs. For instance, the targeting group in the ego-Facebook graph includes all members when the message is very viral ($\alpha=1.0$) and these members are reachable by at least 73.30% (2961 members) other nodes in the graph. In this case, there are no non-targeting members. However, when the message has low virality, with $\alpha=3.0$, the targeting group includes 98.54% members (3980 nodes), but these members are only reachable by 15.21% other nodes in the graph. The remaining 1.46% non-targeting members (59 nodes) can be reached by no more than 0.07% of the nodes, that is, at most 3 members. The targeting class, in particular, includes a large variety of members, all of them acting as message consumers (or collectors). This group of members may possibly concentrate the usual targets of fake news that can be used to start the tracking process of fake news’ origins. To further explore the selection within the targeting members’ group, we may use the betweenness centrality metric as an additional filter. If we focus again on the $\alpha=3.0$ virality level and still on the ego-Facebook network, the nodes with largest betweenness centrality value (larger than 1 million) are, again, members 108, 1685, 3438, 1913, 1086 and 1, among those in the targeting nodes’ class. Curiously, all these nodes are both good launchers and good targets. Also, they are among top degree members, except node 1086 ($g(1086)=66$), so with no novelty, they are strong players in the graph. However, they are also very heavy players that can belong to the expensive nodes’ class. To conclude, an interesting and expected observation in these tests shows that the percentage of launchers is almost the same as the mtbp proportion; and the percentage of targets is also close to the mibp proportion. This illustrates that the nodes able to reach the target members are basically the launchers; and the nodes that are reached by the launchers are mostly the targets. ## 4 Conclusions In the present paper, we have considered an entirely deterministic process for characterizing adoption and influence using an activating condition based on the Linear Threshold (LT) model. The activating condition uses the topological information of the graph, namely the nodes’ out-degree (or degree if it is undirected) and a single parameter - the hurdle coefficient - for classifying the viral propagation strength of the message/product under consideration. Thus, it does not depend on personal information of the users and hence can be applied easily in practice. Based on this process, we have proposed an algorithm that produces two influence propagation indices for online social networks: the ILP (Individual Launching Power) and the ITP (Individual Target Potential). The ILP provides a clear division of the nodes into launchers and non-launchers, with very low distinction inside each group. Each of the launchers can cover most of the graph through influential cascades, reaching more than 70% of the members in most social networks used in our tests. The size of the launchers’ group diminishes significantly with the hurdle coefficient increase, reflecting the natural virality variation of the message or marketing campaign (message virality weakens with the hurdle coefficient increase). It also depends on the social network topology. When the message has low virality (with $\alpha>3.0$), the launchers group’s size is possibly lower than 1%, being in line with [Goel et al. 2016] that observed that less than 1% of the influential cascades are able to pass beyond the very next neighbors. This partition of the nodes into launchers and non-launchers can be used as a first filter for the selection of the best candidates for influence propagation. Then, we can use other metrics in the literature to choose the members with the best characteristics. In effect, if we choose influencers on a social network without this filter, considering, for instance, just the nodes’ degree information, we would possibly tend to select the members with largest degree (individuals with more followers). However, those are also typically the more expensive if used in a marketing campaign or to pass a message. As we have observed in the computational tests here conducted, there are other members in the launchers’ group that are able to reach a similar performance as the mentioned very strong members, but having a significantly lower degree (having much less followers). These lower degree influencers are probably much less expensive, being able to perform almost as well as the highest degree members on the given network. So, instead of searching for the largest degree nodes, we recommend looking for the lower degree ones but belonging to the launchers’ group, earlier found by the ILP index. These are probably the ”ordinary influencers” (individuals who exert average or even less-than-average influence) observed in [Bakshy et al. 2011] when studying influence propagation on Twitter. This process can be conducted using other centralized or link topological metrics besides nodes’ degree, depending on the kind of members we are looking for. We have also observed that the ILP index information obtained from artificial small-world networks (instances WS, generated according to [Watts and Strogatz 1998]) produce lower sized launchers’ groups when the density of the graph increases, especially when the hurdle coefficient is higher (lower virality messages). Although higher density graphs have more connections, offering more chances for message propagation; it also makes the nodes stronger in their own hurdle, which turns them harder to collaborate on message transmission. This last observation may justify the mentioned unexpected behavior found on the small-world graphs, which is possibly justified by the low variation of their nodes’ degrees. This low variation is not typical on social networks, which may cast doubt on the suitability of small-world artificial graphs to simulate social networks. The ITP index, instead, divides the nodes into targeting and non-targeting members. According to our tests, most nodes in the graphs belong to the targets group (more than 70% on most social networks). However, each of those targeting nodes are targeted by a small number of members (no more than 25% of the nodes, in most cases), falling sharply with the hurdle coefficient increase, that is, when the message becomes less viral. These targeting nodes can represent compulsive consumers of information in online networks, which may include fake news’ easy targets. As mentioned above, the targeting group is very large, so, once again, we can use it as a first filter and then resort to other available metrics to assist on the search for specific members selection profiles. On another perspective, the ILP and ITP indices could be used to restrict the set of candidate seed nodes in most Influence Maximization problems based on the Linear Threshold model. That restricted subset should focus on strong influencers (launchers) with low chance of sharing common nodes in their propagation cascades. An additional aspect to stress involves the choice of adequate values for the hurdle parameter $\alpha$. The present work introduces a brief discussion on this matter, but further work is needed, specially on real-practice environments for adequately tuning this parameter. In the meanwhile, we recommend considering sensitivity analysis on $\alpha$ in any new real-world instance. In future works, the ILP and ITP indices can be discussed using the Independent Cascade model [Kempe et al. 2003] instead of the Linear Threshold model here considered. Also, and in this line, instead of assuming that all individuals are equally receptive for a message/product to be launched, we could consider these assumptions to be ruled stochastically. ## Acknowledgements Pedro Martins acknowledges support from the Portuguese National Funding: Fundação para a Ciência e a Tecnologia - FCT, under the project UIDB/04561/2020. ## References * [Bakshy et al. 2011] Bakshy, E., Hofman, J. M., Mason, W.A., Watts, D.J., 2011. Everyone’s an influencer: quantifying influence on twitter. Proceedings of the fourth ACM international conference on Web search and data mining (pp. 65-74). doi: 10.1145/1935826.1935845 * [Barabási 2016, Chap. 3] Barabási, A.L., 2016. Network science. Cambridge university press, Chapter 3. * [Bastian et al. 2009] Bastian, M., Heymann, S., and Jacomy M., 2009. Gephi: an open source software for exploring and manipulating networks. International AAAI Conference on Weblogs and Social Media. * [Boccaletti et al. 2006] Boccaletti, S., Latora, V., Moreno, Y., Chavez, M., Hwang, D.U., 2006. Complex networks: Structure and dynamics. Physics reports 424(4-5), 175-308. doi: 10.1016/j.physrep.2005.10.009 * [Bonacich 2007] Bonacich, P., 2007. Some unique properties of eigenvector centrality. Social networks 29(4), 555-564. doi: 10.1016/j.socnet.2007.04.002 * [Brandes 2001] Brandes, U., 2001. A faster algorithm for betweenness centrality. Journal of Mathematical Sociology 25(2), 163-177. doi: 10.1080/0022250X.2001.9990249 * [Brin and Page 1998] Brin, S., and Page, L., 1998. The anatomy of a large-scale hypertextual web search engine. Computer Networks and ISDN Systems 30, 107-117. * [Brin and Page 2012] Brin, S., and Page, L., 2012. Reprint of: The anatomy of a large-scale hypertextual web search engine. Computer networks 56(18), 3825-3833. doi: 10.1016/j.comnet.2012.10.007 * [Chen 2009] Chen, N., 2009. On the approximability of influence in social networks. SIAM Journal on Discrete Mathematics 23(3), 1400-1415. doi: 10.1137/08073617X * [Chen et al. 2013] Chen, W., Lakshmanan, L. V., and Castillo, C., 2013. Information and influence propagation in social networks. Synthesis Lectures on Data Management 5(4), 1-177. doi: 10.2200/S00527ED1V01Y201308DTM037 * [Fazio et al. 2015] Fazio, L.K., Brashier, N.M., Payne, B.K., Marsh, E.J., 2015. Knowledge does not protect against illusory truth. Journal of Experimental Psychology: General 144(5), 993-1002. doi: 10.1037/xge0000098 * [Fire et al. 2013] Fire, M., Puzis, R., Elovici, Y., 2013. Link prediction in highly fractional data sets. In: Subrahmanian V. (eds) Handbook of computational approaches to counterterrorism, Springer, New York, NY (pp. 283-300). doi: 10.1007/978-1-4614-5311-6_14 * [Fischetti et al. 2018] Fischetti, M., Kahr, M., Leitner, M., Monaci, M., and Ruthmair, M., 2018. Least cost influence propagation in (social) networks. Mathematical Programming 170(1), 293-325. doi: 10.1007/s10107-018-1288-y * [Foulds 2012] Foulds, L.R., 2012. Graph theory applications. Springer-Verlag, New York. doi: 10.1007/s10107-018-1288-y * [Freeman 1977] Freeman, L.C., 1977. A set of measures of centrality based on betweenness. Sociometry 35-41. doi: 10.2307/3033543 * [Gleich 2015] Gleich, D.F., 2015. PageRank beyond the Web. Siam Review 57(3), 321-363. doi: 10.1137/140976649 * [Goel et al. 2016] Goel, S., Anderson, A., Hofman, J., and Watts, D.J., 2016. The structural virality of online diffusion. Management Science 62(1), 180-196. doi: 10.1287/mnsc.2015.2158 * [Günneç and Raghavan 2017] Gunnec, D., Raghavan, S., 2017. Integrating social network effects in the share-of-choice problem. Decision Sciences 48(6), 1098-1131. doi: 10.1111/deci.12246 * [Günneç et al. 2020] Günneç, D., Raghavan, S., and Zhang, R., 2020. Least-Cost Influence Maximization on Social Networks. INFORMS Journal on Computing 32(2), 289-302. doi: 10.1287/ijoc.2019.0886 * [Hage and Harary 1995] Hage, P., Harary, F., 1995. Eccentricity and centrality in networks. Social networks 17(1), 57-63. doi: 10.1016/0378-8733(94)00248-9 * [Huang et al. 2012] Huang, J., Cheng, X. Q., Shen, H. W., Zhou, T., Jin, X., 2012. Exploring social influence via posterior effect of word-of-mouth recommendations. In Proceedings of the fifth ACM international conference on Web search and data mining (pp. 573-582). doi: 10.1145/2124295.2124365 * [Kempe et al. 2003] Kempe, D., Kleinberg, J., and Tardos, É., 2003. Maximizing the spread of influence through a social network. In Proceedings of the ninth ACM SIGKDD international conference on knowledge discovery and data mining, 137-146. doi: 10.1145/956750.956769 * [Kiss and Bichler 2008] Kiss, C., Bichler, M., 2008. Identification of influencers-measuring influence in customer networks. Decision Support Systems 46(1), 233-253. doi: 10.1016/j.dss.2008.06.007 * [Kleinberg 1999] Kleinberg, J.M., 1999. Authoritative sources in a hyperlinked environment. Journal of the ACM (JACM) 46(5), 604-632. doi: 10.1145/324133.324140 * [Kunegis 2013] Kunegis, J., 2013. KONECT - THe Koblenz Network Collection. Proc. Int. Conf. on World Wide Web Companion, 1343-1350. * [Kunegis 2017] Kunegis, J., 2017. Konect network dataset - KONECT. Accessed March 27, 2020, http://konect.uni-koblenz.de/networks/konect. * [Leskovec et al. 2007] Leskovec, J., Kleinberg, J., Faloutsos, C., 2007. Graph evolution: Densification and shrinking diameters. ACM transactions on Knowledge Discovery from Data (TKDD) 1(1), Article 2. doi: 10.1145/1217299.1217301 * [Leskovec and Krevl 2014] Leskovec, J., Krevl, A., 2014. SNAP datasets: Stanford large network dataset collection. Accessed June 30, 2020, http://snap.stanford.edu/data. * [Leskovec and McAuley 2012] Leskovec, J., McAuley, J.J., 2012. Learning to discover social circles in ego networks. In Proceedings of the Neural Information Processing Systems Conference 2012 (NIPS 2012), Advances in Neural Information Processing Systems 25, (pp. 539-547). * [Lesser et al. 2013] Lesser, O., Tenenboim-Chekina, L., Rokach, L., Elovici, Y., 2013. Intruder or welcome friend: Inferring group membership in online social networks. In International Conference on Social Computing, Behavioral-Cultural Modeling, and Prediction (LNCS 7812), Springer, Berlin, Heidelberg (pp. 368-376). doi: 10.1007/978-3-642-37210-0_40 * [Liu et al. 2017] Liu, Q., Xiang, B., Yuan, N.J., Chen, E., Xiong, H., Zheng, Y., Yang, Y., 2017. An influence propagation view of pagerank. ACM Transactions on Knowledge Discovery from Data (TKDD) 11(3), 1-30. doi: 10.1145/3046941 * [Lü et al. 2016] Lü, L., Chen, D., Ren, X.L., Zhang, Q.M., Zhang, Y.C., Zhou, T., 2016. Vital nodes identification in complex networks. Physics Reports 650:1-63. doi: 10.1016/j.physrep.2016.06.007 * [Massa et al. 2009] Massa, P., Salvetti, M., Tomasoni, D., 2009. Bowling alone and trust decline in social network sites. In 2009 Eighth IEEE International Conference on Dependable, Autonomic and Secure Computing (pp. 658-663). doi: 10.1109/DASC.2009.130 * [Okamoto et al. 2008] Okamoto, K., Chen, W., Li, X.Y., 2008. Ranking of closeness centrality for large-scale social networks. In International workshop on frontiers in algorithmics (LNCS 5059). Springer, Berlin, Heidelberg (pp. 186-195). doi: 10.1007/978-3-540-69311-6_21 * [Panzarasa et al. 2009] Panzarasa, P., Opsahl, T., Carley, K.M., 2009. Patterns and dynamics of users’ behavior and interaction: Network analysis of an online community. Journal of the American Society for Information Science and Technology 60(5), 911-932. doi: 10.1002/asi.21015 * [Peng et al. 2018] Peng, S., Zhou, Y., Cao, L., Yu, S., Niu, J., and Jia, W., 2018. Influence analysis in social networks: A survey. Journal of Network and Computer Applications 106, 17-32. doi: 10.1016/j.jnca.2018.01.005 * [Raghavan and Zhang 2019] Raghavan, S., and Zhang, R., 2019. A branch-and-cut approach for the weighted target set selection problem on social networks. INFORMS Journal on Optimization 1(4), 304-322. doi: 10.1287/ijoo.2019.0012 * [Ripeanu et al. 2002] Ripeanu, M., Foster, I., Iamnitchi, A., 2002. Mapping the gnutella network: Properties of large-scale peer-to-peer systems and implications for system design. arXiv preprint cs/0209028. * [Watts and Strogatz 1998] Watts, D.J., Strogatz, S.H., 1998. Collective dynamics of ”small-world” networks. Nature 393(6684), 440-442. doi: 10.1038/30918 * [Yin et al. 2017] Yin, H., Benson, A.R., Leskovec, J., Gleich, D.F., 2017. Local higher-order graph clustering. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 555-564). doi: 10.1145/3097983.3098069 * [Zachary 1977] Zachary, W.W., 1977. An information flow model for conflict and fission in small groups. Journal of anthropological research 33(4), 452-473. doi: 10.1086/jar.33.4.3629752 ## Appendix Table 8 shows the ILP and ITP indices’ values for the Zachary’s club graph. It also includes the results of the metrics described in Subsection 3.2, calculated using Gephi [Bastian et al. 2009]. nodes | $ilp(i)$ | $itp(i)$ | centralized | link topological ---|---|---|---|--- $i$ | $\alpha=1.5$ | $\alpha=3.0$ | $\alpha=1.5$ | $\alpha=3.0$ | $g(i)$ | $ec(i)$ | $cc(i)$ | $bc(i)$ | $ev(i)$ | $pr(i)$ | $au(i)$ | $hu(i)$ 1 | 1 | 0.61 | 0.12 | 0 | 16 | 3 | 0.57 | 231.07 | 0.96 | 0.1 | 0.36 | 0.36 2 | 1 | 0.09 | 0.12 | 0.03 | 9 | 3 | 0.49 | 28.48 | 0.7 | 0.05 | 0.27 | 0.27 3 | 1 | 0.06 | 0.12 | 0.06 | 10 | 3 | 0.56 | 75.85 | 0.84 | 0.06 | 0.32 | 0.32 4 | 0.06 | 0.03 | 0.15 | 0.03 | 6 | 3 | 0.46 | 6.29 | 0.56 | 0.04 | 0.21 | 0.21 5 | 0 | 0 | 0.15 | 0.03 | 3 | 4 | 0.38 | 0.33 | 0.21 | 0.02 | 0.08 | 0.08 6 | 0.06 | 0 | 0.18 | 0.03 | 4 | 4 | 0.38 | 15.83 | 0.23 | 0.03 | 0.08 | 0.08 7 | 0.06 | 0 | 0.18 | 0.03 | 4 | 4 | 0.38 | 15.83 | 0.23 | 0.03 | 0.08 | 0.08 8 | 0 | 0 | 0.18 | 0.03 | 4 | 4 | 0.44 | 0 | 0.45 | 0.02 | 0.17 | 0.17 9 | 0 | 0 | 0.15 | 0.09 | 5 | 3 | 0.52 | 29.53 | 0.61 | 0.03 | 0.23 | 0.23 10 | 0 | 0 | 0.15 | 0.09 | 2 | 4 | 0.43 | 0.45 | 0.27 | 0.01 | 0.1 | 0.1 11 | 0 | 0 | 0.15 | 0.03 | 3 | 4 | 0.38 | 0.33 | 0.21 | 0.02 | 0.08 | 0.08 12 | 0 | 0 | 0.15 | 0.03 | 1 | 4 | 0.37 | 0 | 0.14 | 0.01 | 0.05 | 0.05 13 | 0 | 0 | 0.18 | 0.06 | 2 | 4 | 0.37 | 0 | 0.22 | 0.01 | 0.08 | 0.08 14 | 0 | 0 | 0.15 | 0.06 | 5 | 3 | 0.52 | 24.22 | 0.6 | 0.03 | 0.23 | 0.23 15 | 0 | 0 | 0.15 | 0.06 | 2 | 5 | 0.37 | 0 | 0.27 | 0.01 | 0.1 | 0.1 16 | 0 | 0 | 0.15 | 0.06 | 2 | 5 | 0.37 | 0 | 0.27 | 0.01 | 0.1 | 0.1 17 | 0 | 0 | 0.21 | 0.03 | 2 | 5 | 0.28 | 0 | 0.07 | 0.02 | 0.02 | 0.02 18 | 0 | 0 | 0.15 | 0.06 | 2 | 4 | 0.38 | 0 | 0.25 | 0.01 | 0.09 | 0.09 19 | 0 | 0 | 0.15 | 0.06 | 2 | 5 | 0.37 | 0 | 0.27 | 0.01 | 0.1 | 0.1 20 | 0 | 0 | 0.15 | 0.09 | 3 | 3 | 0.5 | 17.15 | 0.4 | 0.02 | 0.15 | 0.15 21 | 0 | 0 | 0.15 | 0.06 | 2 | 5 | 0.37 | 0 | 0.27 | 0.01 | 0.1 | 0.1 22 | 0 | 0 | 0.15 | 0.06 | 2 | 4 | 0.38 | 0 | 0.25 | 0.01 | 0.09 | 0.09 23 | 0 | 0 | 0.15 | 0.06 | 2 | 5 | 0.37 | 0 | 0.27 | 0.01 | 0.1 | 0.1 24 | 0.03 | 0 | 0.15 | 0.06 | 5 | 5 | 0.39 | 9.3 | 0.41 | 0.03 | 0.15 | 0.15 25 | 0 | 0 | 0.18 | 0.03 | 3 | 4 | 0.38 | 1.17 | 0.16 | 0.02 | 0.06 | 0.06 26 | 0 | 0 | 0.21 | 0.03 | 3 | 4 | 0.38 | 2.03 | 0.17 | 0.02 | 0.06 | 0.06 27 | 0 | 0 | 0.18 | 0.03 | 2 | 5 | 0.36 | 0 | 0.2 | 0.02 | 0.08 | 0.08 28 | 0 | 0 | 0.15 | 0.03 | 4 | 4 | 0.46 | 11.79 | 0.36 | 0.03 | 0.13 | 0.13 29 | 0 | 0 | 0.18 | 0.09 | 3 | 4 | 0.45 | 0.95 | 0.35 | 0.02 | 0.13 | 0.13 30 | 0.03 | 0 | 0.15 | 0.06 | 4 | 5 | 0.38 | 1.54 | 0.36 | 0.03 | 0.13 | 0.13 31 | 0 | 0 | 0.15 | 0.09 | 4 | 4 | 0.46 | 7.61 | 0.46 | 0.02 | 0.17 | 0.17 32 | 0.09 | 0 | 0.15 | 0.06 | 6 | 3 | 0.54 | 73.01 | 0.52 | 0.04 | 0.19 | 0.19 33 | 1 | 0.27 | 0.12 | 0.03 | 12 | 4 | 0.52 | 76.69 | 0.83 | 0.07 | 0.31 | 0.31 34 | 1 | 0.61 | 0.12 | 0 | 17 | 4 | 0.55 | 160.55 | 1 | 0.1 | 0.37 | 0.37 Table 8: ILP and ITP indices and the results of the metrics described in Subsection 3.2 for the Zachary’s club graph
# Homogenization of Schrödinger equations. Extended Effective Mass Theorems for non-crystalline matter Vernny Ccajma1, Wladimir Neves1, Jean Silva2 ###### Abstract This paper concerns the homogenization of Schrödinger equations for non- crystalline matter, that is to say the coefficients are given by the composition of stationary functions with stochastic deformations. Two rigorous results of so-called effective mass theorems in solid state physics are obtained: a general abstract result (beyond the classical stationary ergodic setting), and one for quasi-perfect materials (i.e. the disorder in the non- crystalline matter is limited). The former relies on the double-scale limits and the wave function is spanned on the Bloch basis. Therefore, we have extended the Bloch Theory which was restrict until now to crystals (periodic setting). The second result relies on the Perturbation Theory and a special case of stochastic deformations, namely stochastic perturbation of the identity. 11footnotetext: Instituto de Matemática, Universidade Federal do Rio de Janeiro, C.P. 68530, Cidade Universitária 21945-970, Rio de Janeiro, Brazil. E-mail<EMAIL_ADDRESS><EMAIL_ADDRESS> Departamento de Matemática, Universidade Federal de Minas Gerais. E-mail: <EMAIL_ADDRESS> Key words and phrases. Homogenization theory, stochastic Schrödinger equations, initial value problem, extended effective mass theorems. ###### Contents 1. 1 Introduction 1. 1.1 Contextualization 2. 1.2 Summary of the main results 2. 2 Preliminaries and Background 1. 2.1 Anisotropic Schrödinger equations 2. 2.2 Stochastic configuration 1. 2.2.1 Ergodic theorems 2. 2.2.2 Analysis of stationary functions 3. 2.3 $\Phi_{\omega}-$Two-scale Convergence 4. 2.4 Perturbations of bounded operators 3. 3 Bloch Waves Analysis 1. 3.1 The WKB method 2. 3.2 Sobolev spaces on groups 1. 3.2.1 Groups and Dynamical systems 2. 3.2.2 Rellich–Kondrachov type Theorem 3. 3.2.3 On a class of Quasi-periodic functions 3. 3.3 Auxiliary celular equations 4. 4 ​​ On Schrödinger Equations Homogenization 1. 4.1 The Abstract Theorem. 2. 4.2 Radom Perturbations of the Quasi-Periodic Case 5. 5 ​ ​ ​Homogenization of Quasi-Perfect Materials 1. 5.1 Perturbed Periodic Case: Spectral Analysis 2. 5.2 Homogenization Analysis of the Perturbed Model 1. 5.2.1 Expansion of the effective coefficients ## Part I: Conceptual Framework ## 1 Introduction In this work we study the homogenization (assymptotic limit as $\varepsilon\to 0$) of the anisotropic Schrödinger equation in the following Cauchy problem $\left\\{\begin{aligned} &i\displaystyle\frac{\partial u_{\varepsilon}}{\partial t}-{\rm div}{\big{(}A(\Phi^{-1}{\big{(}\frac{x}{\varepsilon},\omega\big{)}},\omega)\nabla u_{\varepsilon}\big{)}}+\frac{1}{\varepsilon^{2}}V(\Phi^{-1}{\big{(}\displaystyle\frac{x}{\varepsilon},\omega\big{)}},\omega)\;u_{\varepsilon}\\\\[5.0pt] &\hskip 100.0pt+U(\Phi^{-1}{\big{(}\frac{x}{\varepsilon},\omega\big{)}},\omega)\;u_{\varepsilon}=0,\quad\text{in $\mathbb{R}^{n+1}_{T}\\!\times\\!\Omega$},\\\\[3.0pt] &u_{\varepsilon}=u_{\varepsilon}^{0},\quad\text{in $\mathbb{R}^{n}\\!\times\\!\Omega$},\end{aligned}\right.$ (1.1) where $\mathbb{R}^{n+1}_{T}:=(0,T)\times\mathbb{R}^{n}$, for any real number $T>0$, $\Omega$ is a probability space, and the unknown function $u_{\varepsilon}(t,x,\omega)$ is complex-value. The coefficients in (1.1), that is the matrix-value function $A$, the real- value (potencial) functions $V$, $U$ are random perturbations of stationary functions accomplished by stochastic diffeomorphisms $\Phi:\mathbb{R}^{n}\times\Omega\to\mathbb{R}^{n}$, (called stochastic deformations). The stationarity property of random functions will be precisely defined in Section 2.2, also the definition of stochastic deformations which were introduced by X. Blanc, C. Le Bris, P.-L. Lions (see [9, 10]). In that paper they consider the homogenization problem of an elliptic operator whose coefficients are periodic or stationary functions perturbed by stochastic deformations. In particular, we assume that $A=(A_{k\ell})$, $V$ and $U$ are measurable and bounded functions, i.e. for $k,\ell=1,\ldots,n$ $A_{k\ell},\;V,\;U\in L^{\infty}(\mathbb{R}^{n}\times\Omega).$ (1.2) Moreover, the matrix $A$ is symmetric and uniformly positive defined, that is, there exists $a_{0}>0$, such that, for a.a. $(y,\omega)\in\mathbb{R}^{n}\times\Omega$, and each $\xi\in\mathbb{R}^{n}$ $\sum_{k,\ell=1}^{n}A_{k\ell}(y,\omega)\,\xi_{k}\,\xi_{\ell}\geqslant a_{0}{|\xi|}^{2}.$ (1.3) This paper is the second part of the Project initiated with T. Andrade, W. Neves, J. Silva [6] (Homogenization of Liouville Equations beyond stationary ergodic setting) concerning the study of moving electrons in non-crystalline matter, which justify the form considered for the coefficients in (1.1). We recall that crystalline materials, also called perfect materials, are described by periodic functions. Thus any homogenization result for Schrödinger equations with periodic coefficients is restrict to crystalline matter. Moreover, perfect materials are rare in Nature, there exist much more non-crystalline than crystalline materials. For instance, there exists a huge class called quasi-perfect materials (see Section 5, also [6]), which are closer to perfect ones. Indeed, the concept of stochastic deformations are very suitable to describe interstitial defects in materials science (see Cances, Le Bris [12], and Myers [28]). One remarks that, the homogenization of the Schrödinger equation in (1.1), when the stochastic deformation $\Phi(y,\omega)$ is the identity mapping and the coefficients are periodic, were studied by Allaire, Piatnitski [4]. Notably, that paper presents the discussion about the differences between the scaling considered in (1.1) and the one called semi-classical limit. We are not going to rephrase this point here, and address the reader to Chapter 4 in [8] for more general considerations about that. It should be mentioned that, to the best of our knowledge the present work is the first to study the homogenization of the Schrödinger equations beyond the periodic setting, applying the double-scale limits and the wave function is spanned on the Bloch basis. Therefore, we have extended the Bloch Theory, which was restrict until now to periodic potentials. Last but not least, one observes that the initial data $u_{\varepsilon}^{0}$ shall be considered well-prepared, see equation (4.85). This assumption is fundamental for the abstract homogenization result established in Theorem 4.2, where the limit function obtained from $u_{\varepsilon}$ satisfies a simpler Schrödinger equation, called the effective mass equation, with effective constant coefficients, namely matrix $A^{*}$, and potential $V^{*}$. This homogenization procedure is well known in solid state physics as Effective Mass Theorems, see Section 4. Finally, we stress Section 5 which is related to the homogenization of the Schrödinger equation for quasi-perfect materials, and it is also an important part of this paper. Indeed, a very special case occurs in situations where the amount of randomness is small, more specifically the disorder in the material is limited. In particular, this section is interesting for numerical applications, where specific computationally efficient techniques already designed to deal with the homogenization of the Schrödinger equation in the periodic setting, can be employed to treat the case of quasi-perfect materials. ### 1.1 Contextualization Let us briefly recall that the homogenization’s problem for (1.1) has been treated for the periodic case ($A_{\rm per}(y)$, $V_{\rm per}(y)$, $U_{\rm per}(y)$), and $\Phi(y,\omega)=y$ by some authors. Besides the paper by G. Allaire, A.Piatnitski [4] already mentioned, we address the following papers for the case of $A_{\rm per}=I_{n\times n}$, i.e. isotropic Schrödinger equation in (1.1): G. Allaire, M.Vanninathan [3], L. Barletti, N. Ben Abdallah [7], V. Chabu, C. Fermanian-Kammerer, F. Marcià [14], and we observe that this list is by no means exhaustive. In [3], the authors study a semiconductors model excited by an external potencial $U_{\rm per}(t,x)$, which depends on the time $t$ and macroscopic variable $x$. In [7] the authors treat the homogenization’s problem when the external potential $U_{\rm per}(x,y)$ depends also on the macroscopic variable $x$. Finally, in [14] it was considered an external potential $U_{\rm per}(t,x)$ which model the effects of impurities on the otherwise perfect matter. All the references cited above treat the homogenization’s problem for (1.1), studying the spectrum of the associated Bloch spectral cell equation, that is, for each $\theta\in\mathbb{R}^{n}$, find the eigenvalue-eigenfunction pair $(\lambda,\psi)$, satisfying $\left\\{\begin{aligned} L_{\rm per}(\theta){\big{[}\psi\big{]}}&=\lambda\,\psi,\quad\text{in $[0,1)^{n}$},\\\\[5.0pt] \psi(y)&\not=0,\quad\text{periodic function},\end{aligned}\right.$ (1.4) where $L_{\rm per}(\theta)$ is the Hamiltonian given by $L_{\rm per}(\theta){\big{[}f\big{]}}=-{\big{(}{\rm div}_{\\!y}+2i\pi\theta\big{)}}{\big{[}A_{\rm per}(y){(\nabla_{\\!\\!y}+2i\pi\theta)}f\big{]}}+V_{\rm per}(y)f.$ The above eigenvalue problem is precisely stated (in the more general context studied in this paper) in Section 3. Here, concerning the periodic setting mathematical solutions to (1.4), we address the reader to C. H. Wilcox [37], (see in particular Section 2: A discussion of related literature). Then, once this eigenvalue problem is resolved, the goal is to pass to the limit as $\varepsilon\to 0$. One remarks that, there does not exist an uniform estimate in $H^{1}(\mathbb{R}^{n})$ for the family of solutions $\\{u_{\varepsilon}\\}$ of (1.1), due to the scale $\varepsilon^{-2}$ multiplying the internal potential $V_{\rm per}(y)$. To accomplish the desired asymptotic limit, under this lack of compactness, a nice strategy is to use the two-scale convergence, for instance see the proof of Theorem 3.2 in [4]. Let us now focus on the stochastic setting proposed in this paper, more precisely when the coefficients of the Schrödinger equation in (1.1) are the composition of stationary functions with stochastic deformations. Hence we have the following natural questions: $(Q.1)$ Is it possible to obtain an analogously Bloch spectral cell equation to this stochastic setting? $(Q.2)$ This new stochastic spectral problem can be resolved, such that, the eigenvalues do not depend on $\omega\in\Omega$ (see Remark 3.2)? $(Q.3)$ Is it feasible to adapt the two-scale convergence to this new proposed stochastic setting? We remark that, the approach of stochastic two-scale convergence developed by Bourgeat, Mikelic, Wright [11], and also by Zhikov, Pyatnitskii [38] do not fit to the present context because of the presence of the stochastic deformation $\Phi$. The former question $(Q.1)$ is answered in Section 3.1. Indeed, assuming that the solution of equation (1.1) is given by a plane wave, the stochastic spectral Bloch cell equation (3.56) is obtained applying the asymptotic expansion WKB method, (developed by Wentzel, Kramers, and Brillouin, see G. Allaire [2]). More specifically, the Hamiltonian in (3.56) is given by $L^{\Phi}(\theta)\big{[}F\big{]}\\!\\!=-\big{(}{\rm div}_{\\!z}+2i\pi\theta\big{)}{\left[A{(\Phi^{-1}(z,\omega),\omega)}{\big{(}\nabla_{\\!\\!z}+2i\pi\theta\big{)}}F\right]}+V(\Phi^{-1}(z,\omega),\omega)F,$ for each $F(z,\omega)=f\left(\Phi^{-1}(z,\omega),\omega\right)$, where $f(y,\omega)$ is a stationary function. To answer $(Q.2)$, we have to study the spectrum of the operator $L^{\Phi}(\theta)$, for each $\theta\in\mathbb{R}^{n}$ fixed. The first idea is to follow the techniques applied for the periodic setting, that is, for the operator $L_{\rm per}(\theta)$ in (1.4), where the fundamental tool is the compact embedding of $H^{1}_{\rm per}([0,1)^{n})$ in $L^{2}([0,1)^{n})$. Although, since $\omega\in\Omega$ can not be treat as a fixed parameter, we have to consider the more general theory of Sobolev spaces on locally compact Abelian groups, which is developed in Section 3.2. In fact, we have established in details a Rellich-Kondrachov type Theorem, (see Theorem 3.22), such that together with the study of continuous dynamic systems on compact Abelian groups enable us to answer positively this question, at least, when $\Omega$ has some structure. The second applied strategy here to answer $(Q.2)$ is the Perturbation Theory, that is to say, taking the advantage of the well known spectrum for $L_{\rm per}(\theta)$. To this end, we first consider that the coefficients of the Schrödinger equation in (1.1) are the composition of the periodic functions $A_{\rm per}$, $V_{\rm per}$ and $U_{\rm per}$ with a special case of stochastic deformations, namely stochastic perturbation of the identity (see Definition 5.1), which is given by $\Phi_{\eta}(y,\omega):=y+\eta\,Z(y,\omega)+\mathrm{O}(\eta^{2}),$ where $Z$ is some stochastic deformation and $\eta\in(0,1)$. This concept was introduced by X. Blanc, C. Le Bris, P.-L. Lions [10], and applied for the first time to evolutionary equations in T. Andrade, W. Neves, J. Silva [6]. Then, taking this special case $\Phi_{\eta}$, the operator $L^{\Phi_{\eta}}(\theta)$ has the following expansion in a neighborhood of $(0,\theta_{0})\in\mathbb{R}^{n+1}$, $L^{\Phi_{\eta}}(\theta)=L_{\rm per}(\theta_{0})+\sum_{{|\varrho|}=1}^{3}((\eta,\theta)-(0,\theta_{0}))^{\varrho}L_{\varrho}+\mathrm{O}(\eta^{2}),$ where $\varrho=(\varrho_{1},\ldots,\varrho_{n},\varrho_{n+1})\in\mathbb{N}^{n+1}$, ${|\varrho|}=\sum_{k=1}^{n+1}\varrho_{k}$, and $L_{\varrho}$ is a bounded operator, see Section 5. From the above equation, it follows that the point spectrum (i.e. the set of eigenvalues) of $L^{\Phi_{\eta}}(\theta)$ is not empty in a neighborhood of $(0,\theta_{0})$, when $\lambda_{\rm per}(\theta_{0})$ is an isolated eigenvalue with finite multiplicity. This last property is studied in details in Section 2.4, see Theorem 2.26. The question $(Q.3)$ is answered positively in Section 2.3, that is, we have established in this section a two-scale convergence in a stochastic setting, which is beyond the classical stationary ergodic setting. Indeed, the main difference here with the earlier stochastic extensions of the periodic setting is that, the test functions used are random perturbations of stationary functions accomplished by the stochastic deformations. These compositions are beyond the stationary class, thus we have a lack of the stationarity property in this kind of test functions (see the introduction section in [6] for a deep discussion about this subject). It was introduced a compactification argument that, preserves the ergodic nature of the setting involved and allow us to overcome these difficulties. ### 1.2 Summary of the main results In this section we summarize the main results on this paper. Since some of the theorems (cited below) have its on interested, we describe shortly the main issue of each one. First, Theorem 2.18 allows us to overcome the lack of topological structure of a given probability space reducing it to a separable compact space whose topological basis is dictated by the coefficients of the problem (1.1). Then, the Theorem 2.21 uses all topological features brought forth by the Theorem 2.18 in order to give us a result about two-scale convergence where the test functions are random perturbations accomplished by stochastic diffeomorphisms of stationary functions. It is worth mentioning that this result generalizes the corresponding one for deterministic case in [16] and the corresponding one for the stochastic case in [11]. Theorem 2.26 consider a sequence of bounded operators in a Hilbert space, which defines a symmetric operator via the power series of multidimensional complex variables. It is stated that, if the first coefficient operator of this series has isolated eigenvalues of finite multiplicity, then the holomorphic defined operator inherited from it similar point spectrum analysis. The Theorem 3.15 established a necessary condition such that, the Rellich–Kondrachov Theorem on compact Abelian groups holds true. More precisely, the dual group must be an enumerable set. A complete characterization of the Rellich–Kondrachov Theorem on compact Abelian groups is given by Theorem 3.22. Moreover, as a byproduct of this characterization, we provide a proof of the Rellich–Kondrachov Theorem in a precise context. The Theorem 4.2 is one of the main results of this paper. It is an abstract homogenization result for Schrödinger equations that encompasses the corresponding one given by Allaire and Piatnistski [4] in the periodic context. The Theorem 5.9 shows how the periodic setting can be used to deal with homogenization of the equation (1.1) for materials when the amount of randomness is small. This has importants numerical implications. The Theorem 5.11 reveals an interesting splitting property of the solution of the homogenized equation associated to (1.1) in the specific case of the quasi-perfect materials. ## 2 Preliminaries and Background This section introduces the basement theory, which will be used through the paper. To begin we fix some notations, and collect some preliminary results. The material which is well-known or a direct extension of existing work are giving without proofs, otherwise we present them. We denote by $\mathbb{G}$ the group $\mathbb{Z}^{n}$ (or $\mathbb{R}^{n}$), with $n\in\mathbb{N}$. The set $[0,1)^{n}$ denotes the unit cube, which is also called the unitary cell and will be used as the reference period for periodic functions. The symbol $\left\lfloor x\right\rfloor$ denotes the unique number in $\mathbb{Z}^{n}$, such that $x-\left\lfloor x\right\rfloor\in[0,1)^{n}$. Let $H$ be a complex Hilbert space, we denote by $\mathcal{B}(H)$ the Banach space of linear bounded operators from $H$ to $H$. Let $U\subset\mathbb{R}^{n}$ be an open set, $p\geqslant 1$, and $s\in\mathbb{R}$. We denote by $L^{p}(U)$ the set of (real or complex) $p-$summable functions with respect to the Lebesgue measure (vector ones should be understood componentwise). Given a Lebesgue measurable set $E\subset\mathbb{R}^{n}$, $|E|$ denotes its $n-$dimensional Lebesgue measure. Moreover, we will use the standard notations for the Sobolev spaces $W^{s,p}(U)$ and $H^{s}(U)\equiv W^{s,2}(U)$. ### 2.1 Anisotropic Schrödinger equations The aim of this section is to present the well-posedness for the solutions of the Schrödinger equation, and some properties of them. Most of the material can be found in Cazenave, Haraux [13]. First, let us consider the following Cauchy problem, which is driven by a linear anisotropic Schrödinger equation, that is $\left\\{\begin{aligned} &i\;\partial_{t}u(t,x)-{\rm div}\big{(}A(x)\nabla u(t,x)\big{)}+V(x)\,u(t,x)=0\quad\text{in $\mathbb{R}^{n+1}_{T}$},\\\\[5.0pt] &u(0,x)=u_{0}(x)\quad\text{in $\mathbb{R}^{n}$},\end{aligned}\right.$ (2.5) where the unknown $u(t,x)$ is a complex value function, and $u_{0}$ is a given initial datum. The coefficient $A(x)$ is a symmetric real $n\times n$-matrix value function, and the potential $V(x)$ is a real function. We always assume that $A(x),V(x)\quad\text{are measurable bounded functions}.$ (2.6) One recalls that, a matrix $A$ is called (uniformly) coercive, when, there exists $a_{0}>0$, such that, for each $\xi\in\mathbb{R}^{n}$, and almost all $x\in\mathbb{R}^{n}$, $A(x)\xi\cdot\xi\geqslant a_{0}|\xi|^{2}$. The following definition tell us in which sense a complex function $u(t,x)$ is a mild solution to (2.5). ###### Definition 2.1. Let $A,V$ be coefficients satisfying (2.6). Given $u_{0}\in H^{1}(\mathbb{R}^{n})$, a function $u\in C([0,T];H^{1}(\mathbb{R}^{n}))\cap C^{1}((0,T);H^{-1}(\mathbb{R}^{n}))$ is called a mild solution to the Cauchy problem (2.5), when for each $t\in(0,T)$, it follows that $i\partial_{t}u(t)-{\rm div}\big{(}A\nabla u(t)\big{)}+Vu(t)=0\quad\text{in $H^{-1}(\mathbb{R}^{n})$},$ (2.7) and $u(0)=u_{0}$ in $H^{1}(\mathbb{R}^{n})$. Then, we state the following ###### Proposition 2.2. Let $A$ be a coercive matriz value function, $V$ a potential and $u_{0}\in H^{1}(\mathbb{R}^{n})$ a given initial data. Assume that $A,V$ satisfy (2.6). Then, there exist a unique mild solution of the Cauchy problem (2.5). ###### Proof. The proof follows applying Lemma 4.1.5 and Corollary 4.1.2 in [13]. ∎ ###### Remark 2.3. It is very important in the homogenization procedure of the Schrödinger equation, when the coefficients $A$ and $V$ in (2.5) are constants, the matrix $A$ is not necessarily coercive, and the initial data $u_{0}\in L^{2}(\mathbb{R}^{n})$. Then, a function $u\in L^{2}(\mathbb{R}^{n+1}_{T})$ is called a weak solution to (2.5), if it satisfies $i\partial_{t}u-{\rm tr}(AD^{2}u)+Vu=0\quad\text{in distribution sense}.$ Since $A,V$ are constant, we may apply the Fourier Transform, and obtain the existence of a unique solution $u\in H^{1}((0,T);L^{2}(\mathbb{R}^{n}))$. Therefore, the solution $u\in C([0,T];L^{2}(\mathbb{R}^{n}))$ after being redefined in a set of measure zero, and we have $u(0)=u_{0}$ in $L^{2}(\mathbb{R}^{n})$. Now, let us recall the standard a priori estimates for the solutions of the Cauchy problem (2.5). First, under the conditions of Proposition 2.2, a function $u\in C([0,T];H^{1}(\mathbb{R}^{n}))\cap C^{1}((0,T);H^{-1}(\mathbb{R}^{n}))$, which is the mild solution of (2.5), satisfies for each $t\in[0,T]$ $\displaystyle(i)\ \int_{\mathbb{R}^{n}}|u(t)|^{2}dx=\int_{\mathbb{R}^{n}}|u_{0}|^{2}dx,$ (2.8) $\displaystyle(ii)\ \int_{\mathbb{R}^{n}}|\nabla u(t)|^{2}dx\leqslant C\ \big{(}\int_{\mathbb{R}^{n}}|\nabla u_{0}|^{2}dx+\int_{\mathbb{R}^{n}}|u_{0}|^{2}dx\big{)},$ where $C=C(\|V\|_{L^{\infty}},\|A\|_{L^{\infty}},a_{0})$ is a positive constant. Clearly, for the constant coefficients case, with $A$ non-coercive and $u_{0}\in L^{2}(\mathbb{R}^{n})$, a function $u\in C([0,T];L^{2}(\mathbb{R}^{n}))$, which is the weak solution of (2.5), just satisfies the item $(i)$ above. These estimates follow by density argument. ### 2.2 Stochastic configuration Here we present the stochastic context, which will be used thoroughly in the paper. To begin, let $(\Omega,\mathcal{F},\mathbb{P})$ be a probability space. For each random variable $f$ in $L^{1}(\Omega;\mathbb{P})$, ($L^{1}(\Omega)$ for short), we denote its expectation value by $\mathbb{E}[f]=\int_{\Omega}f(\omega)\ d\mathbb{P}(\omega).$ A mapping $\tau:\mathbb{G}\times\Omega\to\Omega$ is said a $n-$dimensional dynamical system if: 1. (i) (Group Property) $\tau(0,\cdot)=id_{\Omega}$ and $\tau(x+y,\omega)=\tau(x,\tau(y,\omega))$ for all $x,y\in\mathbb{G}$ and $\omega\in\Omega$. 2. (ii) (Invariance) The mappings $\tau(x,\cdot):\Omega\to\Omega$ are $\mathbb{P}$-measure preserving, that is, for each $x\in\mathbb{G}$ and every $E\in\mathcal{F}$, we have $\tau(x,E)\in\mathcal{F},\qquad\mathbb{P}(\tau(x,E))=\mathbb{P}(E).$ For simplicity, we shall use $\tau(k)\omega$ to denote $\tau(k,\omega)$. Moreover, it is usual to say that $\tau(k)$ is a discrete (continuous) dynamical system if $k\in\mathbb{Z}^{n}$ ($k\in\mathbb{R}^{n}$), but we only stress this when it is not obvious from the context. A measurable function $f$ on $\Omega$ is called $\tau$-invariant, if for each $k\in\mathbb{G}$ $f(\tau(k)\omega)=f(\omega)\quad\text{for almost all $\omega\in\Omega$}.$ Hence a measurable set $E\in\mathcal{F}$ is $\tau$-invariant, if its characteristic function $\chi_{E}$ is $\tau$-invariant. In fact, it is a straightforward to show that, a $\tau$-invariant set $E$ can be equivalently defined by $\tau(k)E=E\quad\text{for each $k\in\mathbb{G}$}.$ Moreover, we say that the dynamical system $\tau$ is ergodic, when all $\tau$-invariant sets $E$ have measure $\mathbb{P}(E)$ of either zero or one. Equivalently, we may characterize an ergodic dynamical system in terms of invariant functions. Indeed, a dynamical system is ergodic if each $\tau$\- invariant function is constant almost everywhere, that is to say $\Big{(}f(\tau(k)\omega)=f(\omega)\quad\text{for each $k\in\mathbb{G}$ and a.e. $\omega\in\Omega$}\Big{)}\Rightarrow\text{ $f(\cdot)=const.$ a.e.}.$ ###### Example 2.4. Let $\Omega=[0,1)^{n}$ be a sample space, $\mathcal{F}$ the appropriate $\sigma$-algebra on $\Omega$, and $\mathbb{P}$ the probability measure, i.e. the Lebesgue measure restrict to $\Omega$. Then, we consider the $n$-dimensional dynamical system $\tau:\mathbb{R}^{n}\times\Omega\to\Omega$, defined by $\tau(x)\omega:=x+\omega-\left\lfloor x+\omega\right\rfloor.$ The group property for $\tau(x)$ follows from the greatest integer function properties, and its invariance from the translation invariance of the Lebesgue measure. ###### Example 2.5. Let $(\Omega_{0},\mathscr{F}_{0},\mathbb{P}_{0})$ be a probability space. For $m\in\mathbb{N}$ fixed, we consider the set $S=\\{0,1,2,\ldots,m\\}$ and the real numbers $\text{$p_{0},p_{1},p_{2},\ldots,p_{m}$ in $(0,1)$, such that $\sum_{\ell=0}^{m}p_{\ell}=1$}.$ If $\\{X_{k}:\Omega_{0}\to S\\}_{k\in\mathbb{Z}^{n}}$ is a family of random variables, then it is induced a probability measure from it on the measurable space $\big{(}S^{\mathbb{Z}^{n}},\bigotimes_{k\in\mathbb{Z}^{n}}2^{S}\big{)}$. Indeed, we may define the probability measure $\mathbb{P}(E):=\mathbb{P}_{0}{\left\\{X\in E\right\\}},\;\;E\in\bigotimes_{k\in\mathbb{Z}^{n}}2^{S},$ where the mapping $X:\Omega_{0}\to S^{\mathbb{Z}^{n}}$ is given by $X(\omega_{0})=(X_{k}(\omega_{0}))_{k\in\mathbb{Z}^{n}}$. Now, we denote for convenience $\Omega=S^{\mathbb{Z}^{n}}$ and $\mathscr{F}=\bigotimes_{k\in\mathbb{Z}^{n}}2^{S}$, that is, $\mathscr{F}=\sigma(\mathscr{A})$, where $\mathscr{A}$ is the algebra given by the finite union of sets (cylinders of finite base) of the form $\prod_{k\in\mathbb{Z}^{n}}E_{k},$ (2.9) where $E_{k}\in 2^{S}$ is different from $S$ for a finite number of indices $k$. Additionally we assume that, the family $\\{X_{k}\\}_{k\in\mathbb{Z}^{n}}$ is independent, and for each $k\in\mathbb{Z}^{n}$, we have $\mathbb{P}_{0}{\\{X_{k}=0\\}}=p_{0},\,\,\mathbb{P}_{0}{\\{X_{k}=1\\}}=p_{1},\,\,\ldots,\,\,\mathbb{P}_{0}{\\{X_{k}=m\\}}=p_{m}.$ (2.10) Then, we may define an ergodic dynamical system $\tau:\mathbb{Z}^{n}\times\Omega\to\Omega$, by ${\left(\tau(\ell)\omega\right)}(k):=\omega(k+\ell),\quad\text{for any $k,\ell\in\mathbb{Z}^{n}$},$ where $\omega=(\omega(k))_{k\in\mathbb{Z}^{n}}$. $i)$ The group property follows from the definition. Indeed, for each $\omega\in\Omega$ and $\ell_{1},\ell_{2}\in\mathbb{Z}^{n}$, it follows that ${\big{(}\tau(\ell_{1}+\ell_{2})\omega\big{)}}(k)=\omega(k+\ell_{1}+\ell_{2})={\big{(}\tau(\ell_{1})\tau(\ell_{2})\omega\big{)}}(k),$ for any $k\in\mathbb{Z}^{n}$. $(ii)$ The mappings $\tau(\ell,\cdot):\Omega\to\Omega$ are $\mathbb{P}$-measure preserving. First, we observe from (2.9) that, for all $\ell\in\mathbb{Z}^{n}$ $\tau(\ell)\big{(}\prod_{k\in\mathbb{Z}^{n}}E_{k}\big{)}=\prod_{k\in\mathbb{Z}^{n}}E_{k+\ell}.$ Therefore, for any $\ell\in\mathbb{Z}^{n}$ $\displaystyle\mathbb{P}{\Big{(}\tau(\ell){\big{(}\prod_{k\in\mathbb{Z}^{n}}E_{k}\big{)}}\Big{)}}$ $\displaystyle=\mathbb{P}{\big{(}\prod_{k\in\mathbb{Z}^{n}}E_{k+\ell}\big{)}}=\mathbb{P}_{0}{\big{(}\bigcap_{k\in\mathbb{Z}^{n}}\\{X_{k}\in E_{k+\ell}\\}\big{)}}$ $\displaystyle=\prod_{k\in\mathbb{Z}^{n}}\mathbb{P}_{0}{\left\\{X_{k}\in E_{k+\ell}\right\\}}$ $\displaystyle=\prod_{k\in\mathbb{Z}^{n}}\mathbb{P}_{0}{\left\\{X_{k+\ell}\in E_{k+\ell}\right\\}}=\prod_{k\in\mathbb{Z}^{n}}\mathbb{P}_{0}{\left\\{X_{k}\in E_{k}\right\\}},$ where we have used in the second line that the family of random variables is independent and in the third line it has the same distribution, equation (2.10). Then, the measure preserving is satisfied for each element of the algebra $\mathscr{A}$, and hence for each element of $\mathscr{F}$. $(iii)$ The ergodicity. Given the cylinders ${\prod_{k\in\mathbb{Z}^{n}}E_{k}}$ and ${\prod_{k\in\mathbb{Z}^{n}}F_{k}}$, there exists $\ell_{0}\in\mathbb{Z}^{n}$, such that $\mathbb{P}{\Big{(}\tau(\ell_{0}){\big{(}\prod_{k\in\mathbb{Z}^{n}}E_{k}\big{)}}\cap{\big{(}\prod_{k\in\mathbb{Z}^{n}}F_{k}\big{)}}\Big{)}}=\mathbb{P}{\big{(}\prod_{k\in\mathbb{Z}^{n}}E_{k}\big{)}}\,\mathbb{P}{\big{(}\prod_{k\in\mathbb{Z}^{n}}F_{k}\big{)}}.$ Indeed, let us define $e_{0}:={\rm max}{\\{{|k|}\,;\,k\in\mathbb{Z}^{n},\,E_{k}\not=S\\}},\,\quad f_{0}:={\rm max}{\\{{|k|}\,;\,k\in\mathbb{Z}^{n},\,F_{k}\not=S\\}},$ and observe that, if $\ell_{0}\in\mathbb{Z}^{n}$ satisfies ${{|\ell_{0}|}>e_{0}+f_{0}}$, then $E_{k+\ell_{0}}\cap F_{k}=\left\\{\begin{array}[]{ll}F_{k}&\text{if}\;{|k|}\leqslant f_{0},\\\\[5.0pt] E_{k}&\text{if}\;f_{0}<{|k|}\leqslant e_{0}+f_{0},\\\\[5.0pt] S&\text{if}\;{|k|}>e_{0}+f_{0}.\end{array}\right.$ Therefore, we have $\displaystyle\mathbb{P}{\big{(}\tau(\ell_{0}){\big{(}\prod_{k\in\mathbb{Z}^{n}}E_{k}\big{)}}\cap{\big{(}\prod_{k\in\mathbb{Z}^{n}}F_{k}\big{)}}\big{)}}$ $\displaystyle=$ $\displaystyle\mathbb{P}{\big{(}{\big{(}\prod_{k\in\mathbb{Z}^{n}}E_{k+\ell_{0}}\big{)}}\cap{\big{(}\prod_{k\in\mathbb{Z}^{n}}F_{k}\big{)}}\big{)}}$ $\displaystyle=$ $\displaystyle\mathbb{P}{\big{(}\prod_{k\in\mathbb{Z}^{n}}{\big{(}E_{k+\ell_{0}}\cap F_{k}\big{)}}\big{)}}$ $\displaystyle=$ $\displaystyle\prod_{k\in\mathbb{Z}^{n}}\mathbb{P}_{0}{\left\\{X_{k}\in E_{k+\ell_{0}}\cap F_{k}\right\\}}$ $\displaystyle=$ $\displaystyle\mathbb{P}{\big{(}\prod_{k\in\mathbb{Z}^{n}}E_{k}\big{)}}\mathbb{P}{\big{(}\prod_{k\in\mathbb{Z}^{n}}F_{k}\big{)}}.$ The above property follows for finite unions of cylinders, that is to say, given $E_{1},E_{2}\in\mathscr{A}$, there exists $\ell_{0}\in\mathbb{Z}^{n}$, such that $\mathbb{P}{\left(\tau(\ell_{0}){E}_{1}\cap{E}_{2}\right)}=\mathbb{P}({E}_{1})\,\mathbb{P}({E}_{2}).$ Now, let $E\in\mathscr{F}$ be a $\tau$-invariant set. For each $\varepsilon>0$, there exists ${E}_{0}\in\mathscr{A}$ such that, $\mathbb{P}{\left({E}\Delta\,{E}_{0}\right)}<\varepsilon$. Then, since $E$ is $\tau$-invariant we have for each $\ell\in\mathbb{Z}^{n}$ $\displaystyle\mathbb{P}{\big{(}\tau(\ell){E}_{0}\,\Delta\,{E}_{0}\big{)}}$ $\displaystyle\leq\mathbb{P}{\big{(}\tau(\ell){E}_{0}\,\Delta\,\tau(\ell){E}\big{)}}+\mathbb{P}{\big{(}\tau(\ell){E}\,\Delta\,{E}\big{)}}+\mathbb{P}{\big{(}{E}\Delta\,{E}_{0}\big{)}}$ (2.11) $\displaystyle=2\,\mathbb{P}{\big{(}{E}\Delta\,{E}_{0}\big{)}}\leq 2\varepsilon.$ On the other hand, since ${E}_{0}\in\mathscr{A}$, for some $\ell_{0}\in\mathbb{Z}^{n}$, it follows that $\mathbb{P}{\left(\tau(\ell_{0}){E}_{0}\cap{E}_{0}^{c}\right)}=\mathbb{P}({E}_{0})\mathbb{P}({E}_{0}^{c})\quad\text{and}\quad\mathbb{P}{\left(\tau(\ell_{0}){E}_{0}^{c}\cap{E}_{0}\right)}=\mathbb{P}({E}_{0}^{c})\mathbb{P}({E}_{0}),$ and thus $\displaystyle\mathbb{P}{\big{(}\tau(\ell_{0}){E}_{0}\,\Delta\,{E}_{0}\big{)}}$ $\displaystyle=\mathbb{P}{\left(\tau(\ell_{0}){E}_{0}\cap{E}_{0}^{c}\right)}+\mathbb{P}{\left(\tau(\ell_{0}){E}_{0}^{c}\cap{E}_{0}\right)}$ (2.12) $\displaystyle=2\mathbb{P}({E}_{0})(1-\mathbb{P}({E}_{0})).$ From (2.11) and (2.12), it follows for each $\varepsilon>0$ $\mathbb{P}({E}_{0})(1-\mathbb{P}({E}_{0}))<\varepsilon.$ Consequently, we obtain that $\mathbb{P}({E})=0$ or $\mathbb{P}({E})=1$. Now, let $(\Gamma,\mathcal{G},\mathbb{Q})$ be a given probability space. We say that a measurable function $g:\mathbb{R}^{n}\times\Gamma\to\mathbb{R}$ is stationary, if for any finite set consisting of points $x_{1},\ldots,x_{j}\in\mathbb{R}^{n}$, and any $k\in\mathbb{G}$, the distribution of the random vector $\Big{(}g(x_{1}+k,\cdot),\cdots,g(x_{j}+k,\cdot)\Big{)}$ is independent of $k$. Further, subjecting the stationary function $g$ to some natural conditions it can be showed that, there exists other probability space $(\Omega,\mathcal{F},\mathbb{P})$, a $n-$dimensional dynamical system $\tau:\mathbb{G}\times\Omega\to\Omega$ and a measurable function $f:\mathbb{R}^{n}\times\Omega\to\mathbb{R}$ satisfying * • For all $x\in\mathbb{R}^{n}$, $k\in\mathbb{G}$ and $\mathbb{P}-$almost every $\omega\in\Omega$ $f(x+k,\omega)=f(x,\tau(k)\omega).$ (2.13) * • For each $x\in\mathbb{R}^{n}$ the random variables $g(x,\cdot)$ and $f(x,\cdot)$ have the same law. We recall that, the equality almost surely implies equality in law, but the converse is not true. One remarks that, the set of stationary functions forms an algebra, and also is stable by limit process. For instance, the product of two stationaries functions is a stationary one, and the derivative of a stationary function is stationary. Moreover, the stationarity concept is the most general extension of the notions of periodicity and almost periodicity for a function to have some ”self-averaging” behaviour. ###### Example 2.6. Under the conditions of Example 2.4, let $F\\!:\Omega\\!\to\mathbb{C}$ be a measurable function. Then, the function $f\\!:\\!\mathbb{R}^{n}\\!\times\Omega\\!\to\\!\mathbb{C}$, defined by $f(x,\omega):=F(\tau(x)\omega)$ is a stationary function. In fact, considering continuous dynamical systems, any stationary function can be written in this way. Therefore, even if $f(\cdot,\omega)$ is just a measurable function, it makes sense to write, for instance, $f(0,\cdot)$ due to the stationary property. ###### Example 2.7. Under the conditions of Example 2.5, we take $m=1$, and consider the following functions, $\varphi_{0}=0$ and $\varphi_{1}$ be a Lipschitz vector field, such that, $\varphi_{1}$ is periodic, ${\rm supp}\,\varphi_{1}\subset(0,1)^{n}$. Consequently, the function $f(y,\omega):=\varphi_{\omega({\lfloor y\rfloor})}(y),\;\;(y,\omega)\in\mathbb{R}^{n}\times\\!\Omega$ satisfies, ${f(y,\cdot)}$ is ${\mathscr{F}}$-measurable, ${f(\cdot,\omega)}$ is continuous, and for each $k\in\mathbb{Z}^{n}$, $f(y+k,\omega)=f(y,\tau(k)\omega).$ Therefore, ${f}$ is a stationary function. Now, we present the precise definition of the stochastic deformation as presented in [6]. ###### Definition 2.8. A mapping $\Phi:\mathbb{R}^{n}\times\Omega\to\mathbb{R}^{n},(y,\omega)\mapsto z=\Phi(y,\omega)$, is called a stochastic deformation (for short $\Phi_{\omega}$), when satisfies: * i) For $\mathbb{P}-$almost every $\omega\in\Omega$, $\Phi(\cdot,\omega)$ is a bi–Lipschitz diffeomorphism. * ii) There exists $\nu>0$, such that $\underset{\omega\in\Omega,\,y\in\mathbb{R}^{n}}{\rm ess\,inf}\big{(}{\rm det}\big{(}\nabla\Phi(y,\omega)\big{)}\big{)}\geq\nu.$ * iii) There exists a $M>0$, such that $\underset{\omega\in\Omega,\,y\in\mathbb{R}^{n}}{\rm ess\,sup}\big{(}|\nabla\Phi(y,\omega)|\big{)}\leq M<\infty.$ * iv) The gradient of $\Phi$, i.e. $\nabla\Phi(y,\omega)$, is stationary in the sense (2.13). Here, we first recall from [6] a general example of stochastic deformations $\Phi:\mathbb{R}^{n}\times\Omega\to\mathbb{R}^{n}$ associated to a dynamical system $T:\mathbb{R}^{n}\times\Omega\to\Omega$, where the sample space $\Omega$ is arbitrary. Then, following the idea of Example 2.7, we present an example of stochastic deformation $\Phi:\mathbb{R}^{n}\times\Omega\to\mathbb{R}^{n}$ associated to a dynamical system $T:\mathbb{Z}^{n}\times\Omega\to\Omega$, where $\Omega$ is prescribed. Let $(\Omega_{i},\mathcal{F}_{i},\mathbb{P}_{i})_{i=1}^{n}$ be probability spaces, and $f_{i}:\Omega_{i}\to\mathbb{R}$ be a measurable function, such that $0<c_{0}\leq f_{i}(\omega)\leq c_{1}$ for a.e. $\omega\in\Omega_{i}$, for $i=1,\ldots,n$. Let $T_{i}:\mathbb{R}\times\Omega_{i}\to\Omega_{i}$ be a $1-$dimensional dynamical system such that the function $f_{i}\left(T_{i}(\cdot)\omega\right)$ is continuous. Then, we define $\Phi_{i}(\lambda,\omega):=\text{sgn}(\lambda)\int_{\min{\\{\lambda,0\\}}}^{\max{\\{\lambda,0}\\}}f_{i}\left(T_{i}(s)\omega\right)\,ds.$ Then, $\Phi_{i}:\mathbb{R}\times\Omega_{i}\to\mathbb{R}$ is not a stationary function and verifies all conditions of the Definition 2.8. Finally, define the probability space $(\Omega,\mathcal{F},\mathbb{P})$, where $\Omega:=\otimes_{i=1}^{n}\Omega_{i}$, $\mathcal{F}:=\otimes_{i=1}^{n}\mathcal{F}_{i}$ and $\mathbb{P}:=\otimes_{i=1}^{n}\mathbb{P}_{i}$ and the following $n-$dimensional dynamical system $T:\mathbb{R}^{n}\times\Omega\to\Omega$ $T(x,\omega):=\big{(}T_{1}(x_{1},\omega_{1}),\ldots,T_{n}(x_{n},\omega_{n})\big{)},$ where we denote $x=(x_{1},\cdots,x_{n})$ and $\omega=(\omega_{1},\ldots,\omega_{n})$. Thus, the function $\Phi(x,\omega)=(\Phi_{1}(x_{1},\omega_{1}),\ldots,\Phi_{n}(x_{n},\omega_{n}))$ fulfills the conditions of Definition 2.8. Moreover, for any orthogonal $n\times n-$matrix $Q$, with $\det Q=1$, the mapping $Q\Phi$ is also a stochastic deformation, which gradient is not necessary diagonal. Now, let us present a second example of stochastic deformations. ###### Example 2.9. Under the conditions of Example 2.7, let us consider (for $\eta>0$) the following map $\Phi(y,\omega):=y+\eta\,\varphi_{\omega({\lfloor y\rfloor})}(y),\;\;(y,\omega)\in\mathbb{R}^{n}\times\Omega.$ Then, $\nabla_{\\!\\!y}\Phi(y,\omega)=I_{\mathbb{R}^{n\times n}}+\eta\,\nabla\varphi_{\omega({\lfloor y\rfloor})}(y)$, and for $\eta$ sufficiently small all the conditions in the Definition 2.8 are satisfied. Then, ${\Phi}$ is a stochastic deformation. Given a stochastic deformation $\Phi$, let us consider the following spaces $\mathcal{L}_{\Phi}:={\big{\\{}F(z,\omega)=f(\Phi^{-1}(z,\omega),\omega);f\in L^{2}_{\rm loc}(\mathbb{R}^{n};L^{2}(\Omega))\;\;\text{stationary}\big{\\}}}$ (2.14) and $\mathcal{H}_{\Phi}:={\big{\\{}F(z,\omega)=f(\Phi^{-1}(z,\omega),\omega);\;f\in H^{1}_{\rm loc}(\mathbb{R}^{n};L^{2}(\Omega))\;\;\text{stationary}\big{\\}}}$ (2.15) which are Hilbert spaces, endowed respectively with the inner products $\displaystyle{\langle F,G\rangle}_{\mathcal{L}_{\Phi}}$ $\displaystyle:=\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}\\!\\!F(z,\omega)\,\overline{G(z,\omega)}\,dz\,d\mathbb{P}(\omega),$ $\displaystyle{\langle F,G\rangle}_{\mathcal{H}_{\Phi}}$ $\displaystyle:=\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}\\!\\!F(z,\omega)\,\overline{G(z,\omega)}\,dz\,d\mathbb{P}(\omega)$ $\displaystyle\;\quad+\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}\\!\\!\nabla_{\\!\\!z}F(z,\omega)\cdot\overline{\nabla_{\\!\\!z}G(z,\omega)}\,dz\,d\mathbb{P}(\omega).$ ###### Remark 2.10. Under the above notations, when $\Phi=Id$ we denote $\mathcal{L}_{\Phi}$ and $\mathcal{H}_{\Phi}$ by $\mathcal{L}$ and $\mathcal{H}$ respectively. Moroever, a function $F\in{\mathcal{H}}_{\Phi}$ if, and only if, $F\circ\Phi\in{\mathcal{H}}$, and there exist constants $C_{1},C_{2}>0$, such that $C_{1}\|F\circ\Phi\|_{{\mathcal{H}}}\leq\|F\|_{{\mathcal{H}}_{\Phi}}\leq C_{2}\|F\circ\Phi\|_{{\mathcal{H}}}.$ Analogously, $F\in{\mathcal{L}}_{\Phi}$ if, and only if, $F\circ\Phi\in{\mathcal{L}}$, and there exist constants $C_{1},C_{2}>0$, such that $C_{1}\|F\circ\Phi\|_{{\mathcal{L}}}\leq\|F\|_{{\mathcal{L}}_{\Phi}}\leq C_{2}\|F\circ\Phi\|_{{\mathcal{L}}}.$ Indeed, let us show the former equivalence. Applying a change of variables, we obtain $\displaystyle\|F\|^{2}_{{\mathcal{H}}_{\Phi}}$ $\displaystyle=\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}\\!\\!|F(z,\omega)|^{2}\,dz\,d\mathbb{P}(\omega)+\int_{\Omega}\int_{\Phi(\mathsf{Y},\omega)}\\!\\!|\nabla_{\\!\\!z}F(z,\omega)|^{2}\,dz\,d\mathbb{P}(\omega)$ $\displaystyle=\int_{\Omega}\int_{[0,1)^{n}}\\!\\!|f(y,\omega)|^{2}\det[\nabla\Phi(y,\omega)]\,dy\,d\mathbb{P}(\omega)$ $\displaystyle\quad+\int_{\Omega}\int_{[0,1)^{n}}\\!\\!|[\nabla\Phi(y,\omega)]^{-1}\nabla_{\\!\\!z}f(y,\omega)|^{2}\det[\nabla\Phi(y,\omega)]\,dy\,d\mathbb{P}(\omega).$ The equivalence follows from the properties of the stochastic deformation $\Phi$. #### 2.2.1 Ergodic theorems We begin this section with the concept of mean value, which is in connection with the notion of stationarity. A function $f\in L^{1}_{\text{loc}}(\mathbb{R}^{n})$ is said to possess a mean value if the sequence $\\{f(\cdot/\varepsilon){\\}}_{\varepsilon>0}$ converges in the duality with $L^{\infty}$ and compactly supported functions to a constant $M(f)$. This convergence is equivalent to $\lim_{t\to\infty}\frac{1}{t^{n}|A|}\int_{A_{t}}f(x)\,dx=M(f),$ (2.16) where $A_{t}:=\\{x\in\mathbb{R}^{n}\,:\,t^{-1}x\in A\\}$, for $t>0$ and any $A\subset\mathbb{R}^{n}$, with $|A|\neq 0$. ###### Remark 2.11. Unless otherwise stated, we assume that the dynamical system $\tau:\mathbb{G}\times\Omega\to\Omega$ is ergodic and we will also use the notation $\mkern 12.0mu\mbox{\vrule height=4.0pt,depth=-3.2pt,width=5.0pt}\mkern-16.5mu\int\nolimits_{\mathbb{R}^{n}}f(x)\ dx\quad\text{for $M(f)$}.$ Now, we state the result due to Birkhoff, which connects all the notions considered before, see [27]. ###### Theorem 2.12 (Birkhoff Ergodic Theorem). Let $f\in L^{1}_{\text{loc}}(\mathbb{R}^{n};L^{1}(\Omega))$, $($also $f\in L^{\infty}(\mathbb{R}^{n};L^{1}(\Omega)))$, be a stationary random variable. Then, for almost every $\widetilde{\omega}\in\Omega$ the function $f(\cdot,\widetilde{\omega})$ possesses a mean value in the sense of (2.16). Moreover, the mean value $M\left(f(\cdot,\widetilde{\omega})\right)$ as a function of $\widetilde{\omega}\in\Omega$ satisfies for almost every $\widetilde{\omega}\in\Omega$: i) Discrete case (i.e. $\tau:\mathbb{Z}^{n}\times\Omega\to\Omega$); $\mkern 12.0mu\mbox{\vrule height=4.0pt,depth=-3.2pt,width=5.0pt}\mkern-16.5mu\int\nolimits_{\mathbb{R}^{n}}f(x,\widetilde{\omega})\ dx=\mathbb{E}\left[\int_{[0,1)^{n}}f(y,\cdot)\,dy\right].$ ii) Continuous case (i.e. $\tau:\mathbb{R}^{n}\times\Omega\to\Omega$); $\mkern 12.0mu\mbox{\vrule height=4.0pt,depth=-3.2pt,width=5.0pt}\mkern-16.5mu\int\nolimits_{\mathbb{R}^{n}}f(x,\widetilde{\omega})\ dx=\mathbb{E}\left[f(0,\cdot)\right].$ The following lemma shows that, the Birkhoff Ergodic Theorem holds if a stationary function is composed with a stochastic deformation. ###### Lemma 2.13. Let $\Phi$ be a stochastic deformation and $f\in L^{\infty}_{\text{loc}}(\mathbb{R}^{n};L^{1}(\Omega))$ be a stationary random variable in the sense (2.13). Then for almost $\widetilde{\omega}\in\Omega$ the function $f\left(\Phi^{-1}(\cdot,\widetilde{\omega}),\widetilde{\omega}\right)$ possesses a mean value in the sense of (2.16) and satisfies: i) Discrete case; $\text{$\mkern 12.0mu\mbox{\vrule height=4.0pt,depth=-3.2pt,width=5.0pt}\mkern-16.5mu\int\nolimits_{\mathbb{R}^{n}}f\left(\Phi^{-1}(z,\widetilde{\omega}),\widetilde{\omega}\right)\,dz=\frac{\mathbb{E}\left[\int_{\Phi([0,1)^{n},\cdot)}f{\left(\Phi^{-1}\left(z,\cdot\right),\cdot\right)}\,dz\right]}{\det\left(\mathbb{E}\left[\int_{[0,1)^{n}}\nabla_{\\!\\!y}\Phi(y,\cdot)\,dy\right]\right)}$ \quad for a.a. $\widetilde{\omega}\in\Omega$}.$ ii) Continuous case; $\text{$\mkern 12.0mu\mbox{\vrule height=4.0pt,depth=-3.2pt,width=5.0pt}\mkern-16.5mu\int\nolimits_{\mathbb{R}^{n}}f\left(\Phi^{-1}(z,\widetilde{\omega}),\widetilde{\omega}\right)\,dz=\frac{\mathbb{E}\left[f(0,\cdot)\det\left(\nabla\Phi(0,\cdot)\right)\right]}{\det\left(\mathbb{E}\left[\nabla\Phi(0,\cdot)\right]\right)}$ \qquad for a.a. $\widetilde{\omega}\in\Omega$}.$ ###### Proof. See Blanc, Le Bris, Lions [9], also Andrade, Neves, Silva [6]. ∎ #### 2.2.2 Analysis of stationary functions In the rest of this paper, unless otherwise explicitly stated, we assume discrete dynamical systems and therefore, stationary functions are considered in this discrete sense. We begin the analysis of stationary functions with the concept of realization. ###### Definition 2.14. Let $f:\mathbb{R}^{n}\\!\times\\!\Omega\to\mathbb{R}$ be a stationary function. For $\omega\in\Omega$ fixed, the function $f(\cdot,\omega)$ is called a realization of $f$. Due to Theorem 2.12, almost every realization $f(\cdot,\omega)$ possesses a mean value in the sense of (2.16). On the other hand, if $f$ is a stationary function, then the mapping $y\in\mathbb{R}^{n}\mapsto\int_{\Omega}f(y,\omega)\,d\mathbb{P}(\omega)$ is a periodic function. In fact, it is enough to consider the realizations to study some properties of stationary functions. For instance, the following theorem will be used more than once through this paper. ###### Theorem 2.15. For $p>1$, let $u,v\in L^{1}_{\rm loc}(\mathbb{R}^{n};L^{p}(\Omega))$ be stationary functions. Then, for any $i\in\\{1,\ldots,n\\}$ fixed, the following sentences are equivalent: $(A)\quad\int_{[0,1)^{n}}\int_{\Omega}u(y,\omega)\frac{\partial{\zeta}}{\partial y_{i}}(y,\omega)\,d\mathbb{P}(\omega)\,dy=-\int_{[0,1)^{n}}\int_{\Omega}v(y,\omega)\,{\zeta}(y,\omega)\,d\mathbb{P}\,dy,\hskip 20.0pt$ (2.17) for each stationary function $\zeta\in C^{1}(\mathbb{R}^{n};L^{q}(\Omega))$, with $1/p+1/q=1$. $(B)\quad\int_{\mathbb{R}^{n}}u(y,\omega)\frac{\partial{\varphi}}{\partial y_{i}}(y)\,dy=-\int_{\mathbb{R}^{n}}v(y,\omega)\,{\varphi}(y)\,dy,\hskip 87.0pt$ (2.18) for any $\varphi\in C^{1}_{\rm c}(\mathbb{R}^{n})$, and almost sure $\omega\in\Omega$. ###### Proof. 1\. First, let us show that $(A)$ implies $(B)$. To begin, given $\gamma\in\mathbb{R}^{n}$, there exists a $\mathscr{F}$-measurable set $N_{\gamma}$ such that, $\mathbb{P}(N_{\gamma})=0$ and $\int_{\mathbb{R}^{n}}u(y,\omega)\,\frac{\partial{\varphi}}{\partial y_{i}}(y)\,dy=-\int_{\mathbb{R}^{n}}v(y,\omega)\,{\varphi}(y)\,dy,$ for each $\varphi\in C^{1}_{\rm c}((0,1)^{n}+\gamma)$ and $\omega\in\Omega\setminus N_{\gamma}$. Indeed, for $\varphi\in C^{1}_{\rm c}((0,1)^{n}+\gamma)$ and ${\rho\in L^{q}(\Omega)}$, let us define $\zeta_{\gamma}:\mathbb{R}^{n}\\!\times\\!\Omega\to\mathbb{R}$, by $\zeta_{\gamma}(y,\omega):=\varphi(y-\left\lfloor y-\gamma\right\rfloor)\rho(\tau(\left\lfloor y-\gamma\right\rfloor)\omega),$ where $\tau:\mathbb{Z}^{n}\times\Omega\to\Omega$ is a (discrete) dynamical system. Then, $\zeta_{\gamma}(\cdot,\omega)$ is a continuous functions, $\zeta_{\gamma}(y,\cdot)$ is a $\mathscr{F}$-measurable function, and for each $k\in\mathbb{Z}^{n}$, it follows that $\zeta_{\gamma}(y+k,\omega)=\zeta_{\gamma}(y,\tau(k)\omega).$ Consequently, $\zeta_{\gamma}\in C^{1}(\mathbb{R}^{n};L^{q}(\Omega))$ is a stationary Caratheodory function. Moreover, since $\left\lfloor y-\gamma\right\rfloor=0$ for each $y\in(0,1)^{n}+\gamma$, we have $\zeta_{\gamma}(y,\omega)=\varphi(y)\rho(\omega),$ (2.19) for each $(y,\omega)\in((0,1)^{n}+\gamma)\times\Omega$. Therefore, taking $\zeta_{\gamma}$ as a test function in (2.17), we obtain $\int_{[0,1)^{n}}\int_{\Omega}u(y,\omega)\,\frac{\partial{\zeta_{\gamma}}}{\partial y_{i}}(y,\omega)\,d\mathbb{P}(\omega)\,dy=-\int_{[0,1)^{n}}\int_{\Omega}v(y,\omega)\,{\zeta_{\gamma}}(y,\omega)\,d\mathbb{P}(\omega)\,dy.$ (2.20) Due to the space of stationary functions form an algebra, the functions $y\mapsto\int_{\Omega}u(y,\omega)\,\frac{\partial{\zeta_{\gamma}}}{\partial y_{i}}(y,\omega)\,d\mathbb{P}(\omega)\quad\text{and}\quad y\mapsto\int_{\Omega}v(y,\omega)\,{\zeta_{\gamma}}(y,\omega)\,d\mathbb{P}(\omega)$ are periodic, and hence translation invariants. Then, we have from (2.20) $\int_{(0,1)^{n}+\gamma}\int_{\Omega}u\,\frac{\partial{\varphi}}{\partial y_{i}}(y)\,{\rho}(\omega)\,d\mathbb{P}(\omega)\,dy=-\int_{(0,1)^{n}+\gamma}\int_{\Omega}v\,{\varphi}(y){\rho}(\omega)\,d\mathbb{P}(\omega)\,dy,$ where we have used (2.19). Applying Fubini’s Theorem, it follows that $\int_{\Omega}{\left(\int_{(0,1)^{n}+\gamma}u(y,\omega)\,\frac{\partial{\varphi}}{\partial y_{i}}(y)\,dy+\int_{(0,1)^{n}+\gamma}v(y,\omega)\,{\varphi}(y)\,dy\right)}{\rho}(\omega)\,d\mathbb{P}(\omega)=0$ for each $\rho\in L^{p}(\Omega)$. Therefore, for each $\varphi\in C^{1}_{\rm c}((0,1)^{n}+\gamma)$ there exists a set $N_{\gamma}\in\mathscr{F}$ with $\mathbb{P}(N_{\varphi})=0$, (which may depend on $\varphi$), such that, for each $\omega\in\Omega\setminus N_{\gamma}$ we have $\int_{(0,1)^{n}+\gamma}u(y,\omega)\,\frac{\partial{\varphi}}{\partial y_{i}}(y)\,dy=-\int_{(0,1)^{n}+\gamma}v(y,\omega)\,{\varphi}(y)\,dy.$ From a standard argument, we may remove the dependence on $N_{\gamma}$ of the test function $\varphi$. 2\. Finally, to pass from $\varphi\in C^{1}_{c}((0,1)^{n}+\gamma)$ to the case where $\varphi\in C^{1}_{\rm c}(\mathbb{R}^{n})$, we are going to use a standard procedure of partition of unity. We made it here, to become clear the argument in our case. Given $\varphi\in C^{1}_{\rm c}(\mathbb{R}^{n})$, since ${\rm supp}\,\varphi$ is a compact set, there exists $\left\\{\gamma_{j}\right\\}_{j=1}^{m}$ a finite subset of $\mathbb{R}^{n}$, such that ${\rm supp}\,\varphi\subset\bigcup_{j=1}^{m}\left((0,1)^{n}+\gamma_{j}\right).$ Then, we consider a partition of unity $\\{\theta_{j}\\}_{j=0}^{m}$ subordinated to this open covering, that is to say * i) $\theta_{j}\in C^{1}_{c}(\mathbb{R}^{n})$, $0\leqslant\theta_{j}\leqslant 1$, $j=0,\ldots,m$, * ii) ${\sum_{j=0}^{m}\theta_{j}(y)=1}$, for all $y\in\mathbb{R}^{n}$, * iii) ${\rm supp}\,\theta_{j}\subset(0,1)^{n}+\gamma_{j}$, $j=1,\ldots,m$, and ${\rm supp}\,\theta_{0}\subset\mathbb{R}^{n}\setminus{\rm supp}\,\varphi$. Since $\varphi=0$ on the support of $\theta_{0}$, it follows that, for each $y\in\mathbb{R}^{n}$ $\varphi(y)=\varphi(y)\sum_{i=1}^{m}\theta_{i}(y)=\sum_{i=1}^{m}(\varphi\theta_{i})(y).$ (2.21) Moreover, from item 1, there exist sets $N_{\gamma_{1}},\ldots,N_{\gamma_{m}}\in\mathscr{F}$ with $\mathbb{P}(N_{\gamma_{j}})=0$, for any $j\in\\{1,\ldots,m\\}$, such that $\int_{\mathbb{R}^{n}}u(y,\omega)\,\frac{\partial({\varphi\theta_{j}})}{\partial y_{i}}\,dy=-\int_{\mathbb{R}^{n}}v(y,\omega)\,({\varphi\theta_{j}})\,dy,$ for each $\omega\in\Omega\setminus N_{\gamma_{j}}$. To follow, we define $N:=\bigcup_{j=1}^{m}N_{\gamma_{j}}$ (which may depend on $\varphi$), then $\mathbb{P}(N)=0$ and summing from 1 to $m$, we obtain from the above equation $\sum_{j=1}^{m}\int_{\mathbb{R}^{n}}u(y,\omega)\,\frac{\partial({\varphi\theta_{j}})(y)}{\partial y_{i}}\,dy=-\sum_{j=1}^{m}\int_{\mathbb{R}^{n}}v(y,\omega)\,({\varphi\theta_{j}})(j)\,dy,$ for each $\omega\in\Omega\setminus N$. Therefore, since the above sum is finite and using (2.21), we obtain $\int_{\mathbb{R}^{n}}u(y,\omega)\,\frac{\partial{\varphi(y)}}{\partial y_{i}}\,dy=-\int_{\mathbb{R}^{n}}v(y,\omega)\,\varphi(y)\,dy.$ Again, due to a standard argument, we may remove the dependence on $N$ with the test function $\varphi$. Consequently, we have obtained (2.18), more precisely sentence $(B)$. 3\. Now, let us show sentence $(A)$ from $(B)$. For each $\ell\in\mathbb{N}$, we define the set ${Q}_{\ell}:=(-\ell,\ell)^{n}$ and the function $\chi_{\ell}\in C^{1}_{c}(\mathbb{R}^{n})$, such that $\chi_{\ell}\equiv 1$ in ${Q}_{\ell}$, $\chi_{\ell}\equiv 0$ in $\mathbb{R}^{n}\setminus{Q}_{\ell+1}$, and ${\|\nabla\chi_{\ell}\|_{\infty}\leqslant 2}$. Then, given $\zeta\in C^{1}(\mathbb{R}^{n};L^{q}(\Omega))$ and $i\in\\{1,\ldots,n\\}$, we consider $\zeta(\cdotp,\omega)\chi_{\ell}$, (for $\ell\in\mathbb{N}$ and $\omega\in\Omega$ fixed), as test function in (2.18), that is $\int_{\mathbb{R}^{n}}u(y,\omega)\,\frac{\partial}{\partial y_{i}}{\left({\zeta}(y,\omega)\chi_{\ell}(y)\right)}\,dy=-\int_{\mathbb{R}^{n}}v(y,\omega)\,{{\zeta}(y,\omega)\chi_{\ell}(y)}\,dy.$ From the definition of $\chi_{\ell}$, and applying the product rule we obtain $\displaystyle\int_{Q_{\ell+1}}u(y,\omega)\frac{\partial{\zeta(y,\omega)}}{\partial y_{i}}\,\chi_{\ell}(y)\,dy$ $\displaystyle+\int_{Q_{\ell+1}\setminus Q_{\ell}}u(y,\omega){\zeta}(y,\omega)\,\frac{\partial\chi_{\ell}(y)}{\partial y_{i}}\,dy$ $\displaystyle=-\int_{Q_{\ell+1}}v(y,\omega)\,{{\zeta}(y,\omega)\chi_{\ell}(y)}\,dy,$ or conveniently using that $Q_{\ell+1}=Q_{\ell}\cup(Q_{\ell+1}\setminus Q_{\ell})$, we have $\displaystyle\int_{Q_{\ell}}u(y,\omega)\,\frac{\partial{\zeta}}{\partial y_{i}}(y,\omega)\,dy$ $\displaystyle+\int_{Q_{\ell}}v(y,\omega)\,{\zeta}(y,\omega)\,dy$ (2.22) $\displaystyle=-\int_{Q_{\ell+1}\setminus Q_{\ell}}u(y,\omega)\,\frac{\partial{\zeta}}{\partial y_{i}}(y,\omega)\,\chi_{\ell}(y)\,dy$ $\displaystyle\quad-\int_{Q_{\ell+1}\setminus Q_{\ell}}u(y,\omega)\,{\zeta}(y,\omega)\,\frac{\partial\chi_{\ell}(y)}{\partial y_{i}}\,dy$ $\displaystyle\quad-\int_{Q_{\ell+1}\setminus Q_{\ell}}v(y,\omega)\,{\zeta}(y,\omega)\,\chi_{\ell}(y)\,dy$ $\displaystyle=I_{1}(\omega)+I_{2}(\omega)+I_{3}(\omega),$ with obvious notation. Claim: For $j=1,2,3$, $\lim_{\ell\to\infty}\int_{\Omega}\frac{|I_{j}(\omega)|}{|Q_{\ell}|}d\mathbb{P}(\omega)=0.$ Proof of Claim: Let us show for $j=2$, that is $\lim_{\ell\to\infty}\int_{\Omega}\frac{1}{|Q_{\ell}|}{\Big{|}\int_{Q_{\ell+1}\setminus Q_{\ell}}u(y,\omega)\,{\zeta}(y,\omega)\,\frac{\partial\chi_{\ell}(y)}{\partial y_{i}}\,dy\Big{|}}d\mathbb{P}(\omega)=0,$ the others are similar. Then, applying Fubini’s Theorem $\displaystyle\int_{\Omega}\frac{1}{|Q_{\ell}|}\Big{|}\int_{Q_{\ell+1}\setminus Q_{\ell}}$ $\displaystyle u(y,\omega)\,{\zeta}(y,\omega)\,\frac{\partial\chi_{\ell}(y)}{\partial y_{i}}\,dy\Big{|}d\mathbb{P}(\omega)$ $\displaystyle\leq\,\frac{1}{|Q_{\ell}|}\int_{Q_{\ell+1}\setminus Q_{\ell}}\int_{\Omega}{|u(\cdotp,\omega)\,{\zeta}(\cdotp,\omega)|}\,{\|\nabla\chi_{\ell}\|}_{\infty}\,d\mathbb{P}\,dy$ $\displaystyle\leq\,\frac{2}{|Q_{\ell}|}\int_{Q_{\ell+1}\setminus Q_{\ell}}\int_{\Omega}{|u(y,\omega)\,{\zeta}(y,\omega)|}\,d\mathbb{P}(\omega)\,dy$ $\displaystyle=\frac{2\,{((2(\ell+1))^{n}-(2\ell)^{n})}}{(2\ell)^{n}}\\!\\!\int_{[0,1)^{n}}\int_{\Omega}{|u(y,\omega)\,{\zeta}(y,\omega)|}\,d\mathbb{P}(\omega)\,dy$ $\displaystyle=\,{2\,{{((1+\ell^{-1})^{n}-1)}}}\int_{[0,1)^{n}}\int_{\Omega}{|u(y,\omega)\,{\zeta}(y,\omega)|}\,d\mathbb{P}(\omega)\,dy,$ from which, passing to the limit as $\ell\to 0$, follows the claim. Then, dividing equation (2.22) by $|Q_{\ell}|$ and integrating in $\Omega$, we obtain $\liminf_{\ell\to\infty}\int_{\Omega}{\Big{|}\frac{1}{|Q_{\ell}|}\int_{Q_{\ell}}(u\,\frac{\partial{\zeta}}{\partial y_{i}})(y,\omega)dy+\frac{1}{|Q_{\ell}|}\int_{Q_{\ell}}(v\,{\zeta})(y,\omega)dy\Big{|}}d\mathbb{P}(\omega)=0,$ and applying Fato’s Lemma $\int_{\Omega}\liminf_{\ell\to\infty}\int_{\Omega}{\Big{|}\frac{1}{|Q_{\ell}|}\int_{Q_{\ell}}(u\,\frac{\partial{\zeta}}{\partial y_{i}})(y,\omega)dy+\frac{1}{|Q_{\ell}|}\int_{Q_{\ell}}(v\,{\zeta})(y,\omega)\,dy\Big{|}}d\mathbb{P}(\omega)=0.$ Therefore, there exists a $\mathscr{F}$-measurable set $\widetilde{\Omega}\subset\Omega$ of full measure, such that, for each $\omega\in\widetilde{\Omega}$, we have $\liminf_{\ell\to\infty}{\Big{|}\frac{1}{|Q_{\ell}|}\int_{Q_{\ell}}u(y,\omega)\,\frac{\partial{\zeta}}{\partial y_{i}}(y,\omega)\,dy+\frac{1}{|Q_{\ell}|}\int_{Q_{\ell}}v(y,\omega)\,{\zeta}(y,\omega)\,dy\Big{|}}=0.$ Then, applying Theorem 2.12 and from equation (2.16), it follows that $\int_{[0,1)^{n}}\int_{\Omega}u(y,\omega)\,\frac{\partial{\zeta}}{\partial y_{i}}(y,\omega)\,d\mathbb{P}(\omega)\,dy=-\int_{[0,1)^{n}}\int_{\Omega}v(y,\omega)\,{\zeta}(y,\omega)\,d\mathbb{P}(\omega)\,dy,$ which finish the proof of the theorem. ∎ Similarly to the above theorem, we have the characterization of weak derivatives of stationary functions composed with stochastic deformations, given by the following ###### Theorem 2.16. Let $u,v\in L^{1}_{\rm loc}(\mathbb{R}^{n};L^{p}(\Omega))$ be stationary functions, $(p>1)$. Then, for any $i\in\\{1,\ldots,n\\}$ fixed, the following sentences are equivalent: $(A)\quad\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}u{\left(\Phi^{-1}\left(z,\omega\right),\omega\right)}\,{\frac{\partial{\left(\zeta{\left(\Phi^{-1}(z,\omega),\omega\right)}\right)}}{\partial z_{k}}}\,dz\,d\mathbb{P}(\omega)\\\\[5.0pt] =-\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}v{\left(\Phi^{-1}\left(z,\omega\right),\omega\right)}\,{\zeta{\left(\Phi^{-1}(z,\omega),\omega\right)}}\,dz\,d\mathbb{P}(\omega),$ for each stationary function $\zeta\in C^{1}(\mathbb{R}^{n};L^{q}(\Omega))$, with $1/p+1/q=1$. $(B)\quad\int_{\mathbb{R}^{n}}u{\left(\Phi^{-1}\left(z,\omega\right),\omega\right)}\,{\frac{\partial\varphi}{\partial z_{k}}(z)}\,dz=-\int_{\mathbb{R}^{n}}v{\left(\Phi^{-1}\left(z,\omega\right),\omega\right)}\,{\varphi(z)}\,dz,\hskip 12.0pt$ for any $\varphi\in C^{1}_{\rm c}(\mathbb{R}^{n})$, and almost sure $\omega\in\Omega$. ###### Proof. The proof follows the same lines as in the proof of Theorem 2.15 after the change of variables $y=\Phi^{-1}(z,\omega)$. ∎ ### 2.3 $\Phi_{\omega}-$Two-scale Convergence In this subsection, we shall consider the two-scale convergence in a stochastic setting that is beyond of the classical stationary ergodic setting. The classical concept of two-scale convergence was introduced by Nguetseng [29] and futher developed by Allaire [1] to deal with periodic problems. The notion of two-scale convergence has been successfully extended to non- periodic settings in several papers as in [20, 16] in the ergodic algebra setting and in [11] in the stochastic setting. The main difference here with the earlier studies is that the test functions used here are random perturbations accomplished by stochastic diffeomorphisms of stationary functions. The main difficulty brought by this kind of test function is the lack of the stationarity property (see [6] for a deep discussion about that) which makes us unable to use the results described in [11] and the lack of a compatible topology with the probability space considered. This is overcome by using a compactification argument that preserves the ergodic nature of the setting involved. For this, we will make use of the following lemma, whose simple proof can be found in [5]. ###### Lemma 2.17. Let $X_{1},X_{2}$ be compact spaces, $R_{1}$ a dense subset of $X_{1}$ and $W:R_{1}\to X_{2}$. Suppose that for all $g\in C(X_{2})$ the function $g\circ W$ is the restriction to $R_{1}$ of some (unique) $g_{1}\in C(X_{1})$. Then $W$ can be uniquely extended to a continuous mapping $\underline{W}:X_{1}\to X_{2}$. Further, suppose in addition that $R_{2}$ is a dense set of $X_{2}$, $W$ is a bijection from $R_{1}$ onto $R_{2}$ and for all $f\in C(X_{1})$, $f\circ W^{-1}$ is the restriction to $R_{2}$ of some (unique) $f_{2}\in C(X_{2})$. Then, $W$ can be uniquely extended to a homeomorphism $\underline{W}:X_{1}\to X_{2}$. Now, we can prove the following result. ###### Theorem 2.18. Let $\mathbb{S}\subset L^{\infty}(\mathbb{R}^{n}\times\Omega)$ be a countable set of stationary functions. Then there exists a compact (separable) topological space $\widetilde{\Omega}$ and one-to-one function $\delta:\Omega\to\widetilde{\Omega}$ with dense image satisfying the following properties: 1. (i) The probability space $\Big{(}\Omega,\mathscr{F},\mathbb{P}\Big{)}$ and the ergodic dynamical system $\tau:\mathbb{Z}^{n}\times\Omega\to\Omega$ acting on it extends respectively to a Radon probability space $\Big{(}\widetilde{\Omega},\mathscr{B},\widetilde{\mathbb{P}}\Big{)}$ and to an ergodic dynamical system $\widetilde{\tau}:\mathbb{Z}^{n}\times\widetilde{\Omega}\to\widetilde{\Omega}$. 2. (ii) The stochastic deformation $\Phi:\mathbb{R}^{n}\times\Omega\to\mathbb{R}^{n}$ extends to a stochastic deformation $\tilde{\Phi}:\mathbb{R}^{n}\times\widetilde{\Omega}\to\mathbb{R}^{n}$ satisfying $\Phi(x,\omega)=\tilde{\Phi}(x,\delta(\omega)),$ for a.e. $\omega\in\Omega$. 3. (iii) Any function $f\in\mathbb{S}$ extends to a $\tilde{\tau}-$stationary function $\tilde{f}\in L^{\infty}(\widetilde{\Omega}\times\mathbb{R}^{n})$ satisfying $\mkern 12.0mu\mbox{\vrule height=4.0pt,depth=-3.2pt,width=5.0pt}\mkern-16.5mu\int\nolimits_{\mathbb{R}^{n}}f\left(\Phi^{-1}(z,\omega),\omega\right)\,dz=\mkern 12.0mu\mbox{\vrule height=4.0pt,depth=-3.2pt,width=5.0pt}\mkern-16.5mu\int\nolimits_{\mathbb{R}^{n}}\tilde{f}\left(\tilde{\Phi}^{-1}(z,\delta(\omega)),\delta(\omega)\right)\,dz,$ for a.e. $\omega\in\Omega$. ###### Proof. 1\. Let $\mathbb{S}$ be the set of the lemma. Given $f\in\mathbb{S}$, define $f_{j}(y,\omega):=\int_{\mathbb{R}^{n}}f(y+x,\omega)\,\rho_{j}(x)\,dx,$ where $\rho_{j}$ is the classical approximation of the identity in $\mathbb{R}^{n}$. Note that for a.e. $y\in\mathbb{R}^{n}$, we have that $f_{j}(y,\cdot)\to f(y,\cdot)$ in $L^{1}(\Omega)$ as $j\to\infty$. Define $\mathcal{A}$ as the closed algebra with unity generated by the set $\Big{\\{}f_{j}(y,\cdot);\,j\geq 1,y\in\mathbb{Q}^{n},f\in\mathbb{S}\Big{\\}}\cap\Big{\\{}\partial_{j}\Phi_{i}(y,\cdot);\,1\leq j,i\leq n,y\in\mathbb{Q}^{n}\Big{\\}}.$ Since $[-1,1]$ is a compact set, by the well known Tychonoff’s Theorem, the set $[-1,1]^{\mathcal{A}}:=\Big{\\{}\text{the functions $\gamma:\mathcal{A}\to[-1,1]$}\Big{\\}}$ is a compact set in the product topology. Define $\delta:\Omega\to[-1,1]^{\mathcal{A}}$ by $\delta(\omega)(g):=\left\\{\begin{array}[]{rc}\frac{g(\omega)}{\|g{\|}_{\infty}},&\mbox{if}\quad g\neq 0,\\\ 0,&\mbox{if}\quad g=0.\end{array}\right.$ We may assume that the algebra $\mathcal{A}$ distinguishes between points of $\Omega$, that is, given any two distinct points $\omega_{1},\omega_{2}\in\Omega$, there exists $g\in\mathcal{A}$ such that $g(\omega_{1})\neq g(\omega_{2})$. In the case that it is not true we may replace $\Omega$ by its quotient by a trivial equivalence relation, in a standard way, and we proceed correspondingly with the $\sigma-$algebra $\mathscr{F}$ and with the probability measure $\mathbb{P}$. Thus, the function $\delta$ is one-to-one. Define $\widetilde{\Omega}:=\overline{\delta(\Omega)}.$ Now, we can see that the set $\Omega$ inherits all topological features of the compact space $\widetilde{\Omega}$ in a natural way which allows us to identify it homeomorphically with the image $\delta(\Omega)$. 2\. Define the mapping $i:\mathcal{A}\to C(\delta(\Omega))$ by $i(g)(\delta(\omega)):=g(\omega).$ We claim that there exists a continuous function $\tilde{g}:\widetilde{\Omega}\to\mathbb{R}$ such that $i(g)=\tilde{g}\,\text{on $\delta(\Omega)$}.$ In fact, take $g\in\mathcal{A}$ and $Y:=\overline{g(\Omega)}$. Define the function $f^{*}:C(Y)\to\mathcal{A}$ by $f^{*}(h):=h\circ g\,\text{(the algebra structure is used!)}$ Hence, we can define $f^{**}:[-1,1]^{\mathcal{A}}\to[-1,1]^{C(Y)}$ by $f^{**}(h):=h\circ f^{*}.$ Note that the function $f^{**}$ is a continuous function. In order to see that, we highlighted that it is known that a function $H$ from a topological space to a product space $\otimes_{\alpha\in\mathcal{I}}X_{\alpha}$ is continuous if and only if each component $\pi_{\alpha}\circ H:=H_{\alpha}$ is continuous. Hence, if $\alpha\in C(Y)$ then the projection function $f^{**}_{\alpha}$ must satisfy $\displaystyle f^{**}_{\alpha}(h):=\left(\pi_{\alpha}\circ f^{**}\right)(h)=\pi_{\alpha}\circ\left(f^{**}(h)\right)=\pi_{\alpha}\left(h\circ f^{*}\right)$ $\displaystyle\qquad=\left(h\circ f^{*}\right)(\alpha)=h\left(f^{*}(\alpha)\right)=h\left(\alpha\circ g\right)=\pi_{\alpha\circ g}(h).$ Now, consider the function $\tilde{\delta}:Y\to[-1,1]^{C(Y)}$ given by $\tilde{\delta}(y)(h):=\left\\{\begin{array}[]{rc}\frac{h(y)}{\|h{\|}_{\infty}},&\mbox{if}\quad h\neq 0,\\\ 0,&\mbox{if}\quad h=0.\end{array}\right.$ Since the algebra $C(Y)$ has the following property: If $F\subset Y$ is a closed set and $y\notin F$ then $f(y)\notin\overline{f(F)}$ for some $f\in C(Y)$, we can conclude that the function $\tilde{\delta}$ is a homeomorphism onto its image. Furthermore, given $\omega\in\Omega$ we have that $\left(f^{**}\circ\delta\right)(\omega)=f^{**}\left(\delta(\omega)\right)=\delta(\omega)\circ f^{*}$. Hence, if $0\neq h\in C(Y)$ it follows that $\displaystyle\left(f^{**}\circ\delta\right)\big{(}\omega\big{)}(h)=\left(\delta(\omega)\circ f^{*}\right)(h)=\delta(\omega)\left(f^{*}(h)\right)$ $\displaystyle\quad=\delta(\omega)\left(h\circ g\right)=\frac{\left(h\circ g\right)(\omega)}{\|h\circ g{\|}_{\infty}}=\tilde{\delta}\left(g(\omega)\right)(h).$ Thus, we see that ${\tilde{\delta}}^{-1}\circ f^{**}=i(g)$. Defining $\tilde{g}:={\tilde{\delta}}^{-1}\circ f^{**}$ we have clearly that $i:\mathcal{A}\to C(\widetilde{\Omega})$ is a one-to-one isometry satisfying $i(g)=\tilde{g}$ and our claim is proved. Moreover, it is easy to see that $i(\mathcal{A})$ is an algebra of functions over $C(\widetilde{\Omega})$ containing the unity. As before, if $i(\mathcal{A})$ does not separate points of $\widetilde{\Omega}$, then we may replace $\widetilde{\Omega}$ by its quotient $\widetilde{\Omega}/\sim$, where $\tilde{\omega_{1}}\sim\tilde{\omega_{2}}\,\Leftrightarrow\,\tilde{g}(\tilde{\omega_{1}})=\tilde{g}(\tilde{\omega_{2}})\,\forall g\in\mathcal{A}.$ Therefore, we can assume that $i(\mathcal{A})$ separates the points of $\widetilde{\Omega}$. Hence, by the Stone’s Theorem (see [Ru2], p. 162, Theorem 7.3.2) we must have $i(\mathcal{A})=C(\widetilde{\Omega})$. 3\. Define $\widetilde{\tau}:\mathbb{Z}^{n}\times\delta(\Omega)\to\delta(\Omega)$ by $\widetilde{\tau}_{k}\left(\delta(\omega)\right):=\delta(\tau_{k}\omega).$ It is easy to see that $\widetilde{\tau}_{k_{1}+k_{2}}(\delta(\omega))=\widetilde{\tau}_{k_{1}}\Big{(}\widetilde{\tau}_{k_{2}}(\delta(\omega))\Big{)},$ for all $k_{1},k_{2}\in\mathbb{Z}^{n}$ and $\omega\in\Omega$. Since $\tilde{g}\circ{\widetilde{\tau}_{k}}=\widetilde{g\circ{\tau}_{k}}$ for all $g\in\mathcal{A}$ and $k\in\mathbb{Z}^{n}$, the lemma 2.17 allows us to extend the mapping $\widetilde{\tau}_{k}$ from $\delta(\Omega)$ to $\widetilde{\Omega}$ satisfying the group property $\widetilde{\tau}_{k_{1}+k_{2}}=\widetilde{\tau}_{k_{1}}\circ{\widetilde{\tau}}_{k_{2}}$. Given a borelian set $\tilde{A}\subset{\widetilde{\Omega}}$ and defining $\widetilde{\mathbb{P}}(\tilde{A}):=\mathbb{P}(\delta^{-1}(\tilde{A}\cap\delta(\Omega)))$, we can deduce that $\widetilde{\mathbb{P}}\circ{\widetilde{\tau}_{k}}=\widetilde{\mathbb{P}}$. Thus, the mapping $\widetilde{\tau}_{k}$ is an ergodic dynamical system over the Radon probability space $\Big{(}\widetilde{\Omega},\mathscr{B},\widetilde{\mathbb{P}}\Big{)}$. Thus, we have concluded the proof of the item (i). 4\. Now, note that for each $\omega\in\widetilde{\Omega}$ and each integer $j\geq 1$, the function $f_{j}(\cdot,\omega)$ is uniformly continuous over $\mathbb{Q}^{n}$. Hence, it can be extended uniquely to a function $\widetilde{f_{j}}(\cdot,\omega)$ defined in $\mathbb{R}^{n}$ that satisfies $\limsup_{j,l\to\infty}\int_{[0,1)^{n}\times\widetilde{\Omega}}|\widetilde{f_{j}}(y,\omega)-\widetilde{f_{l}}(y,\omega)|\,d\widetilde{\mathbb{P}}(\omega)\,dy=0.$ Therefore, there exists a $\widetilde{\tau}$-stationary function $\widetilde{f}\in L^{1}_{\text{loc}}\left(\mathbb{R}^{n}\times\widetilde{\Omega}\right)$, such that $\widetilde{f_{j}}\to\widetilde{f}$ as $j\to\infty$ in $L^{1}_{\text{loc}}(\mathbb{R}^{n}\times\widetilde{\Omega})$. Since $\|\widetilde{f_{j}}{\|}_{\infty}\leq\|{f}{\|}_{\infty}$, for all $j\geq 1$ we have that $\widetilde{f}\in L^{\infty}(\mathbb{R}^{n}\times\widetilde{\Omega})$. In the same way, the stochastic deformation $\Phi:\mathbb{R}^{n}\times\Omega\to\mathbb{R}^{n}$ extends to a stochastic deformation $\tilde{\Phi}:\mathbb{R}^{n}\times\widetilde{\Omega}\to\mathbb{R}^{n}$ satisfying $\Phi(y,\omega)=\tilde{\Phi}(y,\delta(\omega))$ for all $\omega\in\Omega$ and $\mkern 12.0mu\mbox{\vrule height=4.0pt,depth=-3.2pt,width=5.0pt}\mkern-16.5mu\int\nolimits_{\mathbb{R}^{n}}f\left(\Phi^{-1}(z,\omega),\omega\right)\,dz=\mkern 12.0mu\mbox{\vrule height=4.0pt,depth=-3.2pt,width=5.0pt}\mkern-16.5mu\int\nolimits_{\mathbb{R}^{n}}\tilde{f}\left(\tilde{\Phi}^{-1}(z,\delta(\omega)),\delta(\omega)\right)\,dz,$ for a.e. $\omega\in\Omega$. This, completes the proof of the theorem (2.18). ∎ In practice, in our context, the set $\mathbb{S}$ shall be a countable set generated by the coefficients of our equation $\eqref{jhjkhkjhkj765675233}$ and by the eigenfunctions of the spectral equation associated to it. Thus, the Theorem (2.18) allow us to suppose without loss of generality that our probability space $\Big{(}\Omega,\mathscr{F},\mathbb{P}\Big{)}$ is a separable compact space. Using the Ergodic Theorem, given a stationary function $f\in L^{\infty}(\mathbb{R}^{n}\times\Omega)$ there exists a set of full measure $\Omega_{f}\subset\Omega$ such that $\mkern 12.0mu\mbox{\vrule height=4.0pt,depth=-3.2pt,width=5.0pt}\mkern-16.5mu\int\nolimits_{\mathbb{R}^{n}}f\left(\Phi^{-1}(z,\tilde{\omega}),\tilde{\omega}\right)\,dz=c_{\Phi}^{-1}\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}f\left(\Phi^{-1}(z,\omega),\omega\right)\,dz\,d\mathbb{P}(\omega),$ (2.23) for almost all $\tilde{\omega}\in{\Omega}_{f}$. Due to the separability of the probability compact space $\Big{(}\Omega,\mathscr{F},\mathbb{P}\Big{)}$, we can find a set $\mathbb{D}\subset C_{b}(\mathbb{R}^{n}\times\Omega)$ such that: * • Each $f\in\mathbb{D}$ is a stationary function. * • $\mathbb{D}$ is a countable and dense set in $C_{0}\big{(}[0,1)^{n}\times\Omega\big{)}$. In this case, there exists a set $\Omega_{0}\subset\Omega$ of full measure such that the equality (2.23) holds for any $\tilde{\omega}\in\Omega_{0}$ and $f\in\mathbb{D}$. Now, we proceed with the definition of the two-scale convergence in this scenario of stochastically deformed. In what follows, the set $O\subset\mathbb{R}^{n}$ is an open set. ###### Definition 2.19. Let $1<p<\infty$ and $v_{\varepsilon}:O\times\Omega\to\mathbb{C}$ be a sequence such that $v_{\varepsilon}(\cdot,\tilde{\omega})\in L^{p}(O)$. The sequence $\\{v_{\varepsilon}(\cdot,\tilde{\omega}){\\}}_{\varepsilon>0}$ is said to $\Phi_{\omega}-$two-scale converges to a stationary function $V_{\tilde{\omega}}\in L^{p}\left(O\times[0,1)^{n}\times\Omega\right)$, when for a.e. $\tilde{\omega}\in\Omega$ holds the following $\displaystyle\lim_{\varepsilon\to 0}\int_{O}v_{\varepsilon}(x,\tilde{\omega})\,\varphi(x)\,\Theta\left(\Phi^{-1}\left(\frac{x}{\varepsilon},\tilde{\omega}\right),\tilde{\omega}\right)\,dx$ $\displaystyle=c_{\Phi}^{-1}\int_{O\times\Omega}\int_{\Phi([0,1)^{n},\omega)}\\!\\!\\!V_{\tilde{\omega}}\left(x,\Phi^{-1}\left(z,\omega\right),\omega\right)\,\varphi(x)\,\Theta(\Phi^{-1}(z,\omega),\omega)\,dz\,d{\mathbb{P}(\omega)}\,dx,$ for all $\varphi\in C_{c}^{\infty}(O)$ and $\Theta\in L^{q}_{\text{loc}}(\mathbb{R}^{n}\times\Omega)$ stationary. Here, $p^{-1}+q^{-1}=1$ and $c_{\Phi}:=\det\Big{(}\int_{[0,1)^{n}\times\Omega}\nabla\Phi(y,\omega)\,d{\mathbb{P}(\omega)}\,dy\Big{)}$. ###### Remark 2.20. From now on, we shall use the notation $v_{\varepsilon}(x,\widetilde{\omega})\;\xrightharpoonup[\varepsilon\to 0]{2-{\rm s}}\;V_{\widetilde{\omega}}{\left(x,\Phi^{-1}(z,\omega),\omega\right)},$ to indicate that $v_{\varepsilon}(\cdot,\tilde{\omega})$ $\Phi_{\omega}-$two- scale converges to $V_{\tilde{\omega}}$. The most important result about the two-scale convergence needed in this paper is the following compactness theorem which generalize the corresponding one for the deterministic case in [16] (see Theorem 4.8) and the corresponding one for the stochastic case in [11] (see Theorem 3.4). ###### Theorem 2.21. Let $1<p<\infty$ and $v_{\varepsilon}:O\times\Omega\to\mathbb{C}$ be a sequence such that $\sup_{\varepsilon>0}\int_{O}|v_{\varepsilon}(x,\tilde{\omega})|^{p}\,dx<\infty,$ for almost all $\tilde{\omega}\in\Omega$. Then, for almost all $\tilde{\omega}\in\Omega_{0}$, there exists a subsequence $\\{v_{\varepsilon^{\prime}}(\cdot,\tilde{\omega}){\\}}_{\varepsilon^{\prime}>0}$, which may depend on $\tilde{\omega}$, and a stationary function $V_{\tilde{\omega}}\in L^{p}(O\times[0,1)^{n}\times\Omega)$, such that $v_{\varepsilon}(x,\widetilde{\omega})\;\xrightharpoonup[\varepsilon^{\prime}\to 0]{2-{\rm s}}\;V_{\widetilde{\omega}}{\left(x,\Phi^{-1}(z,\omega),\omega\right)}.$ ###### Proof. 1\. We begin by fixing $\tilde{\omega}\in\Omega_{0}$. Due to our assumption, there exists ${c(\widetilde{\omega})>0}$, such that for all $\varepsilon>0$ ${\left\|v_{\varepsilon}(\cdot,\widetilde{\omega})\right\|}_{L^{p}(O)}\leqslant c(\widetilde{\omega}).$ Now, taking $\phi\in\Xi\times\mathbb{D}$ with $\Xi\subset C_{c}^{\infty}(O)$ dense in $L^{q}(O)$, we have after applying the Hölder inequality and the Ergodic Theorem, $\underset{\varepsilon\to 0}{\limsup}{|\int_{O}v_{\varepsilon}(x,\widetilde{\omega})\,{\phi{\left(x,\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega}\right)}}dx|}\\\ \leq c(\widetilde{\omega}){\left[\underset{\varepsilon\to 0}{\limsup}\int_{O}{\left|\phi{\left(x,\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega}\right)}\right|}^{q}dx\right]}^{1/q}\hskip 100.0pt\\\ =c(\widetilde{\omega}){\left[c_{\Phi}^{-1}\int_{O}\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}{\left|\phi{\left(x,\Phi^{-1}(z,\omega),\omega\right)}\right|}^{q}dz\,d\mathbb{P}(\omega)\,dx\right]}^{1/q}.$ (2.24) Thus, the use of the enumerability of the set $\Xi\times\mathbb{D}$ combined with a diagonal argument yields us a subsequence $\\{\varepsilon^{\prime}\\}$ (maybe depending of $\tilde{\omega}$) such that the functional $\mu:\Xi\times\mathbb{D}\to\mathbb{C}$ given by $\langle\mu,\phi\rangle:=\lim_{\varepsilon^{\prime}\to 0}\int_{O}v_{\varepsilon^{\prime}}(x,\widetilde{\omega})\,{\phi{\left(x,\Phi^{-1}{\left(\frac{x}{\varepsilon^{\prime}},\widetilde{\omega}\right)},\widetilde{\omega}\right)}}dx$ (2.25) is well-defined and bounded with respect to the norm $\|\cdot{\|}_{q}$ defined as $\|\phi{\|}_{q}:=\big{[}c_{\Phi}^{-1}\int_{O}\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}{\left|\phi{\left(x,\Phi^{-1}(z,\omega),\omega\right)}\right|}^{q}dz\,d\mathbb{P}(\omega)\,dx\big{]}^{1/q}$ by (2.24). Since the set $\Xi\times\mathbb{D}$ is dense in $L^{q}\left(O\times[0,1)^{n}\times\Omega\right)$, we can extend the functional $\mu$ to a bounded functional $\tilde{\mu}$ over $L^{q}\left(O\times[0,1)^{n}\times\Omega\right)$. Hence, we find $V_{\tilde{\omega}}\in L^{p}(O\times[0,1)^{n}\times\Omega)$ which can be extended to $O\times\mathbb{R}^{n}\times\Omega$ in a stationary way by setting $V_{\tilde{\omega}}(x,y,\omega)=V_{\tilde{\omega}}\left(x,y-\left\lfloor y\right\rfloor,\tau_{\left\lfloor y\right\rfloor}\omega\right),$ and satisfying for all $\phi\in L^{q}\left(O\times[0,1)^{n}\times\Omega\right)$, $\langle\tilde{\mu},\phi\rangle\\!\\!=c_{\Phi}^{-1}\int_{O\times\Omega}\int_{\Phi([0,1)^{n},\omega)}\\!\\!\\!\\!V_{\tilde{\omega}}\left(x,\Phi^{-1}\left(z,\omega\right),\omega\right)\phi\left(x,\Phi^{-1}(z,\omega),\omega\right)dzd\mathbb{P}(\omega)dx.$ 2\. Now, take $\varphi\in C^{\infty}_{c}(O)$ and $\Theta\in L^{q}_{\text{loc}}(\mathbb{R}^{n}\times\Omega)$ a $\tau$-stationary function. Since the set $\Xi\times\mathbb{D}$ is dense in $L^{q}\left(O\times[0,1)^{n}\times\Omega\right)$, we can pick up a sequence $\\{(\varphi_{j},\Theta_{j}){\\}}_{j\geq 1}\subset\Xi\times\mathbb{D}$ such that $\lim_{j\to\infty}(\varphi_{j},\Theta_{j})=(\varphi,\Theta)\quad\text{in $L^{q}\Big{(}O\times[0,1)^{n}\times\Omega\Big{)}$}.$ Then, observing that $\displaystyle\limsup_{\varepsilon^{\prime}\to 0}\Big{|}\int_{O}v_{\varepsilon^{\prime}}(x,\tilde{\omega})\varphi(x)\Theta\left(\Phi^{-1}\left(\frac{x}{\varepsilon^{\prime}},\tilde{\omega}\right),\tilde{\omega}\right)\,dx$ $\displaystyle-\int_{O}v_{\varepsilon^{\prime}}(x,\tilde{\omega})\varphi_{j}(x)\Theta_{j}\left(\Phi^{-1}\left(\frac{x}{\varepsilon^{\prime}},\tilde{\omega}\right),\tilde{\omega}\right)\,dx\Big{|}$ $\displaystyle\leq c\|\varphi-\varphi_{j}{\|}_{L^{q}(O)}[c_{\Phi}^{-1}\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}{\left|(\Theta-\Theta_{j}){\left(\Phi^{-1}(z,\omega),\omega\right)}\right|}^{q}dz\,d\mathbb{P}(\omega)]^{1/q},$ where $c=c(\tilde{\omega})$ is a positive constant. Then, combining the previous equality with the (2.25), we concluded the proof of the theorem. ∎ Let us remember the following space (see Remark 2.10) $\mathcal{H}:=\Big{\\{}w\in H^{1}_{\text{loc}}(\mathbb{R}^{n};L^{2}(\Omega));\,\text{$w$ is a stationary function}\Big{\\}},$ which is a Hilbert space with respect to the following inner product $\displaystyle\langle w,v{\rangle}_{\mathcal{H}}:=\int_{[0,1)^{n}\times\Omega}$ $\displaystyle\nabla_{\\!y}w(y,\omega)\cdot\nabla_{y}v(y,\omega)\,d{\mathbb{P}}(\omega)\,dy$ $\displaystyle+\int_{[0,1)^{n}\times\Omega}.\\!\\!\\!\\!w(y,\omega)v(y,\omega)\,d{\mathbb{P}}(\omega)\,dy.$ The next lemma will be important in the homogenization’s process. ###### Lemma 2.22. Let $O\subset\mathbb{R}^{n}$ be an open set and assume that $\\{u_{\varepsilon}(\cdot,\tilde{\omega}){\\}}_{\varepsilon>0}$ and $\\{\varepsilon\nabla u_{\varepsilon}(\cdot,\tilde{\omega}){\\}}_{\varepsilon>0}$ are bounded sequences in $L^{2}(O)$ and in $L^{2}(O;\mathbb{R}^{n})$ respectively for a.e. $\tilde{\omega}\in\Omega$. Then, for a.e. $\tilde{\omega}\in\Omega$, there exists a subsequence $\\{\varepsilon^{\prime}\\}$(it may depend on $\tilde{\omega}$) and $u_{\tilde{\omega}}\in L^{2}(O;\mathcal{H})$, such that $u_{\varepsilon^{\prime}}(\cdot,\tilde{\omega})\;\xrightharpoonup[\varepsilon\to 0]{2-{\rm s}}\;u_{\tilde{\omega}},$ and $\varepsilon^{\prime}\nabla u_{\varepsilon^{\prime}}(\cdot,\tilde{\omega})\;\xrightharpoonup[\varepsilon\to 0]{2-{\rm s}}\;[\nabla_{y}\Phi]^{-1}\nabla_{y}u_{\tilde{\omega}}.$ ###### Proof. Applying the Theorem 2.21 for the sequences ${\\{u_{\varepsilon}(\cdot,\widetilde{\omega})\\}_{\varepsilon>0}},\quad{\\{\varepsilon\nabla u_{\varepsilon}(\cdot,\widetilde{\omega})\\}_{\varepsilon>0}}$ for a.e. ${\widetilde{\omega}\in\Omega}$, we can find a subsequence $\\{\varepsilon^{\prime}\\}$, and functions ${u_{\widetilde{\omega}}\in L^{2}({O}\\!\times\\![0,1)^{n}\\!\times\\!\Omega)},\quad{V_{\widetilde{\omega}}\in L^{2}({O}\\!\times[0,1)^{n}\\!\times\\!\Omega;\mathbb{R}^{n})}$ with ${V_{\widetilde{\omega}}=(v^{(1)}_{\widetilde{\omega}},\ldots,v^{(n)}_{\widetilde{\omega}})}$ satisfying for ${k\in\\{1,2,\ldots,n\\}}$, $u_{\varepsilon^{\prime}}(\cdot,\widetilde{\omega})\;\xrightharpoonup[\varepsilon^{\prime}\to 0]{2-{\rm s}}\;u_{\widetilde{\omega}},$ (2.26) and $\varepsilon^{\prime}\frac{\partial u_{\varepsilon^{\prime}}}{\partial x_{k}}\;\xrightharpoonup[\varepsilon^{\prime}\to 0]{2-{\rm s}}\;v^{(k)}_{\widetilde{\omega}}.$ (2.27) Hence for each ${k\in\\{1,\ldots,n\\}}$ and performing an integration by parts we have $\int_{{O}}\varepsilon^{\prime}\frac{\partial u_{\varepsilon^{\prime}}}{\partial x_{k}}(x,\widetilde{\omega})\,{\varphi(x)\,\Theta{\left(\Phi^{-1}\left(\frac{x}{\varepsilon^{\prime}},\widetilde{\omega}\right),\widetilde{\omega}\right)}}dx\\\ =-\varepsilon^{\prime}\int_{{O}}u_{\varepsilon^{\prime}}(x,\widetilde{\omega})\,{\frac{\partial\varphi}{\partial x_{k}}(x)\,\Theta{\left(\Phi^{-1}\left(\frac{x}{\varepsilon^{\prime}},\widetilde{\omega}\right),\widetilde{\omega}\right)}}dx\\\ \quad\quad-\int_{{O}}u_{\varepsilon^{\prime}}(x,\widetilde{\omega})\,{\varphi(x)\,{[\nabla_{\\!\\!y}\Phi]}^{-1}{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon^{\prime}},\widetilde{\omega}\right)},\widetilde{\omega}\right)}\,\nabla_{\\!\\!y}\Theta{\left(\Phi^{-1}\left(\frac{x}{\varepsilon^{\prime}},\widetilde{\omega}\right),\widetilde{\omega}\right)}\cdotp e_{k}}\,dx,$ for every $\varphi\in C^{\infty}_{c}(O)$ and $\Theta\in C_{c}^{\infty}\big{(}[0,1)^{n};L^{\infty}(\Omega)\big{)}$ extended in a stationary way to $\mathbb{R}^{n}$. Then, using the relations (2.26)-(2.27) and a density argument in the space of the test functions, we arrive after letting $\varepsilon^{\prime}\to 0$ $\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}v^{(k)}_{\widetilde{\omega}}{\left(x,\Phi^{-1}\left(z,\omega\right),\omega\right)}\,{\Theta{\left(\Phi^{-1}(z,\omega),\omega\right)}}\,d\mathbb{P}\,dz\\\ =-\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}u_{\widetilde{\omega}}\left(x,\Phi^{-1}\left(z,\omega\right),\omega\right){\frac{\partial}{\partial z_{k}}{\left(\Theta{\left(\Phi^{-1}(z,\omega),\omega\right)}\right)}}\,d\mathbb{P}\,dz,$ for a.e. $x\in{O}$ and for any $\Theta\in C_{c}^{\infty}\big{(}[0,1)^{n};L^{\infty}(\Omega)\big{)}$. Hence­ applying Theorem 2.16, we obtain $\int_{\mathbb{R}^{n}}v^{(k)}_{\widetilde{\omega}}{\left(x,\Phi^{-1}\left(z,\omega\right),\omega\right)}\,{\varphi(z)}\,dz\,=\,-\int_{\mathbb{R}^{n}}u_{\widetilde{\omega}}{\left(x,\Phi^{-1}\left(z,\omega\right),\omega\right)}\,{\frac{\partial\varphi}{\partial z_{k}}(z)}\,dz,$ for all ${\varphi\in C_{\rm c}^{\infty}(\mathbb{R}^{n})}$ and a.e. ${\omega\in\Omega}$. This completes the proof of our lemma. ∎ ### 2.4 Perturbations of bounded operators The aim of this section is to study the point spectrum, that is the set of eigenvalues, for perturbations of a given bounded operator. More precisely, given a complex Hilbert space $H$, and a sequence of operators $\\{A_{\alpha}\\}$, with $A_{\alpha}\in\mathcal{B}(H)$ for each $\alpha\in\mathbb{N}^{n}$, we analyse the point spectrum of the power series of $n-$complex variables $\boldsymbol{z}=(z_{1},\ldots,z_{n})$, which is $\sum_{\alpha\in\mathbb{N}^{n}}\boldsymbol{z}^{\alpha}A_{\alpha}\equiv\sum_{\alpha_{1},\ldots,\alpha_{n}=0}^{\infty}z_{1}^{\alpha_{1}}\ldots z_{n}^{\alpha_{n}}A_{\alpha_{1},\ldots,\alpha_{n}},$ (2.28) from the properties of the spectrum $\sigma(A_{0,\ldots,0})$. This subject was studied for instance by T. Kato [25] and F. Rellich [33]. To follow, we define $|\alpha|:=\alpha_{1}+\ldots+\alpha_{n}$, $(\alpha\in\mathbb{N}^{n})$, $r:=\Big{(}\underset{k\in\mathbb{N}}{\rm inf}\big{\\{}\underset{{\|\alpha\|}_{\infty}>k}{\rm sup}\sqrt[{|\alpha|}]{{\|A_{\alpha}\|}}\big{\\}}\Big{)}^{-1},$ (2.29) and for $R>0$ $\Delta_{R}:=\prod_{\nu=1}^{n}B(0,R).$ Then, we have the following ###### Lemma 2.23. Let ${\left\\{A_{\alpha}\right\\}}$ be a sequence of operators, such that $A_{\alpha}\in\mathcal{B}(H)$ for each $\alpha\in\mathbb{N}^{n}$. Then, the series (2.28) is convergent for each $z\in\Delta_{r}$, with $r>0$ given by (2.29). ###### Proof. Given $\boldsymbol{z}\in\Delta_{r}$, there exists $\varepsilon>0$ such that ${\big{(}\frac{1}{r}+\varepsilon\big{)}}{|z_{\nu}|}<1,\quad\text{for any $\nu\in\\{1,\ldots,n\\}$}.$ (2.30) On the other hand, from (2.29) there exists $k_{0}\in\mathbb{N}$, such that, for each $k\geqslant k_{0}$ $\underset{{\|\alpha\|}_{\infty}>k}{\rm sup}\sqrt[|\alpha|]{{\|A_{\alpha}\|}}<\frac{1}{r}+\varepsilon.$ Then, for $\|\alpha\|_{\infty}>k_{0}$ ${\|A_{\alpha}\|}<{\big{(}\frac{1}{r}+\varepsilon\big{)}}^{{|\alpha|}},$ and hence we have ${|z_{1}|}^{\alpha_{1}}\ldots{|z_{n}|}^{\alpha_{n}}{\|A_{\alpha}\|}<{\Big{(}\big{(}\frac{1}{r}+\varepsilon\big{)}{|z_{1}|}\Big{)}}^{\alpha_{1}}\ldots{\Big{(}\big{(}\frac{1}{r}+\varepsilon\big{)}{|z_{n}|}\Big{)}}^{\alpha_{n}}.$ Therefore, we obtain $\displaystyle\sum_{{\|\alpha\|}_{\infty}>k_{0}}{|\boldsymbol{z}|}^{\alpha}{\|A_{\alpha}\|}$ $\displaystyle\leqslant$ $\displaystyle\sum_{\alpha_{1},\ldots,\alpha_{n}=0}^{\infty}{\left\\{{\left[{\left(\frac{1}{r}+\varepsilon\right)}{|z_{1}|}\right]}^{\alpha_{1}}\ldots{\left[{\left(\frac{1}{r}+\varepsilon\right)}{|z_{n}|}\right]}^{\alpha_{n}}\right\\}}$ $\displaystyle=$ $\displaystyle{\left\\{\sum_{\alpha_{1}=0}^{\infty}{\left[{\left(\frac{1}{r}+\varepsilon\right)}{|z_{1}|}\right]}^{\alpha_{1}}\right\\}}\ldots{\left\\{\sum_{\alpha_{n}=0}^{\infty}{\left[{\left(\frac{1}{r}+\varepsilon\right)}{|z_{n}|}\right]}^{\alpha_{n}}\right\\}},$ and due to (2.30) the power series (2.28) is absolutely convergent for each $\boldsymbol{z}\in\Delta_{r}$. ∎ One remarks that, for each $r_{0}<r$ the series (2.28) converges uniformly in $\Delta_{r_{0}}$. Moreover, it follows from Definition (2.29) that, there exists $c>0$ such that, for any $\alpha\in\mathbb{N}^{n}$, ${\|A_{\alpha}\|}\leqslant{c}^{{|\alpha|}+1}$. Now, let us recall the definition of operator value maps of many complex variables, and after that consider some important results. Let $\mathcal{O}\subset\mathbb{C}^{n}$ be an open set. A map $f:\mathcal{O}\to\mathcal{B}(H)$ is called holomorphic in $\boldsymbol{w}\in\mathcal{O}$, when there exists an open set $U\subset\mathcal{O}$, $\boldsymbol{w}\in U$, such that $f$ is equal to the (absolutely convergent) power series in $\boldsymbol{z}-\boldsymbol{w}$, with coefficients $A_{\alpha}\in\mathcal{B}(H)$, that is $\displaystyle f(\boldsymbol{z})\equiv f(z_{1},\ldots,z_{n})$ $\displaystyle=\sum_{\alpha\in\mathbb{N}^{n}}(\boldsymbol{z}-\boldsymbol{w})^{\alpha}A_{\alpha}$ $\displaystyle\equiv\sum_{\alpha_{1},\ldots,\alpha_{n}=0}^{\infty}(z_{1}-w_{1})^{\alpha_{1}}\ldots(z_{n}-w_{n})^{\alpha_{n}}A_{\alpha_{1},\ldots,\alpha_{n}}$ for each $\boldsymbol{z}\in U$. Moreover, the function $f$ is called holomorphic in $\mathcal{O}$, if it is holomorphic for any $\boldsymbol{w}\in\mathcal{O}$. Moreover, assume that $A\in\mathcal{B}(H)$ is a symmetric operator and $\lambda\in\mathbb{R}$ is an eigenvalue of $A$ with finite multiplicity $h$. Therefore, the operator $A-\lambda I$ is not invertible and there exists a symmetric operator $R\in\mathcal{B}(H)$, uniquely defined, such that $\displaystyle R(A-\lambda I)f$ $\displaystyle=f-\sum_{k=1}^{h}{\langle f,\psi_{k}\rangle}\psi_{k},\quad\text{for each $f\in H$, and}$ (2.31) $\displaystyle R\psi_{k}$ $\displaystyle=0,\quad\text{for all $k\in\\{1,\ldots,h\\}$},$ where $\\{\psi_{1},\ldots,\psi_{h}\\}$ is an orthonormal basis of ${\rm Ker}(A-\lambda I)$. The operator $R$ is called a pseudo-inverse of $A-\lambda I$, and one observes that, $AR=RA$. It is also important to consider the following results on complex value functions. ###### Lemma 2.24 (Osgood’s Lemma). Let $\mathcal{O}\subset\mathbb{C}^{n}$ be an open set, and $f:\mathcal{O}\to\mathbb{C}$ a continuous function that is holomorphic in each variable separately. Then, the function $f$ is holomorphic. Then, in order to state the Weierstrass’ Preparation Theorem, let us recall the concept of Weierstrass’ polinomial. A complex function $W(\varrho,\boldsymbol{z})$, which is holomorphic in a neighborhood of $(0,\boldsymbol{0})\in\mathbb{C}\\!\times\\!\mathbb{C}^{n}$, is called a Weirstrass polynomial of degree $m$, when $W(\varrho,\boldsymbol{z})=\varrho^{m}+a_{1}(\boldsymbol{z})\varrho^{m-1}+\ldots+a_{m-1}(\boldsymbol{z})\varrho+a_{m}(\boldsymbol{z}),$ where any $a_{i}(\boldsymbol{z})$, $(i=1,\ldots,m)$, is an holomorphic function in a neighborhood $\boldsymbol{0}\in\mathbb{C}^{n}$ that vanishes at $\boldsymbol{z}=\boldsymbol{0}$. Then, we have the following ###### Theorem 2.25 (Weierstrass Preparation Theorem). Let $m$ be a positive integer, and $F(\varrho,\boldsymbol{z})$ holomorphic in a neighborhood of $(0,\boldsymbol{0})\in\mathbb{C}\\!\times\\!\mathbb{C}^{n}$ such that, the mapping $\varrho\mapsto F(\varrho,\boldsymbol{0})/\varrho^{m}$ is holomorphic in a neighborhood of $0\in\mathbb{C}$ and is non-zero at $0$. Then, there exist a Weierstrass polynomial $W(\varrho,\boldsymbol{z})$ of degree m, and a holomorphic function $E(\varrho,\boldsymbol{z})$ which does not vanish in a neighborhood $U$ of $(0,\boldsymbol{0})$, such that, for all $(\varrho,\boldsymbol{z})\in U$ $F(\varrho,\boldsymbol{z})=W(\varrho,\boldsymbol{z})E(\varrho,\boldsymbol{z}).$ ###### Proof. See S. G. Krantz, H. R. Parks [26, p. 96]. ∎ At this point, we are in condition to establish the main result of this section, that is to say, the perturbation theory for bounded operators with isolated eigenvalues of finite multiplicity. The theorem considered here is a convenient and direct version for our purposes in this paper. ###### Theorem 2.26. Let $H$ be a Hilbert space, and a sequence of operators $\\{A_{\alpha}\\}$, $A_{\alpha}\in\mathcal{B}(H)$ for each $\alpha\in\mathbb{N}^{n}$. Consider the power series of $n-$complex variables $\boldsymbol{z}=(z_{1},\ldots,z_{n})$ with coefficients $A_{\alpha}$, which is absolutely convergent in a neighborhood $\mathcal{O}$ of $\boldsymbol{z}=\boldsymbol{0}$. Define, the holomorphic map $A:\mathcal{O}\to\mathcal{B}(H)$, $A(\boldsymbol{z}):=\sum_{\alpha\in\mathbb{N}^{n}}\boldsymbol{z}^{\alpha}A_{\alpha}$ and assume that, it is symmetric. If $\lambda$ is an eigenvalue of $A_{0}\equiv A(\boldsymbol{0})$ with finite multiplicity $h$ (and respective eigenvectors $\psi_{i}$, $i=1,\ldots,h$), then there exist a neighborhood $U\subset\mathcal{O}$ of ${\boldsymbol{0}}$, and holomorphic functions $\displaystyle\boldsymbol{z}\in U$ $\displaystyle\,\mapsto\,\lambda_{1}(\boldsymbol{z}),\lambda_{2}(\boldsymbol{z}),\ldots,\lambda_{h}(\boldsymbol{z})\in\mathbb{R},$ $\displaystyle\boldsymbol{z}\in U$ $\displaystyle\,\mapsto\,\psi_{1}(\boldsymbol{z}),\psi_{2}(\boldsymbol{z}),\ldots,\psi_{h}(\boldsymbol{z})\in H\setminus\\{0\\},$ satisfying for each $\boldsymbol{z}\in U$ and $i\in\\{1,\ldots,h\\}:$ * $(i)$ $A(\boldsymbol{z})\psi_{i}(\boldsymbol{z})=\lambda_{i}(\boldsymbol{z})\psi_{i}(\boldsymbol{z})$, * $(ii)$ ${\lambda_{i}(\boldsymbol{z}=\boldsymbol{0})=\lambda}$, * $(iii)$ ${{\rm dim}{\\{w\in H\;;\;A(\boldsymbol{z})w=\lambda_{i}(\boldsymbol{z})w\\}}\leqslant h}$. Moreover, if there exists $d>0$ such that $\sigma(A_{0})\cap(\lambda-d,\lambda+d)={\left\\{\lambda\right\\}},$ then for each $d^{\prime}\in(0,d)$ there exists a neighborhood $W\subset U$ of $\boldsymbol{0}$, such that $\sigma(A(\boldsymbol{z}))\cap(\lambda-d^{\prime},\lambda+d^{\prime})={\left\\{\lambda_{1}(\boldsymbol{z}),\ldots,\lambda_{h}(\boldsymbol{z})\right\\}}$ (2.32) for all $\boldsymbol{z}\in W$. ###### Proof. 1\. First, we conveniently define $B(\boldsymbol{z}):=A(\boldsymbol{z})-A_{0}=\sum_{{|\alpha|}\not=0}\boldsymbol{z}^{\alpha}A_{\alpha}.$ (2.33) Then, there exists a neighborhood of $(\varrho,\boldsymbol{z})=(0,\boldsymbol{0})$ such that, the function $(\varrho,\boldsymbol{z})\,\mapsto\,\sum_{l=0}^{\infty}{\left[R{\left(\varrho-B(\boldsymbol{z})\right)}\right]}^{l}\in\mathcal{B}(H)$ is well defined (see equation (2.31)), and holomorphic on it. Indeed, first we recall that there exists $c>0$ such that, for any $\alpha\in\mathbb{N}^{n}$, ${\|A_{\alpha}\|}\leqslant{c}^{{|\alpha|}+1}$. Then, it follows from (2.33) that $\displaystyle{\|B(\boldsymbol{z})\|}$ $\displaystyle\leqslant\sum_{{|\alpha|}\not=0}{|z_{1}|}^{\alpha_{1}}\ldots{|z_{n}|}^{\alpha_{n}}{\|A_{\alpha}\|}\leqslant\sum_{{|\alpha|}\not=0}{|\boldsymbol{z}|}^{{|\alpha|}}c^{{|\alpha|}+1}$ $\displaystyle=\sum_{k=1}^{\infty}{\sum_{{|\alpha|}=k}{|\boldsymbol{z}|}^{{|\alpha|}}c^{{|\alpha|}+1}}=\sum_{k=1}^{\infty}{\sum_{{|\alpha|}=k}{|\boldsymbol{z}|}^{k}c^{k+1}}$ $\displaystyle=\sum_{k=1}^{\infty}{\left(\\#{\left\\{\alpha\in\mathbb{N}^{n}\;;\;{|\alpha|}=k\right\\}}{|\boldsymbol{z}|}^{k}c^{k+1}\right)}$ $\displaystyle\leqslant\sum_{k=1}^{\infty}(k+1)^{n}{|\boldsymbol{z}|}^{k}c^{k+1}={|\boldsymbol{z}|}c^{2}\sum_{k=1}^{\infty}(k+1)^{n}{|\boldsymbol{z}|}^{k-1}c^{k-1}$ $\displaystyle\leqslant{|\boldsymbol{z}|}c^{2}\sum_{k=0}^{\infty}(k+2)^{n}{|\boldsymbol{z}|}^{k}c^{k}.$ Therefore, it follows that $\sum_{k=0}^{\infty}(k+2)^{n}{|\boldsymbol{z}|}^{k}c^{k}$ is absolutely convergent for each $\boldsymbol{z}\in B{\left(\boldsymbol{0},\frac{1}{4^{n}c}\right)}$. Moreover, there exists $\tilde{c}>0$, such that $\big{|}\sum_{k=0}^{\infty}(k+2)^{n}{|\boldsymbol{z}|}^{k}c^{k}\big{|}\leqslant\tilde{c}\quad\text{for each $\boldsymbol{z}\in B(\boldsymbol{0},\frac{1}{4^{n+1}c})$}.$ Hence we have from (2.33) that $\displaystyle{\left\|R(\varrho-B(\boldsymbol{z}))\right\|}$ $\displaystyle\leq{\|R\|}({|\varrho|}+{|\boldsymbol{z}|}c^{2}\tilde{c})$ (2.34) $\displaystyle\leq{\|R\|}\ {\rm max}{\left\\{1,c^{2}\tilde{c}\right\\}}({|\varrho|}+{|\boldsymbol{z}|}),$ for ${\varrho\in\mathbb{C}}$ and ${\boldsymbol{z}\in B{\left(\boldsymbol{0},\frac{1}{4^{n+1}c}\right)}}$. To follow, we define $r:=\min\Big{\\{}\frac{1}{8{\|R\|}{\rm max}{\left\\{1,c^{2}\tilde{c}\right\\}}},\frac{1}{4^{n+1}c}\Big{\\}},\quad\Delta_{r}:=B(0,r)\\!\times\\!B(\boldsymbol{0},r)\subset\mathbb{C}\\!\times\\!\mathbb{C}^{n}.$ (2.35) Then, for any $m,n\in\mathbb{N}$ with $m>n$, and all $(\varrho,\boldsymbol{z})\in\Delta_{r}$, we have $\displaystyle{\|\sum_{l=0}^{m}{\left[R(\varrho-B(\boldsymbol{z}))\right]}^{l}-\sum_{l=0}^{n}{\left[R(\varrho-B(\boldsymbol{z}))\right]}^{l}\|}$ $\displaystyle\leq\sum_{l=n+1}^{m}{\|R(\varrho-B(\boldsymbol{z}))\|}^{l}$ $\displaystyle\leq\sum_{l=n+1}^{m}{\left(\frac{1}{4}\right)}^{l}.$ Consequently, for any $(\varrho,\boldsymbol{z})\in\Delta_{r}$, $\\{\sum_{l=0}^{m}{\left[R(\varrho-B(\boldsymbol{z}))\right]}^{l}\\}_{m\in\mathbb{N}}$ is a Cauchy sequence in $\mathcal{B}(H)$. Therefore, the mapping $(\varrho,\boldsymbol{z})\in\Delta_{r}\;\mapsto\;\sum_{l=0}^{\infty}{\left[R(\varrho-B(\boldsymbol{z}))\right]}^{l}$ (2.36) is holomorphic, since it is the uniform limit of holomorphic functions. 2\. Now, for $i,j=1,\ldots,h$ and $(\varrho,\boldsymbol{z})\in\Delta_{r}$, let us consider $f_{ij}(\varrho,\boldsymbol{z})=\Big{\langle}\sum_{l=0}^{\infty}(\varrho-B(\boldsymbol{z})){\left[R(\varrho-B(\boldsymbol{z}))\right]}^{l}\psi_{i},\psi_{j}\Big{\rangle}.$ Therefore, the function $F:\Delta_{r}\rightarrow\mathbb{C}$, defined by $F(\varrho,\boldsymbol{z}):={\rm det}{\left[{\left(f_{ij}(\varrho,\boldsymbol{z})\right)}\right]}$ is holomorphic. In fact, $F(\varrho,\boldsymbol{z})$ is a real value function, when $\varrho\in\mathbb{R}$. Moreover, $F(\varrho,\boldsymbol{0})={\rm det}{\left[\varrho\,(\delta_{ij})\right]}=\varrho^{h}$ for each $\varrho\in B(0,r)$, where $\delta_{ij}$ is the Kronecker delta. Indeed, we have $\displaystyle f_{ij}(\varrho,\boldsymbol{0})$ $\displaystyle=$ $\displaystyle\Big{\langle}\sum_{l=0}^{\infty}(\varrho-B(\boldsymbol{0})){\left[R(\varrho-B(\boldsymbol{0}))\right]}^{l}\psi_{i},\psi_{j}\Big{\rangle}$ $\displaystyle=$ $\displaystyle\left\langle\sum_{l=0}^{\infty}\varrho^{l+1}R^{l}\psi_{i},\psi_{j}\right\rangle$ $\displaystyle=$ $\displaystyle\left\langle\varrho\,\psi_{i},\psi_{j}\right\rangle+\sum_{l=1}^{\infty}\left\langle\varrho^{l+1}R^{l}\psi_{i},\psi_{j}\right\rangle=\varrho\,\delta_{ij},$ from which follows the result. 3\. At this point, we show that there exist $h$ holomorphic functions $\varrho_{k}(\boldsymbol{z})$, $(k=1,\ldots,h)$, defined in a neighborhood of $\boldsymbol{z}=\boldsymbol{0}$, such that $\lim_{\boldsymbol{z}\to\boldsymbol{0}}\varrho_{k}(\boldsymbol{z})=0,\quad\text{for $k\in\\{1,\ldots,h\\}$}.$ Indeed, applying Theorem 2.25 (Weierstrass Preparation Theorem) there exists a Weirstrass polynomial of degree $h$ $W(\varrho,\boldsymbol{z})=\varrho^{h}+a_{1}(\boldsymbol{z})\varrho^{h-1}+\ldots+a_{h-1}(\boldsymbol{z})\varrho+a_{h}(\boldsymbol{z}),$ and also a holomorphic function $E(\varrho,\boldsymbol{z})$, which does not vanish in a neighborhood $U\times V$ of $(0,\boldsymbol{0})$, with $U\subset B(0,r)\subset\mathbb{C}$ and $V\subset B(\boldsymbol{0},r)\subset\mathbb{C}^{n}$, such that for each $(\varrho,\boldsymbol{z})\in U\\!\times\\!V$ $F(\varrho,\boldsymbol{z})=E(\varrho,\boldsymbol{z}){\left(\varrho^{h}+a_{1}(\boldsymbol{z})\varrho^{h-1}+\ldots+a_{h-1}(\boldsymbol{z})\varrho+a_{h}(\boldsymbol{z})\right)}.$ Since the coefficients of the Weirstrass polynomial are holomorphic functions, which vanish in $\boldsymbol{z}=\boldsymbol{0}$, then there exist holomorphic functions $\varrho_{k}(\boldsymbol{z})$, such that $\displaystyle F(\varrho,\boldsymbol{z})$ $\displaystyle=E(\varrho,\boldsymbol{z})\,\prod_{k=1}^{h}{\left(\varrho-\varrho_{k}(\boldsymbol{z})\right)},$ (2.37) $\displaystyle\lim_{\boldsymbol{z}\to\boldsymbol{0}}\varrho_{k}(\boldsymbol{z})$ $\displaystyle=0,\quad(k=1,\ldots,h).$ 4\. At this point, let us show that, for $k\in\\{1,\ldots,h\\}$ there exists a map $\psi_{k}(\boldsymbol{z})\in H-\\{0\\}$ such that, $A(\boldsymbol{z})\psi_{k}(\boldsymbol{z})=(\lambda+\varrho_{k}(\boldsymbol{z}))\psi_{k}(\boldsymbol{z})$, for each $\boldsymbol{z}$ in a neighborhood of $\boldsymbol{0}$. Indeed, let $k\in\\{1,\ldots,h\\}$ be fix. From item 3, there exists a set $V\subset\mathbb{C}$, which is a neighborhood of $\boldsymbol{z}=\boldsymbol{0}$, such that ${\rm det}{\left[{\left(f_{ij}(\varrho_{k}(\boldsymbol{z}),\boldsymbol{z})\right)}\right]}=0$ for each $\boldsymbol{z}\in V$. Therefore, for each $\boldsymbol{z}\in V$ the linear system $\left(f_{ji}(\varrho_{k}(\boldsymbol{z}),\boldsymbol{z})\right)(c_{1},\ldots,c_{n})^{T}=0$ has a non-trivial solution. Consequently, there exist $h$ holomorphic functions $\boldsymbol{z}\in V\mapsto c_{1}^{k}(\boldsymbol{z}),\ldots,c_{h}^{k}(\boldsymbol{z})$, such that for all $j=1,\ldots,h$ $\sum_{i=1}^{h}f_{ij}(\varrho_{k}(\boldsymbol{z}),\boldsymbol{z})\,c_{i}^{k}(\boldsymbol{z})=0,$ and without loss of generality we may assume $\sum_{i=1}^{h}{|c_{i}^{k}(\boldsymbol{z})|}^{2}=1.$ (2.38) From equation (2.37) it is possible to find a neighborhood $\tilde{V}$ of $\boldsymbol{0}$, which is compactly embedded in $V$, such that $\underset{\boldsymbol{z}\in\tilde{V}}{\rm sup}{|\varrho_{k}(\boldsymbol{z})|}<\frac{1}{8{\|R\|}{\rm max}(1,c^{2}\tilde{c})},$ for each $\boldsymbol{z}\in\tilde{V}$ and all $k\in\\{1,\ldots,n\\}$. Hence we obtain for each $\boldsymbol{z}\in\tilde{V}$ ${\|R(\varrho_{k}(\boldsymbol{z})-B(\boldsymbol{z}))\|}\leq{\|R\|}{\rm max}(1,c^{2}\tilde{c}){\left({|\varrho_{k}(\boldsymbol{z})|}+{|\boldsymbol{z}|}\right)}\leq\frac{1}{4},$ (2.39) and then $\sum_{l=0}^{\infty}{\|R(\varrho_{k}(\boldsymbol{z})-B(\boldsymbol{z}))\|}^{l}\leqslant\frac{4}{3}.$ (2.40) Now, we define for any $\boldsymbol{z}\in\tilde{V}$ $\phi_{k}(\boldsymbol{z}):=\sum_{i=1}^{h}c_{i}^{k}(\boldsymbol{z})\psi_{i},\quad\text{and}\quad\psi_{k}(\boldsymbol{z}):=\sum_{l=0}^{\infty}{\left[R(\varrho_{k}(\boldsymbol{z})-B(\boldsymbol{z}))\right]}^{l}\phi_{k}(\boldsymbol{z}).$ Therefore, we have $\displaystyle\psi_{k}(\boldsymbol{z})$ $\displaystyle=$ $\displaystyle\phi_{k}(\boldsymbol{z})+\sum_{l=1}^{\infty}{\left[R(\mu_{k}(\boldsymbol{z})-B(\boldsymbol{z}))\right]}^{l}\phi_{k}(\boldsymbol{z})$ $\displaystyle=$ $\displaystyle\phi_{k}(\boldsymbol{z})+{\left[R(\varrho_{k}(\boldsymbol{z})-B(\boldsymbol{z}))\right]}\sum_{l=1}^{\infty}{\left[R(\varrho_{k}(\boldsymbol{z})-B(\boldsymbol{z}))\right]}^{l-1}\phi_{k}(\boldsymbol{z})$ $\displaystyle=$ $\displaystyle\phi_{k}(\boldsymbol{z})+{\left[R(\varrho_{k}(\boldsymbol{z})-B(\boldsymbol{z}))\right]}\psi_{k}(\boldsymbol{z}),$ and it follows that $\displaystyle(A_{0}-\lambda)\psi_{k}(\boldsymbol{z})$ $\displaystyle=(A_{0}-\lambda)\phi_{k}(\boldsymbol{z})+(A_{0}-\lambda){\left[R(\varrho_{k}(\boldsymbol{z})-B(\boldsymbol{z}))\right]}\psi_{k}(\boldsymbol{z})$ (2.41) $\displaystyle=\sum_{i=1}^{h}c_{i}^{k}(\boldsymbol{z})(A_{0}-\lambda)\psi_{i}+(A_{0}-\lambda)R{\left[(\varrho_{k}(\boldsymbol{z})-B(\boldsymbol{z}))\psi_{k}(\boldsymbol{z})\right]}$ $\displaystyle=R(A_{0}-\lambda){\left[(\varrho_{k}(\boldsymbol{z})-B(\boldsymbol{z}))\psi_{k}(\boldsymbol{z})\right]}$ $\displaystyle=(\varrho_{k}(\boldsymbol{z})-B(\boldsymbol{z}))\psi_{k}(\boldsymbol{z})-\sum_{j=1}^{h}\left\langle(\varrho_{k}(\boldsymbol{z})-B(\boldsymbol{z}))\psi_{k}(\boldsymbol{z}),\psi_{j}\right\rangle\psi_{j}$ $\displaystyle=(\varrho_{k}(\boldsymbol{z})-B(\boldsymbol{z}))\psi_{k}(\boldsymbol{z})$ since $\displaystyle\left\langle(\varrho_{k}(\boldsymbol{z})\right.$ $\displaystyle-\left.B(\boldsymbol{z}))\psi_{k}(\boldsymbol{z}),\psi_{j}\right\rangle$ $\displaystyle=\sum_{i=1}^{h}\Big{\langle}\sum_{l=0}^{\infty}(\varrho_{k}(\boldsymbol{z})-B(\boldsymbol{z})){\left[R(\varrho_{k}(\boldsymbol{z})-B(\boldsymbol{z}))\right]}^{l}\psi_{i},\psi_{j}\Big{\rangle}c_{i}^{k}(\boldsymbol{z})$ $\displaystyle=\sum_{i=1}^{h}f_{ij}(\varrho_{k}(\boldsymbol{z}),\boldsymbol{z})\,c_{i}^{k}(\boldsymbol{z})=0.$ Thus, for each $\boldsymbol{z}\in\tilde{V}$, $A(\boldsymbol{z})\psi_{k}(\boldsymbol{z})=(\lambda+\varrho_{k}(\boldsymbol{z}))\psi_{k}(\boldsymbol{z})$. On the other hand, $\psi_{k}(\boldsymbol{z})=\phi_{k}(\boldsymbol{z})+{\left[R(\varrho_{k}(\boldsymbol{z})-B(\boldsymbol{z}))\right]}\sum_{l=1}^{\infty}{\left[R(\varrho_{k}(\boldsymbol{z})-B(\boldsymbol{z}))\right]}^{l-1}\phi_{k}(\boldsymbol{z}),$ hence from (2.38), (2.39), and (2.40), we have for each $\boldsymbol{z}\in\tilde{V}$ $\displaystyle{\|\psi_{k}(\boldsymbol{z})-\phi_{k}(\boldsymbol{z})\|}$ $\displaystyle\leq$ $\displaystyle{\|R(\varrho_{k}(\boldsymbol{z})-B(\boldsymbol{z}))\|}{\|\sum_{l=0}^{\infty}{\left[R(\varrho_{k}(\boldsymbol{z})-B(\boldsymbol{z}))\right]}^{l}\|}{\|\phi_{k}(\boldsymbol{z})\|}$ $\displaystyle\leq$ $\displaystyle\frac{4}{3}{\|R(\varrho_{k}(\boldsymbol{z})-B(\boldsymbol{z}))\|}\leq\frac{1}{3}.$ Consequently, for each $\boldsymbol{z}\in\tilde{V}$ we have $\psi_{k}(\boldsymbol{z})\not=0$, since $1={\|\phi_{k}(\boldsymbol{z})\|}\leq{\|\phi_{k}(\boldsymbol{z})-\psi_{k}(\boldsymbol{z})\|}+{\|\psi_{k}(\boldsymbol{z})\|}\leq\frac{1}{3}+{\|\psi_{k}(\boldsymbol{z})\|}.$ 5\. Now, let us show item $(iii)$ of the thesis, that is, ${{\rm dim}{\\{w\in H\;;\;A(\boldsymbol{z})w=\lambda_{i}(\boldsymbol{z})w\\}}\leqslant h}.$ From the previous item, there exists $\lambda_{k}(\boldsymbol{z})=\lambda+\varrho_{k}(\boldsymbol{z})$, $k\in\\{1,\ldots,h\\}$, an eigenvalue of the operator $A(\boldsymbol{z})$, for $\boldsymbol{z}$ in a neighborhood of $\boldsymbol{z}=\boldsymbol{0}$. We set $\lambda(\boldsymbol{z})=\lambda_{k}(\boldsymbol{z})$, for any $k\in\\{1,\ldots,h\\}$ fixed, and let $\psi(\boldsymbol{z})$ be any function satisfying $A(\boldsymbol{z})\psi(\boldsymbol{z})=\lambda(\boldsymbol{z})\psi(\boldsymbol{z}),$ which is not necessarily the eigenfunction $\psi_{k}(\boldsymbol{z})$. Then, we are going to show that, there exist a neighborhood of $\boldsymbol{z}=\boldsymbol{0}$, and for each $\boldsymbol{z}$ in this neighborhood an invertible holomorphic operator $S(\boldsymbol{z})\in\mathcal{B}(H)$, such that, $\psi(\boldsymbol{z})\in{\rm span}{\big{\\{}S(\boldsymbol{z})\psi_{1},S(\boldsymbol{z})\psi_{2},\ldots,S(\boldsymbol{z})\psi_{h}\big{\\}}}.$ (2.42) Indeed, to show (2.42) let us define $\varrho(\boldsymbol{z}):=\lambda(\boldsymbol{z})-\lambda$, then we have ${\big{(}\varrho(\boldsymbol{z})I-B(\boldsymbol{z})\big{)}}\psi(\boldsymbol{z})={\big{(}A_{0}-\lambda\big{)}}\psi(\boldsymbol{z}).$ Hence from the first equation in (2.31), it follows that $R{\big{(}\varrho(\boldsymbol{z})I-B(\boldsymbol{z})\big{)}}\psi(\boldsymbol{z})=\psi(\boldsymbol{z})-\sum_{i=1}^{h}{\langle\psi(\boldsymbol{z}),\psi_{i}\rangle}\psi_{i},$ or equivalently ${\big{[}I-R{\big{(}\varrho(\boldsymbol{z})I-B(\boldsymbol{z})\big{)}}\big{]}}\psi(\boldsymbol{z})=\sum_{i=1}^{h}{\langle\psi(\boldsymbol{z}),\psi_{i}\rangle}\psi_{i}.$ On the other hand, from (2.34) it is possible to find a neighborhood $V$ of $\boldsymbol{z}=\boldsymbol{0}$, such that, $\|R\big{(}\varrho(\boldsymbol{z})I-B(\boldsymbol{z})\big{)}\|<1$ Therefore, it exists an invertible operator $S(\boldsymbol{z})={\big{[}I-R{\big{(}\varrho(\boldsymbol{z})I-B(\boldsymbol{z})\big{)}}\big{]}}^{-1}=\sum_{\nu=0}^{\infty}{\big{[}R{\big{(}\varrho(\boldsymbol{z})I-B(\boldsymbol{z})\big{)}}\big{]}}^{\nu},$ and hence $\psi(\boldsymbol{z})=S(\boldsymbol{z}){\left(\sum_{i=1}^{h}{\langle\psi(\boldsymbol{z}),\psi_{i}\rangle}\psi_{i}\right)}=\sum_{i=1}^{h}{\langle\psi(\boldsymbol{z}),\psi_{i}\rangle}{\big{[}S(\boldsymbol{z})\psi_{i}\big{]}}.$ 6\. Finally, we show that the perturbed eigenvalues are isolated. To this end, we consider $N(\boldsymbol{z}):={\rm span}{\left\\{\psi_{1}(\boldsymbol{z}),\psi_{2}(\boldsymbol{z}),\ldots,\psi_{h}(\boldsymbol{z})\right\\}},$ the operator $P(\boldsymbol{z}):H\to H$, which is a projection on $N(\boldsymbol{z})$, given by $P(\boldsymbol{z})u=\sum_{i=1}^{h}\left\langle u,\psi_{i}(\boldsymbol{z})\right\rangle\psi_{i}(\boldsymbol{z}),$ and for $d>0$ the operator $D(\boldsymbol{z}):H\to H$, defined by $D(\boldsymbol{z}):=A(\boldsymbol{z})-2dP(\boldsymbol{z}).$ One observes that $D(\boldsymbol{z})u=\sum_{i=1}^{h}(\lambda_{i}(\boldsymbol{z})-2d)\left\langle u_{1},\psi_{i}(\boldsymbol{z})\right\rangle\psi_{i}(\boldsymbol{z})+A(\boldsymbol{z})u_{2},$ (2.43) where we have used the direct sum $u=u_{1}+u_{2}$, $u_{1}\in N(\boldsymbol{z})$ and $u_{2}\in N(\boldsymbol{z})^{\perp}$. Claim 1. * a) For $\xi\in\mathbb{R}\setminus{\left\\{\lambda_{1}(\boldsymbol{z}),\lambda_{2}(\boldsymbol{z}),\ldots,\lambda_{h}(\boldsymbol{z})\right\\}}$, $D(\boldsymbol{z})-\xi\;\,\text{is bijective}\;\Rightarrow\;A(\boldsymbol{z})-\xi\;\,\text{is bijective}.$ * b) For $\xi\in\mathbb{R}\setminus{\left\\{\lambda_{1}(\boldsymbol{z})-2d,\lambda_{2}(\boldsymbol{z})-2d,\ldots,\lambda_{h}(\boldsymbol{z})-2d\right\\}}$, $A(\boldsymbol{z})-\xi\;\,\text{is bijective}\;\Rightarrow\;D(\boldsymbol{z})-\xi\;\,\text{is bijective}.$ Proof of Claim 1. First, we show item (a). Let $\xi\in\mathbb{R}\setminus{\left\\{\lambda_{1}(\boldsymbol{z}),\ldots,\lambda_{h}(\boldsymbol{z})\right\\}}$ be such that, $D(\boldsymbol{z})-\xi$ is bijective. Then, we must show that $A(\boldsymbol{z})-\xi$ is injective and surjective: Injective. Let $u\in{\rm Ker}(A(\boldsymbol{z})-\xi)$ and since $\xi\not=\lambda_{i}(\boldsymbol{z})$, for $i\in\\{1,\ldots,n\\}$, we have $\left\langle u,\psi_{i}(\boldsymbol{z})\right\rangle=0$ for all $i\in\\{1,\ldots,n\\}$. Therefore, $u\in N(\boldsymbol{z})^{\perp}$ and from (2.43), $(D(\boldsymbol{z})-\xi)u=(A(\boldsymbol{z})-\xi)u=0.$ Consequently, we obtain $u=0$. Surjective. Applying the surjection of $D(\boldsymbol{z})-\xi$, for each $v\in H$ there exists $u\in H$, such that $(D(\boldsymbol{z})-\xi)u=v.$ (2.44) On the other hand, we write $u=u_{1}+u_{2}$, with $u_{1}\in N(\boldsymbol{z})$ and $u_{2}\in N(\boldsymbol{z})^{\perp}$, hence from equations (2.43) and (2.44), we obtain $v=\sum_{i=1}^{h}(\lambda_{i}(\boldsymbol{z})-2d-\xi)\left\langle u_{1},\psi_{i}(\boldsymbol{z})\right\rangle\psi_{i}(\boldsymbol{z})+(A(\boldsymbol{z})-\xi)u_{2}.$ (2.45) Moreover, since $\xi\not=\lambda_{i}(\boldsymbol{z})$, for $i\in\\{1,\ldots,n\\}$, it follows that $(A(\boldsymbol{z})-\xi){\left[\frac{\psi_{i}(\boldsymbol{z})}{\lambda_{i}(\boldsymbol{z})-\xi}\right]}=\psi_{i}(\boldsymbol{z}),$ and hence applying it in (2.45), we have $\displaystyle v$ $\displaystyle=$ $\displaystyle\sum_{i=1}^{h}(A(\boldsymbol{z})-\xi){\left[{\left(\frac{\lambda_{i}(\boldsymbol{z})-2d-\xi}{\lambda_{i}(\boldsymbol{z})-\xi}\right)}\left\langle u_{1},\psi_{i}(\boldsymbol{z})\right\rangle\psi_{i}(\boldsymbol{z})\right]}+(A(\boldsymbol{z})-\xi)u_{2}$ $\displaystyle=$ $\displaystyle(A(\boldsymbol{z})-\xi){\left[\sum_{i=1}^{h}{\left(\frac{\lambda_{i}(\boldsymbol{z})-2d-\xi}{\lambda_{i}(\boldsymbol{z})-\xi}\right)}\left\langle u_{1},\psi_{i}(\boldsymbol{z})\right\rangle\psi_{i}(\boldsymbol{z})+u_{2}\right]}.$ Thus, the operator $A(\boldsymbol{z}-\xi)$ is surjective. Now, let us show item (b). Let $\xi\in\mathbb{R}\setminus{\left\\{\lambda_{1}(\boldsymbol{z})-2d,\ldots,\lambda_{h}(\boldsymbol{z})-2d\right\\}}$ be such that, $A(\boldsymbol{z})-\xi$ is bijective. Similarly, we must show that $D(\boldsymbol{z})-\xi$ is injective and surjective: Injective. Let $u\in H$ be such that $(D(\boldsymbol{z})-\xi)u=0$. Then writing $u=u_{1}+u_{2}$, with $u_{1}\in N(\boldsymbol{z})$ and $u_{2}\in N(\boldsymbol{z})^{\perp}$, it follows from equation (2.43) that $0=\sum_{i=1}^{h}(\lambda_{i}(\boldsymbol{z})-2d-\xi)\left\langle u_{1},\psi_{i}(\boldsymbol{z})\right\rangle\psi_{i}(\boldsymbol{z})+(A(\boldsymbol{z})-\xi)u_{2},$ (2.46) thus $(A(\boldsymbol{z})-\xi)u_{2}\in N(\boldsymbol{z})$. Consequently, we have $\displaystyle(A(\boldsymbol{z})-\xi)u_{2}$ $\displaystyle=$ $\displaystyle P(\boldsymbol{z}){\left[(A(\boldsymbol{z})-\xi)u_{2}\right]}=P(\boldsymbol{z})A(\boldsymbol{z})u_{2}-\xi P(\boldsymbol{z})u_{2}$ $\displaystyle=$ $\displaystyle\sum_{i=1}^{h}\left\langle A(\boldsymbol{z})u_{2},\psi_{i}(\boldsymbol{z})\right\rangle\psi_{i}(\boldsymbol{z})-\xi P(\boldsymbol{z})u_{2}$ $\displaystyle=$ $\displaystyle\sum_{i=1}^{h}\left\langle u_{2},A(\boldsymbol{z})\psi_{i}(\boldsymbol{z})\right\rangle\psi_{i}(\boldsymbol{z})-\xi P(\boldsymbol{z})u_{2}$ $\displaystyle=$ $\displaystyle\sum_{i=1}^{h}\lambda_{i}(\boldsymbol{z})\left\langle u_{2},\psi_{i}(\boldsymbol{z})\right\rangle\psi_{i}(\boldsymbol{z})-\xi P(\boldsymbol{z})u_{2}=0$ since $u_{2}\in N(\boldsymbol{z})^{\perp}$. By hypothesis $A(\boldsymbol{z})-\xi$ is injective, thus $u_{2}=0$. Then, from equation (2.46) we obtain $\sum_{i=1}^{h}(\lambda_{i}(\boldsymbol{z})-2d-\xi)\left\langle u_{1},\psi_{i}(\boldsymbol{z})\right\rangle\psi_{i}(\boldsymbol{z})=0,$ and since $\\{\psi_{i}(\boldsymbol{z})\\}_{i=1}^{h}$ is a linearly dependent set of vectors, we have for each $i\in\\{1,\ldots,h\\}$, $\left\langle u_{1},\psi_{i}(\boldsymbol{z})\right\rangle=0$, thus $\left\langle u_{1},\psi_{i}(\boldsymbol{z})\right\rangle=0$. Recall that by hypothesis $\lambda_{i}(\boldsymbol{z})-2d-\xi\not=0$, for all $i\in\\{1,\ldots,h\\}$. Therefore, we obtain $u_{1}=0$. Surjective. Again, applying the surjection of $A(\boldsymbol{z})-\xi$, for each $v\in H$ there exists $u\in H$, such that $(A(\boldsymbol{z})-\xi)u=v.$ (2.47) Then, writing $u=u_{1}+u_{2}$, with $u_{1}\in N(\boldsymbol{z})$ and $u_{2}\in N(\boldsymbol{z})^{\perp}$, from equations (2.43) and (2.47), we have $v=\sum_{i=1}^{h}(\lambda_{i}(\boldsymbol{z})-\xi){\left\langle u_{1},\psi_{i}(\boldsymbol{z})\right\rangle}\psi_{i}(\boldsymbol{z})+(D(\boldsymbol{z})-\xi)u_{2}.$ (2.48) Moreover, since $\xi\not=\lambda_{i}(\boldsymbol{z})-2d$, for $i\in\\{1,\ldots,n\\}$, $(D(\boldsymbol{z})-\xi){\left[\frac{\psi_{i}(\boldsymbol{z})}{\lambda_{i}(\boldsymbol{z})-2d-\xi}\right]}=\psi_{i}(\boldsymbol{z})$ and then from (2.48), it follows that $\displaystyle v$ $\displaystyle=$ $\displaystyle\sum_{i=1}^{h}(D(\boldsymbol{z})-\xi){\left[{\left(\frac{\lambda_{i}(\boldsymbol{z})-\xi}{\lambda_{i}(\boldsymbol{z})-2d-\xi}\right)}{\left\langle u_{1},\psi_{i}(\boldsymbol{z})\right\rangle}\phi^{i}(\boldsymbol{z})\right]}+(D(\boldsymbol{z})-\xi)u_{2}$ $\displaystyle=$ $\displaystyle(D(\boldsymbol{z})-\xi){\left[\sum_{i=1}^{h}{\left(\frac{\lambda_{i}(\boldsymbol{z})-\xi}{\lambda_{i}(\boldsymbol{z})-2d-\xi}\right)}{\left\langle u_{1},\psi_{i}(\boldsymbol{z})\right\rangle}\psi_{i}(\boldsymbol{z})+u_{2}\right]}.$ Therefore, the operator $D(\boldsymbol{z})-\xi$ is surjective. Claim 2. The spectrum of the operator $D(\boldsymbol{0})$ does not contain elements of the interval $(\lambda-d,\lambda+d)$, i.e. $\sigma(D(\boldsymbol{0}))\cap(\lambda-d,\lambda+d)=\emptyset.$ Proof Claim 2. From item (b) of Claim 1, we have $\sigma(D(\boldsymbol{0}))\subset\sigma(A_{0})$, and thus $\sigma(D(\boldsymbol{0}))\cap(\lambda-d,\lambda+d)\subset\sigma(A_{0})\cap(\lambda-d,\lambda+d)=\\{\lambda\\}.$ Suppose that $\lambda\in\sigma(D(\boldsymbol{0}))\cap(\lambda-d,\lambda+d)$, that is to say, it is an isolated element of the spectrum of $D(\boldsymbol{0})$. Therefore, $\lambda$ is an eigenvalue of $D(\boldsymbol{0})$, but this is not possible since $D(\boldsymbol{0})-\lambda$ is an injective operator (see the proof of item (b) of Claim 1). Consequently, we have $\sigma(D(\boldsymbol{0}))\cap(\lambda-d,\lambda+d)=\emptyset.$ It remains to show (2.32). First, by definition $P(\boldsymbol{z})u$ is holomorphic for each $\boldsymbol{z}$ in a neighborhood of $\boldsymbol{0}$. Therefore, the mapping $\boldsymbol{z}\mapsto P(\boldsymbol{z})$ is holomorphic in this neighbohood. Then, the mapping $\boldsymbol{z}\ \mapsto D(\boldsymbol{z})\in\mathcal{B}(H)$ is continuous. Moreover, since the subset of invertible operators in $\mathcal{B}(H)$ is an open set, there exists a (small) neighborhood $\boldsymbol{0}$, such that the function $\boldsymbol{z}\mapsto(D(\boldsymbol{z})-\lambda)^{-1}\in\mathcal{B}(H)$ is continuous. On the other hand, there exists $d^{\prime}\in(0,d)$ such that ${\left\|(D(\boldsymbol{0})-\lambda)^{-1}\right\|}\leqslant\frac{1}{{\rm dist}(\lambda,\sigma(D(\boldsymbol{0})))}\leqslant\frac{1}{d}<\frac{1}{d^{\prime}},$ see Reed, Simon [32, Chapter VIII]. Therefore, by the continuity of the map $\boldsymbol{z}\mapsto(D(\boldsymbol{z})-\lambda)^{-1}\in\mathcal{B}(H)$, there exists a neighborhood of $\boldsymbol{0}$, namely $W$, such that for all $\boldsymbol{z}\in W$ ${\left\|(D(\boldsymbol{\boldsymbol{z}})-\lambda)^{-1}\right\|}<\frac{1}{d^{\prime}}.$ Thus for any $u\in H$ and $\boldsymbol{z}\in W$, it follows that $\displaystyle{\|u\|}$ $\displaystyle={\left\|(D(\boldsymbol{\boldsymbol{z}})-\lambda)^{-1}{\left[{\left(D(\boldsymbol{\boldsymbol{z}})-\lambda\right)}u\right]}\right\|}$ $\displaystyle\leq{\left\|(D(\boldsymbol{\boldsymbol{z}})-\lambda)^{-1}\right\|}{\left\|(D(\boldsymbol{\boldsymbol{z}})-\lambda)u\right\|}<\frac{1}{d^{\prime}}\,{\left\|(D(\boldsymbol{\boldsymbol{z}})-\lambda)u\right\|}.$ Hence for $d^{\prime\prime}\in(0,d^{\prime})$ and $\xi\in(\lambda-d^{\prime\prime},\lambda+d^{\prime\prime})$, we have $\displaystyle{\left\|(D(\boldsymbol{\boldsymbol{z}})-\xi)u\right\|}$ $\displaystyle\geq$ $\displaystyle{\left\|(D(\boldsymbol{\boldsymbol{z}})-\lambda)u\right\|}-{\left|\lambda-\xi\right|}{\lVert u\rVert}$ $\displaystyle\geq$ $\displaystyle(d^{\prime}-d^{\prime\prime}){\|u\|}.$ Consequently, for all $\xi\in(\lambda-d^{\prime\prime},\lambda+d^{\prime\prime})$, $\xi$ is an element of the resolvent of $D(\boldsymbol{z})$, that is $\xi\in\rho(D(\boldsymbol{z}))$. Thus for each $\boldsymbol{z}\in W$, we have $(\lambda-d^{\prime},\lambda+d^{\prime})\subset\rho(D(\boldsymbol{z})).$ Finally, since for each $\boldsymbol{z}\in W$ $\sigma(A(\boldsymbol{z}))\setminus\\{\lambda_{1}(\boldsymbol{z}),\ldots,\lambda_{h}(\boldsymbol{z})\\}\subset\sigma(D(\boldsymbol{z}))\setminus\\{\lambda_{1}(\boldsymbol{z}),\ldots,\lambda_{h}(\boldsymbol{z})\\},$ we obtain from item (a) of Claim 1, that $\sigma(A(\boldsymbol{z}))\setminus\\{\lambda_{1}(\boldsymbol{z}),\ldots,\lambda_{h}(\boldsymbol{z})\\}\cap(\lambda-d^{\prime},\lambda+d^{\prime})=\emptyset,$ which finish the proof. ∎ ## 3 Bloch Waves Analysis Bloch waves analysis is important in the theory of solid-state physics. More precisely, the displacement of an electron in a crystal (periodic setting) is often described by Bloch waves, and this application is supported by Bloch’s Theorem which states that, the energy eigenstates for an electron in a crystal can be written as Bloch waves. The aim of this section is to extend the Bloch waves theory, which is known just for periodic functions to the considered stochastic setting, that is, stationary functions composed with stochastic deformations, which is used here to describe non-crystalline matter. Therefore, we would like to show that, the electron waves in a non-crystalline matter can have a basis consisting entirely of Bloch wave energy eigenstates (now solution to a stochastic Bloch spectral cell equation). Consequently, we are extending the concept of electronic band structures to non-crystalline matter. ### 3.1 The WKB method Here we formally obtain the Bloch spectral cell equation (see Definition 3.56), applying the asymptotic Wentzel-Kramers-Brillouin (WKB for short) expansion method, that is, we assume that the solution of equation (1.1) is given by a plane wave. More precisely, for each $\varepsilon>0$ let us assume that, the solution $u_{\varepsilon}(t,x,\omega)$ of the equation (1.1) has the following asymptotic expansion $u_{\varepsilon}(t,x,\omega)=e^{2\pi iS_{\varepsilon}(t,x)}\sum_{k=1}^{\infty}\varepsilon^{k}u_{k}\Big{(}t,x,\Phi^{-1}\left(\frac{x}{\varepsilon},\omega\right),\omega\Big{)},$ (3.49) where the functions $u_{k}(t,x,y,\omega)$ are conveniently stationary in $y$, and $S_{\varepsilon}$ is a real valued function to be established a posteriori (not necessarily a polynomial in $\varepsilon$), which take part of the modulated plane wave (3.49) from $e^{2\pi iS_{\varepsilon}(t,x)}$. The spatial derivative of the above ansatz (3.49) is $\displaystyle\nabla u_{\varepsilon}$ $\displaystyle(t,x,\omega)=e^{2i\pi S_{\varepsilon}(t,x)}\big{(}2i\pi\nabla S_{\varepsilon}(t,x)\sum_{k=0}^{\infty}\varepsilon^{k}\,u_{k}\big{(}t,x,\Phi^{-1}\left(\frac{x}{\varepsilon},\omega\right),\omega\big{)}$ $\displaystyle\qquad+\sum_{k=0}^{\infty}\varepsilon^{k}\Big{\\{}\left(\partial_{x}u_{k}\right)\Big{(}t,x,\Phi^{-1}\left(\frac{x}{\varepsilon},\omega\right),\omega\Big{)}$ $\displaystyle\qquad+\frac{1}{\varepsilon}(\nabla\Phi)^{-1}\left(\Phi^{-1}\left(\frac{x}{\varepsilon},\omega\right),\omega\right)\left(\partial_{y}u_{k}\right)\Big{(}t,x,\Phi^{-1}\left(\frac{x}{\varepsilon},\omega\right),\omega\Big{)}\Big{\\}}\Big{)}$ $\displaystyle=e^{2i\pi S_{\varepsilon}(t,x)}\Big{(}\sum_{k=0}^{\infty}\varepsilon^{k}\left(\frac{{\nabla}_{z}}{\varepsilon}+2i\pi\nabla S_{\varepsilon}(t,x)\right)u_{k}\Big{(}t,x,\Phi^{-1}(\frac{x}{\varepsilon},\omega),\omega\Big{)}$ $\displaystyle\qquad+\sum_{k=0}^{\infty}\varepsilon^{k}\left(\nabla_{x}u_{k}\right)\Big{(}t,x,\Phi^{-1}\left(\frac{x}{\varepsilon},\omega\right),\omega\Big{)}\Big{)}.$ Now, computing the second derivatives of the expansion (3.49) and writing as a cascade of the power of $\varepsilon$, we have $\displaystyle e^{-2i\pi S_{\varepsilon}(t,x)}$ $\displaystyle{\rm div}{\big{(}A{(\Phi^{-1}{\left(\frac{x}{\varepsilon},\omega\right)},\omega)}\nabla u_{\varepsilon}(t,x,\omega)\big{)}}$ (3.50) $\displaystyle=\frac{1}{\varepsilon^{2}}\Big{(}{\rm div}_{\\!z}+2i\pi\varepsilon\nabla S_{\varepsilon}(t,x)\Big{)}\Big{(}A{\left(\Phi^{-1}(\cdot,\omega),\omega\right)}\Big{(}\nabla_{\\!\\!z}+2i\pi\varepsilon\nabla S_{\varepsilon}(t,x)\Big{)}$ $\displaystyle\qquad\qquad\qquad\qquad u_{0}(t,x,\Phi^{-1}(\cdot,\omega),\omega)\Big{)}{\Bigg{\rvert}}_{z=x/\varepsilon}$ $\displaystyle+\frac{1}{\varepsilon}\Big{(}{\rm div}_{\\!z}+2i\pi\varepsilon\nabla S_{\varepsilon}(t,x)\Big{)}\Big{(}A{\left(\Phi^{-1}(\cdot,\omega),\omega\right)}\Big{(}\nabla_{\\!\\!z}+2i\pi\varepsilon\nabla S_{\varepsilon}(t,x)\Big{)}$ $\displaystyle\qquad\qquad\qquad\qquad u_{1}(t,x,\Phi^{-1}(\cdot,\omega),\omega)\Big{)}{\Big{\rvert}}_{z=x/\varepsilon}+I_{\varepsilon},$ where $\displaystyle I_{\varepsilon}=\sum_{k=0}^{\infty}\varepsilon^{k}\Big{(}{\rm div}_{\\!z}+2i\pi\varepsilon\nabla S_{\varepsilon}(t,x)\Big{)}\Big{(}A{\left(\Phi^{-1}(\cdot,\omega),\omega\right)}\Big{(}\nabla_{\\!\\!z}+2i\pi\varepsilon\nabla S_{\varepsilon}(t,x)\Big{)}$ $\displaystyle\qquad\qquad u_{k+2}(t,x,\Phi^{-1}(\cdot,\omega),\omega)\Big{)}{\Big{\rvert}}_{z=x/\varepsilon}$ $\displaystyle+\frac{1}{\varepsilon}\Big{(}{\rm div}_{\\!z}+2i\pi\varepsilon\nabla S_{\varepsilon}(t,x)\Big{)}\Big{(}A{\left(\Phi^{-1}(\cdot,\omega),\omega\right)}\nabla_{x}u_{0}(t,x,\Phi^{-1}(\cdot,\omega),\omega)\Big{)}{\Big{\rvert}}_{z=x/\varepsilon}$ $\displaystyle+\sum_{k=0}^{\infty}\varepsilon^{k}\Big{(}{\rm div}_{\\!z}+2i\pi\varepsilon\nabla S_{\varepsilon}(t,x)\Big{)}\Big{(}A{\left(\Phi^{-1}(\cdot,\omega),\omega\right)}\nabla_{x}u_{k+1}(t,x,\Phi^{-1}(\cdot,\omega),\omega)\Big{)}{\Big{\rvert}}_{z=x/\varepsilon}$ $\displaystyle+\frac{1}{\varepsilon}{\rm div}_{\\!x}\Big{(}A{\left(\Phi^{-1}(\cdot,\omega),\omega\right)}\Big{(}\nabla_{\\!\\!z}+2i\pi\varepsilon\nabla S_{\varepsilon}(t,x)\Big{)}u_{0}(t,x,\Phi^{-1}(\cdot,\omega),\omega)\Big{)}{\Big{\rvert}}_{z=x/\varepsilon}$ $\displaystyle+\sum_{k=0}^{\infty}\varepsilon^{k}{\rm div}_{\\!x}\Big{(}A{\left(\Phi^{-1}(\cdot,\omega),\omega\right)}\Big{(}\nabla_{\\!\\!z}+2i\pi\varepsilon\nabla S_{\varepsilon}(t,x)\Big{)}u_{k+1}(t,x,\Phi^{-1}(\cdot,\omega),\omega)\Big{)}{\Big{\rvert}}_{z=x/\varepsilon}$ $\displaystyle+\sum_{k=0}^{\infty}\varepsilon^{k}{\rm div}_{\\!x}\Big{(}A{\left(\Phi^{-1}(\cdot,\omega),\omega\right)}\nabla_{\\!\\!x}u_{k}{\Big{(}t,x,\Phi^{-1}(\cdot,\omega),\omega\Big{)}}\Big{)}{\Big{\rvert}}_{z=x/\varepsilon}.$ (3.51) Proceeding in the same way with respect to the temporal derivative, we have $\displaystyle e^{-2i\pi S_{\varepsilon}(t,x)}\,{\partial}_{t}u_{\varepsilon}$ $\displaystyle\qquad\qquad=\frac{1}{\varepsilon^{2}}\Big{(}2i\pi\varepsilon^{2}{\partial}_{t}S_{\varepsilon}(t,x)\Big{)}u_{0}\Big{(}t,x,\Phi^{-1}\left(\frac{x}{\varepsilon},\omega\right),\omega\Big{)}$ $\displaystyle\qquad\qquad+\frac{1}{\varepsilon}\Big{(}2i\pi\varepsilon^{2}{\partial}_{t}S_{\varepsilon}(t,x)\Big{)}u_{1}\Big{(}t,x,\Phi^{-1}\left(\frac{x}{\varepsilon},\omega\right),\omega\Big{)}$ $\displaystyle\qquad\qquad+\Big{(}2i\pi\varepsilon^{2}{\partial}_{t}S_{\varepsilon}(t,x)\Big{)}\sum_{k=0}^{\infty}\varepsilon^{k}u_{k+2}\Big{(}t,x,\Phi^{-1}\left(\frac{x}{\varepsilon},\omega\right),\omega\Big{)}$ $\displaystyle\qquad\qquad\qquad\qquad+\sum_{k=0}^{\infty}\varepsilon^{k}{\partial}_{t}u_{k}\Big{(}t,x,\Phi^{-1}\left(\frac{x}{\varepsilon},\omega\right),\omega\Big{)}.$ (3.52) Thus, if we insert the equations (3.50) and (3.1) in (1.1) and compute the $\varepsilon^{-2}$ order term, we arrive at $L^{\Phi}(\varepsilon\nabla S_{\varepsilon}(t,x))u_{0}{\big{(}t,x,\Phi^{-1}(\cdot,\omega),\omega\big{)}}=2\pi\big{(}\varepsilon^{2}\partial_{t}S_{\varepsilon}(t,x)\big{)}u_{0}{\big{(}t,x,\Phi^{-1}(\cdot,\omega),\omega\big{)}},$ where for each $\theta\in\mathbb{R}^{n}$, the linear operator $L^{\Phi}(\theta)$ is defined by $\displaystyle L^{\Phi}(\theta)[\cdot]:=$ $\displaystyle-\big{(}{\rm div}_{\\!z}+2i\pi\theta\big{)}\big{(}A{(\Phi^{-1}(z,\omega),\omega)}{\big{(}\nabla_{\\!\\!z}+2i\pi\theta\big{)}}[\cdot]\big{)}$ (3.53) $\displaystyle+V{\big{(}\Phi^{-1}\left(z,\omega\right),\omega\big{)}}[\cdot].$ Therefore, $2\pi\Big{(}\varepsilon^{2}\partial_{t}S_{\varepsilon}(t,x)\Big{)}$ is an eigenvalue of $L^{\Phi}(\varepsilon\nabla S_{\varepsilon}(t,x))$. Consequently, if $\lambda(\theta)$ is any eigenvalue of $L^{\Phi}(\theta)$ (which is sufficiently regular with respect to $\theta$), then the following (eikonal) Hamilton-Jacobi equation must be satisfied $2\pi\varepsilon^{2}\partial_{t}S_{\varepsilon}(t,x)-\lambda(\varepsilon\nabla S_{\varepsilon}(t,x))=0.$ Thus, if we suppose for $t=0$ (companion to (3.49)) the modulated plane wave initial data $u_{\varepsilon}(0,x,\omega)=e^{2i\pi\frac{\theta\cdot x}{\varepsilon}}\sum_{k=1}^{\infty}\varepsilon^{k}u_{k}\Big{(}0,x,\Phi^{-1}\left(\frac{x}{\varepsilon},\omega\right),\omega\Big{)},$ (3.54) then the unique solution for the above Hamilton-Jacobi equation is, for each parameter $\theta\in\mathbb{R}^{n}$, $S_{\varepsilon}(t,x)=\frac{\lambda(\theta)\ t}{2\pi\varepsilon^{2}}+\frac{\theta\cdot x}{\varepsilon}.$ (3.55) To sum up, the above expansion, that is the solution $u_{\varepsilon}$ of the equation (1.1) with initial data given respectively by (3.49) and (3.54), suggests the following ###### Definition 3.1 (Bloch or shifted spectral cell equation). Let $\Phi$ be a stochastic deformation. For any $\theta\in\mathbb{R}^{n}$ fixed, the following time independent asymptotic equation $\left\\{\begin{array}[]{l}L^{\Phi}(\theta)[\Psi(z,\omega)]=\lambda\ \Psi(z,\omega),\hskip 40.0pt\text{in $\mathbb{R}^{n}\times\Omega$},\\\\[5.0pt] \hskip 32.0pt\Psi(z,\omega)=\psi{\left(\Phi^{-1}(z,\omega),\omega\right)},\quad\text{$\psi$ is a stationary function},\end{array}\right.$ (3.56) is called Bloch’s spectral cell equation companion to the Schrödinger equation in (1.1), where $L^{\Phi}(\theta)$ is given by (3.53). Moreover, each $\theta\in\mathbb{R}^{n}$ is called a Bloch frequency, $\lambda(\theta)$ is called a Bloch energy and the corresponded $\Psi(\theta)$ is called a Bloch wave. Moreover, if $\Phi$ is well understood in the context, then $L\equiv L^{\Phi}$. The unknown $(\lambda,\Psi)$ in (3.56), which is an eigenvalue-eigenfunction pair, is obtained by the associated variational formulation, that is $\displaystyle\langle L(\theta)[F],G\rangle$ (3.57) $\displaystyle=\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}\\!\\!\\!\\!\\!\\!\\!\\!\\!A(\Phi^{-1}(z,\omega),\omega)(\nabla_{\\!\\!z}+2i\pi\theta)F(z,\omega)\cdot\overline{{(\nabla_{\\!\\!z}+2i\pi\theta)}G(z,\omega)}\,dz\,d\mathbb{P}(\omega)$ $\displaystyle+\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}V{(\Phi^{-1}(z,\omega),\omega)}\ F(z,\omega)\,\overline{G(z,\omega)}\,dz\,d\mathbb{P}(\omega).$ ###### Remark 3.2. One remarks that, $\lambda=\lambda(\theta)\in\mathbb{R}$, that is to say, $\lambda$ depends on the parameter $\theta$. However, $\lambda$ could not depend on $\omega$, since the homogeneized effective matrix is obtained from the Hessian of $\lambda$ at some point $\theta^{*}$, and should be constant. Therefore, the probabilistic variable $\omega$ could not be considered as fixed parameter in (3.56). ### 3.2 Sobolev spaces on groups The main motivation to study Sobolev spaces on groups, besides being an elegant and modern mathematical theory, is related to the eigenvalue problem: Find $\lambda(\theta)\in\mathbb{R}$ and $\Psi(\theta)\in\mathcal{H}_{\Phi}\setminus\\{0\\}$ satisfying (3.56). Indeed, we may use a compactness argument, that is the space $\mathcal{H}_{\Phi}$ is compactly embedded in $\mathcal{L}_{\Phi}$, in order to solve the associated variational formulation (3.57). Although, as observed in Remark 3.2, $\omega\in\Omega$ can not be fixed, hence we are going to establish an equivalence between the space $\mathcal{H}_{\Phi}$ and the Sobolev space on groups, and then consider a related Rellich-Kondrachov’s Theorem. This is the main issue of this section. Let us recall that, the subject of Sobolev spaces on Abelian locally compact groups, to the best of our knowledge, was introduced by P. Górka, E. G. Reyes [21]. To begin, we sum up some definitions and properties of topological groups, which will be used along this section. Most of the material could be found in E. Hewitt, A. Ross [23] and G. B. Folland [19] (with more details). A nonempty set $G$ endowed with an application, $\ast:G\\!\times\\!G\to G$, is called a group, when for each $x,y,z\in G$: * 1. ${(x\ast y)\ast z=x\ast(y\ast z)}$; * 2. There exists ${e\in G}$, such that ${x\ast e=e\ast x=e}$; * 3. For all ${y\in G}$, there exists ${y^{-1}\in G}$, such that ${y\ast y^{-1}=y^{-1}\ast y=e}$. Moreover, if $x\ast y=y\ast x$, then $G$ is called an Abelian group. From now on, we write for simplicity $x\,z$ instead of $x\ast z$. A topological group is a group $G$ together with a topology, such that, both the group’s binary operation $(x,y)\mapsto x\,y$, and the function mapping group elements to their respective inverses $x\mapsto x^{-1}$ are continuous functions with respect to the topology. Unless the contrary is explicit stated, any group mentioned here is a locally compact Abelian (LCA for short) group, and we may assume without loss of generality that, the associated topology is Hausdorff (see G. B. Folland [19], Corollary 2.3). A complex value function $\xi:G\to\mathbb{S}^{1}$ is called a character of $G$, when $\xi(x\,y)=\xi(x)\xi(y),\quad\quad\text{(for each $x,y\in G$)}.$ We recall that, the set of characters of $G$ is an Abelian group with the usual product of functions, identity element $e=1$, and inverse element $\xi^{-1}=\overline{\xi}$. The characters’ group of the topological group $G$, called the dual group of $G$ and denoted by $G^{\wedge}$, is the set of all continuous characters, that is to say $G^{\wedge}:=\\{\xi:G\to\mathbb{S}^{1}\;;\;\text{$\xi$ is a continuous homomorphism}\\}.$ Moreover, we may endow $G^{\wedge}$ with a topology with respect to which, $G^{\wedge}$ itself is a LCA group. We denote by $\mu$, $\nu$ the unique (up to a positive multiplicative constant) Haar mesures in $G$ and $G^{\wedge}$ respectively. The $L^{p}$ spaces over $G$ and its dual are defined as usual, with their respective mesures. Let us recall two important properties when $G$ is compact: $\displaystyle i)\quad\text{If $\mu(G)=1$, then $G^{\wedge}$ is an orthonormal set in $L^{2}(G;\mu)$}.$ (3.58) $\displaystyle ii)\quad\text{The dual group $G^{\wedge}$ is discrete, and $\nu$ is the countermeasure}.$ One remarks that, the study of Sobolev spaces on LCA groups uses essentially the concept of Fourier Transform, then we have the following ###### Definition 3.3. Given a complex value function $f\in L^{1}(G;\mu)$, the function $\widehat{f}:G^{\wedge}\to\mathbb{C}$, defined by $\widehat{f}(\xi):=\int_{G}f(x)\,\overline{\xi(x)}\,d\mu(x)$ (3.59) is called the Fourier transform of $f$ on $G$. Usually, the Fourier Transform of $f$ is denoted by ${\mathcal{F}}f$ to emphasize that it is an operator, but we prefer to adopt the usual notation $\widehat{f}$. Moreover, we recall that the Fourier transform is an homomorphism from $L^{1}(G;\mu)$ to $C_{0}(G^{\wedge})$ (or $C(G^{\wedge})$ when $G$ is compact), see Proposition 4.13 in [19]. Also we address the reader to [19], Chapter 4, for the Plancherel Theorem and the Inverse Fourier Transform. Before we establish the definition of (energy) Sobolev spaces on LCA groups, let us consider the following set $\displaystyle{\rm P}=\\{p:G^{\wedge}\times$ $\displaystyle G^{\wedge}\to[0,\infty)/$ $\displaystyle\text{$p$ is a continuous invariant pseudo-metric on $G^{\wedge}$}\\}.$ The Birkhoff-Kakutani Theorem (see [23] p.68) ensures that, the set P is not empty. Any pseudo-metric $p\in{\rm P}$ is well defined for each $(x,y)\in G^{\wedge}\times G^{\wedge}$, hence we may define $\gamma(x):=p(x,e)\equiv p(x,1).$ (3.60) Moreover, one observes that $\gamma(1)=0$. Then, we have the following ###### Definition 3.4 (Energy Sobolev Spaces on LCA Groups). Let $s$ be a non-negative real number and $\gamma(x)$ be given by (3.60) for some fixed $p\in{\rm P}$. The energy Sobolev space $H^{s}_{\gamma}(G)$ is the set of functions $f\in L^{2}(G;\mu)$, such that $\int_{G^{\wedge}}(1+\gamma(\xi)^{2})^{s}\,|\widehat{f}(\xi)|^{2}d\nu(\xi)<\infty.$ (3.61) Moreover, given a function $f\in H^{s}_{\gamma}(G)$ its norm is defined as $\|f\|_{H^{s}_{\gamma}(G)}:=\left(\int_{G^{\wedge}}\left(1+\gamma(\xi)^{2}\right)^{s}|\widehat{f}(\xi)|^{2}d\nu(\xi)\right)^{1/2}.$ (3.62) Below, taking specific functions $\gamma$, the usual Sobolev spaces on $\mathbb{R}^{d}$ and other examples are considered. In particular, the Plancherel Theorem implies that, $H^{0}_{\gamma}(G)=L^{2}(G;\mu)$. ###### Example 3.5. Let $G=(\mathbb{R}^{n},+)$ which is known to be a LCA group, and consider its dual group $(\mathbb{R}^{n})^{\wedge}=\\{\xi_{y}\;;\;y\in\mathbb{R}^{n}\\}$, where for each $x\in\mathbb{R}^{n}$ $\xi_{y}(x)=e^{2\pi i\,y\cdot x},$ (3.63) hence $|\xi_{y}(x)|=1$ and $\xi_{0}(x)=1$. One remarks that, here we denote (without invocation of vector space structure) $a\cdot b=a_{1}b_{1}+a_{2}b_{2}+\ldots+a_{n}b_{n},\quad\text{(for all $a,b\in G$)}.$ For any $x,y\in\mathbb{R}^{n}$ let us consider $p(\xi_{x},\xi_{y})=2\pi\|x-y\|,$ where $\|\cdot\|$ is the Euclidean norm in $\mathbb{R}^{n}$. Hence $\gamma(\xi_{x})=p(\xi_{x},1)=2\pi\|x\|$. Since $(\mathbb{R}^{n})^{\wedge}\cong\mathbb{R}^{n}$, the Sobolev space $H^{s}_{\gamma}(G)$ coincide with the usual Sobolev space on $\mathbb{R}^{n}$. ###### Example 3.6. Let us recall that, the set $[0,1)^{n}$ endowed with the binary operation $(x,y)\in[0,1)^{n}\\!\times\\![0,1)^{n}\;\;\mapsto\;\;x+y-\left\lfloor x+y\right\rfloor\in[0,1)^{d}$ is an Abelian group, and the function $\Lambda:\mathbb{R}^{n}\to[0,1)^{n}$, $\Lambda(x):=x-\left\lfloor x\right\rfloor$ is an homomorphism of groups. Moreover, under the induced topology by $\Lambda$, that is to say $\\{U\subset[0,1)^{n}\;;\;\Lambda^{-1}(U)\;\text{is an open set of}\;\,\mathbb{R}^{n}\\},$ $[0,1)^{n}$ is a compact Abelian group, which is called $n-$dimensional Torus and denoted by $\mathbb{T}^{n}$. Its dual group is characterized by the integers $\mathbb{Z}^{n}$, that is $\text{ $(\mathbb{T}^{n})^{\wedge}=\\{\xi_{m}\;;\;m\in\mathbb{Z}^{n}\\}$, where $\xi_{m}(x)$ is given by \eqref{caracterunitario} for all $x\in\mathbb{R}^{n}$}.$ For each $m,k\in\mathbb{Z}^{n}$, we consider $p(\xi_{m},\xi_{k})=2\pi\sum_{j=1}^{n}{|m_{j}-k_{j}|},\quad\text{and thus $\gamma(\xi_{m})=2\pi\sum_{j=1}^{n}{|m_{j}|}$}.$ Then, the Sobolev space $H^{s}_{\gamma}(\mathbb{T}^{n})$ coincide with the usual Sobolev space on $\mathbb{T}^{n}$. Now, following the above discussion let us consider the infinite Torus $\mathbb{T}^{I}$, where $I$ is an index set. Since an arbitrary product of compact spaces is compact in the product topology (Tychonoff Theorem), $\mathbb{T}^{I}$ is a compact Abelian group. Here, the binary operation on $\mathbb{T}^{I}\times\mathbb{T}^{I}$ is defined coordinate by coordinate, that is, for each $\ell\in I$ $g_{\ell}+h_{\ell}:=g_{\ell}+h_{\ell}-\left\lfloor g_{\ell}+h_{\ell}\right\rfloor.$ Moreover, the set $\mathbb{Z}^{I}_{\rm c}:=\\{m\in\mathbb{Z}^{I};\text{{\rm supp} $m$ is compact}\\}$ characterizes the elements of the dual group $(\mathbb{T}^{I})^{\wedge}$. Indeed, applying Theorem 23.21 in [23], similarly we have $(\mathbb{T}^{I})^{\wedge}={\left\\{\xi_{m}\;;\;m\in\mathbb{Z}^{I}_{\rm c}\right\\}},$ where $\xi_{m}(k)$ is given by (3.63) for each $m,k\in\mathbb{Z}_{\rm c}^{I}$, the pseudo-metric $p(\xi_{m},\xi_{k})=2\pi\sum_{\ell\in I}{|m_{\ell}-k_{\ell}|},\quad\text{and $\gamma(\xi_{m})=2\pi\sum_{\ell\in I}{|m_{\ell}|}$}.$ Consequently, we have establish the Sobolev spaces $H^{s}_{\gamma}(\mathbb{T}^{I})$. #### 3.2.1 Groups and Dynamical systems In this section, we are interested to come together the discussion about dynamical systems studied in Section 2.2 with the theory developed in the last section for LCA groups. To this end, we consider stationary functions in the continuous sense (continuous dynamical systems). Moreover, we recall that all the groups in this paper are assumed to be Hausdorff. To begin, let $G$ be a locally compact group with Haar measure $\mu$, we know that $\mu(G)<\infty$ if, and only if, $G$ is compact. Therefore, we consider from now on that $G$ is a compact Abelian group, hence $\mu$ is a finite measure and, up to a normalization, $(G,\mu)$ is a probability space. We are going to consider the dynamical systems, $\tau:\mathbb{R}^{n}\times G\to G$, defined by $\tau(x)\omega:=\varphi(x)\,\omega,$ (3.64) where $\varphi:\mathbb{R}^{n}\to G$ is a given (continuous) homomorphism. Indeed, first $\tau(0)\omega=\omega$ and $\tau(x+y,\omega)=\varphi(x)\varphi(y)\omega=\tau(x,\tau(y)\omega)$. Moreover, since $\mu$ is a translation invariant Haar measure, the mapping $\tau(x,\cdot):G\to G$ is $\mu-$measure preserving. Recall from Remark 2.11 we have assumed that, the dynamical systems we are interested here are ergodic. Then, it is important to characterize the conditions for the mapping $\varphi$, under which the dynamical system defined by (3.64) is ergodic. To this end, first let us consider the following ###### Lemma 3.7. Let $H$ be a topological group, $F\subset H$ closed, $F\neq H$ and $x\notin F$. Then, there exists a neighborwood $V$ of the identity $e$, such that $FV\cap xV=\emptyset.$ ###### Proof. First, we observe that: i) Since $F\subset H$ is closed and $F\neq H$, there exists a neighborwood $U$ of the identity $e$, such that $F\cap xU=\emptyset$. ii) There exists a symmetric neighborwood $V$ of the identity $e$, such that $VV\subset U$. Now, suppose that $FV\cap xV\neq\emptyset$. Therefore, there exist $v_{1},v_{2}\in V$ and $k_{0}\in F$ such that, $k_{0}v_{1}=xv_{2}$. Consequently, $k_{0}=xv_{2}v_{1}^{-1}$ and from $(ii)$ it follows that, $k_{0}\in xU$. Then, we have a contradiction from $(i)$. ∎ Claim 1: The dynamical system defined by (3.64) is ergodic if, and only if, $\varphi(\mathbb{R}^{n})$ is dense in $G$. Proof of Claim 1: 1. Let us show first the necessity. Therefore, we suppose that $\varphi(\mathbb{R}^{n})$ is not dense in $G$, that is $K:=\overline{\varphi(\mathbb{R}^{n})}\neq G$. Then, applying Lemma 3.7 there exists a neighborhood $V$ of $e$, such that $KV\cap xV=\emptyset$, for some $x\notin K$. Recall that the Haar measure on open sets are positive, moreover $KV=\bigcup_{k\in K}kV,$ which is an open set, thus we have $0<\mu(KV)+\mu(xV)\leq 1.$ Consequently, it follows that $0<\mu(\varphi(\mathbb{R}^{n})V)<1$. For convenience, le us denote $E=\varphi(\mathbb{R}^{n})V$, hence $\tau(x)E=E$ for each $x\in\mathbb{R}^{n}$. Then, the dynamical system $\tau$ is not ergodic, since $E\subset G$ is a $\tau$-invariant set with $0<\mu(E)<1$. 2\. It remains to show the sufficiency. Let $E\subset G$ be a $\mu-$measurable $\tau$-invariant set, hence $\omega E=E$ for each $\omega\in\varphi(\mathbb{R}^{n})$. Assume by contradiction that, $0<\mu(E)<1$, thus $\mu(G\setminus E)>0$. Denote by $\mathcal{B}_{G}$ the Borel $\sigma-$algebra on $G$, and define, $\lambda:=\mu_{\lfloor E}$, that is $\lambda(A)=\mu(E\cap A)$ for all $A\in\mathcal{B}_{G}$. Recall that $G$ is not necessarily metric, therefore, it is not clear if each Borel set is $\mu-$measurable. Then, it follows that: $(i)$ For any $A\in\mathcal{B}_{G}$ fixed, the mapping $\omega\in G\mapsto\lambda(\omega A)$ is continuous. Indeed, for $\omega\in G$ and $A\in\mathcal{B}_{G}$, we have $\displaystyle\lambda(\omega A)$ $\displaystyle=\int_{G}1_{E}(\varpi)1_{\omega A}(\varpi)d\mu(\varpi)$ $\displaystyle=\int_{G}1_{E}(\varpi)1_{A}(\omega^{-1}\varpi)d\mu(\varpi)=\int_{G}1_{E}(\omega\varpi)1_{A}(\varpi)d\mu(\varpi).$ Therefore, for $\omega,\omega_{0}\in G$ $\displaystyle|\lambda(\omega A)-\lambda(\omega_{0}A)|$ $\displaystyle=\big{|}\int_{G}\big{(}1_{E}(\omega\varpi)-1_{E}(\omega_{0}\varpi)\big{)}1_{A}(\varpi)d\mu(\varpi)\big{|}$ $\displaystyle\leq\big{(}\mu(A)\big{)}^{1/2}\big{(}\int_{G}|1_{E}(\omega\varpi)-1_{E}(\omega_{0}\varpi)|^{2}d\mu(\varpi)\big{)}^{1/2}\raisebox{-7.3194pt}{$\stackrel{{\scriptstyle\textstyle{\longrightarrow}}}{{\scriptstyle{\omega\to\omega_{0}}}}$}0.$ $(ii)$ $\lambda$ is invariant, i.e. for all $\omega\in G$, and $A\in\mathcal{B}_{G}$, $\lambda(\omega A)=\lambda(A)$. Indeed, for each $\omega\in\varphi(\mathbb{R}^{d})$, and $A\in\mathcal{B}_{G}$, we have $(\omega A)\cap E=(\omega A)\cap(\omega E)=\omega(A\cap E).$ Thus since $\mu$ is invariant, $\mu_{\lfloor E}(\omega A)=\mu_{\lfloor E}(A)$. Consequently, due to item $(i)$ and $\overline{\varphi(\mathbb{R}^{d})}=G$, it follows that $\lambda$ is invariant. From item $(ii)$ the Radon measure $\lambda$ is a Haar measure on $G$. By the uniqueness of the Haar measure on $G$, there exists $\alpha>0$, such that for all $A\in\mathcal{B}_{G}$, $\alpha\lambda(A)=\mu(A)$. In particular, $\alpha\lambda(G\setminus E)=\mu(G\setminus E)$. But $\lambda(G\setminus E)=0$ by definition and $\mu(G\setminus E)>0$, which is a contradiction and hence $\tau$ is ergodic. ###### Remark 3.8. 1\. One remarks that, in order to show that $\tau$ given by (3.64) is ergodic, it was not used that $\varphi$ is continuous, nor that $G$ is metric. Compare with the statement in [24] p.225 (after Theorem 7.2). 2\. From now on, we assume that $\varphi(\mathbb{R}^{n})$ is dense in $G$. Now, for the dynamical system established before, the main issue is to show how the Sobolev space $H^{1}_{\gamma}(G)$ is related with the space $\mathcal{H}_{\Phi}$ given by (2.15) for $\Phi=Id$, that is $\mathcal{H}={\big{\\{}f(y,\omega);\;f\in H^{1}_{\rm loc}(\mathbb{R}^{n};L^{2}(G))\;\;\text{stationary}\big{\\}}},$ which is a Hilbert space endowed with the following inner product ${\langle f,g\rangle}_{\mathcal{H}}=\int_{G}f(0,\omega)\,\overline{g(0,\omega)}\,d\mu(\omega)+\int_{G}\nabla_{\\!\\!y}f(0,\omega)\cdot\overline{\nabla_{\\!\\!y}g(0,\omega)}\,d\mu(\omega).$ Let $\chi$ be a character on $G$, i.e. $\chi\in G^{\wedge}$. Since $\varphi:\mathbb{R}^{n}\to G$ is a continuous homomorphism, the function $(\chi\circ\varphi):\mathbb{R}^{n}\to\mathbb{C}$ is a continuous character in $\mathbb{R}^{n}$. More precisely, given any fixed $\chi\in G^{\wedge}$ we may find $y\in\mathbb{R}^{n}$, $(y\equiv y(\chi))$, such that, for each $x\in\mathbb{R}^{n}$ $\big{(}\chi\circ\varphi\big{)}(x)=:\xi_{y(\chi)}(x)=e^{2\pi i\,y(\chi)\cdot x}.$ Following Example 3.5 we define the pseudo-metric $p_{\varphi}:G^{\wedge}\times G^{\wedge}\to[0,\infty)$ by $p_{\varphi}(\chi_{1},\chi_{2}):=p(\xi_{y_{1}(\chi_{1})},\xi_{y_{2}(\chi_{2})})=2\pi\|y_{1}(\chi_{1})-y_{2}(\chi_{2})\|.$ (3.65) Then, we have $\gamma(\chi)=p_{\varphi}(\chi,1)=2\pi\|y(\chi)\|.$ Let us observe that, we have used in the above construction of $\gamma$ the continuity of the homomorphism $\varphi:\mathbb{R}^{n}\to G$, that is to say, it was essential the continuity of $\varphi$. In fact, the function $\gamma$ was given by the pseudo-metric $p_{\varphi}$, which is not necessarily a metric. Although, we have the following Claim 2: The pseudo-metric $p_{\varphi}:G^{\wedge}\times G^{\wedge}\to[0,\infty)$ given by (3.65) is a metric if, and only if, $\varphi(\mathbb{R}^{n})$ is dense in $G$. Proof of Claim 2: 1. First, let us assume that $\overline{\varphi(\mathbb{R}^{n})}\neq G$, and then show that $p_{\varphi}$ is not a metric. Therefore, we have the necessity proved. From Corollary 24.12 in [23], since $\overline{\varphi(\mathbb{R}^{n})}$ is a closer proper subgroup of $G$, hence there exists $\xi\in G^{\wedge}\setminus\\{1\\}$, such that $\xi(\overline{\varphi(\mathbb{R}^{n})})=\\{1\\}$. Hence there exists $\xi\in G^{\wedge}\setminus\\{1\\}$, such that, $\xi(\varphi(x))=1$, for each $x\in\mathbb{R}^{n}$, i.e. $y(\xi)=0$. Therefore, we have $p_{\varphi}(\xi,1)=0$, which implies that $p_{\varphi}$ is not a metric. 2\. Now, let us assume that $\overline{\varphi(\mathbb{R}^{n})}=G$, and it is enough to show that if $p_{\varphi}(\xi,1)=0$, then $\xi=1$. Indeed, if $0=p_{\varphi}(\xi,1)=2\pi\|y(\xi)\|$, then $y(\xi)=0$. Therefore, $\xi(\varphi(x))=1$ for each $x\in\mathbb{R}^{d}$, since $\xi$ is continuous and $\overline{\varphi(\mathbb{R}^{n})}=G$, it follows that, for each $\omega\in G$, $\xi(\omega)=1$, which finishes the proof of the claim. ###### Remark 3.9. Since we have already assumed that $\varphi(\mathbb{R}^{n})$ is dense in $G$, it follows that $p_{\varphi}$ is indeed a metric, which does not imply necessarily that $G$, itself, is metric. Under the assumptions considered above, we have the following ###### Lemma 3.10. If $f\in\mathcal{H}$, then for $j\in\\{1,\ldots,d\\}$ and all $\xi\in G^{\wedge}$ $\widehat{\partial_{j}f(0,\xi)}=2\pi i\;y_{j}(\xi)\widehat{f(0,\xi)}.$ (3.66) ###### Proof. First, for each $x\in\mathbb{R}^{d}$ and $\omega\in G$, define $\displaystyle\xi_{\tau}(x,\omega)$ $\displaystyle:=\xi(\tau(x,\omega))=\xi(\varphi(x)\omega)=\xi(\varphi(x))\;\xi(\omega)$ $\displaystyle=e^{2\pi ix\cdot y(\xi)}\;\xi(\omega).$ Therefore $\xi_{\tau}\in C^{\infty}(\mathbb{R}^{d};L^{2}(G))$, and we have for $j\in\\{1,\ldots,d\\}$ $\partial_{j}\xi_{\tau}(0,\omega)=2\pi i\;y_{j}(\xi)\;\xi(\omega).$ (3.67) Finally, applying Theorem 2.15 we obtain $\displaystyle\int_{G}\partial_{j}f(0,\omega)\;\overline{\xi_{\tau}}(0,\omega)d\mu(\omega)$ $\displaystyle=-\int_{G}f(0,\omega)\;\partial_{j}\overline{\xi_{\tau}}(0,\omega)d\mu(\omega)$ $\displaystyle=2\pi i\;y_{j}(\xi)\int_{G}f(0,\omega)\;\overline{\xi}(\omega)d\mu(\omega),$ where we have used (3.67). From the above equation and the definition of the Fourier transform on groups we obtain (3.66), and the lemma is proved. ∎ Now we are able to state the equivalence between the spaces $\mathcal{H}$ and $H^{1}_{\gamma}(G)$, which is to say, we have the following ###### Theorem 3.11. A function $f\in\mathcal{H}$ if, and only if, $f(0,\cdot)\in H^{1}_{\gamma}(G)$, and $\|f\|_{\mathcal{H}}=\|f(0,\cdot)\|_{H_{\gamma}^{1}(G)}.$ ###### Proof. 1\. Let us first show that, if $f\in\mathcal{H}$ then $f\in H^{1}_{\gamma}(G)$. To follow we observe that $\displaystyle\int_{G^{\wedge}}(1+\gamma(\xi)^{2})|\widehat{f(0,\xi)}|^{2}\;d\nu(\xi)$ $\displaystyle=\int_{G^{\wedge}}|\widehat{f(0,\xi)}|^{2}\;d\nu(\xi)$ $\displaystyle+\int_{G^{\wedge}}|2\pi i\;y(\xi)\widehat{f(0,\xi)}|^{2}\;d\nu(\xi)$ $\displaystyle=\int_{G^{\wedge}}|\widehat{f(0,\xi)}|^{2}\;d\nu(\xi)+\int_{G^{\wedge}}|\widehat{\nabla_{\\!\\!y}f(0,\xi)}|^{2}\;d\nu(\xi),$ where we have used (3.66). Therefore, applying Plancherel theorem $\int_{G^{\wedge}}\\!(1+\gamma(\xi)^{2})|\widehat{f(0,\xi)}|^{2}\;d\nu(\xi)=\\!\\!\int_{G}\\!|{f(0,\omega)}|^{2}\;d\mu(\omega)+\\!\int_{G}|\nabla_{\\!\\!y}{f(0,\omega)}|^{2}\;d\mu(\omega)\\!<\\!\infty,$ and thus $f(0,\cdot)\in H^{1}_{\gamma}(G)$. 2\. Now, let $f(x,\omega)$ be a stationary function, such that $f(0,\cdot)\in H^{1}_{\gamma}(G)$, then we show that $f\in\mathcal{H}$. Given a stationary function $\zeta\in C^{1}(\mathbb{R}^{d};L^{2}(G))$, applying the Palncherel theorem and polarization identity $\int_{G}\partial_{j}\zeta(0,\omega)\;\overline{f(0,\omega)}d\mu(\omega)=\int_{G^{\wedge}}\widehat{\partial_{j}\zeta(0,\xi)}\;\overline{\widehat{f(0,\xi)}}d\nu(\xi)$ for $j\in\\{1,\ldots,d\\}$. Due to (3.66), we may write $\displaystyle\int_{G}\partial_{j}\zeta(0,\omega)\;\overline{f(0,\omega)}d\mu(\omega)$ $\displaystyle=\int_{G^{\wedge}}2\pi i\;y_{j}(\xi)\widehat{\zeta(0,\xi)}\;\overline{\widehat{f(0,\xi)}}d\nu(\xi)$ (3.68) $\displaystyle=-\int_{G^{\wedge}}\widehat{\zeta(0,\xi)}\;\overline{2\pi i\;y_{j}(\xi)\widehat{f(0,\xi)}}d\nu(\xi).$ For $j\in\\{1,\ldots,d\\}$ we define, $g_{j}(\omega):=\big{(}2\pi i\;y_{j}(\xi)\widehat{f(0,\xi)}\big{)}^{\vee}$, then $g_{j}\in L^{2}(G)$. Indeed, we have $\int_{G}|g_{j}(\omega)|^{2}d\mu(\omega)=\int_{G^{\wedge}}|\widehat{g_{j}(\xi)}|^{2}d\nu(\xi)\leq\int_{G^{\wedge}}(1+\gamma(\xi)^{2})|\widehat{f(0,\xi)}|^{2}d\nu(\xi)<\infty.$ Therefore, we obtain from (3.68) $\int_{G}\partial_{j}\zeta(0,\omega)\;\overline{f(0,\omega)}d\mu(\omega)=-\int_{G}\zeta(0,\omega)\;\overline{g_{j}(\omega)}d\mu(\omega)$ for any stationary function $\zeta\in C^{1}(\mathbb{R}^{d};L^{2}(G))$, and $j\in\\{1,\ldots,d\\}$. Then $f\in\mathcal{H}$ due to Theorem 2.15. ∎ ###### Corollary 3.12. Let $f\in L^{2}_{\text{loc}}(\mathbb{R}^{d};L^{2}(G))$ be a stationary function and $\Phi$ a stochastic deformation. Then, $f\circ\Phi^{-1}\in{\mathcal{H}}_{\Phi}$ if, and only if, $f(0,\cdot)\in H^{1}_{\gamma}(G)$, and there exist constants $C_{1},C_{2}>0$, such that $C_{1}\|f\circ\Phi^{-1}\|_{\mathcal{H}_{\Phi}}\leq\|f(0,\cdot)\|_{H_{\gamma}^{1}(G)}\leq C_{2}\|f\circ\Phi^{-1}\|_{\mathcal{H}_{\Phi}}.$ ###### Proof. Follows from Theorem 3.11 and Remark 2.10. ∎ #### 3.2.2 Rellich–Kondrachov type Theorem The aim of this section is to characterize when the Sobolev space $H^{1}_{\gamma}(G)$ is compactly embedded in $L^{2}(G)$, written $H^{1}_{\gamma}(G)\subset\subset L^{2}(G)$, where $G$ is considered a compact Abelian group and $\gamma:G^{\wedge}\to[0,\infty)$ is given by (3.60). We observe that, $H^{1}_{\gamma}(G)\subset\subset L^{2}(G)$ is exactly the Rellich–Kondrachov Theorem on compact Abelian groups, which was established under some conditions on $\gamma$ in [21]. Nevertheless, as a byproduct of the characterization established here, we provide the proof of this theorem in a more precise context. To start the investigation, let $(G,\mu)$ be a probability space and consider the operator $T:L^{2}(G^{\wedge})\to L^{2}(G^{\wedge})$, defined by $[T(f)](\xi):=\frac{f(\xi)}{\sqrt{(1+\gamma(\xi)^{2})}}.$ (3.69) We remark that, $T$ as defined above is a bounded linear, ($\|T\|\leqslant 1$), self-adjoint operator, which is injective and satisfies for each $f\in L^{2}(G^{\wedge})$ $\int_{G^{\wedge}}\left(1+\gamma(\xi)^{2}\right)\,{|[T(f)](\xi)|}^{2}d\nu(\xi)=\int_{G^{\wedge}}|f(\xi)|^{2}d\nu(\xi).$ (3.70) Moreover, a function $f\in H^{1}_{\gamma}(G)$ if, and only if, $\widehat{f}\in T(L^{2}(G^{\wedge}))$, that is to say $f\in H^{1}_{\gamma}(G)\Leftrightarrow\widehat{f}\in T(L^{2}(G^{\wedge})).$ (3.71) Indeed, if $f\in H^{1}_{\gamma}(G)$ then, we have $f\in L^{2}(G)$ and $\int_{G^{\wedge}}\left(1+\gamma(\xi)^{2}\right)|\widehat{f}(\xi)|^{2}d\nu(\xi)=\int_{G^{\wedge}}|\sqrt{\left(1+\gamma(\xi)^{2}\right)}\,\widehat{f}(\xi)|^{2}d\nu(\xi)<\infty.$ Therefore, defining $g(\xi):=\sqrt{\left(1+\gamma(\xi)^{2}\right)}\widehat{f(\xi)}$, hence $g\in L^{2}(G^{\wedge})$ and we have $\widehat{f}\in T(L^{2}(G^{\wedge}))$. Now, if $\widehat{f}\in T(L^{2}(G^{\wedge}))$ let us show that, $f\in H^{1}_{\gamma}(G)$. First, there exists $g\in L^{2}(G^{\wedge})$ such that, $\widehat{f}=T(g)$. Thus from equation (3.70), we obtain $\int_{G^{\wedge}}(1+\gamma(\xi)^{2})\,|\widehat{f}(\xi)|^{2}d\nu(\xi)=\int_{G^{\wedge}}|g(\xi)|^{2}d\nu(\xi)<\infty,$ that is, by definition $f\in H^{1}_{\gamma}(G)$. Then we have the following Equivalence Theorem: ###### Theorem 3.13. The Sobolev space $H^{1}_{\gamma}(G)$ is compactly embedded in $L^{2}(G)$ if, and only if, the operator $T$ defined by (3.69) is compact. ###### Proof. 1\. First, let us assume that $H^{1}_{\gamma}(G)\subset\subset L^{2}(G)$, and take a bounded sequence $\\{f_{m}\\}$, $f_{m}\in L^{2}(G^{\wedge})$ for each $m\in\mathbb{N}$. Thus $T(f_{m})\in L^{2}(G^{\wedge})$, and defining $g_{m}:=T(f_{m})^{\vee}$, we obtain by Plancherel Theorem that $g_{m}\in L^{2}(G)$ for each $m\in\mathbb{N}$. Moreover, from equation (3.70), we have for any $m\in\mathbb{N}$ $\displaystyle\infty>\int_{G^{\wedge}}|f_{m}(\xi)|^{2}d\nu(\xi)$ $\displaystyle=\int_{G^{\wedge}}(1+\gamma(\xi)^{2})\,|[T(f_{m})](\xi)|^{2}d\nu(\xi)$ $\displaystyle=\int_{G^{\wedge}}(1+\gamma(\xi)^{2})\,|\widehat{g_{m}(\xi)}|^{2}d\nu(\xi).$ Therefore, the sequence $\\{g_{m}\\}$ is uniformly bounded in $H^{1}_{\gamma}(G)$, with respect to $m\in\mathbb{N}$. By hypothesis there exists a subsequence of $\\{g_{m}\\}$, say $\\{g_{m_{j}}\\}$, and a function $g\in L^{2}(G)$ such that, $g_{m_{j}}$ converges strongly to $g$ in $L^{2}(G)$ as $j\to\infty$. Consequently, we have $T(f_{m_{j}})=\widehat{g_{m_{j}}}\to\widehat{g}\quad\text{in $L^{2}(G^{\wedge})$ as $j\to\infty$},$ that is, the operator $T$ is compact. 2\. Now, let us assume that the operator $T$ is compact and then show that $H^{1}_{\gamma}(G)\subset\subset L^{2}(G)$. To this end, we take a sequence $\\{f_{m}\\}_{m\in\mathbb{N}}$ uniformly bounded in $H^{1}_{\gamma}(G)$. Then, due to the equivalence (3.71) there exists for each $m\in\mathbb{N}$, $g_{m}\in L^{2}(G^{\wedge})$, such that $\widehat{f_{m}}=T(g_{m})$. Thus for any $m\in\mathbb{N}$, we have from equation (3.70) that $\displaystyle\int_{G^{\wedge}}|g_{m}(\xi)|^{2}d\nu(\xi)$ $\displaystyle=\int_{G^{\wedge}}(1+\gamma(\xi)^{2})\,|[T(g_{m})](\xi)|^{2}d\nu(\xi)$ $\displaystyle=\int_{G^{\wedge}}(1+\gamma(\xi)^{2})\,|\widehat{f_{m}(\xi)}|^{2}d\nu(\xi)<\infty.$ Then, the sequence $\\{g_{m}\\}$ is uniformly bounded in $L^{2}(G)$. Since the operator $T$ is compact, there exist $\\{m_{j}\\}_{j\in\mathbb{N}}$ and $g\in L^{2}(G^{\wedge})$, such that $\widehat{f_{m_{j}}}=T(g_{m_{j}})\xrightarrow[j\to\infty]{}g\quad\text{in $L^{2}(G^{\wedge})$}.$ Consequently, the subsequence $\\{f_{m_{j}}\\}$ converges to $g^{\vee}$ strongly in $L^{2}(G)$, and thus $H^{1}_{\gamma}(G)$ is compactly embedded in $L^{2}(G)$. ∎ ###### Remark 3.14. Due to Theorem 3.13 the compactness characterization, that is $H^{1}_{\gamma}(G)\subset\subset L^{2}(G)$, follows once we show the conditions that the operator $T$ is compact. The study of the dual space of $G$, i.e. $G^{\wedge}$, and $\gamma$ it will be essential for this characterization. Recall from (3.58) item $(ii)$ that, $G^{\wedge}$ is discrete since $G$ is compact. Then, $\nu$ is a countermeasure, and $\nu(\\{\chi\\})=1$ for each singleton $\\{\chi\\}$, $\chi\in G^{\wedge}$. Now, for any $\chi\in G^{\wedge}$ fixed, we define the point mass function at $\chi$ by $\delta_{\chi}(\xi):=1_{\\{\chi\\}}(\xi),\quad\text{for each $\xi\in G^{\wedge}$}.$ Hence the set $\\{\delta_{\xi}\;;\;\xi\in G^{\wedge}\\}$ is an orthonornal basis for $L^{2}(G^{\wedge})$. Indeed, we first show the orthonormality. For each $\chi,\pi\in G^{\wedge}$, we have $\langle\delta_{\chi},\delta_{\pi}\rangle_{L^{2}(G^{\wedge})}=\int_{G^{\wedge}}\delta_{\chi}(\xi)\;\delta_{\pi}(\xi)\,d\nu(\xi)=\left\\{\begin{array}[]{ccl}1,&\text{if}&\chi=\pi,\\\ 0,&\text{if}&\chi\not=\pi.\end{array}\right.$ (3.72) Now, let us show the density, that is $\overline{\\{\delta_{\xi}\;;\;\xi\in G^{\wedge}\\}}=L^{2}(G^{\wedge})$, or equivalently $\\{\delta_{\xi}\;;\;\xi\in G^{\wedge}\\}^{\perp}=\\{0\\}$. For any $w\in\\{\delta_{\xi}\;;\;\xi\in G^{\wedge}\\}^{\perp}$, we obtain $0=\langle\delta_{\xi},w\rangle_{L^{2}(G^{\wedge})}=\int_{G^{\wedge}}\delta_{\xi}(\chi)w(\chi)\,d\nu(\chi)=\int_{\\{\xi\\}}w(\chi)\,d\nu(\chi)=w(\xi)$ for any $\xi\in G^{\wedge}$, which proves the density. From the above discussion, it is important to study the operator $T$ on elements of the set $\\{\delta_{\xi}\;;\;\xi\in G^{\wedge}\\}$. Then, we have the following ###### Theorem 3.15. If the operator $T$ defined by (3.69) is compact, then $G^{\wedge}$ is an enumerable set. ###### Proof. 1\. First, let $\\{\delta_{\xi}\;;\;\xi\in G^{\wedge}\\}$ be the orthonormal basis for $L^{2}(G^{\wedge})$, and $T$ the operator defined by (3.69). Then, the function $\delta_{\xi}\in L^{2}(G^{\wedge})$ is an eigenfunction of $T$ corresponding to the eigenvalue $(1+\gamma^{2})^{-1/2}$, that is $\delta_{\xi}\neq 0$, and $T(\delta_{\xi})=\frac{\delta_{\xi}}{\sqrt{1+\gamma(\xi)^{2}}}.$ (3.73) 2\. Now, since $T$ is compact and $\\{\delta_{\xi}\;;\;\xi\in G^{\wedge}\\}$ is a basis for $L^{2}(G^{\wedge})$, it must be enumerable from (3.73). On the other hand, the function $\xi\in G^{\wedge}\mapsto\delta_{\xi}\in L^{2}(G^{\wedge})$ is injective, hence $G^{\wedge}$ is enumerable. ∎ ###### Corollary 3.16. If the operator $T$ defined by (3.69) is compact, then $L^{2}(G)$ is separable. ###### Proof. First, the Hilbert space $L^{2}(G^{\wedge})$ is separable, since $\\{\delta_{\xi}\;;\;\xi\in G^{\wedge}\\}$ is an enumerable orthonormal basis of it. Then, the proof follows applying the Plancherel Theorem. ∎ ###### Corollary 3.17. Let $G_{B}$ be the Bohr compactification of $\mathbb{R}^{n}$ (see A. Pankov [30]). Then $H^{1}_{\gamma}(G_{B})$ is not compactly embedded in $L^{2}(G_{B})$. ###### Proof. Indeed, $G_{B}^{\wedge}$ is non enumerable. ∎ Consequently, $G^{\wedge}$ be enumerable is a necessarily condition for the operator $T$ be compact, which is not sufficient as shown by the Example 3.20 below. Indeed, it might depend on the chosen $\gamma$, see also Example 3.23. To follow, we first recall the ###### Definition 3.18. Let $G$ be a group (not necessarily a topological one) and $S$ a subset of it. The smallest subgroup of G containing every element of S, denoted $\langle S\rangle$, is called the subgroup generated by $S$. Equivalently, see Dummit, Foote [17] p.63, $\langle S\rangle=\big{\\{}g^{\varepsilon_{1}}_{1}g^{\varepsilon_{2}}_{2}\ldots g^{\varepsilon_{k}}_{k}/\text{$k\in\mathbb{N}$ and for each $j$, $g_{j}\in S,\varepsilon_{j}=\pm 1$}\big{\\}}.$ Moreover, if a group $G=\langle S\rangle$, then $S$ is called a generator of $G$, and in this case when S is finite, $G$ is called finitely generated. ###### Theorem 3.19. If the operator $T$ defined by (3.69) is compact and there exists a generator of $G^{\wedge}$ such that $\gamma$ is bounded on it, then $G^{\wedge}$ is finite generated. ###### Proof. Let $S_{0}$ be a generator of $G^{\wedge}$, such that $\gamma$ is bounded on it. Therefore, there exists $d_{0}\geq 0$ such that, $\text{for each $\xi\in S_{0}$, $\;\gamma(\xi)\leq d_{0}$}.$ Now, since $T$ is compact and $\|T\|\leq 1$, there exists $0<c\leq 1$ such that, the set of eigenvectors $\Big{\\{}\delta_{\xi}\;;\;\xi\in G^{\wedge}\;\;\text{and}\;\;\frac{1}{\sqrt{1+\gamma(\xi)^{2}}}\geq c\Big{\\}}\equiv\Big{\\{}\delta_{\xi}\;;\;\xi\in G^{\wedge}\;\;\text{and}\;\;\gamma(\xi)\leq\sqrt{\frac{1}{c^{2}}-1}\Big{\\}}$ is finite, where we have used the Spectral Theorem for bounded compact operators. Therefore, since $\left\\{\delta_{\xi}\;;\;\xi\in S_{0}\right\\}\subset\left\\{\delta_{\xi}\;;\;\xi\in G^{\wedge}\;\;\text{and}\;\;\gamma(\xi)\leq d_{0}\right\\}$ it follows that $S_{0}$ is a finite set, and thus $G^{\wedge}$ is finite generated. ∎ ###### Example 3.20 (Infinite enumerable Torus). Let us recall the Sobolev space $H^{1}_{\gamma}(\mathbb{T}^{\mathbb{N}})$, where $\mathbb{T}^{\mathbb{N}}$ is the infinite enumerable Torus. We claim that: $H^{1}_{\gamma}(\mathbb{T}^{\mathbb{N}})$ is not compactly embedded in $L^{2}(\mathbb{T}^{\mathbb{N}})$, for $\gamma$ defined in Exemple 3.6. Indeed, given $k\in\mathbb{N}$ we define $1_{k}\in\mathbb{Z}^{\mathbb{N}}$, such that it is zero for any coordinate $\ell\neq k$, and one in $k-$coordinate. Therefore, the set $S_{0}:=\\{\xi_{1_{k}}\;;\;k\in\mathbb{N}\\}$ is an infinite generator of the dual group $(\mathbb{T}^{\mathbb{N}})^{\wedge}$. Since for each $k\in\mathbb{N}$, $\gamma(\xi_{1_{k}})=1$, i.e. bounded in $S_{0}$, applying Theorem 3.19 it follows that $H^{1}_{\gamma}(\mathbb{T}^{\mathbb{N}})$ is not compactly embedded in $L^{2}(\mathbb{T}^{\mathbb{N}})$. ###### Remark 3.21. The above discussion in the Example 3.20 follows as well to the Sobolev space $H^{1}_{\gamma}(\mathbb{T}^{I})$, where $I$ is an index set (enumerable or not). Clearly, the Sobolev space $H^{1}_{\gamma}(\mathbb{T}^{I})$ in not compactly embedded in $L^{2}(\mathbb{T}^{I})$, when $I$ is a non enumerable index set. Indeed, the set $(\mathbb{T}^{I})^{\wedge}$ is non enumerable. Now, we charactherize the condition on $\gamma:G^{\wedge}\to[0,\infty)$, in order to $T$ be compact. More precisely, let us consider the following property: ${\bf C}.\quad\text{For each $d>0$, the set $\left\\{\xi\in G^{\wedge}\;;\;\gamma(\xi)\leq d\right\\}$ is finite}.$ (3.74) ###### Theorem 3.22. If $\gamma:G^{\wedge}\to[0,\infty)$ satisfies ${\bf C}$, then the operator $T$ defined by (3.69) is compact. ###### Proof. By hypothesis, $\\{\xi\in G^{\wedge}\;;\;\gamma(\xi)\leq d\\}$ is finite, then we have $G^{\wedge}=\bigcup_{k\in\mathbb{N}}\left\\{\xi\in G^{\wedge}\;;\;\gamma(\xi)\leq k\right\\}.$ Consequently, the set $G^{\wedge}$ is enumerable and we may write $G^{\wedge}=\\{\xi_{i}\\}_{i\in\mathbb{N}}$. Again, due to condition ${\bf C}$ for each $c\in(0,1)$ the set $\Big{\\{}\xi\in G^{\wedge}\;;\;\frac{1}{\sqrt{1+\gamma(\xi)^{2}}}\geq c\Big{\\}}$ (3.75) is finite. Since the function $\xi\in G^{\wedge}\mapsto\delta_{\xi}\in L^{2}(G^{\wedge})$ is injective, the set $\\{\delta_{\xi_{i}}\;;\;i\in\mathbb{N}\\}$ is an enumerable orthonormal basis of eigenvectors for $T$, which corresponding eigenvalues satisfy $\lim_{i\to\infty}\frac{1}{\sqrt{1+\gamma(\xi_{i})^{2}}}=0,$ where we have used (3.75). Consequently, $T$ is a compact operator. ∎ ###### Example 3.23 (Bis: Infinite enumerable Torus). There exists a function $\gamma_{0}$ such that, $H^{1}_{\gamma_{0}}(\mathbb{T}^{\mathbb{N}})$ is compactly embedded in $L^{2}(\mathbb{T}^{\mathbb{N}})$. Indeed, we are going to show that, $\gamma_{0}$ satisfies ${\bf C}$. Let $\alpha\equiv(\alpha_{\ell})_{\ell\in\mathbb{N}}$ be a sequence in $\mathbb{R}^{\mathbb{N}}$, such that for each $\ell\in\mathbb{N}$, $\alpha_{\ell}\geq 0$ and $\lim_{\ell\to\infty}\alpha_{\ell}=+\infty.$ (3.76) Then, we define the following pseudo-metric in the dual group $(\mathbb{T}^{\mathbb{N}})^{\wedge}$ as follows $p_{0}(\xi_{m},\xi_{n}):=2\pi\sum_{\ell=1}^{\infty}\alpha_{\ell}\;{|m_{\ell}-n_{\ell}|},\quad(m,n\in\mathbb{Z}^{\mathbb{N}}_{\rm c}),$ and consider $\gamma_{0}(\xi_{m})=p_{0}(\xi_{m},1)$. Thus for each $d>0$, the set $\\{m\in\mathbb{Z}^{\mathbb{N}}_{\rm c}\;;\;\gamma_{0}(\xi_{m})\leq d\\}\quad\text{is finite.}$ Indeed, from (3.76) there exists $\ell_{0}\in\mathbb{N}$, such that $\alpha_{\ell}>d$, for each $\ell\geq\ell_{0}$. Therefore, if $m\in\mathbb{Z}^{\mathbb{N}}_{\rm c}$ and the support of $m$ is not contained in $\\{1,\ldots,\ell_{0}-1\\}$, that is to say, there exists $\tilde{\ell}\geq\ell_{0}$, such that, $m_{\tilde{\ell}}\neq 0$. Then, $2\pi\sum_{\ell=1}^{\infty}\alpha_{\ell}\;{|m_{\ell}|}\geq\alpha_{\tilde{\ell}}>d.$ Consequently, we have $\\{m\in\mathbb{Z}^{\mathbb{N}}_{\rm c}\;;\;\gamma_{0}(\xi_{m})\leq d\\}\subset\\{m\in\mathbb{Z}^{\mathbb{N}}_{\rm c}\;;\;{\rm supp}\ m\subset\\{1,\ldots,\ell_{0}-1\\}\\},$ which is a finite set. Finally, applying Theorem 3.22 we obtain that, the Sobolev space $H^{1}_{\gamma_{0}}(\mathbb{T}^{\mathbb{N}})$ is compactly embedded in $L^{2}(\mathbb{T}^{\mathbb{N}})$. #### 3.2.3 On a class of Quasi-periodic functions In this section we consider an important class of quasi-periodic functions, which includes for instance the periodic functions. Let $\lambda_{1},\lambda_{2},\ldots,\lambda_{m}\in\mathbb{R}^{n}$ be $m-$linear independent vectors with respect to $\mathbb{Z}$, and consider the following matrix $\Lambda:={\left(\begin{array}[]{c}\lambda_{1}\\\ \lambda_{2}\\\ \vdots\\\ \lambda_{m}\end{array}\right)}_{m\times n}$ such that, for each $d>0$ the set $\\{k\in\mathbb{Z}^{m}\;;\;{|\Lambda^{T}k|}\leqslant d\\}\quad\text{is finite.}$ (3.77) Therefore, we are considering the class of quasi-periodic functions satisfying condition (3.77). This set is not empty, for instance let us define the matrix $B:=\Lambda\Lambda^{T}$, such that $\det B>0$, which is called here positive quasi-periodic functions. It is not difficult to see that, positive quasi- periodic functions satisfies (3.77). Indeed, it is sufficiently to observe that, for each $k\in\mathbb{Z}^{m}$, we have $|k|=|B^{-1}Bk|\leq\|B^{-1}\|\|\Lambda\||\Lambda^{T}k|.$ Moreover, since $\lambda_{1},\lambda_{2},\ldots,\lambda_{m}\in\mathbb{R}^{n}$ are $m-$linear independent vectors with respect to $\mathbb{Z}$, (this property does not imply $\det B>0$), the dynamical system $\tau:\mathbb{R}^{n}\times\mathbb{T}^{m}\to\mathbb{T}^{m}$, given by $\tau(x)\omega:=\omega+\Lambda x-\left\lfloor\omega+\Lambda x\right\rfloor$ (3.78) is a ergodic. Now we remark that, the application ${\varphi:\mathbb{R}^{n}\to\mathbb{T}^{m}}$, $\varphi(x):=\Lambda x-\left\lfloor\Lambda x\right\rfloor$, is a continuous homeomorphism of groups. Then, we have $\tau(x)\omega=\varphi(x)\omega\equiv\omega+\Lambda x-\left\lfloor\omega+\Lambda x\right\rfloor.$ Consequently, under the conditions of the previous sections, we obtain for each $k\in\mathbb{Z}^{m}$ $\gamma(\xi_{k})=2\pi{|\Lambda^{T}k|},$ and applying Theorem 3.22 (recall (3.77)), it follows that $H^{1}_{\gamma}{\left(\mathbb{T}^{m}\right)}\subset\\!\subset L^{2}{\left(\mathbb{T}^{m}\right)}.$ Therefore, given a stochastic deformation $\Phi$, we have $\mathcal{H}_{\Phi}\subset\\!\subset\mathcal{L}_{\Phi}$ for the class of quasi-periodic functions satisfying (3.77), and it follows a solution to Bloch’s spectral cell equation. ### 3.3 Auxiliary celular equations The preposition below, which is an immediate consequence of Theorem 2.26, give us the necessaries characteristics to deduce from the cell equation (3.56) other equations (called here auxiliary cellular equations), that will be essential in our homogenization analysis. ###### Proposition 3.24. Given ${\theta\in\mathbb{R}^{n}}$, let $\big{(}\lambda(\theta),\Psi(\theta)\big{)}$ be a spectral point of the cell equation (3.56). Suppose that for some $\theta_{0}\in\mathbb{R}^{n}$ the corresponding eigenvalue ${\lambda(\theta_{0})}$ has finite multiplicity. Then, there exists a neighborhood ${{\mathcal{U}}\subset\mathbb{R}^{n}}$ of ${\theta_{0}}$, such that the following functions $\theta\in{\mathcal{U}}\mapsto\Psi(\theta)\in\mathcal{H}_{\Phi}\;\;\;\text{and}\;\;\;\theta\in{\mathcal{U}}\mapsto\lambda(\theta)\in\mathbb{R}-\\{0\\},$ are analytical. Now, introducing the operator ${\mathbb{A}(\theta)}$, (${\theta\in\mathbb{R}^{n}}$), defined on $\mathcal{H}_{\Phi}$ by $\displaystyle\mathbb{A}(\theta)[F]=-({\rm div}_{\\!z}+2i\pi\theta){\Big{(}A{\left(\Phi^{-1}(z,\omega),\omega\right)}{(\nabla_{\\!\\!z}+2i\pi\theta)}F\Big{)}}$ $\displaystyle\qquad\qquad\qquad\qquad+V{\left(\Phi^{-1}(z,\omega),\omega\right)}F-\lambda(\theta)F,$ and writing $\theta=(\theta_{1},\cdots,\theta_{n})$, we obtain for $k=1,\ldots,n$, $\displaystyle\mathbb{A}(\theta){\left[\frac{\partial\Psi(\theta)}{\partial\theta_{k}}\right]}=({\rm div}_{\\!z}+2i\pi\theta){\Big{(}A{\left(\Phi^{-1}(z,\omega),\omega\right)}{(2i\pi e_{k}\Psi(\theta))}\Big{)}}$ $\displaystyle\qquad\qquad+{(2i\pi e_{k})}{\Big{(}A{\left(\Phi^{-1}(z,\omega),\omega\right)}{({\rm div}_{\\!z}+2i\pi\theta)}\Psi(\theta)\Big{)}}$ $\displaystyle\hskip 170.71652pt+\frac{\partial\lambda}{\partial\theta_{k}}(\theta)\Psi(\theta),$ (3.79) where $\\{e_{k}{\\}}_{1\leq k\leq n}$ is the canonical basis of $\mathbb{R}^{n}$. The equation (3.3) is called the first auxiliary cellular equation (or f.a.c. equation, in short). In the same way, we have for $k,\ell=1,\ldots,n$, $\displaystyle\mathbb{A}(\theta){\left[\frac{\partial^{2}\Psi(\theta)}{\partial\theta_{\ell}\,\partial\theta_{k}}\right]}=({\rm div}_{\\!z}+2i\pi\theta){\Big{(}A{\left(\Phi^{-1}(z,\omega),\omega\right)}{2i\pi e_{\ell}\frac{\partial\Psi(\theta)}{\partial\theta_{k}}}\Big{)}}$ $\displaystyle\qquad\qquad+({\rm div}_{\\!z}+2i\pi\theta){\Big{(}A{\left(\Phi^{-1}(z,\omega),\omega\right)}{2i\pi e_{k}\frac{\partial\Psi(\theta)}{\partial\theta_{\ell}}}\Big{)}}$ $\displaystyle\qquad\qquad+{(2i\pi e_{k})}{\Big{(}A{\left(\Phi^{-1}(z,\omega),\omega\right)}{({\rm div}_{\\!z}+2i\pi\theta)}\frac{\partial\Psi(\theta)}{\partial\theta_{\ell}}\Big{)}}$ $\displaystyle\qquad\qquad+{(2i\pi e_{\ell})}{\Big{(}A{\left(\Phi^{-1}(z,\omega),\omega\right)}{({\rm div}_{\\!z}+2i\pi\theta)}\frac{\partial\Psi(\theta)}{\partial\theta_{k}}\Big{)}}$ $\displaystyle\qquad\qquad\qquad+{(2i\pi e_{k})}{\Big{(}A{\left(\Phi^{-1}(z,\omega),\omega\right)}{\left(2i\pi e_{\ell}\Psi(\theta)\right)}\Big{)}}$ (3.80) $\displaystyle\qquad\qquad\qquad\qquad+{(2i\pi e_{\ell})}{\Big{(}A{\left(\Phi^{-1}(z,\omega),\omega\right)}{\left(2i\pi e_{k}\Psi(\theta)\right)}\Big{)}}$ $\displaystyle\qquad\qquad\qquad\qquad+\frac{\partial\lambda(\theta)}{\partial\theta_{k}}\frac{\partial\Psi(\theta)}{\partial\theta_{\ell}}+\,\frac{\partial\lambda(\theta)}{\partial\theta_{\ell}}\frac{\partial\Psi(\theta)}{\partial\theta_{k}}+\frac{\partial^{2}\lambda(\theta)}{\partial\theta_{\ell}\,\partial\theta_{k}}\Psi(\theta),$ which we call the second auxiliary cellular equation (or s.a.c. equation, in short). In order to make clear in which sense the auxiliary cellular equations are understood, we note that if ${G\in\mathcal{H}_{\Phi}}$ then the variational formulation of the f.a.c. equation (3.3) is given by $\displaystyle\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}\Big{\\{}A{\left(\Phi^{-1}(z,\omega),\omega\right)}{\left(\nabla_{\\!\\!z}+2i\pi\theta\right)}\frac{\partial\Psi(\theta)}{\partial\theta_{k}}\cdot\overline{{\left(\nabla_{\\!\\!z}+2i\pi\theta\right)}G}$ $\displaystyle\qquad\qquad+V{\left(\Phi^{-1}(z,\omega),\omega\right)}\frac{\partial\Psi(\theta)}{\partial\theta_{k}}\,\overline{G}-\lambda(\theta)\frac{\partial\Psi(\theta)}{\partial\theta_{k}}\,\overline{G}\Big{\\}}\,dz\,d\mathbb{P}(\omega)$ $\displaystyle\qquad\qquad\qquad=:\Big{\langle}\mathbb{A}(\theta){\left[\frac{\partial\Psi(\theta)}{\partial\theta_{k}}\right]},G\Big{\rangle}$ (3.81) $\displaystyle\qquad=-\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}\Big{\\{}A{\left(\Phi^{-1}(z,\omega),\omega\right)}{\left(\nabla_{\\!\\!z}+2i\pi\theta\right)}\Psi(\theta)\cdot\overline{{(2i\pi e_{k}G)}}$ $\displaystyle\qquad\qquad\qquad\qquad-A{\left(\Phi^{-1}(z,\omega),\omega\right)}{(2i\pi e_{k}\Psi(\theta))}\cdot\overline{{\left(\nabla_{\\!\\!z}+2i\pi\theta\right)}G}$ $\displaystyle\qquad\qquad\qquad\qquad\qquad\qquad+\frac{\partial\lambda(\theta)}{\partial\theta_{k}}\,\Psi(\theta)\,\overline{G}\Big{\\}}\,dz\,d\mathbb{P}(\omega).$ Similar reasoning can be made with the s.a.c. equation (3.3). In the following, we highlight an important fact that is fundamental to determine the hessian nature of the effective tensor in our homogenization analysis concerning the Schrödinger equation (1.1). This fact is brought out by choosing ${\theta\in{\mathcal{U}}}$, ${k\in\\{1,\ldots,n\\}}$ and defining $\Lambda_{k}(z,\omega,\theta):=\frac{1}{2i\pi}\frac{\partial\Psi}{\partial\theta_{k}}(z,\omega,\theta)$. Hence, taking ${\Psi(\theta)}$ as a test function in the s.a.c. equation (3.3), we get $\displaystyle\frac{1}{4\pi^{2}}\frac{\partial^{2}\lambda(\theta)}{\partial\theta_{\ell}\,\partial\theta_{k}}\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}{|\Psi(\theta)|}^{2}\,dz\,d\mathbb{P}(\omega)$ $\displaystyle\qquad=\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}\Big{\\{}-A{\left(\Phi^{-1}(z,\omega),\omega\right)}{(e_{\ell}\Lambda_{k}(\theta))}\cdot\overline{{\left(\nabla_{\\!\\!z}+2i\pi\theta\right)}\Psi(\theta)}$ $\displaystyle\qquad\qquad\qquad\,-A{\left(\Phi^{-1}(z,\omega),\omega\right)}{(e_{k}\Lambda_{\ell}(\theta))}\cdot\overline{{\left(\nabla_{\\!\\!z}+2i\pi\theta\right)}\Psi(\theta)}$ $\displaystyle\qquad\qquad\qquad\,+A{\left(\Phi^{-1}(z,\omega),\omega\right)}{\left(\nabla_{\\!\\!z}+2i\pi\theta\right)}\Lambda_{k}(\theta)\cdot\overline{{(e_{\ell}\,\Psi(\theta))}}$ $\displaystyle\qquad\qquad\qquad\,+A{\left(\Phi^{-1}(z,\omega),\omega\right)}{\left(\nabla_{\\!\\!z}+2i\pi\theta\right)}\Lambda_{\ell}(\theta)\cdot\overline{{(e_{k}\,\Psi(\theta))}}$ $\displaystyle\qquad\qquad\qquad\qquad\,+A{\left(\Phi^{-1}(z,\omega),\omega\right)}{(e_{k}\Psi(\theta))}\cdot\overline{{(e_{\ell}\,\Psi(\theta))}}$ (3.82) $\displaystyle\qquad\qquad\qquad\qquad\,+A{\left(\Phi^{-1}(z,\omega),\omega\right)}{(e_{\ell}\Psi(\theta))}\cdot\overline{{(e_{k}\,\Psi(\theta))}}$ $\displaystyle\qquad\qquad\qquad\qquad+\frac{1}{2i\pi}\Big{(}\frac{\partial\lambda(\theta)}{\partial\theta_{\ell}}\Lambda_{k}(\theta)+\frac{\partial\lambda(\theta)}{\partial\theta_{k}}\Lambda_{\ell}(\theta)\Big{)}\,\overline{\Psi(\theta)}\Big{\\}}\,dz\,d\mathbb{P}(\omega).$ On the other hand, using $\Lambda_{k}(z,\omega,\theta)$ as a test function in the f.a.c. equation (3.3) and due to Theorem 2.16, we arrive at $\begin{array}[]{l}\displaystyle\int_{\mathbb{R}^{n}}A{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\omega\right)},\omega\right)}{\left(\nabla+2i\pi\frac{\theta}{\varepsilon}\right)}\Lambda_{k,\varepsilon}(\theta)\cdot\overline{{\left(\nabla+2i\pi\frac{\theta}{\varepsilon}\right)}\varphi}\,dx\\\\[12.0pt] \displaystyle+\frac{1}{\varepsilon^{2}}\int_{\mathbb{R}^{n}}V{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\omega\right)},\omega\right)}\,\Lambda_{k,\varepsilon}(\theta)\,\overline{\varphi}\,dx-\frac{\lambda(\theta)}{\varepsilon^{2}}\int_{\mathbb{R}^{n}}\Lambda_{k,\varepsilon}(\theta)\,\overline{\varphi}\,dx\\\\[12.0pt] \hskip 56.9055pt\displaystyle=-\frac{1}{\varepsilon}\int_{\mathbb{R}^{n}}A{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\omega\right)},\omega\right)}{\left(\nabla+2i\pi\frac{\theta}{\varepsilon}\right)}\Psi_{\varepsilon}(\theta)\cdot\overline{{(e_{k}\varphi)}}\,dx\\\\[12.0pt] \hskip 71.13188pt\displaystyle-\frac{1}{\varepsilon}\int_{\mathbb{R}^{n}}A{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\omega\right)},\omega\right)}{(e_{k}\Psi_{\varepsilon}(\theta))}\cdot\overline{{\left(\nabla+2i\pi\frac{\theta}{\varepsilon}\right)}\varphi}\,dx\\\\[12.0pt] \hskip 78.24507pt\displaystyle+\frac{1}{\varepsilon^{2}}\frac{1}{2i\pi}\frac{\partial\lambda}{\partial\theta_{k}}(\theta)\int_{\mathbb{R}^{n}}\Psi_{\varepsilon}(\theta)\,\overline{\varphi}\,dx,\end{array}$ (3.83) for any ${\varphi\in C^{\infty}_{\rm c}(\mathbb{R}^{n})}$ and a.e ${\omega\in\Omega}$. Here, $\Lambda_{k,\varepsilon}(x,\omega,\theta):=\Lambda_{k}{\left(\frac{x}{\varepsilon},\omega,\theta\right)}$. Proceeding in a similar way with the cell equation (3.56), we can find $\int_{\mathbb{R}^{n}}A{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\omega\right)},\omega\right)}{\left(\nabla+2i\pi\frac{\theta}{\varepsilon}\right)}\Psi_{\varepsilon}(\theta)\cdot\overline{{\left(\nabla+2i\pi\frac{\theta}{\varepsilon}\right)}\varphi}\,dx\\\ +\frac{1}{\varepsilon^{2}}\int_{\mathbb{R}^{n}}V{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\omega\right)},\omega\right)}\,\Psi_{\varepsilon}(\theta)\,\overline{\varphi}\,dx-\frac{\lambda(\theta)}{\varepsilon^{2}}\int_{\mathbb{R}^{n}}\Psi_{\varepsilon}(\theta)\,\overline{\varphi}\,dx=0,$ (3.84) for any ${\varphi\in C^{\infty}_{\rm c}(\mathbb{R}^{n})}$ and a.e. ${\omega\in\Omega}$. ## Part II: Asymptotic Equations ## 4 ​​ On Schrödinger Equations Homogenization In this section, we shall describe the asymptotic behaviour of the family of solutions $\\{u_{\varepsilon}{\\}_{\varepsilon>0}}$ of the equation (1.1), this is the content of Theorem 4.2 below. It generalizes the similar result of Allaire, Piatnitski [4] where they consider the similar problem in the periodic setting. Our scenario is much different from one considered by them. Here, the coefficients of equation (1.1) are random perturbations accomplished by stochastic diffeomorphisms of stationary functions. Since the two-scale convergence technique is the best tool to deal with asymptotic analysis of linear operators, we make use of it in analogous way done in [4]. Although, the presence of the stochastic deformation in the coefficients brings out several complications, which we were able to overcome. To begin, some important and basic a priori estimates of the solution of the Schrödinger equation (1.1) are needed. Then, we have the following ###### Lemma 4.1 (Energy Estimates). Assume that the conditions (1.2), (1.3) hold and let ${u_{\varepsilon}}\in C\big{(}[0,T);H^{1}(\mathbb{R}^{n})\big{)}$ be the solution of the equation (1.1) with initial data $u_{\varepsilon}^{0}$. Then, for all ${t\in[0,T]}$ and a.e. ${\omega\in\Omega}$, the following a priori estimates hold: * (i) $($Energy Conservation$.)$ ${\displaystyle\int_{\mathbb{R}^{n}}{|u_{\varepsilon}(t,x,\omega)|}^{2}dx=\int_{\mathbb{R}^{n}}{|u_{\varepsilon}^{0}(x,\omega)|}^{2}dx}$. * (ii) $(\varepsilon\nabla-$ Estimate$.)$ $\displaystyle\int_{\mathbb{R}^{n}}|\varepsilon\nabla u_{\varepsilon}(t,x,\omega)|^{2}\,dx\leq C\int_{\mathbb{R}^{n}}\Big{\\{}|\varepsilon\nabla u_{\varepsilon}^{0}(x,\omega)|^{2}+|u_{\varepsilon}^{0}(x,\omega)|^{2}\Big{\\}}\,dx,$ where ${C:=C\big{(}\Lambda,{\|A\|}_{\infty},{\|V\|}_{\infty},{\|U\|}_{\infty}}\big{)}$ is a positive constant which does not depend on $\varepsilon>0$. ###### Proof. 1\. If we multiply the Eq. (1.1) by $\overline{u_{\varepsilon}}$ and take the imaginary part, then we obtain $\frac{d}{dt}\int_{\mathbb{R}^{n}}|u_{\varepsilon}(t,x,\omega)|^{2}\,dx=0,$ which gives the proof of the item $(i)$. 2\. Now, multiplying the Eq. (1.1) by $\overline{\partial_{t}u_{\varepsilon}}$ and taking the real part, we get $\displaystyle\frac{1}{2}\frac{d}{dt}\int_{\mathbb{R}^{n}}\Big{\\{}\varepsilon^{2}A\left(\Phi^{-1}\left(\frac{x}{\varepsilon},\omega\right),\omega\right)\nabla u_{\varepsilon}\cdot\nabla\overline{u_{\varepsilon}}$ $\displaystyle\qquad\qquad+\Big{(}V\left(\Phi^{-1}\left(\frac{x}{\varepsilon},\omega\right),\omega\right)+\varepsilon^{2}U\left(\Phi^{-1}\left(\frac{x}{\varepsilon},\omega\right),\omega\right)\Big{)}|u_{\varepsilon}|^{2}\Big{\\}}\,dx=0,$ which provides the proof of the item $(ii)$. ∎ Its important to remember the followings facts that will be necessary in this section: * • The initial data of the equation (1.1) is assumed to be well-prepared, that is, for $(x,\omega)\in\mathbb{R}^{n}\\!\times\\!\Omega$, and ${\theta^{\ast}\in\mathbb{R}^{n}}$ $u_{\varepsilon}^{0}(x,\omega)=e^{2i\pi\frac{\theta^{\ast}\cdot x}{\varepsilon}}\psi\left(\Phi^{-1}(x/\varepsilon,\omega),\omega,\theta^{\ast}\right)v^{0}(x),$ (4.85) where ${v^{0}\in C_{\rm c}^{\infty}(\mathbb{R}^{n})}$, and ${\psi(\theta^{\ast})}$ is an eigenfunction of the cell problem (3.56). * • Using the Ergodic Theorem, it is easily seen that the sequences ${\\{u_{\varepsilon}^{0}(\cdot,\omega)\\}_{\varepsilon>0}}\quad\text{and}\quad{\\{\varepsilon\nabla u_{\varepsilon}^{0}(\cdot,\omega)\\}_{\varepsilon>0}}$ are bounded in ${L^{2}(\mathbb{R}^{n})}$ and ${[L^{2}(\mathbb{R}^{n})]^{n}}$, respectively. One observes that, the main importance of the well preparedness of the initial data is the following: Trivially, the sequence of solutions $\\{u_{\varepsilon}{\\}_{\varepsilon>0}}$ of the equation (1.1) two-scale converges to zero. However, if our initial data is well-prepared, we are able to correct the oscillations present in $u_{\varepsilon}$ in such a way that, after this correction we can strengthen the weak convergence to the solution of a nontrivial homogenized Schrödinger Equation. For instance, we invite the readers to Allaire, Piatnistski [4], Bensoussan, Lions, Papanicolaou [8, Chapter 4] and Poupaud, Ringhofer [31]. ### 4.1 The Abstract Theorem. In the next, we establish an abstract homogenization theorem for Schrödinger equations. ###### Theorem 4.2. Let $\Phi(y,\omega)$ be a stochastic deformation, and $\tau:\mathbb{Z}^{n}\times\Omega\to\Omega$ an ergodic $n-$dimensional dynamical system. Assume that the conditions (1.2), (1.3) hold, and there exists a Bloch frequence ${\theta^{\ast}\\!\in\mathbb{R}^{n}}$ which is a critical point of ­$\lambda(\cdot)$, that is ${\nabla_{\\!\\!\theta}\,\lambda(\theta^{\ast})=0}$, where ${\lambda(\theta^{\ast})}$ is a simple eigenvalue of the spectral cell equation (3.56) associated to the eigenfunction $\Psi(z,\omega,\theta^{\ast})\equiv\psi{\left(\Phi^{-1}(z,\omega),\omega,\theta^{\ast}\right)}$. Assume also that, the initial data is well-prepared in the sense of (4.85). If ${u_{\varepsilon}}\in C\big{(}[0,T);H^{1}(\mathbb{R}^{n})\big{)}$ is the solution of (1.1) for each $\varepsilon>0$ fixed, then the sequence ${v_{\varepsilon}}$ defined by $v_{\varepsilon}(t,x,\omega):=e^{-{\left(i\frac{\lambda(\theta^{\ast})t}{\varepsilon^{2}}+2i\pi\frac{\theta^{\ast}\\!\cdot x}{\varepsilon}\right)}}u_{\varepsilon}(t,x,\omega),\;\,(t,x)\in\mathbb{R}^{n+1}_{T},\;\omega\in\Omega,$ $\Phi_{\omega}-$two-scale converges to ${v(t,x)\,\psi{\left(\Phi^{-1}(z,\omega),\omega,\theta^{\ast}\right)}}$, and satisfies for a.e. ${\omega\in\Omega}$ $\lim_{\varepsilon\to 0}\iint_{\mathbb{R}^{n+1}_{T}}\\!{\left|v_{\varepsilon}(t,x,\omega)-v(t,x)\,\psi{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\omega\right)},\omega,\theta^{\ast}\right)}\right|}^{2}dx\,dt\,=\,0,$ where the function ${v\in C\big{(}[0,T);L^{2}(\mathbb{R}^{n})\big{)}}$ is the unique solution of the homogenized Schrödinger equation $\left\\{\begin{aligned} &i\displaystyle\frac{\partial v}{\partial t}-{\rm div}{\left(A^{\ast}\nabla v\right)}+U^{\ast}v=0\,,\;\,\text{in}\;\,\mathbb{R}^{n+1}_{T},\\\\[5.0pt] &v(0,x)=v^{0}(x)\,,\;\,x\in\mathbb{R}^{n},\end{aligned}\right.$ (4.86) with effective (constant) coefficients: matrix ${A^{\ast}=D_{\theta}^{2}\lambda(\theta^{\ast})}$, and potential $U^{\ast}=c^{-1}_{\psi}\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}U{\left(\Phi^{-1}(z,\omega),\omega\right)}{\left|\psi{\left(\Phi^{-1}(z,\omega),\omega,\theta^{\ast}\right)}\right|}^{2}dz\,d\mathbb{P}(\omega),$ where $c_{\psi}=\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}{\left|\psi{\left(\Phi^{-1}(z,\omega),\omega,\theta^{\ast}\right)}\right|}^{2}dz\,d\mathbb{P}(\omega).$ ###### Proof. In order to better understand the main difficulties brought by the presence of the stochastic deformation $\Phi$, we split our proof in five steps. 1.(A priori estimates and $\Phi_{\omega}-$two-scale convergence.) First, we define $v_{\varepsilon}(t,x,\widetilde{\omega}):=e^{-{\left(i\frac{\lambda(\theta^{\ast})t}{\varepsilon^{2}}+2i\pi\frac{\theta^{\ast}\\!\cdot x}{\varepsilon}\right)}}u_{\varepsilon}(t,x,\widetilde{\omega}),\;\,(t,x,\widetilde{\omega})\in\mathbb{R}^{n+1}_{T}\\!\times\\!\Omega.$ (4.87) Then, computing the first derivatives with respect to the variable $x$, we get $\varepsilon\nabla u_{\varepsilon}(t,x,\widetilde{\omega})\,e^{-{\left(i\frac{\lambda(\theta^{\ast})t}{\varepsilon^{2}}+2i\pi\frac{\theta^{\ast}\\!\cdot x}{\varepsilon}\right)}}=(\varepsilon\nabla+2i\pi\theta^{\ast})v_{\varepsilon}(t,x,\widetilde{\omega}).$ (4.88) Applying Lemma 4.1 yields: * • ${\displaystyle\int_{\mathbb{R}^{n}}{|v_{\varepsilon}(t,x,\widetilde{\omega})|}^{2}dx=\int_{\mathbb{R}^{n}}{|u_{\varepsilon}^{0}(x,\widetilde{\omega})|}^{2}dx},$ * • ${\displaystyle\int_{\mathbb{R}^{n}}{|\varepsilon\nabla v_{\varepsilon}(t,x,\widetilde{\omega})|}^{2}dx\leq\widetilde{C}{\displaystyle\int_{\mathbb{R}^{n}}\Big{(}{|\varepsilon\nabla u_{\varepsilon}^{0}(x,\widetilde{\omega})|}^{2}+{|u_{\varepsilon}^{0}(x,\widetilde{\omega})|}^{2}\Big{)}dx}}$ for all ${t\in[0,T)}$ and a.e. $\widetilde{\omega}\in\Omega$, where the constant ${\widetilde{C}}$ depends on $\|A{\|}_{\infty}$, $\|V{\|}_{\infty}$, $\|U{\|}_{\infty}$ and $\theta^{\ast}$. Then, from the uniform boundedness of the sequences ${\\{u_{\varepsilon}^{0}(\cdot,\widetilde{\omega})\\}_{\varepsilon>0}}$ and ${\\{\varepsilon\nabla u_{\varepsilon}^{0}(\cdot,\widetilde{\omega})\\}_{\varepsilon>0}}$, we deduce that the sequences ${{\\{v_{\varepsilon}(\cdot,\cdot\cdot,\widetilde{\omega})\\}}_{\varepsilon>0}}\quad\text{and}\quad{{\\{\varepsilon\nabla v_{\varepsilon}(\cdot,\cdot\cdot,\widetilde{\omega})\\}}_{\varepsilon>0}}$ are bounded, respectively, in ${L^{2}(\mathbb{R}^{n+1}_{T})}$ and ${{[L^{2}(\mathbb{R}^{n+1}_{T})]}^{n}}$ for a.e. ${\widetilde{\omega}\in\Omega}$. Therefore, applying Lemma 2.22, there exists a subsequence ${\\{\varepsilon^{\prime}\\}}$(which may dependent on $\tilde{\omega}$), and a stationary function ${v^{\ast}_{\widetilde{\omega}}\in L^{2}(\mathbb{R}^{n+1}_{T},\mathcal{H})}$, for a.e. ${\widetilde{\omega}\in\Omega}$, such that $v_{\varepsilon^{\prime}}(t,x,\widetilde{\omega})\;\xrightharpoonup[\varepsilon^{\prime}\to 0]{2-{\rm s}}\;v^{\ast}_{\widetilde{\omega}}{\left(t,x,\Phi^{-1}(z,\omega),\omega\right)},$ and $\varepsilon^{\prime}\frac{\partial v_{\varepsilon^{\prime}}}{\partial x_{k}}(t,x,\widetilde{\omega})\;\xrightharpoonup[\varepsilon^{\prime}\to 0]{2-{\rm s}}\;\frac{\partial}{\partial z_{k}}{\big{(}v^{\ast}_{\widetilde{\omega}}{\left(t,x,\Phi^{-1}{(z,\omega)},\omega\right)}\big{)}},$ which means that, for ${k\in\\{1,\ldots,n\\}}$, we have $\displaystyle\lim_{\varepsilon^{\prime}\to 0}\iint_{\mathbb{R}^{n+1}_{T}}$ $\displaystyle v_{\varepsilon^{\prime}}\left(t,x,\widetilde{\omega}\right)\,\overline{\varphi(t,x)\,\Theta\left(\Phi^{-1}{\left(\frac{x}{\varepsilon^{\prime}},\widetilde{\omega}\right)},\widetilde{\omega}\right)}\,dx\,dt$ (4.89) $\displaystyle=c_{\Phi}^{-1}\\!\\!\iint_{\mathbb{R}^{n+1}_{T}}\\!\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}\\!\\!\\!\\!\\!v^{\ast}_{\widetilde{\omega}}{\left(t,x,\Phi^{-1}{(z,\omega)},\omega\right)}\,$ $\displaystyle\hskip 90.0pt\times\;\overline{\varphi(t,x)\,\Theta\left(\Phi^{-1}(z,\omega),\omega\right)}\,dz\,d\mathbb{P}\,dx\,dt$ and $\displaystyle\lim_{\varepsilon^{\prime}\to 0}\iint_{\mathbb{R}^{n+1}_{T}}$ $\displaystyle\varepsilon^{\prime}\frac{\partial v_{\varepsilon^{\prime}}}{\partial x_{k}}\left(t,x,\widetilde{\omega}\right)\,\overline{\varphi(t,x)\,\Theta\left(\Phi^{-1}{\left(\frac{x}{\varepsilon^{\prime}},\widetilde{\omega}\right)},\widetilde{\omega}\right)}\,dx\,dt$ (4.90) $\displaystyle=c_{\Phi}^{-1}\\!\\!\iint_{\mathbb{R}^{n+1}_{T}}\\!\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}\frac{\partial}{\partial z_{k}}{\left(v^{\ast}_{\widetilde{\omega}}{\left(t,x,\Phi^{-1}{(z,\omega)},\omega\right)}\right)}\,$ $\displaystyle\hskip 90.0pt\times\,\overline{\varphi(t,x)\,\Theta\left(\Phi^{-1}(z,\omega),\omega\right)}\,dz\,d\mathbb{P}\,dx\,dt,$ for all functions $\varphi\in C^{\infty}_{\rm c}((-\infty,T)\times\mathbb{R}^{n})$ and $\Theta\in L^{2}_{\text{loc}}\left(\mathbb{R}^{n}\times\Omega\right)$ stationary. Moreover, the sequence ${{\\{v_{\varepsilon}^{0}(\cdot,\widetilde{\omega})\\}}_{\varepsilon>0}}$ defined by, $v_{\varepsilon}^{0}(x,\omega):=\psi{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\omega\right)},\omega,\theta^{\ast}\right)}v^{0}(x),\;\;(x,\omega)\in\mathbb{R}^{n}\\!\times\\!\Omega,$ (4.91) satisfies $v_{\varepsilon}^{0}(\cdot,\widetilde{\omega})\;\xrightharpoonup[\varepsilon\to 0]{2-{\rm s}}\;v^{0}(x)\,\psi{\left(\Phi^{-1}(z,\omega),\omega,\theta^{\ast}\right)},$ (4.92) for each stationary function ${\psi(\theta^{\ast})}$. 2.(The Split Process.) We consider the following Claim: There exists ${v_{\widetilde{\omega}}\in L^{2}(\mathbb{R}^{n+1}_{T})}$, such that $\displaystyle v^{\ast}_{\widetilde{\omega}}{\left(t,x,\Phi^{-1}(z,\omega),\omega\right)}$ $\displaystyle=v_{\widetilde{\omega}}(t,x)\,\psi{\left(\Phi^{-1}(z,\omega),\omega,\theta^{\ast}\right)}$ $\displaystyle\equiv v_{\widetilde{\omega}}(t,x)\,\Psi(z,\omega,\theta^{\ast}).$ Proof of Claim: First, for any $\widetilde{\omega}\in\Omega$ fixed, we take the function $Z_{\varepsilon}(t,x,\widetilde{\omega})=\varepsilon^{2}e^{i\frac{\lambda(\theta^{\ast})t}{\varepsilon^{2}}+2i\pi\frac{\theta^{\ast}\\!\cdot x}{\varepsilon}}\varphi(t,x)\,\Theta{\big{(}\Phi^{-1}\big{(}\frac{x}{\varepsilon},\widetilde{\omega}\big{)},\widetilde{\omega}\big{)}}$ (4.93) as a test function in the associated variational formulation of the equation (1.1), where ${\varphi\in C^{\infty}_{\rm c}((-\infty,T)\\!\times\\!\mathbb{R}^{n})}$ and $\Theta\in L^{\infty}\left(\mathbb{R}^{n}\times\Omega\right)$ stationary, with $\Theta(\cdot,\omega)$ smooth. Therefore, we obtain $\displaystyle-i\iint_{\mathbb{R}^{n+1}_{T}}u_{\varepsilon}(t,x,\widetilde{\omega})\,\frac{\partial\overline{Z_{\varepsilon}}}{\partial t}(t,x,\widetilde{\omega})\,dx\,dt+i\int_{\mathbb{R}^{n}}u_{\varepsilon}^{0}(x,\widetilde{\omega})\,\overline{Z_{\varepsilon}}(0,x,\widetilde{\omega})\,dx$ $\displaystyle+\iint_{\mathbb{R}^{n+1}_{T}}A{\left(\Phi^{-1}\left(\frac{x}{\varepsilon},\widetilde{\omega}\right),\widetilde{\omega}\right)}\nabla u_{\varepsilon}(t,x,\widetilde{\omega})\cdot\nabla\overline{Z_{\varepsilon}}(t,x,\widetilde{\omega})\,dx\,dt$ $\displaystyle+\frac{1}{\varepsilon^{2}}\iint_{\mathbb{R}^{n+1}_{T}}V{\left(\Phi^{-1}\left(\frac{x}{\varepsilon},\widetilde{\omega}\right),\widetilde{\omega}\right)}u_{\varepsilon}(t,x,\widetilde{\omega})\,\overline{Z_{\varepsilon}}(t,x,\widetilde{\omega})\,dx\,dt$ $\displaystyle+\iint_{\mathbb{R}^{n+1}_{T}}U{\left(\Phi^{-1}\left(\frac{x}{\varepsilon},\widetilde{\omega}\right),\widetilde{\omega}\right)}u_{\varepsilon}(t,x,\widetilde{\omega})\,\overline{Z_{\varepsilon}}(t,x,\widetilde{\omega})\,dx\,dt=0,$ and since $\displaystyle\frac{\partial Z_{\varepsilon}}{\partial t}(t,x,\widetilde{\omega})$ $\displaystyle=i\lambda(\theta^{\ast})\,e^{i\frac{\lambda(\theta^{\ast})t}{\varepsilon^{2}}+2i\pi\frac{\theta^{\ast}\\!\cdot x}{\varepsilon}}\,\varphi(t,x)\,\Theta{(\Phi^{-1}(\frac{x}{\varepsilon},\widetilde{\omega}),\widetilde{\omega})}+\mathrm{O}(\varepsilon^{2}),$ $\displaystyle\nabla Z_{\varepsilon}(t,x,\widetilde{\omega})$ $\displaystyle=\varepsilon\,e^{i\frac{\lambda(\theta^{\ast})t}{\varepsilon^{2}}+2i\pi\frac{\theta^{\ast}\\!\cdot x}{\varepsilon}}\,(\varepsilon\nabla+2i\pi\theta^{\ast})\big{(}\varphi(t,x)\,\Theta{(\Phi^{-1}{(\frac{x}{\varepsilon},\widetilde{\omega})},\widetilde{\omega})}\big{)},$ it follows that $\displaystyle-\lambda(\theta^{\ast})\iint_{\mathbb{R}^{n+1}_{T}}v_{\varepsilon}(t,x,\widetilde{\omega})\,\overline{\varphi(t,x)\,\Theta{(\Phi^{-1}\big{(}\frac{x}{\varepsilon},\widetilde{\omega}\big{)},\widetilde{\omega})}}\,dx\,dt$ $\displaystyle+\iint_{\mathbb{R}^{n+1}_{T}}A{(\Phi^{-1}{\big{(}\frac{x}{\varepsilon},\widetilde{\omega}\big{)}},\widetilde{\omega})}\,{(\varepsilon\nabla+2i\pi\theta^{\ast})}v_{\varepsilon}(t,x,\widetilde{\omega})$ $\displaystyle\hskip 60.0pt\cdot\overline{{(\varepsilon\nabla+2i\pi\theta^{\ast})}{\left(\varphi(t,x)\,\Theta{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega}\right)}\right)}}\,dx\,dt$ $\displaystyle+\iint_{\mathbb{R}^{n+1}_{T}}V{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega}\right)}\,v_{\varepsilon}(t,x,\widetilde{\omega})\,\overline{\varphi(t,x)\,\Theta{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega}\right)}}\,dx\,dt=\mathrm{O}(\varepsilon^{2}),$ where we have used (4.87), (4.88), (4.91), and (4.93). Although, it is more convenient to rewrite as $\displaystyle-\lambda(\theta^{\ast})\iint_{\mathbb{R}^{n+1}_{T}}v_{\varepsilon}(t,x,\widetilde{\omega})\,\overline{\varphi(t,x)\,\Theta{(\Phi^{-1}\big{(}\frac{x}{\varepsilon},\widetilde{\omega}\big{)},\widetilde{\omega})}}\,dx\,dt$ (4.94) $\displaystyle+\iint_{\mathbb{R}^{n+1}_{T}}{(\varepsilon\nabla+2i\pi\theta^{\ast})}v_{\varepsilon}(t,x,\widetilde{\omega})$ $\displaystyle\hskip 20.0pt\cdot\overline{A{(\Phi^{-1}{\big{(}\frac{x}{\varepsilon},\widetilde{\omega}\big{)}},\widetilde{\omega})}\,{(\varepsilon\nabla+2i\pi\theta^{\ast})}{\big{(}\varphi(t,x)\,\Theta{(\Phi^{-1}{\big{(}\frac{x}{\varepsilon},\widetilde{\omega}\big{)}},\widetilde{\omega})}\big{)}}}\,dx\,dt$ $\displaystyle+\iint_{\mathbb{R}^{n+1}_{T}}v_{\varepsilon}(t,x,\widetilde{\omega})\,\overline{\varphi(t,x)\,V{(\Phi^{-1}{\big{(}\frac{x}{\varepsilon},\widetilde{\omega}\big{)}},\widetilde{\omega})}\,\Theta{(\Phi^{-1}{\big{(}\frac{x}{\varepsilon},\widetilde{\omega}\big{)}},\widetilde{\omega})}}\,dx\,dt=\mathrm{O}(\varepsilon^{2}).$ Now, making $\varepsilon={\varepsilon^{\prime}}$, letting $\varepsilon^{\prime}\to 0$ and using the Definition 2.19, we have for a.e. ${\widetilde{\omega}\in\Omega}$, for all ${\varphi\in C^{\infty}_{\rm c}((-\infty,T)\\!\times\\!\mathbb{R}^{n})}$, $\Theta\in L^{\infty}\left(\mathbb{R}^{n}\times\Omega\right)$ stationary and $\Theta(\cdot,\omega)$ smooth, $\displaystyle-\lambda(\theta^{\ast})\,c_{\Phi}^{-1}\\!\\!\\!\iint_{\mathbb{R}^{n+1}_{T}}\\!\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}v^{\ast}_{\widetilde{\omega}}{\left(t,x,\Phi^{-1}(z,\omega),\omega\right)}$ $\displaystyle\hskip 90.0pt\times\overline{\varphi(t,x)\,\Theta{\left(\Phi^{-1}(z,\omega),\omega\right)}}\,dz\,d\mathbb{P}(\omega)\,dx\,dt$ $\displaystyle+c_{\Phi}^{-1}\\!\\!\\!\iint_{\mathbb{R}^{n+1}_{T}}\\!\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}{(\nabla_{\\!\\!z}+2i\pi\theta^{\ast})}{\left(v^{\ast}_{\widetilde{\omega}}{\left(t,x,\Phi^{-1}(z,\omega),\omega\right)}\right)}$ $\displaystyle\cdot\overline{\varphi(t,x)\,A{\left(\Phi^{-1}(z,\omega),\omega\right)}{\left[{(\nabla_{\\!\\!z}+2i\pi\theta^{\ast})}{\left(\Theta{\left(\Phi^{-1}(z,\omega),\omega\right)}\right)}\right]}}\,dz\,d\mathbb{P}(\omega)\,dx\,dt$ $\displaystyle+c_{\Phi}^{-1}\\!\\!\\!\iint_{\mathbb{R}^{n+1}_{T}}\\!\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}v^{\ast}_{\widetilde{\omega}}{\left(t,x,\Phi^{-1}(z,\omega),\omega\right)}\,$ $\displaystyle\hskip 60.0pt\times\overline{\varphi(t,x)\,V{\left(\Phi^{-1}(z,\omega),\omega\right)}\,\Theta{\left(\Phi^{-1}(z,\omega),\omega\right)}}\,dz\,d\mathbb{P}(\omega)\,dx\,dt=0.$ Therefore, due to an argument of density in the test functions (thanks to the topological structure of $\Omega$), we can conclude that $\displaystyle-\lambda(\theta^{\ast})\,\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}v^{\ast}_{\widetilde{\omega}}{\left(t,x,\Phi^{-1}(z,\omega),\omega\right)}\,\overline{\Theta{\left(\Phi^{-1}(z,\omega),\omega\right)}}\,dz\,d\mathbb{P}(\omega)$ $\displaystyle+\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}A{\left(\Phi^{-1}(z,\omega),\omega\right)}{(\nabla_{\\!\\!z}+2i\pi\theta^{\ast})}{\left(v^{\ast}_{\widetilde{\omega}}{\left(t,x,\Phi^{-1}(z,\omega),\omega\right)}\right)}$ $\displaystyle\hskip 60.0pt\cdot\overline{{(\nabla_{\\!\\!z}+2i\pi\theta^{\ast})}{\left(\Theta{\left(\Phi^{-1}(z,\omega),\omega\right)}\right)}}\,dz\,d\mathbb{P}(\omega)$ $\displaystyle+\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}\\!\\!\\!V{(\Phi^{-1}(z,\omega),\omega)}\,v^{\ast}_{\widetilde{\omega}}{(t,x,\Phi^{-1}(z,\omega),\omega)}\,$ $\displaystyle\hskip 120.0pt\times\overline{\Theta{(\Phi^{-1}(z,\omega),\omega)}}\,dz\,d\mathbb{P}(\omega)=0,$ for a.e. ${(t,x)\in\mathbb{R}^{n+1}_{T}}$ and for all ${\Theta}$ as above. Thus, the simplicity of the eigenvalue­ $\lambda(\theta^{\ast})$ assures us that for a.e. $\mathbb{R}^{n+1}_{T}$, the function ${(z,\omega)\mapsto v^{\ast}_{\widetilde{\omega}}{\left(t,x,\Phi^{-1}(z,\omega),\omega\right)}}$ (which belongs to the space ${\mathcal{H}}$) is parallel to the function ${\Psi(\theta^{\ast})}$, i.e., we can find ${v_{\widetilde{\omega}}(t,x)\in\mathbb{C}}$, such that $\displaystyle v^{\ast}_{\widetilde{\omega}}{\left(t,x,\Phi^{-1}(z,\omega),\omega\right)}$ $\displaystyle=$ $\displaystyle v_{\widetilde{\omega}}(t,x)\,\Psi(z,\omega,\theta^{\ast})$ $\displaystyle\equiv$ $\displaystyle v_{\widetilde{\omega}}(t,x)\,\psi{\left(\Phi^{-1}(z,\omega),\omega,\theta^{\ast}\right)}.$ Finally, since ${v^{\ast}_{\widetilde{\omega}}\in L^{2}(\mathbb{R}^{n+1}_{T};\mathcal{H})}$, we conclude that ${v_{\widetilde{\omega}}\in L^{2}(\mathbb{R}^{n+1}_{T})}$, which completes the proof of our claim. 3.(Homogenization Process.) Let ${\Lambda_{k}(\theta^{\ast})}$, for any $k\in\\{1,\ldots,n\\}$, be the function defined by $\Lambda_{k}(z,\omega,\theta^{\ast})=\frac{1}{2i\pi}\frac{\partial\Psi}{\partial\theta_{k}}(z,\omega,\theta^{\ast})=\frac{1}{2i\pi}\frac{\partial\psi}{\partial\theta_{k}}{\left(\Phi^{-1}(z,\omega),\omega,\theta^{\ast}\right)},\;(z,\omega)\in\mathbb{R}^{n}\\!\times\\!\Omega,$ where the function $\Psi(z,\omega,\theta^{\ast})=\psi{\left(\Phi^{-1}(z,\omega),\omega,\theta^{\ast}\right)}$ is the eigenfunction of the spectral cell problem (3.56). Then, we consider the following test function $Z_{\varepsilon}(t,x,\widetilde{\omega})=e^{i\frac{\lambda(\theta^{\ast})t}{\varepsilon^{2}}+2i\pi\frac{\theta^{\ast}\\!\cdot x}{\varepsilon}}{\big{(}\varphi(t,x)\,\Psi_{\varepsilon}(x,\widetilde{\omega},\theta^{\ast})+\varepsilon\sum_{k=1}^{n}\frac{\partial\varphi}{\partial x_{k}}(t,x)\,\Lambda_{k,\varepsilon}(x,\widetilde{\omega},\theta^{\ast})\big{)}},$ where ${\varphi\in C^{\infty}_{\rm c}((-\infty,T)\\!\times\\!\mathbb{R}^{n})}$ and $\Psi_{\varepsilon}(x,\widetilde{\omega},\theta^{\ast})=\Psi{\left(\frac{x}{\varepsilon},\widetilde{\omega},\theta^{\ast}\right)},\quad\;\;\Lambda_{k,\varepsilon}(x,\widetilde{\omega},\theta^{\ast})=\Lambda_{k}{\left(\frac{x}{\varepsilon},\widetilde{\omega},\theta^{\ast}\right)}.$ Using the function ${Z_{\varepsilon}}$ as test function in the variational formulation of the equation (1.1), we obtain $\displaystyle\big{[}i\int_{\mathbb{R}^{n}}u_{\varepsilon}^{0}(x,\widetilde{\omega})\,\overline{Z_{\varepsilon}}(0,x,\widetilde{\omega})\,dx-i\iint_{\mathbb{R}^{n+1}_{T}}u_{\varepsilon}(t,x,\widetilde{\omega})\,\frac{\partial\overline{Z_{\varepsilon}}}{\partial t}(t,x,\widetilde{\omega})\,dx\,dt\big{]}$ (4.95) $\displaystyle+\big{[}\iint_{\mathbb{R}^{n+1}_{T}}A{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega}\right)}\,\nabla u_{\varepsilon}(t,x,\widetilde{\omega})\cdot\nabla\overline{Z_{\varepsilon}}(t,x,\widetilde{\omega})\,dx\,dt\big{]}$ $\displaystyle+\big{[}\frac{1}{\varepsilon^{2}}\iint_{\mathbb{R}^{n+1}_{T}}V{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega}\right)\,u_{\varepsilon}(t,x,\widetilde{\omega})\,\overline{Z_{\varepsilon}}(t,x,\widetilde{\omega})\,dx\,dt}$ $\displaystyle\hskip 30.0pt+\iint_{\mathbb{R}^{n+1}_{T}}U{\left(x,\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega}\right)}\,u_{\varepsilon}(t,x,\widetilde{\omega})\,\overline{Z_{\varepsilon}}(t,x,\widetilde{\omega})\,dx\,dt\big{]}=0.$ In order to simplify the manipulation of the above equation, we shall denote by $I_{k}^{\varepsilon}(k=1,2,3)$ the respective term in the $k^{\text{th}}$ brackets, so that we can rewrite the equation (4.95) as $I_{1}^{\varepsilon}+I_{2}^{\varepsilon}+I_{3}^{\varepsilon}=0$. The analysis of the $I_{1}^{\varepsilon}$ term is triggered by the following computation $\displaystyle\frac{\partial Z_{\varepsilon}}{\partial t}(t,x,\widetilde{\omega})$ $\displaystyle=e^{i\frac{\lambda(\theta^{\ast})t}{\varepsilon^{2}}+2i\pi\frac{\theta^{\ast}\\!\cdot x}{\varepsilon}}{\Big{[}i\frac{\lambda(\theta^{\ast})}{\varepsilon^{2}}{\big{(}\varphi(t,x)\,\Psi_{\varepsilon}(x,\widetilde{\omega},\theta^{\ast})}}$ $\displaystyle+\,{{\varepsilon\sum_{k=1}^{n}\frac{\partial\varphi}{\partial x_{k}}(t,x)\,\Lambda_{k,\varepsilon}(x,\widetilde{\omega},\theta^{\ast})\big{)}}+\frac{\partial\varphi}{\partial t}(t,x)\,\Psi_{\varepsilon}(x,\widetilde{\omega},\theta^{\ast})}$ $\displaystyle{+\,\varepsilon\sum_{k=1}^{n}\frac{\partial^{2}\varphi}{\partial t\,\partial x_{k}}(t,x)\,\Lambda_{k,\varepsilon}(x,\widetilde{\omega},\theta^{\ast})\Big{]}},$ therefore we have $\displaystyle I_{1}^{\varepsilon}$ $\displaystyle=i\int_{\mathbb{R}^{n}}v_{\varepsilon}^{0}\,\overline{{\big{(}\varphi(0,x)\,\Psi_{\varepsilon}(x,\widetilde{\omega},\theta^{\ast})+\varepsilon\sum_{k=1}^{n}\frac{\partial\varphi}{\partial x_{k}}(0,x)\,\Lambda_{k,\varepsilon}(x,\widetilde{\omega},\theta^{\ast})\big{)}}}dx$ $\displaystyle-\frac{\lambda(\theta^{\ast})}{\varepsilon^{2}}\iint_{\mathbb{R}^{n+1}_{T}}v_{\varepsilon}\,\overline{{\big{(}\varphi(t,x)\,\Psi_{\varepsilon}(x,\widetilde{\omega},\theta^{\ast})+\varepsilon\sum_{k=1}^{n}\frac{\partial\varphi}{\partial x_{k}}(t,x)\,\Lambda_{k,\varepsilon}(x,\widetilde{\omega},\theta^{\ast})\big{)}}}dx\,dt$ $\displaystyle-i\iint_{\mathbb{R}^{n+1}_{T}}v_{\varepsilon}\,\overline{{\big{(}\frac{\partial\varphi}{\partial t}(t,x)\,\Psi_{\varepsilon}(x,\widetilde{\omega},\theta^{\ast})+\varepsilon\sum_{k=1}^{n}\frac{\partial^{2}\varphi}{\partial t\,\partial x_{k}}(t,x)\Lambda_{k,\varepsilon}(x,\widetilde{\omega},\theta^{\ast})\big{)}}}.$ For the analysis of the term $I_{2}^{\varepsilon}$, we need to make the following computations $\displaystyle\nabla Z_{\varepsilon}(t,x,\widetilde{\omega})$ $\displaystyle=e^{i\frac{\lambda(\theta^{\ast})t}{\varepsilon^{2}}+2i\pi\frac{\theta^{\ast}\\!\cdot x}{\varepsilon}}{\big{[}\nabla\varphi(t,x)\,\Psi_{\varepsilon}(x,\widetilde{\omega},\theta^{\ast})+\varphi(t,x)\,\nabla\Psi_{\varepsilon}(z,\widetilde{\omega},\theta^{\ast})}$ $\displaystyle+\varepsilon\sum_{k=1}^{n}\nabla{\big{(}\frac{\partial\varphi}{\partial x_{k}}(t,x)\big{)}}\,\Lambda_{k,\varepsilon}(x,\widetilde{\omega},\theta^{\ast})+\varepsilon\sum_{k=1}^{n}\frac{\partial\varphi}{\partial x_{k}}(t,x)\,\nabla\Lambda_{k,\varepsilon}(x,\widetilde{\omega},\theta^{\ast})$ $\displaystyle+\,2i\pi\frac{\theta^{\ast}}{\varepsilon}{{\big{(}\varphi(t,x)\,\Psi_{\varepsilon}(x,\widetilde{\omega},\theta^{\ast})+\varepsilon\sum_{k=1}^{n}\frac{\partial\varphi}{\partial x_{k}}(t,x)\,\Lambda_{k,\varepsilon}(x,\widetilde{\omega},\theta^{\ast})\big{)}}\big{]}}$ $\displaystyle=e^{i\frac{\lambda(\theta^{\ast})t}{\varepsilon^{2}}+2i\pi\frac{\theta^{\ast}\\!\cdot x}{\varepsilon}}{\big{[}\varphi(t,x)\,{\big{(}\nabla+2i\pi\frac{\theta^{\ast}}{\varepsilon}\big{)}}\Psi_{\varepsilon}(x,\widetilde{\omega},\theta^{\ast})}$ $\displaystyle+{\,\varepsilon\sum_{k=1}^{n}\frac{\partial\varphi}{\partial x_{k}}(t,x)\,{\big{(}\nabla+2i\pi\frac{\theta^{\ast}}{\varepsilon}\big{)}}\Lambda_{k,\varepsilon}(x,\widetilde{\omega},\theta^{\ast})+\nabla\varphi(t,x)\,\Psi_{\varepsilon}(x,\widetilde{\omega},\theta^{\ast})}$ $\displaystyle+\,{\varepsilon\sum_{k=1}^{n}\nabla{\big{(}\frac{\partial\varphi}{\partial x_{k}}(t,x)\big{)}}\,\Lambda_{k,\varepsilon}(x,\widetilde{\omega},\theta^{\ast})\big{]}},$ and from this, we have $\displaystyle A{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega}\right)}\nabla u_{\varepsilon}(t,x,\widetilde{\omega})\cdot\overline{\nabla Z_{\varepsilon}}(t,x,\widetilde{\omega})$ $\displaystyle=A{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega}\right)}{\big{[}\nabla u_{\varepsilon}(t,x,\widetilde{\omega})\,e^{-\left({i\frac{\lambda(\theta^{\ast})t}{\varepsilon^{2}}+2i\pi\frac{\theta^{\ast}\\!\cdot x}{\varepsilon}}\right)}\big{]}}$ $\displaystyle\cdot\big{[}\overline{\varphi}(t,x)\,(\nabla-2i\pi\frac{\theta^{\ast}}{\varepsilon})\overline{\Psi_{\varepsilon}}(x,\widetilde{\omega},\theta^{\ast})}+\varepsilon\\!\sum_{k=1}^{n}\frac{\partial\overline{\varphi}}{\partial x_{k}}(t,x)\,{(\nabla\\!\\!-2i\pi\frac{\theta^{\ast}}{\varepsilon})\overline{\Lambda_{k,\varepsilon}}(x,\widetilde{\omega},\theta^{\ast})$ $\displaystyle+\nabla\overline{\varphi}(t,x)\,\overline{\Psi_{\varepsilon}}(x,\widetilde{\omega},\theta^{\ast})+\varepsilon\sum_{k=1}^{n}\nabla{\big{(}\frac{\partial\overline{\varphi}}{\partial x_{k}}(t,x)\big{)}}\,\overline{\Lambda_{k,\varepsilon}}(x,\widetilde{\omega},\theta^{\ast})\big{]}.$ Then, from equation (4.88) and using the terms ${\big{(}\nabla+2i\pi\frac{\theta^{\ast}}{\varepsilon}\big{)}}(v_{\varepsilon}(t,x,\widetilde{\omega})\,\overline{\varphi}(t,x)),\quad\big{(}\nabla+2i\pi\frac{\theta^{\ast}}{\varepsilon}\big{)}(v_{\varepsilon}(t,x,\widetilde{\omega})\,\frac{\partial\overline{\varphi}}{\partial x_{k}}(t,x)),$ it follows from the above equation that $A{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega}\right)}\nabla u_{\varepsilon}\cdot\overline{\nabla Z_{\varepsilon}}=\sum_{k=1}^{n}\left(I_{2,1}^{\varepsilon,k}+I_{2,2}^{\varepsilon,k}+I_{2,3}^{\varepsilon,k}\right)(t,x,\widetilde{\omega}),$ where $\displaystyle I_{2,1}^{\varepsilon,k}(t,x,\widetilde{\omega}):=\varepsilon\,A{(\Phi^{-1}{\big{(}\frac{x}{\varepsilon},\widetilde{\omega}\big{)}},\widetilde{\omega})}\big{[}{\big{(}\nabla+2i\pi\frac{\theta^{\ast}}{\varepsilon}\big{)}}{(v_{\varepsilon}(t,x,\widetilde{\omega})\,\frac{\partial\overline{\varphi}}{\partial x_{k}}(t,x))}\big{]}$ $\displaystyle\quad\cdot\big{[}{\big{(}\nabla-2i\pi\frac{\theta^{\ast}}{\varepsilon}\big{)}}\overline{\Lambda_{k,\varepsilon}}(x,\widetilde{\omega},\theta^{\ast})\big{]}$ $\displaystyle-A(\Phi^{-1}{\big{(}\frac{x}{\varepsilon},\widetilde{\omega}\big{)}},\widetilde{\omega})\big{[}v_{\varepsilon}(t,x,\widetilde{\omega})\,\frac{\partial\overline{\varphi}}{\partial x_{k}}(t,x)\,e_{k}\big{]}\\!\cdot\\!\big{[}{\big{(}\nabla-2i\pi\frac{\theta^{\ast}}{\varepsilon}\big{)}}\overline{\Psi_{\varepsilon}}(x,\widetilde{\omega},\theta^{\ast})\big{]}$ $\displaystyle+A(\Phi^{-1}{\big{(}\frac{x}{\varepsilon},\widetilde{\omega}\big{)}},\widetilde{\omega}){\big{[}{\big{(}\nabla+2i\pi\frac{\theta^{\ast}}{\varepsilon}\big{)}}(v_{\varepsilon}(t,x,\widetilde{\omega})\,\frac{\partial\overline{\varphi}}{\partial x_{k}}(t,x))}\big{]}\\!\cdot\\!\big{[}{e_{k}}\overline{\Psi_{\varepsilon}}(x,\widetilde{\omega},\theta^{\ast})\big{]},$ $\displaystyle I_{2,2}^{\varepsilon,k}(t,x,\widetilde{\omega}):=\frac{1}{n}A{(\Phi^{-1}{\big{(}\frac{x}{\varepsilon},\widetilde{\omega}\big{)}},\widetilde{\omega})}{\big{[}{\big{(}\nabla+2i\pi\frac{\theta^{\ast}}{\varepsilon}\big{)}}(v_{\varepsilon}(t,x,\widetilde{\omega})\overline{\varphi}(t,x))\big{]}}$ $\displaystyle\quad\quad\cdot\big{[}{\big{(}\nabla-2i\pi\frac{\theta^{\ast}}{\varepsilon}\big{)}}\overline{\Psi_{\varepsilon}}(x,\widetilde{\omega},\theta^{\ast})\big{]}$ $\displaystyle\quad\quad-A(\Phi^{-1}{\big{(}\frac{x}{\varepsilon},\widetilde{\omega}\big{)}},\widetilde{\omega}){\big{[}v_{\varepsilon}(t,x,\widetilde{\omega})\,\nabla\frac{\partial\overline{\varphi}}{\partial x_{k}}(t,x)\big{]}}\cdot{\big{[}e_{k}\,\overline{\Psi_{\varepsilon}}(x,\widetilde{\omega},\theta^{\ast})\big{]}},$ and $\displaystyle I_{2,3}^{\varepsilon,k}(t,x,\widetilde{\omega}):=A(\Phi^{-1}{\big{(}\frac{x}{\varepsilon},\widetilde{\omega}\big{)}},\widetilde{\omega})\big{[}{\left(\varepsilon\nabla+2i\pi\theta^{\ast}\right)}v_{\varepsilon}(t,x,\widetilde{\omega})\big{]}$ $\displaystyle\quad\cdot\big{[}\nabla{\big{(}\frac{\partial\overline{\varphi}}{\partial x_{k}}(t,x)\big{)}}\,\overline{\Lambda_{k,\varepsilon}}(x,\widetilde{\omega},\theta^{\ast})\big{]}$ $\displaystyle-A(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega})\big{[}v_{\varepsilon}(t,x,\widetilde{\omega})\,\nabla\frac{\partial\overline{\varphi}}{\partial x_{k}}(t,x)\big{]}\\!\cdot\\!\big{[}{\left(\varepsilon\nabla-2i\pi\theta^{\ast}\right)}\overline{\Lambda_{k,\varepsilon}}(x,\widetilde{\omega},\theta^{\ast})\big{]}.$ Thus integrating in $\mathbb{R}^{n+1}_{T}$ we recover the $I_{2}^{\varepsilon}$ term, that is $\displaystyle I_{2}^{\varepsilon}=\iint_{\mathbb{R}^{n+1}_{T}}A{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega}\right)}\nabla u_{\varepsilon}(t,x,\widetilde{\omega})\cdot\overline{\nabla Z_{\varepsilon}}(t,x,\widetilde{\omega})\,dx\,dt$ $\displaystyle\qquad\qquad=\sum_{k=1}^{n}\iint_{\mathbb{R}^{n+1}_{T}}\left(I_{2,1}^{\varepsilon,k}+I_{2,2}^{\varepsilon,k}+I_{2,3}^{\varepsilon,k}\right)(t,x,\widetilde{\omega})\,dx\,dt.$ (4.96) Now, with the help of the first auxiliar cell equation (3.3), we intend to simplify the expression of $I_{2}^{\varepsilon}$ to a more comely one. For this aim, we shall take ${v_{\varepsilon}(t,\cdot,\widetilde{\omega})\,\overline{\varphi}(t,\cdot)}$, $t\in(0,T)$, as a test function in equation (3.84). Then, we obtain $\displaystyle\int_{\mathbb{R}^{n}}\\!\\!A(\Phi^{-1}{\big{(}\frac{x}{\varepsilon},\widetilde{\omega}\big{)}},\widetilde{\omega})[\big{(}\nabla\\!+2i\pi\frac{\theta^{\ast}}{\varepsilon}\big{)}{(v_{\varepsilon}(t,x,\widetilde{\omega})\,\overline{\varphi})}]\\!\cdot\\![\big{(}\nabla\\!-2i\pi\frac{\theta^{\ast}}{\varepsilon}\big{)}\overline{\Psi_{\varepsilon}}(x,\widetilde{\omega},\theta^{\ast})]dx$ $\displaystyle\quad=\frac{\lambda(\theta^{\ast})}{\varepsilon^{2}}\int_{\mathbb{R}^{n}}{(v_{\varepsilon}(t,x,\widetilde{\omega})\,\overline{\varphi}(t,x))}\,\overline{\Psi_{\varepsilon}}(x,\widetilde{\omega},\theta^{\ast})\,dx$ $\displaystyle\quad-\frac{1}{\varepsilon^{2}}\int_{\mathbb{R}^{n}}V(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega})\,{(v_{\varepsilon}(t,x,\widetilde{\omega})\,\overline{\varphi}(t,x))}\,\overline{\Psi_{\varepsilon}}(x,\widetilde{\omega},\theta^{\ast})\,dx.$ Therefore, comparing with $I_{2,2}^{\varepsilon,k}(t,x,\widetilde{\omega})$ term obtained before, we have $\displaystyle\iint_{\mathbb{R}^{n+1}_{T}}I_{2,2}^{\varepsilon,k}(t,x,\widetilde{\omega})\,dx\,dt$ (4.97) $\displaystyle=\frac{\lambda(\theta^{\ast})}{n\varepsilon^{2}}\iint_{\mathbb{R}^{n+1}_{T}}{(v_{\varepsilon}(t,x,\widetilde{\omega})\,\overline{\varphi}(t,x))}\,\overline{\Psi_{\varepsilon}}(x,\widetilde{\omega},\theta^{\ast})\,dx\,dt$ $\displaystyle-\frac{1}{n\varepsilon^{2}}\iint_{\mathbb{R}^{n+1}_{T}}V{(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega})}\,(v_{\varepsilon}(t,x,\widetilde{\omega})\,\overline{\varphi}(t,x))\,\overline{\Psi_{\varepsilon}}(x,\widetilde{\omega},\theta^{\ast})\,dx\,dt$ $\displaystyle-\iint_{\mathbb{R}^{n+1}_{T}}\\!\\!\\!A(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega}){[v_{\varepsilon}(t,x,\widetilde{\omega})\,\nabla\frac{\partial\overline{\varphi}}{\partial x_{k}}(t,x)]}\cdot{[e_{k}\,\overline{\Psi_{\varepsilon}}(x,\widetilde{\omega},\theta^{\ast})]}\,dx\,dt.$ Analogously, taking ${v_{\varepsilon}(t,\cdot,\widetilde{\omega})\,\displaystyle\frac{\partial\overline{\varphi}}{\partial x_{k}}(t,\cdot)}$, $t\in(0,T)$, with ${k\in\\{1,\ldots,n\\}}$ as a test function in the equation (3.83), taking into account that $\nabla_{\\!\theta}\lambda(\theta^{\ast})=0$ and comparing this expression with $I_{2,1}^{\varepsilon,k}(t,x,\widetilde{\omega})$, we deduce that $\displaystyle\iint_{\mathbb{R}^{n+1}_{T}}I_{2,1}^{\varepsilon,k}(t,x,\widetilde{\omega})\,dx\,dt$ $\displaystyle\quad=\frac{\lambda(\theta^{\ast})}{\varepsilon}\iint_{\mathbb{R}^{n+1}_{T}}{(v_{\varepsilon}(t,x,\widetilde{\omega})\,\frac{\partial\overline{\varphi}}{\partial x_{k}}(t,x))}\,\overline{\Lambda_{k,\varepsilon}}(x,\widetilde{\omega},\theta^{\ast})\,dx\,dt$ (4.98) $\displaystyle\quad-\frac{1}{\varepsilon}\iint_{\mathbb{R}^{n+1}_{T}}V{(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega})}\,{(v_{\varepsilon}(t,x,\widetilde{\omega})\,\frac{\partial\overline{\varphi}}{\partial x_{k}}(t,x))}\,\overline{\Lambda_{k,\varepsilon}}(x,\widetilde{\omega},\theta^{\ast})\,dx\,dt.$ Therefore, summing equations (4.97), (4.1), we arrive at $\displaystyle\sum_{k=1}^{n}\iint_{\mathbb{R}^{n+1}_{T}}\Big{(}I_{2,1}^{\varepsilon,k}+I_{2,2}^{\varepsilon,k}\Big{)}(t,x,\widetilde{\omega})\,dxdt$ $\displaystyle\quad\\!\\!={\frac{\lambda(\theta^{\ast})}{\varepsilon^{2}}\iint_{\mathbb{R}^{n+1}_{T}}\\!\\!\\!v_{\varepsilon}(t,x,\widetilde{\omega})\,\overline{\varphi(t,x)\,\Psi_{\varepsilon}(x,\widetilde{\omega},\theta^{\ast})+\varepsilon\\!\sum_{k=1}^{n}\frac{\partial\varphi}{\partial x_{k}}(t,x)\,\Lambda_{k,\varepsilon}(x,\widetilde{\omega},\theta^{\ast})}\,dxdt}$ $\displaystyle\quad-\frac{1}{\varepsilon^{2}}\iint_{\mathbb{R}^{n+1}_{T}}\\!V{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega}\right)}\,v_{\varepsilon}(t,x,\widetilde{\omega})$ (4.99) $\displaystyle\hskip 85.35826pt\times\,\overline{\varphi(t,x)\,\Psi_{\varepsilon}(x,\widetilde{\omega},\theta^{\ast})+\varepsilon\sum_{k=1}^{n}\frac{\partial\varphi}{\partial x_{k}}(t,x)\,\Lambda_{k,\varepsilon}(x,\widetilde{\omega},\theta^{\ast})}\,dxdt$ $\displaystyle\quad-\sum_{k=1}^{n}\iint_{\mathbb{R}^{n+1}_{T}}\\!A{(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega})}{[v_{\varepsilon}(t,x,\widetilde{\omega})\,\nabla\frac{\partial\overline{\varphi}}{\partial x_{k}}(t,x)]}\\!\cdot\\!{[e_{k}\,\overline{\Psi_{\varepsilon}}(x,\widetilde{\omega},\theta^{\ast})]}\,dxdt.$ Moreover, expressing the $I_{3}^{\varepsilon}$ term as $\displaystyle I_{3}^{\varepsilon}=\iint_{\mathbb{R}^{n+1}_{T}}\frac{1}{\varepsilon^{2}}\,V{(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega})}\,v_{\varepsilon}(t,x,\widetilde{\omega})$ $\displaystyle\qquad\qquad\qquad\qquad\times\,\overline{\varphi(t,x)\,\Psi_{\varepsilon}(x,\widetilde{\omega},\theta^{\ast})+\varepsilon\sum_{k=1}^{n}\frac{\partial\varphi}{\partial x_{k}}(t,x)\Lambda_{k,\varepsilon}(x,\widetilde{\omega},\theta^{\ast})}\,dx\,dt$ $\displaystyle\qquad\quad+\iint_{\mathbb{R}^{n+1}_{T}}U{(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega})}\,v_{\varepsilon}(t,x,\widetilde{\omega})$ $\displaystyle\qquad\qquad\qquad\qquad\times\,\overline{\varphi(t,x)\,\Psi_{\varepsilon}(x,\widetilde{\omega},\theta^{\ast})+\varepsilon\sum_{k=1}^{n}\frac{\partial\varphi}{\partial x_{k}}(t,x)\Lambda_{k,\varepsilon}(x,\widetilde{\omega},\theta^{\ast})}\,dx\,dt,$ adding with (4.1) and $I_{1}^{\varepsilon}$, we obtain $\displaystyle I_{1}^{\varepsilon}+\sum_{k=1}^{n}\iint_{\mathbb{R}^{n+1}_{T}}\Big{(}I_{2,1}^{\varepsilon,k}+I_{2,2}^{\varepsilon,k}\Big{)}(t,x,\widetilde{\omega})\,dx\,dt+I_{3}^{\varepsilon}$ $\displaystyle={i\int_{\mathbb{R}^{n}}\\!\\!\\!v_{\varepsilon}^{0}(x,\widetilde{\omega})\,\overline{\varphi(0,x)\,\Psi_{\varepsilon}(x,\widetilde{\omega},\theta^{\ast})}dx-i\\!\iint_{\mathbb{R}^{n+1}_{T}}\\!\\!\\!\\!v_{\varepsilon}(t,x,\widetilde{\omega})\,\overline{\frac{\partial\varphi}{\partial t}(t,x)\,\Psi_{\varepsilon}(x,\widetilde{\omega},\theta^{\ast})}}$ $\displaystyle-\sum_{k,\ell=1}^{n}\iint_{\mathbb{R}^{n+1}_{T}}\\!\\!\\!{v_{\varepsilon}(t,x,\widetilde{\omega})\,e_{\ell}\,\frac{\partial^{2}\overline{\varphi}}{\partial x_{\ell}\,\partial x_{k}}(t,x)}\cdot\overline{A{(\Phi^{-1}{(\frac{x}{\varepsilon},\widetilde{\omega})},\widetilde{\omega})}{\;e_{k}\Psi_{\varepsilon}(x,\widetilde{\omega},\theta^{\ast})}}\,dxdt$ $\displaystyle+\iint_{\mathbb{R}^{n+1}_{T}}U{(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega})}\,v_{\varepsilon}(t,x,\widetilde{\omega})\,\overline{\varphi(t,x)\,\Psi_{\varepsilon}(x,\widetilde{\omega},\theta^{\ast})}\,dxdt+\,\mathrm{O}(\varepsilon).$ Thus, for $\varepsilon=\varepsilon^{\prime}(\widetilde{\omega})$ and due to Step 2, that is to say $v_{\varepsilon^{\prime}}(t,x,\widetilde{\omega})\;\xrightharpoonup[\varepsilon^{\prime}\to 0]{2-{\rm s}}\;v_{\widetilde{\omega}}(t,x)\,\Psi(z,\omega,\theta^{\ast}),$ we obtain after letting $\varepsilon^{\prime}\to 0$, from the previous equation $\displaystyle\lim_{\varepsilon^{\prime}\to 0}\Big{(}I_{1}^{\varepsilon^{\prime}}+\sum_{k=1}^{n}\iint_{\mathbb{R}^{n+1}_{T}}\Big{(}I_{2,1}^{\varepsilon^{\prime},k}+I_{2,2}^{\varepsilon^{\prime},k}\Big{)}(t,x,\widetilde{\omega})\,dx\,dt+I_{3}^{\varepsilon^{\prime}}\Big{)}$ (4.100) $\displaystyle=i\int_{\mathbb{R}^{n}}{\left(\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}{\left|\Psi(z,\omega,\theta^{\ast})\right|}^{2}dz\,d\mathbb{P}\right)}v^{0}(x)\,\overline{\varphi}(0,x)\,dx$ $\displaystyle-i\iint_{\mathbb{R}^{n+1}_{T}}\\!\\!{(\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}{\left|\Psi(z,\omega,\theta^{\ast})\right|}^{2}dz\,d\mathbb{P})}\,v_{\widetilde{\omega}}(t,x)\,\frac{\partial\overline{\varphi}}{\partial t}(t,x)\,dx\,dt$ $\displaystyle-\sum_{k,\ell=1}^{n}\iint_{\mathbb{R}^{n+1}_{T}}\\!\\!{(\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}\\!\\!A{\left(\Phi^{-1}(z,\omega),\omega\right)}\,{(e_{\ell}\,\Psi)}\cdot{(e_{k}\,\overline{\Psi})}\,dz\,d\mathbb{P})}$ $\displaystyle\quad\times\,v_{\widetilde{\omega}}(t,x)\,\frac{\partial^{2}\overline{\varphi}}{\partial x_{\ell}\,\partial x_{k}}(t,x)\,dx\,dt$ $\displaystyle+\iint_{\mathbb{R}^{n+1}_{T}}\\!\\!{(\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}\\!\\!\\!\\!U{\left(\Phi^{-1}(z,\omega),\omega\right)}{|\Psi|}^{2}dz\,d\mathbb{P})}\,v_{\widetilde{\omega}}(t,x)\,\overline{\varphi}(t,x)\,dx\,dt.$ Proceeding in the same way with respect to the term $I_{2,3}^{\varepsilon,k}(t,x,\widetilde{\omega})$, we obtain $\displaystyle\lim_{\varepsilon^{\prime}\to 0}\sum_{k=1}^{n}\iint_{\mathbb{R}^{n+1}_{T}}I_{2,3}^{\varepsilon^{\prime},k}(t,x,\widetilde{\omega})\,dx\,dt$ $\displaystyle\quad=\sum_{k,\ell=1}^{n}\iint_{\mathbb{R}^{n+1}_{T}}\\!\\!\Big{(}\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}\\!\\!A{\left(\Phi^{-1}(z,\omega),\omega\right)}\,{\left({\left(\nabla_{\\!\\!z}+2i\pi\theta^{\ast}\right)}\Psi(z,\omega,\theta^{\ast})\right)}$ $\displaystyle\hskip 56.9055pt\cdot{\left(e_{\ell}\,\overline{\Lambda_{k}}(z,\omega,\theta^{\ast})\right)}\,dz\,d\mathbb{P}\Big{)}v_{\widetilde{\omega}}(t,x)\,\frac{\partial^{2}\overline{\varphi}}{\partial x_{\ell}\,\partial x_{k}}(t,x)\,dx\,dt$ $\displaystyle\quad-\sum_{k,\ell=1}^{n}\iint_{\mathbb{R}^{n+1}_{T}}\\!\\!\Big{(}\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}\\!\\!A{\left(\Phi^{-1}(z,\omega),\omega\right)}\,{\left(e_{\ell}\,\Psi(z,\omega,\theta^{\ast})\right)}$ (4.101) $\displaystyle\hskip 56.9055pt\cdot{\left(\nabla_{\\!\\!z}-2i\pi\theta^{\ast}\right)}\overline{\Lambda_{k}}(z,\omega,\theta^{\ast})\,dz\,d\mathbb{P}\Big{)}v_{\widetilde{\omega}}(t,x)\,\frac{\partial^{2}\overline{\varphi}}{\partial x_{\ell}\,\partial x_{k}}(t,x)\,dx\,dt.$ Therefore, since $I_{1}^{\varepsilon^{\prime}}+I_{2}^{\varepsilon^{\prime}}+I_{3}^{\varepsilon^{\prime}}=0$, (see (4.95)), combining the two last equations we conclude that, the function $v_{\widetilde{\omega}}$ is a distribution solution of the following homogenized Schrödinger equation $\left\\{\begin{array}[]{c}i\displaystyle\frac{\partial v_{\widetilde{\omega}}}{\partial t}(t,x)-{\rm div}{\big{(}B^{\ast}\nabla v_{\widetilde{\omega}}(t,x)\big{)}}+U^{\ast}v_{\widetilde{\omega}}(t,x)=0,\;\,(t,x)\in\mathbb{R}^{n+1}_{T},\\\\[5.0pt] v_{\widetilde{\omega}}(0,x)=v^{0}(x),\;\,x\in\mathbb{R}^{n},\end{array}\right.$ (4.102) where the effective tensor $\displaystyle B_{k,\ell}^{\ast}=\frac{1}{c_{\psi}}{\int_{\Omega}\int_{\Phi\left([0,1)^{n},\omega\right)}\\!\\!\\!\big{\\{}A{\left(\Phi^{-1}(z,\omega),\omega\right)}\,{\left(e_{\ell}\,\Psi(z,\omega,\theta^{\ast})\right)}}\cdot{\left(e_{k}\,\overline{\Psi}(z,\omega,\theta^{\ast})\right)}$ $\displaystyle\qquad\qquad+A{\left(\Phi^{-1}(z,\omega),\omega\right)}\,{\left(e_{\ell}\,\Psi(z,\omega,\theta^{\ast})\right)}\cdot{\left((\nabla_{\\!\\!z}-2i\pi\theta^{\ast})\overline{\Lambda_{k}}(z,\omega,\theta^{\ast})\right)}$ $\displaystyle\qquad\qquad\qquad-A{\left(\Phi^{-1}(z,\omega),\omega\right)}\,{\Big{(}(\nabla_{\\!\\!z}+2i\pi\theta^{\ast})\Psi(z,\omega,\theta^{\ast})\Big{)}}$ $\displaystyle\hskip 170.71652pt\cdot{\left(e_{\ell}\,\overline{\Lambda_{k}}(z,\omega,\theta^{\ast})\right)}\big{\\}}\,dz\,d\mathbb{P}(\omega),$ (4.103) for ${k,\ell\in\\{1,\ldots,n\\}}$, and the effective potential $U^{\ast}=c_{\psi}^{-1}\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}U{\left(\Phi^{-1}(z,\omega),\omega\right)}{|\Psi(z,\omega,\theta^{\ast})|}^{2}dz\,d\mathbb{P}(\omega)$ with $\displaystyle c_{\psi}=\\!\\!\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}$ $\displaystyle\\!\\!\\!{|\Psi(z,\omega,\theta^{\ast})|}^{2}dz\,d\mathbb{P}(\omega)$ $\displaystyle\equiv\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}\\!\\!\\!{|\psi{(\Phi^{-1}{(\frac{x}{\varepsilon},\omega)},\omega)}|}^{2}dz\,d\mathbb{P}(\omega).$ Moreover, we are allowed to change the tensor $B^{\ast}$ in the equation (4.102) by the corresponding symmetric part of it, that is $A^{\ast}=\big{(}B^{\ast}+(B^{\ast})^{t}\big{)}/2.$ 4.(The Form of the Matrix $A^{\ast}$.) Now, we show that the homogenized tensor $A^{\ast}$ is a real value matrix, and it coincides with the hessian matrix of the function ${\theta\mapsto\lambda(\theta)}$ in the point $\theta^{\ast}$. In fact, using that ${\nabla_{\\!\\!\theta}\lambda(\theta^{\ast})=0}$ the equation (3.3) can be written as $\displaystyle\quad\frac{1}{4\pi^{2}}\frac{\partial^{2}\lambda(\theta^{\ast})}{\partial\theta_{\ell}\,\partial\theta_{k}}\,c_{\psi}$ $\displaystyle\qquad\qquad=\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}\big{\\{}A{\left(\Phi^{-1}(z,\omega),\omega\right)}{\left(e_{\ell}\,\Psi(z,\omega,\theta^{\ast})\right)}\cdot{\left(e_{k}\,\overline{\Psi}(z,\omega,\theta^{\ast})\right)}$ $\displaystyle\qquad\qquad\qquad+A{\left(\Phi^{-1}(z,\omega),\omega\right)}\,{\left(e_{\ell}\,\Psi(z,\omega,\theta^{\ast})\right)}\cdot{\left((\nabla_{\\!\\!z}-2i\pi\theta^{\ast})\overline{\Lambda_{k}}(z,\omega,\theta^{\ast})\right)}$ $\displaystyle\qquad\qquad\qquad-A{\left(\Phi^{-1}(z,\omega),\omega\right)}\,{\left[(\nabla_{\\!\\!z}+2i\pi\theta^{\ast})\Psi(z,\omega,\theta^{\ast})\right]}\cdot{\left(e_{\ell}\,\overline{\Lambda_{k}}(z,\omega,\theta^{\ast})\right)}$ $\displaystyle\qquad\qquad\qquad+A{\left(\Phi^{-1}(z,\omega),\omega\right)}\,{\left(e_{k}\,\Psi(z,\omega,\theta^{\ast})\right)}\cdot{\left(e_{\ell}\,\overline{\Psi}(z,\omega,\theta^{\ast})\right)}$ $\displaystyle\qquad\qquad\qquad+A{\left(\Phi^{-1}(z,\omega),\omega\right)}\,{\left(e_{k}\,\Psi(z,\omega,\theta^{\ast})\right)}\cdot{\left((\nabla_{\\!\\!z}-2i\pi\theta^{\ast})\overline{\Lambda_{\ell}}(z,\omega,\theta^{\ast})\right)}$ $\displaystyle\qquad\qquad\qquad-A{\left(\Phi^{-1}(z,\omega),\omega\right)}\,{\left((\nabla_{\\!\\!z}+2i\pi\theta^{\ast})\Psi(z,\omega,\theta^{\ast})\right)}$ (4.104) $\displaystyle\hskip 227.62204pt\cdot{\left(e_{k}\,\overline{\Lambda_{\ell}}(z,\omega,\theta^{\ast})\right)}\big{\\}}\,dz\,d\mathbb{P}(\omega),$ from which we obtain $A^{\ast}=\frac{1}{8\pi^{2}}\,D^{2}_{\\!\theta}\lambda(\theta^{\ast}).$ Therefore, from Remark 2.3 we deduce the well-posedness of the homogenized Schrödinger (4.102). Hence the function ${v_{\widetilde{\omega}}\in L^{2}(\mathbb{R}^{n+1}_{T})}$ does not depend on ${\widetilde{\omega}\in{\Omega}}$. Moreover, denoting by $v$ the unique solution of the problem (4.102), we have that the sequence ${\\{v_{\varepsilon}(t,x,\widetilde{\omega})\\}_{\varepsilon>0}\subset L^{2}(\mathbb{R}^{n+1}_{T})}$ $\Phi_{\omega}-$two-scale converges to the function $v(t,x)\,\Psi(z,\omega,\theta^{\ast})\equiv v(t,x)\,\psi{\left(\Phi^{-1}(z,\omega),\omega,\theta^{\ast}\right)}.$ 5.(A Corrector-type Result.) Finally, we show the following corrector type result, that is, for a.e. ${\widetilde{\omega}\in\Omega}$ $\lim_{\varepsilon\to 0}\iint_{\mathbb{R}^{n+1}_{T}}\big{|}v_{\varepsilon}(t,x,\widetilde{\omega})-v(t,x)\,\psi{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega},\theta^{\ast}\right)}\big{|}^{2}dx\,dt=0.$ We begin by the simple observation $\displaystyle\iint_{\mathbb{R}^{n+1}_{T}}|v_{\varepsilon}(t,x,\widetilde{\omega})-v(t,x)\,\psi{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega},\theta^{\ast}\right)}|^{2}dxdt$ (4.105) $\displaystyle\quad=\iint_{\mathbb{R}^{n+1}_{T}}{\left|v_{\varepsilon}(t,x,\widetilde{\omega})\right|}^{2}dx\,dt$ $\displaystyle\quad-\iint_{\mathbb{R}^{n+1}_{T}}v_{\varepsilon}(t,x,\widetilde{\omega})\,\overline{v(t,x)\,\psi{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega},\theta^{\ast}\right)}}\,dxdt$ $\displaystyle\quad-\iint_{\mathbb{R}^{n+1}_{T}}\overline{v_{\varepsilon}(t,x,\widetilde{\omega})}\,v(t,x)\,\psi{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega},\theta^{\ast}\right)}\,dx\,dt$ $\displaystyle\quad+\iint_{\mathbb{R}^{n+1}_{T}}{\left|v(t,x)\,\psi{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega},\theta^{\ast}\right)}\right|}^{2}dx\,dt.$ From Lemma 4.1 we see that, the first integral of the right hand side of the above equation satisfies, for all ${t\in[0,T]}$ and a.e. ${\widetilde{\omega}\in\Omega}$ $\displaystyle\int_{\mathbb{R}^{n}}{\left|v_{\varepsilon}(t,x,\widetilde{\omega})\right|}^{2}dx$ $\displaystyle=$ $\displaystyle\int_{\mathbb{R}^{n}}{\left|u_{\varepsilon}(t,x,\widetilde{\omega})\right|}^{2}dx$ $\displaystyle=$ $\displaystyle\int_{\mathbb{R}^{n}}{\left|u_{\varepsilon}^{0}(x,\widetilde{\omega})\right|}^{2}dx\;\,=\;\,\int_{\mathbb{R}^{n}}{\left|v_{\varepsilon}^{0}(x,\widetilde{\omega})\right|}^{2}dx$ $\displaystyle=$ $\displaystyle\int_{\mathbb{R}^{n}}{\left|v^{0}(x)\,\psi{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega},\theta^{\ast}\right)}\right|}^{2}dx.$ Using the elliptic regularity theory (see E. De Giorgi [15], G. Stampacchia [36]), it follows that $\psi(\theta)\in L^{\infty}(\mathbb{R}^{n};L^{2}(\Omega))$ and we can apply the Ergodic Theorem to obtain $\displaystyle\lim_{\varepsilon\to 0}$ $\displaystyle\iint_{\mathbb{R}^{n+1}_{T}}{\left|v_{\varepsilon}(t,x,\widetilde{\omega})\right|}^{2}dx\,dt$ $\displaystyle=\lim_{\varepsilon\to 0}\iint_{\mathbb{R}^{n+1}_{T}}{\left|v^{0}(x)\right|}^{2}{\left|\psi{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega},\theta^{\ast}\right)}\right|}^{2}dxdt$ $\displaystyle=c_{\Phi}^{-1}\iint_{\mathbb{R}^{n+1}_{T}}\\!\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}{\left|v^{0}(x)\,\psi{\left(\Phi^{-1}(z,\omega),\omega,\theta^{\ast}\right)}\right|}^{2}dz\,d\mathbb{P}\,dxdt.$ Similarly, we have $\lim_{\varepsilon\to 0}\iint_{\mathbb{R}^{n+1}_{T}}{\left|v(t,x)\,\psi{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega},\theta^{\ast}\right)}\right|}^{2}dxdt\\\ =c_{\Phi}^{-1}\iint_{\mathbb{R}^{n+1}_{T}}\\!\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}{\left|v(t,x)\,\psi{\left(\Phi^{-1}(z,\omega),\omega,\theta^{\ast}\right)}\right|}^{2}dz\,d\mathbb{P}\,dxdt.$ Moreover, seeing that for a.e. ${\widetilde{\omega}\in\Omega}$ $\displaystyle\lim_{\varepsilon\to 0}$ $\displaystyle\iint_{\mathbb{R}^{n+1}_{T}}v_{\varepsilon}(t,x,\widetilde{\omega})\,\overline{v(t,x)\,\psi{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega},\theta^{\ast}\right)}}\,dxdt$ $\displaystyle=c_{\Phi}^{-1}\iint_{\mathbb{R}^{n+1}_{T}}\\!\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}\\!\\!\\!v(t,x)\,\psi{\left(\Phi^{-1}(z,\omega),\omega,\theta^{\ast}\right)}\,$ $\displaystyle\qquad\qquad\qquad\qquad\times\overline{v(t,x)\,\psi{\left(\Phi^{-1}(z,\omega),\omega,\theta^{\ast}\right)}}\,dz\,d\mathbb{P}\,dxdt,$ we can make ­${\varepsilon\to 0}$ in the equation (4.105) to find $\displaystyle\lim_{\varepsilon\to 0}\iint_{\mathbb{R}^{n+1}_{T}}{\left|v_{\varepsilon}(t,x,\widetilde{\omega})-v(t,x)\,\psi{\left(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega},\theta^{\ast}\right)}\right|}^{2}dxdt$ $\displaystyle\qquad=c_{\Phi}^{-1}{\big{(}\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}{\left|\psi{\left(\Phi^{-1}(z,\omega),\omega,\theta^{\ast}\right)}\right|}^{2}dz\,d\mathbb{P}(\omega)\big{)}}$ $\displaystyle\qquad\qquad\qquad\qquad\times\big{(}{\iint_{\mathbb{R}^{n+1}_{T}}{\left|v^{0}(x)\right|}^{2}dx\,dt}-{\iint_{\mathbb{R}^{n+1}_{T}}{\left|v(t,x)\right|}^{2}dx\,dt}\big{)},$ for a.e. ${\widetilde{\omega}\in\Omega}$. Therefore, using the energy conservation of the homogenized Schrödinger equation (4.86), that is, for all ${t\in[0,T]}$ $\int_{\mathbb{R}^{n}}{\left|v(t,x)\right|}^{2}dx=\int_{\mathbb{R}^{n}}{\left|v^{0}(x)\right|}^{2}dx,$ we obtain that, for a.e. ${\widetilde{\omega}\in\Omega}$ $\lim_{\varepsilon\to 0}\iint_{\mathbb{R}^{n+1}_{T}}|v_{\varepsilon}(t,x,\widetilde{\omega})-v(t,x)\,\psi{(\Phi^{-1}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega},\theta^{\ast})}|^{2}dxdt=0,$ completing the proof of the theorem. ∎ ### 4.2 Radom Perturbations of the Quasi-Periodic Case In this section, we shall give a nice application of the framework introduced in this paper, which can be used to homogenize a model beyond the periodic settings considered by Allaire and Piatnitski in [4]. For reaching this aim, we shall make use of some results discussed in Section 3.2 (Sobolev spaces on groups), in particular, Section 3.2.3. Other interesting application will be given in the last section. Let $n,m\geq 1$ be integers numbers and $\lambda_{1},\cdots,\lambda_{m}$ be vectors in $\mathbb{R}^{n}$ linearly independent over the set $\mathbb{Z}$ satisfying the condition that $\big{\\{}k\in\mathbb{Z}^{m};\,|k_{1}\lambda_{1}+\cdots+k_{m}\lambda_{m}|<d\big{\\}}$ is a finite set for any $d>0$. Let $\left(\Omega_{0},\mathcal{F}_{0},\mathbb{P}_{0}\right)$ be a probability space and $\tau_{0}:\mathbb{Z}^{n}\times\Omega_{0}\to\Omega_{0}$ be a discrete ergodic dynamical system and $\mathbb{R}^{m}/{\mathbb{Z}^{m}}$ be the $m-$dimensional torus which can be identified with the cube $[0,1)^{m}$. For $\Omega:=\Omega_{0}\times[0,1)^{m}$, consider the following continuous dynamical system $T:\mathbb{R}^{n}\times\Omega\to\Omega$, defined by $T(x)(\omega_{0},s):=\Big{(}\tau_{\left\lfloor s+Mx\right\rfloor}\omega_{0},s+Mx-\left\lfloor s+Mx\right\rfloor\Big{)},$ where $M$ is the matrix $M=\Big{(}\lambda_{i}\cdot e_{j}{\Big{)}}_{i=1,j=1}^{m,n}$ and $\left\lfloor y\right\rfloor$ denotes the unique element in $\mathbb{Z}^{m}$ such that $y-\left\lfloor y\right\rfloor\in[0,1)^{m}$. Now, we consider $[0,1)^{m}-$periodic functions $A_{\rm per}:\mathbb{R}^{m}\to\mathbb{R}^{n^{2}},\,V_{\rm per}:\mathbb{R}^{m}\to\mathbb{R}$ and $U_{\rm per}:\mathbb{R}^{m}\to\mathbb{R}$ such that * • There exists $a_{0},a_{1}>0$ such that for all $\xi\in\mathbb{R}^{n}$ and for a.e $y\in\mathbb{R}^{m}$ we have $a_{0}|\xi|^{2}\leq A_{\rm per}(y)\xi\cdot\xi\leq a_{1}|\xi|^{2}.$ * • $V_{\rm per},\,U_{\rm per}\in L^{\infty}(\mathbb{R}^{m})$. Let $B_{\rm per}:\mathbb{R}^{m}\to\mathbb{R}^{n^{2}}$ be a $[0,1)^{m}-$periodic matrix and $\Upsilon:\mathbb{R}^{n}\times[0,1)^{m}\to\mathbb{R}^{n}$ be any stochastic diffeomorphism satisfying $\nabla\Upsilon(x,s)=B_{\rm per}\Big{(}T(x)(\omega_{0},s)\Big{)}.$ Thus, we define the following stochastic deformation $\Phi:\mathbb{R}^{n}\times\Omega\to\mathbb{R}^{n}$ by $\Phi(x,\omega)=\Upsilon(x,s)+{\bf X}(\omega_{0}),$ where we have used the notation $\omega$ for the pair $(\omega_{0},s)\in\Omega$ and ${\bf X}:\Omega_{0}\to\mathbb{R}^{n}$ is a random vector. Now, taking $A(x,\omega):=A_{\rm per}\left(T(x)\omega\right),\,V(x,\omega):=V_{\rm per}\left(T(x)\omega\right),\;U(x,\omega):=U_{\rm per}\left(T(x)\omega\right)$ in the equation (1.1), it can be seen after some computations that the spectral equation correspondent is $\left\\{\begin{array}[]{l}-{\Big{(}{\rm div}_{\rm{QP}}+2i\pi\theta\Big{)}}{\left[A_{\rm per}{\left(\cdot\right)}{\Big{(}\nabla^{\rm{QP}}+2i\pi\theta\Big{)}}{\Psi}_{\rm per}(\cdot)\right]}\\\\[7.5pt] \hskip 56.9055pt+V_{\rm per}{\left(\cdot\right)}{\Psi}_{\rm per}(\cdot)=\lambda{\Psi}_{\rm per}(\cdot)\;\;\text{in}\,\;[0,1)^{m},\\\\[7.5pt] \hskip 42.67912pt{\Psi}_{\rm per}(\cdot)\;\;\;\psi\;\,\text{is a $[0,1)^{m}-$periodic function},\end{array}\right.$ (4.106) where the operators ${\rm div}_{\rm{QP}}$ and $\nabla^{\rm{QP}}$ are defined as * • $\left(\nabla^{\rm{QP}}u_{\rm per}\right)(y):=B_{\rm per}^{-1}(y)M^{\ast}\left(\nabla u_{\rm per}\right)(y)$; * • $\left(\rm{div}_{\rm{QP}}\,a\right)(y):=\rm{div}\left(MB_{\rm per}^{-1}(\cdot)a(\cdot)\right)(y)$. Although the coefficients of the spectral equation (4.106) can be seen as periodic functions, its analysis is possible thanks to the results developed in the Section 3.2.3. This happens due to the fact that, the bilinear form associated to the problem (4.106) can lose its coercivity which unable us to apply the classic theory. Assume that for some $\theta^{\ast}\in\mathbb{R}^{n}$, the spectral equation (4.106) admits a solution $\big{(}\lambda(\theta^{\ast}),\Psi_{\rm per}(\theta^{\ast})\big{)}\in\mathbb{R}\times H^{1}\left([0,1)^{m}\right)$, such that * • $\lambda(\theta^{\ast})$ is a simple eigenvalue; * • $\nabla\lambda(\theta^{\ast})=0$. Now, we consider the problem (1.1) with new coefficients as highlighted above and with well-prepared initial data, that is, $u_{\varepsilon}(x,\omega):=e^{2\pi i\frac{\theta^{\ast}\cdot x}{\varepsilon}}\,{\Psi}_{\rm per}\Big{(}T\left(\Phi^{-1}\left(\frac{x}{\varepsilon},\omega\right)\right)\omega,\theta^{\ast}\Big{)}v^{0}(x),$ for $(x,\omega)\in\mathbb{R}^{n}\times\Omega$ and $v^{0}\in C^{\infty}_{c}(\mathbb{R}^{n})$. Applying Theorem 4.2, the function $v_{\varepsilon}(t,x,\omega):=e^{-{\left(i\frac{\lambda(\theta^{\ast})t}{\varepsilon^{2}}+2i\pi\frac{\theta^{\ast}\\!\cdot x}{\varepsilon}\right)}}u_{\varepsilon}(t,x,\omega),\;\,(t,x)\in\mathbb{R}^{n+1}_{T},\;\omega\in\Omega,$ $\Phi_{\omega}-$two-scale converges strongly to ${v(t,x)\,{\Psi}_{\rm per}\Big{(}{T\left(\Phi^{-1}(z,\omega)\right)\omega,\theta^{\ast}}}\Big{)}$, where ${v\in C([0,T],L^{2}(\mathbb{R}^{n}))}$ is the unique solution of the homogenized Schrödinger equation $\left\\{\begin{array}[]{c}i\displaystyle\frac{\partial v}{\partial t}-{\rm div}{\left(A^{\ast}\nabla v\right)}+U^{\ast}v=0\,,\;\,\text{em}\;\,\mathbb{R}^{n+1}_{T},\\\\[7.5pt] v(0,x)=v^{0}(x)\,,\;\,x\in\mathbb{R}^{n},\end{array}\right.$ with effective matrix ${A^{\ast}=D_{\theta}^{2}\lambda(\theta^{\ast})}$ and effective potential $U^{\ast}=c^{-1}_{\psi}\int_{[0,1)^{m}}U_{\rm per}{\left(y\right)}\,{\left|{\Psi}_{\rm per}{\left(y,\theta^{\ast}\right)}\right|}^{2}|\det\left(B_{\rm per}(y)\right)|\,dy,$ where $c_{\psi}=\int_{[0,1)^{m}}{\left|{\Psi}_{\rm per}{\left(y,\theta^{\ast}\right)}\right|}^{2}\,|\det\left(B_{\rm per}(y)\right)|\,dy.$ It is worth highlighting that this singular example encompasses the settings considered by Allaire-Piatnitski in [4]. For this, it is enough to take $n=m,\,\lambda_{j}=e_{j},\,\Upsilon(\cdot,s)\equiv I_{n\times n},\;\text{and ${\bf X}(\cdot)\equiv 0$.}$ Moreover, we consider $[0,1)^{m}-$periodic functions: $V_{\rm per},U_{\rm per}:\mathbb{R}^{m}\to\mathbb{R}$, and $A_{\rm per}:\mathbb{R}^{m}\to\mathbb{R}^{n^{2}}$, such that * • There exists $a_{0},a_{1}>0$ such that for all $\xi\in\mathbb{R}^{n}$ and for a.e $y\in\mathbb{R}^{m}$ we have $a_{0}|\xi|^{2}\leq A_{\rm per}(y)\xi\cdot\xi\leq a_{1}|\xi|^{2};$ * • $V_{\rm per},\,U_{\rm per}\in L^{\infty}(\mathbb{R}^{m})$. ## 5 ​ ​ ​Homogenization of Quasi-Perfect Materials Perfect materials (which represent the periodic setting) are rare in nature. However, there is a huge class of materials which have small deviation from perfect ones, called here quasi-perfect materials. We consider in this section an interesting context, which is the small random perturbation of the periodic setting. In particular, this context is important for numerical applications. To begin, we remember the reader that in the previous section, it was seen that our homogenization analysis (see Theorem 4.2) of the equation (1.1) rely on the spectral study of the operator $L^{\Phi}(\theta)(\theta\in\mathbb{R}^{n})$ posed in the dual space ${\mathcal{H}^{\ast}}$ and with domain ­${D(L^{\Phi}(\theta))=\mathcal{H}}$ and defined by $\begin{array}[]{l}L^{\Phi}(\theta)[f]:=-{\big{(}{\rm div}_{\\!z}+2i\pi\theta\big{)}}{\Big{[}A{\big{(}\Phi^{-1}(\cdot,{\cdot\cdot}),{\cdot\cdot}\big{)}}{\big{(}\nabla_{\\!\\!z}+2i\pi\theta\big{)}}f{\big{(}\Phi^{-1}(\cdot,{\cdot\cdot}),{\cdot\cdot}\big{)}}\Big{]}}\\\\[10.0pt] \hskip 113.81102pt+\,V{\big{(}\Phi^{-1}(\cdot,{\cdot\cdot}),{\cdot\cdot}\big{)}}f{\big{(}\Phi^{-1}(\cdot,{\cdot\cdot}),{\cdot\cdot}\big{)}},\end{array}$ (5.107) where $\Phi:\mathbb{R}^{n}\times\Omega\to\mathbb{R}^{n}$ is a stochastic deformation, $A:\mathbb{R}^{n}\times\Omega\to\mathbb{R}^{n^{2}}$ and $V:\mathbb{R}^{n}\times\Omega\to\mathbb{R}$ are stationary functions. Also, remember that the variational formulation of the operator $L^{\Phi}(\theta)$ is given by: $\begin{split}&{\left\langle L^{\Phi}(\theta)[f],g\right\rangle}:=\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}A{\left(\Phi^{-1}(z,\omega),\omega\right)}{\big{(}\nabla_{\\!\\!z}+2i\pi\theta\big{)}}f{\left(\Phi^{-1}(z,\omega),\omega\right)}\cdot\\\ &\hskip 199.16928pt\overline{{\big{(}\nabla_{\\!\\!z}+2i\pi\theta\big{)}}g{\left(\Phi^{-1}(z,\omega),\omega\right)}}\,dz\,d\mathbb{P}(\omega)\\\ &+\int_{\Omega}\int_{\Phi([0,1)^{n},\omega)}V{\left(\Phi^{-1}(z,\omega),\omega\right)}f{\left(\Phi^{-1}(z,\omega),\omega\right)}\,\overline{g{\left(\Phi^{-1}(z,\omega),\omega\right)}}\,dz\,d\mathbb{P}(\omega),\end{split}$ for ${f,g\in\mathcal{H}}$. More precisely, it was required the existence of a pair ${{\big{(}\theta^{\ast},\lambda(\theta^{\ast})\big{)}}\in\mathbb{R}^{n}\times\mathbb{R}}$ that satisfies: $\left\\{\,\begin{split}&\lambda(\theta^{\ast})\;\text{is a simple eigenvalue of}\;L^{\Phi}(\theta^{\ast}),\\\ &\theta^{\ast}\;\text{is a critical point of}\;\lambda(\cdot),\,\text{that is},\nabla_{\\!\\!\theta}\lambda(\theta^{\ast})=0.\end{split}\right.$ (5.108) As observed before, it is not clear the existence of a pair $(\theta^{\ast},\lambda(\theta^{\ast}))$, in general stochastic environments, satisfying the two above conditions. The reason is due mainly to the lack of compact embedding of ${\mathcal{H}}$ in ${\mathcal{L}}$. However, in the periodic settings there are concrete situations where such conditions take place (see, for instance, [4, 7, 8]). Our aim in this section is to show realistic models whose spectral nature is inherited from the periodic ones. ### 5.1 Perturbed Periodic Case: Spectral Analysis In this section we shall study the spectral properties of the operator ${L^{\Phi}(\theta)}$, when the diffeomorphism ${\Phi}$ is a stochastic perturbation of the identity. This concept was introduced in [10], and well- developed by T. Andrade, W. Neves, J. Silva [6] for modelling quasi-perfect materials. Let $(\Omega,\mathcal{F},\mathbb{P})$ be a probability space, $\tau:\mathbb{Z}^{n}\times\Omega\to\Omega$ a discrete dynamical system, and $Z$ any fixed stochastic deformation. Then, we consider the concept of stochastic perturbation of the identity given by the following ###### Definition 5.1. Given $\eta\in(0,1)$, let $\Phi_{\eta}:\mathbb{R}^{n}\times\Omega\to\mathbb{R}^{n}$ be a stochastic deformation. Then $\Phi_{\eta}$ is said a stochastic perturbation of the identity, when it can be written as $\Phi_{\eta}(y,\omega)=y+\eta\,Z(y,\omega)+\mathrm{O}(\eta^{2}),$ (5.109) for some stochastic deformation $Z$. We emphasize that the equality (5.109) is understood in the sense of ${\rm Lip}_{\text{loc}}\big{(}\mathbb{R}^{n};L^{2}(\Omega)\big{)}$, i.e. for each bounded open subset ${\mathcal{O}\subset\mathbb{R}^{n}}$, there exist $\delta,C>0$, such that for all ${\eta\in(0,\delta)}$ $\displaystyle\underset{y\in\mathcal{O}}{\rm sup}\,{\left\|\Phi_{\eta}(y,\cdotp)-y-\eta Z(y,\cdotp)\right\|}_{L^{2}(\Omega)}$ $\displaystyle\qquad+\,\underset{y\in\mathcal{O}}{\rm ess\,sup}\,{\left\|\nabla_{\\!\\!y}\Phi_{\eta}(y,\cdotp)-I-\eta\,\nabla_{\\!\\!y}Z(y,\cdotp)\right\|}_{L^{2}(\Omega)}\leqslant C\,\eta^{2}.$ Moreover, after some computations, we have $\left\\{\begin{aligned} \nabla_{y}^{-1}\Phi_{\eta}&=I-\eta\,\nabla_{y}Z+O(\eta^{2}),\\\\[5.0pt] \det\big{(}\nabla_{y}\Phi_{\eta}\big{)}&=1+\eta\,{\rm div}_{y}Z+O(\eta^{2}).\end{aligned}\right.$ (5.110) Now, we consider the periodic functions $A_{\rm per}:\mathbb{R}^{n}\to\mathbb{R}^{n^{2}},\,V_{\rm per}:\mathbb{R}^{n}\to\mathbb{R}$ and $U_{\rm per}:\mathbb{R}^{n}\to\mathbb{R}$, such that * • There exists $a_{0},a_{1}>0$ such that for all $\xi\in\mathbb{R}^{n}$ and for a.e $y\in\mathbb{R}^{n}$ we have $a_{0}|\xi|^{2}\leq A_{\rm per}(y)\xi\cdot\xi\leq a_{1}|\xi|^{2}.$ * • $V_{\rm per},\,U_{\rm per}\in L^{\infty}(\mathbb{R}^{n})$. The following lemma is well-known and it is stated explicitly here only for reference. For a proof, we recommend the reader to [18]. ###### Lemma 5.2. For $\theta\in\mathbb{R}^{n}$ and $f\in H_{\rm per}^{1}([0,1)^{n})$, let ${L_{\rm per}(\theta)}$ be the operator defined by $L_{\rm per}(\theta){[f]}:=-({\rm div}_{\\!y}+2i\pi\theta){\big{[}A_{\rm per}(y){(\nabla_{\\!\\!y}+2i\pi\theta)}f(y)\big{]}}+V_{\rm per}(y)f(y),$ (5.111) with variational formulation $\begin{array}[]{c}\displaystyle{\left\langle L_{\rm per}(\theta){\big{[}f\big{]}},g\right\rangle}:=\int_{[0,1)^{n}}A_{\rm per}(y){\left(\nabla_{\\!\\!y}+2i\pi\theta\right)}f(y)\cdot\overline{{\left(\nabla_{\\!\\!y}+2i\pi\theta\right)}g(y)}\,dy\\\\[10.0pt] \displaystyle\hskip 48.36958pt+\int_{[0,1)^{n}}V_{\rm per}(y)\,f(y)\,\overline{g(y)}\,dy,\end{array}$ for ${f,g\in H_{\rm per}^{1}({[0,1)^{n}})}$. Then ${L_{\rm per}(\theta)}$ has the following properties: 1. (i) There exist ${\gamma_{0},b_{0}>0}$, such that ${L_{\gamma_{0}}:=L_{\rm per}(\theta)+{\gamma_{0}}I}$ satisfies for all $f\in H_{\rm per}^{1}({[0,1)^{n}})$, ${\langle L_{\gamma_{0}}{\big{[}f\big{]}},f\rangle}\geq b_{0}{\|f\|}_{H_{\rm per}^{1}({[0,1)^{n}})}^{2}.$ 2. (ii) The point spectrum of ${L_{\rm per}(\theta)}$ is not empty and their eigenspaces have finite dimension, that is, the set $\sigma_{\rm point}{\big{(}L_{\rm per}(\theta)\big{)}}=\\{\lambda\in\mathbb{C}\;;\;\lambda\;\text{an eigenvalue of}\;L_{\rm per}(\theta)\\}$ is not empty and for all ${\lambda\in\sigma_{\rm point}{\big{(}L_{\rm per}(\theta)\big{)}}}$ fixed, ${\rm dim}{\big{\\{}f\in H^{1}_{\rm per}({[0,1)^{n}})\;;\;L_{\rm per}(\theta){\big{[}f\big{]}}=\lambda f\big{\\}}}<\infty.$ 3. (iii) Every point in ${\sigma_{\rm point}\big{(}L_{\rm per}(\theta)\big{)}}$ is isolated. ###### Remark 5.3. We observe that, the properties of the ${L_{\rm per}(\theta)}$, ${\theta\in\mathbb{R}^{n}}$, given by the Lemma 5.2 can be conveyed to the space ${\mathcal{H}}$ in a natural way. In whats follow, we are interested in the study of spectral properties of the operator ${L^{\Phi_{\eta}}(\theta)}$ whose variational formulation is given by $\begin{split}&{\left\langle L^{\Phi_{\eta}}(\theta)[f],g\right\rangle}:=\int_{\Omega}\int_{\Phi_{\eta}([0,1)^{n},\omega)}A_{\rm per}{\left(\Phi_{\eta}^{-1}(z,\omega)\right)}{\big{(}\nabla_{\\!\\!z}+2i\pi\theta\big{)}}f{\left(\Phi_{\eta}^{-1}(z,\omega),\omega\right)}\cdot\\\ &\hskip 199.16928pt\overline{{\big{(}\nabla_{\\!\\!z}+2i\pi\theta\big{)}}g{\left(\Phi_{\eta}^{-1}(z,\omega),\omega\right)}}\,dz\,d\mathbb{P}(\omega)\\\ &+\int_{\Omega}\int_{\Phi_{\eta}([0,1)^{n},\omega)}V_{\rm per}{\left(\Phi_{\eta}^{-1}(z,\omega)\right)}f{\left(\Phi_{\eta}^{-1}(z,\omega),\omega\right)}\,\overline{g{\left(\Phi_{\eta}^{-1}(z,\omega),\omega\right)}}\,dz\,d\mathbb{P}(\omega),\end{split}$ (5.112) for ${f,g\in\mathcal{H}}$. As we shall see in the next theorem, some of the spectral properties of the operator ${L^{\Phi_{\eta}}(\theta)}$ are inherited from the periodic case. ###### Theorem 5.4. Let ${\Phi_{\eta}}$, ${\eta\in(0,1)}$ be a stochastic perturbation of identity and ${\theta_{0}\in\mathbb{R}^{n}}$. If ${\lambda_{0}}$ is an eigenvalue of ${L_{\rm per}(\theta_{0})}$ with multiplicity ${k_{0}\in\mathbb{N}}$, that is, ${\rm dim}{\big{\\{}f\in H^{1}_{\rm per}([0,1)^{n})\;;\;L_{\rm per}(\theta_{0}){\big{[}f\big{]}}=\lambda_{0}f\big{\\}}}=k_{0},$ then there exist a neighbourhood ${\mathcal{U}}$ of ${(0,\theta_{0})}$, ${k_{0}}$ real analytic functions $(\eta,\theta)\in\mathcal{U}\;\mapsto\;\lambda_{k}(\eta,\theta)\in\mathbb{R},\;\;k\in\\{1,\ldots,k_{0}\\},$ and ${k_{0}}$ vector-value analytic maps $(\eta,\theta)\in\mathcal{U}\;\mapsto\;\psi_{k}(\eta,\theta)\in\mathcal{H}\setminus\\{0\\},\;\;k\in\\{1,\ldots,k_{0}\\},$ such that, for all ${k\in\\{1,\ldots,k_{0}\\}}$, * (i) ${\lambda_{k}(0,\theta_{0})=\lambda_{0}}$, * (ii) ${L^{\Phi_{\eta}}(\theta){\big{[}\psi_{k}(\eta,\theta)\big{]}}=\lambda_{k}(\eta,\theta)\,\psi_{k}(\eta,\theta)}$, ${\forall(\eta,\theta)\in\mathcal{U}}$, * (iii) ${{\rm dim}{\big{\\{}f\in\mathcal{H}\;;\;L^{\Phi_{\eta}}(\theta){\big{[}f\big{]}}=\lambda_{k}(\eta,\theta)f\big{\\}}}\leqslant k_{0}}$, ${\forall(\eta,\theta)\in\mathcal{U}}$. ###### Proof. 1\. The aim of this step is to rewrite the operator ${L^{\Phi_{\eta}}(\theta)\in\mathcal{B}(\mathcal{H},\mathcal{H}^{\ast})}$, for ${\eta\in(0,1)}$ and ${\theta\in\mathbb{R}^{n}}$ as an expansion in the variable ${(\eta,\theta)}$ of operators in ${\mathcal{B}(\mathcal{H},\mathcal{H}^{\ast})}$ around the point ${(\eta,\theta)=(0,\theta_{0})}$. For this, using the variational formulation (5.112), a change of variables and the expansions (5.110) we obtain $\begin{split}&\\!\\!\\!{\langle L^{\Phi_{\eta}}(\theta){\big{[}f\big{]}},g\rangle}=\\\ &{\left[\int_{[0,1)^{n}}\int_{\Omega}A_{\rm per}(y){\left(\nabla_{\\!\\!y}+2i\pi\theta\right)}f\cdot\overline{{\left(\nabla_{\\!\\!y}+2i\pi\theta\right)}g}\,d\mathbb{P}\,dy+\int_{[0,1)^{n}}\int_{\Omega}V_{\rm per}(y)\,f\,\overline{g}\,d\mathbb{P}\,dy\right]}\\\ &\hskip 7.11317pt+\eta{\left[\int_{[0,1)^{n}}\int_{\Omega}A_{\rm per}(y){\left(-[\nabla_{\\!\\!y}Z](y,\omega)\nabla_{\\!\\!y}f\right)}\cdot\overline{{\left(\nabla_{\\!\\!y}+2i\pi\theta\right)}g}\,d\mathbb{P}\,dy\right.}\\\ &\hskip 42.67912pt{+\int_{[0,1)^{n}}\int_{\Omega}A_{\rm per}(y){\left(\nabla_{\\!\\!y}+2i\pi\theta\right)}f\cdot\overline{{\left(-[\nabla_{\\!\\!y}Z](y,\omega)\nabla_{\\!\\!y}g\right)}}\,d\mathbb{P}\,dy}\\\ &\hskip 59.75095pt{\left.+\int_{[0,1)^{n}}\int_{\Omega}A_{\rm per}(y){\left(\nabla_{\\!\\!y}+2i\pi\theta\right)}f\cdot\overline{{\left(\nabla_{\\!\\!y}+2i\pi\theta\right)}g}\,\,{\rm div}_{\\!y}Z(y,\omega)\,d\mathbb{P}\,dy\right]}\\\ &\hskip 177.82971pt+\mathrm{O}(\eta^{2}),\end{split}$ in $\mathbb{C}$ as ${\eta\to 0}$, for ${f,g\in\mathcal{H}}$. Now, making the expansion of it in the variable ${\theta}$ about the point ${\theta=\theta_{0}}$, it is convenient to rewrite the above expansion in the form $\begin{split}&\\!{\left\langle L^{\Phi_{\eta}}(\theta){\big{[}f\big{]}},g\right\rangle}=\\\ &((\eta,\theta)-(0,\theta_{0}))^{(0,\boldsymbol{0})}{\left(\int_{[0,1)^{n}}\int_{\Omega}A_{\rm per}(y){(\nabla_{\\!\\!y}+2i\pi\theta_{0})}f\cdot\overline{{(\nabla_{\\!\\!y}+2i\pi\theta_{0})}g}\,d\mathbb{P}\,dy\right.}\\\ &\hskip 256.0748pt{\left.+\int_{[0,1)^{n}}\int_{\Omega}V_{\rm per}(y)\,f\,\overline{g}\,d\mathbb{P}\,dy\right)}\\\ &+\sum_{k=1}^{n}((\eta,\theta)-(0,\theta_{0}))^{(0,e_{k})}{\left(\int_{[0,1)^{n}}\int_{\Omega}A_{\rm per}(y){(\nabla_{\\!\\!y}+2i\pi\theta_{0})}f\cdot\overline{(2i\pi e_{k}g)}\,d\mathbb{P}\,dy\right.}\\\ &\hskip 149.37697pt{\left.+\int_{[0,1)^{n}}\int_{\Omega}A_{\rm per}(y)(2i\pi e_{k}f)\cdot\overline{{(\nabla_{\\!\\!y}+2i\pi\theta_{0})}g}\,d\mathbb{P}\,dy\right)}\end{split}$ $\begin{split}&+\sum_{k,\ell=1}^{n}((\eta,\theta)-(0,\theta_{0}))^{(0,e_{k}+e_{\ell})}{\left(\int_{[0,1)^{n}}\int_{\Omega}A_{\rm per}(y)(2i\pi e_{k}f)\cdot\overline{(2i\pi e_{\ell}g)}\,d\mathbb{P}\,dy\right)}\\\ &+((\eta,\theta)-(0,\theta_{0}))^{(1,\boldsymbol{0})}{\left(\int_{[0,1)^{n}}\int_{\Omega}A_{\rm per}(y){\left(-[\nabla_{\\!\\!y}Z](y,\omega)\nabla_{\\!\\!y}f\right)}\cdot\overline{{(\nabla_{\\!\\!y}+2i\pi\theta_{0})}g}\,d\mathbb{P}\,dy\right.}\\\ &\hskip 61.17325pt+\int_{[0,1)^{n}}\int_{\Omega}A_{\rm per}(y){(\nabla_{\\!\\!y}+2i\pi\theta_{0})}f\cdot\overline{{\left(-[\nabla_{\\!\\!y}Z](y,\omega)\nabla_{\\!\\!y}g\right)}}\,d\mathbb{P}\,dy\\\ &\hskip 61.17325pt{\left.+\int_{[0,1)^{n}}\int_{\Omega}A_{\rm per}(y){(\nabla_{\\!\\!y}+2i\pi\theta_{0})}f\cdot\overline{{(\nabla_{\\!\\!y}+2i\pi\theta_{0})}g}\,\,{\rm div}_{\\!y}Z(y,\omega)\,d\mathbb{P}\,dy\right)}\\\ &+\sum_{k=1}^{n}((\eta,\theta)-(0,\theta_{0}))^{(1,e_{k})}{\left(\int_{[0,1)^{n}}\int_{\Omega}A_{\rm per}(y){\left(-[\nabla_{\\!\\!y}Z](y,\omega)\nabla_{\\!\\!y}f\right)}\cdot\overline{(2i\pi e_{k}g)}\,d\mathbb{P}\,dy\right.}\\\ &\hskip 85.35826pt+\int_{[0,1)^{n}}\int_{\Omega}A_{\rm per}(y){(2i\pi e_{k}f)}\cdot\overline{{\left(-[\nabla_{\\!\\!y}Z](y,\omega)\nabla_{\\!\\!y}g\right)}}\,d\mathbb{P}\,dy\\\ &\hskip 85.35826pt+\int_{[0,1)^{n}}\int_{\Omega}A_{\rm per}(y){(\nabla_{\\!\\!y}+2i\pi\theta_{0})}f\cdot\overline{(2i\pi e_{k}g)}\,\,{\rm div}_{\\!y}Z(y,\omega)\,d\mathbb{P}\,dy\\\ &\hskip 85.35826pt{\left.+\int_{[0,1)^{n}}\int_{\Omega}A_{\rm per}(y){(2i\pi e_{k}f)}\cdot\overline{{(\nabla_{\\!\\!y}+2i\pi\theta_{0})}g}\,\,{\rm div}_{\\!y}Z(y,\omega)\,d\mathbb{P}\,dy\right)}\\\ &+\sum_{k,\ell=1}^{n}((\eta,\theta)-(0,\theta_{0}))^{(1,e_{k}+e_{\ell})}{\left(\int_{[0,1)^{n}}\int_{\Omega}A_{\rm per}(y){(2i\pi e_{k}f)}\cdot\overline{(2i\pi e_{\ell}g)}\,\,{\rm div}_{\\!y}Z(y,\omega)\,d\mathbb{P}\,dy\right)}\\\ &\hskip 177.82971pt+\mathrm{O}(\eta^{2}),\end{split}$ in ${\mathbb{C}}$ as ${\eta\to 0}$, for ${f,g\in\mathcal{H}}$, which is the expansion in the variable ${(\eta,\theta)}$ around the point ${(0,\theta_{0})}$. Here, for ${(\alpha,\beta)\in\mathbb{N}\times\mathbb{N}^{n}}$ and ${\beta=(\beta_{1},\ldots,\beta_{n})}$, we are using the multi-index notation ${((\eta,\theta)-(0,\theta_{0}))^{(\alpha,\beta)}=\eta^{\alpha}\prod_{k=1}^{n}(\theta_{k}-\theta_{0k})^{\beta_{k}}}$. Now, noting that the term of order ${(\eta,\theta)^{(0,\boldsymbol{0})}}$ is the variational formulation of ${L_{\rm per}(\theta_{0})}$ as in (5.111), we can rewrite the above expansion in the form $L^{\Phi_{\eta}}(\theta)=L_{\rm per}(\theta_{0})+\sum_{{|(\alpha,\beta)|}=1}^{3}((\eta,\theta)-(0,\theta_{0}))^{(\alpha,\beta)}L_{(\alpha,\beta)}+\mathrm{O}(\eta^{2}),$ (5.113) in ${\mathcal{B}(\mathcal{H},\mathcal{H}^{\ast})}$ as ${\eta\to 0}$, where ${L_{(\alpha,\beta)}\in\mathcal{B}(\mathcal{H},\mathcal{H}^{\ast})}$ and ${{|(\alpha,\beta)|}=\alpha+\sum_{k=1}^{n}\beta_{k}}$. Clearly, we can consider the parameters ${(\eta,\theta)}$ in the set $B(0,1)\times\mathbb{C}^{n}$. 2\. In this step, we shall modify the expansion (5.113) conveniently in order to obtain an holomorphic invertible operator in the variable ${(\eta,\theta)}$. For this, remember that according to the item ${(i)}$ in Lemma 5.2, there exists $\gamma_{0}>0$ such that the operator ${L_{\rm per}(\theta_{0})+{\gamma_{0}}I}$ is invertible. Then there exists ${\delta>0}$ such that the expansion $L^{\Phi_{\eta}}(\theta)+{\gamma_{0}}I=(L_{\rm per}(\theta_{0})+{\gamma_{0}}I)+\sum_{{|(\alpha,\beta)|}=1}^{3}((\eta,\theta)-(0,\theta_{0}))^{(\alpha,\beta)}L_{(\alpha,\beta)}+\mathrm{O}(\eta^{2}),$ (5.114) in ${\mathcal{B}(\mathcal{H},\mathcal{H}^{\ast})}$ as ${\eta\to 0}$, is invertible for all ${(\eta,\theta)\in B(0,\delta)\times B(\theta_{0},\delta)}$, since the set of invertible bounded operators ${GL(\mathcal{H},\mathcal{H}^{\ast})}$ is an open subset of ${\mathcal{B}(\mathcal{H},\mathcal{H}^{\ast})}$. Now, we denote by ${S(\eta,\theta)}$ the inverse operator of ${L^{\Phi_{\eta}}(\theta)+{\gamma_{0}}I}$, ${(\eta,\theta)\in B(0,\delta)\times B(\theta_{0},\delta)}$. Since the map ${L\in GL(\mathcal{H},\mathcal{H}^{\ast})\mapsto L^{-1}\in\mathcal{B}(\mathcal{H}^{\ast},\mathcal{H})}$ is continuous, the map $(\eta,\theta)\in B(0,\delta)\times B(\theta_{0},\delta)\mapsto S(\eta,\theta)\in\mathcal{B}(\mathcal{H}^{\ast},\mathcal{H})$ is continuous. As a consequence of this, for ${(\widetilde{\eta},\widetilde{\theta})\in B(0,\delta)\times B(\theta_{0},\delta)}$ fixed, the limit of $\frac{S(\eta,\widetilde{\theta})-S(\widetilde{\eta},\widetilde{\theta})}{\eta-\widetilde{\eta}}=-S(\eta,\widetilde{\theta}){\left[\frac{(L^{\Phi_{\eta}}(\widetilde{\theta})+{\gamma_{0}}I)-(L^{\Phi_{\widetilde{\eta}}}(\widetilde{\theta})+{\gamma_{0}}I)}{\eta-\widetilde{\eta}}\right]}S(\widetilde{\eta},\widetilde{\theta}),$ as ${\widetilde{\eta}\not=\eta\to 0}$, exists. Thus, ${\eta\in B(0,\delta)\mapsto S(\eta,\widetilde{\theta})}$ is an holomorphic map. In analogy with it, for ${j\in{\\{1,\ldots,n\\}}}$, we can prove that $\theta_{j}\mapsto S(\widetilde{\eta},\widetilde{\theta}_{1},\ldots,\widetilde{\theta}_{j-1},\theta_{j},\widetilde{\theta}_{j+1},\ldots,\widetilde{\theta}_{n})$ is an holomorphic map. Therefore, by Osgood’s Lemma, see for instance [22], we conclude that $(\eta,\theta)\in B(0,\delta)\times B(\theta_{0},\delta)\mapsto S(\eta,\theta)\in\mathcal{B}(\mathcal{H}^{\ast},\mathcal{H})$ (5.115) is a holomorphic function. 3\. Finally, we are in conditions to prove items ${(i)}$, ${(ii)}$ and ${(iii)}$ (the spectral analysis of the operator ${S(\eta,\theta)}$). First, we shall note that for ${(\eta,\theta)}$ in a neighbourhood of ${(0,\theta_{0})}$, the map ${(\eta,\theta)\mapsto S(\eta,\theta)}$ satisfies the assumptions of the Theorem 2.26. We begin recalling that the restriction operator ${T\in\mathcal{B}(\mathcal{H}^{\ast},\mathcal{H})\mapsto T\big{|}_{\mathcal{L}}\in\mathcal{B}(\mathcal{L},\mathcal{L})}$ is continuous and it satisfies ${\|T\|}_{\mathcal{B}(\mathcal{L},\mathcal{L})}\leqslant{\|T\|}_{\mathcal{B}(\mathcal{H}^{\ast},\mathcal{H})}\,,\;\,\forall T\in\mathcal{B}(\mathcal{H}^{\ast},\mathcal{H}).$ (5.116) Then, by (5.115), the map ${(\eta,\theta)\in B(0,\delta)\times B(\theta_{0},\delta)\mapsto S(\eta,\theta)\in\mathcal{B}(\mathcal{L},\mathcal{L})}$ is holomorphic. Since holomorphic maps are, locally, analytic maps there exists a neighbourhood ${\mathcal{U}}$ of ${(0,\theta_{0})}$, ${(0,\theta_{0})\in\mathcal{U}\subset\mathbb{C}\times\mathbb{C}^{n}}$, and a family ${\\{S_{\sigma}\\}_{\sigma\in\mathbb{N}\times\mathbb{N}^{n}}}$ contained in ${\mathcal{B}(\mathcal{L},\mathcal{L})}$, such that $S(\eta,\theta)=S_{0}+\sum_{\begin{subarray}{c}\sigma\in\mathbb{N}\times\mathbb{N}^{n}\\\ {|\sigma|}\neq 0\end{subarray}}(\eta,\theta)^{\sigma}S_{\sigma}\,,\;\,\forall(\eta,\theta)\in\mathcal{U}.$ (5.117) Using (5.114) and (5.117), it is easy to see that ${S_{0}=(L_{\rm per}(\theta_{0})+{\gamma_{0}}I)^{-1}\big{|}_{\mathcal{L}}}$. Notice also that ${\mu_{0}:={\left(\lambda_{0}+\gamma_{0}\right)}^{-1}}$ is an eigenvalue of ${S_{0}}$ if and only if ${\lambda_{0}}$ is an eigenvalue of ${L_{\rm per}(\theta_{0})}$ that is $g\in{\\{f\in\mathcal{L}\;;\;S_{0}{\big{[}f\big{]}}=\mu_{0}f\\}}\;\Leftrightarrow\;g\in{\\{f\in\mathcal{L}\;;\;L_{\rm per}(\theta_{0}){\big{[}f\big{]}}=\lambda_{0}f\\}}.$ The final part of the proof is a direct application of the Theorem 2.26. Due to our assumption, $\mu_{0}$ is a real eigenvalue of the operator $S_{0}$ with multiplicity $k_{0}$. Hence, by the Theorem 2.26, there exists a neighbourhood ${\widetilde{\mathcal{U}}}$ of ${(0,\theta_{0})}$, with ${\widetilde{\mathcal{U}}\subset\mathcal{U}}$ and analytic maps $\begin{array}[]{l}(\eta,\theta)\in\widetilde{\mathcal{U}}\;\longmapsto\;\mu_{01}(\eta,\theta),\mu_{02}(\eta,\theta),\ldots,\mu_{0k_{0}}(\eta,\theta)\in(0,\infty),\\\\[5.0pt] (\eta,\theta)\in\widetilde{\mathcal{U}}\;\longmapsto\;\psi_{01}(\eta,\theta),\psi_{02}(\eta,\theta),\ldots,\psi_{0k_{0}}(\eta,\theta)\in\mathcal{L}-\\{0\\},\end{array}$ such that * • ${\mu_{0\ell}(0,\theta_{0})=\mu_{0}}$, * • ${S(\eta,\theta){\big{[}\psi_{0\ell}(\eta,\theta)\big{]}}=\mu_{0\ell}(\eta,\theta)\psi_{0\ell}(\eta,\theta)}$, ${\forall(\eta,\theta)\in\widetilde{\mathcal{U}}}$, * • ${{\rm dim}{\\{f\in\mathcal{L}\;;\;S(\eta,\theta){\big{[}f\big{]}}=\mu_{0\ell}(\eta,\theta)f\\}}\leqslant k_{0}}$, ${\forall(\eta,\theta)\in\widetilde{\mathcal{U}}}$, for all ${\ell\in\\{1,\ldots,k_{0}\\}}$. Thus, the proof of the item ${(i)}$ is clear. Using the second equality above, we obtain $\displaystyle(L^{\Phi_{\eta}}(\theta)+{\gamma_{0}}I){\big{[}\psi_{0\ell}(\eta,\theta)\big{]}}$ $\displaystyle=$ $\displaystyle\frac{1}{\mu_{0\ell}(\eta,\theta)}(L^{\Phi_{\eta}}(\theta)+{\gamma_{0}}I){\left\\{S(\eta,\theta){\big{[}\psi_{0\ell}(\eta,\theta)\big{]}}\right\\}}$ $\displaystyle=$ $\displaystyle\frac{1}{\mu_{0\ell}(\eta,\theta)}\psi_{0\ell}(\eta,\theta),$ which implies that ${L^{\Phi_{\eta}}(\theta){\big{[}\psi_{0\ell}(\eta,\theta)\big{]}}=\lambda_{0\ell}(\eta,\theta)\psi_{0\ell}(\eta,\theta)}$, for ${(\eta,\theta)\in\widetilde{\mathcal{U}}}$, ${\ell\in\\{1,\ldots,m_{0}\\}}$ and ${\lambda_{0\ell}(\eta,\theta):=[\mu_{0\ell}(\eta,\theta)]^{-1}-{\gamma_{0}}}$. This finish the proof of the item ${(ii)}$. Finally, note that ${S(\eta,\theta)\big{[}\mathcal{L}\big{]}\subset\mathcal{H}}$ and $g\\!\in\\!\big{\\{}f\in\mathcal{H};S(\eta,\theta){\big{[}f\big{]}}\\!=\\!\mu_{0\ell}(\eta,\theta)f\big{\\}}\Leftrightarrow g\\!\in\\!\big{\\{}f\in\mathcal{H}\;;\;L^{\Phi_{\eta}}(\theta){\big{[}f\big{]}}\\!=\\!\lambda_{0\ell}(\eta,\theta)f\big{\\}},$ which concludes the proof of the item ${(iii)}$. Therefore, the proof is completed. ∎ ### 5.2 Homogenization Analysis of the Perturbed Model In this section, we shall investigate in which way the stochastic perturbation of the identity characterize the form of the coefficients, during the asymptotic limit of the Schrödinger equation $\left\\{\begin{array}[]{l}i\displaystyle\frac{\partial u_{\eta\varepsilon}}{\partial t}-{\rm div}{\bigg{(}A_{\rm per}{\left(\Phi_{\eta}^{-1}{\left(\frac{x}{\varepsilon},\omega\right)}\right)}\nabla u_{\eta\varepsilon}\bigg{)}}\\\\[14.0pt] +{\bigg{(}\displaystyle\frac{1}{\varepsilon^{2}}V_{\rm per}{\left(\Phi_{\eta}^{-1}{\left(\displaystyle\frac{x}{\varepsilon},\omega\right)}\right)}+U_{\rm per}{\left(\Phi_{\eta}^{-1}{\left(\displaystyle\frac{x}{\varepsilon},\omega\right)}\right)}\bigg{)}}u_{\eta\varepsilon}=0\quad\text{in}\;\,\mathbb{R}^{n+1}_{T}\\!\times\\!\Omega,\\\\[14.0pt] u_{\eta\varepsilon}(0,x,\omega)=u_{\eta\varepsilon}^{0}(x,\omega),\;\;(x,\omega)\in\mathbb{R}^{n}\\!\times\\!\Omega,\end{array}\right.$ (5.118) where ${0<T<\infty}$, ${\mathbb{R}^{n+1}_{T}=(0,T)\times\mathbb{R}^{n}}$. The coefficients are accomplishing of the periodic functions ${A_{\rm per}(y)}$, ${V_{\rm per}(y)}$, ${U_{\rm per}(y)}$ (as defined in the last subsection) with a stochastic perturbation of identity ${\Phi_{\eta}}$, ${\eta\in(0,1)}$, presenting an rate of oscillation ${\varepsilon^{-1}}$, ${\varepsilon>0}$. The function ${u_{\eta\varepsilon}^{0}(x,\omega)}$ is a well prepared initial data(see (5.126)) and this well-preparedness is triggered by natural periodic conditions, to wit, on the existence of a pair ${\big{(}\theta^{\ast},\lambda_{\rm per}(\theta^{\ast})\big{)}\in\mathbb{R}^{n}\times\mathbb{R}}$ such that $\begin{split}(i)&\;\;\,\lambda_{\rm per}(\theta^{\ast})\;\text{is a simple eigenvalue of}\;L_{\rm per}(\theta^{\ast}),\\\ (ii)&\;\;\,\theta^{\ast}\;\text{is a critical point of}\;\lambda_{\rm per}(\cdot),\,\text{that is},\nabla_{\\!\\!\theta}\lambda_{\rm per}(\theta^{\ast})=0.\end{split}$ (5.119) By the condition ${(i)}$ and the Theorem 5.4, there exists a neighborhood ${\mathcal{U}}$ of ${(0,\theta^{\ast})}$ and the analytic maps $\begin{split}(i)&\;\;\,(\eta,\theta)\in\mathcal{U}\;\mapsto\;\lambda(\eta,\theta)\in\mathbb{R},\\\ (ii)&\;\;\,(\eta,\theta)\in\mathcal{U}\;\mapsto\;\psi(\eta,\theta)\in\mathcal{H}\setminus\\{0\\},\end{split}$ (5.120) such that ${\lambda(0,\theta^{\ast})=\lambda_{\rm per}(\theta^{\ast})}$, ${L^{\Phi_{\eta}}(\theta)\big{[}\psi(\eta,\theta)\big{]}=\lambda(\eta,\theta)\,\psi(\eta,\theta)}$ and ${\rm dim}\big{\\{}f\in\mathcal{H}\;;\;L^{\Phi_{\eta}}(\theta)=\lambda(\eta,\theta)\,f\big{\\}}=1,\;\forall(\eta,\theta)\in\mathcal{U}.$ Thus, $\lambda(\eta,\theta)\;\text{is a simple eigenvalue of}\;L^{\Phi_{\eta}}(\theta),\forall(\eta,\theta)\in\mathcal{U}.$ (5.121) Additionally, as ${\lambda(0,\theta^{\ast})=\lambda_{\rm per}(\theta^{\ast})}$ is an isolated point of ${\sigma_{\rm point}\big{(}L_{\rm per}(\theta^{\ast})\big{)}}$ (any point has this property ), ${\lambda(\eta,\theta)}$ is an isolated point of ${\sigma_{\rm point}\big{(}L^{\Phi_{\eta}}(\theta^{\ast})\big{)}}$ for each ${(\eta,\theta)\in\mathcal{U}}$. Thus, we have ${\lambda(0,\cdot)=\lambda_{\rm per}(\cdot)}$ in a neighbourhood of ${\theta^{\ast}}$. We now denote ${\psi_{\rm per}(\cdot):=\psi(0,\cdot)}$. Without loss of generality, we assume ${\int_{[0,1)^{n}}{|\psi_{\rm per}(\theta^{\ast})|}^{2}dy=1}$. Moreover, we shall assume that the homogenized (periodic) matrix ${A_{\rm per}^{\ast}=D_{\\!\theta}^{2}\lambda_{\rm per}(\theta^{\ast})}$ is invertible which happens if $\theta=\theta^{\ast}$ is a point of local minimum or local maximum strict of $\mathbb{R}^{n}\ni\theta\mapsto\lambda_{\rm per}(\theta)$. Thus, an immediate application of the Implicit Function Theorem gives us the following lemma: ###### Lemma 5.5. Let the condition (5.119) be satisfied and ${A_{\rm per}^{\ast}}$ be an invertible matrix. Then, there exists a neighborhood ${\mathcal{V}}$ of ${0}$, ${0\in\mathcal{V}\subset\mathbb{R}}$, and a ${\mathbb{R}^{n}}$-value analytic map $\theta(\cdot):\eta\in\mathcal{V}\mapsto\theta(\eta)\in\mathbb{R}^{n},$ such that ${\theta(0)=\theta^{\ast}}$ and $\nabla_{\\!\\!\theta}\lambda\big{(}\eta,\theta(\eta)\big{)}=0,\;\;\forall\eta\in\mathcal{V}.$ (5.122) By the analytic structure of the functions in (5.120) and the Lemma 5.5, there exists a neighborhood ${\mathcal{V}}$ of ${0}$, ${0\in\mathcal{V}\subset\mathbb{R}}$, such that $\begin{split}(i)&\;\;\,\eta\in\mathcal{V}\;\mapsto\;\lambda\big{(}\eta,\theta(\eta)\big{)}\in\mathbb{R},\\\ (ii)&\;\;\,\eta\in\mathcal{V}\;\mapsto\;\psi\big{(}\eta,\theta(\eta)\big{)}\in\mathcal{H}\setminus\\{0\\},\\\ (iii)&\;\;\,\eta\in\mathcal{V}\;\mapsto\;\xi_{k}\big{(}\eta,\theta(\eta)\big{)}\in\mathcal{H},\forall\\{1,\ldots,n\\},\end{split}$ (5.123) are analytic functions, where ${\xi_{k}(\eta,\theta):=(2i\pi)^{-1}{\partial_{\theta_{k}}\psi}(\eta,\theta)}$, for ${k\in\\{1,\ldots,n\\}}$. We also consider ${\xi_{k,{\rm per}}(\cdot)=\xi_{k}(0,\cdot)}$. Furthermore, by (5.121) and (5.122), for each fixed ${\eta\in\mathcal{V}}$ we have that the pair ${\big{(}\theta(\eta),\lambda\big{(}\eta,\theta(\eta)\big{)}\big{)}\in\mathbb{R}^{n}\times\mathbb{R}}$ satisfies: $\begin{split}(i)&\;\;\,\lambda(\eta,\theta(\eta))\;\text{is a simple eigenvalue of}\;L^{\Phi_{\eta}}\big{(}\theta(\eta)\big{)},\\\ (ii)&\;\;\,\theta(\eta)\;\text{is a critical point of}\;\lambda(\eta,\cdot),\,\text{that is},\nabla_{\\!\\!\theta}\lambda(\eta,\theta(\eta))=0.\end{split}$ (5.124) This means that the Theorem 4.2 can be used. Before, we establish a much simplified notations for the functions in (5.123) as follows: $\begin{split}(i)&\;\;\,\theta_{\eta}:=\theta(\eta),\\\ (ii)&\;\;\,\lambda_{\eta}:=\lambda\big{(}\eta,\theta(\eta)\big{)},\\\ (iii)&\;\;\,\psi_{\eta}:=\psi\big{(}\eta,\theta(\eta)\big{)},\\\ (iv)&\;\;\,\xi_{k,\eta}:=\xi_{k}\big{(}\eta,\theta(\eta)\big{)},\,k\in\\{1,\ldots,n\\}.\end{split}$ (5.125) Finally, from (5.124), for each fixed ${\eta\in\mathcal{V}}$, the notion of well-preparedness for the initial data $u_{\eta\varepsilon}^{0}$ is given as below. $u_{\eta\varepsilon}^{0}(x,\omega)=e^{2i\pi\frac{\theta_{\eta}\cdot x}{\varepsilon}}\,v^{0}(x)\,\psi_{\eta}{\left(\Phi_{\eta}^{-1}{\left(\frac{x}{\varepsilon},\omega\right)},\omega\right)},\;(x,\omega)\in\mathbb{R}^{n}\times\Omega,$ (5.126) where ${v^{0}\in C_{\rm c}^{\infty}(\mathbb{R}^{n})}$. Thus, applying the Theorem 4.2, if ${u_{\eta\varepsilon}}$ is solution of (5.118), the sequence in ${\varepsilon>0}$ $v_{\eta\varepsilon}(t,x,\widetilde{\omega})=e^{-{\left(i\frac{\lambda_{\eta}t}{\varepsilon^{2}}+2i\pi\frac{\theta_{\eta}\cdot x}{\varepsilon}\right)}}u_{\eta\varepsilon}(t,x,\widetilde{\omega}),\;\,(t,x,\widetilde{\omega})\in\mathbb{R}^{n+1}_{T}\times\Omega,$ $\Phi_{\omega}-$two-scale converges to the limit ${v_{\eta}(t,x)\,\psi_{\eta}{\big{(}\Phi^{-1}(z,\omega),\omega\big{)}}}$ with $\lim_{\varepsilon\to 0}\iint_{\mathbb{R}^{n+1}_{T}}\\!{\left|v_{\eta\varepsilon}(t,x,\widetilde{\omega})-v_{\eta}(t,x)\,\psi_{\eta}{\left(\Phi^{-1}_{\eta}{\left(\frac{x}{\varepsilon},\widetilde{\omega}\right)},\widetilde{\omega}\right)}\right|}^{2}dx\,dt\,=\,0,$ for a.e. ${\widetilde{\omega}\in\Omega}$, where ${v_{\eta}\in C\big{(}[0,T],L^{2}(\mathbb{R}^{n})\big{)}}$ is the unique solution of the homogenized Schrödinger equation $\left\\{\begin{array}[]{c}i\displaystyle\frac{\partial v_{\eta}}{\partial t}-{\rm div}{\left(A^{\ast}_{\eta}\nabla v_{\eta}\right)}+U_{\\!\eta}^{\ast}v_{\eta}=0\,,\;\,\text{in}\;\,\mathbb{R}^{n+1}_{T},\\\\[7.5pt] v_{\eta}(0,x)=v^{0}(x)\,,\;\,x\in\mathbb{R}^{n},\end{array}\right.$ (5.127) with effective coefficients ${A^{\ast}_{\eta}=D_{\\!\theta}^{2}\lambda\big{(}\eta,\theta(\eta)\big{)}}$ and $U^{\ast}_{\\!\eta}=c^{-1}_{\eta}\int_{\Omega}\int_{\Phi_{\eta}([0,1)^{n},\omega)}U_{\rm per}{\big{(}\Phi^{-1}_{\eta}(z,\omega)\big{)}}{\left|\psi_{\eta}{\big{(}\Phi^{-1}_{\eta}(z,\omega),\omega\big{)}}\right|}^{2}dz\,d\mathbb{P}(\omega),$ (5.128) where $c_{\eta}=\int_{\Omega}\int_{\Phi_{\eta}([0,1)^{n},\omega)}{\left|\psi_{\eta}{\big{(}\Phi^{-1}_{\eta}(z,\omega),\omega\big{)}}\right|}^{2}dz\,d\mathbb{P}(\omega).$ (5.129) ###### Remark 5.6. We remember that, using the equality (4.1), we have for each ${\eta}$ fixed that the matrix ${B_{\eta}\in\mathbb{R}^{n\times n}}$ must satisfy $\begin{split}&(B_{\eta})_{k\ell}:=c_{\eta}^{-1}\bigg{[}\int_{\Omega}\int_{\Phi_{\eta}([0,1)^{n},\omega)}A_{\rm per}{\left(\Phi_{\eta}^{-1}(z,\omega)\right)}{\left(e_{\ell}\,\psi_{\eta}{\left(\Phi_{\eta}^{-1}(z,\omega),\omega\right)}\right)}\cdot\\\ &\hskip 223.3543pt\overline{\left(e_{k}\,\psi_{\eta}{\left(\Phi_{\eta}^{-1}(z,\omega),\omega\right)}\right)}\,dz\,d\mathbb{P}(\omega)\\\ &+\int_{\Omega}\int_{\Phi_{\eta}([0,1)^{n},\omega)}A_{\rm per}{\left(\Phi_{\eta}^{-1}(z,\omega)\right)}{\left(e_{\ell}\,\psi_{\eta}{\left(\Phi_{\eta}^{-1}(z,\omega),\omega\right)}\right)}\cdot\\\ &\hskip 170.71652pt\overline{{\left(\nabla_{\\!\\!z}+2i\pi\theta_{\eta}\right)}{\left(\xi_{k,\eta}{\left(\Phi_{\eta}^{-1}(z,\omega),\omega\right)}\right)}}\,dz\,d\mathbb{P}(\omega)\\\ &-\int_{\Omega}\int_{\Phi_{\eta}([0,1)^{n},\omega)}A_{\rm per}{\left(\Phi_{\eta}^{-1}(z,\omega)\right)}{\left(\nabla_{\\!\\!z}+2i\pi\theta_{\eta}\right)}{\left(\psi_{\eta}{\left(\Phi_{\eta}^{-1}(z,\omega),\omega\right)}\right)}\cdot\\\ &\hskip 221.93158pt\overline{{\left(e_{\ell}\,\xi_{k,\eta}{\left(\Phi_{\eta}^{-1}(z,\omega),\omega\right)}\right)}}\,dz\,d\mathbb{P}(\omega)\bigg{]},\end{split}$ (5.130) for ${k,\ell\in\\{1,\ldots,n\\}}$ and the homogenized matrix can be written as ${A_{\eta}^{\ast}=2^{-1}{\big{(}B_{\eta}+B_{\eta}^{t}\big{)}}}$. #### 5.2.1 Expansion of the effective coefficients As a consequence of the formula of the effective coefficients of the homogenized equation (5.127), we have the following proposition: ###### Proposition 5.7. The maps ${\eta\mapsto A_{\eta}^{\ast}\in\mathbb{R}^{n\times n}}$, ${\eta\mapsto B_{\eta}\in\mathbb{R}^{n\times n}}$ and ${\eta\mapsto U_{\eta}^{\ast}\in\mathbb{R}}$ are analytics in a neighbourhood of ${\eta=0}$. ###### Proof. Let us assume ${\mathcal{U}}$ and ${\mathcal{V}}$ as in (5.120) and (5.123), respectively. For each ${\eta\in\mathcal{V}}$, the above arguments give us the formula ${A^{\ast}_{\eta}=D_{\\!\theta}^{2}\lambda\big{(}\eta,\theta(\eta)\big{)}}$. Thus, as ${(\eta,\theta)\in\mathcal{U}\mapsto D_{\\!\theta}^{2}\lambda(\eta,\theta)\in\mathbb{R}^{n\times n}}$ and ${\eta\in\mathcal{V}\mapsto\theta(\eta)\in\mathbb{R}^{n}}$ are analytic maps, we conclude that ${\eta\in\mathcal{V}\mapsto D_{\theta}^{2}\lambda(\eta,\theta(\eta))\in\mathbb{R}^{n\times n}}$ is also an analytic map. This means that ${\eta\mapsto A_{\eta}^{\ast}}$ is an analytic map. From (5.128) and (5.129), making a change of variables, we have $U^{\ast}_{\\!\eta}=c^{-1}_{\eta}\int_{\Omega}\int_{[0,1)^{n}}U_{\rm per}(y){\left|\psi_{\eta}(y,\omega)\right|}^{2}{\rm det}[\nabla_{\\!\\!y}\Phi_{\eta}(y,\omega)]\,dz\,d\mathbb{P}(\omega)$ and $c_{\eta}=\int_{\Omega}\int_{[0,1)^{n}}{\left|\psi_{\eta}(y,\omega)\right|}^{2}{\rm det}[\nabla_{\\!\\!y}\Phi_{\eta}(y,\omega)]\,dz\,d\mathbb{P}(\omega)\not=0.$ Then, as the map ${\eta\mapsto\psi_{\eta}\in\mathcal{H}\setminus\\{0\\}}$ is analytic, the map ${\eta\mapsto c_{\eta}\not=0}$ is also analytic. Hence the map ${\eta\mapsto c_{\eta}^{-1}}$ is analytic. Therefore, ${\eta\mapsto U_{\eta}^{\ast}}$ is analytic. ∎ As a consequence of this proposition, there exist ${\\{A^{(j)},\,B^{(j)}\\}_{j\in\mathbb{N}}\subset\mathbb{R}^{n\times n}}$ and ${\\{U^{(j)}\\}_{j\in\mathbb{N}}\subset\mathbb{R}}$ such that $\left\\{\begin{array}[]{lll}A_{\eta}^{\ast}&=&A^{(0)}+\eta A^{(1)}+\eta^{2}A^{(2)}+\ldots,\\\\[5.0pt] U_{\eta}^{\ast}&=&U^{(0)}+\eta U^{(1)}+\eta^{2}U^{(2)}+\ldots,\\\\[5.0pt] B_{\eta}^{\ast}&=&B^{(0)}+\eta B^{(1)}+\eta^{2}B^{(2)}+\ldots.\end{array}\right.$ (5.131) Now, the object of our interest is determine the terms of order ${\eta^{0}}$ and ${\eta}$ of these homogenized coefficients. For this purpose, guided by 5.110 and by the formulas (5.130), (5.128) and (5.129), we shall analyse the expansion of the analytic functions in (5.125). By analytic property, there exist the sequences ${{\\{\theta^{(j)}\\}}_{j\in\mathbb{N}}\subset\mathbb{R}^{n}}$, ${{\\{\lambda^{(j)}\\}}_{j\in\mathbb{N}}\subset\mathbb{R}}$, ${{\\{\psi^{(j)}\\}}_{j\in\mathbb{N}}\subset\mathcal{H}}$ and ${{\\{\xi_{k}^{(j)}\\}}_{j\in\mathbb{N}}\subset\mathcal{H}}$, ${k\in\\{1,\ldots,n\\}}$, such that $\displaystyle\theta_{\eta}$ $\displaystyle=$ $\displaystyle\theta^{(0)}+\eta\theta^{(1)}+\eta^{2}\theta^{(2)}+\ldots=\theta^{(0)}+\eta\theta^{(1)}+\mathrm{O}(\eta^{2}),$ (5.132) $\displaystyle\lambda_{\eta}$ $\displaystyle=$ $\displaystyle\lambda^{(0)}+\eta\lambda^{(1)}+\eta^{2}\lambda^{(2)}+\ldots=\lambda^{(0)}+\eta\lambda^{(1)}+\mathrm{O}(\eta^{2}),$ (5.133) $\displaystyle\psi_{\eta}$ $\displaystyle=$ $\displaystyle\psi^{(0)}+\eta\psi^{(1)}+\eta^{2}\psi^{(2)}+\ldots=\psi^{(0)}+\eta\psi^{(1)}+\mathrm{O}(\eta^{2}),$ (5.134) $\displaystyle\xi_{k,\eta}$ $\displaystyle=$ $\displaystyle\xi_{k}^{(0)}+\eta\xi_{k}^{(1)}+\eta^{2}\xi_{k}^{(2)}+\ldots=\xi_{k}^{(0)}+\eta\xi_{k}^{(1)}+\mathrm{O}(\eta^{2}),$ (5.135) where ${k\in\\{1,\ldots,n\\}}$. At first glance, in order to determine the coefficients of the expansions in (5.131) we should solve, a priori, auxiliary problems that involves both, the deterministic and stochastic variables. This can be a disadvantage from the point of view of numerical analysis. Our aim hereafter is to prove that, we can simplify the computations of this coefficients working in a periodic environment which is computationally cheaper. In order to do this, note that ${\theta^{(0)}=\theta^{\ast}}$, ${\lambda^{(0)}=\lambda_{\rm per}(\theta^{\ast})}$, ${\psi^{(0)}=\psi_{\rm per}(\theta^{\ast})}$ and ${\xi_{k}^{(0)}=\xi_{k,{\rm per}}(\theta^{\ast})}$, ${k\in\\{1,\ldots,n\\}}$, which satisfy $\displaystyle\left\\{\begin{array}[]{l}{\left(L_{\rm per}(\theta^{\ast})-\lambda_{\rm per}(\theta^{\ast})\right)}{\big{[}\psi_{\rm per}(\theta^{\ast})\big{]}}=0\;\,\text{in}\;\,[0,1)^{n},\\\\[6.0pt] \hskip 42.67912pt\psi_{\rm per}(\theta^{\ast})\;\;[0,1)^{n}\text{-periodic},\end{array}\right.$ (5.138) $\displaystyle\left\\{\begin{array}[]{l}{\left(L_{\rm per}(\theta^{\ast})-\lambda_{\rm per}(\theta^{\ast})\right)}{\big{[}\xi_{k,{\rm per}}(\theta^{\ast})\big{]}}=\mathcal{X}{\big{[}\psi_{\rm per}(\theta^{\ast})\big{]}}\;\,\text{in}\;\,[0,1)^{n},\\\\[6.0pt] \hskip 42.67912pt\xi_{k,{\rm per}}(\theta^{\ast})\;\;[0,1)^{n}\text{-periodic},\end{array}\right.$ (5.141) where $\mathcal{X}{\big{[}f\big{]}}:={\left({\rm div}_{\\!y}+2i\pi\theta^{\ast}\right)}{\big{\\{}A_{\rm per}(y){(e_{k}f)}\big{\\}}}+{(e_{k})}{\big{\\{}A_{\rm per}(y){(\nabla_{\\!\\!y}+2i\pi\theta^{\ast})}f\big{\\}}},$ for ${f\in\mathcal{H}}$. The equation (5.138) is the spectral cell equation and (5.141) is the first auxiliary cell equation related to the periodic case (see Section 3.3). The following theorem show us that, the terms ${\psi^{(1)}}$ and ${\xi_{k}^{(1)}}$, $k\in\\{1,\ldots,n\\}$, given by (5.134) and (5.135) respectively, satisfy auxiliary type cell equations. ###### Theorem 5.8. Let ${\psi^{(1)}}$ and ${\xi_{k}^{(1)}}$, ${k\in\\{1,\ldots,n\\}}$, be as above. Then these functions satisfy the following equations: $\displaystyle\left\\{\begin{array}[]{l}{\left(L_{\rm per}(\theta^{\ast})-\lambda_{\rm per}(\theta^{\ast})\right)}{\big{[}\psi^{(1)}\big{]}}=\mathcal{Y}{\big{[}\psi_{\rm per}(\theta^{\ast})\big{]}}\;\,\text{in}\;\,[0,1)^{n}\times\Omega,\\\\[6.0pt] \hskip 42.67912pt\psi^{(1)}\;\text{stationary},\end{array}\right.$ (5.144) $\displaystyle\left\\{\begin{array}[]{l}{\left(L_{\rm per}(\theta^{\ast})-\lambda_{\rm per}(\theta^{\ast})\right)}{\big{[}\xi_{k}^{(1)}\big{]}}=\mathcal{X}{\big{[}\psi^{(1)}\big{]}}\\\\[6.0pt] \hskip 56.9055pt+\,\mathcal{Y}{\big{[}\xi_{k,{\rm per}}(\theta^{\ast})\big{]}}+\mathcal{Z}_{k}{\big{[}\psi_{\rm per}(\theta^{\ast})\big{]}}\;\,\text{in}\;\,[0,1)^{n}\times\Omega,\\\\[6.0pt] \hskip 42.67912pt\xi_{k}^{(1)}\;\text{stationary},\end{array}\right.$ (5.148) where the operators ${\mathcal{Y}}$ and ${\mathcal{Z}_{k}}$, ${k\in\\{1,\ldots,n\\}}$, are defined by $\displaystyle\mathcal{Y}{\big{[}f\big{]}}$ $\displaystyle\\!\\!:=\\!\\!$ $\displaystyle{\left({\rm div}_{\\!y}+2i\pi\theta^{\ast}\right)}{\big{\\{}A_{\rm per}(y){\big{(}-[\nabla_{\\!\\!y}Z](y,\omega)\nabla_{\\!\\!y}f+2i\pi\theta^{(1)}f\big{)}}\big{\\}}}$ $\displaystyle-\,{\rm div}_{\\!y}{\big{\\{}[\nabla_{\\!\\!y}Z]^{t}(y,\omega)A_{\rm per}(y){(\nabla_{\\!\\!y}+2i\pi\theta^{\ast})}f\big{\\}}}$ $\displaystyle+{\left(2i\pi\theta^{(1)}\right)}{\big{\\{}A_{\rm per}(y){\left(\nabla_{\\!\\!y}+2i\pi\theta^{\ast}\right)}f\big{\\}}}$ $\displaystyle+\,{\left({\rm div}_{\\!y}+2i\pi\theta^{\ast}\right)}{\big{\\{}{\left[{\rm div}_{\\!y}Z(y,\omega)A_{\rm per}(y)\right]}{(\nabla_{\\!\\!y}+2i\pi\theta^{\ast})}f\big{\\}}}+\lambda^{(1)}f$ $\displaystyle+\,{\big{\\{}{\rm div}_{\\!y}Z(y,\omega)\,{\left[\lambda_{\rm per}(\theta^{\ast})-V_{\rm per}(y)\right]}\big{\\}}}f,$ $\displaystyle\mathcal{Z}_{k}{\big{[}f\big{]}}$ $\displaystyle\\!\\!:=\\!\\!$ $\displaystyle{\left({\rm div}_{\\!y}+2i\pi\theta^{\ast}\right)}{\big{\\{}{\left[{\rm div}_{\\!y}Z(y,\omega)A_{\rm per}(y)\right]}{(e_{k}f)}\big{\\}}}$ $\displaystyle-\,{\rm div}_{\\!y}{\big{\\{}[\nabla_{\\!\\!y}Z]^{t}(y,\omega)A_{\rm per}(y){(e_{k}f)}\big{\\}}}+{\left(2i\pi\theta^{(1)}\right)}{\left\\{A_{\rm per}(y){(e_{k}f)}\right\\}}$ $\displaystyle+\,{\left(e_{k}\right)}{\big{\\{}{\big{[}{\rm div}_{\\!y}Z(y,\omega)A_{\rm per}(y)\big{]}}{\left(\nabla_{\\!\\!y}+2i\pi\theta^{\ast}\right)}f\big{\\}}}$ $\displaystyle-\,{\left(e_{k}\right)}{\left\\{A_{\rm per}(y){\left[\nabla_{\\!\\!y}Z\right]}(y,\omega)\nabla_{\\!\\!y}f\right\\}}+{\left(e_{k}\right)}{\big{\\{}A_{\rm per}(y){(2i\pi\theta^{(1)}f)}\big{\\}}},$ for ${f\in\mathcal{H}}$. For the proof of this theorem, we shall use essentially the structure of the spectral cell equation (3.56) and of the f.a.c. equation (3.3) with periodic coefficients accomplished by stochastic deformation of identity ${\Phi_{\eta}}$ together with the identities (5.110). ###### Proof. 1\. For begining, let us consider the set ${\mathcal{V}}$ as in (5.123). Then, making change of variables in the spectral cell equation (3.56) adapted to this context, we find $\displaystyle\int_{[0,1)^{n}}\int_{\Omega}\Big{\\{}A_{\rm per}(y)\big{(}[\nabla_{\\!\\!y}\Phi_{\eta}]^{-1}\nabla_{\\!\\!y}\psi_{\eta}+2i\pi\theta_{\eta}\psi_{\eta}\big{)}\cdot\overline{\big{(}[\nabla_{\\!\\!y}\Phi_{\eta}]^{-1}\nabla_{\\!\\!y}\zeta+2i\pi\theta_{\eta}\zeta\big{)}}\,$ $\displaystyle\qquad\qquad\qquad+{\left(V_{\rm per}(y)-\lambda_{\eta}\right)}\,\psi_{\eta}\,\overline{\zeta}\,\Big{\\}}{\rm det}[\nabla_{\\!\\!y}\Phi_{\eta}]\,d\mathbb{P}(\omega)\,dy=0,$ (5.149) for all ${\eta\in\mathcal{V}}$ and ${\zeta\in\mathcal{H}}$. If we insert the equations (5.110), (5.132), (5.133) and (5.134) in equation (5.2.1) and compute the term $\eta$, we arrive at $\begin{array}[]{l}\displaystyle{\int_{[0,1)^{n}}\\!\int_{\Omega}\\!\Big{\\{}A_{\rm per}(y){\left(\nabla_{\\!\\!y}\psi^{(1)}+2i\pi\theta^{\ast}\psi^{(1)}\right)}\cdot\overline{{\left(\nabla_{\\!\\!y}\zeta+2i\pi\theta^{\ast}\zeta\right)}}+{\left(V_{\rm per}(y)-\lambda_{\rm per}(\theta^{\ast})\right)}\psi^{(1)}\,\overline{\zeta}}\\\\[15.0pt] \displaystyle\qquad\qquad\qquad+\,\\!A_{\rm per}(y){\left(-[\nabla_{\\!\\!y}Z]\,\nabla_{\\!\\!y}\psi_{\rm per}(\theta^{\ast})+2i\pi\theta^{(1)}\psi_{\rm per}(\theta^{\ast})\right)}\\!\cdot\\!\overline{{\left(\nabla_{\\!\\!y}\zeta+2i\pi\theta^{\ast}\zeta\right)}}\\\\[15.0pt] \displaystyle\qquad\qquad\qquad+A_{\rm per}(y){\left(\nabla_{\\!\\!y}\psi_{\rm per}(\theta^{\ast})+2i\pi\theta^{\ast}\psi_{\rm per}(\theta^{\ast})\right)}\cdot\overline{{\left(-[\nabla_{\\!\\!y}Z]\nabla_{\\!\\!y}\zeta+2i\pi\theta^{(1)}\zeta\right)}}\\\\[15.0pt] \displaystyle\qquad\qquad\qquad+A_{\rm per}(y){\left(\nabla_{\\!\\!y}\psi_{\rm per}(\theta^{\ast})+2i\pi\theta^{\ast}\psi_{\rm per}(\theta^{\ast})\right)}\cdot\overline{{\left(\nabla_{\\!\\!y}\zeta+2i\pi\theta^{\ast}\zeta\right)}}\,{\rm div}_{\\!y}Z\\\\[15.0pt] \displaystyle\qquad\qquad\qquad-\lambda^{(1)}\,\psi_{\rm per}(\theta^{\ast})\,\overline{\zeta}+{\left(V_{\rm per}(y)-\lambda_{\rm per}(\theta^{\ast})\right)}\psi^{(0)}\,\overline{\zeta}\,{\rm div}_{\\!y}Z\Big{\\}}\,d\mathbb{P}(\omega)\,dy=0,\end{array}$ for all ${\eta\in\mathcal{V}}$ and ${\zeta\in\mathcal{H}}$. This equation is the variational formulation of the equation (5.144), which concludes the first part of the proof. 2\. For the second part of the proof, we proceed similarly with respect to the f.a.c. equation (3.3) and obtain $\begin{split}&\displaystyle\int_{[0,1)^{n}}\int_{\Omega}\Big{\\{}A_{\rm per}(y)\big{(}[\nabla_{\\!\\!y}\Phi_{\eta}]^{-1}\nabla_{\\!\\!y}\xi_{k,\eta}+2i\pi\theta_{\eta}\xi_{k,\eta}\big{)}\cdot\overline{\big{(}[\nabla_{\\!\\!y}\Phi_{\eta}]^{-1}\nabla_{\\!\\!y}\zeta+2i\pi\theta_{\eta}\zeta\big{)}}\\\ &\displaystyle\qquad\qquad\qquad+A_{\rm per}(y){\left(e_{k}\,\psi_{\eta}\right)}\cdot\overline{\big{(}[\nabla_{\\!\\!y}\Phi_{\eta}]^{-1}\nabla_{\\!\\!y}\zeta+2i\pi\theta_{\eta}\zeta\big{)}}\\\ &\displaystyle\qquad\qquad\qquad-A_{\rm per}(y)\big{(}[\nabla_{\\!\\!y}\Phi_{\eta}]^{-1}\nabla_{\\!\\!y}\psi_{\eta}+2i\pi\theta_{\eta}\psi_{\eta}\big{)}\cdot\overline{{\left(e_{k}\,\zeta\right)}}\\\ &\displaystyle\qquad\quad+{\left(V_{\rm per}(y)-\lambda_{\eta}\right)}\,\xi_{k,\eta}\,\overline{\zeta}-\,\frac{1}{2i\pi}\frac{\partial\lambda}{\partial\theta_{k}}(\eta,\theta(\eta))\,\psi_{\eta}\,\overline{\zeta}\Big{\\}}\,{\rm det}[\nabla_{\\!\\!y}\Phi_{\eta}]\,d\mathbb{P}(\omega)\,dy=0,\end{split}$ (5.150) for all ${\eta\in\mathcal{V}}$, ${\zeta\in\mathcal{H}}$ and ${k\in\\{1,\ldots,n\\}}$. Hence, taking into account the Lemma 5.5 and inserting the equations (5.110), (5.132), (5.133), (5.134) and (5.135) in equation (5.150), a computation of the term of order ${\eta}$ lead us to $\begin{array}[]{l}\displaystyle\int_{[0,1)^{n}}\int_{\Omega}\Big{\\{}A_{\rm per}(y)\big{(}\nabla_{\\!\\!y}\xi_{k}^{(1)}+2i\pi\theta^{\ast}\xi_{k}^{(1)}\big{)}\cdot\overline{\big{(}\nabla_{\\!\\!y}\zeta+2i\pi\theta^{\ast}\zeta\big{)}}+{\left(V_{\rm per}(y)-\lambda_{\rm per}(\theta^{\ast})\right)}\xi_{k}^{(1)}\,\overline{\zeta}\\\\[15.0pt] \displaystyle\qquad\qquad+A_{\rm per}(y){\left(e_{k}\,\psi^{(1)}\right)}\cdot\overline{\big{(}\nabla_{\\!\\!y}\zeta+2i\pi\theta^{\ast}\zeta\big{)}}-A_{\rm per}(y)\big{(}\nabla_{\\!\\!y}\psi^{(1)}+2i\pi\theta^{\ast}\psi^{(1)}\big{)}\cdot\overline{{\left(e_{k}\,\zeta\right)}}\\\\[15.0pt] \displaystyle\qquad\qquad+A_{\rm per}(y)\big{(}-[\nabla_{\\!\\!y}Z]\nabla_{\\!\\!y}\xi_{k,{\rm per}}(\theta^{\ast})+2i\pi\theta^{(1)}\xi_{k,{\rm per}}(\theta^{\ast})\big{)}\cdot\overline{\big{(}\nabla_{\\!\\!y}\zeta+2i\pi\theta^{\ast}\zeta\big{)}}\\\\[15.0pt] \displaystyle\qquad\qquad+A_{\rm per}(y)\big{(}\nabla_{\\!\\!y}\xi_{k,{\rm per}}(\theta^{\ast})+2i\pi\theta^{\ast}\xi_{k,{\rm per}}(\theta^{\ast})\big{)}\cdot\overline{\big{(}-[\nabla_{\\!\\!y}Z]\nabla_{\\!\\!y}\zeta+2i\pi\theta^{(1)}\zeta\big{)}}\\\\[15.0pt] \displaystyle\qquad\qquad+A_{\rm per}(y)\big{(}\nabla_{\\!\\!y}\xi_{k,{\rm per}}(\theta^{\ast})+2i\pi\theta^{\ast}\xi_{k,{\rm per}}(\theta^{\ast})\big{)}\cdot\overline{\big{(}\nabla_{\\!\\!y}\zeta+2i\pi\theta^{\ast}\zeta\big{)}}\,{\rm div}_{\\!y}Z\\\\[15.0pt] \displaystyle\qquad\qquad\qquad\qquad-\lambda^{(1)}\,\xi_{k,{\rm per}}(\theta^{\ast})\,\overline{\zeta}+{\left(V_{\rm per}(y)-\lambda_{\rm per}(\theta^{\ast})\right)}\xi_{k,{\rm per}}(\theta^{\ast})\,\overline{\zeta}\,{\rm div}_{\\!y}Z\\\\[15.0pt] \displaystyle\qquad\qquad\qquad+\,A_{\rm per}(y){\left(e_{k}\,\psi_{\rm per}(\theta^{\ast})\right)}\cdot\Big{(}\overline{\big{(}\nabla_{\\!\\!y}\zeta+2i\pi\theta^{\ast}\zeta\big{)}\,{\rm div}_{\\!y}Z-[\nabla_{\\!\\!y}Z]\nabla_{\\!\\!y}\zeta+2i\pi\theta^{(1)}\zeta}\Big{)}\\\\[15.0pt] \displaystyle\qquad\qquad\qquad\qquad\qquad\qquad-A_{\rm per}(y)\big{(}\nabla_{\\!\\!y}\psi_{\rm per}(\theta^{\ast})+2i\pi\theta^{\ast}\psi_{\rm per}(\theta^{\ast})\big{)}\cdot\overline{{\left(e_{k}\,\zeta\right)}}\,{\rm div}_{\\!y}Z\\\\[15.0pt] \displaystyle\qquad\qquad-\,A_{\rm per}(y)\big{(}-[\nabla_{\\!\\!y}Z]\,\nabla_{\\!\\!y}\psi_{\rm per}(\theta^{\ast})+2i\pi\theta^{(1)}\psi_{\rm per}(\theta^{\ast})\big{)}\cdot\overline{{\left(e_{k}\,\zeta\right)}}\Big{\\}}\,d\mathbb{P}(\omega)\,dy=0,\end{array}$ for all ${\zeta\in\mathcal{H}}$. Noting that this is the variational formulation of the equation (5.148), we conclude the proof. ∎ We remember the reader that if $f:\mathbb{R}^{n}\times\Omega\to\mathbb{R}$ is a stationary function, then we shall use the following notation $\mathbb{E}[f(x,\cdot)]=\int_{\Omega}f(x,\omega)\,d\mathbb{P}(\omega),$ for any $x\in\mathbb{R}^{n}$. Roughly speaking, the theorem below tell us that the homogenized matrix of the problem (5.118) can be obtained by solving periodic problems. ###### Theorem 5.9. Let ${A_{\eta}^{\ast}}$ be the homogenized matrix as in (5.127). Then $A_{\eta}^{\ast}=A_{\rm per}^{\ast}+\eta A^{(1)}+\mathrm{O}(\eta^{2}).$ Moreover, the term of order ${\eta^{0}}$ is given by the homogenized matrix of the periodic case, that is, ${A_{\rm per}^{\ast}=2^{-1}{\left(B^{(0)}+(B^{(0)})^{t}\right)}}$, where the matrix ${B^{(0)}}$ is the term of order ${\eta^{0}}$ in (5.131) and it is defined by $\begin{array}[]{r}\displaystyle(B^{(0)})_{k\ell}:=\int_{[0,1)^{n}}A_{\rm per}(y){(e_{\ell}\,\psi_{\rm per}(\theta^{\ast}))}\cdot{(e_{k}\,\overline{\psi_{\rm per}(\theta^{\ast})})}\,dy\\\\[10.0pt] \displaystyle+\,\int_{[0,1)^{n}}A_{\rm per}(y){(e_{\ell}\,\psi_{\rm per}(\theta^{\ast}))}\cdot\overline{\left(\nabla_{\\!\\!y}\xi_{k,{\rm per}}(\theta^{\ast})+2i\pi\theta^{\ast}\xi_{k,{\rm per}}(\theta^{\ast})\right)}\,dy\\\\[10.0pt] \displaystyle-\,\int_{[0,1)^{n}}A_{\rm per}(y){\Big{(}{\left(\nabla_{\\!\\!y}\psi_{\rm per}(\theta^{\ast})+2i\pi\theta^{\ast}\psi_{\rm per}(\theta^{\ast})\right)}\Big{)}}\cdot\overline{{(e_{\ell}\,\xi_{k,{\rm per}}(\theta^{\ast}))}}\,dy.\end{array}$ The term of order ${\eta}$ is given by ${A^{(1)}=2^{-1}{\left(B^{(1)}+(B^{(1)})^{t}\right)}}$, where the matrix ${B^{(1)}}$ is the term of order ${\eta}$ in (5.131) and it is defined by $\begin{array}[]{l}\displaystyle(B^{(1)})_{k\ell}={{\bigg{[}\int_{[0,1)^{n}}A_{\rm per}(y){(e_{\ell}\,\psi_{\rm per}(\theta^{\ast}))}\cdot{(e_{k}\,\overline{\psi_{\rm per}(\theta^{\ast})})}\,\mathbb{E}\Big{[}{\rm div}_{\\!y}Z(y,\cdot)\Big{]}\,dy}}\\\\[11.0pt] \displaystyle+\,\int_{[0,1)^{n}}A_{\rm per}(y){(e_{\ell}\,\psi_{\rm per}(\theta^{\ast}))}\cdot{(e_{k}\,\overline{\mathbb{E}\big{[}\psi^{(1)}(y,\cdot)\big{]}})}\,dy\\\\[11.0pt] \displaystyle+\int_{[0,1)^{n}}A_{\rm per}(y){\left(e_{\ell}\,{\mathbb{E}\big{[}\psi^{(1)}(y,\cdot)\big{]}}\right)}\cdot{(e_{k}\,\overline{\psi_{\rm per}(\theta^{\ast})})}\,dy\\\\[11.0pt] \displaystyle+\int_{[0,1)^{n}}A_{\rm per}(y){(e_{\ell}\,\psi_{\rm per}(\theta^{\ast}))}\cdot\overline{\left(\nabla_{\\!\\!y}\xi_{k,{\rm per}}(\theta^{\ast})+2i\pi\theta^{\ast}\xi_{k,{\rm per}}(\theta^{\ast})\right)}\,{\mathbb{E}\big{[}{\rm div}_{\\!y}Z(y,\cdot)\big{]}}\,dy\\\\[11.0pt] \displaystyle+\int_{[0,1)^{n}}A_{\rm per}(y){(e_{\ell}\,\psi_{\rm per}(\theta^{\ast}))}\cdot\overline{(\nabla_{\\!\\!y}{\mathbb{E}\big{[}\xi_{k}^{(1)}(y,\cdot)\big{]}}+2i\pi\theta^{\ast}{\mathbb{E}\Big{[}\xi_{k}^{(1)}(y,\cdot)\Big{]}}}\\\\[11.0pt] \hskip 113.81102pt\overline{+\,2i\pi\theta^{(1)}\xi_{k,{\rm per}}(\theta^{\ast})-{\mathbb{E}\Big{[}[\nabla_{\\!\\!y}Z](y,\cdot)\Big{]}}\nabla_{\\!\\!y}\xi_{k,{\rm per}}(\theta^{\ast})})\,dy\\\\[11.0pt] \displaystyle+\int_{[0,1)^{n}}\\!\\!\\!A_{\rm per}(y){\left(e_{\ell}\,{\mathbb{E}\Big{[}\psi^{(1)}(y,\cdot)\Big{]}}\right)}\cdot\overline{\left(\nabla_{\\!\\!y}\xi_{k,{\rm per}}(\theta^{\ast})+2i\pi\theta^{\ast}\xi_{k,{\rm per}}(\theta^{\ast})\right)}\,dy\\\\[11.0pt] \displaystyle-\int_{[0,1)^{n}}\\!\\!\\!A_{\rm per}(y){({\left(\nabla_{\\!\\!y}\psi_{\rm per}(\theta^{\ast})+2i\pi\theta^{\ast}\psi_{\rm per}(\theta^{\ast})\right)})}\cdot\overline{{(e_{\ell}\,\xi_{k,{\rm per}}(\theta^{\ast}))}}\,{\mathbb{E}\big{[}{\rm div}_{\\!y}Z(y,\cdot)\big{]}}\,dy\\\\[11.0pt] \displaystyle-\int_{[0,1)^{n}}\\!\\!\\!A_{\rm per}(y){\Big{(}{\left(\nabla_{\\!\\!y}\psi_{\rm per}(\theta^{\ast})+2i\pi\theta^{\ast}\psi_{\rm per}(\theta^{\ast})\right)}\Big{)}}\cdot\overline{{\Big{(}e_{\ell}\,{\mathbb{E}\Big{[}\xi_{k}^{(1)}(y,\cdot)\Big{]}}\Big{)}}}\,dy\end{array}$ $\begin{array}[]{l}\displaystyle\hskip 21.33955pt-\int_{[0,1)^{n}}A_{\rm per}(y){\left(\nabla_{\\!\\!y}{\mathbb{E}\Big{[}\psi^{(1)}(y,\cdot)\Big{]}}+2i\pi\theta^{\ast}{\mathbb{E}\Big{[}\psi^{(1)}(y,\cdot)\Big{]}}\right.}\\\\[10.0pt] \hskip 56.9055pt{{\left.+\,2i\pi\theta^{(1)}\psi_{\rm per}(\theta^{\ast})-{\mathbb{E}\Big{[}[\nabla_{\\!\\!y}Z](y,\cdot)\Big{]}}\nabla_{\\!\\!y}\psi_{\rm per}(\theta^{\ast})\right)}\cdot\overline{{(e_{\ell}\,\xi_{k,{\rm per}}(\theta^{\ast}))}}\,dy\bigg{]}}\\\\[11.0pt] \displaystyle-{\bigg{[}\int_{[0,1)^{n}}{|\psi_{\rm per}(\theta^{\ast})|}^{2}{\mathbb{E}\Big{[}{\rm div}_{\\!y}Z(y,\cdot)\Big{]}}\,dy+\int_{[0,1)^{n}}\psi_{\rm per}(\theta^{\ast})\,\overline{\mathbb{E}\Big{[}\psi^{(1)}(y,\cdot)\Big{]}}}\,dy\\\\[11.0pt] \displaystyle\hskip 21.33955pt{+\int_{[0,1)^{n}}{\mathbb{E}\Big{[}\psi^{(1)}(y,\cdot)\Big{]}}\,\overline{\psi_{\rm per}(\theta^{\ast})}\,dy\bigg{]}}\cdot{\bigg{[}\int_{[0,1)^{n}}A_{\rm per}(y){(e_{\ell}\,\psi_{\rm per}(\theta^{\ast}))}\cdot{(e_{k}\,\overline{\psi_{\rm per}(\theta^{\ast})})}\,dy}\\\\[11.0pt] \displaystyle\hskip 21.33955pt+\int_{[0,1)^{n}}A_{\rm per}(y){(e_{\ell}\,\psi_{\rm per}(\theta^{\ast}))}\cdot\overline{\left(\nabla_{\\!\\!y}\xi_{k,{\rm per}}(\theta^{\ast})+2i\pi\theta^{\ast}\xi_{k,{\rm per}}(\theta^{\ast})\right)}\,dy\\\\[11.0pt] \displaystyle\hskip 21.33955pt-\,{{\int_{[0,1)^{n}}A_{\rm per}(y){\left[{\left(\nabla_{\\!\\!y}\psi_{\rm per}(\theta^{\ast})+2i\pi\theta^{\ast}\psi_{\rm per}(\theta^{\ast})\right)}\right]}\cdot\overline{{(e_{\ell}\,\xi_{k,{\rm per}}(\theta^{\ast}))}}\,dy\bigg{]}}}.\end{array}$ ###### Proof. 1\. Taking into account ${\mathcal{V}}$ as in (5.123), we get from (5.130), for ${\eta\in\mathcal{V}}$, that the homogenized matrix is given by ${A_{\eta}^{\ast}=2^{-1}(B_{\eta}+B_{\eta}^{t})}$. Thus, in order to describe the terms of the expansion of ${A_{\eta}^{\ast}}$, we only need to determine the terms in the expansion of ${B_{\eta}}$. 2\. Using the equations (5.110) and (5.134), the map ${\eta\mapsto c_{\eta}\in(0,+\infty)}$ has an expansion about ${\eta=0}$. Remembering that ${\int_{[0,1)^{n}}{|\psi_{\rm per}(\theta^{\ast})|}^{2}dy=1}$, we have $\begin{split}c_{\eta}^{-1}=&1-\eta{\left[\int_{\Omega}\int_{[0,1)^{n}}{|\psi_{\rm per}(\theta^{\ast})|}^{2}{\rm div}_{\\!y}Z(y,\omega)\,dy\,d\mathbb{P}\right.}\\\ &{\left.+\int_{\Omega}\int_{{[0,1)^{n}}}\psi_{\rm per}(\theta^{\ast})\overline{\psi^{(1)}}\,dy\,d\mathbb{P}+\int_{\Omega}\int_{[0,1)^{n}}\psi^{(1)}\overline{\psi_{\rm per}(\theta^{\ast})}\,dy\,d\mathbb{P}\right]}+\;\mathrm{O}(\eta^{2}),\end{split}$ in ${\mathbb{C}}$ as ${\eta\to 0}$. Thus, using the expansions (5.110), (5.132), (5.134) and (5.135) in the formula (5.130), the computation of the resulting term of order ${\eta^{0}}$ of ${B_{\eta}}$ give us the desired expression for $(B^{(0)})_{k,\ell}$. The same reasoning with a little more computations, which is an exercise that we leave to the reader, allow us to obtain the expression for $(B^{(1)})_{k\ell}$. ∎ ###### Remark 5.10. We next record the observation that the computation of the coefficients of ${A_{\rm per}^{\ast}}$ is performed by solving the equations (5.138) and (5.141), which are equations with periodic boundary conditions. In order to compute the coefficients of ${A^{(1)}}$, we need to know the functions ${\psi^{(1)}}$ and ${\xi_{k}^{(1)}}$, ${k\in\\{1,\ldots,n\\}}$, which are a priori stochastic in nature (see the equations (5.144) and (5.148), respectively). But in Theorem 5.9, it has seen that we only need their expectation values, ${\mathbb{E}\Big{[}\psi^{(1)}(y,\cdot)\Big{]}}$ and ${\mathbb{E}\Big{[}\xi_{k}^{(1)}(y,\cdot)\Big{]}}$, ${k\in\\{1,\ldots,n\\}}$, which are ${[0,1)^{n}}$-periodic functions and, respectively, solutions of the following equations: $\displaystyle\left\\{\begin{array}[]{l}{\Big{(}L_{\rm per}(\theta^{\ast})-\lambda_{\rm per}(\theta^{\ast})\Big{)}}\,{\mathbb{E}\Big{[}\psi^{(1)}(y,\cdot)\Big{]}}=\mathcal{Y}_{\rm per}{\big{[}\psi_{\rm per}(\theta^{\ast})\big{]}}\;\,\text{in}\;\,[0,1)^{n},\\\\[6.0pt] \hskip 56.9055pt{\mathbb{E}\Big{[}\psi^{(1)}(y,\cdot)\Big{]}}\;\text{is $[0,1)^{n}$-periodic},\end{array}\right.$ $\displaystyle\left\\{\begin{array}[]{l}{\left(L_{\rm per}(\theta^{\ast})-\lambda_{\rm per}(\theta^{\ast})\right)}{\mathbb{E}\Big{[}\xi_{k}^{(1)}(y,\cdot)\Big{]}}=\mathcal{X}\Big{[}{\mathbb{E}\Big{[}\psi^{(1)}(y,\cdot)\Big{]}}\Big{]}\\\\[6.0pt] \hskip 49.79231pt+\,\mathcal{Y}_{\rm per}{\big{[}\xi_{k,{\rm per}}(\theta^{\ast})\big{]}}+\mathcal{Z}_{k,{\rm per}}{\big{[}\psi_{\rm per}(\theta^{\ast})\big{]}}\;\,\text{in}\;\,[0,1)^{n},\\\\[6.0pt] \hskip 56.9055pt{\mathbb{E}\Big{[}\xi_{k}^{(1)}(y,\cdot)\Big{]}}\;\text{is $[0,1)^{n}$-periodic},\end{array}\right.$ where $\displaystyle\mathcal{Y}_{\rm per}{\big{[}f\big{]}}$ $\displaystyle\\!\\!:=\\!\\!$ $\displaystyle{\left({\rm div}_{\\!y}+2i\pi\theta^{\ast}\right)}{\Big{\\{}A_{\rm per}(y){\big{(}-\mathbb{E}\Big{[}[\nabla_{\\!\\!y}Z](y,\cdot)\Big{]}\nabla_{\\!\\!y}f+2i\pi\theta^{(1)}f\big{)}}\Big{\\}}}$ $\displaystyle-\,{\rm div}_{\\!y}{\Big{\\{}\mathbb{E}\Big{[}[\nabla_{\\!\\!y}Z](y,\cdot)\Big{]}^{t}A_{\rm per}(y){(\nabla_{\\!\\!y}+2i\pi\theta^{\ast})}f\Big{\\}}}$ $\displaystyle+{\left(2i\pi\theta^{(1)}\right)}{\big{\\{}A_{\rm per}(y){\left(\nabla_{\\!\\!y}+2i\pi\theta^{\ast}\right)}f\big{\\}}}$ $\displaystyle+\,{\left({\rm div}_{\\!y}+2i\pi\theta^{\ast}\right)}{\Big{\\{}{\left[\mathbb{E}\Big{[}{\rm div}_{\\!y}Z(y,\cdot)\Big{]}A_{\rm per}(y)\right]}{(\nabla_{\\!\\!y}+2i\pi\theta^{\ast})}f\Big{\\}}}+\lambda^{(1)}f$ $\displaystyle+\,{\Big{\\{}\mathbb{E}\Big{[}{\rm div}_{\\!y}Z(y,\cdot)\Big{]}\,{\left[\lambda_{\rm per}(\theta^{\ast})-V_{\rm per}(y)\right]}\Big{\\}}}f,$ $\displaystyle\mathcal{Z}_{k,{\rm per}}{\big{[}f\big{]}}$ $\displaystyle\\!\\!:=\\!\\!$ $\displaystyle{\left({\rm div}_{\\!y}+2i\pi\theta^{\ast}\right)}{\Big{\\{}{\left[\mathbb{E}\Big{[}{\rm div}_{\\!y}Z(y,\cdot)\Big{]}A_{\rm per}(y)\right]}{(e_{k}f)}\Big{\\}}}$ $\displaystyle-\,{\rm div}_{\\!y}{\Big{\\{}\mathbb{E}\Big{[}[\nabla_{\\!\\!y}Z](y,\cdot)\Big{]}^{t}A_{\rm per}(y){(e_{k}f)}\Big{\\}}}+{\left(2i\pi\theta^{(1)}\right)}{\left\\{A_{\rm per}(y){(e_{k}f)}\right\\}}$ $\displaystyle+\,{\left(e_{k}\right)}{\Big{\\{}{\Big{[}\mathbb{E}\Big{[}{\rm div}_{\\!y}Z(y,\cdot)\Big{]}A_{\rm per}(y)\Big{]}}{\left(\nabla_{\\!\\!y}+2i\pi\theta^{\ast}\right)}f\Big{\\}}}$ $\displaystyle-\,{\left(e_{k}\right)}{\left\\{A_{\rm per}(y)\mathbb{E}\Big{[}[\nabla_{\\!\\!y}Z](y,\cdot)\Big{]}\nabla_{\\!\\!y}f\right\\}}+{\left(e_{k}\right)}{\big{\\{}A_{\rm per}(y){(2i\pi\theta^{(1)}f)}\big{\\}}}.$ for ${f\in H^{1}_{\rm per}([0,1)^{n})}$. Summing up, the determination of the homogenized coefficients for (1.1) is a stochastic problem in nature. However, when we consider the interesting context of materials which have small deviation from perfect ones (modeled by periodic functions), this problem, in the specific case (5.1) reduces, at the first two orders in $\eta$, to the simpler solution to the two periodic problems above. Both of them are of the same nature. Importantly, note that $Z$ in (5.1) is only present through $\mathbb{E}\Big{[}{\rm div}_{\\!y}Z(y,\cdot)\Big{]}$ and $\mathbb{E}\Big{[}[\nabla_{\\!\\!y}Z](y,\cdot)\Big{]}$. In the theorem below, we assume that the homogenized matrix of the periodic case satisfies the uniform coercive condition, that is, $A_{\rm per}^{\ast}\xi\cdot\xi\geq\Lambda|\xi|^{2},$ for some $\Lambda>0$ and for all $\xi\in\mathbb{R}^{n}$, which has experimental evidence for metals and semiconductors. Therefore, due to Theorem 5.9 the homogenized matrix of the perturbed case ${A_{\eta}^{\ast}}$ has similar property for $\eta\sim 0$. ###### Theorem 5.11. Let ${v_{\eta}}$ be the solution of homogenized equation (5.127). Then $v_{\eta}\Big{(}t,\sqrt{A_{\eta}^{\ast}}\,x\Big{)}=v_{\rm per}\Big{(}t,\sqrt{A_{\rm per}^{\ast}}\,x\Big{)}+\eta\,v^{(1)}\Big{(}t,\sqrt{A_{\rm per}^{\ast}}\,x\Big{)}+\mathrm{O}(\eta^{2}),$ weakly in ${L^{2}(\mathbb{R}^{n}_{T})}$ as ${\eta\to 0}$, that means, $\displaystyle\int_{\mathbb{R}^{n}_{T}}{\Bigg{(}v_{\eta}\Big{(}t,\sqrt{A_{\eta}^{\ast}}\,x\Big{)}-v_{\rm per}\Big{(}t,\sqrt{A_{\rm per}^{\ast}}\,x\Big{)}-\eta\,v^{(1)}\Big{(}t,\sqrt{A_{\rm per}^{\ast}}\,x\Big{)}\Bigg{)}}\,h(t,x)\,dx\,dt$ $\displaystyle\qquad\qquad\qquad=\mathrm{O}(\eta^{2}),$ for each ${h\in L^{2}(\mathbb{R}^{n}_{T})}$, where ${v_{\rm per}}$ is the solution of the periodic homogenized problem $\left\\{\begin{array}[]{c}i\displaystyle\frac{\partial v_{\rm per}}{\partial t}-{\rm div}{\left(A_{\rm per}^{\ast}\nabla v_{\rm per}\right)}+U_{\\!\rm per}^{\ast}v_{\rm per}=0,\;\,\text{in}\;\,\mathbb{R}^{n+1}_{T},\\\\[7.5pt] v_{\rm per}(0,x)=v_{0}(x)\,,\;\,x\in\mathbb{R}^{n},\end{array}\right.$ (5.153) and ${v^{(1)}}$ is the solution of $\left\\{\begin{array}[]{c}i\displaystyle\frac{\partial v^{(1)}}{\partial t}-{\rm div}{\left(A_{\rm per}^{\ast}\nabla v^{(1)}\right)}+U_{\\!\rm per}^{\ast}v^{(1)}={\rm div}{\left(A_{\rm per}^{\ast}\nabla v_{\rm per}\right)}-U^{(1)}v_{\rm per},\;\,\text{in}\;\,\mathbb{R}^{n+1}_{T},\\\\[7.5pt] v^{(1)}(0,x)=v^{1}_{0}(x)\,,\;\,x\in\mathbb{R}^{n},\end{array}\right.$ (5.154) where ${U^{(1)}}$ is the coefficient of the term of order ${\eta}$ of the expansion ${U_{\eta}^{\ast}}$ and $v_{0}^{1}\in C_{c}^{\infty}(\mathbb{R}^{n})$ is given by the limit $v_{0}^{1}\Big{(}\sqrt{A_{\rm per}^{\ast}}\,x\Big{)}:=\lim_{\eta\to 0}\frac{v_{0}\Big{(}\sqrt{A_{\eta}^{\ast}}\,x\Big{)}-v_{0}\Big{(}\sqrt{A_{\rm per}^{\ast}}\,x\Big{)}}{\eta}.$ ###### Proof. 1\. Taking into account the set ${\mathcal{V}}$ as in (5.123), we have for ${\eta\in\mathcal{V}}$ and from the conservation of energy of the homogenized Schrödinger equation (5.127), that the solution ${v_{\eta}:\mathbb{R}^{n}_{T}\to\mathbb{C}}$ satisfies ${\|v_{\eta}\|}_{L^{2}(\mathbb{R}^{n+1}_{T})}=T{\|v_{0}\|}_{L^{2}(\mathbb{R}^{n})},\;\,\forall\eta\in\mathcal{V}.$ Thus, after possible extraction of a subsequence, we have the existence of a function ${v^{(0)}\in L^{2}(\mathbb{R}^{n+1}_{T})}$ such that $v_{\eta}\;\xrightharpoonup[\eta\to 0]{}\;v^{(0)}\;\text{em}\;L^{2}(\mathbb{R}^{n+1}_{T}).$ (5.155) By the variational formulation of the equation (5.127), we find $\begin{array}[]{l}0=\displaystyle i\int_{\mathbb{R}^{n}}v_{0}(x)\,\overline{\varphi}(0,x)\,dx-i\int_{\mathbb{R}^{n}_{T}}v_{\eta}(t,x)\,\frac{\partial\overline{\varphi}}{\partial t}(t,x)\,dx\,dt\\\\[15.0pt] \displaystyle\qquad\qquad+\int_{\mathbb{R}^{n}_{T}}\Bigg{\\{}-{\left\langle A_{\eta}^{\ast}v_{\eta}(t,x),D^{2}{\varphi}(t,x)\right\rangle}+U_{\eta}^{\ast}v_{\eta}(t,x)\,\overline{\varphi}(t,x)\Bigg{\\}}\,dx\,dt,\end{array}$ (5.156) for all ${\varphi\in C_{\rm c}^{1}((-\infty,T))\otimes C_{\rm c}^{2}(\mathbb{R}^{n})}$. Recall that ${{\left\langle P,Q\right\rangle}:={\rm tr}(P\overline{Q}^{t})}$, for ${P,Q}$ in ${\mathbb{C}^{n\times n}}$. Then, using (5.155), the Theorem 5.7, making ${\eta\to 0}$ and invoking the uniqueness property of the equation (5.153), we conclude that ${v^{(0)}=v_{\rm per}}$. 2\. Now, using that ${U_{\eta}^{\ast}=U_{\rm per}^{\ast}+\eta\,U^{(1)}+\mathrm{O}(\eta^{2})}$ as ${\eta\to 0}$, defining $V_{\eta}(t,x):=v_{\eta}\Big{(}t,\sqrt{A_{\eta}^{\ast}}\,x\Big{)}$ and using the homogenized equation (5.127), we arrive at $\left\\{\begin{array}[]{c}i\displaystyle\frac{\partial V_{\eta}}{\partial t}-\Delta V_{\eta}+U_{\rm per}^{\ast}V_{\eta}=-\Big{(}\eta\,U^{(1)}+\mathrm{O}(\eta^{2})\Big{)}\,V_{\eta}\,,\;\,\text{in}\;\,\mathbb{R}^{n+1}_{T},\\\\[7.5pt] V_{\eta}(0,x)=v^{0}\Big{(}\sqrt{A_{\eta}^{\ast}}\,x\Big{)}\,,\;\,x\in\mathbb{R}^{n},\end{array}\right.$ (5.157) Proceeding similarly with respect to $V(t,x):=v_{\rm per}\Big{(}t,\sqrt{A_{\rm per}^{\ast}}\,x\Big{)}$, we obtain $\left\\{\begin{array}[]{c}i\displaystyle\frac{\partial V}{\partial t}-\Delta V+U_{\rm per}^{\ast}V_{\eta}=0\,,\;\,\text{in}\;\,\mathbb{R}^{n+1}_{T},\\\\[7.5pt] V(0,x)=v^{0}\Big{(}\sqrt{A_{\rm per}^{\ast}}\,x\Big{)}\,,\;\,x\in\mathbb{R}^{n}.\end{array}\right.$ (5.158) Now, the difference between the equations (5.157) and (5.158) yields, $\left\\{\begin{array}[]{c}i\displaystyle\frac{\partial(V_{\eta}-V)}{\partial t}-\Delta(V_{\eta}-V)+U_{\rm per}^{\ast}(V_{\eta}-V)=-\Big{(}\eta\,U^{(1)}+\mathrm{O}(\eta^{2})\Big{)}\,V_{\eta}\,,\;\,\text{in}\;\,\mathbb{R}^{n+1}_{T},\\\\[7.5pt] (V_{\eta}-V)(0,x)=v^{0}\Big{(}\sqrt{A_{\eta}^{\ast}}\,x\Big{)}-v^{0}\Big{(}\sqrt{A_{\rm per}^{\ast}}\,x\Big{)}\,,\;\,x\in\mathbb{R}^{n}.\end{array}\right.$ (5.159) Hence, multiplying the last equation by $\overline{V_{\eta}-V}$, integrating over $\mathbb{R}^{n}$ and taking the imaginary part yields $\frac{d}{dt}\|V_{\eta}-V{\|}_{L^{2}(\mathbb{R}^{n})}\leq\mathrm{O}(\eta),$ for $\eta\in\mathcal{V}$. Defining $W_{\eta}(t,x):=\frac{V_{\eta}(t,x)-V(t,x)}{\eta},\;\,\eta\in\mathcal{V},$ the last inequality provides $\sup_{\eta\in\mathcal{V}}\|W_{\eta}{\|}_{L^{2}(\mathbb{R}^{n+1}_{T})}<+\infty.$ Thus, taking a subsequence if necessary, there exists ${v^{(1)}\in L^{2}(\mathbb{R}^{n+1}_{T})}$ such that $W_{\eta}(t,x)\;\xrightharpoonup[\eta\to 0]{}\;v^{(1)}\Big{(}t,\sqrt{A_{\rm per}^{\ast}}\,x\Big{)},\;\text{in}\;L^{2}(\mathbb{R}^{n+1}_{T}).$ (5.160) Hence, multiplying the equation (5.159) by $\eta^{-1}$, letting $\eta\to 0$ and performing a change of variables, we reach the equation (5.154) finishing the proof of the theorem. ∎ ## Acknowledgements Conflict of Interest: Author Wladimir Neves has received research grants from CNPq through the grant 308064/2019-4. Author Jean Silva has received research grants from CNPq through the Grant 302331/2017-4. ## References * [1] Allaire G., Homogenization and two-scale convergence, SIAM J. Math. Anal. 23(6), 1482-1518, 1992. * [2] Allaire G., Periodic homogenization and effective mass theorems for the Schrödinger Equation. In: Abdallah N. B., Frosali G. (eds) Quantum transport. Lecture Notes in Mathematics, vol 1946. Springer, Berlin, Heidelberg, 2008. * [3] Allaire G., Vanninathan M., Homogenization of the Schrödinger equation with a time oscillating potential, Discrete Contin. Dyn. Syst.-Ser. B 6(1), 1-16, 2006. * [4] Allaire G., Piatnistki A., Homogenization of the Schrödinger Equation and Effective Mass Theorems, Commun. Math. Phys. 258(1), 1-22, 2005. * [5] Ambrosio L., Frid H., Multiscale Young Measure in almost periodic homogenization and applications, Arch. Ration. Mech. Anal. 192(1), 37-85, 2009. * [6] Andrade T., Neves W., Silva J., Homogenization of Liouville Equations beyond stationary ergodic setting, Arch. Ration. Mech. Anal. 237(2), 999-1040, 2020. * [7] Barletti L., Abdallah N. B., Quantum transport in crystals: effective mass theorem and K$\cdot$P Hamiltonians, Comm. Math. Phys. 307(3), 567-607, 2011. * [8] Bensoussan A., Lions J.-L., Papanicolaou G., Asymptotic analysis for periodic structures, Amsterdam: North-Holland Pub. Co., 1978. * [9] Blanc X., Le Bris C., Lions P.-L., Une variante de la théorie de l’homogénéisation stochastique des opérateurs elliptiques, C. R. Math. Acad. Sci. Paris 343(11-12), 717-724, 2006. * [10] Blanc X., Le Bris C., Lions P.-L., Stochastic homogenization and random lattices, J. Math. Pures Appl. 88(1), 34-63, 2007. * [11] Bourgeat A., Mikelić A., Wright S., Stochastic two-scale convergence in the mean and applications, J. Reine Angew. Math. 456, 19-51, 1994. * [12] Cancès E., Le Bris C., Mathematical modeling of point defects in materials science, Math. Models Methods Appl. Sci. 23(10), 1795-1859, 2013. * [13] Cazenave T., Haraux A., An introduction to semilinear evolution equations, Oxford Lecture Series in Mathematics and Its Applications, 13. Clarendon, Oxford University Press, New York, 1998. * [14] Chabu V., Fermanian-Kammerer C., Marcià F., Wigner measures and effective mass theorems, Ann. H. Lebesgue 3, 1049-1089, 2020. * [15] De Giorgi E., Sulla differenziabilità e l’analiticità delle estremali degli integrali multipli regolari, Mem. Accad. Sci. Torino Cl. Sci. Fis. Mat. Nat. (3) 3, 25-43, 1957. * [16] Diaz J. C., Gayte I., The two-scale convergence method Applied to generalized Besicovitch spaces, R. Soc. Lond. Proc. Ser. A Math. Phys. Eng. Sci. 458(2028), 2925-2946, 2002. * [17] Dummit D., Foote R., Abstrac algebra, Third edition, John Wiley and Sons, New York, 2004. * [18] Evans L. C., Partial differential equations, Second edition. Graduate Studies in Mathematics, vol. 19. AMS, Providence, 2010. * [19] Folland G. B., A course in abstract harmonic analysis, Studies in Advanced Mathematics, CRC Press, Boca Raton, FL, 1995. * [20] Frid H., Silva J., Versieux H., Homogenization of a generalized Stefan problem in the context of ergodic algebra, J. Funct. Anal. 268(11), 3232-3277, 2015. * [21] Górka P., Reyes E. G., Sobolev spaces on locally compact Abelian groups and the Bosonic string equation, J. Aust. Math. Soc. 98(1), 39-53, 2015. * [22] Gunning R. C., Rossi H., Analytic functions of several complex variables, Prentice-Hall, Inc., Englewood Cliffs, N.J., 1965. * [23] Hewitt E., Ross K. A., Abstract harmonic analysis. Vol. I, Springer-Verlag, Berlin, 1963. * [24] Jikov V. V., Kozlov S. M., Oleinik O. A., Homogenization of differential operators and integral functionals, Springer-Verlag, Berlin, 1994. * [25] Kato T., Perturbation theory for linear operators, Springer-Verlag, Berlin, 1995. * [26] Krantz S. G., Parks H. R., The implicit function theorem. History, theory and applications, Birkhäuser, Boston, 2002. * [27] Krengel U., Ergodic theorems, Gruyter Studies in Mathematics, vol. 6, de Gruyter, Berlin, 1985. * [28] Myers H. P., Introductory solid state physics, London: Taylor & Francis, 1990. * [29] Nguetseng G., A General convergence result for a functional related to the theory of homogenization, SIAM J. Math. Anal. 20(3), 608-623, 1989. * [30] Pankov A. A., Bounded and almost periodic solutions of nonlinear operator differential equations, Dordrecht, Springer, 2013. * [31] Poupaud F., Ringhofer C., Semi-classical limits in a crystal with exterior potentials and effective mass theorems, Comm. Partial Differential Equations 21(11-12), 1897-1918, 1996. * [32] Reed M., Simon B., Methods of modern mathematical physics. Vol. I, Academic Press, Inc., New York, 1980. * [33] Rellich F., Perturbation theory of eigenvalue problems, Gordon & Breach, New York, 1969. * [34] Rodríguez-Vega J. J., Zúñiga-Galindo W. A., Elliptic pseudodifferential equations and Sobolev spaces over $p$-adic fields, Pacific J. Math. 246(2), 407-420, 2010. * [35] Rudin W., Real and complex analysis, Third edition. McGraw-Hill Book Co., New York, 1987. * [36] Stampacchia G., Le problème de Dirichlet pour les équations elliptiques du second ordre à coefficients discontinus, Ann. Inst. Fourier (Grenoble) 15, 189-258, 1965. * [37] Wilcox C. H., Theory of Bloch waves, J. Analyse Math. 33, 146-167, 1978. * [38] Zhikov V. V., Pyatnitskii A. L., Homogenization of random singular structures and random measures, Izv. Math. 70(1), 19-67, 2006.
# A diffuse interface box method for elliptic problems G. Negrinia, N. Parolinia and M. Verania ###### Abstract We introduce a diffuse interface box method (DIBM) for the numerical approximation on complex geometries of elliptic problems with Dirichlet boundary conditions. We derive a priori $H^{1}$ and $L^{2}$ error estimates highlighting the rôle of the mesh discretization parameter and of the diffuse interface width. Finally, we present a numerical result assessing the theoretical findings. Keywords: box method, diffuse interface, complex geometries a MOX, Dipartimento di Matematica, Politecnico di Milano, Piazza Leonardo da Vinci 32, I-20133 Milano, Italy ## 1 Introduction The finite volume method (FVM) is a popular numerical strategy for solving partial differential equations modelling real life problems. One crucial and attractive property of FVM is that, by construction, many physical conservation laws possessed in a given application are naturally preserved. Besides, similar to the finite element method, the FVM can be used to deal with domains with complex geometries. In this respect, one crucial issue is the construction of the computational grid. To face this problem, one can basically resort to two different types of approaches. In the first approach, a mesh is constructed on a sufficiently accurate approximation of the exact physical domain (see, e.g., isoparametric finite elements [9], isogeometric analysis [10], or Arbitrary Lagrangian-Eulerian formulation [11, 16, 17]), while in the second approach (see, e.g., Immersed Boundary methods [19], the Penalty Methods [2], the Fictitious Domain/Embedding Domain Methods [6, 5, 4], the cut element method [7, 8] and the Diffuse Interface Method [18]) one embeds the physical domain into a simpler computational mesh whose elements can intersect the boundary of the given domain. Clearly, the mesh generation process is extremely simplified in the second approach, while the imposition of boundary conditions requires extra work. Among the methods sharing the second approach, in this paper we focus on the diffuse interface approach developed in [20]. In parallel, we consider, for its simplicity, the piecewise linear FVM, or box method, that has been the object of an intense study in the literature (see, e.g., the pioneering works [3, 15] and the more recent [12, 21]). The goal of this paper is to propose and analyse a diffuse interface variant of the box method, in the sequel named DIBM (diffuse interface box method), obtaining a priori $H^{1}$ and $L^{2}$ error estimates depending both on the discretization parameter $h$ (dictating the accuracy of the approximation of the PDE) and the width $\epsilon$ of the diffuse interface (dictating the accuracy of the domain approximation). Up to our knowledge, this is new in the literature. Besides, the study of DIBM for elliptic problems, despite its simplicity, opens the door to the study of more complicated differential problems and to the analysis of diffuse interface variants of more sophisticated finite volume schemes. The outline of the paper is as follows. In section 2 we briefly recall the box method, while in section 3, we present the diffuse interface box method (DIBM) along with a priori error estimates. Finally in section 4 we will provide a numerical test to validate the theoretical results. The numerical results have been obtained using the open-source library OpenFOAM®. ## 2 The box method In this section, we recall (see [3, 15, 21]) the box method for the solution of an elliptic problem. Let $D\subset\mathbb{R}^{2}$ be a polygonal bounded domain (in the following section this hypothesis will be relaxed). We consider the following problem: $\begin{cases}-\Delta u=f,\quad&\mathrm{in}\leavevmode\nobreak\ D\\\ u=g,\quad&\mathrm{on}\leavevmode\nobreak\ \Gamma=\partial D,\end{cases}$ (2.1) where $f\in L^{2}(\Omega)$ and $g\in H^{1/2}(\Gamma)$. Let $\mathcal{T}_{h}=\\{t_{i}\\}$ be a conforming and shape regular triangulation of $D$. We denote by $h_{t}$ the diameter of $t\in\mathcal{T}_{h}$ and we introduce the set $\textsf{V}_{h}=\\{\textsf{v}_{i}\\}$ of vertices of $\mathcal{T}_{h}$ with $\textsf{V}_{h}=\textsf{V}_{h}^{\partial}\cup\textsf{V}_{h}^{o}$, the set $\textsf{V}_{h}^{o}$ containing the interior vertices of $\mathcal{T}_{h}$. We denote by $w_{\textsf{v}}$ the set of triangles sharing the vertex v. On $\mathcal{T}_{h}$ we define the space of linear finite elements $\mathcal{V}_{h,g_{h}}=\left\\{v_{h}\in C^{0}(\bar{D}):v_{h}|_{t}\in\mathbb{P}^{1}(t)\leavevmode\nobreak\ \forall t\in\mathcal{T}_{h}\leavevmode\nobreak\ \mathrm{and}\leavevmode\nobreak\ v_{h}=g_{h}\leavevmode\nobreak\ \mathrm{on}\leavevmode\nobreak\ \partial D\right\\},$ where $g_{h}$ is a suitable piecewise linear approximation of $g$ on $\partial D$. Let $\mathcal{B}_{h}=\\{b_{\textsf{v}}\\}_{\textsf{v}\in{\textsf{V}}_{h}^{o}}$ be the “box mesh” (or dual mesh) associated to $\mathcal{T}_{h}$. Each box $b_{\textsf{v}}$ is a polygon with a boundary consisting of two straight lines in each related triangle $t\in w_{\textsf{v}}$. These lines are defined by the mid-points of the edges and the barycentres of the triangles in $w_{\textsf{v}}$. On $\mathcal{B}_{h}$ we introduce the space of piecewise constant functions, $\mathcal{W}_{h}=\left\\{w_{h}\in L^{2}(D):w_{h}\in\mathbb{P}^{0}(b_{\textsf{v}})\leavevmode\nobreak\ \forall b_{\textsf{v}}\in\mathcal{B}_{h}\right\\}.$ The box method for the approximation of (2.1) reads as follows: find $u_{B,h}\in\mathcal{V}_{h,g_{h}}$ such that $a_{\mathcal{T}_{h}}(u_{B,h},w_{h})=(f,w_{h})_{D}\quad\forall w_{h}\in\mathcal{W}_{h},$ (2.2) where $a_{\mathcal{T}_{h}}(u_{B,h},w_{h})=-\sum_{\textsf{v}\in\textsf{V}_{h}^{o}}\int_{\partial b_{\textsf{v}}}\frac{\partial v_{h}}{\partial\boldsymbol{\textbf{n}}_{b}}w_{h}\mathrm{d}s,$ (2.3) being $\boldsymbol{\textbf{n}}_{b}$ the outer normal to $b_{\textsf{v}}$ and $(\cdot,\cdot)_{D}$ is the usual $L^{2}$ scalar product on $D$. Note that there holds (see [3, 15] for the two dimensional case and [21] for the extension to any dimension) $\int_{\partial b_{\textsf{v}}}\frac{\partial\phi_{\textsf{v}^{\prime}}}{\partial\boldsymbol{\textbf{n}}_{b}}w_{h}\mathrm{d}s=\int_{D}\nabla\phi_{\textsf{v}}\cdot\nabla\phi_{\textsf{v}^{\prime}}\mathrm{d}x,\quad\forall\textsf{v}\in\textsf{V}_{h}^{o},\forall\textsf{v}^{\prime}\in\textsf{V}_{h},$ (2.4) where $\phi_{\textsf{v}}$ is the usual hat basis function with support equal to $w_{\textsf{v}}$. The relation (2.4) is crucial to show the following perturbation results (see [15, 21]): $\displaystyle\left\lVert\nabla(u_{B,h}-u_{G,h})\right\rVert_{L^{2}(D)}$ $\displaystyle\leq Ch\left\lVert f\right\rVert_{L^{2}(D)},$ (2.5) $\displaystyle\left\lVert u_{B,h}-u_{G,h}\right\rVert_{L^{2}(D)}$ $\displaystyle\leq Ch^{2}\left\lVert f\right\rVert_{L^{2}(D)},$ where $h=\max_{t\in\mathcal{T}_{h}}h_{t}$ and $u_{G,h}\in\mathcal{V}_{h,g_{h}}$ is the linear finite element approximation to the solution of problem (2.1). ## 3 The box method with diffuse interface (DIBM) Figure 1: Diffuse interface representation: $D$ is a surrogate domain of $\Omega$; $\Gamma$ is the Dirichlet boundary and $S^{\epsilon}$ is its tubular neighbour. The aim of this section is to introduce a variant of the box method for the approximate solution of problem (2.1) in case of a general (non-polygonal) domain $D\subset\mathbb{R}^{2}$, where in the spirit of [20] the Dirichlet boundary condition is treated with a diffuse interface approach. To this aim we introduce an hold-all domain $\Omega$ such that $D\subset\Omega$. In the sequel we will work under the hypothesis $\Gamma=\partial D\in C^{1,1}$. With a slight abuse of notation we denote by $\mathcal{T}_{h}$ a shape regular triangulation of $\Omega$. It is worth noting that $\mathcal{T}_{h}$ is not conforming with $D$. Following [20] we first select a tubular neighbourhood $S^{\epsilon}$ of $\Gamma$, where $\epsilon$ denotes the width of $S^{\epsilon}$ (see Figure 1). Then we introduce the set $S^{\epsilon}_{h}$ which contain s all the triangles of $\mathcal{T}_{h}$ having non-empty intersection with $S^{\epsilon}$. Note that the width of the discrete tubular neighbourhood $S^{\epsilon}_{h}$ is $\delta+\epsilon$ where $\delta$ is the maximum diameter of triangles crossed by $\partial S^{\epsilon}$. To proceed, we assume that there exists an extension $\tilde{g}\in H^{2}(\Omega)$ of the boundary data g. We set $D^{\epsilon}_{h}=D\backslash S^{\epsilon}_{h}$ and introduce the function $u^{\epsilon,h}\in H^{1}(D_{h}^{\epsilon})$ such that $u^{\epsilon,h}=g$ on $\partial D^{\epsilon}_{h}$, which solves the following continuos problem: $\int_{D^{\epsilon}_{h}}\nabla u^{\epsilon,h}\cdot\nabla v=\int_{D^{\epsilon}_{h}}fv\quad\forall v\in H^{1}_{0}(D^{\epsilon}_{h}).$ (3.1) The solution $u^{\epsilon,h}$ is then extended to $S^{\epsilon}_{h}$ by setting $u^{\epsilon,h}=\tilde{g}$ in $S^{\epsilon}_{h}$. The following results have been proved in [20, Thm 1.2]: $\frac{1}{\epsilon+\delta}\left\lVert u-u^{\epsilon,h}\right\rVert_{L^{2}(D)}+\frac{1}{\sqrt{\epsilon+\delta}}\left\lVert\nabla u-\nabla u^{\epsilon,h}\right\rVert_{L^{2}(D)}\leq C\left(\left\lVert f\right\rVert_{L^{2}(D)}+\left\lVert g\right\rVert_{H^{2}(D)}\right).$ (3.2) Let $\mathcal{V}_{h,\tilde{g}_{h}}^{\epsilon}=\left\\{v_{h}|_{D^{\epsilon}_{h}}:v_{h}\in\mathbb{P}^{1}(t)\forall t\in\mathcal{T}_{h}\leavevmode\nobreak\ \mathrm{and}\leavevmode\nobreak\ v_{h}=\tilde{g}_{h}\leavevmode\nobreak\ \mathrm{on}\leavevmode\nobreak\ \partial D^{\epsilon}_{h}\right\\}$, with $\tilde{g}_{h}$ the Lagrangian piecewise linear interpolant of $\tilde{g}$. It has been proved (cf. [20, Thms 5.1 and 5.3]) that the linear finite element approximation $u^{\epsilon}_{G,h}\in\mathcal{V}_{h,\tilde{g}_{h}}^{\epsilon}$ of $u^{\epsilon,h}$ satisfies the following estimates: $\displaystyle\left\lVert\nabla(u^{\epsilon,h}-u^{\epsilon}_{G,h})\right\rVert_{L^{2}(D)}$ $\displaystyle\leq C(\sqrt{\delta}+\kappa^{\frac{2}{3}}+h)\left(\left\lVert f\right\rVert_{L^{2}(D)}+\left\lVert\tilde{g}\right\rVert_{H^{2}(D)}\right),$ (3.3) $\displaystyle\left\lVert u^{\epsilon,h}-u^{\epsilon}_{G,h}\right\rVert_{L^{2}(D)}$ $\displaystyle\leq C(\delta+\kappa^{\frac{4}{3}}+h^{2})\left(\left\lVert f\right\rVert_{L^{2}(D)}+\left\lVert\tilde{g}\right\rVert_{H^{2}(D)}\right),$ where $\kappa$ is the maximum diameter of the triangles intersection $\partial S^{\epsilon+h}$ and $u^{\epsilon}_{G,h}$ has been extended to $D^{\epsilon}_{h}$ by setting $u^{\epsilon}_{G,h}=\tilde{g}_{h}$ on $S^{\epsilon}_{h}$. Let us now introduce the box method with diffuse interface (DIBM). We denote by $u^{\epsilon}_{B,h}\in\mathcal{V}_{h,\tilde{g}_{h}}^{\epsilon}$, the approximation obtained from applying the box method to (3.1) (cf. (2.2)). The solution $u^{\epsilon}_{B,h}$ is then extended to $D$ by setting $u^{\epsilon}_{B,h}=\tilde{g}_{h}$ in $S^{\epsilon}_{h}$. Then employing the triangle inequality in combination with (3.2), (3.3) and (2.5) we get the following estimates for DIBM: $\displaystyle\left\lVert\nabla(u-u^{\epsilon}_{B,h})\right\rVert_{L^{2}(D)}$ $\displaystyle\lesssim\sqrt{\epsilon+\delta}+\sqrt{\delta}+k^{\frac{2}{3}}+h,$ (3.4) $\displaystyle\left\lVert u-u^{\epsilon}_{B,h}\right\rVert_{L^{2}(D)}$ $\displaystyle\lesssim\epsilon+\delta+k^{\frac{4}{3}}+h^{2}.$ Figure 2: Discrete diffuse interface representation on triangulation (left) and on box mesh (right). Constrained cells are marked with red dots while the continuous and discrete diffuse interfaces are coloured by darker an lighter red respectively. ## 4 Numerical experiments In this section we numerically assess the theoretical estimates obtained in Section 3. To this aim, we consider the test case originally introduced in [20, Section 6] that is briefly recalled in the sequel. Let $\Omega=(-1,1)^{2}$ and let $\Gamma$ be the boundary of the circle $B_{1}(0)$ with centre $(0,0)$ and unitary radius. Thus, $\Gamma$ splits the domain $\Omega$ into two subregions: $D_{1}=B_{1}(0)$ and $D_{2}=\Omega\setminus\overline{D}_{1}$. Let $u$ be the solution of the following problem $-\Delta u=f\leavevmode\nobreak\ \leavevmode\nobreak\ \text{in\leavevmode\nobreak\ }\Omega,\qquad u=g\leavevmode\nobreak\ \leavevmode\nobreak\ \text{on\leavevmode\nobreak\ }\Gamma,\qquad u=0\leavevmode\nobreak\ \leavevmode\nobreak\ \text{on\leavevmode\nobreak\ }\partial\Omega,$ (4.1) where $g(x,y)=(4-x^{2})(4-y^{2})$ on $\Gamma$ and extended to $\Omega$ as $\tilde{g}(x,y)=(4-x^{2})(4-y^{2})\cos(1-x^{2}-y^{2}).$ Setting the solution equal to: $u(x,y)=(4-x^{2})(4-y^{2})\left(\chi_{D_{2}}+\exp(1-x^{2}-y^{2})\chi_{\bar{D}_{1}}\right),$ (4.2) where $\chi_{D_{i}}$, $i=1,2$ are the characteristic functions of the two parts of $\Omega$, the source term $f$ is chosen as: $f=\begin{cases}-\Delta u&\quad\mathrm{in}\leavevmode\nobreak\ \Omega\backslash\Gamma,\\\ 0&\quad\mathrm{on}\leavevmode\nobreak\ \Gamma.\end{cases}$ All the computations have been performed employing a Voronoi dual mesh of a Delaunay triangulation (i.e., the dual mesh is obtained by connecting the barycentres of the triangles with straight lines). To validate the estimates (3.4) we consider in a separate way the influence of $h$ and $\epsilon$ on the error. More precisely, we first explore the convergence with respect to $h$ and then we study the convergence with respect to $\epsilon$. In both cases we consider a uniform discretization of the domain $\Omega$ so to have $\kappa=\delta=h$. ### Convergence w.r.t. $\boldsymbol{h}$ We set $\epsilon=2^{-20}\ll h$ while we let $h$ vary as $h=0.056,0.028,0.0139,0.00694.$ From Figure 3 we observe that the $L^{2}$-norm of the error decreases with order $1$ while the error decreases with order 1/2 in the $H^{1}$-norm. These rates of convergence are in agreement with (3.4). ###### Remark 4.1. If a local refinement of the diffuse interface region is performed in such a way that $\delta\simeq\kappa\simeq h^{2}$ (Figure 5), then first and second order of convergence are recovered for $H^{1}$ and $L^{2}$ norms, respectively (cf. [20, Section 6]). ### Convergence w.r.t. $\boldsymbol{\epsilon}$ We employ a fine mesh ($h=0.00694$) and let the value of $\epsilon$ vary as: $\epsilon=2^{i},\leavevmode\nobreak\ i=-1,...,-20.$ The results are collected in Figure 4. The theoretical rates of convergence with respect to $\epsilon$ (cf. (3.4)) are obtained both in the $L^{2}$-norm (order $1$) and in the $H^{1}$\- norm (order $1/2$ ). It is worth noticing that when the value of $\epsilon$ becomes smaller than the chosen value of $h$, a plateau is observed as the (fixed) contribution from the discretization of the PDE (related to $h$) dominates over the contribution from the introduction of the diffuse interface (related to $\epsilon$). Figure 3: Error behaviour with respect to $h$ (fixed $\epsilon=2^{-20}$): (left) $L^{2}$-norm error, (right) $H^{1}$-norm error. Dashed lines are theoretical convergence orders. Figure 4: Error behaviour with respect to $\epsilon$ (fixed $h=0.00694$): (left) $L^{2}$-norm error, (right) $H^{1}$-norm error. Dotted lines are theoretical convergence orders. Figure 5: On the left: example of a dual mesh with local mesh refinement around surrogate boundary. On the right: error behaviour with respect to $h$ with local mesh refinement around the interface (fixed $\epsilon=2^{-20}$): (left) $L^{2}$-norm error, (right) $H^{1}$-norm error. Dashed lines are theoretical convergence orders. ## 5 Conclusions In this paper we introduced a diffuse interface variant of a finite volume method, namely of the the so-called box method and obtained $L^{2}$ and $H^{1}$ error estimates highlighting the contributions from the discretization parameter $h$ associated to the polygonal computational mesh and the width $\epsilon$ of the diffuse interface. Despite the simplicity of the method, the present contribution seems to be novel in the literature. Moreover, the present work may represent the first step towards the study of the diffuse interface variant of more sophisticated finite volume schemes (possibly for more complex differential problems). This work opens fictitious boundary methods analysis to the box method and finite volume framework. Possible extensions of this research could be to being able to apply the plenty of penalization methods that are mostly thought for finite element implementations such as shifted boundary, Nitsche penalty, cut-fem or Brinkman penalization. ## 6 Acknowledgements The first author acknowledges the financial support of Fondazione Politecnico. The third author acknowledges the financial support of PRIN research grant number 201744KLJL “ _Virtual Element Methods: Analysis and Applications_ ” funded by MIUR. The second and third authors acknowledge the financial support of INdAM-GNCS. ## References * [1] Ivo Babuška. The finite element method with Lagrangian multipliers. Numer. Math., 20:179–192, 1972/73. * [2] Ivo Babuška. The finite element method with penalty. Math. Comp., 27:221–228, 1973. * [3] Randolph E. Bank and Donald J. Rose. Some error estimates for the box method. SIAM Journal on Numerical Analysis, 24(4):777–787, 1987. * [4] Stefano Berrone, Andrea Bonito, Rob Stevenson, and Marco Verani. An optimal adaptive fictitious domain method. Math. Comp., 88(319):2101–2134, 2019. * [5] Daniele Boffi and Lucia Gastaldi. A finite element approach for the immersed boundary method. volume 81, pages 491–501. 2003. In honour of Klaus-Jürgen Bathe. * [6] Christoph Börgers and Olof B. Widlund. On finite element domain imbedding methods. SIAM J. Numer. Anal., 27(4):963–978, 1990. * [7] Erik Burman and Peter Hansbo. Fictitious domain finite element methods using cut elements: I. A stabilized Lagrange multiplier method. Comput. Methods Appl. Mech. Engrg., 199(41-44):2680–2686, 2010\. * [8] Erik Burman and Peter Hansbo. Fictitious domain finite element methods using cut elements: II. A stabilized Nitsche method. Appl. Numer. Math., 62(4):328–341, 2012. * [9] Philippe G. Ciarlet. The finite element method for elliptic problems, volume 40 of Classics in Applied Mathematics. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 2002. Reprint of the 1978 original [North-Holland, Amsterdam; MR0520174 (58 #25001)]. * [10] J. Austin Cottrell, Thomas J. R. Hughes, and Yuri Bazilevs. Isogeometric analysis. John Wiley & Sons, Ltd., Chichester, 2009. Toward integration of CAD and FEA. * [11] J. Donea, S. Giuliani, and J.P. Halleux. An arbitrary lagrangian-eulerian finite element method for transient dynamic fluid-structure interactions. Computer Methods in Applied Mechanics and Engineering, 33(1):689 – 723, 1982. * [12] Richard E. Ewing, Tao Lin, and Yanping Lin. On the accuracy of the finite volume element method based on piecewise linear polynomials. SIAM J. Numer. Anal., 39(6):1865–1888, 2002. * [13] V. Girault and R. Glowinski. Error analysis of a fictitious domain method applied to a Dirichlet problem. Japan J. Indust. Appl. Math., 12(3):487–514, 1995. * [14] Roland Glowinski, Tsorng-Whay Pan, and Jacques Périaux. A fictitious domain method for Dirichlet problem and applications. Comput. Methods Appl. Mech. Engrg., 111(3-4):283–303, 1994. * [15] W. Hackbusch. On first and second order box schemes. Computing, 41(4):277–296, 1989. * [16] C. W. Hirt, A. A. Amsden, and J. L. Cook. An arbitrary Lagrangian-Eulerian computing method for all flow speeds [J. Comput. Phys. 14 (1974), no. 3, 227–253]. volume 135, pages 198–216. 1997. With an introduction by L. G. Margolin, Commemoration of the 30th anniversary {of J. Comput. Phys.}. * [17] Thomas J. R. Hughes, Wing Kam Liu, and Thomas K. Zimmermann. Lagrangian-Eulerian finite element formulation for incompressible viscous flows. Comput. Methods Appl. Mech. Engrg., 29(3):329–349, 1981. * [18] X. Li, J. Lowengrub, A. Rätz, and A. Voigt. Solving PDEs in complex geometries: a diffuse domain approach. Commun. Math. Sci., 7(1):81–107, 2009. * [19] Charles S. Peskin. The immersed boundary method. Acta Numer., 11:479–517, 2002. * [20] Matthias Schlottbom. Error analysis of a diffuse interface method for elliptic problems with Dirichlet boundary conditions. Appl. Numer. Math., 109:109–122, 2016. * [21] Jinchao Xu and Qingsong Zou. Analysis of linear and quadratic simplicial finite volume methods for elliptic equations. Numer. Math., 111(3):469–492, 2009.
acoustics actuators addictions admsci aerospace agriculture agriengineering agronomy ai algorithms animals antibiotics antibodies antioxidants applmech applsci arts asc asi atmosphere atoms axioms batteries bdcc behavsci beverages bioengineering biology biomedicines biomimetics biomolecules biosensors brainsci buildings cancers carbon catalysts cells ceramics challenges chemengineering chemistry chemosensors children civileng cleantechnol climate clockssleep cmd coatings colloids computation computers condensedmatter cosmetics cryptography crystals dairy data dentistry designs diagnostics diseases diversity drones econometrics economies education ejbc ejihpe electrochem electronics endocrines energies entropy environments epigenomes est fermentation fibers fire fishes fluids foods forecasting forests fractalfract futureinternet futurephys galaxies games gastrointestdisord gels genealogy genes geohazards geosciences geriatrics hazardousmatters healthcare hearts heritage highthroughput horticulturae humanities hydrology ijerph ijfs ijgi ijms ijtpp informatics information infrastructures inorganics insects instruments inventions iot j jcdd jcm jcp jcs jdb jfb jfmk jimaging jintelligence jlpea jmmp jmse jne jnt jof joitmc jpm jrfm jsan land languages laws life literature logistics lubricants machines magnetochemistry make marinedrugs materials mathematics mca medicina medicines medsci membranes metabolites metals microarrays micromachines microorganisms minerals modelling molbank molecules mps mti nanomaterials ncrna ijns neurosci neuroglia nitrogen notspecified nutrients oceans ohbm optics particles pathogens pharmaceuticals pharmaceutics pharmacy philosophies photonics physics plants plasma pollutants polymers polysaccharides preprints proceedings processes prosthesis proteomes psych publications quantumrep quaternary qubs reactions recycling religions remotesensing reports reprodmed resources risks robotics safety sci scipharm sensors separations sexes signals sinusitis smartcities sna societies socsci soilsystems sports standards stats surfaces surgeries sustainability sustainableworld symmetry systems technologies telecom test tourismhosp toxics toxins transplantology tropicalmed universe urbansci vaccines vehicles vetsci vibration viruses vision water wem wevj
# Coverage Analysis of Broadcast Networks with Users Having Heterogeneous Content/Advertisement Preferences Kanchan K. Chaurasia, Reena Sahu, Abhishek K. Gupta The authors are with the Department of Electrical Engineering, Indian Institute of Technology Kanpur, Kanpur, India 208016. Email<EMAIL_ADDRESS> ###### Abstract This work is focused on the system-level performance of a broadcast network. Since all transmitters in a broadcast network transmit the identical signal, received signals from multiple transmitters can be combined to improve system performance. We develop a stochastic geometry based analytical framework to derive the coverage of a typical receiver. We show that there may exist an optimal connectivity radius that maximizes the rate coverage. Our analysis includes the fact that users may have their individual content/advertisement preferences. We assume that there are multiple classes of users with each user class prefers a particular type of content/advertisements and the users will pay the network only when then can see content aligned with their interest. The operator may choose to transmit multiple contents simultaneously to cater more users’ interests to increase its revenue. We present revenue models to study the impact of the number of contents on the operator revenue. We consider two scenarios for users’ distribution- one where users’ interest depends on their geographical location and the one where it doesn’t. With the help of numerical results and analysis, we show the impact of various parameters including content granularity, connectivity radius, and rate threshold and present important design insights. ###### Index Terms: Stochastic geometry, Broadcast networks, Coverage. ## I Introduction Broadcasting networks provide society with various services including TV communication, delivery of critical information and alerts, general entertainment, and educational services and thus have been a key wireless technology. With the recent advancement in wireless technologies and handheld electronic devices including smartphone and tablets, the use of broadcasting services has been extended to include many modern applications including delivery of traffic information to vehicles in vehicle-to-infrastructure networks, advertisement industry and mobile TV services. Many broadcasting standards have been recently proposed including the digital video broadcast- terrestrial standard (DVB-T2), the advanced television systems committee standard (ATSC 3.0), and the DVB-next generation handheld standard (DVB-NGH) to assist delivering of TV broadcasting services to mobile devices [1][2][3]. The advent of digital broadcasting has lead to a significant increase in the demand for multimedia services for handheld devices including mobile TV, live video streaming, and on-demand video in the last decade [1]. From these applications’ perspective, broadcasting based multimedia services can provide better data rate and performance compared to the uni-cast cellular network based mobile TV. Note that in a cellular network where the desired data is transmitted to each user via orthogonal resources, users may suffer from spectral congestion in regions with high density due to limit bandwidth resulting in the performance degradation. But, in a broadcasting network providing multimedia services, all transmitters transmit identical data to all users and hence, do not require orthogonal resource. In these networks, each transmitter can use the complete spectrum to serve their users, and hence, these are also called single frequency networks (SFN). Due to this, users may experience a better quality of service. ### I-A Related Work Given increasing demand of broadcasting services, it is very interesting to analyze the broadcast networks in terms of signal-to-noise-ratio (SINR) and achievable data rate to understand their limitations and potential to meet these demands. There have been some recent works in the system-level analysis of broadcast networks. In [4], the authors evaluated the blocking probability for users accessing a network delivering mobile TV services over a hybrid broadcast unicast communication. In [5], the authors studied a cellular network with uni-cast and multicast-broadcast deployments. However, these works didn’t include the effect of transmitters’ locations in the evaluations which is required for the system-level analysis of broadcast networks. Stochastic geometry framework can be utilized to analyze wireless networks from system level perspective [6, 7, 8]. Stochastic geometry based models have been validated for various types of networks including cellular networks,and ad hoc networks [9, 6, 10, 11]. In [12], the authors describe the analytical approach to calculate the coverage probability of a hybrid broadcast and uni- cast network, however, the authors have only considered a single broadcast transmitter along with many uni-cast transmitters. In our past work [13], we have considered a broadcast network with multiple broadcasting transmitters to compute the coverage performance of users. However, the work assumed a static connectivity region around the user where transmitters need to be located to be able to serve the users. As shown in this paper, this connectivity region is of variable size depending on the location of the first closest transmitter. To the best of our knowledge, there exists no other past work which analyzes the SINR and rate performance of a broadcast network with multiple broadcasting transmitters which is one of the main focuses of this paper. Another important metric to evaluate broadcast networks is the revenue earned by the network operator. In a broadcast network, the revenue is generated either from subscribers as network access fees for viewing content of their choice or from advertisers to show their advertisements to interested subscribers. With digital broadcast, subscriptions and user-targeted advertisements added a new dimension in the revenue. Due to advancements in technologies over the past few decades, the advertising has become more user targeted and location-adaptive which can be planned according to the user demographics and their preferences to improve network revenue. In [14], the authors studied location-based mobile marketing and advertising to show the positive interest of mobile consumers in receiving relevant promotions. It is intuitive that a targeted and localized content will have a better engagement factor. It is interesting to analyze the network revenue earned from the users with their preference dependent on their choices and geographical location. As far we know, there does not exist any past work that analyses the network revenue of a broadcast network with subscribers having preferences for content and advertisement which is another focus of this paper. ### I-B Contributions In this paper, we derive an analytical framework to evaluate the performance of a broadcast network with multiple broadcasting transmitters with users having content preference. We also present a revenue model to quantify the network revenue earned by the network operator. In particular, the contributions of this paper are as follows: 1. 1. We consider a broadcast network with multiple transmitters. Since all transmitters in a broadcast network are transmitting the same signal, received signals from multiple transmitters from a certain connectivity region around the user can be combined to improve the coverage at this user. Using tools from stochastic geometry, we derive the expression for SINR coverage and rate coverage of a typical receiver located at the origin. Due to the contribution in the desired signal power from multiple transmitters, the analysis is significantly different and difficult than their cellular counterpart. Our main contribution lies in developing the framework and deriving techniques to evaluate the analytical expressions of SINR and rate coverage. We show that this connectivity region depends on network bandwidth. 2. 2. We present some numerical results to validate our analysis and present design insights. We show the impact of connectivity region size, path-loss exponents, and the network density on the SINR and rate coverage. We also find that there exists an optimal size of connectivity region that maximizes the rate coverage. 3. 3. In this paper, we also include the fact that users may have their individual content or advertisement preferences. We assume that there are multiple classes of users with each class of users prefers a particular type of content/advertisements and the users will pay the network only when then can see a particular content of their interest. We assume that one unit of revenue comes to the network from a particular class of users if every user of this class can see the content as per the preference of this class. We study the revenue thus obtained by the network from users. The broadcast operator may choose to transmit multiple contents simultaneously to cater more users’ interest to increase its revenue. However, given the limited resources, the network can cater only to few classes and this capability depends on how these user classes are distributed spatially. There are two scenarios considered for users’ distribution. In one scenario, users’ interest depend on their geographical position in the network and in the second scenario it does not. We calculate the analytical expression for SINR coverage and rate coverage at a typical user and evaluate the total revenue. We present many important design insights via numerical results. Notation: Let $\mathcal{B}({\mathbf{x}},r)$ denote the ball of radius $r$ with center at ${\mathbf{x}}$. $\|{\mathbf{x}}\|$ denotes the norm of the vector ${\mathbf{x}}$ and $\|{\mathbf{x}}_{i}\|=r_{i}$ denotes the random distance of BBS located at ${{\mathbf{x}}_{i}}$. Let $\mathbf{o}$ denote the origin. $\mathsf{B}\left(x,y;z\right)$ is the incomplete Beta function which is defined as $\mathsf{B}\left(x,y;z\right)=\int_{0}^{z}u^{x-1}(1-u)^{y-1}\mathrm{d}u.$ Let $c$ denote the speed of EM waves in the media. $\mathsf{A}^{\complement}$ denotes the complement of set $\mathsf{A}$. ## II System Model In this paper, we consider a broadcast network with multiple broadcasting base stations (BBSs), deployed in the 2D region $\mathsf{R}=\mathbb{R}^{2}$. The considered system model is as follows: Figure 1: Illustration of system model of a broadcast network. A typical user is considered at the origin. ${X_{0}}$ is the distance of the nearest BS from the typical user. The 2D region $\mathcal{B}(\mathbf{o},{X_{0}}+R_{\mathrm{s}})$ denotes the connectivity region of the user. ### II-A Network Model The location of BBSs can be modeled as a homogeneous Poisson point process $\Phi=\\{\mathbf{X}_{i}\in\mathbb{R}^{2}\\}$ with density $\lambda$ in the region $\mathsf{R}$ (See Fig. 1). Let $R_{e}=1/\sqrt{\lambda\pi}$ which represents the cell radius of an average cell. The subscribers (users) of the broadcasting service are assumed to form a stationary point process. We assume a typical user located at the origin $\mathbf{o}$. Consider each BS is operating in the same frequency band with the transmission bandwidth $W$. Let $T_{\mathrm{s}}$ is the symbol time of the transmitted symbol which is inversely proportional to the bandwidth $W$. Assume the transmit power of each BBS be $p_{\mathsf{t~{}\\!}}$ and all devices are equipped with a single isotropic transmit antenna. The analysis can be extended for finite networks by taking $\mathsf{R}=\mathcal{B}(\mathbf{o},R)$ with a finite $R$. ### II-B Channel Model We assume the standard path-loss model. Hence, the received signal power from the $i$th BBS at the typical user at origin is given as $\displaystyle P_{i}$ $\displaystyle=p_{\mathsf{t~{}\\!}}a\beta_{i}{\|\mathbf{X}_{i}\|}^{-\alpha},$ (1) where $X_{i}=\|\mathbf{X}_{i}\|$ denotes the random distance of this BBS from the typical user. Here, $\alpha$ is the path-loss exponent and $a$ is near- field gain which depends on the propagation environment. $\beta_{i}$ denotes the fading between the $i^{th}$ BBS and the user. We assume Rayleigh fading, i.e. , $\beta_{i}\sim\mathrm{Exp}(1)$ for tractability. ### II-C Serving Signal and Interference Model In a broadcast system, multiple BBSs may transmit the same data at the one frequency band (as suggested by the name SFN). Therefore, at the receiver end, it can be seen as a single transmission with multi-path propagation and signals transmitted from multiple BBSs can be combined at the user. However, since the signals from different BBSs are delayed according to time delays dependent on their distance, some of these signals may be delayed significantly and may overlap with the next transmission slots. Therefore, only those signals that have delay within a certain limit can be combined to successfully decode the received symbol [15]. The rest BBSs contribute to the ISI (inter-symbol interference) which can be significant depending on the BBSs density. Let $\mathbf{X}_{0}$ denotes the nearest serving BBS. The probability density function of the distance $X_{0}=\|\mathbf{X}_{0}\|$ to the nearest BS from the user is given as[7] $\displaystyle f_{X_{0}}(u)=2\pi\lambda ue^{-\pi\lambda u^{2}}\mathbbm{1}\left({u\geq 0}\right).$ (2) The time taken by the signal to reach from the $i^{th}$ BBS located at $\mathbf{X}_{i}$ to the typical user at $\mathbf{o}$ be $T_{i}={X_{i}}/c$. In particular, $T_{0}$ denotes the time taken by signal to reach from $\mathbf{X}_{0}$ to a typical user at $\mathbf{o}$. Let the propagation delay of transmitted signal from $i^{th}$ BBS compared to the nearest serving BBS is $\Delta_{i}=T_{i}-T_{0}$. We assume that the receiver design allows the maximum delay of $\delta T_{\mathrm{s}}$ for the received signals to be combined at the user where $\delta\in[0\ 1]$ is a design parameter. This means that the received signal from the $i^{th}$ BBS may contribute in the serving signal power if $\Delta_{i}\leq\delta T_{\mathrm{s}}$. This condition is equivalent to the condition $\|\mathbf{X}_{i}\|-\|\mathbf{X}_{0}\|\leq R_{\mathrm{s}}\stackrel{{\scriptstyle\Delta}}{{=}}T_{\mathrm{s}}\delta c$ on the BBSs location $\mathbf{X}_{i}$. In other words, this means that all the BBSs that are located in the 2D region $\\{\mathbf{X}:\|\mathbf{X}\|\leq\|\mathbf{X}_{0}\|+R_{\mathrm{s}}\\}=\mathcal{B}(\mathbf{o},{X_{0}}+R_{\mathrm{s}})$ can contribute to the serving signal at the typical receiver at origin $\mathbf{o}$. We term this region $\mathcal{B}(\mathbf{o},{X_{0}}+R_{\mathrm{s}})$ as the connectivity region for the user and ${X_{0}}+R_{\mathrm{s}}$ can be termed as the connectivity radius. Let $\overline{m}^{2}={\lambda\pi R_{\mathrm{s}}^{2}}$ denote the mean number of BBSs in this connectivity radius. On the other hand, all the BBSs located outside $\mathcal{B}(\mathbf{o},{X_{0}}+R_{\mathrm{s}})$ i.e. all the BBSs with ${X_{i}}\geq{X_{0}}+R_{\mathrm{s}}$ will contribute to the interference power even when they are transmitting the same data as their signal will be delayed beyond the specified limit. ### II-D Modeling Content Preferences of Users In this paper, we also include the fact that users may have their individual content or advertisement preferences. We assume that there are $N_{\mathrm{c}}$ classes of users. Here, $N_{\mathrm{c}}$ is termed content/advertisement granularity. Each class of users prefers a particular type of content/ advertisements. We assume that the users will pay the network only when then can see a particular content of their interest. Each class consists of some quanta of users. For simplicity, we assume that each class has the same number of users, however, the presented framework can be trivially extended to include user classes with unequal sizes. We assume that one unit of revenue comes to the network from a particular class of users if every user of this class can see the content as per the preference of this class. Given the limited resources, the network can cater only to few classes and this capability depends on how these user classes are distributed spatially. We will consider two types of users class distributions over the geographical space. We will also discuss a revenue model to characterize the network’s revenue to help us understand optimal scheduling policies for the two scenarios. ## III Coverage Analysis for Common Content Transmission We first start with the scenario that all users seek the same content, hence, all BBSs are transmitting the same content to everyone. Examples include systems transmitting emergency information, or traffic data which is common to every user. In this section, we will derive the SINR and rate coverage probability for a typical user at the origin $\mathbf{o}$ for such system. ### III-A SINR Since all BBSs are transmitting the same content, all BBSs located inside the connectivity region $\mathcal{B}(0,X_{0}+\mathbb{R}_{\mathrm{s}})$ contribute to the signal power. Therefore, the desired received signal power for the typical user at origin is given as $\displaystyle S^{\prime}$ $\displaystyle=p_{\mathsf{t~{}\\!}}a\beta_{0}{\|\mathbf{X}_{0}\|}^{-\alpha}+\sum_{\mathbf{X}_{i}\in\Phi\cap\mathcal{B}(0,{X_{0}}+R_{\mathrm{s}})\setminus\mathbf{X}_{0}}p_{\mathsf{t~{}\\!}}a\beta_{i}{\|\mathbf{X}_{i}\|}^{-\alpha}.$ (3) Similarly, the total interference can be given as $\displaystyle I^{\prime}$ $\displaystyle=\sum_{\mathbf{X}_{j}\in\Phi\cap\mathcal{B}(0,\,X_{0}+R_{\mathrm{s}})^{\complement}}p_{\mathsf{t~{}\\!}}a\beta_{j}{\|\mathbf{X}_{j}\|}^{-\alpha}.$ (4) The signal-to-interference-plus-noise ratio (SINR) at the typical receiver is given as $\displaystyle\mathtt{SINR}$ $\displaystyle=\frac{S^{\prime}}{I^{\prime}+N}=\frac{\beta_{0}{{X_{0}}}^{-\alpha}+\sum_{\mathbf{X}_{i}\in\Phi\cap\mathcal{B}(0,{X_{0}}+R_{\mathrm{s}})\setminus\mathbf{X}_{0}}\beta_{i}{{X_{i}}}^{-\alpha}}{\sum_{\mathbf{X}_{j}\in\Phi\cap\mathcal{B}(0,{X_{0}}+R_{\mathrm{s}})^{\complement}}\beta_{j}{{X_{j}}}^{-\alpha}+\sigma^{2}}.$ (5) Here, $\sigma^{2}$ is the normalized noise power given as $\sigma^{2}=N/(p_{\mathsf{t~{}\\!}}a)$ where $N$ is the noise power. Similarly normalized desired received signal power and interference are denoted by $S$ and $I$ which are given as $S=S^{\prime}/(p_{\mathsf{t~{}\\!}}a)$ and $I=I^{\prime}/(p_{\mathsf{t~{}\\!}}a)$. Hence, the SINR is equal to $\displaystyle\mathtt{SINR}$ $\displaystyle=\frac{S}{I+\sigma^{2}}.$ (6) Let $K=\sigma^{2}/R_{e}^{-\alpha}$ which represents the SNR at cell edge of an average cell. ### III-B SINR Coverage Probability The SINR coverage probability $\mathrm{p_{c}}(\tau,\lambda)$ of a user is defined as the probability that the SINR at the user is above the threshold $\tau$ i.e. $\displaystyle\mathrm{p_{c}}(\tau,\lambda)$ $\displaystyle=\mathbb{P}\left[\mathtt{SINR}>\tau\right]$ (7) Using the conditioning on the nearest serving BBS’s location $\mathbf{X}_{0}$, the SINR coverage for typical user at $\mathbf{o}$ is given as $\displaystyle\mathrm{p_{c}}(\tau,\lambda)$ $\displaystyle=\mathbb{E}_{\mathbf{X}_{0}}\left[\mathbb{P}\left(\mathtt{SINR}>\tau\right)|\,\mathbf{X}_{0}\right]$ $\displaystyle=\mathbb{E}_{\mathbf{X}_{0}}\left[\mathbb{P}\left(\frac{S}{I+\sigma^{2}}>\tau\right)|\,\mathbf{X}_{0}\right]=\mathbb{E}_{\mathbf{X}_{0}}\left[\mathbb{P}\left(S>(I+\sigma^{2})\tau|\,\mathbf{X}_{0}\right)\right].$ (8) Using the distribution of $\|\mathbf{X}_{0}\|=X_{0}$, the SINR coverage probability can be further written as $\displaystyle\mathrm{p_{c}}(\tau,\lambda)$ $\displaystyle=\mathbb{E}_{\mathbf{X}_{0}}\left[\mathbb{P}\left(S>\tau\left(I+N\right)\,|\,\mathbf{X}_{0}\right)\right]$ $\displaystyle=\int_{0}^{\infty}2\pi\lambda ue^{-\pi\lambda u^{2}}\mathbb{P}\left(S>\tau\left(I+\sigma^{2}\right)\,|\,\|\mathbf{X}_{0}\|=u\right)\mathrm{d}u.$ (9) To solve the inner term further, we will use Gil Pelaez’s Lemma [16] which states that the CDF of a random variable $Y$ can be written in term of its Laplace transform $\mathcal{L}_{Y}(t)$ as $\displaystyle F_{Y}(s)=\mathbb{P}\left[Y\leq s\right]=\frac{1}{2}-\frac{1}{\pi}\int_{0}^{\infty}\frac{1}{t}\mathsf{Im}\left[e^{-jts}\mathcal{L}_{Y}(-jt)\right]\mathrm{d}t.$ (10) Using this Lemma, we get, $\displaystyle\mathbb{P}\left(S>(I+\sigma^{2})\tau|\,\mathbf{X}_{0}\right)$ $\displaystyle=\mathbb{E}_{I|\,\mathbf{X}_{0}}\left[\frac{1}{2}+\frac{1}{\pi}\int_{0}^{\infty}\frac{1}{t}\mathsf{Im}\left[e^{-jt\tau(I+\sigma^{2})}\mathcal{L}_{S}(-jt)\right]\mathrm{d}t\right]$ $\displaystyle=\frac{1}{2}+\frac{1}{\pi}\int_{0}^{\infty}\frac{1}{t}\mathsf{Im}\left[\mathbb{E}_{I|\,\mathbf{X}_{0}}\left[e^{-jt\tau(I+\sigma^{2})}\right]\mathcal{L}_{S|\,\mathbf{X}_{0}}(-jt)\right]\mathrm{d}t$ $\displaystyle=\frac{1}{2}+\frac{1}{\pi}\int_{0}^{\infty}\frac{1}{t}\mathsf{Im}\left[\mathcal{L}_{I|\,\mathbf{X}_{0}}(jt\tau)e^{-jt\tau\sigma^{2}}\mathcal{L}_{S|\,\mathbf{X}_{0}}(-jt)\right]\mathrm{d}t.$ (11) Now, using in (9), the SINR coverage probability is $\displaystyle\mathrm{p_{c}}(\tau,\lambda)$ $\displaystyle=\frac{1}{2}+\frac{1}{\pi}\int_{0}^{\infty}\int_{0}^{\infty}2\pi\lambda ue^{-\pi\lambda u^{2}}\frac{1}{t}\mathsf{Im}\left[\mathcal{L}_{I|\,\mathbf{X}_{0}}(jt\tau)e^{-jt\tau\sigma^{2}}\mathcal{L}_{S|\,\mathbf{X}_{0}}(-jt)\right]\mathrm{d}t\ \mathrm{d}u.$ (12) Here, $\mathcal{L}_{I|\,\mathbf{X}_{0}}(.)$ and $\mathcal{L}_{S|\,\mathbf{X}_{0}}(.)$ are the Laplace transform of the sum interference $I$ and of the desired received signal power $S$ respectively which are given in the following Lemma. ###### Lemma 1. The Laplace transforms of the desired signal power and the sum interference at the receiver located at origin $\mathbf{o}$ are given as $\displaystyle\mathcal{L}_{S\,|\,\mathbf{X}_{0}}(s)$ $\displaystyle=\frac{1}{1+sX_{0}^{-\alpha}}\exp\left(-2\pi\lambda\int_{X_{0}}^{X_{0}+R_{\mathrm{s}}}\frac{sr^{-\alpha}}{1+sr^{-\alpha}}r\,\mathrm{d}r\right)$ (13) $\displaystyle\mathcal{L}_{I}(s)$ $\displaystyle=\exp{\left(-2\pi\lambda\int_{X_{0}+R_{\mathrm{s}}}^{\infty}\frac{sr^{-\alpha}}{1+sr^{-\alpha}}r\,\mathrm{d}r\right)}$ (14) ###### Proof. See Appendix A. ∎ Using Lemma 1 in (12), we can get the SINR coverage which is given in Theorem 1. ###### Theorem 1. The probability of the SINR coverage for the user located at the origin in a broadcast network with $\lambda$ density of BBSs, is given as $\displaystyle\mathrm{p_{c}}(\tau,\lambda)=$ $\displaystyle\frac{1}{2}+\frac{1}{\pi}\int_{0}^{\infty}2\pi\lambda ue^{-\pi\lambda u^{2}}\int_{0}^{\infty}\frac{1}{t}\mathsf{Im}\left[{\color[rgb]{0,0,0}\frac{e^{-jt\tau\sigma^{2}}}{1-jtu^{-\alpha}}}\right.$ $\displaystyle\hskip 56.9055pt\times\left.\exp\left(-2\pi\lambda\left(\int_{u}^{u+R_{\mathrm{s}}}\frac{-jtr^{-\alpha}}{1-jtr^{-\alpha}}r\,\mathrm{d}r+\int_{u+R_{\mathrm{s}}}^{\infty}\frac{jt\tau r^{-\alpha}}{1+jt\tau r^{-\alpha}}r\,\mathrm{d}r\right)\right)\right]\mathrm{d}t\ \mathrm{d}u$ $\displaystyle=$ $\displaystyle\frac{1}{2}+\frac{1}{\pi}\int_{0}^{\infty}\int_{0}^{\infty}2v\frac{1}{s}\left[\frac{1}{1+s^{2}v^{-2\alpha}}\right]e^{-v^{2}}e^{-s^{\frac{2}{\alpha}}M_{d}(s,v)}$ $\displaystyle\times\left[sv^{-\alpha}\cos\left(s^{\frac{2}{\alpha}}N_{d}(s,v)+\tau sK\right)-\sin\left(s^{\frac{2}{\alpha}}N_{d}(s,v)+\tau sK\right)\right]\mathrm{d}v\;\mathrm{d}s$ (15) where $M_{d}(t,\,u)$ and $N_{d}(t,\,u)$ are given as $\displaystyle M_{d}(s,v)$ $\displaystyle=\frac{1}{\alpha}\left[Q\left(\frac{1}{\alpha},s^{2}(v+\overline{m})^{-2\alpha},s^{2}v.^{-2\alpha}\right)+\tau^{2/\alpha}Q\left(\frac{1}{\alpha},0,\tau^{2}s^{2}(v+\overline{m})^{-2\alpha}\right)\right]$ (16) $\displaystyle N_{d}(s,v)$ $\displaystyle=\frac{1}{\alpha}\left[-Q\left(\frac{1}{\alpha}+\frac{1}{2},s^{2}(v+\overline{m})^{-2\alpha},s^{2}v^{-2\alpha}\right)+\tau^{\frac{2}{\alpha}}Q\left(\frac{1}{\alpha}+\frac{1}{2},0,\tau^{2}s^{2}(v+\overline{m})^{-2\alpha}\right)\right]$ (17) with $\displaystyle Q\left(z,a,b\right)=\mathsf{B}\left(z,-z+1;\frac{1}{1+a}\right)-\mathsf{B}\left(z,-z+1;\frac{1}{1+b}\right).$ (18) ###### Proof. See Appendix B. ∎ Theorem 1 provides the SINR coverage in terms of two parameters: $K$ which denotes the inverse of SNR at the cell edge and $\overline{m}^{2}$ which denotes the mean number of BBSs in connectivity radius circle. Further we can derive the following remarks. ###### Remark 1. For interference limited scenario, $K=0$, which means the coverage probability is a function of $\overline{m}$ only. In case $\overline{m}$ is fixed, individual variation of $\lambda$ and $R_{\mathrm{s}}$ will not change the coverage. ###### Remark 2. For a broadcast network, an increase in the BBS density $\lambda$ improves both the desired signal power and the interference power. However, due to increase in number of serving BBSs due to increase in $\lambda$ which improve the overall SINR coverage (which can also be seen in the numerical results). This behavior is different than conventional cellular case. Recall that with single serving BS density, the SINR in an interference-limited cellular network does not get affected by any increase in the BS density which is known as SINR invariance [6]. This can be shown from (15) by performing a comparative study between $\lambda$ and $\lambda(1+\epsilon)$ with $\epsilon<1$ for some $\lambda$. ###### Remark 3. It can been observed that the SINR coverage probability increases with an increases connectivity radius $R_{\mathrm{s}}$ as it increases the serving power and decreases the interference. ### III-C Rate Coverage Probability The rate coverage probability of a user is defined as the probability that the maximum achievable rate for the considered user is above some threshold $\rho$ i.e. $\displaystyle\mathrm{r_{c}}(\rho)$ $\displaystyle=\mathbb{P}\left[\mathtt{Rate}>\rho\right].$ Note that the maximum achievable rate for the typical user is given as $\displaystyle\mathtt{Rate}$ $\displaystyle=\xi W\log_{2}(1+\mathtt{SINR})$ (19) where $\xi$ is some coefficient that denotes the spectrum utilization. $W$ denotes the system bandwidth available to each BBS. Hence, the rate coverage for the typical user is $\displaystyle\mathrm{r_{c}}(\rho)$ $\displaystyle=\mathbb{P}\left[\mathtt{Rate}>\rho\right]$ $\displaystyle=\mathbb{P}\left[\xi W\log_{2}(1+\mathtt{SINR})>\rho\right]$ $\displaystyle=\mathbb{P}\left[\mathtt{SINR}>2^{\rho/(\xi W)}-1\right]=\mathrm{p_{c}}(2^{\rho/(\xi W)}-1)$ (20) where $\mathrm{p_{c}}$ is the SINR coverage probability given in (15). Note that the available bandwidth $W$ affects $T_{\mathrm{s}}$ and hence, $R_{\mathrm{s}}$. If the BBSs use orthogonal frequency division multiplexing (OFDM) for transmission with FFT size $N_{\mathrm{s}}$, then, $W$ is related to $T_{\mathrm{s}}$ as $W=\frac{N_{\mathrm{s}}}{T_{\mathrm{s}}}.$ Hence, the connectivity radius is $\displaystyle R_{\mathrm{s}}=T_{\mathrm{s}}\delta c=\frac{N_{\mathrm{s}}\delta c}{W}.$ (21) Hence, an increase in the system bandwidth increases the pre-log factor in (19), however, it also decreases the connectivity radius resulting in the lower SINR coverage probability. Therefore, we can observe a trade-off on the rate coverage with increasing bandwidth. ### III-D Numerical Results We now validate our results for SINR and rate coverage probabilities through numerical simulation. We will also explore the impact of different parameters on the coverage probabilities via numerical evaluations of derived expressions to develop design insights. The default parameters are given in Table I which are according to [17, 18]. TABLE I: Default parameters for numerical evaluations Parameters | Numerical value | Parameters | Numerical value ---|---|---|--- $R_{\mathrm{s}}$ | 19.18 km | $\lambda$ | 0.0014 BBSs / $\text{km}^{2}$ $N$ | 0 | Path-loss $a,\ \alpha$ | $1.6\times 10^{-3},\ 4$ $W$ | 8 MHz | $p_{\mathsf{t~{}\\!}}$ | 20 dB $N_{\mathrm{s}}\delta$ | 512 | Coefficients | $a=10^{-3},\ \xi=1$ $N_{\mathrm{c}}$ | 15 | Simulation radius | $800$ km Figure 2: SINR coverage vs. SINR threshold ($\tau$) for various BBS density $\lambda$ in a broadcast system with multiple BBSs. Here, the solid lines represent the analytical expression and markers represent simulation values. The parameters are according to Table I. It can be seen that the analysis matches with simulation results. Figure 3: Rate coverage vs. rate threshold ($\rho$) for various BBS density $\lambda$ in a broadcast system with multiple BBSs. Here, the solid lines represent the analytical expression and markers represent simulation values. The parameters are according to Table I. It can be seen that the analysis matches with simulation results. Validation of results: Fig. 2 shows the SINR coverage probability vs SINR threshold ($\tau$) for different values of BBSs density ($\lambda$). Here, the solid lines represent the analytical expression and markers represent simulation values. It can be seen that the analysis matches with simulation results which establishes the validity of the presented analytical results. From Fig. 2, it can be seen that SINR coverage increases with an increase in the BBS density consistent with Remark 2. Similarly, Fig. 3 shows the rate coverage probability vs rate threshold ($\rho$) for different values of $\lambda$. It can be seen that the rate coverage increases with an increase in $\lambda$ which is expected due to the SINR coverage behavior with $\lambda$. Figure 4: SINR coverage vs. $R_{\mathrm{s}}$ for different values of SINR threshold ($\tau$) in a broadcast network. Here, bandwidth varies with $R_{\mathrm{s}}$ according to (21) with maximum value at 80 MHz. The rest of the parameters are according to Table I. It is observed that the SINR improves with $R_{\mathrm{s}}$. Figure 5: Rate coverage vs. $R_{\mathrm{s}}$ for different values of rate threshold ($\rho$ in Mbps) in a broadcast network. Here, bandwidth varies with $R_{\mathrm{s}}$ according to (21) with maximum value at 50 MHz. The rest of the parameters are according to Table I. A trade- off is seen in rate the with varying $R_{\mathrm{s}}$. Impact of connectivity radius on SINR and rate coverage: Fig. 4 shows the variation of SINR coverage with the connectivity radius ($R_{\mathrm{s}}$) for different values of target SINR threshold. It is observed that the SINR coverage increases with $R_{\mathrm{s}}$. It can be justified as $R_{\mathrm{s}}$ increases the number of contributing BBSs increases and the number of interfering BBSs decreases. Fig. 5 shows the variation of the rate coverage with $R_{\mathrm{s}}$. We can observe that with $R_{\mathrm{s}}$, the rate coverage first increases up to a certain value of $R_{\mathrm{s}}$ and afterwards starts decreasing again. From (21), increase in $R_{\mathrm{s}}$ requires a decrease in the bandwidth $W$ in order to allow larger symbol time. This causes a trade-off in the system performance. As bandwidth is a pre-log factor in the rate expression, it has a negative and larger impact on the rate coverage. Hence, beyond a certain value of $R_{\mathrm{s}}$, the impact of decrease in $W$ dominates the increase in SINR caused by increased $R_{\mathrm{s}}$ which results in the decrease in the rate. Due to the same reasons, there may exist an optimal $R_{\mathrm{s}}$ that maximizes the rate coverage. The knowledge of optimal $R_{\mathrm{s}}$ can be helpful in designing the broadcast network. Impact of network density on SINR coverage: Fig. 6(a) and (b) show the variation of SINR and rate coverage with the network density $\lambda$ for different value of target SINR threshold and connectivity radius (which is achieved by changing bandwidth while keeping other parameters the same as Table I). The coverage while ignoring the noise is also shown. From Fig. 5(a), it can be seen that densification of the network helps in SINR coverage. When the BBS density is very small, network is noise limited. As $\lambda$ increases, BBSs comes closer to the user improving SINR coverage while SIR coverage remains constant. After a certain $\lambda$, the increase in $\lambda$ reduces the interference also. Hence, both SIR and SINR coverage improves. At high value of $\lambda$, coverage becomes 1 as all dominant BBSs provide serving power. The behavior of SIR coverage is similar to as seen in networks with dual-slope pathloss [19]. Figure 6: Coverage vs. BBS density $\lambda$ for different values of $R_{\mathrm{s}}$ in a broadcast network. Here, rest of the parameters are according to Table I. (a) SINR coverage. Dashed lines indicate SIR coverage while ignoring noise. (b) Rate coverage. Dashed lines indicate rate coverage ignoring the noise. ## IV Scenario I: Coverage Analysis for Networks with Users having a High Level of Spatial Heterogeneity in Content Preference We now extend the system model to include networks with users having their individual content or advertisement preferences. In this section, we consider the first scenario where there is high level of spatial heterogeneity in users. This means that all classes of users are present in any region. Given the limitation of resources, network selects $n$ classes of users and shows $n$ contents (one for each class) at any point of time. Here, $n$ is a design parameter decided by the broadcast network. Since the user classes are spatially inseparable, each BBSs should transmit to the same $n$ contents to improve SINR coverage. We have assumed OFDM based transmission where a BBS transmits the $n$ number of contents on orthogonal resources. ### IV-A SINR and SINR Coverage To improve coverage, the network can use the same bands for a particular content across all BBSs. Since all the BBSs are transmitting the same data in a band, the SINR for a typical user is the same as given in (5). Similarly, in this case, the SINR coverage probability of a typical user is the same as given in Theorem 1. ### IV-B Rate Coverage Now, the available resources are divided among $n$ contents. If the total available bandwidth is $W$, the bandwidth available for each content is $W/n$. The instantaneous achievable rate for a typical user located at origin, for each content is $\displaystyle\mathtt{Rate}$ $\displaystyle=\xi\frac{W}{n}\log_{2}(1+\mathtt{SINR}).$ (22) From (20), the rate coverage probability is given as: $\displaystyle\mathrm{r_{c}}(\rho)$ $\displaystyle=\mathbb{P}\left[\mathtt{SINR}>2^{n\rho/(\xi W)}-1\right]=\mathrm{p_{c}}(2^{n\rho/(\xi W)}-1)$ (23) where $\mathrm{p_{c}}$ is given in (15). ### IV-C Network Revenue Let $\rho$ denote the minimum rate required for a user to be able to view the content. Then, the rate coverage $\mathrm{r_{c}}$ at $\rho$ denotes the fraction of users that are able to view this content. Therefore, $\mathrm{r}_{c}$ unit of revenue will be earned by the network from a particular class, since only $\mathrm{r_{c}}$ fraction of users can watch it. Therefore, the network’s total revenue $\mathrm{R}_{\mathrm{n}}$ can be given as: $\displaystyle\mathrm{R}_{\mathrm{n}}$ $\displaystyle=n\mathrm{r}_{c}(\rho).$ (24) Figure 7: Variation of total revenue with respect to allowed number of user classes for different rate threshold $\rho$ (Mbps) (which is a proxy for content quality requested). Content granularity is $N_{\mathrm{c}}=15$. Other parameters are according to Table I. It is observed that an optimal value of $n$ can provide the maximum revenue to the network which depends on the content quality requested. ### IV-D Numerical Results: Impact of $n$ on Total Revenue Fig. 7 shows the variation of the network revenue $\mathrm{R}_{\mathrm{n}}$ with $n$ for a system with $N_{\mathrm{c}}=15$. We can observe that initially, the revenue increases with an increase in $n$ up to a certain value, and thereafter, starts decreasing. This can be justified in the following way. If $n$ increases, the following two effects take place– (1) more user classes are served, causing a linear increase in $\mathrm{R}_{\mathrm{n}}$, and (2) available bandwidth $W/n$ for each content/advertisement decreases causing the rate coverage to drop. As a combined effect dictated by (24), the revenue is optimal at a particular value of $n$. However, this behavior also depends on the target rate threshold. At a higher rate threshold, $\mathrm{R}_{\mathrm{n}}$ decreases with $n$, showing $n=1$ as the optimal choice. This implies that the optimal number of user classes that can be served, depends on the quality of the content. If the quality requested is high, then it may be better to serve less user classes, while more user classes can be served when the quality requested is lower. ## V Scenario II: Coverage Analysis for Networks with Users having Spatially Separated Classes for Content Preference The second scenario we consider corresponds to the case where user classes are spatially separated. For tractability, we assume that coverage area of each BBS comprises of users from a single class. Thus, users of different classes are spatially separated. Let $S_{i}$ denote the class of the $i$th BBS located at $\mathbf{X}_{i}$ which is a uniform discrete random variable with PMF given as $\displaystyle p_{S_{i}}(k)$ $\displaystyle=\frac{1}{N_{\mathrm{c}}}\mathbbm{1}\left({1\leq k\leq N_{\mathrm{c}}}\right).$ (25) We assume that $S_{i}$’s are independent of each other. The network can cater to all user classes by letting each BBS to transmit content according to the class of users lying in its coverage area. Note that for a typical user, only those BBSs that transmit the same content, can contribute to the serving signal power at this user. Therefore, this strategy will reduce the number of serving BSs and hence reduce the SINR. On the other hand, network can decide to cater only one user class by forcing all BBSs to transmit only one content, will reduce the revenue as users of only one class will receive their preferred content. It will be interesting to find the optimal number of user-classes that can be catered by the network. As a general problem, we consider that the network decides to cater $n$ user classes out of total $N_{\mathrm{c}}$ classes. Let us denote the set of all selected index by $\mathcal{M}$. Let us denote the content transmitted by $i$th BS by $M_{i}$. If the BBS’s user class is one of the $n$ selected classes (i.e. $S_{i}\in\mathcal{M}$), it will transmit the content corresponding to its user class i.e. $M_{i}=S_{i}$. If the BBS’s user class is not one of the selected classes, it will transmit the content corresponding to a randomly selected user class to help boost its signal strength. Let us consider a typical user at $\mathbf{o}$. Without loss of generality, assume that its user class is 1. The probability that its class is one of the $n$ selected classes is $n/N_{\mathrm{c}}$. Let us condition on the fact that it is one of the selected classes. Let the tagged BS of this typical user transmits the content $M_{0}=1$. Then, for the $i$th BBS, $M_{i}=M_{0}$ if 1. 1. $S_{i}=M_{0}$ which occurs with probability $\frac{1}{N_{\mathrm{c}}}$, or 2. 2. $S_{i}\notin\mathcal{M}$ and $S_{i}=M_{i}$ which occurs with probability $\frac{N_{\mathrm{c}}-n}{N_{\mathrm{c}}}\cdot\frac{1}{n}$. Therefore, the probability that the $i$th BBS is transmitting the content as per the preference of this typical user is $\displaystyle p$ $\displaystyle=\mathbb{P}\left[M_{i}=M_{0}\right]=\frac{1}{N_{\mathrm{c}}}+\frac{N_{\mathrm{c}}-n}{N_{\mathrm{c}}}\cdot\frac{1}{n}=\frac{1}{n}.$ (26) ### V-A SINR Now, note that the BBSs that are transmitting the same content as $M_{0}$ and are located inside $\mathcal{B}(\mathbf{o},{X_{0}}+R_{\mathrm{s}})$ will contribute to the desired signal power to the typical user at origin $\mathbf{o}$. Therefore, the desired signal power for the typical user at origin is given as $\displaystyle S=$ $\displaystyle p_{\mathsf{t~{}\\!}}a\beta_{0}{{X_{0}}}^{-\alpha}+\sum_{\mathbf{X}_{i}\in\Phi\cap\mathcal{B}(0,X_{0}+R_{\mathrm{s}})\setminus\mathbf{X}_{0}}p_{\mathsf{t~{}\\!}}a\beta_{i}{{X_{i}}}^{-\alpha}\mathbbm{1}\left({M_{i}=M_{0}}\right).$ (27) Similarly, the interference for the typical user is caused by the BBSs that are either located outside $\mathcal{B}(\mathbf{o},{u}+R_{\mathrm{s}})$ or located inside the $\mathcal{B}(\mathbf{o},{X_{0}}+R_{\mathrm{s}})$ but transmitting a different content than $M_{0}$. Hence, the total interference is given as $\displaystyle I=$ $\displaystyle\sum_{\mathbf{X}_{i}\in\Phi\cap\mathcal{B}(0,X_{0}+R_{\mathrm{s}})\setminus\mathbf{X}_{0}}p_{\mathsf{t~{}\\!}}a\beta_{i}{{X_{i}}}^{-\alpha}\mathbbm{1}\left({M_{i}\neq M_{0}}\right)+\sum_{\mathbf{X}_{j}\in\Phi\cap\mathcal{B}(0,{X_{0}}+R_{\mathrm{s}})^{\complement}}p_{\mathsf{t~{}\\!}}a\beta_{j}{{X_{j}}}^{-\alpha}.$ (28) Now, the SINR for this user is given as $\displaystyle\mathtt{SINR}=\frac{S}{I+N}$ $\displaystyle=\frac{\beta_{0}{\|\mathbf{X}_{0}\|}^{-\alpha}+\sum_{\mathbf{X}_{i}\in\Phi\cap\mathcal{B}(0,u+R_{\mathrm{s}})\setminus\mathbf{X}_{0}}\beta_{i}{\|\mathbf{X}_{i}\|}^{-\alpha}\mathbbm{1}\left({M_{i}=M_{0}}\right)}{\sum_{\mathbf{X}_{i}\in\Phi\cap\mathcal{B}(0,u+R_{\mathrm{s}})\setminus\mathbf{X}_{0}}\beta_{i}{\|\mathbf{X}_{i}\|}^{-\alpha}\mathbbm{1}\left({M_{i}\neq M_{0}}\right)+\sum_{\mathbf{X}_{j}\in\Phi\cap\mathcal{B}(0,u+R_{\mathrm{s}})^{\complement}}\beta_{j}{\|\mathbf{X}_{j}\|}^{-\alpha}+\sigma^{2}}.$ (29) Here, $\sigma^{2}$ is the normalized noise power which is given as $\sigma^{2}=N/(p_{\mathsf{t~{}\\!}}a)$ where $N$ is the noise power. ### V-B SINR Coverage Probability We now calculate the SINR coverage for a typical user. Similar to Section III-B, the SINR coverage probability is given as $\displaystyle\mathrm{p_{c}}(\tau,\lambda)$ $\displaystyle=\frac{1}{2}+\frac{1}{\pi}\int_{0}^{\infty}\int_{0}^{\infty}2\pi\lambda ue^{-\pi\lambda u^{2}}\frac{1}{t}\mathsf{Im}\left[\mathcal{L}_{I|\,\mathbf{X}_{0}}(jt\tau)e^{-jt\tau\sigma^{2}}\mathcal{L}_{S|\,\mathbf{X}_{0}}(-jt)\right]\mathrm{d}t\ \mathrm{d}u$ (30) where $S$ and $I$ are given in (27) and (28). ###### Lemma 2. Conditioned on the location of the closest serving BBS, the Laplace transforms of the desired signal power and the sum interference at the receiver are given as $\displaystyle\mathcal{L}_{S|\,\mathbf{X}_{0}}(s)=\frac{1}{1+su^{-\alpha}}\exp\left(-2\pi\lambda p\int_{X_{0}}^{X_{0}+R_{\mathrm{s}}}\frac{sr^{-\alpha}}{1+sr^{-\alpha}}r\mathrm{d}r\right)$ (31) $\displaystyle\mathcal{L}_{I|\mathbf{X}_{0}}(s)=\exp\left(-2\pi\lambda(1-p)\int_{X_{0}}^{X_{0}+R_{\mathrm{s}}}\frac{sr^{-\alpha}}{1+sr^{-\alpha}}r\mathrm{d}r-2\pi\lambda\int_{X_{0}+R_{\mathrm{s}}}^{\infty}\frac{sr^{-\alpha}}{1+sr^{-\alpha}}r\mathrm{d}r\right).$ (32) ###### Proof. See Appendix C. ∎ Using Lemma 2 and (30) we can calculate the SINR coverage which is given in Theorem 2. ###### Theorem 2. The probability of SINR coverage for the user located at the origin in a broadcast network with $\lambda$ density of BBSs, is given as: $\displaystyle\mathrm{p_{c}}(\tau,\lambda)=$ $\displaystyle\frac{1}{2}+\frac{1}{\pi}\int_{0}^{\infty}2\pi\lambda ue^{-\pi\lambda u^{2}}\int_{0}^{\infty}\frac{1}{t}\mathsf{Im}\left[{\color[rgb]{0,0,0}\frac{e^{-jt\tau\sigma^{2}}}{1-jtu^{-\alpha}}}\right.$ $\displaystyle\indent\times\left.\exp\left(-2\pi\lambda\left(p\int_{u}^{u+R_{\mathrm{s}}}\frac{-jtr^{-\alpha}}{1-jtr^{-\alpha}}r\,\mathrm{d}r+(1-p)\int_{u}^{u+R_{\mathrm{s}}}\frac{jt\tau r^{-\alpha}}{1+jt\tau r^{-\alpha}}r\,\mathrm{d}r\right.\right.\right.$ $\displaystyle\hskip 42.67912pt\left.\left.\left.+\int_{u+R_{\mathrm{s}}}^{\infty}\frac{jt\tau r^{-\alpha}}{1+jt\tau r^{-\alpha}}r\,\mathrm{d}r\right)\right)\right]\mathrm{d}t\mathrm{d}u$ $\displaystyle=$ $\displaystyle\frac{1}{2}+\frac{1}{\pi}\int_{0}^{\infty}\int_{0}^{\infty}2v\frac{1}{s}\left[\frac{1}{1+s^{2}v^{-2\alpha}}\right]e^{-v^{2}}e^{-s^{\frac{2}{\alpha}}M^{\prime}_{d}(s,v)}$ $\displaystyle\times\left[sv^{-\alpha}\cos\left(s^{\frac{2}{\alpha}}N^{\prime}_{d}(s,v)+\tau sK\right)-\sin\left(s^{\frac{2}{\alpha}}N^{\prime}_{d}(s,v)+\tau sK\right)\right]\mathrm{d}v\;\mathrm{d}s$ (33) where $M^{\prime}_{d}(t,\,u)$ and $N^{\prime}_{d}(t,\,u)$ is given as $\displaystyle M_{d}(s,v)=$ $\displaystyle\frac{1}{\alpha}\left[pQ\left(\frac{1}{\alpha},s^{2}(v+\overline{m})^{-2\alpha},s^{2}v^{-2\alpha}\right)+(1-p)\tau^{2/\alpha}Q\left(\frac{1}{\alpha},\tau^{2}s^{2}(v+\overline{m})^{-2\alpha},\tau^{2}s^{2}v^{-2\alpha}\right)\right.$ $\displaystyle\left.+\tau^{2/\alpha}Q\left(\frac{1}{\alpha},0,\tau^{2}s^{2}(v+\overline{m})^{-2\alpha}\right)\right]$ (34) $\displaystyle N_{d}(s,v)=$ $\displaystyle\frac{1}{\alpha}\left[-pQ\left(\frac{1}{\alpha}+\frac{1}{2},s^{2}(v+\overline{m})^{-2\alpha},s^{2}v^{-2\alpha}\right)\right.$ $\displaystyle\left.+(1-p)\tau^{2/\alpha}Q\left(\frac{1}{\alpha}+\frac{1}{2},\tau^{2}s^{2}(v+\overline{m})^{-2\alpha},\tau^{2}s^{2}v^{-2\alpha}\right)\right.$ $\displaystyle\left.+\tau^{2/\alpha}Q\left(\frac{1}{\alpha}+\frac{1}{2},0,\tau^{2}s^{2}(v+\overline{m})^{-2\alpha}\right)\right]$ (35) ###### Proof. See Appendix D. ∎ ### V-C Rate Coverage Since each BBS shows only one content, it can use the total available bandwidth $W$. The instantaneous achievable rate for a typical user located at origin, while receiving the content, is $\displaystyle\mathtt{Rate}$ $\displaystyle=\xi{W}\log_{2}(1+\mathtt{SINR}).$ (36) From (20), the rate coverage probability is given as: $\displaystyle\mathrm{r_{c}}(\rho)$ $\displaystyle=\mathbb{P}\left[\mathtt{SINR}>2^{\rho/(\xi W)}-1\right]=\mathrm{p_{c}}(2^{\rho/(\xi W)}-1)$ (37) where $\mathrm{p_{c}}$ is given in Theorem 2. ### V-D Network Revenue Let $\rho$ denote the minimum rate required for a user to be able to view the content. Then, the rate coverage $\mathrm{r_{c}}$ at $\rho$ denotes the fraction of users that are able to view this content. Therefore, $\mathrm{r}_{c}$ unit of revenue will be earned by the network from a particular class, since only $\mathrm{r_{c}}$ fraction of users can watch it. The probability that the typical user receives the content as per its preference is $n/N_{\mathrm{c}}$ which is also the probability that the network will receive revenue from this typical user. Similar to previous sections, the total revenue can be computed as $\displaystyle\mathrm{R}_{\mathrm{n}}$ $\displaystyle=\frac{n}{N_{\mathrm{c}}}\cdot\mathrm{r_{c}}(\rho),$ (38) where the rate coverage $\mathrm{r_{c}}$ for the typical user is given by (37). ### V-E Numerical Results We now present numerical results for the considered scenario II. The parameters are stated in Table I. Figure 8: Variation of SINR and rate coverage probability with respect to allowed number of user classes for different values of connectivity radius ($R_{\mathrm{s}}$ in km) at a typical user in a broadcast system with geographically separated user classes. Content granularity is $N_{\mathrm{c}}=15$, SINR threshold $\tau=10$ dB and rate threshold $\rho=15$ Mbps. Network density is $\lambda=.014$/km2. The bandwidth varies with $R_{\mathrm{s}}$ according to (21) with maximum value at 80 MHz. Other parameters are according to Table I. The coverage decreases with $n$ due to increased interference at the typical user. Impact of $n$ on SINR and rate coverage probability: Fig. 8(a) shows the variation of SINR coverage with $n$ for different values of connectivity radius $R_{\mathrm{s}}$. Here, SINR threshold $\tau=10$ dB and there are $N_{\mathrm{c}}=15$ user classes. From Fig. 8, we observe that for a fix value of $R_{\mathrm{s}}$, SINR coverage decreases with increase in $n$. This is due to the fact that more BBSs interfere as $n$ increases. However, after a certain $n$, the coverage doesn’t changes much with $n$. This is because additional fraction of BBSs that interfere when $n$ increases by 1, is equal to $\frac{n}{n+1}-\frac{n-1}{n}=\frac{1}{n(n+1)}$ which decreases very fast with $n$. Therefore, after a certain $n$, there will not be a significant increase in the interference, which makes the SINR constant with $n$. It is also observed that the decrease in the SINR coverage probability is faster when $R_{\mathrm{s}}$ is large. This can be justified as follows. First note that $n$ only affects the BBSs that can either be a interferer or a serving BBS, depending on the content they are showing. These BBSs lie in the ring of $R_{\mathrm{s}}$ width denoting the region and their number approximately scales as $\lambda\pi R_{\mathrm{s}}^{2}$. Note that this number is large when $R_{\mathrm{s}}$ is large. When we allow BBSs to show more advertisements/contents, a large number of these BBSs means that there is a larger number of potential interferes. When $R_{\mathrm{s}}$ is small, there are less number of these potential BBSs (or even 0), hence allowing more advertisements doesn’t affect the coverage significantly. Fig. 8(b) shows the variation of rate coverage with respect to $n$ for different size of connectivity region. The rate coverage also follows similar behavior as SINR coverage as described in (37). Figure 9: Variation of the total network revenue with respect to allowed number of user classes $n$ for different values of rate threshold $\rho$ (in Mbps) with $R_{\mathrm{s}}=150\,\text{km}$ for a broadcast system with geographically separated user classes. Content granularity is $N_{\mathrm{c}}=15$. Here, bandwidth varies with $R_{\mathrm{s}}$ according to (21). Other parameters are according to Table I. Impact of $n$ on the total network revenue: Fig. 9 shows the impact of $n$ on the total network revenue for different values of rate threshold. Increase in $n$ means catering to more number of user classes. From Fig. 9, we can observe that the revenue initially decreases and then, increases with $n$. The initial decrease in the revenue seen from $n=1$ to $n=2$ for some configurations is due to the decrease in rate coverage from $n=1$ to $n=2$, as observed in Fig. 8(b) which dominated the increase in the revenue generated due to catering to an additional user class. However, this behavior may depend on the target rate threshold and the value of $R_{\mathrm{s}}$. Figure 10: Variation of the total network revenue with respect to the connectivity radius $R_{\mathrm{s}}$ (in km) for different values of $n$ with $\rho=5\,\text{Mbps}$ for a broadcast system with geographically separated user classes. Content granularity is $N_{\mathrm{c}}=15$. Here, bandwidth varies with $R_{\mathrm{s}}$ according to (21) with maximum value at 50 MHz. Other parameters are according to Table I. Impact of $R_{\mathrm{s}}$ on the total network revenue Fig. 10 shows the total revenue with respect to size of connectivity region $R_{\mathrm{s}}$ for different $n$ with $\rho=10\,\text{Mbps}$ and $N_{\mathrm{c}}=15$. We can observe that, the revenue decreases with increase in $R_{\mathrm{s}}$. Also the revenue increases with $n$ for lower values of $R_{\mathrm{s}}$. For a higher values of $R_{\mathrm{s}}$, the behavior with $n$ is not monotonic. The revenue for $n\geq 1$ may fall below the revenue for $n=1$. As discussed previously, the rate coverage decreases drastically from $n=1$ to $n=2$ for large $R_{\mathrm{s}}$ which can dominate the increase in the revenue generated due to catering to an additional user class. ## VI Conclusions In this paper, we presented an analytical framework for the system performance of a broadcast network using stochastic geometry. Since all BBSs in the broadcast network are transmitting the same signal, signals from multiple BSs can be used to improve the coverage. We show that there exists a region such that all BBSs lying in this region may contribute to the desired signal power. We computed the SINR and rate coverage probability for a typical user located at the origin. We validated our results using numerical analysis. Using these results, we found that there exists an optimal region size which maximized the rate coverage. When users consists of many user classes having heterogenous content preference, network can schedule content to maximize its revenue. We presented an analytical model of revenue thus obtained from users. The results are validated through numerical analysis. We also present the variation of total revenue with respect to various parameters including number of user classes to be catered, size of connectivity region, and rate threshold. We show how content quality also affects the network decision on variety of content shown by the operator. ## Appendix A Proof for Lemma 1 Using (4), the Laplace transform of the sum interference $I|\mathbf{X}_{0}$ is given as $\displaystyle\mathcal{L}_{I\,|\,\mathbf{X}_{0}}(s)=\mathbb{E}\left[e^{-sI}|\mathbf{X}_{0}\right]=\mathbb{E}\left[\exp{\left(-s\sum_{\mathbf{X}_{j}\in\Phi\cap\mathcal{B}(0,{X_{0}}+R_{\mathrm{s}})^{\complement}}\beta_{j}{\|\mathbf{X}_{j}\|}^{-\alpha}\right)}\right]$ $\displaystyle\overset{(a)}{=}\exp{\left(-\lambda\int_{\Phi\cap\mathcal{B}(0,{X_{0}}+R_{\mathrm{s}})^{\complement}}\left(1-\mathbb{E}_{\beta}\left[e^{-s\beta\|\mathbf{X}\|^{-\alpha}}\right]\right)\mathrm{d}{\mathbf{x}}\right)}$ $\displaystyle\stackrel{{\scriptstyle(b)}}{{=}}\exp{\left(-2\pi\lambda\int_{X_{0}+R_{\mathrm{s}}}^{\infty}\left(1-\mathbb{E}_{\beta}\left[e^{-s\beta r^{-\alpha}}\right]\right)r\,\mathrm{d}r\right)}$ (39) where (a) is due to the probability generating functional (PGFL) of homogeneous PPP [7] and (b) is due to conversion to polar coordinates. Now, since $\beta$’s are exponetially distributed, using their MGF, we get $\displaystyle\mathcal{L}_{I}(s)=\exp{\left(-2\pi\lambda\int_{u+R_{\mathrm{s}}}^{\infty}\frac{sr^{-\alpha}}{1+sr^{-\alpha}}r\,\mathrm{d}r\right)}.$ (40) Similarly, the Laplace transform of desired signal power conditioned on the nearest is BS located at $\mathbf{X}_{0}$, is given as: $\displaystyle\mathcal{L}_{S\,|\,\mathbf{X}_{0}}(s)=\mathbb{E}_{S|\,\mathbf{X}_{0}}\left[e^{-sS}|\mathbf{X}_{0}\right]$ $\displaystyle=$ $\displaystyle\mathbb{E}_{\beta_{i},\mathbf{X}_{i}}\left[\exp{\left(-s\left(\beta_{0}\|\mathbf{X}_{0}\|^{-\alpha}+\sum_{\mathbf{X}_{i}\in\Phi\cap\mathcal{B}{(\mathbf{o},X_{0}+R_{\mathrm{s}})}\setminus\mathbf{X}_{0}}\beta_{i}\|\mathbf{X}_{i}\|^{-\alpha}\right)\right)}\right]$ $\displaystyle=$ $\displaystyle\mathbb{E}_{\Phi\,|\,\mathbf{X}_{0}}\left[\mathbb{E}_{\beta_{0}|\,\mathbf{X}_{0}}\left[e^{-s\beta_{0}\|\mathbf{X}_{0}\|^{-\alpha}}\right]\right.$ $\displaystyle\times\left.\prod_{\mathbf{X}_{i}\in\Phi\cap\mathcal{B}{(\mathbf{o},X_{0}+R_{\mathrm{s}})}\setminus\mathbf{X}_{0}}\mathbb{E}_{\beta_{i}|X_{i},\mathbf{X}_{0}}\left[e^{-s\beta_{i}\|\mathbf{X}_{i}\|^{-\alpha})}\right]\right]$ (41) Now, from Slivnyak theorem [7], we know that conditioned on $\mathbf{X}_{0}$, $\\{\mathbf{X}_{i}:\mathbf{X}_{i}\in\Phi\\}\setminus\\{\mathbf{X}_{0}\\}$ is a PPP with the same density. Therefore, using the PGFL of a PPP and noting that $\beta_{i}$’s are exponential RVs, we get $\displaystyle\mathcal{L}_{S\,|\,\mathbf{X}_{0}}(s)=$ $\displaystyle\frac{1}{1+s\|\mathbf{X}_{0}\|^{-\alpha}}\exp\left(-2\pi\lambda\int_{X_{0}}^{X_{0}+R_{\mathrm{s}}}\left(1-\frac{1}{1+sr^{-\alpha}}\right)r\mathrm{d}r\right)$ $\displaystyle=$ $\displaystyle\frac{1}{1+sX_{0}^{-\alpha}}\exp\left(-2\pi\lambda\int_{X_{0}}^{X_{0}+R_{\mathrm{s}}}\frac{sr^{-\alpha}}{1+sr^{-\alpha}}r\mathrm{d}r\right).$ (42) ## Appendix B Proof of Theorem 1 Using Lemma 1, we get $\displaystyle\mathcal{L}_{S|\,X_{0}=u}(-jt)$ $\displaystyle=\frac{1+jtu^{-\alpha}}{1+t^{2}u^{-2\alpha}}\exp\left(-2\pi\lambda\int_{u}^{u+R_{\mathrm{s}}}\frac{-jtr^{-\alpha}}{1-jtr^{-\alpha}}r\mathrm{d}r\right)$ (43) $\displaystyle\mathcal{L}_{I|\,X_{0}=u}(jt\tau)$ $\displaystyle=\exp{\left(-2\pi\lambda\int_{u+R_{\mathrm{s}}}^{\infty}\frac{jt\tau r^{-\alpha}}{1+jt\tau r^{-\alpha}}r\,\mathrm{d}r\right)}.$ (44) Substituting the above values in (11), we get $\displaystyle\mathbb{P}\left(S>(I+\sigma^{2})\tau|\,X_{0}=u\right)$ $\displaystyle=\frac{1}{2}+\frac{1}{\pi}\int_{0}^{\infty}\frac{1}{t}\mathsf{Im}\left[\left(\frac{1+jtu^{-\alpha}}{1+t^{2}u^{-2\alpha}}\right)e^{-jt\tau\sigma^{2}}\times\right.$ $\displaystyle\hskip 14.22636pt\left.\exp\left(-2\pi\lambda\left(\int_{u+R_{\mathrm{s}}}^{\infty}\frac{jt\tau r^{-\alpha}}{1+jt\tau r^{-\alpha}}r\,\mathrm{d}r+\int_{u}^{u+R_{\mathrm{s}}}\frac{-jtr^{-\alpha}}{1-jtr^{-\alpha}}r\,\mathrm{d}r\right)\right)\right]\mathrm{d}t$ $\displaystyle=\frac{1}{2}+\frac{1}{\pi}\int_{0}^{\infty}\frac{1}{t}\mathsf{Im}\left[\left(\frac{1+jtu^{-\alpha}}{1+t^{2}u^{-2\alpha}}\right)e^{-jt\tau\sigma^{2}}\times\right.$ $\displaystyle\left.\exp\left(-2\pi\lambda\left(\int_{u+R_{\mathrm{s}}}^{\infty}\frac{t^{2}\tau^{2}r^{-2\alpha+1}}{1+t^{2}\tau^{2}r^{-2\alpha}}\mathrm{d}r+\int_{u}^{u+R_{\mathrm{s}}}\frac{t^{2}r^{-2\alpha+1}}{1+t^{2}r^{-2\alpha}}\mathrm{d}r\right.\right.\right.$ $\displaystyle\left.\left.\left.+j\int_{u+R_{\mathrm{s}}}^{\infty}\frac{t\tau r^{-\alpha+1}}{1+t^{2}\tau^{2}r^{-2\alpha}}\mathrm{d}r\,-\,j\int_{u}^{u+R_{\mathrm{s}}}\frac{tr^{-\alpha+1}}{1+t^{2}r^{-2\alpha}}\mathrm{d}r\right)\right)\right]\mathrm{d}t$ (45) where the last step is obtained using multiplication of conjugate terms. Now, if we define $\displaystyle M(t,u)=2\alpha{t^{-2/\alpha}}\left[\int_{u}^{u+R_{\mathrm{s}}}\frac{t^{2}r^{-2\alpha+1}}{1+t^{2}r^{-2\alpha}}\mathrm{d}r+\,\int_{u+R_{\mathrm{s}}}^{\infty}\frac{t^{2}\tau^{2}r^{-2\alpha+1}}{1+t^{2}\tau^{2}r^{-2\alpha}}\mathrm{d}r\right]$ (46) $\displaystyle N(t,u)=2\alpha{t^{-2/\alpha}}\left[\int_{u+R_{\mathrm{s}}}^{\infty}\frac{t\tau r^{-\alpha+1}}{1+t^{2}\tau^{2}r^{-2\alpha}}\mathrm{d}r\,-\,\int_{u}^{u+R_{\mathrm{s}}}\frac{tr^{-\alpha+1}}{1+t^{2}r^{-2\alpha}}\mathrm{d}r\right]$ (47) (45) can be written as $\displaystyle\mathbb{P}\left(S>(I+\sigma^{2})\tau|\,X_{0}=u\right)$ $\displaystyle=\frac{1}{2}+\frac{1}{\pi}\int_{0}^{\infty}\frac{1}{t}\mathsf{Im}\left[\left(\frac{1+jtu^{-\alpha}}{1+t^{2}u^{-2\alpha}}\right)\right.$ $\displaystyle\times\left.\exp\left(-2\pi\lambda{\color[rgb]{0,0,0}t^{2/\alpha}}M(t,u)-j2\pi\lambda{\color[rgb]{0,0,0}t^{2/\alpha}}N(t,u)-jt\tau\sigma^{2}\right)\right]\mathrm{d}t.$ (48) Now, with some trivial manipulations and substituting (48) in (9), we get $\displaystyle\mathrm{p_{c}}(\tau,\lambda)=$ $\displaystyle\frac{1}{2}+\frac{1}{\pi}\int_{0}^{\infty}\int_{0}^{\infty}2\pi\lambda ue^{-\pi\lambda u^{2}}\cdot\frac{1}{t}\cdot\left[\frac{1}{1+t^{2}u^{-2\alpha}}\right]e^{-2\pi\lambda t^{2/\alpha}M(t,\,u)/2\alpha}$ $\displaystyle\times\left[tu^{-\alpha}\cos{\left(\frac{\pi}{\alpha}\lambda{t^{2/\alpha}}N(t,\,u)+\,t\tau\sigma^{2}\right)}-\sin{\left(\frac{\pi}{\alpha}\lambda{t^{2/\alpha}}N(t,\,u)+\,t\tau\sigma^{2}\right)}\right]\mathrm{d}t\mathrm{d}u.$ (49) Further, the forms of $M$ and $N$ can be simplified by trivial manipulations and definition of incomplete Beta function in (46) and (47) to get $\displaystyle M(t,u)=Q\left(\frac{1}{\alpha},t^{2}(u+R_{\mathrm{s}})^{-2\alpha},t^{2}u^{-2\alpha}\right)+\tau^{2/\alpha}Q\left(\frac{1}{\alpha},0,t^{2}\tau^{2}(u+R_{\mathrm{s}})^{-2\alpha}\right),\text{ and }$ (50) $\displaystyle N(t,\,u)=-Q\left(\frac{1}{\alpha}+\frac{1}{2},t^{2}(u+R_{\mathrm{s}})^{-2\alpha},t^{2}u^{-2\alpha}\right)+\tau^{2/\alpha}Q\left(\frac{1}{\alpha}+\frac{1}{2},0,t^{2}\tau^{2}(u+R_{\mathrm{s}})^{-2\alpha}\right)$ (51) Now, we can substitute $\displaystyle t$ $\displaystyle\rightarrow s/{(\lambda\pi)}^{\alpha/2}$ $\displaystyle u$ $\displaystyle\rightarrow v/\sqrt{\lambda\pi}\ $ (52) in (49), (50) and (51) to get the desired result. ## Appendix C Proof of Lemma 2 From (23), the Laplace transform of desired signal power $S$ conditioned that the nearest BS located at $\mathbf{X}_{0}$ is given as: $\displaystyle\mathcal{L}_{S|\,\mathbf{X}_{0}}(s)=\mathbb{E}_{S\,|\,\mathbf{X}_{0}}\left[e^{-sS}\right]$ $\displaystyle=\mathbb{E}_{\\{\beta_{i},\mathbf{X}_{i}\\}|\,\mathbf{X}_{0}}\left[\exp\left(-s\beta_{0}{X_{0}}^{-\alpha}-\sum_{\mathbf{X}_{i}\in\Phi\cap\mathcal{B}{(\mathbf{o},X_{0}+R_{\mathrm{s}})}\setminus\mathbf{X}_{0}}s\beta_{i}{X_{i}}^{-\alpha}\mathbbm{1}\left(M_{i}=M_{0}\right)\right)\right]$ which is similar to (41) except the fact that the summation in the last term is over only those points that satisfy an additional condition $M_{i}=M_{0}$. From the independent thinning theorem, these points also form a PPP with density $\lambda\mathbb{P}\left[M_{i}=M_{0}\right]=\lambda p$. Now using the PGFL of this PPP, we get $\displaystyle\mathcal{L}_{S|\,\mathbf{X}_{0}}(s)$ $\displaystyle=\frac{1}{1+sX_{0}^{-\alpha}}\exp\left(-2\pi\lambda p\int_{X_{0}}^{X_{0}+R_{\mathrm{s}}}\frac{sr^{-\alpha}}{1+sr^{-\alpha}}r\mathrm{d}r\right).$ (53) Now, from (24), the Laplace transform of sum interference is $\displaystyle\mathcal{L}_{I}(s)=\mathbb{E}_{I}\left[e^{-sI}\right]$ $\displaystyle=\mathbb{E}_{I}\left[\exp\left(-\sum_{\mathbf{X}_{i}\in\Phi\cap\mathcal{B}(0,X_{0}+R_{\mathrm{s}})\setminus\mathbf{X}_{0}}s\beta_{i}{{X_{i}}}^{-\alpha}\mathbbm{1}\left(M_{i}\neq M_{0}\right)-\sum_{\mathbf{X}_{j}\in\Phi\cap\mathcal{B}(0,X_{0}+R_{\mathrm{s}})^{\complement}}s\beta_{j}{X_{j}}^{-\alpha}\right)\right]$ $\displaystyle\stackrel{{\scriptstyle(a)}}{{=}}\exp\left(-2\pi\lambda(1-p)\int_{X_{0}}^{X_{0}+R_{\mathrm{s}}}\left(1-\mathbb{E}_{\beta}\left[e^{-s\beta r^{-\alpha}}\right]\right)r\mathrm{d}r-2\pi\lambda\int_{X_{0}+R_{\mathrm{s}}}^{\infty}\left(1-\mathbb{E}_{\beta}\left[e^{-s\beta r^{-\alpha}}\right]\right)r\mathrm{d}r\right)$ $\displaystyle\overset{(b)}{=}\exp\left(-2\pi\lambda(1-p)\int_{u}^{X_{0}+R_{\mathrm{s}}}\frac{sr^{-\alpha}}{1+sr^{-\alpha}}r\mathrm{d}r-2\pi\lambda\int_{X_{0}+R_{\mathrm{s}}}^{\infty}\frac{sr^{-\alpha}}{1+sr^{-\alpha}}r\mathrm{d}r\right),$ (54) where (a) is due to the probability generating functional of homogeneous PPP and independent thinning theorem and (b) is due to MGF of exponentially distributed $\beta_{i}$’s. ## Appendix D Proof of Theorem 2 Using Lemma 2, we get $\displaystyle\mathcal{L}_{S|\,X_{0}=u}(-jt)=\frac{1+jtu^{-\alpha}}{1+t^{2}u^{-2\alpha}}\exp\left(-2\pi\lambda p\int_{u}^{u+R_{\mathrm{s}}}\frac{-jtr^{-\alpha}}{1-jtr^{-\alpha}}rdr\right)$ (55) $\displaystyle\mathcal{L}_{I|\,X_{0}=u}(jt\tau)=\exp\left(-2\pi\lambda\left[(1-p)\int_{u}^{u+R_{\mathrm{s}}}\frac{jt\tau r^{-\alpha}}{1+jt\tau r^{-\alpha}}r\,\mathrm{d}r+\int_{u+R_{\mathrm{s}}}^{\infty}\frac{jt\tau r^{-\alpha}}{1+jt\tau r^{-\alpha}}r\,\mathrm{d}r\right]\right).$ (56) Substituting the above values in (11), we get $\displaystyle\mathbb{P}\left(S>(I+\sigma^{2})\tau|\,X_{0}=u\right)$ $\displaystyle=$ $\displaystyle\frac{1}{2}+\frac{1}{\pi}\int_{0}^{\infty}\frac{1}{t}\mathsf{Im}\left[\left(\frac{1+jtu^{-\alpha}}{1+t^{2}u^{-2\alpha}}\right)e^{-jt\tau\sigma^{2}}\cdot\right.$ $\displaystyle\left.\exp\left(-2\pi\lambda\left(\int_{u+R_{\mathrm{s}}}^{\infty}\frac{jt\tau r^{-\alpha}}{1+jt\tau r^{-\alpha}}r\,\mathrm{d}r+(1-p)\int_{u}^{u+R_{\mathrm{s}}}\frac{jt\tau r^{-\alpha}}{1+jt\tau r^{-\alpha}}r\,\mathrm{d}r\right.\right.\right.$ $\displaystyle\indent\left.\left.\left.+p\int_{u}^{u+R_{\mathrm{s}}}\frac{-jtr^{-\alpha}}{1-jtr^{-\alpha}}r\,\mathrm{d}r\right)\right)\right]\mathrm{d}t$ $\displaystyle=$ $\displaystyle\frac{1}{2}+\frac{1}{\pi}\int_{0}^{\infty}\frac{1}{t}\mathsf{Im}\left[\left(\frac{1+jtu^{-\alpha}}{1+t^{2}u^{-2\alpha}}\right)e^{-jt\tau\sigma^{2}}\cdot\right.$ $\displaystyle\left.\exp\left(-2\pi\lambda\left(\int_{u}^{u+R_{\mathrm{s}}}\frac{p\,t^{2}r^{-2\alpha+1}}{1+t^{2}r^{-2\alpha}}\mathrm{d}r\,\int_{u+R_{\mathrm{s}}}^{\infty}\frac{t^{2}\tau^{2}r^{-2\alpha+1}}{1+t^{2}\tau^{2}r^{-2\alpha}}\mathrm{d}r\,+\,\int_{u}^{u+R_{\mathrm{s}}}\frac{(1-p)t^{2}\tau^{2}r^{-2\alpha+1}}{1+t^{2}\tau^{2}r^{-2\alpha}}\mathrm{d}r\right.\right.\right.$ $\displaystyle\left.\left.\left.+\,j\left[\int_{u+R_{\mathrm{s}}}^{\infty}\frac{t\tau r^{-\alpha+1}}{1+t^{2}\tau^{2}r^{-2\alpha}}\mathrm{d}r\,+\,\int_{u}^{u+R_{\mathrm{s}}}\frac{(1-p)\,t\tau r^{-\alpha+1}}{1+t^{2}\tau^{2}r^{-2\alpha}}\mathrm{d}r\,-\,\int_{u}^{u+R_{\mathrm{s}}}\frac{p\,tr^{-\alpha+1}}{1+t^{2}r^{-2\alpha}}\mathrm{d}r\right]\right)\right)\right]\mathrm{d}t$ (57) where the last step is obtained using multiplication of conjugate terms and rearranging into the real and imaginary parts. Now, if we define $\displaystyle M^{\prime}(t,u)=$ $\displaystyle 2{\alpha t^{-\alpha/2}}\left[\int_{u}^{u+R_{\mathrm{s}}}\frac{p\,t^{2}r^{-2\alpha+1}}{1+t^{2}r^{-2\alpha}}\mathrm{d}r\right.$ $\displaystyle\left.+\int_{u+R_{\mathrm{s}}}^{R}\frac{t^{2}\tau^{2}r^{-2\alpha+1}}{1+t^{2}\tau^{2}r^{-2\alpha}}\mathrm{d}r+\int_{u}^{u+R_{\mathrm{s}}}\frac{(1-p)t^{2}\tau^{2}r^{-2\alpha+1}}{1+t^{2}\tau^{2}r^{-2\alpha}}\mathrm{d}r\right],$ (58) $\displaystyle N^{\prime}(t,u)=$ $\displaystyle 2\alpha{t^{-\alpha/2}}\left[\int_{u+R_{\mathrm{s}}}^{R}\frac{t\tau r^{-\alpha+1}}{1+t^{2}\tau^{2}r^{-2\alpha}}\mathrm{d}r\right.$ $\displaystyle+\left.\int_{u}^{u+R_{\mathrm{s}}}\frac{(1-p)\,t\tau r^{-\alpha+1}}{1+t^{2}\tau^{2}r^{-2\alpha}}\mathrm{d}r-\int_{u}^{u+R_{\mathrm{s}}}\frac{p\,tr^{-\alpha+1}}{1+t^{2}r^{-2\alpha}}\mathrm{d}r\right],$ (59) (57) can be written as $\displaystyle\mathbb{P}\left(S>(I+\sigma^{2})\tau|\,X_{0}=u\right)$ $\displaystyle=\frac{1}{2}+\frac{1}{\pi}\int_{0}^{\infty}\frac{1}{t}\mathsf{Im}\left[\left(\frac{1+jtu^{-\alpha}}{1+t^{2}u^{-2\alpha}}\right)\right.$ $\displaystyle\times\left.\exp\left(-2\pi\lambda{\color[rgb]{0,0,0}t^{2/\alpha}}M^{\prime}(t,u)-j2\pi\lambda{\color[rgb]{0,0,0}t^{2/\alpha}}N^{\prime}(t,u)-jt\tau\sigma^{2}\right)\right]\mathrm{d}t.$ (60) Now, with some trivial manipulations and substituting (60) in (9), we get $\displaystyle\mathrm{p_{c}}(\tau,\lambda)=$ $\displaystyle\frac{1}{2}+\frac{1}{\pi}\int_{0}^{\infty}\int_{0}^{\infty}2\pi\lambda ue^{-\pi\lambda u^{2}}\cdot\frac{1}{t}\cdot\left[\frac{1}{1+t^{2}u^{-2\alpha}}\right]\cdot e^{-\frac{\pi}{\alpha}\lambda t^{2/\alpha}M^{\prime}(t,\,u)}$ $\displaystyle\times\left[tu^{-\alpha}\cos{\left(\frac{\pi}{\alpha}\lambda{t^{2/\alpha}}N^{\prime}(t,u)+\,t\tau\sigma^{2}\right)}-\sin{\left(\frac{\pi}{\alpha}\lambda{t^{2/\alpha}}N^{\prime}(t,u)+\,t\tau\sigma^{2}\right)}\right]\mathrm{d}t\mathrm{d}u.$ (61) Further, the forms of $M$ and $N^{\prime\prime}$ can be simplified by trivial manipulations and definition of incomplete Beta function in (58) and (59) to get $\displaystyle M^{\prime}(t,u)=pQ\left(\frac{1}{\alpha},t^{2}(u+R_{\mathrm{s}})^{-2\alpha},t^{2}u^{-2\alpha}\right)$ $\displaystyle+\tau^{2/\alpha}(1-p)Q\left(\frac{1}{\alpha},(t\tau)^{2}(u+R_{\mathrm{s}})^{-2\alpha},(t\tau)^{2}u^{-2\alpha}\right)+\tau^{2/\alpha}Q\left(\frac{1}{\alpha},0,t^{2}\tau^{2}(u+R_{\mathrm{s}})^{-2\alpha}\right).$ (62) and, $\displaystyle N^{\prime}(t,u)=-pQ\left(\frac{1}{\alpha}+\frac{1}{2},t^{2}(u+R_{\mathrm{s}})^{-2\alpha},t^{2}u^{-2\alpha}\right)$ $\displaystyle+\tau^{2/\alpha}(1-p)Q\left(\frac{1}{\alpha}+\frac{1}{2},(t\tau)^{2}(u+R_{\mathrm{s}})^{-2\alpha},(t\tau)^{2}u^{-2\alpha}\right)+\tau^{2/\alpha}Q\left(\frac{1}{\alpha}+\frac{1}{2},0,t^{2}\tau^{2}(u+R_{\mathrm{s}})^{-2\alpha}\right).$ (63) Now, we can substitute $\displaystyle t$ $\displaystyle\rightarrow s/{(\lambda\pi)}^{\alpha/2}$ $\displaystyle u$ $\displaystyle\rightarrow v/\sqrt{\lambda\pi}\ $ (64) in (61), (62) and (63) to get the desired result. ## References * [1] M. El-Hajjar and L. Hanzo., “A survey of digital television broadcast transmission techniques,” _IEEE Communication Surveys and Tutorials_ , vol. 15, no. 4, p. 1924–1949., 2013. * [2] L. Fay and L. Michael and D. Gmez-Barquero and N. Ammar and M.W. Caldwell, “An overview of the ATSC 3.0 physical layer specification,” _IEEE Transactions on Broadcasting_ , vol. 62, pp. 159–171, 2016. * [3] D. Gómez-Barquero, C. Douillard, P. Moss, and V. Mignone, “DVB-NGH: The next generation of digital broadcast services to handheld devices,” _IEEE Transactions on Broadcasting_ , vol. 60, no. 2, pp. 246–257, 2014. * [4] D. Catrein, J. Huschke, U. Horn, “Analytic evaluation of a hybrid broadcast-unicast TV offering,” in _Proc. IEEE Vehicular Technology Conference-Spring (VTC-spring)_ , 2008, pp. 2864–2868. * [5] L. Rong, S. E. Elayoubi, and O. B. Haddada, “Performance evaluation of cellular networks offering TV services,” _IEEE Transactions on Vehicular Technology_ , vol. 60, no. 2, 2011. * [6] J. G. Andrews, F. Baccelli, and R. K. Ganti, “A tractable approach to coverage and rate in cellular networks,” _IEEE Transactions on communications_ , vol. 59, no. 11, pp. 3122–3134, 2011. * [7] J. G. Andrews, A. K. Gupta, and H. S. Dhillon, “A primer on cellular network analysis using stochastic geometry,” _arXiv preprint arXiv:1604.03183_ , 2016. * [8] M. Haenggi, _Stochastic Geometry for Wireless Networks_. Cambridge: Cambridge University Press, 2012. * [9] A. Guo and M. Haenggi, “Asymptotic deployment gain: A simple approach to characterize the SINR distribution in general cellular networks,” _IEEE Transactions on Communications_ , vol. 63, pp. 962–976, Mar. 2015. * [10] J. G. Andrews, T. Bai, M. N. Kulkarni, A. Alkhateeb, A. K. Gupta, and R. W. Heath, “Modeling and analyzing millimeter wave cellular systems,” _IEEE Trans. Commun._ , vol. 65, no. 1, pp. 403–430, Jan 2017. * [11] W. Lu and M. Di Renzo, “Stochastic geometry modeling of cellular networks: Analysis, simulation and experimental validation,” in _Proc. ACM International Conference on Modeling, Analysis and Simulation of Wireless and Mobile Systems (MSWiM)_ , 2015, p. 179–188. * [12] A. Shokair and Y. Nasser and M. Crussière and J. Hélard and O. Bazzi, “Analytical study of the probability of coverage in hybrid broadcast-unicast networks,” in _Proc. Annual International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC)_ , 2018, pp. 1–6. * [13] R. Sahu, K. K. Chaurasia, and A. K. Gupta, “SINR and rate coverage of broadcast networks using stochastic geometry,” in _Proc. International Conference on Signal Processing and Communications (SPCOM)_ , July 2020, pp. 1–5. * [14] L. Peterson and R. Groot, “Location-based advertising: The key to unlocking the most value in the mobile advertising and location-based services markets,” _Peterson Mobility Solutions_ , 2009. * [15] L. Shi, E. Obregon, K. W. Sung, J. Zander, and J. Bostrom, “CellTV—on the benefit of TV distribution over cellular networks: A case study,” _IEEE Transactions on Broadcasting_ , vol. 60, no. 1, pp. 73–84, 2014. * [16] J. Gil-Pelaez, “Note on the inversion theorem,” _Biometrika_ , vol. 38,no. 3-4, p. 481–482, 1951. * [17] International Telecommunication Union, “Recommendation ITU-R P.1546-6, method for point-to-area predictions for terrestrial services in the frequency range 30 mhz to 4 000 MHz,” _P Series Radiowave Propagation_ , August 2019. * [18] ——, _Handbook on Digital Terrestrial Television Broadcasting Networks and Systems Implementation_ , 2016. * [19] A. K. Gupta, X. Zhang, and J. G. Andrews, “SINR and throughput scaling in ultradense urban cellular networks,” _IEEE Wireless Communications Letters_ , vol. 4, no. 6, pp. 605–608, 2015.
# Fully developed anelastic convection with no-slip boundaries Chris A. Jones1<EMAIL_ADDRESS>Krzysztof A. Mizerski2 Mouloud Kessar3 1Department of Applied Mathematics, University of Leeds, Leeds, LS2 9JT, UK 2 Department of Magnetism, Institute of Geophysics, Polish Academy of Sciences, ul. Ksiecia Janusza 64, 01-452 Warsaw, Poland 3Université de Paris, Institut de physique du globe de Paris, CNRS, IGN, F-75005 Paris, France ###### Abstract Anelastic convection at high Rayleigh number in a plane parallel layer with no slip boundaries is considered. Energy and entropy balance equations are derived, and they are used to develop scaling laws for the heat transport and the Reynolds number. The appearance of an entropy structure consisting of a well-mixed uniform interior, bounded by thin layers with entropy jumps across them, makes it possible to derive explicit forms for these scaling laws. These are given in terms of the Rayleigh number, the Prandtl number, and the bottom to top temperature ratio, which measures how stratified the layer is. The top and bottom boundary layers are examined and they are found to be very different, unlike in the Boussinesq case. Elucidating the structure of these boundary layers plays a crucial part in determining the scaling laws. Physical arguments governing these boundary layers are presented, concentrating on the case in which the boundary layers are thin even when the stratification is large, the incompressible boundary layer case. Different scaling laws are found, depending on whether the viscous dissipation is primarily in the boundary layers or in the bulk. The cases of both high and low Prandtl number are considered. Numerical simulations of no-slip anelastic convection up to a Rayleigh number of $10^{7}$ have been performed and our theoretical predictions are compared with the numerical results. Manuscript at ## 1 Introduction The problem of the influence of density stratification on developed convection is of great importance from the astrophysical point of view. Giant planets and stars are typically strongly stratified and the anelastic approximation, see e.g. Ogura & Phillips (1962), Gough (1969) and Lantz & Fan (1999), is commonly used to describe convection in their interiors, e.g. Toomre et al. (1976), Glatzmaier & Roberts (1995), Brun & Toomre (2002), Browning et al. (2006), Miesch et al. (2008), Jones & Kuzanyan (2009), Jones et al. (2011), Verhoeven et al. (2015), Kessar et al. (2019) and many others. The anelastic approximation is based on the convective system being a small departure from the adiabatic state, which is appropriate for large scale systems with developed, turbulent and thus strongly-mixing convective regions. The small departure from adiabaticity induces convective velocities much smaller than the speed of sound, so sound waves are neglected in the dynamics. We consider a plane layer of fluid between two parallel plates distance $d$ apart, and we assume the turbulent flow is spatially homogeneous in the horizontal direction. We also assume that the convection is in a statistically steady state, so that the time-averages of time-derivative terms can be neglected. Most astrophysical applications are in spherical geometry, but the simpler plane layer problem is a natural place to start our investigation of high Rayleigh number anelastic convection. In convection theory, we try to determine the dependencies of the superadiabatic heat flux and the convective velocities (measured by the Nusselt, $Nu$, and Reynolds $Re$ numbers) on the driving force measured by the imposed temperature difference between the top and bottom plates, i.e. on the Rayleigh number, $Ra$, and on the Prandtl number, $Pr$ (the ratio of the fluid kinematic viscosity to its thermal diffusivity). Here we aim to discover how these dependencies are affected by the stratification. We rely strongly on the theory of Grossmann & Lohse (2000) (further developed later and updated in Stevens et al. 2013) developed for Boussinesq, i.e. non-stratified, convection. However, compressible convection differs strongly from the Boussinesq case, with the latter mostly corresponding to experimental situations. There are two crucial differences, which have very important consequences for the dynamics of convection. Firstly, in the compressible case the viscous heating and the work of the buoyancy force are no longer negligible compared to the heat transport. Secondly, in stratified convection the boundary layers and the flow velocities are different at the top of the layer and the bottom of the layer, unlike the Boussinesq case where the top and bottom boundary layers have the same structure and the temperature of the well-mixed interior is exactly half way between the temperature of the top and bottom plates. So although our approach is based on that of Grossmann & Lohse (2000), there are additional novel features required in the compressible convection case. We develop the theory of fully developed convection in stratified systems and study the dependence of the total superadiabatic heat flux and the amplitude of convective flow on the number of density scale heights in the layer. The scaling laws, i.e. the dependencies of the Nusselt and Reynolds numbers on the Rayleigh and Prandtl numbers are the same as in the Boussinesq convection. In this paper, we concentrate on the convective regimes which seem to be most relevant to current numerical capabilities, i.e. regimes most easily achieved by numerical experiments. These are the regimes where the thermal dissipation is dominated by the thermal boundary layer contribution. It is those regimes, denoted by $I_{u}$, $I_{l}$, $II_{u}$ and $II_{l}$ on the phase diagram, figure 2 of Grossmann & Lohse (2000), that in the Boussinesq case correspond to Rayleigh numbers less than about 1012. It must be noted, however, that contrary to the Boussinesq case, which is well established by numerous experimental and numerical investigations, there are to date no experiments on fully turbulent stratified convection, due to the difficulties of achieving significant density stratification in laboratory settings. Some experiments are being developed using the centrifugal acceleration in rapid rotating systems to enhance the effective gravity (Menaut et al. 2019). There have also been some numerical investigations of anelastic convection in a plane layer, mostly focussed on elucidating how well the anelastic approximation performs compared to fully compressible convection, Verhoeven et al. (2015) and Curbelo et al. (2019). This latter paper notes that the top and bottom boundary layer structures that occur in the case of high Prandtl number are different. In addition to the dependence on Rayleigh and Prandtl number, our problem depends on how stratified the layer is, which can be estimated by the ratio $\Gamma$ of the temperature at the bottom of the the layer $T_{B}$ to the temperature at the top $T_{T}$. When $\Gamma$ is close to unity the layer is nearly Boussinesq, but when $\Gamma$ is large there are many temperature and density scale heights within the layer. We aim to derive the functional form of $Nu(\Gamma,\,Ra,\,Pr)$ and $Re(\Gamma,\,Ra,\,Pr)$, but we cannot derive reliable numerical values for the prefactors in anelastic convection. Since experiments are not available, this will require high resolution high Ra simulations, which are being developed, but are not yet in a state to determine the prefactors accurately. In § 2 we derive the anelastic equations and the reference states we use, and outline the structure of high Rayleigh number anelastic convection. Further details of the form of the anrelastic temperature perturbation are given in appendix A. In § 3 we derive the energy and entropy production integrals, which are the fundamental building blocks for developing convective scaling laws. In sections § 4 and § 5 we derive the physical arguments used to get the key properties of the top and bottom boundary layers. In § 6 we derive the scaling laws for the case where the viscous dissipation is primarily in the boundary layers. The case where the dissipation is mainly in the bulk is dealt with in appendix B. § 7 gives the results of our numerical simulations, comparing them with our theoretical results. Our conclusions are in § 8. ## 2 Fully developed compressible convection under the anelastic approximation Consider a layer of compressible perfect gas between two parallel plates, of thickness $d$, periodic in horizontal directions, the evolution of which is described by the set of the Navier-Stokes, mass conservation and energy equations under the anelastic approximation, $\frac{\partial\mathbf{u}}{\partial t}+\left(\mathbf{u}\cdot\nabla\right)\mathbf{u}=-\nabla\left(\frac{p}{\bar{\varrho}}\right)+\frac{g}{c_{p}}s\hat{\mathbf{e}}_{z}+\frac{\mu}{\bar{\rho}}\left[\nabla^{2}\mathbf{u}+\frac{1}{3}\nabla\left(\nabla\cdot\mathbf{u}\right)\right],$ (1) $\nabla\cdot\left(\bar{\varrho}\mathbf{u}\right)=0,$ (2) $\bar{\rho}\bar{T}\left[\frac{\partial s}{\partial t}+\mathbf{u}\cdot\nabla s\right]=k\nabla^{2}T+\mu\left[q+\partial_{j}u_{i}\partial_{i}u_{j}-\left(\nabla\cdot\mathbf{u}\right)^{2}\right],$ (3) $\frac{p}{\bar{p}}=\frac{\rho}{\bar{\rho}}+\frac{T}{\bar{T}},\qquad s=c_{v}\frac{p}{\bar{p}}-c_{p}\frac{\rho}{\bar{\rho}},\quad\gamma=\frac{c_{p}}{c_{v}},\quad c_{p}-c_{v}=R,$ $None$ where $q=(\partial_{j}u_{i})(\partial_{j}u_{i})+\frac{1}{3}\left(\nabla\cdot\mathbf{u}\right)^{2},$ (5) $\mathbf{u}$ being the fluid velocity, $p$ the pressure, $\rho$ the density, $T$ the temperature and $s$ the entropy. Barred variables are adiabatic reference state variables, unbarred variables denote the perturbation from the reference state due to convection. The dynamic viscosity $\mu=\bar{\rho}\nu$, the thermal conductivity $k$, gravity $g$ and the specific heats at constant pressure, $c_{p}$, and constant volume, $c_{v}$, are all assumed constant. The bounding plates are no-slip and impenetrable, so $\bf u=0$ there. We consider the constant entropy boundary conditions $s=\Delta S\quad\textrm{at}\quad z=0,\qquad s=0\quad\textrm{at}\quad z=d.$ (6) Note that we do not replace the thermal diffusion term in (3) by an entropy diffusion term, as is often done in anelastic approaches. With our no-slip boundaries, there will be boundary layers which may be laminar even at very high Rayleigh number, and entropy diffusion is not appropriate if laminar boundary layers are present. We discuss the additional issues raised by adopting constant temperature boundary conditions rather than constant entropy conditions in Appendix A. In the anelastic approximation, the full variables, $\hat{p}$, $\hat{\rho}$ and $\hat{T}$, are expanded in terms of the small parameter $\epsilon$, which is defined precisely in equation (10) below, so $\hat{p}=\bar{p}+\epsilon p,\quad\hat{\rho}=\bar{\rho}+\epsilon\rho,\quad\hat{T}=\bar{T}+\epsilon T,\quad u\sim(\epsilon gd)^{1/2},\quad t\sim\left(\frac{\epsilon g}{d}\right)^{-1/2},\quad\hat{s}=\bar{s}+\epsilon s.$ (7) where $\bar{p}$, $\bar{\rho}$ and $\bar{T}$ comprise the adiabatic reference state. Here $\bar{s}$ is simply a constant, and since $s=c_{v}\ln{\hat{p}}/{\hat{\rho}}^{\gamma}$ \+ const., and $\bar{p}/{\bar{\rho}}^{\gamma}$ is constant, we obtain ($None$b) by choosing the constant appropriately. ### 2.1 The adiabatic reference state The reference state is the adiabatic static equilibrium governed by $\mathrm{d}\bar{p}/\mathrm{d}z=-g\bar{\rho}$, $\bar{p}=R\bar{\rho}\bar{T}$ and $\bar{p}=K{\bar{\rho}}^{\gamma}$, where $R$ is the gas constant, $z=0$ is the bottom of the layer and $z=d$ the top. It follows that $\bar{T}=T_{B}\left(1-\theta\frac{z}{d}\right),\quad\bar{\rho}=\rho_{B}\left(1-\theta\frac{z}{d}\right)^{m},\quad\bar{p}=\frac{gd\rho_{B}}{\theta\left(m+1\right)}\left(1-\theta\frac{z}{d}\right)^{m+1},$ $None$ $\frac{gd}{c_{p}}=\Delta{\bar{T}}=T_{B}-T_{T}>0,\quad\theta=\frac{\Delta{\bar{T}}}{T_{B}},\quad m=\frac{1}{\gamma-1},\quad\Gamma=\frac{T_{B}}{T_{T}}=\frac{1}{1-\theta},$ $None$ which defines $\theta$, and the polytropic index $m$. We use subscripts $T$ and $B$ to denote conditions at the top and bottom boundary respectively. The temperature ratio $\Gamma>1$, is a convenient measure of the compressibility. $\Gamma\to 1$ is the Boussinesq limit, and highly compressible layers have $\Gamma$ large. Note that $\Gamma^{m}=\rho_{B}/\rho_{T}$ is the ratio of the highest to lowest density in the layer. The density ratio can be very large in astrophysical applications, the density of the bottom of the solar convection zone being $\sim 10^{6}$ times the density at the top. Sometimes the number of density scale heights, $N_{\rho}$, (the scale height being defined at the top of the layer) that fit into the layer is used to measure compressibility, $N_{\rho}=m(\Gamma-1)$. ### 2.2 The conduction state The adiabatic reference state satisfies $\nabla^{2}{\bar{T}}=0$, but since it is isentropic is does not satisfy the entropy boundary conditions. The anelastic conduction state is also a polytrope, but with a slightly more negative temperature gradient, so ${\tilde{T}}_{B}=T_{B},\ {\tilde{T}}_{T}<T_{T}$. The conduction state is $\tilde{T}=T_{B}\left(1-{\tilde{\theta}}\frac{z}{d}\right),\quad\tilde{\rho}=\rho_{B}\left(1-{\tilde{\theta}}\frac{z}{d}\right)^{\tilde{m}},\quad\tilde{p}=\frac{gd\rho_{B}}{{\tilde{\theta}}\left({\tilde{m}}+1\right)}\left(1-{\tilde{\theta}}\frac{z}{d}\right)^{{\tilde{m}}+1},$ $None$ ${\widetilde{\Delta T}}=T_{B}-{\tilde{T}}_{T}>0,\quad{\tilde{\theta}}=\frac{{\widetilde{\Delta T}}}{T_{B}},\quad{\tilde{m}}=\frac{gd}{R{\widetilde{\Delta T}}}-1.$ $None$ The small anelastic parameter $\epsilon$ is now defined as $\epsilon={\tilde{\theta}}\frac{{\tilde{m}}+1-\gamma{\tilde{m}}}{\gamma}=-\frac{d}{T_{B}}\left[\frac{\mathrm{d}\tilde{T}}{\mathrm{d}z}+\frac{g}{c_{p}}\right]\ll 1,$ (10) and the entropy in the conduction state is ${\tilde{s}}=c_{v}\ln\frac{\tilde{p}}{\tilde{\rho}^{\gamma}}+\mathrm{const}=\frac{\epsilon c_{p}}{\theta}\ln\left(1-\theta\frac{z}{d}\right)+\mathrm{const},\ \textrm{so}\ s=\frac{c_{p}}{\theta}\ln\left(1-\theta\frac{z}{d}\right)+\mathrm{const}$ $None$ which is the scaled entropy, see (7), correct to $O(\epsilon)$ since $\tilde{\theta}$ and $\theta$ differ by only $O(\epsilon)$. Since the boundaries have fixed entropy, the entropy at the boundaries in the conduction state defines the entropy drop across the layer for all Rayleigh numbers, so $\Delta S=\frac{c_{p}}{\theta}\ln\Gamma=\frac{c_{p}\Gamma\ln\Gamma}{\Gamma-1},$ (12) relating to the entropy boundary conditions (6). Note that as our entropy variable $s$ is scaled by $\epsilon$, the entropy drop is $O(\epsilon)$. Some anelastic papers take the conduction state as the reference state, and some take the adiabatic state as the reference state. Taking the conduction state as the reference state is appropriate when convection near critical Rayleigh number is considered, but at the large Rayleigh numbers considered here, the conduction state is less relevant. Although the conduction state tends to the adiabatic reference state as $\epsilon\to 0$, the thermodynamic variables are not the same in the two systems: $T=0$ with respect to the adiabatic state corresponds to $T=T_{B}z/d$ if the conduction state is chosen as the reference state. In equation (1) we have made use of ($None$b), ($None$) and (10) to write (Braginsky & Roberts, 1995; Lantz & Fan, 1999) $\displaystyle-\frac{\nabla p}{\bar{\rho}}-\frac{\rho}{\bar{\rho}}g\hat{\mathbf{e}}_{z}=-\nabla\left(\frac{p}{\bar{\rho}}\right)+\frac{g}{c_{p}}s\hat{\mathbf{e}}_{z}+O(\epsilon).$ (13) ### 2.3 The Nusselt and Rayleigh numbers in anelastic convection Next we consider the superadiabatic heat flux. The horizontal average at level $z$ is denoted by $\langle\ \rangle_{h}$. At the boundaries, all the heat is carried by conduction, and if the total temperature $\hat{T}=\bar{T}+\epsilon T$, then the total heat flux at the boundaries is $-kd{\left\langle\hat{T}\right\rangle_{h}}/dz=-kd\bar{T}/dz-k\epsilon d\left\langle T\right\rangle_{h}/dz$, but the superadiabatic part is obtained by subtracting off the heat flux carried along the adiabat, so we let $\epsilon F^{super}=-\epsilon k\frac{d\left\langle T\right\rangle_{h}}{dz}\Big{|}_{z=0},\quad{\rm so}\quad F^{super}=-k\frac{d\left\langle T\right\rangle_{h}}{dz}\Big{|}_{z=0}.$ (14) The Nusselt number in anelastic convection is defined as the ratio of the superadiabatic heat flux divided by the heat conducted down the conduction state superadiabatic gradient. Note that the flux conducted down the adiabatic gradient is ignored in this definition, so $Nu=\frac{F^{super}d}{kT_{B}},$ (15) so $Nu$ is close to unity near onset, and is large in fully developed convection. For fixed entropy boundary conditions, the Rayleigh number is defined as $Ra=\frac{g\Delta Sd^{3}\rho_{B}^{2}}{\mu k}\approx\frac{c_{p}\Delta S\Delta{\bar{T}}d^{2}\rho_{B}^{2}}{\mu k}.$ (16) The anelastic approximation is asymptotically valid in the limit $\epsilon\to 0$. Note that small superadiabatic temperature gradient does not imply small Rayleigh number $Ra$, since the diffusion coefficients can be small, in fact to get $Ra\sim O(1)$ in the limit $\epsilon\to 0$ we must have $\frac{k}{\rho_{B}c_{p}}\sim\left(gd^{3}\epsilon\right)^{1/2},\quad\frac{\mu}{\rho_{B}}\sim\left(gd^{3}\epsilon\right)^{1/2}.$ (17) allowing large but finite $Ra$ even when the superadiabatic gradient is small. To derive the anelastic equations (1) to (5), we insert (7) into the full compressible equations and divide the momentum equation by $\epsilon$, the mass conservation equation by $\epsilon^{1/2}$ and the energy equation by $\epsilon^{3/2}$. Having taken this limit, the parameter $\epsilon$ no longer appears in our analysis. However, if anelastic work is compared to fully compressible situations, then a finite value of $\epsilon$ must be chosen, and the anelastic results are only approximate, though there is a growing body of evidence that the anelastic approximation does capture the main features of subsonic compressible convection. ### 2.4 High Rayleigh number convection We have the following physical picture in mind. In strongly turbulent convection we expect the entropy $s$ to be well-mixed away from boundary layers near $z=0,d$. We denote the global spatial average over the convecting layer by $||\ \ ||$ and the horizontal average at level $z$ by $\langle\ \rangle_{h}$. The total entropy drop is the conduction state value $\Delta S=c_{p}\ln\Gamma/\theta.$ Since the entropy is constant in the bulk interior, we define the entropy drops $\Delta S_{B}$ and $\Delta S_{T}$ across the bottom and top boundary layers respectively. These will not be equal, with $\Delta S_{T}$ normally considerably larger than $\Delta S_{B}$. We must however have $\Delta S_{B}+\Delta S_{T}=\Delta S.$ (18) We consider only the case where both the top and bottom boundary layers are laminar. At extremely high $Ra$ these layers may become turbulent as can happen in the Boussinesq case. The laminar boundary layer case is simpler, and gives predictions which can be broadly compared with numerical simulations, though it is difficult for numerical simulations to get into the fully developed large Rayleigh and Nusselt number regime we are aiming at here. A schematic picture of the horizontally-averaged entropy profile expected in highly supercritical anelastic convection is shown in figure 1(a). Since the heat flux through the boundary layers is determined by thermal diffusion rather than entropy diffusion, we need to express the temperature jumps across the thermal boundary layers in terms of the entropy jumps. From ($None$) we obtain $\frac{\left(\Delta\rho\right)_{i}}{\rho_{i}}\approx\frac{1}{\gamma-1}\left[\frac{\left(\Delta T\right)_{i}}{T_{i}}-\gamma\frac{\left(\Delta s\right)_{i}}{c_{p}}\right],\qquad\frac{\left(\Delta p\right)_{i}}{p_{i}}\approx\frac{\gamma}{\gamma-1}\left[\frac{\left(\Delta T\right)_{i}}{T_{i}}-\frac{\left(\Delta s\right)_{i}}{c_{p}}\right],$ $None$ where the $\Delta$ quantities refer to the jump in the horizontally averaged value across the boundary layer and the subscript $i$ stands either for $B$ or $T$. We also define the thermal and viscous boundary layer thicknesses, $\delta_{i}^{th}$ and $\delta_{i}^{\nu}$ with $i=B,T$, which we use to obtain scaling estimates. Numerical simulations indicate that the horizontal velocity $U_{H}=\left(\left\langle u_{x}^{2}\right\rangle_{h}+\left\langle u_{y}^{2}\right\rangle_{h}\right)^{1/2}$ (20) has local maxima close to both boundaries (see e.g. figure 3(b) below), so these maxima are a convenient way to define the velocity jumps across the viscous boundary layers, $\Delta U_{i}$, so $\Delta U_{B}=U_{H}(z=z_{max,B}),\quad\Delta U_{T}=U_{H}(z=z_{max,T}),$ (21) where $z=z_{max,B}$, $z=z_{max,T}$ are the locations of the local maxima. The thermal boundary layer thickness for the entropy, $\delta_{i}^{th}$, and the viscous boundary layer thickness, $\delta_{i}^{\nu}$, can be defined as $\delta_{i}^{th}=\left\\{-\frac{1}{\Delta S_{i}}\frac{d\left\langle S\right\rangle_{h}}{dz}{\Big{|}}_{z=z_{i}}\right\\}^{-1},\quad\delta_{i}^{\nu}=\left\\{\pm\frac{1}{\Delta U_{i}}\frac{dU_{H}}{dz}{\Big{|}}_{z=z_{i}}\right\\}^{-1},\quad z_{i}=z_{B},\,z_{T}$ (22) (a) (b) z z $\langle s\rangle_{h}$ $\langle T\rangle_{h}$ Figure 1: (a) A schematic picture of the entropy profile in developed convection. (b) a schematic picture of the anelastic temperature perturbation in developed convection. In the boudary layers, the dominant balance in the $z$-component of the Navier-Stokes equation occurs between the pressure gradient and the buoyancy force. Mass conservation in the boundary layers means $u_{z,i}\sim O(\delta_{i}^{\nu})$ so the vertical component of inertia is small. The boundary layers are therefore approximately hydrostatic, $\left(\Delta p\right)_{i}\approx\frac{g}{c_{p}}\rho_{i}\Delta s_{i}\delta_{i}^{th}.$ (23) Inserting (23) into ($None$b) leads to $\frac{\left(\Delta T\right)_{i}}{T_{i}}\approx\frac{\left(\Delta s\right)_{i}}{c_{p}}\left(1+\theta\frac{\delta_{i}^{th}}{d}\frac{T_{B}}{T_{i}}\right).$ (24) Typically the term $(\theta\delta_{i}^{th}T_{B})/(T_{i}d)$ resulting from the pressure jump across the boundary layers is expected to be small because the boundary layer is thin. However, in simulations where the Rayleigh number is bounded above by numerical constraints, the top boundary layer may not be as thin as desired for accurate asymptotics to apply, and the factor $T_{B}/T_{T}$ can be large in layers containing many scale heights. We refer to the case where the pressure term is not negligible as the compressible boundary layer case. However, in this work we assume the boundary layers are incompressible, which is valid provided $T_{B}/T_{T}$ remains finite as the Rayleigh number increases and the boundary layers become very thin. Then the pressure fluctuation term is constant in both boundary layers, so that in the boundary layers $\frac{\left(\Delta T\right)_{i}}{T_{i}}\approx\frac{\left(\Delta s\right)_{i}}{c_{p}},\ \ \textrm{and}\ \ \frac{\left(\nabla T\right)_{i}}{T_{i}}\approx\frac{\left(\nabla s\right)_{i}}{c_{p}}$ (25) and defining the temperature boundary layer thicknesses similarly to (22), $\delta_{i}^{T}=\left\\{-\frac{1}{\Delta T_{i}}\frac{d\left\langle T\right\rangle_{h}}{dz}{\Big{|}}_{z=z_{i}}\right\\}^{-1},$ (26) the temperature boundary layer thicknesses are the same as the entropy boundary layer thicknesses. Note that in the compressible boundary layer case the entropy and temperature boundary layer thicknesses will be different. For incompressible boundary layers and high Rayleigh number, the Nusselt number can be written in terms of the boundary layer thicknesses, using (15), (14), (26) and (25), $Nu=\frac{d}{\delta^{th}_{T}}\frac{\Delta S_{T}}{\Gamma c_{p}}=\frac{d}{\delta^{th}_{B}}\frac{\Delta S_{B}}{c_{p}}.$ (27) In figure 1(b) we sketch the horizontally-averaged anelastic temperature perturbation $\langle T\rangle_{h}$. This is sometimes referred to as the superadiabatic temperature (e.g. Verhoeven et al. 2015). Note that with our constant entropy boundary conditions, $\langle T\rangle_{h}$ is not zero at the boundaries. We show in appendix A that the offsets, $\langle T\rangle_{hB}$ at $z=0$ and $\langle T\rangle_{hT}$ at $z=d$, are both positive and we show also that in the bulk, turbulent pressure effects make the gradient of $T$ positive as shown in figure 1(b). This means that the total horizontally averaged temperature gradient in the presence of convection is less negative than the adiabatic reference state, so that on horizontal average the layer is subadiabatically stratified (e.g. Korre et al. 2017). Of course, in some parts of the layer the local temperature gradient must be superadiabatic to drive the convection, but other parts are subadiabatic so that the horizontal average can be subadiabatic. To obtain the anelastic temperature fluctuation as sketched in figure 1(b), we need to make use of equations ($None$a) and ($None$b), so we need to make a specific choice for entropy at the boundaries. Here we have chosen to take the entropy at the top boundary as zero, so the entropy at the bottom boundary is $s=\Delta S$. A different choice of entropy constant adds an easily found function of $z$ to $T$, $\rho$ and $p$ but this does not affect the velocity field obtained. One further point is that if (1) is horizontally averaged, the horizontal average satisfies a first order differential equation in $z$ (see appendix A for details), so a boundary condition on $\langle p\rangle_{h}$ is required. Here we choose the natural condition that the anelastic density perturbation vanishes when integrated over the layer, that is $||\ \rho\ ||=0\quad\Rightarrow\quad\langle p\rangle_{h,T}=\langle p\rangle_{h,B}.$ (28) This means that the total mass in the layer is the same as in the adiabatic reference state. As we see in appendix A, this means the horizontally averaged anelastic pressure perturbations at the top and bottom of the layer must be equal. ## 3 Energy and entropy production integrals Understanding the energy transfer and entropy production in convective flow is the key to understanding the physics of compressible convection. Therefore we derive now a few exact relations which will allow us to study some general aspects of the dynamics of developed compressible convection. We assume any initial transients in the convection have been eliminated, and we are in a statistically steady state. Formally, this means we consider time-averaged quantities throughout the paper. ### 3.1 Energy balance By multiplying the Navier-Stokes equation (1) by $\bar{\rho}{u}$ and averaging over the entire volume (recalling that horizontal averages of $x$ and $y$ derivatives are zero) we obtain the following relation $\frac{g}{c_{p}}||\bar{\rho}u_{z}s||=\mu||q||,$ (29) stating that the total work per unit volume of the buoyancy force is equal to the total viscous dissipation in the fluid per unit volume. In deriving (29) use has been made of the no-slip boundary conditions and equation (2). We derive the superadiabatic heat flux in the system at every $z$ by averaging over a horizontal plane and integrating the heat equation (3) from $0$ to $z$, $\displaystyle F^{super}$ $\displaystyle=$ $\displaystyle-k\frac{d\left\langle T\right\rangle_{h}}{dz}\Big{|}_{z=0}=\left\langle\bar{\rho}\bar{T}u_{z}s\right\rangle_{h}-k\frac{d\left\langle T\right\rangle_{h}}{dz}$ (30) $\displaystyle+$ $\displaystyle\frac{g}{c_{p}}\int_{0}^{z}\left\langle\bar{\rho}u_{z}s\right\rangle_{h}\mathrm{d}z-\mu\int_{0}^{z}\left\langle q\right\rangle_{h}\mathrm{d}z-2\mu\left[\left\langle u_{z}\frac{du_{z}}{\mathrm{d}z}\right\rangle_{h}-\frac{m\Delta{\bar{T}}}{\bar{T}d}\left\langle u_{z}^{2}\right\rangle_{h}\right].$ In deriving this expression, we have made use of ($None$a,d) and $(\partial_{j}u_{i})(\partial_{i}u_{j})-(\nabla\cdot{\bf u})^{2}=\partial_{j}\partial_{i}(u_{i}u_{j})-2\partial_{j}(u_{j}\nabla\cdot{\bf u})=\partial_{j}\partial_{i}(u_{i}u_{j})+2\partial_{j}(u_{j}u_{z}\frac{m\Delta{\bar{T}}}{\bar{T}})$ (31) since the continuity equation gives $\nabla\cdot{\bf u}=\partial_{i}u_{i}=\frac{u_{z}m\Delta{\bar{T}}}{{\bar{T}}d},$ (32) and we recall that $x$ or $y$ derivatives vanish on horizontal averaging. As $z\to d$ all terms with a factor $u_{z}$ tend to zero, so on using (29) we obtain the overall energy balance equation, $F^{super}=-k\frac{d\left\langle T\right\rangle_{h}}{dz}\Big{|}_{z=0}=-k\frac{d\left\langle T\right\rangle_{h}}{dz}\Big{|}_{z=d},$ (33) thus in a stationary state the heat flux entering the fluid volume at $z=0$ must be equal to the outgoing heat flux $z=d$. ### 3.2 Entropy balance This energy balance equation alone is not very helpful for evaluating the Nusselt number. We need the entropy balance equation, obtained by dividing the energy equation (3) by $\bar{T}$, horizontally averaging, and integrating from $0$ to $z$, $\displaystyle\left\langle\bar{\rho}u_{z}s\right\rangle_{h}$ $\displaystyle=$ $\displaystyle-\frac{k}{T_{B}}\frac{\mathrm{d}}{\mathrm{d}z}\left\langle T\right\rangle_{h}\Big{|}_{z=0}+\frac{k}{\bar{T}}\frac{\mathrm{d}}{\mathrm{d}z}\left\langle T\right\rangle_{h}\Big{|}_{z}-\int_{0}^{z}\frac{k\Delta{\bar{T}}}{{\bar{T}}^{2}d}\frac{\mathrm{d}}{\mathrm{d}z}\left\langle T\right\rangle_{h}\,dz+\int_{0}^{z}\frac{\mu}{\bar{T}}\left\langle q\right\rangle_{h}\,dz$ (34) $\displaystyle+$ $\displaystyle\int_{0}^{z}\frac{\mu}{\bar{T}}\left\langle\partial_{j}(\partial_{i}(u_{i}u_{j}))-2\partial_{j}(u_{j}(\partial_{i}u_{i}))\right\rangle_{h}\,dz$ where use has been made of equation (31). The overall entropy balance equation comes from taking the limit $z\to d$ of (34), noting $u_{z}\to 0$ in this limit, $\displaystyle\frac{k}{T_{B}}\frac{\mathrm{d}}{\mathrm{d}z}\left\langle T\right\rangle_{h}\Big{|}_{z=0}-\frac{k}{T_{T}}\frac{\mathrm{d}}{\mathrm{d}z}\left\langle T\right\rangle_{h}\Big{|}_{z=d}$ $\displaystyle=$ $\displaystyle-\int_{0}^{d}\frac{k\Delta{\bar{T}}}{{\bar{T}}^{2}d}\frac{\mathrm{d}}{\mathrm{d}z}\left\langle T\right\rangle_{h}\,dz+\int_{0}^{d}\frac{\mu}{\bar{T}}\left\langle q\right\rangle_{h}\,dz$ Equations (34, LABEL:eq:3_7) look complicated, but they simplify considerably when the top and bottom boundary layers are thin. We start with (34), which we write as $\displaystyle S_{conv}=S_{diff}+S_{visc}$ (36) where $\displaystyle S_{conv}=\left\langle\bar{\rho}u_{z}s\right\rangle_{h},$ (37) the net entropy carried out of the region $(0,z)$ by the convective velocity at level $z$, $\displaystyle S_{diff}=-\frac{k}{T_{B}}\frac{\mathrm{d}}{\mathrm{d}z}\left\langle T\right\rangle_{h}\Big{|}_{z=0}+\frac{k}{\bar{T}}\frac{\mathrm{d}}{\mathrm{d}z}\left\langle T\right\rangle_{h}\Big{|}_{z}-\int_{0}^{z}\frac{k\Delta{\bar{T}}}{{\bar{T}}^{2}d}\frac{\mathrm{d}}{\mathrm{d}z}\left\langle T\right\rangle_{h}\,dz$ (38) so that $S_{diff}$ is the entropy balance in region $(0,z)$ of the entropy change due to thermal diffusion. This is divided into the first term, which represents the positive entropy being conducted into our region at the bottom boundary (the gradient of $\langle T\rangle_{h}$ is negative there, see figure 1b), the second term is the entropy conducted across level $z$, and the third term is the entropy production by internal diffusion in our given region. By looking at figure 1b it is apparent that when the boundary layers are thin, the first term is much larger than the other two except when $z$ is in the boundary layers. If $z$ is in the bulk, the gradient of $\langle T\rangle_{h}$ is $O(\Delta{\bar{T}}/d)$ whereas at the boundary it is $O(\Delta{\bar{T}}/\delta^{th})$, much larger since the boundary layer is thin. The third term is always small compared to the first, because the gradient is $O(\Delta{\bar{T}}/d)$ outside the boundary layers. The integrand is of order $O(\Delta{\bar{T}}/\delta^{th})$ in the boundary layers, but because they are thin this only contributes a small amount to the integral. So when the boundary layers are thin $\displaystyle S_{diff}\approx-\frac{k}{T_{B}}\frac{\mathrm{d}}{\mathrm{d}z}\left\langle T\right\rangle_{h}\Big{|}_{z=0}\quad\textrm{if $z$ is in the bulk},$ (39) $\displaystyle S_{diff}\approx-\frac{k}{T_{B}}\frac{\mathrm{d}}{\mathrm{d}z}\left\langle T\right\rangle_{h}\Big{|}_{z=0}+\frac{k}{T_{T}}\frac{\mathrm{d}}{\mathrm{d}z}\left\langle T\right\rangle_{h}\Big{|}_{z=d}\quad\textrm{if $z=d$ . }$ (40) We now turn to $S_{visc}=\int_{0}^{z}\frac{\mu}{\bar{T}}\left\langle q\right\rangle_{h}\,dz+\int_{0}^{z}\frac{\mu}{\bar{T}}\left\langle\partial_{j}(\partial_{i}(u_{i}u_{j}))-2\partial_{j}(u_{j}(\partial_{i}u_{i}))\right\rangle_{h}\,dz.$ (41) Because of the horizontal averaging, and using equations (2), ($None$) and (32), the second integral can be written $\displaystyle\int_{0}^{z}\frac{\mu}{\bar{T}}\left\langle\partial_{j}(\partial_{i}(u_{i}u_{j}))-2\partial_{j}(u_{j}(\partial_{i}u_{i}))\right\rangle_{h}\,dz=\qquad\qquad$ $\displaystyle\int_{0}^{z}\frac{\mu}{\bar{T}}\left[\frac{\partial^{2}}{\partial z^{2}}\left\langle u_{z}^{2}\right\rangle_{h}-\frac{m\Delta{\bar{T}}}{{\bar{T}}d}\frac{\partial}{\partial z}\left\langle u_{z}^{2}\right\rangle_{h}-\frac{m(\Delta{\bar{T}})^{2}}{{\bar{T}}^{2}d^{2}}\left\langle u_{z}^{2}\right\rangle_{h}\right]\,dz.$ (42) We now consider the magnitude of the terms in equation (41). If the root mean square velocity in the bulk is $U$, we expect the horizontal velocity to vary from 0 to $O(U)$ across the boundary layers of thickness $\delta_{i}^{\nu}$, so the velocity gradients appearing in $q$ are of magnitude $O(U/\delta_{i}^{\nu})$. $q$ itself is therefore $O(U^{2}/{(\delta_{i}^{\nu}})^{2})$, and since the boundary layers are thin their contribution to the first integral in $S_{visc}$ is $O(\mu U^{2}/{\bar{T}}\delta_{i}^{\nu})$. In the boundary layers $u_{z}$ is small, but in the bulk we expect $u_{z}$ to be $O(U)$ and so $\langle u_{z}^{2}\rangle_{h}$ is $O(U^{2})$. Because $\langle u_{z}^{2}\rangle_{h}$ is horizontally averaged, it will vary on a length-scale $d$ with $z$, so the gradient of $\langle u_{z}^{2}\rangle_{h}$ is $O(U^{2}/d)$ and the second derivative is $O(U^{2}/d^{2})$. From (42), the order of magnitude of the second term in (41) is then $O(\mu U^{2}/Td)$, which is $O(\delta_{i}^{\nu}/d)$ smaller than the contribution from the term due to $q$. Therefore provided the Rayleigh number is high enough for the boundary layers to be thin, equation (34) is asymptotically equivalent to $\displaystyle\left\langle\bar{\rho}u_{z}s\right\rangle_{h}\sim-\frac{k}{T_{B}}\frac{\mathrm{d}}{\mathrm{d}z}\left\langle T\right\rangle_{h}\Big{|}_{z=0}+\int_{0}^{z}\frac{\mu}{\bar{T}}\left\langle q\right\rangle_{h}\,dz.$ (43) when $z$ is in the bulk. Note that this simplification still holds if the dissipation in the bulk is larger than the dissipation in the boundary layers, which can happen, as noted by Grossmann & Lohse (2000). When the dissipation is primarily in the boundary layers, the left-hand-side of ( 43) is constant in the bulk, which we exploit later. In either case, as $z\to d$ we get $\displaystyle\frac{k}{T_{B}}\frac{\mathrm{d}}{\mathrm{d}z}\left\langle T\right\rangle_{h}\Big{|}_{z=0}-\frac{k}{T_{T}}\frac{\mathrm{d}}{\mathrm{d}z}\left\langle T\right\rangle_{h}\Big{|}_{z=d}=\frac{F^{super}\Delta{\bar{T}}}{T_{B}T_{T}}\sim\int_{0}^{d}\frac{\mu}{\bar{T}}\left\langle q\right\rangle_{h}\,dz.$ (44) Note also that in these expressions, the term $(\nabla\cdot{\bf u})^{2}/3$ in equation (5) makes a negligible contribution to (44) compared to the gradient terms by using (32). ### 3.3 The Boussinesq limit At first sight, it appears that our entropy balance formulation of the equation for dissipation (LABEL:eq:3_7) is fundamentally different from that used by Grossman & Lohse (2000) in the Boussinesq case, who start from equation (2.5) of Siggia (1994), $(Nu-1)Ra=\langle(\nabla v)^{2}\rangle,\ \textrm{or}\quad g\alpha\kappa\Delta T(Nu-1)=\nu\int_{0}^{d}\langle q\rangle_{h}\,dz$ (45) in our dimensional units. Here $\Delta T$ is the superadiabatic temperature difference between the boundaries. In the Boussinesq limit $\Gamma\to 1$, the basic state temperature and density tend to constant values $T_{B}$ and $\rho_{B}$ respectively, so the thermal diffusivity $\kappa$ and kinematic viscosity $\nu$ are constants, $\kappa=k/\rho_{B}c_{p}$ and $\nu=\mu/\rho_{B}$. For a perfect gas the coefficient of expansion $\alpha=1/T_{B}$. In this subsection we show that (45) is in fact the Boussinesq limit of the entropy balance equation (LABEL:eq:3_7), which we use to derive the scaling laws in §6 below. Our entropy balance formulation is a generalization of the Grossmann & Lohse (2000) formulation, which is now seen to be a limiting case of the more general entropy balance approach. Following Spiegel & Veronis (1960) we note that from the $z$-component of (1) $p/{\bar{\rho}}d\sim gs/c_{p}$, so $\frac{p}{\bar{p}}\sim\frac{g{\bar{\rho}}d}{{\bar{p}}}\frac{s}{c_{p}}\sim\frac{d}{H}\frac{s}{c_{p}}$ (46) where $H$ is the pressure scale height. In the Boussinesq limit $d/H$ becomes small; the Boussinesq limit $\Gamma\to 1$ is the thin layer limit (Spiegel & Veronis, 1960). Then ($None$a,b) become $\frac{T}{T_{B}}\sim-\frac{\rho}{\rho_{B}},\quad\frac{s}{c_{p}}\sim\alpha T,$ (47) so the entropy variable becomes equivalent to the superdiabatic temperature variable. The entropy jump $\Delta S$ across the layer can be written as a superadiabatic temperature jump $\Delta T=\Delta S/\alpha c_{p}$. As $\Gamma\to 1$, from (12) $\Delta S/c_{p}\to 1$ so $\Delta T\to 1/\alpha=T_{B}$. From energy conservation (33) the gradient of $\langle T\rangle_{h}$ is the same at the top and bottom of the layer, so (LABEL:eq:3_7) can be written $-\frac{\Delta{\bar{T}}}{T_{B}^{2}}\frac{d}{dz}\langle T\rangle_{h}{\Big{|}}_{z=0}=-\frac{k\Delta{\bar{T}}}{T_{B}^{2}d}\int_{0}^{d}\frac{d}{dz}\langle T\rangle_{h}\,dz+\frac{\mu}{T_{B}}\int_{0}^{d}\langle q\rangle_{h}\,dz,$ (48) using the constancy of $\bar{T}$ in the Boussinesq limit. From (14) and (15) $Nu=-\frac{d}{T_{B}}\frac{d}{dz}\langle T\rangle_{h}{\Big{|}}_{z=0}\ \textrm{or}\quad Nu=-\frac{d}{\Delta T}\frac{d}{dz}\langle T\rangle_{h}{\Big{|}}_{z=0},$ (49) which is the familiar form of the Boussinesq Nusselt number, the ratio of the total heat flux at the bottom to the conducted heat flux $-k\Delta T/d$. From ($None$d) $\Delta{\bar{T}}$ can be written $gd/c_{p}$ so (48) becomes $\frac{k}{c_{p}T_{B}}Nug\alpha\Delta T=\frac{k}{c_{p}T_{B}}g\alpha\Delta T+\frac{\mu}{T_{B}}\int_{0}^{d}\langle q\rangle_{h}\,dz\ \ \textrm{or}\ \ (Nu-1)ga\Delta T\kappa=\nu\int_{0}^{d}\langle q\rangle_{h}\,dz,$ (50) which is (45), showing that the dissipation integral which plays a key role in Grossmann & Lohse’s (2000) analysis is indeed the Boussinesq limit of the entropy balance equation (LABEL:eq:3_7). ## 4 The boundary layers and Prandtl number effects (a) (b) Figure 2: (a) Thermal and viscous boundary layers in the case $Pr>1$. The thermal diffusion is smaller, so the thermal boundary layer is nested inside the viscous boundary layer. (b) The case $Pr<1$, where the viscous boundary layer is nested inside the thermal boundary layer. As in the Boussinesq case, the thermal and viscous boundary layers can be nested inside each other when the Prandtl number is different from unity. A central idea in the theory of fully developed Boussinesq convection is based on the assumption that the structure of turbulent convective flow is always characterized by the presence of a large-scale convective roll called the _wind of turbulence_ , Grossmann & Lohse (2000). This idea, which in the non- stratified case stems from vast numerical and experimental evidence, is retained in the case of anelastic convection. However, the significant stratification in the anelastic case breaks the Boussinesq up-down symmetry, and thus we must distinguish between the magnitude of the wind of turbulence near the bottom of the bulk and its magnitude near the top of the bulk, denoted by $U_{B}$ and $U_{T}$ respectively, which can now significantly differ. So the label $U_{i}$ can denote either the horizontal velocity just outside the top or bottom viscous boundary layers. We also assume that this wind of turbulence forms boundary layers with a horizontal length scale comparable to the layer depth $d$. It is of course an assumption that such layers form in anelastic convection, but they are observed to occur in incompressible flow, and the limited simulations we have available gives this idea some support. Whereas the results in §2 and §3 are asymptotically valid in the anelastic framework in the limit of large $Ra$, what follows is dependent on the Grossmann-Lohse (2000) approach being valid for anelastic convection. The Prandtl number is a constant in this problem, given by $Pr=\frac{\mu c_{p}}{k}.$ (51) In figure (2a) the high Prandtl number case is shown, with the thinner thermal boundary layer nested inside the viscous boundary layer. The velocity at the edge of the thermal boundary layer is then $\delta_{i}^{th}U_{i}/\delta_{i}^{\nu}$, assuming the velocity falls off linearly inside the viscous boundary layer over the thinner thermal boundary layer. In the boundary layers, advection balances diffusion, so from the momentum equation (1) we estimate that $\frac{\rho_{i}U_{i}^{2}}{d}\sim\frac{\mu U_{i}}{\left(\delta_{i}^{\nu}\right)^{2}}\ \ \textrm{so}\ \ U_{i}\sim\frac{\mu d}{\rho_{i}\left(\delta_{i}^{\nu}\right)^{2}}.$ (52) For the entropy boundary layers, from (3) $\frac{\rho_{i}T_{i}U_{i}s}{d}\frac{\delta_{i}^{th}}{\delta_{i}^{\nu}}\sim k\nabla^{2}T\approx\frac{kT_{i}}{c_{p}}\nabla^{2}s\sim\frac{kT_{i}s}{c_{p}\left(\delta_{i}^{th}\right)^{2}},\ \ \textrm{so}\ \ U_{i}=\frac{kd\delta_{i}^{\nu}}{\rho_{i}c_{p}\left(\delta_{i}^{th}\right)^{3}},$ (53) where (25) has been used and the factor $\delta_{i}^{th}/{\delta_{i}^{\nu}}$ arises because the horizontal velocity is reduced because the entropy boundary layer is thinner than the viscous boundary layer. Dividing (52) by (53) we obtain $\frac{\delta_{i}^{\nu}}{\delta_{i}^{th}}\sim Pr^{1/3}$ (54) giving the ratio of the viscous to thermal boundary layer thickness for the high Prandtl number case. Note that although the top and bottom boundary layers have different thicknesses, this ratio is the same for both layers. For the low Prandtl number case, the viscous boundary layer lies inside the thermal boundary layer, see figure (2b). Now the velocity at the edge of the thermal boundary layer is the same as that at the edge of the viscous boundary layer, so the velocity reduction factor $\delta_{i}^{th}/\delta_{i}^{\nu}$ is no longer required, so (53) becomes $\frac{\rho_{i}T_{i}U_{i}s}{d}\sim k\nabla^{2}T\approx\frac{kT_{i}}{c_{p}}\nabla^{2}s\sim\frac{kT_{i}s}{c_{p}\left(\delta_{i}^{th}\right)^{2}},\ \ \textrm{so}\ \ U_{i}=\frac{kd}{\rho_{i}c_{p}\left(\delta_{i}^{th}\right)^{2}},$ (55) giving the ratio of the boundary layer thicknesses as $\frac{\delta_{i}^{\nu}}{\delta_{i}^{th}}\sim Pr^{1/2}$ (56) in the low Prandtl number case. ## 5 The boundary layer ratio problem In Boussinesq convection, there is a symmetry about the mid-plane which means that the top and bottom boundary layers have the same thickness and structure, and the temperature of the bulk interior is midway between that of the boundaries. In anelastic convection, this symmetry no longer holds, and the top and bottom boundary layers can be very different, and the entropy of the bulk interior is significantly different from $\Delta S/2$. This raises the question of how the ratios of the thicknesses of the top and bottom boundary layers, the ratio of the bulk horizontal velocities just outside the boundary layers, and the ratio of the entropy jumps across the layers are actually determined. We assume the incompressible boundary layer case holds throughout this section. We write the ratio of the entropy jumps across the boundary layers as $r_{s}=\frac{\Delta S_{T}}{\Delta S_{B}},\ \ \textrm{so}\ \ \Delta S_{B}=\frac{\Delta S}{1+r_{s}}\ \ \textrm{and the entropy in the bulk is}\quad\frac{r_{s}}{1+r_{s}}\Delta S,$ (57) and the ratio of the anelastic temperature jumps across the boundary layers as $r_{T}=\frac{\Delta T_{T}}{\Delta T_{B}}.$ (58) We define the ratio of the velocities at the edge of the boundary layers as $r_{u}=\frac{U_{T}}{U_{B}}.$ (59) The last important ratio is the ratio of the thicknesses of the boundary layers. In general the viscous and thermal boundary layers will be of different thickness, but here we start with the thermal boundary layers which have thicknesses at the top and bottom of $\delta^{th}_{T}$ and $\delta^{th}_{B}$ with ratio $r_{\delta}=\frac{\delta^{th}_{T}}{\delta^{th}_{B}}.$ (60) We have four unknown ratios, so we need four equations to determine them. Our first equation uses the fact that the heat flux passing through the bottom boundary must equal the heat flux passing through the top boundary. Since this heat flux is entirely by conduction close to the boundary, $-k\frac{dT}{dz}|_{i}\sim k\frac{\Delta T_{i}}{\delta^{th}_{i}}\Rightarrow r_{\delta}=r_{T}.$ (61) We can use the balance of advection and diffusion in the boundary layers exploited in §4 to obtain another equation relating the boundary layer ratios. In §4 we saw that the ratio of the thermal boundary layer thicknesses was the same as the ratio of the viscous boundary layer thicknesses, $\frac{\delta^{th}_{T}}{\delta^{th}_{B}}=\frac{\delta^{\nu}_{T}}{\delta^{\nu}_{B}}=r_{\delta},$ (62) so we use the viscous boundary balance equation (52) to estimate $\frac{\rho_{B}U_{B}^{2}}{d}\sim\frac{\mu U_{B}}{{\delta^{\nu}_{B}}^{2}},\quad\frac{\rho_{T}U_{T}^{2}}{d}\sim\frac{\mu U_{T}}{{\delta^{\nu}_{T}}^{2}}\quad\Rightarrow r_{u}{r_{\delta}}^{2}\sim\frac{\rho_{B}}{\rho_{T}}=\Gamma^{m}.$ (63) We now need an equation for the ratio of bulk large scale flow velocities at the top and bottom of the layer, $r_{U}$. We consider first the case where the viscous dissipation occurs primarily in the boundary layers, which is likely to be true in numerical simulations with no-slip boundaries. Since the entropy production occurs in the boundary layers and is relatively small in the interior, and entropy diffusion is small in the bulk interior, the convected entropy flux $\langle\bar{\rho}u_{z}s\rangle_{h}$ is approximately constant in the bulk interior. We now multiply the equation of motion (1) by ${\bar{\rho}}{\bf u}$ and average over the bulk interior, ignoring the small viscous term in the bulk, to get $\frac{1}{2}\frac{\partial}{\partial z}\left(\bar{\rho}\left\langle u_{z}u^{2}\right\rangle_{h}\right)\approx-\frac{\partial}{\partial z}\langle u_{z}p\rangle_{h}+\frac{g}{c_{p}}\left\langle{\bar{\rho}}u_{z}s\right\rangle_{h}.$ (64) Near the boundary layers, the pressure term $p$ will be approximately constant as shown in (23), and since $\langle u_{z}\rangle_{h}=0$, the term involving $\langle u_{z}p\rangle_{h}$ will be small there, and we ignore it. Since we expect $\langle\bar{\rho}u_{z}s\rangle_{h}$ to be approximately the same just outside the two boundary layers, $\frac{\partial}{\partial z}\left(\bar{\rho}\left\langle u_{z}u^{2}\right\rangle_{h}\right)|_{T}\approx\frac{\partial}{\partial z}\left(\bar{\rho}\left\langle u_{z}u^{2}\right\rangle_{h}\right)|_{B},$ (65) where here $T$ and $B$ refer to conditions at the top of the bulk and the bottom of the bulk respectively. In the turbulent bulk interior (unlike the boundary layers), we expect the three components of velocity to have similar magnitudes. It remains to estimate the length-scale associated with the $z$-derivative, and this is perhaps the most uncertain part of the analysis. Astrophysical mixing length theory uses the pressure scale height, or a multiple of the pressure scale height, as the mixing length for the vertical length scale. Since we are only interested in the top and bottom ratios here, our results are independent of what multiple of the scale height is chosen. Some support for the mixing length idea can be derived from Kessar et al. (2019), which shows that convective length scales decrease in the bulk towards the top of the layer. We also note that because we are only concerned with ratios, it doesn’t matter whether the pressure scale height or the density scale height is used. Adopting the pressure scale height, $H_{T}=\frac{d}{(m+1)(\Gamma-1)},\quad H_{B}=\frac{\Gamma d}{(m+1)(\Gamma-1)}\quad\Rightarrow\quad\frac{H_{T}}{H_{B}}=\Gamma^{-1}$ (66) so that equation (65) gives $\frac{\rho_{T}u_{T}^{3}}{H_{T}}\sim\frac{\rho_{B}u_{B}^{3}}{H_{B}}\Rightarrow r_{u}^{3}\sim\Gamma^{m-1}\Rightarrow r_{u}\sim\Gamma^{\frac{m-1}{3}}.$ (67) From the incompressible boundary layer equation (25) we have $r_{s}=\Gamma r_{T}$, so with the other three ratio equations (61), (63) and (67) we have $r_{s}=\Gamma r_{T},\quad r_{\delta}=r_{T},\quad r_{U}{r_{\delta}}^{2}=\Gamma^{m},\quad{r_{u}}=\Gamma^{\frac{m-1}{3}},$ $None$ with solution $r_{u}=\Gamma^{\frac{m-1}{3}},\quad r_{\delta}=\Gamma^{\frac{2m+1}{6}},\quad r_{s}=\Gamma^{\frac{2m+7}{6}},\quad r_{T}=\Gamma^{\frac{2m+1}{6}}.$ (69) In the case where $m=3/2$, appropriate for ideal monatomic gas, $r_{u}=\Gamma^{\frac{1}{6}},\quad r_{\delta}=\Gamma^{\frac{2}{3}},\quad r_{s}=\Gamma^{\frac{5}{3}},\quad r_{T}=\Gamma^{\frac{2}{3}}.$ (70) Since $\Gamma$ is always greater than 1 and can be large, we see that the entropy in the bulk is much closer to the entropy of the bottom boundary than to the entropy of the top boundary. The top boundary layer is thicker than the bottom boundary layer. The challenge for numerical simulations is to get to sufficiently high Rayleigh number that the top boundary layer is truly thin, as required by our asymptotic analysis. The ratio of the bulk velocities at the top and bottom, is only weakly dependent on $\Gamma$, so again rather large values of $Ra$ are required to establish the asymptotic trend. Note that in deriving equation (63) we assumed that the horizontal length scale for advection along the boundary layer was $d$, as did Grossmann & Lohse (2000). We found that choosing the vertical length scales in the bulk to be based on the pressure scale height in equation (66) agreed reasonably with the numerics, see § 7 below, so a natural question is whether the horizontal length scale near the top boundary becomes less than $d$ when the layer is strongly stratified. The picture from our numerics is somewhat mixed, and is discussed further in § 7 below. For moderate stratification, the numerics suggest the boundary layers do appear to extend to $d$ at both the top and bottom boundary; including a factor $\Gamma$ in the horizontal length scales in the boundary layers gives poorer agreement with numerical estimates of the boundary layer thickness ratio. However, at the largest values of $\Gamma$ we did find a departure from the (70) scalings which could be accounted for by some reduction in the horizontal length scale near the top boundary. In the case where the viscous dissipation is mainly in the bulk, which happens at low $Pr$ (Grossmann & Lohse, 2000) the equations (57-63) still hold, but the argument for equation (67) breaks down because the entropy flux is no longer approximately constant in the bulk because viscous dissipation in the bulk is no longer negligible. This case is discussed in Appendix B. ## 6 The Nusselt number and Reynolds number scaling laws When the boundary layers are thin, the overall entropy balance reduced to (44), $\frac{F^{super}\Delta{\bar{T}}}{T_{B}T_{T}}\sim\int_{0}^{d}\frac{\mu}{\bar{T}}\left\langle q\right\rangle_{h}\,dz.$ (71) If the dissipation is mainly in laminar boundary layers, the $z$-derivatives of the horizontal velocities will dominate terms in the expression for $q$, so $q=(\partial_{j}u_{i})(\partial_{j}u_{i})+\frac{1}{3}\left(\nabla\cdot\mathbf{u}\right)^{2}\approx\left(\frac{\partial u_{x}}{\partial z}\right)^{2}+\left(\frac{\partial u_{y}}{\partial z}\right)^{2}\sim\frac{U_{i}^{2}}{(\delta_{i}^{\nu})^{2}},$ (72) in these layers. So integrating over the boundary layers of thickness $\delta_{i}^{\nu}$ and assuming $T$ varies little in these layers, $\frac{F^{super}\Delta{\bar{T}}}{T_{B}T_{T}}\sim\frac{\mu U_{B}^{2}}{T_{B}\delta_{B}^{\nu}}+\frac{\mu U_{T}^{2}}{T_{T}\delta_{T}^{\nu}}=\frac{\mu U_{B}^{2}}{T_{B}\delta_{B}^{\nu}}\left(1+\frac{\Gamma r_{u}^{2}}{r_{\delta}}\right).$ (73) where we have used the ratios (59), (60) and ($None$g). Now we can write the superadiabatic flux in terms of the thermal boundary layer thicknesses, using (14), (26), (25), (57) and (18) to get $F^{super}=\frac{k\Delta T_{B}}{\delta_{B}^{th}}=\frac{kT_{B}\Delta S}{c_{p}(1+r_{s})\delta_{B}^{th}}.$ (74) Inserting this into (73) and using the definition (16) for the Rayleigh number, the entropy balance equation can be written $\frac{k^{2}Ra\Gamma}{c_{p}^{2}(1+r_{s})d^{2}\rho_{B}^{2}}\frac{\delta_{B}^{\nu}}{\delta_{B}^{th}}\sim U_{B}^{2}\left(1+\frac{\Gamma r_{u}^{2}}{r_{\delta}}\right).$ (75) Now we introduce the Reynolds number near the bottom boundary $Re_{B}=\frac{\rho_{B}U_{B}d}{\mu},$ (76) noting that the Reynolds number near the top, $Re_{T}$, is given by $r_{u}Re_{B}$. We also use the definition of the Prandtl number, (51), to write (75) as $\frac{Ra\Gamma}{Pr^{2}(1+r_{s})}\frac{\delta_{B}^{\nu}}{\delta_{B}^{th}}\sim Re_{B}^{2}\left(1+\frac{\Gamma r_{u}^{2}}{r_{\delta}}\right).$ (77) The entropy balance equation has thus given us a relation between the Reynolds number and the Rayleigh number, which is similar to that of regime I of Grossmann & Lohse (2000) but with additional factors of $\Gamma$. In the high Prandtl number limit applying (54) gives $Re_{B}\sim Ra^{1/2}Pr^{-5/6}\Gamma^{1/2}(1+r_{s})^{-1/2}\left(1+\frac{\Gamma r_{u}^{2}}{r_{\delta}}\right)^{-1/2},$ (78) while in the low Prandtl number case we get using (55) $Re_{B}\sim Ra^{1/2}Pr^{-3/4}\Gamma^{1/2}(1+r_{s})^{-1/2}\left(1+\frac{\Gamma r_{u}^{2}}{r_{\delta}}\right)^{-1/2}.$ (79) We now use the viscous boundary layer balance between advection and diffusion, (52), $\rho_{B}U_{B}/d\sim\mu/(\delta^{\nu}_{B})^{2}$, to obtain a balance between $Nu$ and $Re_{B}$. The boundary layer balance becomes $Re_{B}\sim\left(\frac{d}{\delta_{B}^{\nu}}\right)^{2},\ \textrm{but}\ Nu=\frac{d}{\delta_{B}^{th}}\frac{\Gamma\ln\Gamma}{(1+r_{s})(\Gamma-1)}$ (80) using the expression (27) for the Nusselt number together with (12) and (57). Eliminating $d/\delta_{B}^{th}$ between these, $Re_{B}^{1/2}=\frac{\delta_{B}^{th}}{\delta_{B}^{\nu}}\frac{Nu(\Gamma-1)(1+r_{s})}{\Gamma\ln\Gamma}.$ (81) As above, the ratio of the boundary layer thicknesses can be evaluated in terms of the Prandtl number, and (81) allows us to eliminate $Re_{B}$ from (77) to obtain the Nusselt number as a function of Rayleigh number, $\frac{Ra\Gamma}{Pr^{2}(1+r_{s})}\left(\frac{\delta_{B}^{\nu}}{\delta_{B}^{th}}\right)^{5}\sim\left(\frac{Nu(\Gamma-1)(1+r_{s})}{\Gamma\ln\Gamma}\right)^{4}\left(1+\frac{\Gamma r_{u}^{2}}{r_{\delta}}\right).$ (82) At large $Pr$, (54) gives the Nusselt number – Rayleigh number scaling in the form $Nu\sim Ra^{1/4}Pr^{-1/12}\frac{\Gamma^{5/4}\ln\Gamma}{\Gamma-1}(1+r_{s})^{-5/4}\left(1+\frac{\Gamma r_{u}^{2}}{r_{\delta}}\right)^{-1/4},$ (83) while at low $Pr$ (56) gives $Nu\sim Ra^{1/4}Pr^{1/8}\frac{\Gamma^{5/4}\ln\Gamma}{\Gamma-1}(1+r_{s})^{-5/4}\left(1+\frac{\Gamma r_{u}^{2}}{r_{\delta}}\right)^{-1/4}.$ (84) If we accept the ratio scalings derived in §5, in the case of a monatomic ideal gas, $\gamma=5/3,m=3/2$, we can write $(1+r_{s})^{-5/4}\left(1+\frac{\Gamma r_{u}^{2}}{r_{\delta}}\right)^{-1/4}=\left(1+\Gamma^{5/3}\right)^{-5/4}\left(1+\Gamma^{2/3}\right)^{-1/4},$ (85) so as $\Gamma$ becomes large, $Nu$ decreases as $\ln\Gamma\Gamma^{-2}$. So we expect the Nusselt number, the dimensionless measure of the heat transport, to be considerably smaller when the layer is strongly compressible, i.e. when $\Gamma$ is large and there are many density scale heights in the layer at a given $Ra$ and $Pr$. Analogous results for the case where the dissipation is mainly in the bulk rather than in the boundary layers, as can happen at low $Pr$, are given in Appendix B. In the Boussinesq limit, $\Gamma$ is close to unity and $\theta=(\Gamma-1)/\Gamma$ is small, so $\ln\Gamma/(\Gamma-1)\to 1$ and $\rho_{B}\to\rho_{T}$ and $T_{B}\to T_{T}$. Equation (1) reduces to the usual Boussinesq equation with $s/c_{p}$ replaced by $\alpha T$, where for a perfect gas $\alpha=1/{\bar{T}}$ is the coefficient of expansion, consistent with (25). The total jump of entropy across the layer, $\Delta S$, is replaced by the total temperature jump $\Delta T=\Delta S/\alpha c_{p}$ so the Rayleigh number (16) reduces to the familiar form $Ra=g\alpha\Delta Td^{3}/\kappa\nu$ where $\kappa=k/{\bar{\rho}}c_{p}$ and $\nu=\mu/{\bar{\rho}}$ are the thermal diffusivity and kinematic viscosity respectively. These are both constant in the Boussinesq limit, and the ratios $r_{u}$, $r_{\delta}$ and $r_{s}$ all go to unity, see §5. Our scaling laws (78), (79), (83) and (84) all reduce to those of regimes $I_{u}$ and $I_{l}$ of Grossmann & Lohse (2000). Grossmann & Lohse also give suggested prefactors for their scaling laws in table 2 of their paper, and since our anelastic scaling laws reduce to theirs in the Boussinesq limit, our prefactors should in theory be consistent with theirs. In practice, the prefactors depend on the aspect ratio of the experiments (or numerical experiments) used to determine them (see e.g. Chong et al., 2018). For the case of the high Prandtl number regime $I_{u}$, their values were $Nu\approx 0.33Ra^{1/4}Pr^{-1/12}$ and $Re\approx 0.039Ra^{1/2}Pr^{-5/6}$, so (83) becomes $Nu\approx C_{Nu}Ra^{1/4}Pr^{-1/12}\frac{\Gamma^{5/4}\ln\Gamma}{\Gamma-1}(1+r_{s})^{-5/4}\left(1+\frac{\Gamma r_{u}^{2}}{r_{\delta}}\right)^{-1/4},\quad C_{Nu}=0.93.$ (86) In the low $Pr$ limit where (84) applies, regime $I_{l}$ of Grossmann & Lohse (2000), they suggest a prefactor corresponding to $C_{Nu}=0.76$ rather than 0.93. For the Reynolds number, (78) becomes $Re_{B}\approx C_{Re}Ra^{1/2}Pr^{-5/6}\Gamma^{1/2}(1+r_{s})^{-1/2}\left(1+\frac{\Gamma r_{u}^{2}}{r_{\delta}}\right)^{-1/2},\quad C_{Re}=0.078.$ (87) There is some uncertainty about the prefactor $C_{Re}$, discussed in §7 below. Prefactors in the case $I_{l}$ and in the case where dissipation is mainly in the bulk, their case $II_{l}$, (see Appendix B) can also be found. ## 7 The numerical results and discussion (a) (b) (c) (d) (e) (f) Figure 3: Horizontally averaged entropy (in units of $c_{p}$) and horizontal mean velocity profiles (Peclet number units) from the numerical simulations for $\Gamma=1.9438$, $Ra=10^{6}$. (a) Entropy profile at $Pr=1$. (b) Horizontal velocity profile at $Pr=1$. (c) Entropy profile at $Pr=10$. (d) Horizontal velocity profile at $Pr=10$. (e) Entropy profile at $Pr=0.25$. (f) Horizontal velocity profile at $Pr=0.25$. (a) (b) (c) (d) Figure 4: Horizontally averaged entropy ${\langle}S{\rangle}_{h}$ and horizontal mean velocity $U_{H}$ profiles for (a,b) $\Gamma=2.924$, $Ra=3\times 10^{6}$, $Pr=1$: (c,d) $\Gamma=4.6416$, $Ra=6\times 10^{6}$, $Pr=1$. We have tested the theoretical predictions of our asymptotic theory using numerical simulations of high Rayleigh number plane layer anelastic convection. The numerical code differs from the theory in one respect, as it uses entropy diffusion $k_{s}$ rather than temperature diffusion in the heat equation, so $\bar{\rho}\bar{T}\left[\frac{\partial s}{\partial t}+\mathbf{u}\cdot\nabla s\right]=k_{s}\nabla\cdot{\bar{T}}\nabla s+\mu\left[q+\partial_{j}u_{i}\partial_{i}u_{j}-\left(\nabla\cdot\mathbf{u}\right)^{2}\right],$ (88) where $k_{s}$ is constant, replaces (3). This simplifies the code because entropy is the only anelastic thermodynamic variable computed, and it can be justified in circumstances where turbulent diffusion dominates laminar diffusion (Braginsky & Roberts, 1995). Constant entropy boundary conditions were used in the code. The energy balance equation (33) is only changed by replacing $-kd{\langle T\rangle}_{h}/dz$ by $-k_{s}{\bar{T}}d{\langle s\rangle}_{h}/dz$. In the entropy balance equation, entropy diffusion changes $S_{diff}$ to $\displaystyle S_{diff}=-{k_{s}}\frac{\mathrm{d}}{\mathrm{d}z}\left\langle s\right\rangle_{h}\Big{|}_{z=0}+{k_{s}}\frac{\mathrm{d}}{\mathrm{d}z}\left\langle s\right\rangle_{h}\Big{|}_{z}-\int_{0}^{z}\frac{k_{s}\Delta{\bar{T}}}{{\bar{T}}d}\frac{\mathrm{d}}{\mathrm{d}z}\left\langle s\right\rangle_{h}\,dz.$ (89) Just as in the temperature diffusion case, the last two terms are negligible when the entropy boundary layers are thin, except when $z$ is in the boundary layers, when the second term is significant. In the case when the viscous dissipation is primarily in the boundary layers, the argument leading to (65) still holds, so the ratios satisfy (69) in the entropy diffusion case as well as in the temperature diffusion case. It is therefore reasonable to compare with the numerical results for the entropy diffusion case in the expectation that they will be reasonably close to the temperature diffusion case. Run | A1 | A2 | A3 | A4 | A5 | ---|---|---|---|---|---|--- | B1 | B2 | C1 | D1 | D2 | D3 Ra | $10^{6}$ | $3\times 10^{6}$ | $10^{7}$ | $3\times 10^{6}$ | $6\times 10^{6}$ | | $10^{6}$ | $3\times 10^{6}$ | $10^{6}$ | $10^{6}$ | $10^{6}$ | $10^{6}$ Pr | $1$ | $1$ | $1$ | $1$ | $1$ | | $10$ | $10$ | $0.25$ | $1$ | $10$ | $0.25$ $\Gamma$ | $1.9438$ | $1.9438$ | $1.9438$ | $2.924$ | $4.6416$ | | $1.9438$ | $4.6416$ | $1.9438$ | $1$ | $1$ | $1$ $\rho_{b}/\rho_{T}$ | $2.71$ | $2.71$ | $2.71$ | $5$ | $10$ | | $2.71$ | $10$ | $2.71$ | $1$ | $1$ | $1$ $r_{\delta}$ | $1.65\pm 0.01$ | $1.70\pm 0.01$ | $1.69\pm 0.01$ | $2.05\pm 0.01$ | $2.27\pm 0.02$ | | $1.48\pm 0.01$ | $2.17\pm 0.01$ | $2.05\pm 0.04$ | 1 | 1 | 1 $r_{s}$ | $3.21\pm 0.02$ | $3.29\pm 0.01$ | $3.31\pm 0.01$ | $5.99\pm 0.02$ | $10.62\pm 0.03$ | | $2.91\pm 0.02$ | $10.0\pm 0.02$ | $4.04\pm 0.04$ | 1 | 1 | 1 $r_{u}$ | $0.92\pm 0.01$ | $0.92\pm 0.01$ | $0.93\pm 0.01$ | $0.96\pm 0.01$ | $1.06\pm 0.02$ | | $1.17\pm 0.01$ | $1.38\pm 0.01$ | $0.82\pm 0.02$ | 1 | 1 | 1 $Nu$ | $5.90\pm 0.02$ | $7.75\pm 0.03$ | $10.69\pm 0.03$ | $5.35\pm 0.02$ | $4.26\pm 0.02$ | | $6.18\pm 0.02$ | $4.09\pm 0.02$ | $5.00\pm 0.04$ | $8.78\pm 0.03$ | $8.76\pm 0.04$ | $8.17\pm 0.04$ $U_{T}$ | $186\pm 3$ | $321\pm 3$ | $570\pm 3$ | $298\pm 3$ | $368\pm 3$ | | $287\pm 2$ | $410\pm 3$ | $125\pm 4$ | $200\pm 3$ | $260\pm 5$ | $149\pm 3$ $U_{B}$ | $202\pm 3$ | $348\pm 3$ | $613\pm 3$ | $311\pm 3$ | $348\pm 3$ | | $244\pm 2$ | $298\pm 3$ | $154\pm 4$ | $200\pm 3$ | $260\pm 5$ | $149\pm 3$ $\Gamma^{2/3}$ | 1.557 | 1.557 | 1.557 | 2.045 | 2.783 | | 1.557 | 2.783 | 1.557 | 1 | 1 | 1 $\Gamma^{5/3}$ | 3.028 | 3.028 | 3.028 | 5.979 | 12.915 | | 3.028 | 12.915 | 3.028 | 1 | 1 | 1 $\Gamma^{1/6}$ | 1.117 | 1.117 | 1.117 | 1.196 | 1.292 | | 1.117 | 1.292 | 1.117 | 1 | 1 | 1 $Nu$-theory | 5.56 | 7.32 | 9.89 | 4.65 | 2.98 | | 5.55 | 2.50 | 5.18 | 8.78 | 8.76 | 8.17 $Nu$-nblr | 5.60 | 7.22 | 9.67 | 4.97 | 3.86 | | 5.63 | 3.11 | 4.37 | 8.78 | 8.76 | 8.17 $Pe_{T}$-theory | 194 | 336 | 614 | 307 | 376 | | 252 | 345 | 144 | 200 | 260 | 149 $Pe_{B}$-theory | 174 | 301 | 549 | 257 | 291 | | 226 | 267 | 129 | 200 | 260 | 149 $Pe_{T}$-nblr | 177 | 306 | 559 | 283 | 361 | | 256 | 358 | 118 | 200 | 260 | 149 $Pe_{B}$-nblr | 192 | 332 | 601 | 295 | 341 | | 219 | 260 | 144 | 200 | 260 | 149 Table 1: Data from the numerical runs all corresponding to $m=3/2$ polytropes. The first four rows are the input parameters. $r_{\delta}$, $r_{s}$ and $r_{u}$ are the measured boundary layer ratios for each run. The velocities $U_{T}$ and $U_{B}$ are the local maxima at the edge of the boundary layers, measured in velocity units of $k/d\rho_{B}c_{p}$, i.e they are Peclet numbers based on the diffusivity at the base of the layer. The theoretical predictions for the boundary layer ratios are given in the next three rows, see equation (70). The $Nu$-theory entries are based on (86) with the prefactors $C_{Nu}$ as given in the text, and the boundary ratios come from (70). The $Nu$-nblr entries also use (86) with the same prefactors, but instead of using (70), the numerical boundary layer ratios (nblr) above are used. The $Pe_{T}$-theory and $Pe_{B}$-theory entries come from (78) and (79). The prefactors used are not those of Grossmann & Lohse (2000), see (87), but those given in the text. Again, (70) is used to determine the boundary layer ratios. The $Pe_{T}$-nblr and $Pe_{B}$-nblr entries use the numerical boundary layer ratios. The numerical code is described in Kessar et al. (2019), though for that paper stress-free boundary conditions were applied, whereas no-slip boundaries where imposed in the runs described here. The unit of length is taken as $d$, the unit of time is $\rho_{B}d^{2}c_{p}/k$, the thermal diffusion time at the bottom of the layer. The velocities are in units of $k/\rho_{B}dc_{p}$, so they correspond to local Peclet numbers, where $Pe=RePr$. All the runs have polytropic index $m=3/2$. The code assumes periodic boundary conditions in the horizontal $x$ and $y$ directions, with aspect ratio 2, that is the period in the horizontal directions is $2d$. Table 1 gives the parameters used in the eleven runs, which span a range of Prandtl numbers and are at Rayleigh numbers which are as large as is numerically feasible, bearing in mind the need to resolve the small scale structures that develop. The last three runs are for Boussinesq cases, $\Gamma=1$, for comparison with the anelastic cases. The density stratification measured by $\Gamma$ varies over a moderate range only, because for the modest values of $Ra$ that are numerically accessible, large $\Gamma$ leads to a top boundary layer which is no longer thin, so our theory will no longer be valid. In figure 3, the entropy profiles (in units of $c_{p}$) and the horizontal velocity profiles are shown for the three runs A1, B1 and C1, and the profiles for A4 and A5 are shown in figure 4. The entropy profiles are constructed by horizontal averaging and time averaging the vertical profiles. From (12) the entropy difference between the boundaries is $\Gamma\ln\Gamma/(\Gamma-1)$ and the constant is chosen so that it is zero at the bottom boundary $z=0$. From figure 3 it is immediately apparent that the entropy is indeed rather constant in the well-mixed bulk interior, consistent with a key assumption of the theory. It is also clear that the jump in entropy across the top boundary layer is greater than that across the bottom boundary layer, and that the top entropy boundary layer is thicker than the bottom entropy boundary layer. This is consistent with the boundary layer ratios found in §5. The velocity profiles have local maxima near the boundaries, which gives a convenient definition for the top and bottom Reynolds numbers, $Re_{B}$ and $Re_{T}$, after converting Peclet numbers to Reynolds numbers using $RePr=Pe$. We note that there is no great difference between the top and bottom horizontal velocities, consistent with the weak scaling with $\Gamma$ found in (67). This result is slightly surprising, because astrophysical mixing length theory predicts faster velocities where the fluid is less dense, but in our problem the boundaries play an important role. The low Prandtl number case, figures 3(e) and 3(f), has a slightly different entropy boundary layer profile from those of the $Pr=1$ and $Pr=10$ cases, with a more gradual matching on to the uniform entropy interior, which is particularly noticeable in the upper boundary layer. This suggests it may be necessary to go to higher $Ra$ at low $Pr$ before accurate agreement with a theory that assumes thin boundary layers can be obtained. The cases shown in figure 4, together with figures 3(a) and 3(b), form a sequence at constant $Pr=1$ with increasing $\Gamma$. The most noticeable feature is that at larger $\Gamma$ the entropy of the mixed interior becomes close to that of the bottom boundary, so the boundary layer ratio $r_{S}$ increases rapidly, consistent with the prediction of (70). Also notable is that the velocity ratio of the maximum horizontal velocities, $r_{U}$, is never far from unity, again consistent with the weak scaling with $\Gamma$ in (70). As expected, the boundary layers become thicker at larger $\Gamma$, so that at fixed $Ra$ the thin boundary layer assumption breaks down at large $\Gamma$. In table 1 we compare various results of the simulations against the theoretical predictions of §5 and §6. To evaluate the entropy boundary layer thicknesses for our numerical data we use the definition (22), so the gradients $d\langle S\rangle_{h}/dz$ at $z=0$ and $z=1$ were obtained by differentiating a cubic spline representation of the entropy, and the entropy jumps were obtained by averaging the entropy over the well mixed region, assuming constant entropy there. The ratios of the top and bottom entropy thicknesses and entropy jumps are denoted by $r_{\delta}$ and $r_{s}$ in table 1. The velocity ratios at the top and bottom are denoted by $r_{u}$. We can compare these with the predicted ratios in (70). We see that there is some variation in the numerical results, but they are roughly in agreement with the predicted results. Considering that the top boundary layer is not that thin, as can be seen in figures 3 and 4, these results are as good as can be expected. We also tested how well the equations leading to our boundary layer ratios compare individually with the numerical results. Equation (61) expresses the fact that the heat flux is the same at the top and bottom, and together with the incompressible boundary layer equation (25), gives $r_{s}=\Gamma r_{\delta}$, which agrees with our numerics rather well, to the 1% level. The viscous boundary layer balance, (63), has less accurate agreement with the numerics. All the runs at $\Gamma=1.9438$ had errors less than 10% (except the low $Pr$ run where as we saw in figure 3 the boundary layer structure is slightly different), but the runs with density ratio 10, A5 and B2, have $r_{\delta}$ and $r_{s}$ large, but not quite as large as predicted by (70). The likely explanation is that the horizontal length scale at the top boundary layer is getting smaller than at the bottom boundary, though not by as much as the factor $\Gamma$ predicted for the vertical length scale ratio. There is therefore some doubt as to whether it is correct to have $d$ as the horizontal length scale in the boundary layers when $\Gamma$ is large. Further research is needed to elucidate this issue. The velocity ratio equation (67) correctly predicts that the velocity ratio is always close to unity, but is less reliable at predicting whether it is above or below unity. However, the higher density ratio runs do have $r_{u}>1$, as predicted by (67). We also evaluated the Nusselt number from the data, using $\displaystyle Nu=-\frac{d}{c_{p}}\frac{d\langle S\rangle_{h}}{dz}\Big{|}_{z=0}=-\frac{d}{\Gamma c_{p}}\frac{d\langle S\rangle_{h}}{dz}\Big{|}_{z=d},$ (90) again determining the gradients from our spline representation. When the run has been integrated long enough, initial transients in the numerical run are eliminated and these two estimates of the Nusselt number become close, and the average value is used in table 1. Because the flow is turbulent, the Nusselt number fluctuates continuously at about the 10% level, so a long time average is used. The finite length of the run means there is a small uncertainty due to the fluctuations not exactly cancelling, which we estimate as error bars in table 1. In table 1 we also give the value of the Nusselt number calculated by our theory. We used the Boussinesq runs, D1, D2 and D3 to determine the prefactors $C_{Nu}$ appropriate for each Prandtl number used. This gives $C_{Nu}=0.949$ for $Pr=10$, $C_{Nu}=0.785$ for $Pr=1$ and $C_{Nu}=0.869$ for $Pr=0.25$. Note that in table 1 this means that for the Boussinesq runs D1, D2, and D3, the $Nu$-theory entries are the same as the actual $Nu$ entries by construction. None of these prefactors is very far from the values suggested by Grossmann & Lohse (2000), who had $C_{Nu}=0.93$ at large $Pr$ and $C_{Nu}=0.76$ at small Prandtl number, suggesting that the differences in aspect ratio and geometry only make relatively small changes to the Nusselt number. For consistency, we use these same prefactors in all runs. Since our main interest is in the compressible cases, we have not explored why the prefactor for the $Pr=1$ case is slightly lower than the other values of $C_{Nu}$. We first consider the Nusselt number for the cases where the density ratio is only 2.71, runs A1, A2, A3, B1 and C1. We note that all the predicted values are not too far off the numerical values, though the predicted values are generally a little lower than the actual numerical values. Using the numerically calculated boundary layer ratios in formula (86) rather than the theoretically predicted ones gives the result $Nu$-nblr in table 1. These are only available after the simulation is run, but they are helpful for testing whether small errors are due to slightly inaccurate boundary layer ratios, or whether the formula (86) is inaccurate. For the density ratio 2.71 runs with $Pr\geq 1$, the boundary layer ratios were close to the predicted values, so not surprisingly, $Nu$-nblr are not significantly better than $Nu$-theory. We conclude that the under-prediction of the Nusselt number in these cases, which is less than about the 10% level, is due to the viscous dissipation not being completely in the boundary layers, as required by the theory. We believe that at higher $Ra$, where the dissipation progressively goes into the boundary layers, and using longer runs to average out the fluctuations, the small discrepancy will disappear. Using runs A1, A2 and A3, where only $Ra$ varies, we can test the $Ra^{1/4}$ power law predicted in (86). The least squares fit to a straight line in $\log(Nu)$ vs $\log(Ra)$ space has a slope of 0.257, rather close to the predicted slope. We now look at the larger $\Gamma$ cases, runs A4, A5 and B2, corresponding to the more compressible cases. In the case run A4, at density ratio 5, the predicted and numerical Nusselt numbers are reasonably close, but in the most extreme cases of density ratio 10 the predicted $Nu$ is only 61% of the numerical value for run B2, and in the case A5 predicted $Nu$ is 71% of the numerical $Nu$. Part of this discrepancy is down to the boundary layer ratios, which become very large at high $\Gamma$, and so small inaccuracy can affect the Nusselt number significantly. If we use the numerical boundary layer ratios in (86) rather than the theoretical ones for run B2, the predicted $Nu$-nblr rises to 3.12, but even this is only 76% of the numerical value, and A5 similarly improves but still is too low. We again conclude that in runs B2 and A5 the assumption that the dissipation occurs dominantly in the boundary layers is suspect (particularly near the top boundary) and that higher $Ra$ is needed before it becomes robustly valid. We now consider the Reynolds number formula, (87), though it is convenient to express this in terms of Peclet number $Pe=RePr$ . If the table 1 parameter values are inserted into (87) with the value of $C_{Re}$ quoted there, the values of the Peclet number are consistently a factor of about 5 too small compared with the numerical values of $U_{T}$ and $U_{B}$ in table 1. There are a number of reasons for this, but the two most important are (i) the power law dependence of the Peclet number with Rayleigh number in Boussinesq convection is slightly less than the predicted 0.5 (Grossmann & Lohse 2002), and (ii) our runs are for aspect ratio 2, whereas experiments, and the numerical simulations that simulate them (e.g. Silano et al. 2010), use aspect ratios of 1 or less. The prefactor in (87) is based on experiments at large $Ra$ which mostly used aspect ratios less than unity. The experiments of Qiu & Tong (2001), see also figure 1 of Grossmann & Lohse (2002), using water (with $Pr=5.5$) in a cylinder of aspect ratio unity found $Re=0.085Ra^{0.455}$. At the run A1 parameters this formula gives $Pe=251$ consistent with a prefactor of $C_{Re}=0.38$ in (87), much larger than the Grossmann & Lohse (2002) value. They found very similar $Re$ prefactors for both high and low $Pr$. If we adopt the same procedure as we did to get the Nusselt number prefactors, and normalise using the Boussinesq runs D1, D2 and D3 we obtain $C_{Re}=0.354$ for $Pr=10$, $C_{Re}=0.400$ at $Pr=1$ and $C_{Re}=0.421$ at $Pr=0.25$. Reassuringly, these are all quite close to the Qiu & Tong (2001) value of $C_{Re}=0.38$. We therefore use these three values of $C_{Re}$ at the appropriate Prandtl number in all our theory calculations. With these prefactors, the numerical results for $U_{T}$ and $U_{B}$ agree reasonably well with the predicted $Pe_{T}$-theory and $Pe_{B}$-theory results. The results have some scatter, which seems to reflect the scatter in our computed $r_{U}$. If we use the computed boundary layer ratios, rather than the asymptotically predicted ratios, there is less scatter in the comparison between computed and theoretical Peclet numbers, though the theoretical Peclet numbers are generally a few percent lower that the computed Peclet numbers. Given that the boundary layers are not very thin, these small discrepancies are not unexpected, and overall the predicted Reynolds numbers are in reasonable agreement with those of our §6 asymptotic theory. ## 8 Conclusions The scaling laws for heat flux and Reynolds number at high Rayleigh number convection have been derived from the energy balance and entropy balance equations derived in §3. These scaling laws are derived in terms of the Rayleigh number, the Prandtl number and the temperature ratio $\Gamma$ which measures the strength of the stratification. In the Boussinesq limit, $\Gamma\to 1$, they reduce to the scaling laws of Grossmann & Lohse (2000). The existence of the well-mixed entropy state, with the entropy changes being mainly confined to thin boundary layers, makes it possible to estimate the terms in the entropy balance equation, so allowing Nusselt number and Reynolds number relationships to be established. The cases treated are those where the viscous dissipation occurs in the boundary layers, the cases labelled as $I_{u}$ and $I_{l}$ by Grossmann & Lohse (2000), the subscripts referring to the high and low Prandtl number regimes, and the cases where the viscous dissipation is primarily in the bulk, the cases $II_{u}$ and $II_{l}$. A limitation of the theory is that both the entropy boundary layers do have to be thin for the theory to be valid. For the top boundary layer to be thin when the stratification is strong, the Rayleigh number has to be very large, which is numerically difficult, so the range of $\Gamma$ which can be tested both numerically and asymptotically is quite limited. This condition that the top boundary layer is thin is equivalent to the condition that the boundary layers are incompressible, so that a rather simple relationship holds between temperature and entropy within the boundary layers. The more difficult case where the boundary layers are compressible has not yet been solved in closed form, but it is likely to be significantly different from our solutions. A feature of this high Rayleigh number anelastic problem is that the top and bottom boundary layers have a different structure, so to determine the scaling laws, boundary layer ratios for the top and bottom boundary layers have to be established. The three key ratios are those for the boundary layer widths, the boundary layer entropy jumps and the horizontal velocities just outside the boundary layers. In §5 we proposed formulae based on a simple physical picture for these ratios. We have performed some numerical simulations to test these proposed boundary layer ratios, and within the constraints imposed by the numerics, namely not very high $Ra$, we find broad agreement between the theory and the numerics. Another important assumption for Grossmann-Lohse theory to be valid is the existence of a wind of turbulence. Our numerics suggest that this feature persists in our simulations. There is, however, still some uncertainty about whether the horizontal length scale of that wind, which controls the boundary layers, remains at the vertical length scale $d$ as the stratification $\Gamma$ increases, or whether it becomes smaller at the top boundary than the bottom boundary at large $\Gamma$. We have also tested the theoretically derived Nusselt number and Reynolds number relationships against the numerics, in the case where the viscous dissipation is mainly in the boundary layers, the only numerically accessible case. Using the prefactors determined in the Boussinesq case, which are the only free parameters in the theory, the Nusselt numbers obtained are in reasonable agreement with the theory, again noting the numerical limitations preventing accurate agreement. A problem was encountered when comparing with the theoretical Reynolds numbers, in that the theory using the original Grossmann-Lohse prefactors gave smaller $Re$ than did the numerics. However, the disagreement seems to be due more to issues with the Boussinesq problem rather than to its extension to the anelastic case, in particular to the difficulty of establishing a single prefactor over a huge range of $Ra$ and to dependence of $Re$ on the aspect ratio. When the prefactors were determined by normalizing on our Boussinesq runs, the issue was resolved. We have focussed here on the case of no-slip boundaries, as this seems the simplest case in which scaling laws can be derived from first principles without introducing arbitrary constants into the formulae. There are, however, many similar problems which could be addressed which are of great astrophysical interest: the case of stress-free boundaries is thought to be particularly relevant to stellar convection zones. Even within the context of our simplified no-slip problem, the case of compressible boundary layers would be of interest. We found it most convenient to consider fixed entropy boundary conditions, but other boundary conditions, such as fixed temperature or fixed flux are of interest too. Another issue that could be explored are the differences between temperature diffusion and entropy diffusion cases. In our particular problem, with incompressible boundary layers, the differences appear to be quite minor, but this is not necessarily the case if more challenging cases are addressed. Given the growing importance of the anelastic approximation in exploring a very wide range of exciting astrophysical problems, a firmer understanding of the fundamental behaviour of high Rayleigh number anelastic convection would be very valuable. ###### Acknowledgements. Acknowledgements This work was partially funded by the STFC grant ST/S00047X/1 held at the University of Leeds. The partial funding from the subvention of the Ministry of Science and Higher Education in Poland as a part of the statutory activity and the support of the National Science Centre of Poland (grant No. 2017/26/E/ST3/00554) is gratefully acknowledged. The computational work was performed on the ARC clusters, part of the high performance computing facilities at the University of Leeds, and on the COSMA Data Centre system at Durham University, operated on behalf of the STFC DiRAC HPC. ## Appendix A Form of the anelastic temperature perturbation Taking the horizontal average of the anelastic continuity equation (2.2), and using the $u_{z}=0$ boundary conditions gives $\langle u_{z}\rangle_{h}=0.$ (91) Using (1) and (13), the $z$-component of the anelastic equation of motion can be written ${\bar{\rho}}\frac{\partial u_{z}}{\partial t}+\nabla\cdot({\bar{\rho}}u_{z}{\bf u})+\nabla p=-g\rho+\mu\left(\nabla^{2}u_{z}+\frac{1}{3}\frac{\partial}{\partial z}\nabla\cdot{\bf u}\right).$ (92) Taking the horizontal average of (92) we see that, using (91), the viscous term vanishes to leave $\frac{\mathrm{d}}{\mathrm{d}z}\left\langle\bar{\rho}u_{z}^{2}\right\rangle_{h}+\frac{\mathrm{d}\left\langle p\right\rangle_{h}}{\mathrm{d}z}=-g\left\langle\rho\right\rangle_{h}=-g{\bar{\rho}}\left(\frac{\langle p\rangle_{h}}{\bar{p}}-\frac{\langle T\rangle_{h}}{\bar{T}}\right)$ (93) using ($None$a). In the bulk, entropy is well-mixed, so it is constant there so differentiating ($None$b) and using ($None$a), $R\frac{d}{dz}\left(\frac{\langle p\rangle_{h}}{\bar{p}}\right)=c_{p}\frac{d}{dz}\left(\frac{\langle T\rangle_{h}}{\bar{T}}\right)$ (94) in the bulk. Using the adiabatic reference state hydrostatic and perfect gas equations, with ($None$), this can be written $\frac{d\langle p\rangle_{h}}{dz}=c_{p}{\bar{\rho}}\frac{d\langle T\rangle_{h}}{dz}-\frac{g{\bar{\rho}}}{\bar{p}}\langle p\rangle_{h}+\frac{g{\bar{\rho}}}{\bar{T}}\langle T\rangle_{h}$ (95) and on substituting this into (93) we obtain $\frac{d{\langle T\rangle_{h}}}{dz}=-\frac{1}{c_{p}{\bar{\rho}}}\frac{d}{dz}\langle\bar{\rho}u_{z}^{2}\rangle_{h},$ (96) which is valid in the bulk. Integrating this across the bulk from $z=\delta^{th}_{B}$ to $z=d-\delta^{th}_{T}$, and assuming $u_{z}$ is negligible close to the boundaries, $\langle T_{bulk}(d-\delta^{th}_{T})\rangle_{h}-\langle T_{bulk}(\delta^{th}_{B})\rangle_{h}=\int_{\delta^{th}_{B}}^{d-\delta^{th}_{T}}\langle\bar{\rho}u_{z}^{2}\rangle_{h}\frac{d}{dz}\left(\frac{1}{c_{p}\bar{\rho}}\right)\,dz=\Delta T_{vel}>0,$ (97) since $\bar{\rho}$ is monotonic decreasing with $z$. This establishes that in the bulk the gradient $d{\langle T\rangle_{h}}/dz$ is positive on average, corresponding to a subadiabatic horizontally averaged temperature gradient. We denote this jump in $T$ across the bulk by $\Delta T_{vel}$ because it is physically connected to the pressure changes induced by the fluid velocity. A natural question is how large $\Delta T_{vel}$ is compared to the jumps in $\langle T\rangle_{h}$ across the boundary layers, $\Delta T_{B}$ and $\Delta T_{T}$. Formally they are both of same order of magnitude in the anelastic approximation, but $\Delta T_{vel}$ will be small if we are close to Boussinesq or if the Rayleigh number is small. Numerical evidence is sparse, but figure 4 from Verhoeven et al. (2015) suggests that for their parameters, $Ra=10^{6}$, $\rho_{B}/\rho_{T}=2.72$ and $Pr=0.7$, their $\Delta T_{vel}$ was small. ### A.1 Positivity of the temperature offsets We now consider the temperature offsets at the bottom and top boundaries, $\left\langle T\right\rangle_{h,T}$ and $\left\langle T\right\rangle_{h,B}$. Without numerical simulations, we cannot determine their magnitude, but we can show that they must both be positive, a useful check on future simulations. By examining the sum of the temperature jumps across the layer in Figure 1b we can see that $\left\langle T\right\rangle_{h,T}-\left\langle T\right\rangle_{h,B}+\Delta T_{B}+\Delta T_{T}=\Delta T_{vel}.$ (98) Using the incompressible boundary layer forms for the temperature jumps across the boundary layers, (25), and the formulae for the ratios of these jumps, $r_{s}=\frac{\Delta S_{T}}{\Delta S_{B}},\quad r_{T}=\frac{\Delta T_{T}}{\Delta T_{B}},\quad r_{s}=\Gamma r_{T},$ (99) equation (98) becomes $\left\langle T\right\rangle_{h,T}-\left\langle T\right\rangle_{h,B}+\frac{T_{B}\Delta S}{c_{p}}\frac{(1+r_{T})}{(1+\Gamma r_{T})}=\Delta T_{vel}.$ (100) A second equation for the temperature offsets can be derived from the boundary conditions. From ($None$a) and ($None$b) we can deduce $s=\frac{c_{p}T}{\bar{T}}-\frac{p}{\bar{\rho}\bar{T}}$ (101) At $z=0$ and $z=d$ this gives $\Delta S=\frac{c_{p}\left\langle T\right\rangle_{h,B}}{T_{B}}-\frac{\left\langle p\right\rangle_{h,B}}{\rho_{B}T_{B}},\quad 0=\frac{c_{p}\left\langle T\right\rangle_{h,T}}{T_{T}}-\frac{\left\langle p\right\rangle_{h,T}}{\rho_{T}T_{T}}.$ (102) We now use the mass conservation equation (28) to set the pressure perturbations on the top and bottom boundary equal, giving $\frac{T_{B}\Delta S}{c_{p}}=\left\langle T\right\rangle_{h,B}-\left\langle T\right\rangle_{h,T}\frac{\rho_{T}}{\rho_{B}}.$ (103) Equations (100) and (103) are two equations for the temperature offsets $\left\langle T\right\rangle_{h,B}$ and $\left\langle T\right\rangle_{h,T}$, and using $\rho_{B}/\rho_{T}=\Gamma^{m}$ the solutions are $\left\langle T\right\rangle_{h,B}=\frac{\Delta T_{vel}}{\Gamma^{m}-1}+\frac{T_{B}\Delta S}{c_{p}}\left\\{\frac{\Gamma^{m}(1+\Gamma r_{T})-(1+r_{T})}{(\Gamma^{m}-1)(1+\Gamma r_{T})}\right\\},$ (104) $\left\langle T\right\rangle_{h,T}=\frac{\Gamma^{m}\Delta T_{vel}}{\Gamma^{m}-1}+\frac{T_{B}\Delta S}{c_{p}}\left\\{\frac{(\Gamma-1)\Gamma^{m}r_{T}}{(\Gamma^{m}-1)(1+\Gamma r_{T})}\right\\}.$ (105) Since $\Gamma>1$ and $\Delta T_{vel}>0$ it follows that both quantities are positive whatever $\Delta T_{vel}$ is. It is not possible to decide which offset is larger without having more information about $\Delta T_{vel}$, but these results confirm that Figure 1b is a plausible sketch of the temperature perturbation, and will be helpful in testing numerical simulations. ## Appendix B The case when the dissipation in the bulk dominates the dissipation in the boundary layers Grossmann & Lohse (2000) point out that at low $Pr$ and large $Ra$ it is possible for the viscous dissipation in the bulk to be larger than the viscous dissipation in the boundary layers. When this occurs, our arguments about the boundary layer ratios in §5 and the scaling laws in §6 need revising. We now consider this scenario. ### B.1 The boundary layer ratios When the viscous dissipation is mainly in the bulk, equations (57-63) still hold, but the argument for equation (67) breaks down because the entropy flux is no longer approximately constant in the bulk because viscous dissipation in the bulk is no longer negligible. We can however use the energy flux equation (30) because when the viscous dissipation is mainly in the bulk, the work done by buoyancy must balance the viscous dissipation in the bulk, since now the viscous dissipation in the boundary layers is negligible. So $\displaystyle\frac{g}{c_{p}}\int_{bulk}\left\langle\bar{\rho}u_{z}s\right\rangle_{h}\mathrm{d}z=\mu\int_{bulk}\left\langle q\right\rangle\mathrm{d}z,$ (106) and since thermal diffusion and the last two terms in (30) are negligible in the bulk when the boundary layers are thin, it follows that $\langle{\bar{\rho}}{\bar{T}}u_{z}s\rangle_{h}$ will be approximately the same just outside the two boundary layers at $z=\delta_{B}^{\nu}$ and $z=d-\delta_{T}^{\nu}$, so $\displaystyle\rho_{B}T_{B}\langle u_{z}s\rangle_{h}|_{z=\delta_{B}^{\nu}}\approx\rho_{T}T_{T}\langle u_{z}s\rangle_{h}|_{z=d-\delta_{T}^{\nu}}.$ (107) Note this is different from the case where the dissipation was mainly in the boundary layers, when $\langle{\bar{\rho}}u_{z}s\rangle_{h}$ is approximately constant. As we did in §5,we horizontally average the dot product of $\bf u$ and (1), and apply it just outside the boundary layers, at $z=\delta_{B}^{\nu}$ and $z=d-\delta_{T}^{\nu}$. Here we are justified in neglecting the pressure term as we did in §5, and we also neglect the viscous term. This is not obvious when most of the viscous dissipation is in the bulk, but following Grossmann & Lohse (2000), we envisage a turbulent cascade, where the dissipation at larger scales is dominated by the inertial term. We therefore adopt $\frac{1}{2}\frac{\partial}{\partial z}\left(\bar{\rho}\left\langle u_{z}u^{2}\right\rangle_{h}\right)\approx-\frac{\partial}{\partial z}\langle u_{z}p\rangle_{h}+\frac{g}{c_{p}}\left\langle{\bar{\rho}}u_{z}s\right\rangle_{h}\approx\frac{g}{c_{p}}\left\langle{\bar{\rho}}u_{z}s\right\rangle_{h}$ (108) at $z=\delta_{B}^{\nu}$ and $z=d-\delta_{T}^{\nu}$. Then since in the bulk we expect all velocity components to be of similar magnitude in the bulk, using (107) $\displaystyle\rho_{B}T_{B}\frac{U_{B}^{3}}{H_{B}}\approx\rho_{T}T_{T}\frac{U_{T}^{3}}{H_{T}}\Rightarrow r_{u}\sim\Gamma^{\frac{m}{3}},$ (109) where the pressure scale heights are defined in (66). This result differs from (67), where the dissipation is in the boundary layers, so that now the horizontal velocity at the top is expected to be considerably faster than the velocity at the bottom, whereas (69) predicts only a weak dependence on $\Gamma$. ### B.2 The scaling laws when dissipation is in the bulk We still expect thin boundary layers even when the dissipation is mainly in the bulk, so (71), $\frac{F^{super}\Delta{\bar{T}}}{T_{B}T_{T}}\sim\int_{0}^{d}\frac{\mu}{\bar{T}}\left\langle q\right\rangle_{h}\,dz,$ (110) still applies, but unlike the boundary layer dissipation case, we do not know how the dissipation is distributed over the interior. We therefore assume that the dissipation in the interior can be written as $\langle q\rangle_{h}\sim U_{H}^{3}/2H$ where $U_{H}(z)$ is the horizontally averaged horizontal velocity and $H$ is the local pressure scale height. We don’t know how $U_{H}(z)$ is distributed in $z$, but we argued in §B1 above that $\rho U_{H}^{3}$ is approximately the same at the edge of both boundary layers, so a reasonable assumption for the purposes of estimation is that $\rho U_{H}(z)^{3}\sim\ \textrm{constant}\ \approx\rho_{B}U_{B}^{3}\approx\rho_{T}U_{T}^{3}.$ (111) The form of $U_{H}(z)$ from our numerical results suggests this might overestimate the dissipation integrated over the whole layer, but nevertheless we adopt (111) for the rest of this section. Equation (110) then becomes $\frac{F^{super}\Delta{\bar{T}}}{T_{B}T_{T}}\sim\int_{0}^{d}\frac{\bar{\rho}U_{H}^{3}}{2\bar{T}H}\,dz=\int_{0}^{d}\frac{\rho_{B}U_{B}^{3}(m+1)\Delta{\bar{T}}}{2d{\bar{T}}^{2}}\,dz=\frac{\rho_{B}U_{B}^{3}(m+1)(\Gamma-1)}{2T_{b}}.$ (112) From (15) $F^{super}=NukT_{B}/d$, and writing $U_{B}$ in terms of the bottom Reynolds number using (76) $\frac{Nuk\Delta{\bar{T}}}{dT_{T}}\sim\frac{\mu^{3}Re_{B}^{3}(m+1)(\Gamma-1)}{2T_{b}\rho_{B}^{2}d^{2}}.$ (113) Combining (12) and (16), we can write the Rayleigh number as $Ra=\frac{c_{p}^{2}\Delta{\bar{T}}d^{2}\rho_{B}^{2}\Gamma\ln\Gamma}{\mu k(\Gamma-1)},$ (114) and combining this with (113) and using the definition of the Prandtl number (51) we obtain $\frac{NuRa}{Pr^{2}}\sim\frac{(m+1)\ln\Gamma}{2}Re_{B}^{3},$ (115) which is the entropy balance equation in the case where the dissipation is mainly in the bulk rather than the boundary layers. We now use the same boundary layer balance equation as before, but since we expect bulk dissipation only to dominate at low $Pr$, we only use (55) for the boundary layer ratio, so (81) becomes $\left(Re_{B}Pr\right)^{1/2}=\frac{Nu(\Gamma-1)(1+r_{s})}{\Gamma\ln\Gamma}.$ (116) Combining (115) and (116) we get the Nusselt number in terms of the Rayleigh number in this case, $Nu\sim Ra^{1/5}Pr^{1/5}\left(\frac{2}{m+1}\right)^{1/5}\frac{\Gamma^{6/5}\ln\Gamma}{(\Gamma-1)^{6/5}}(1+r_{s})^{-6/5}.$ (117) ## References * Braginsky and Roberts (1995) BRAGINSKY, S. I. & ROBERTS, P. H. 1995 Equations governing convection in Earth’s core and the geodynamo, Geophys. Astrophys. Fluid Dyn. 79, 1–97. * Browning et al. (2006) BROWNING, M. K., MIESCH, M. S., BRUN, A. S. & TOOMRE, J. 2006 Dynamo action in the Solar convection zone and tachocline: pumping and organization of toroidal fields. Astrophys. J. 648, L157–L160. * Brun and Toomre (2002) BRUN, A. S. & TOOMRE, J. 2002 Turbulent convection under the influence of rotation: sustaining a strong differential rotation. Astrophys. J. 570, 865–885. * Chong et al (2018) CHONG, K. L., WAGNER, S., KACZOROWSKI, M., SHISKINA, O. & XIA, K-Q. 2018 Effect of Prandtl number on heat transport enhancement in Rayleigh-Bénard convection under geometrical confinement. Phys. Rev. Fluids 3, 013501\. * Curbelo et al. (2019) CURBELO, J., DUARTE L., ALBOUSSIÈRE, T., DUBUFFET, F., LABROSSE, S. & RICARD, Y. 2019 Numerical solutions of compressible convection with an infinite Prandtl number: comparison of the anelastic and anelastic liquid models with the exact equations. J. Fluid Mech. 3, 646–687. * Deardorff (1970) DEARDORFF, J. W. 1970 Convective velocity and temperature scales for the unstable planetary boundary layer and for Rayleigh convection. J. Atmos. Sci. 27, 1211–1213, * Glatzmaier and Roberts (1995) GLATZMAIER, G. A. & ROBERTS, P. H. 1995 A three-dimensional convective dynamo solution with rotating and finitely conducting inner core and mantle. Phys. Earth Planet. Inter. 91, 63–75. * Gough (1969) GOUGH, D. O. 1969 The anelastic approximation for thermal convection. J. Atmos. Sci. 26, 448–456. * Grossmann and Lohse (2000) GROSSMANN, S. & LOHSE, D. 2000 Scaling in thermal convection: a unifying theory. J. Fluid Mech. 407, 27–56, * Grossmann and Lohse (2002) GROSSMANN, S. & LOHSE, D. 2002 Prandtl and Rayleigh number dependence of the Reynolds number in turbulent thermal convection. Phys. Rev. E 66 (1), 016305\. * Jones and Kuzanyan (2009) JONES, C. A. & KUZANYAN, K. 2009 Compressible convection in the deep atmospheres of giant planets. Icarus 204, 227–238. * Jones et al. (2011) JONES, C. A., BORONSKI, P., BRUN, A. S., GLATZMAIER, G. A., GASTINE, T., MIESCH, M. S. & WICHT, J. 2011 Anelastic convection-driven dynamo benchmarks. Icarus 216, 120–135. * Kessar et al. (2019) KESSAR, M., HUGHES, D.W., KERSALÉ, E., MIZERSKI, K. A. & TOBIAS, S. M. 2019 Scale selection in the stratified convection of the solar photosphere. Astrophys. J. 874, 103–117. * Kolmogorov (1941a) KOLMOGOROV, A. N. 1941a The local structure of turbulence in incompressible viscous fluid for very large Reynolds numbers. Proceedings of the USSR Academy of Sciences 30, 299–303. * Kolmogorov (1941b) KOLMOGOROV, A. N., 1941b Dissipation of Energy in the Locally Isotropic Turbulence. Proceedings of the USSR Academy of Sciences 32, 16–18. * Korre et al. (2017) KORRE, L., BRUMMELL, N. & GARAUD, P. 2017 Weakly non-Boussinesq convection in a gaseous spherical shell. Phys. Rev. E 96 (3), 033104\. * Lantz and Fan (1999) LANTZ, S. R. & FAN, Y. 1999 Anelastic magnetohydrodynamic equations for modelling solar and stellar convection zones. Astroph. J. Supp. Series 121, 247–264. * Menaut et al. (2019) MENAUT, R., CORRE, Y., HUGUET, L., Le REUN, T., ALBOUSSIÈRE, T., BERGMAN, M., DEGUEN, R., LABROSSE, S. & MOULIN, M. 2019 Experimental study of convection in the compressible regime. Phys. Rev. Fluids 4, 033502\. * Miesch et al. (2000) MIESCH, M. S., ELIOTT, J. R., TOOMRE, J., CLUNE, T. L., GLATZMAIER, G. A. & GILMAN, P. A. 2000 Three-dimensional Spherical Simulations of Solar Convection. I. Differential Rotation and Pattern Evolution Achieved with Laminar and Turbulent States. Astrophys. J. 532, 593–615. * Ogura and Phillips (1962) OGURA, Y. & PHILLIPS, N. A. 2000 Scale analysis of deep and shallow convection in the atmosphere. J. Atmos. Sci. 19, 173–179. * Qiu and Tong (2001) QIU, X.-L. & TONG, P. 2001 Large-scale velocity structures in turbulent thermal convection. Phys. Rev. E 64 (3), 036304\. * Siggia (1994) SIGGIA. E. D. 1994 High Rayleigh number convection. Ann. Rev. Fluid Mech. 26, 137–168. * Silano et al. (2010) SILANO, G., SREENIVASAN, K. R. & VERZICCO, R. 2010 Numerical simulations of Rayleigh-Bénard convection for Prandtl numbers between $10^{-1}$ and $10^{4}$ and Rayleigh numbers between $10^{5}$ and $10^{9}$. J. Fluid Mech. 662, 409–446. * Spiegel and Veronis (1960) SPIEGEL, E. A. & VERONIS, G. 1960 On the Boussinesq approximation for a compressible fluid. Astrophys. J., 131, 442–447. * Stevens et al. (2013) STEVENS, R. J. A. M., VAN DER POEL, E. P., GROSSMANN, S. & LOHSE, D. (2013) The unifying theory of scaling in thermal convection: updated prefactors. J. Fluid Mech. 730, 295–308. * Toomre et al. (1976) TOOMRE, J., ZAHN, J.-P., LATOUR, J. & SPIEGEL, E. A. 1976 Stellar convection theory. II - Single-mode study of the second convection zone in an A-type star. Astrophys. J. 207, 545–563. * Verhoeven et al. (2015) VERHOEVEN, J., WIESEHÖFER, T. & STELLMACH, S. 2015 Anelastic versus fully compressible turbulent Rayleigh-Bénard convection. Astrophys. J. 805, 62–75.
# An Empirical Study of Cross-Lingual Transferability in Generative Dialogue State Tracker Yen-Ting Lin, Yun-Nung Chen ###### Abstract There has been a rapid development in data-driven task-oriented dialogue systems with the benefit of large-scale datasets. However, the progress of dialogue systems in low-resource languages lags far behind due to the lack of high-quality data. To advance the cross-lingual technology in building dialog systems, DSTC9 introduces the task of cross-lingual dialog state tracking, where we test the DST module in a low-resource language given the rich- resource training dataset. This paper studies the transferability of a cross-lingual generative dialogue state tracking system using a multilingual pre-trained seq2seq model. We experiment under different settings, including joint-training or pre-training on cross-lingual and cross-ontology datasets. We also find out the low cross- lingual transferability of our approaches and provides investigation and discussion. ## Introduction Dialogue state tracking is one of the essential building blocks in the task- oriented dialogues system. With the active research breakthrough in the data- driven task-oriented dialogue technology and the popularity of personal assistants in the market, the need for task-oriented dialogue systems capable of doing similar services in low-resource languages is expanding. However, building a new dataset for task-oriented dialogue systems for low-resource language is even more laborious and costly. It would be desirable to use existing data in a high-resource language to train models in low-resource languages. Therefore, if cross-lingual transfer learning can be applied effectively and efficiently on dialogue state tracking, the development of task-oriented dialogue systems on low-resource languages can be accelerated. The Ninth Dialog System Technology Challenge (DSTC9) Track2 (Gunasekara et al. 2020) proposed a cross-lingual multi-domain dialogue state tracking task. The main goal is to build a cross-lingual dialogue state tracker with a rich resource language training set and a small development set in the low resource language. The organizers adopt MultiWOZ 2.1 (Eric et al. 2019) and CrossWOZ (Zhu et al. 2020) as the dataset and provide the automatic translation of these two datasets for development. In this paper’s settings, our task is to build a cross-lingual dialogue state tracker in the settings of CrossWOZ-en, the English translation of CrossWOZ. In the following, we will refer cross- lingual datasets to datasets in different languages, such as MultiWOZ-zh and CrossWOZ-en, and cross-ontology datasets to datasets with different ontologies, such as MultiWOZ-en and CrossWOZ-en. The cross-lingual transfer learning claims to transfer knowledge across different languages. However, in our experiments, we experience tremendous impediments in joint training on cross-lingual or even cross-ontology datasets. To the best of our knowledge, all previous cross-lingual dialogue state trackers approach DST as a classification problem (Mrkšić et al. 2017)(Liu et al. 2019), which does not guarantee the success of transferability on our generative dialogue state tracker. The contributions of this paper are three-fold: * • This paper explores the cross-lingual generative dialogue state tracking system’s transferability. * • This paper compares joint training and pre-train then finetune method with cross-lingual and cross-ontology datasets. * • This paper analyzes and open discussion on colossal performance drop when training with cross-lingual or cross-ontology datasets. ## Problem Formulation In this paper, we study the cross-lingual multi-domain dialogue state tracking task. Here we define the multi-domain dialogue state tracking problem and introduce the cross-lingual DST datasets. ### Multi-domain Dialogue State Tracking The dialogue state in the multi-domain dialogue state tracking is a set of (domain, slot name, value) triplets, where the domain indicates the service that the user is requesting, slot name represents the goal from the user, and value is the explicit constraint of the goal. For dialogue states not mentioned in the dialogue context, we assign a null value, $\emptyset$, to the corresponding values. For example, (Hotel, type, luxury) summarizes one of the user’s constraints of booking a luxury hotel, and (Attraction, fee, 20 yuan or less) means the user wants to find a tourist attraction with a ticket price equal to or lower than 20 dollars. An example is presented in Figure1. Our task is to predict the dialogue state at the $t^{th}$ turn, $\mathcal{B}_{t}=\\{(\mathcal{D}^{i},\mathcal{S}^{i},\mathcal{V}^{i})\,|\,1\leq i\leq I\\}$ where $I$ is the number of states to be tracked, given the historical dialogue context until now, defined as $\mathcal{C}_{t}=\\{\mathcal{U}_{1},\mathcal{R}_{1},\mathcal{U}_{2},\mathcal{R}_{2},\dots,\mathcal{R}_{t-1},\mathcal{U}_{t}\\}$ where $\mathcal{U}_{i}$ and $\mathcal{R}_{i}$ is the user utterance and system response, respectively, at the $i^{th}$ turn. Figure 1: Illustration of dialogue state tracking. The dialogue is sampled from CrossWOZ-en. ### Dataset MultiWOZ is the task-oriented dataset often used as the benchmark dataset for task-oriented dialogue system tasks, including dialogue state tracking, dialogue policy optimization, and NLG. MultiWOZ 2.1 is a cleaner version of the previous counterpart with more than 30% updates in dialogue state annotations. CrossWOZ is a Chinese multi-domain task-oriented dataset with more than 6,000 dialogues, five domains, and 72 slots. Both of the above datasets collects human-to-human dialogues in Wizard-of-Oz settings. Table 1 lists the details of the dataset. In DSTC9 Track 2, the organizers translate MultiWOZ and CrossWOZ into Chinese and English, respectively, and we refer the translated version of MultiWOZ and CrossWOZ as MultiWOZ-zh and CrossWOZ-en, respectively. The public and private test of CrossWOZ-en in DSTC9 has 250 dialogues, but only the public test set has annotations. Therefore, we use the public one as the test set in our experiments. Metric | MultiWOZ | CrossWOZ ---|---|--- Language | English | Chinese (Simplified) # Dialogues | 8,438 | 5,012 Total # turns | 113,556 | 84,692 # Domains | 7 | 5 # Slots | 24 | 72 # Values | 4,510 | 7,871 Table 1: Statistics of MultiWOZ and CrossWOZ. Note that the translated version of these two datasets have the same metrics ## Related Work ### Dialogue State Tracker Traditionally, dialogue state tracking depends on fixed vocabulary approaches where retrieval-based models ranks slot candidates from a given slot ontology. (Ramadan, Budzianowski, and Gašić 2018)(Lee, Lee, and Kim 2019)(Shan et al. 2020) However, recent research efforts in DST have moved towards generation- based approaches where the models generate slot value given the dialogue history. (Wu et al. 2019) proposed a generative multi-domain DST model with a copy mechanism which ensures the capability to generate unseen slot values. (Kim et al. 2019) introduced a selectively overwriting mechanism, a memory- based approach to increase efficiency in training and inference. (Le, Socher, and Hoi 2020) adopted a non-autoregressive architecture to model potential dependencies among (domain, slot) pairs and reduce real-time DST latency significantly. (Hosseini-Asl et al. 2020) took advantage of the powerful generation ability of large-scale auto-regressive language model and formulated the DST problem as a casual language modeling problem. ### Multilingual Transfer Learning in Task-oriented Dialogue (Schuster et al. 2019) introduced a multilingual multi-domain NLU dataset. (Mrkšić et al. 2017) annotated two additional languages to WOZ 2.0 (Mrkšic et al. 2017) and (Liu et al. 2019) proposed a mixed-language training for cross- lingual NLU and DST tasks. Noted that all previous multilingual DST methods modeled the dialogue state tracking task as a classification problem. (Mrkšić et al. 2017)(Liu et al. 2019) ## Methods This paper considers the multi-domain dialogue state track-ing as a sequence generation task by adopting a sequence-to-sequence framework. ### Architecture Following (Liu et al. 2020), we use the sequence-to-sequence Transformer architecture (Vaswani et al. 2017) with 12 layers in each encoder and decoder. We denote seq2seq as our model in the following. ### DST as Sequence Generation The input sequence is composed of the concatenation of dialogue context $\mathbf{x^{t}}=\\{\mathcal{U}_{1};\mathcal{R}_{1};\mathcal{U}_{2};\mathcal{R}_{2};\dots;\mathcal{R}_{t-1};\mathcal{U}_{t}\\}$ where ; denote the concatenation of texts. For the target dialogue state, we only consider the slots where the values are non-empty. The target sequence is consist of the concatenation of the (domain, slot, value) triplets with a non-empty value, $\mathbf{y^{t}}=\\{\mathcal{D}^{i};\mathcal{S}^{i};\mathcal{V}^{i}|1\leq i\leq I\wedge\mathcal{S}^{i}\neq\emptyset\\}$. $\mathbf{\hat{y}^{t}}=seq2seq(\mathbf{x^{t}})$ We fix the order of the (domain, slot name, value) triplets for consistency. The training objective is to minimize the cross-entropy loss between the ground truth sequence $\mathbf{y^{t}}$ and the predicted sequence $\mathbf{\hat{y}^{t}}$. ### Post-processing The predicted sequence $\mathbf{\hat{y}^{t}}$ is then parsed by heuristic rules to construct $\hat{\mathcal{B}_{t}}=\\{\mathcal{D}^{i};\mathcal{S}^{i};\hat{\mathcal{V}}^{i}|1\leq i\leq I\\}$. By utilizing the possible values of slots in the ontology, for predicted slot values $\hat{\mathcal{V}}$ that do not appears in the ontology, we choose the one with the best match to our predicted value. 111This is implemented by difflib.get_close_matches in Python ## Experiments In the following section, we describe evaluation metrics, experiment setting and introduce experimental results. ### Evaluation Metrics We use joint goal accuracy and slot F1 as our metrics to evaluate our dialogue state tracking system. * • Joint Goal Accuracy: The proportion of dialogue turns where predicted dialogue states match entirely to the ground truth dialogue states. * • Slot F1: The macro-averaged F1 score for all slots in each turn. ### Experiments Settings We want to examine how different settings affect the performance of the target low-resource dataset: CrossWOZ-en.222In our experimental circumstance, English is the low-resource language since the original language of CrossWOZ is Chinese. We will conduct our experiments in the settings below. * • Direct Fine-tuning * • Cross-Lingual Training (CLT) * • Cross-Ontology Training (COT) * • Cross-Lingual Cross-Ontology Training (CL/COT) * • Cross-Lingual Pre-Training (CLPT) * • Cross-Ontology Pre-Training (COPT) * • Cross-Lingual Cross-Ontology Pre-Training (CL/COPT) Table 2 and 3 show the datasets for training and pre-training in different settings. For experiments with pre-training, all models are pre-trained on the pre-training dataset and then fine-tuned on CrossWOZ-en. The baseline model provided by DSTC9 is SUMBT (Lee, Lee, and Kim 2019), the ontology-based model trained on CrossWOZ-en. ### Multilingual Denoising Pre-training All of our models initialize from mBART25. (Liu et al. 2020) mBART25 is trained with denoising auto-encoding task on mono-lingual data in 25 languages, including English and Simplified Chinese. (Liu et al. 2020) shows pre-training of denoising autoencoding on multiple languages improves the performance on low resource machine translation. We hope using mBART25 as initial weights would improve the cross-lingual transferability. ### Implementation Details In all experiments, the models are optimized with AdamW (Loshchilov and Hutter 2017) with learning rate set to $1e^{-4}$ for 4 epochs. The best model is selected from the validation loss and is used for testing. During training, the decoder part of our model is trained in the teacher forcing fashion (Williams and Zipser 1989). Greedy decoding (Vinyals and Le 2015) is applied when inference. Following mBART (Liu et al. 2020), we use sentencespiece tokenizer. For GPU memory constraints, source sequences longer than 512 tokens are truncated at the front and target sequences longer than 256 tokens are truncated at the back. The models are implemented in Transformers (Wolf et al. 2019), PyTorch (Paszke et al. 2019) and PyTorch Lightning (Falcon 2019). ## Results and Discussion The results for all experiment settings are shown in Table 2 and 3. Experiment | Training Data | JGA | SF1 ---|---|---|--- MultiWOZ | CrossWOZ en | zh | en | zh Baseline | | | ✓ | | 7.41 | 55.27* Direct Fine-tuning | | | ✓ | | 16.82 | 66.35 CL/COT | ✓ | ✓ | ✓ | ✓ | 4.10 | 26.50 COT | ✓ | | ✓ | | 0.95 | 19.60 CLT | | | ✓ | ✓ | 0.53 | 13.45 Table 2: Experimental results on CrossWOZ-en with different training data (%). *: This slot f1 is averaged over both the public and private test dialogues. JGA: Joint Goal Accuracy. SF1: Slot F1. Experiment | Pre-training Data | JGA | SF1 ---|---|---|--- MultiWOZ | CrossWOZ en | zh | en | zh Direct Fine-tuning | | | | | 16.82 | 66.35 CL/COPT | ✓ | ✓ | | | 5.94 | 38.36 COPT | ✓ | | | | 2.52 | 27.01 CLPT | | | | ✓ | 0.11 | 15.01 Table 3: Experimental results on CrossWOZ-en with pre-training (%). ### Additional Training Data Cause Degeneration Direct Fine-tuning significantly outperforms other settings, including the official baseline. We assume English and Chinese data with the same ontology to train the mBART would bridge the gap between the two languages and increase the performance. However, in Cross-Lingual Training, training on English and Chinese version of CrossWOZ leads to catastrophic performance on CrossWOZ-en. In the Cross-Ontology Training where combine two data in the same language. However, with different ontologies, the performance marginally increases from Cross-Lingual Training, which shows more extensive mono-lingual data with the unmatched domain, slots, and ontology confuses the model during inference. In the Cross-Lingual Cross-Ontology Training, we collect all four datasets for training, and the performance is still far from Direct Fine-tuning. In conclusion, additional data deteriorate the performance on CrossWOZ-en even whether the language or ontology matches or not. ### Does ”First Pre-training, then fine-tuning” Help? We hypothesize that training with additional data causes performance degeneration, and therefore one possible improvement could be first pre- training the model on cross-lingual / cross-ontology data and then fine-tuning on the target dataset CrossWOZ-en. Table 3 shows the results. By comparing COPT to COT and CL/COPT to CL/COP, the relative performance gain by over 37% with regards to slot F1. ”Pre-training, fine-tuning” framework may partially alleviate the problem of catastrophic performance drop in joint training. ### Domain Performance Difference across Experiment Settings? This section further investigates the cause of the performance decrease by comparing the slot F1 of different models across five domains in Figure 2. Generally speaking, in attraction, restaurant, and hotel domains, ”pre-train then fine-tune” methods beat their ”joint training” counterparts by an observable margin. By contrast, in metro and taxi domains, despite poor performance among all, ”joint training” settings beat their”pre-train then fine-tune” counterparts. The only two trackable slots in the metro and taxi domain, ”from” and ”to,” usually take the address or name of buildings, are highly non-transferable across datasets. We conjecture that pretraining on cross-lingual or cross- ontology datasets does not help or even hurt those non-transferable slots. Figure 2: Slot F1 across 5 domains in CrossWOZ-en in different settings. ## Conclusion In this paper, we build a cross-lingual multi-domain generative dialogue state tracker with multilingual seq2seq to test on CrossWOZ-en and investigate our tracker’s transferability under different training settings. We find that jointly trained the dialogue state tracker on cross-lingual or cross-ontology data degenerates the performance. Pre-training on cross-lingual or cross- ontology data, then fine-tuning framework may alleviate the problem, and we find empirically evidence on relative improvement in slot F1. A finding from the domain performance shift is that performance on some non-transferable slots, such as name, from, to, may be limited by the previous pretraining approach. A future research direction would investigate why such a significant performance declines in joint training and tries to bridge it. ## References * Eric et al. (2019) Eric, M.; Goel, R.; Paul, S.; Kumar, A.; Sethi, A.; Ku, P.; Goyal, A. K.; Agarwal, S.; Gao, S.; and Hakkani-Tur, D. 2019. MultiWOZ 2.1: A Consolidated Multi-Domain Dialogue Dataset with State Corrections and State Tracking Baselines. * Falcon (2019) Falcon, W. 2019. PyTorch Lightning. _GitHub. Note: https://github.com/PyTorchLightning/pytorch-lightning_ 3\. * Gunasekara et al. (2020) Gunasekara, C.; Kim, S.; D’Haro, L. F.; Rastogi, A.; Chen, Y.-N.; Eric, M.; Hedayatnia, B.; Gopalakrishnan, K.; Liu, Y.; Huang, C.-W.; Hakkani-Tür, D.; Li, J.; Zhu, Q.; Luo, L.; Liden, L.; Huang, K.; Shayandeh, S.; Liang, R.; Peng, B.; Zhang, Z.; Shukla, S.; Huang, M.; Gao, J.; Mehri, S.; Feng, Y.; Gordon, C.; Alavi, S. H.; Traum, D.; Eskenazi, M.; Beirami, A.; Eunjoon; Cho; Crook, P. A.; De, A.; Geramifard, A.; Kottur, S.; Moon, S.; Poddar, S.; and Subba, R. 2020. Overview of the Ninth Dialog System Technology Challenge: DSTC9 URL https://arxiv.org/abs/2011.06486. * Hosseini-Asl et al. (2020) Hosseini-Asl, E.; McCann, B.; Wu, C.-S.; Yavuz, S.; and Socher, R. 2020. A Simple Language Model for Task-Oriented Dialogue URL http://arxiv.org/abs/2005.00796. * Kim et al. (2019) Kim, S.; Yang, S.; Kim, G.; and Lee, S.-W. 2019. Efficient Dialogue State Tracking by Selectively Overwriting Memory. _arXiv_ URL http://arxiv.org/abs/1911.03906. * Le, Socher, and Hoi (2020) Le, H.; Socher, R.; and Hoi, S. C. H. 2020. Non-Autoregressive Dialog State Tracking 1–21. URL http://arxiv.org/abs/2002.08024. * Lee, Lee, and Kim (2019) Lee, H.; Lee, J.; and Kim, T.-Y. 2019. SUMBT: Slot-Utterance Matching for Universal and Scalable Belief Tracking 5478–5483. doi:10.18653/v1/p19-1546. * Liu et al. (2020) Liu, Y.; Gu, J.; Goyal, N.; Li, X.; Edunov, S.; Ghazvininejad, M.; Lewis, M.; and Zettlemoyer, L. 2020. Multilingual Denoising Pre-training for Neural Machine Translation URL https://arxiv.org/abs/2001.08210. * Liu et al. (2019) Liu, Z.; Winata, G. I.; Lin, Z.; Xu, P.; and Fung, P. 2019. Attention-Informed Mixed-Language Training for Zero-shot Cross-lingual Task-oriented Dialogue Systems. _arXiv_ URL http://arxiv.org/abs/1911.09273. * Loshchilov and Hutter (2017) Loshchilov, I.; and Hutter, F. 2017. Decoupled Weight Decay Regularization URL http://arxiv.org/abs/1711.05101. * Mrkšic et al. (2017) Mrkšic, N.; Séaghdha, D.; Wen, T. H.; Thomson, B.; and Young, S. 2017\. Neural belief tracker: Data-driven dialogue state tracking. _ACL 2017 - 55th Annual Meeting of the Association for Computational Linguistics, Proceedings of the Conference (Long Papers)_ 1: 1777–1788. doi:10.18653/v1/P17-1163. * Mrkšić et al. (2017) Mrkšić, N.; Vulić, I.; Séaghdha, D. Ó.; Leviant, I.; Reichart, R.; Gašić, M.; Korhonen, A.; and Young, S. 2017. Semantic Specialisation of Distributional Word Vector Spaces using Monolingual and Cross-Lingual Constraints. _arXiv_ URL http://arxiv.org/abs/1706.00374. * Paszke et al. (2019) Paszke, A.; Gross, S.; Massa, F.; Lerer, A.; Bradbury, J.; Chanan, G.; Killeen, T.; Lin, Z.; Gimelshein, N.; Antiga, L.; Desmaison, A.; Kopf, A.; Yang, E.; DeVito, Z.; Raison, M.; Tejani, A.; Chilamkurthy, S.; Steiner, B.; Fang, L.; Bai, J.; and Chintala, S. 2019. PyTorch: An Imperative Style, High-Performance Deep Learning Library. In Wallach, H.; Larochelle, H.; Beygelzimer, A.; d'Alché-Buc, F.; Fox, E.; and Garnett, R., eds., _Advances in Neural Information Processing Systems 32_ , 8024–8035. Curran Associates, Inc. URL http://papers.neurips.cc/paper/9015-pytorch-an-imperative-style-high-performance-deep-learning-library.pdf. * Ramadan, Budzianowski, and Gašić (2018) Ramadan, O.; Budzianowski, P.; and Gašić, M. 2018. Large-Scale Multi-Domain Belief Tracking with Knowledge Sharing. In _Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers)_ , 432–437. Melbourne, Australia: Association for Computational Linguistics. doi:10.18653/v1/P18-2069. URL https://www.aclweb.org/anthology/P18-2069. * Schuster et al. (2019) Schuster, S.; Gupta, S.; Shah, R.; and Lewis, M. 2019. Cross-lingual Transfer Learning for Multilingual Task Oriented Dialog. In _Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers)_ , 3795–3805. Minneapolis, Minnesota: Association for Computational Linguistics. doi:10.18653/v1/N19-1380. URL https://www.aclweb.org/anthology/N19-1380. * Shan et al. (2020) Shan, Y.; Li, Z.; Zhang, J.; Meng, F.; Feng, Y.; Niu, C.; and Zhou, J. 2020. A Contextual Hierarchical Attention Network with Adaptive Objective for Dialogue State Tracking 6322–6333. URL http://arxiv.org/abs/2006.01554. * Vaswani et al. (2017) Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, L.; and Polosukhin, I. 2017. Attention Is All You Need. * Vinyals and Le (2015) Vinyals, O.; and Le, Q. 2015. A neural conversational model. _arXiv preprint arXiv:1506.05869_ . * Williams and Zipser (1989) Williams, R. J.; and Zipser, D. 1989. A learning algorithm for continually running fully recurrent neural networks. _Neural computation_ 1(2): 270–280. * Wolf et al. (2019) Wolf, T.; Debut, L.; Sanh, V.; Chaumond, J.; Delangue, C.; Moi, A.; Cistac, P.; Rault, T.; Louf, R.; Funtowicz, M.; Davison, J.; Shleifer, S.; von Platen, P.; Ma, C.; Jernite, Y.; Plu, J.; Xu, C.; Scao, T. L.; Gugger, S.; Drame, M.; Lhoest, Q.; and Rush, A. M. 2019. HuggingFace’s Transformers: State-of-the-art Natural Language Processing. _ArXiv_ abs/1910.03771. * Wu et al. (2019) Wu, C.-S.; Madotto, A.; Hosseini-Asl, E.; Xiong, C.; Socher, R.; and Fung, P. 2019\. Transferable Multi-Domain State Generator for Task-Oriented Dialogue Systems 808–819. doi:10.18653/v1/p19-1078. * Zhu et al. (2020) Zhu, Q.; Zhang, W.; Liu, T.; and Wang, W. Y. 2020. Counterfactual Off-Policy Training for Neural Response Generation URL http://arxiv.org/abs/2004.14507.
# Marangoni instability of a drop in a stably stratified liquid Yanshen Li<EMAIL_ADDRESS>Physics of Fluids group, Max-Planck Center Twente for Complex Fluid Dynamics, Department of Science and Technology, Mesa+ Institute, and J. M. Burgers Centre for Fluid Dynamics, University of Twente, P.O. Box 217, 7500 AE Enschede, The Netherlands Christian Diddens Physics of Fluids group, Max-Planck Center Twente for Complex Fluid Dynamics, Department of Science and Technology, Mesa+ Institute, and J. M. Burgers Centre for Fluid Dynamics, University of Twente, P.O. Box 217, 7500 AE Enschede, The Netherlands Department of Mechanical Engineering, Eindhoven University of Technology, P.O. Box 513, 5600 MB Eindhoven, The Netherlands Andrea Prosperetti Physics of Fluids group, Max-Planck Center Twente for Complex Fluid Dynamics, Department of Science and Technology, Mesa+ Institute, and J. M. Burgers Centre for Fluid Dynamics, University of Twente, P.O. Box 217, 7500 AE Enschede, The Netherlands Department of Mechanical Engineering, University of Houston, Texas 77204-4006, USA Detlef Lohse<EMAIL_ADDRESS>Physics of Fluids group, Max-Planck Center Twente for Complex Fluid Dynamics, Department of Science and Technology, Mesa+ Institute, and J. M. Burgers Centre for Fluid Dynamics, University of Twente, P.O. Box 217, 7500 AE Enschede, The Netherlands Max Planck Institute for Dynamics and Self-Organization, Am Faßberg 17, 37077 Göttingen, Germany ###### Abstract Marangoni instabilities can emerge when a liquid interface is subjected to a concentration or temperature gradient. It is generally believed that for these instabilities bulk effects like buoyancy are negligible as compared to interfacial forces, especially on small scales. Consequently, the effect of a stable stratification on the Marangoni instability has hitherto been ignored. Here we report, for an immiscible drop immersed in a stably stratified ethanol-water mixture, a new type of oscillatory solutal Marangoni instability which is triggered once the stratification has reached a critical value. We experimentally explore the parameter space spanned by the stratification strength and the drop size and theoretically explain the observed crossover from levitating to bouncing by balancing the advection and diffusion around the drop. Finally, the effect of the stable stratification on the Marangoni instability is surprisingly amplified in confined geometries, leading to an earlier onset. A concentration or temperature gradient applied to an interface can induce a Marangoni instability of the motionless state, resulting in a steady convection. Similarly, the steady state Marangoni convection can undergo another instability, leading to an oscillatory motion rednikov1998two . Since the first quantitative analysis in 1958 pearson1958convection , Marangoni instabilities have been studied extensively due to their relevance for liquid extraction sternling1959interfacial ; groothuis1960influence ; rother1999effect ; berejnov2002spontaneous ; jain2011recent , coating techniques pearson1958convection ; yarin1995surface ; demekhin2006suppressing , metal processing gupta1992pore ; ratke2005theoretical ; zhang2006indirect and crystal growth schwabe1978experiments ; schwabe1979some ; chang1979thermocapillary ; chun1979experiments ; schwabe1982studies ; preisser1983steady ; kamotani1984oscillatory , etc. Marangoni instabilities are also the main mechanism to drive the self-propulsion of active drops rednikov1994active ; rednikov1994drop ; herminghaus2014interfacial ; yoshinaga2014spontaneous ; ryazantsev2017thermo ; maass2016swimming ; morozov2019self , which have attracted lots of recent interest. Such drops are an example of the rich physicochemical hydrodynamics of droplets far from equilibrium lohse2020physicochemical which are very relevant for food processing degner2013influence ; degner2014factors and modelling biological systems maass2016swimming , etc. Depending on the application, Marangoni instabilities have been investigated in different configurations, such as a horizontal interface between two fluid layers pearson1958convection ; sternling1959interfacial ; reichenbach1981linear ; takashima1981surface ; levchenko1981instability ; nepomnyashchii1983thermocapillary ; chu1988sustained ; chu1989transverse ; hennenberg1992transverse ; rednikov1998two , the surface of a falling film on a tilted plate nepomnyashchy1976wavy ; chang1994wave ; kliakhandler1997viscous ; miladinova2005effects ; demekhin2006suppressing , a vertical interface of a liquid column schwabe1978experiments ; schwabe1979some ; chang1979thermocapillary ; chun1979experiments ; schwabe1982studies ; preisser1983steady ; kamotani1984oscillatory , and for drops submerged in a solution rednikov1994active ; rednikov1994drop ; herminghaus2014interfacial ; yoshinaga2014spontaneous ; ryazantsev2017thermo ; maass2016swimming ; morozov2019self ; thanasukarn2004impact ; ghosh2008factors ; degner2013influence ; degner2014factors ; dedovets2018five , etc. In many of these situations, these systems are subjected to a stabilizing temperature/concentration gradient takashima1981surface ; levchenko1981instability ; demekhin2006suppressing ; schwabe1978experiments ; schwabe1979some ; chang1979thermocapillary ; chun1979experiments ; schwabe1982studies ; preisser1983steady ; kamotani1984oscillatory ; chu1988sustained ; chu1989transverse , which induces a continuously stable density stratification. However, except for a few cases for the horizontal interface configuration welander1964convective ; wierschem2000internal ; rednikov2000rayleigh , the effect of such a stable density stratification on Marangoni convection has always been ignored, due to the generally accepted view that on small scales bulk effects like buoyancy are negligible nepomnyashchy2012interfacial . Here we report, for an immiscible drop immersed in an ethanol-water mixture, that the stable stratification could actually trigger an oscillatory instability once it is above a critical value. Surprisingly, this critical value will decrease in a confined geometry, implying that the effect of the stable stratification is actually amplified on small scales. Our findings demonstrate that stable stratification can strongly affect Marangoni convection and ask for further studies in related geometries. Figure 1: Bouncing and levitating drops in a linearly and stably stratified mixture of ethanol (lighter) and water (heavier). (a) Snapshots of two 5 cSt silicone oil drops at the given time after they were released in the mixture. The larger drop bounces at $h<$3\text{\,}\mathrm{m}\mathrm{m}$$, while the smaller drop is levitating at a higher position $h\approx$8.7\text{\,}\mathrm{m}\mathrm{m}$$. The snapshots are taken from one experiment with two drops. To better show them, the upper/lower half of the snapshots are shown with different scales. (b) Drop’s height $h$ as functions of time $t$ for different drop radii $R$ after the initial sinking period. $h=0$ is the position where the density of the drop equals that of the mixture. The filled circles represent the relative size of the drops. (c) Flow field around a levitating drop ($R=31\pm$1\text{\,}\mathrm{\SIUnitSymbolMicro m}$$) measured by PIV, in a mixture with $\mathrm{d}w_{\mathrm{e}}/\mathrm{d}y\approx$5\text{\,}\mathrm{m}^{-1}$$. The resolution is not high enough to resolve the velocity close to the drop’s surface. (d) The ethanol weight fraction $w_{\mathrm{e}}$ at the corresponding height, with $\mathrm{d}w_{\mathrm{e}}/\mathrm{d}y\approx$5\text{\,}\mathrm{m}^{-1}$$. (e) A sketch of the levitating drop (with radius $R$ and density $\rho^{\prime}$) and the ethanol concentration around it. Deeper red means higher ethanol concentration. The shaded ring inside the dashed circle represents the kinematic boundary layer with thickness $\delta$, set by the Marangoni velocity $V_{\mathrm{M}}$. The ethanol concentration inside this layer is enhanced & homogenized by Maragnoni advection bringing down the ethanol rich liquid. The Marangoni flow is represented by the solid arrows. Dashed arrows represent diffusion across this layer. $\rho$ is the representative density inside this layer, and $\rho^{*}$ is the undisturbed density in the far field. $\mu$ and $\mu^{\prime}$ are the viscosities of the mixture and the drop, respectively. (f) Interfacial tension $\sigma(w_{\mathrm{e}})$ between $5\text{\,}\mathrm{c}\mathrm{S}\mathrm{t}$ silicone oil and the ethanol-water mixture. Each point is an average of six measurements and the error bar is the standard deviation. The solid line is a polynomial fit to the data points. To determine the onset of the Marangoni instability, we experimentally explore the parameter space spanned by the concentration gradient and the drop radius $R$. Using the double-bucket method oster1965density , linearly stratified liquid mixtures are prepared in a cubic glass container (Hellma, 704.001-OG, Germany) with inner width of $L=$30\text{\,}\mathrm{m}\mathrm{m}$$ filled to different depth, depending on the degree of stratification. The ethanol weight fraction $w_{\mathrm{e}}$ at each height is measured by laser deflection lin2013one ; li2019bouncing , from which the gradient of ethanol weight fraction $\mathrm{d}w_{\mathrm{e}}/\mathrm{d}y$ is calculated. The concentration gradient $\mathrm{d}w_{\mathrm{e}}/\mathrm{d}y$ is varied from $\sim$3\text{\,}\mathrm{m}^{-1}$$ to $\sim$130\text{\,}\mathrm{m}^{-1}$$, corresponding to density gradients ranging from $-480\text{\,}\mathrm{k}\mathrm{g}\mathrm{/}\mathrm{m}^{4}$ to $-4200\text{\,}\mathrm{k}\mathrm{g}\mathrm{/}\mathrm{m}^{4}$. $5\text{\,}\mathrm{c}\mathrm{S}\mathrm{t}$ Silicone oil (Sigma-Aldrich, Germany) is injected through a thin needle (with outer-diameter $0.515\text{\,}\mathrm{m}\mathrm{m}$) to generate drops of different radii $R$. The drops are released from the top of the stratified mixtures, and their trajectories are recorded by a sideview camera. During the measurements, only one single drop exists in the container at a time. The silicone oil has density $\rho^{\prime}=$913\text{\,}\mathrm{k}\mathrm{g}\mathrm{/}\mathrm{m}^{3}$$ and viscosity $\mu^{\prime}=$4.6\text{\,}\mathrm{m}\mathrm{Pa}\cdot\mathrm{s}$$. Two typical behaviors are observed after the initial sinking phase. See Fig. 1(a) for the successive snapshots of two silicone oil drops in a mixture with $\mathrm{d}w_{\mathrm{e}}/\mathrm{d}y\approx$5\text{\,}\mathrm{m}^{-1}$$: While a smaller drop ($R=69\pm$2\text{\,}\mathrm{\SIUnitSymbolMicro m}$$) stays at a fixed position around $h\approx$8.7\text{\,}\mathrm{m}\mathrm{m}$$, a larger drop ($R=454\pm$2\text{\,}\mathrm{\SIUnitSymbolMicro m}$$) bounces continuously in the range $$0\text{\,}\mathrm{m}\mathrm{m}$<h<$3\text{\,}\mathrm{m}\mathrm{m}$$. Here $h=0$ marks the position where the density of the oil ($\rho^{\prime}=$913\text{\,}\mathrm{k}\mathrm{g}\mathrm{/}\mathrm{m}^{3}$$) equals that of the mixture (at $w_{\mathrm{e}}\approx$49\text{\,}\%$$). The drop’s position $h(t)$ as a function of time $t$ in the same stratified liquid and the ethanol weight fraction $w_{\mathrm{e}}$ at the corresponding height are respectively shown in Fig. 1(b) and (d). The smallest drop ($R_{1}\approx$44\text{\,}\mathrm{\SIUnitSymbolMicro m}$$) is levitating at $h\approx$9.1\text{\,}\mathrm{m}\mathrm{m}$$. As the drop size increases, it levitates at a lower position, until above a critical radius $R_{\mathrm{cr}}$ it starts to bounce instead of levitating. If its size is further increased, the drop bounces around a lower position (but still with $h>0$). The smaller drops are able to levitate above the density matched position $h=0$ against gravity because of a stable Marangoni flow around it, as shown in Fig. 1(c). The flow field is obtained by PIV measurements for a drop levitating in the gradient $\mathrm{d}w_{\mathrm{e}}/\mathrm{d}y\approx$5\text{\,}\mathrm{m}^{-1}$$. The interfacial tension of the drop $\sigma$ decreases with increasing ethanol concentration of the mixture $w_{\mathrm{e}}$, as shown in Fig. 1(f). This interfacial tension gradient at the drop’s surface pulls liquid downwards, generating a viscous force acting against gravity, which levitates the drop. When the drop becomes large enough, however, the equilibrium becomes oscillatory, and the drop starts to bounce between two different levels. Thus, the transition from a levitating drop to a bouncing one signals the onset of the instability. While exploring the parameter space, we use an easily distinguishable criterion to determine whether a drop is bouncing: If the drop’s bouncing amplitude $h_{\mathrm{A}}$ is larger than its radius $R$, then the drop is considered to be bouncing (see Supplemental Material for more details). The results are shown in Fig. 2. Surprisingly, while for weak gradients (like for $\mathrm{d}w_{\mathrm{e}}/\mathrm{d}y\approx$10\text{\,}\mathrm{m}^{-1}$$) there is a critical radius $R_{\mathrm{cr}}$ ($\approx$80\text{\,}\mathrm{\SIUnitSymbolMicro m}$$) above which the Marangoni flow becomes unstable, the Marangoni flow is always unstable for stronger gradients $\mathrm{d}w_{\mathrm{e}}/\mathrm{d}y>$23\text{\,}\mathrm{m}^{-1}$$ in all performed experiments. Note that for larger drops ($R>$0.1\text{\,}\mathrm{m}\mathrm{m}$$), we could not explore the full parameter space for $\mathrm{d}w_{\mathrm{e}}/\mathrm{d}y<$3\text{\,}\mathrm{m}^{-1}$$ since it would require an unrealistically large container. Figure 2: Phase diagram of the levitating & bouncing drops in the parameter space of drop radius $R$ vs. concentration gradient $\mathrm{d}w_{\mathrm{e}}/\mathrm{d}y$. Black triangles stand for bouncing drops, red circles for levitating ones. Measurement errors in the $x$ direction are comparable with the size of the symbols. To get a better understanding of the onset of this Marangoni instability, the key is to understand the coupling between the Marangoni flow and the concentration field: The Marangoni flow is induced by the ethanol (solute) concentration gradient around the drop, which is subjected to change by advection (caused by the Marangoni flow itself) and diffusion, see the sketch in Fig. 1(e). The Marangoni flow tends to homogenize the ethanol concentration around the drop, thus weakening the Marangoni flow force and thus itself. At the same time, diffusion acts to restore the ethanol gradient in the vicinity of the drop to its undisturbed value, i.e., the value it takes in the far field. This competition between advection and diffusion around the drop determines whether the Marangoni flow is stable or not. Furthermore, once it becomes unstable, a temporarily strong Marangoni flow homogenizes the concentration field around the drop, consequently weakening itself. Later the Marangoni flow restarts once diffusion has restored the concentration field around the drop, so that the flow is oscillatory and leads to the continuous bouncing of the drop. The liquid layer whose concentration is affected by the Marangoni advection is effectively the Marangoni flow boundary layer with thickness $\delta$ (see Fig. 1(e)). The time scale for advection to change the concentration in this layer is the time needed for the Marangoni flow to bring down the ethanol-rich liquid from the top: $\tau_{\mathrm{a}}\sim R/V_{\mathrm{M}}$, where $V_{\mathrm{M}}$ is the Marangoni flow velocity at the equator of the drop. For the drop in the concentration gradient it holds young1959motion (see Supplementary Material) $V_{\mathrm{M}}\sim-\mathrm{d}\sigma/\mathrm{d}y\cdot R/(\mu+\mu^{\prime})$. The time scale for diffusion to restore the concentration across this layer is $\tau_{\mathrm{d}}\sim\delta^{2}/D$, where $D$ is the diffusivity of ethanol in water. The flow will become unstable when advection is faster than diffusion, $\tau_{\mathrm{a}}<\tau_{\mathrm{d}}$. Substituting the two time scales into this relation, we obtain ${V_{\mathrm{M}}R}/{D}>{R^{2}}/{\delta^{2}}$. The left hand side has the form of a Péclet number, which is the ratio between advection and diffusion, and which in problems of this type is referred to as the Marangoni number $Ma=\frac{V_{\mathrm{M}}R}{D}=-\frac{\mathrm{d}\sigma}{\mathrm{d}w_{\mathrm{e}}}\frac{\mathrm{d}w_{\mathrm{e}}}{\mathrm{d}y}R^{2}\cdot\frac{1}{(\mu+\mu^{\prime})D},$ (1) where we have used above expression for $V_{\mathrm{M}}$ with an equal sign and where $\mathrm{d}\sigma/\mathrm{d}w_{\mathrm{e}}$ is a material property (see Fig. 1(f)) and $\mathrm{d}w_{\mathrm{e}}/\mathrm{d}y$ the undisturbed ethanol gradient of the mixture. The instability criterion thus is $Ma>{R^{2}}/{\delta^{2}}.$ (2) The liquid within the boundary layer is lighter than its surroundings as it is entrained from the top, and it is held in place by the Marangoni induced viscous stress against buoyancy: $\mu\frac{V_{\mathrm{M}}}{\delta^{2}}\sim g\Delta\rho,$ (3) where $\Delta\rho=\rho^{*}-\rho$ is the density difference between the liquid inside and outside of the kinematic boundary layer, see Fig. 1(e). The lighter liquid is brought down by the Marangoni flow along the drop’s surface, so $\Delta\rho\sim-R\cdot\mathrm{d}\rho/\mathrm{d}y$. Cancelling $\delta$ from Eqs.(2)&(3), we obtain the instability criterion $Ma/{Ra}^{1/2}>c,$ (4) where $Ra=-\frac{\mathrm{d}\rho}{\mathrm{d}y}\cdot\frac{gR^{4}}{\mu D}$ (5) is the Rayleigh number for characteristic length $R$ and $c$ is a constant to be determined. To calculate the Marangoni and Rayleigh numbers, ethanol weight fractions at the positions where the drops levitate are used to obtain the viscosity $\mu$, diffusivity $D$ and the interfacial tension $\sigma$ (see Supplemental Material for the concentration dependence of $\mu$ and $D$). In the following, for bouncing drops, we use values corresponding to their lowest position. The phase diagram shown in Fig. 2 is replotted with $Ma/Ra^{1/2}$ vs. $Ra$ in Fig. 3(a). It clearly shows that there is indeed a critical value $(Ma/{Ra}^{1/2})_{\mathrm{cr}}$ above which the drop will always bounce, and the instability threshold $c$ in Eq.(4) is measured to be $c=275\pm 10$ in the range $6\times 10^{-3}\lesssim Ra\lesssim 3$. We cannot carry out experiments for $Ra<6\times 10^{-3}$ because drops with $R<$20\text{\,}\mathrm{\SIUnitSymbolMicro m}$$ are too small to observe. For experiments in the range $Ra>3$, the finite size of the container comes into play. However, we speculate that $c\approx 275$ still holds for $Ra>3$ as long as the container is large enough. The existing data on bouncing are consistent with this value. Figure 3: (a) Phase diagram replotted in dimensionless numbers: $Ma/Ra^{1/2}$ vs. $Ra$. Black triangles stand for bouncing drops, red circles for levitating ones. The blue line is the instability threshold $(Ma/{Ra}^{1/2})_{\mathrm{cr}}=275$, above which the flow is oscillatory and all drops bounce. The blue solid line (in the range $6\times 10^{-3}<Ra<3$) is confirmed by experiments. Measurement errors in the $y$ direction are comparable with the size of the symbols. (b) Phase diagram replotted with $\mathrm{d}w_{\mathrm{e}}/\mathrm{d}y$ vs. $w_{\mathrm{e}}$, where $w_{\mathrm{e}}$ is the ethanol weight fraction at the levitation height. The blue curve is calculated from Eq.(6) with $c=275$. The dashed blue line in the range $w_{\mathrm{e}}>$98\text{\,}\mathrm{w}\mathrm{t}\%$$ ($w_{\mathrm{e}}<$50\text{\,}\mathrm{w}\mathrm{t}\%$$) corresponds to $Ra<6\times 10^{-3}$ ($Ra>3$). Measurement errors are comparable with the size of the symbols. We now express our stability criterion Eq.(4) in dimensional quantities by substituting the definition of $Ma$ and $Ra$ to obtain: $\left(\frac{\mathrm{d}w_{\mathrm{e}}}{\mathrm{d}y}\right)_{\mathrm{cr}}=c^{2}\left(\mu+\mu^{\prime}\right)^{2}\cdot\frac{gD}{\mu}\frac{\mathrm{d}\rho}{\mathrm{d}\sigma}\frac{\mathrm{d}w_{\mathrm{e}}}{\mathrm{d}\sigma}.$ (6) Eq.(6) actually predicts a critical concentration gradient above which the equilibrium is unstable. Note that remarkably the drop radius $R$ does not enter into this equation. All the fluid properties $\mu$, $D$, $\mathrm{d}\rho/\mathrm{d}\sigma$ and $\mathrm{d}w_{\mathrm{e}}/\mathrm{d}\sigma$ depend on $w_{\mathrm{e}}$ – the ethanol weight fraction at the levitation height. Thus the critical gradient $(\mathrm{d}w_{\mathrm{e}}/\mathrm{d}y)_{\mathrm{cr}}$ as a function of $w_{\mathrm{e}}$ is shown in Fig. 3(b) as the blue curve. The data shown in Fig. 2 are also replotted in Fig. 3(b). As can be seen, the blue curve as predicted by Eq.(6) nicely separates the levitating drops and the bouncing ones. The dashed blue line in the range $w_{\mathrm{e}}>$98\text{\,}\mathrm{w}\mathrm{t}\%$$ ($w_{\mathrm{e}}<$50\text{\,}\mathrm{w}\mathrm{t}\%$$) corresponds to $Ra<6\times 10^{-3}$ ($Ra>3$), i.e., the region in which we could not perform experiments. The above results are all obtained in a large enough container. We will now discuss the effect of a geometrical confinement, i.e., the dependence of our findings on the container size $L$. Let $\mathcal{L}$ denote the maximum extent of the flow field induced by the drop. Then $\mathcal{L}>L$ means that the flow is confined. In the case of no confinement, i.e., $\mathcal{L}<L$, the liquid in the far field is not disturbed by the Marangoni flow, so that the density in the far field is maintained at $\rho^{*}$ (see Fig. 1(e)). However, when the flow is confined, i.e., $\mathcal{L}>L$, the liquid close to the side wall is affected by the Marangoni flow. In such a situation, because the liquid is pulled down in the center by the drop, the liquid close to the wall will be pushed up due to mass conservation. This effectively increases the density $\rho^{*}$. Consequently, the density difference $\Delta\rho=\rho^{*}-\rho$ is increased, which means that the effect of buoyancy is amplified. According to Eqs. (3)&(4), the instability threshold $c$ will thus decrease. Either decreasing the container size $L$ or increasing $\mathcal{L}$ both leads to a stronger confinement effect. Since for stable stratifications $\mathcal{L}\sim\left(-{\mathrm{d}\rho}/{\mathrm{d}y}\cdot{\mu D}/{g}\right)^{-1/4}$ phillips1970flows ; wunsch1970oceanic , one can thus also increase the confinement effect by using very weak stratifications. We have performed experiments for weaker gradients $\mathrm{d}w_{\mathrm{e}}/\mathrm{d}y<$3\text{\,}\mathrm{m}^{-1}$$ and also in a larger container to confirm the effect of the confinement. Indeed, for $\mathrm{d}w_{\mathrm{e}}/\mathrm{d}y\approx$2\text{\,}\mathrm{m}^{-1}$$, a cubic container with $L=$50\text{\,}\mathrm{m}\mathrm{m}$$ is already not large enough, and the instability threshold is reduced to $c\approx 172$. A smaller container ($L=$30\text{\,}\mathrm{m}\mathrm{m}$$) further decreases the threshold to $c\approx 157$. For $\mathrm{d}w_{\mathrm{e}}/\mathrm{d}y\approx$1\text{\,}\mathrm{m}^{-1}$$, the geometry is more confined, and the threshold is further decreased to $c\approx 122$ in the larger container and even to $c\approx 117$ in the smaller one. The fact that a weaker stratification leads to a more amplified effect of buoyancy demonstrates that the stable stratification is very relevant for the Marangoni instability, in particular on small scales where the confinement is more relevant. In conclusion, we have discovered a new type of oscillatory Marangoni instability for an immiscible drop immersed in a stably stratified ethanol- water mixture. The commonly ignored stable density stratification induced by the concentration gradient is vital in triggering this instability. Its onset is indicated by the transition from a levitating drop to a bouncing one. By experimentally exploring the parameter space spanned by the concentration gradient $\mathrm{d}w_{\mathrm{e}}/\mathrm{d}y$ and the drop radius $R$, the instability is found to be determined by the balance between the advection and diffusion through the kinetic boundary layer set by the Marangoni flow. This yields a critical concentration gradient as the instability criterion. Remarkably, the critical gradient is decreased in a confined geometry, i.e., the effect of the stable stratification is amplified on small confined scales. Our findings indicate that the stable stratification induced by the corresponding concentration gradient is very relevant, especially in confined geometries, and should be further explored in other geometries. Our results for solutal Marangoni flows can also be extended to thermal Marangoni flows. We found that a stabilizing temperature gradient as low as $3\text{\,}\mathrm{K}\mathrm{/}\mathrm{m}\mathrm{m}$ can trigger a similar oscillatory instability on a bubble immersed in water. Such low temperature gradient is smaller than what is occurring in various applications schwabe1978experiments ; ratke2006destabilisation ; dedovets2018five , where a stabilizing temperature gradient can easily go beyond $10\text{\,}\mathrm{K}\mathrm{/}\mathrm{m}\mathrm{m}$. We thank Chao Sun and Vatsal Sanjay for valuable discussions. We acknowledge support from the Netherlands Center for Multiscale Catalytic Energy Conversion (MCEC), an NWO Gravitation programme funded by the Ministry of Education, Culture and Science of the government of Netherlands, and the ERC-Advanced Grant Diffusive Droplet Dynamics (DDD) with Project No. 740479. ## References * (1) A. Y. Rednikov, P. Colinet, M. G. Velarde, and J. C. Legros, Two-layer Benard-Marangoni instability and the limit of transverse and longitudinal waves, Phys. Rev. E 57, 2872 (1998). * (2) J. R. A. Pearson, On convection cells induced by surface tension, J. Fluid Mech. 4, 489 (1958). * (3) C. V. Sternling and L. E. Scriven, Interfacial turbulence: Hydrodynamic instability and the Marangoni effect, AIChE J. 5, 514 (1959). * (4) H. Groothuis and F. J. Zuiderweg, Influence of mass transfer on coalescence of drops, Chem. Eng. Sci 12, 288 (1960). * (5) M. A. Rother and R. H. Davis, The effect of slight deformation on thermocapillary-driven droplet coalescence and growth, J. Colloid Interface Sci. 214, 297 (1999). * (6) V. Berejnov, A. M. Leshanksy, O. M. Lavrenteva, and A. Nir, Spontaneous thermocapillary interaction of drops: Effect of surface deformation at nonzero capillary number, Phys. Fluids 14, 1326 (2002). * (7) A. Jain and K. K. Verma, Recent advances in applications of single-drop microextraction: A review, Anal. Chim. Acta 706, 37 (2011). * (8) A. L. Yarin, Surface-tension-driven flows at low Reynolds number arising in optoelectronic technology, J. Fluid Mech. 286, 173 (1995). * (9) E. A. Demekhin, S. Kalliadasis, and M. G. Velarde, Suppressing falling film instabilities by Marangoni forces, Phys. Fluids 18, 042111 (2006). * (10) A. K. Gupta, B. K. Saxena, S. N. Tiwari, and S. L. Malhotra, Pore formation in cast metals and alloys, J. Mater. Sci. 27, 853 (1992). * (11) L. Ratke, Theoretical considerations and experiments on microstructural stability regimes in monotectic alloys, Mater. Sci. Eng. A 413, 504 (2005). * (12) L. Zhang, Indirect methods of detecting and evaluating inclusions in steel—a review, J. Iron Steel Res. Int. 13, 1 (2006). * (13) D. Schwabe, A. Scharmann, F. Preisser, and R. Oeder, Experiments on surface tension driven flow in floating zone melting, J. Cryst. Growth 43, 305 (1978). * (14) D. Schwabe and A. Scharmann, Some evidence for the existence and magnitude of a critical Marangoni number for the onset of oscillatory flow in crystal growth melts, J. Cryst. Growth 46, 125 (1979). * (15) C. E. Chang, W. R. Wilcox, and R. A. Lefever, Thermocapillary convection in floating zone melting: Influence of zone geometry and Prandtl number at zero gravity, Mater. Res. Bull. 14, 527 (1979). * (16) C. H. Chun and W. Wuest, Experiments on the transition from the steady to the oscillatory Marangoni-convection of a floating zone under reduced gravity effect, Acta Astronaut. 6, 1073 (1979). * (17) D. Schwabe, A. Scharmann, and F. Preisser, Studies of Marangoni convection in floating zones, Acta Astronaut. 9, 183 (1982). * (18) F. Preisser, D. Schwabe, and A. Scharmann, Steady and oscillatory thermocapillary convection in liquid columns with free cylindrical surface, J. Fluid Mech. 126, 545 (1983). * (19) Y. Kamotani, S. Ostrach, and M. Vargas, Oscillatory thermocapillary convection in a simulated floating-zone configuration, J. Cryst. Growth 66, 83 (1984). * (20) A. Y. Rednikov, Y. S. Ryazantsev, and M. G. Velarde, Active drops and drop motions due to nonequilibrium phenomena, J. Non-Equilib. Thermodyn. 19, 95 (1994). * (21) A. Y. Rednikov, Y. S. Ryazantsev, and M. G. Velarde, Drop motion with surfactant transfer in a homogeneous surrounding, Phys. Fluids 6, 451 (1994). * (22) S. Herminghaus, C. C. Maass, C. Krüger, S. Thutupalli, L. Goehring, and C. Bahr, Interfacial mechanisms in active emulsions, Soft Matter 10, 7008 (2014). * (23) N. Yoshinaga, Spontaneous motion and deformation of a self-propelled droplet, Phys. Rev. E 89, 012913 (2014). * (24) Y. S. Ryazantsev, M. G. Velarde, R. G. Rubio, E. Guzmán, F. Ortega, and P. López, Thermo-and soluto-capillarity: Passive and active drops, Adv. Colloid Interfac. 247, 52 (2017). * (25) C. C. Maass, C. Krüger, S. Herminghaus, and C. Bahr, Swimming droplets, Annu. Rev. Condens. Matter Phys. 7, 171 (2016). * (26) M. Morozov and S. Michelin, Self-propulsion near the onset of Marangoni instability of deformable active droplets, J. Fluid Mech. 860, 711 (2019). * (27) D. Lohse and X. Zhang, Physicochemical Hydrodynamics of Droplets out of Equilibrium, Nat. Rev. Phys. 2, 426 (2020). * (28) B. M. Degner, K. M. Olson, D. Rose, V. Schlegel, R. Hutkins, and D. J. McClements, Influence of freezing rate variation on the microstructure and physicochemical properties of food emulsions, J. Food Eng. 119, 244 (2013). * (29) B. M. Degner, C. Chung, V. Schlegel, R. Hutkins, and D. J. McClements, Factors influencing the freeze-thaw stability of emulsion-based foods, Compr. Rev. Food. Sci. F. 13, 98 (2014). * (30) J. Reichenbach and H. Linde, Linear perturbation analysis of surface-tension-driven convection at a plane interface (Marangoni instability), J. Colloid Interface Sci. 84, 433 (1981). * (31) M. Takashima, Surface tension driven instability in a horizontal liquid layer with a deformable free surface. I. Stationary convection, J. Phys. Soc. Japan 50, 2745 (1981). * (32) E. B. Levchenko and A. L. Chernyakov, Instability of surface waves in a nonuniformly heated liquid, Sov. Phys. JETP 54, 102 (1981). * (33) A. A. Nepomnyashchii and I. B. Simanovskii, Thermocapillary convection in a two-layer system, Fluid Dyn. 18, 629 (1983). * (34) X. L. Chu and M. G. Velarde, Sustained transverse and longitudinal-waves at the open surface of a liquid, PhysicoChem. Hydrodyn. 10, 727 (1988). * (35) X. L. Chu and M. G. Velarde, Transverse and longitudinal waves induced and sustained by surfactant gradients at liquid-liquid interfaces, J. Colloid Interface Sci. 131, 471 (1989). * (36) M. Hennenberg, X. L. Chu, A. Sanfeld, and M. G. Velarde, Transverse and longitudinal waves at the air-liquid interface in the presence of an adsorption barrier, J. Colloid Interface Sci. 150, 7 (1992). * (37) A. A. Nepomnyashchy, Wavy motions in the layer of viscous fluid flowing down the inclined plane, Fluid Dynamics, Part 8, Proc. of Perm State University 362, 114 (1976). * (38) H. Chang, Wave evolution on a falling film, Annu. Rev. Fluid Mech. 26, 103 (1994). * (39) I. L. Kliakhandler and G. I. Sivashinsky, Viscous damping and instabilities in stratified liquid film flowing down a slightly inclined plane, Phys. Fluids 9, 23 (1997). * (40) S. Miladinova and G. Lebon, Effects of nonuniform heating and thermocapillarity in evaporating films falling down an inclined plate, Acta Mech. 174, 33 (2005). * (41) P. Thanasukarn, R. Pongsawatmanit, and D. J. McClements, Impact of fat and water crystallization on the stability of hydrogenated palm oil-in-water emulsions stabilized by whey protein isolate, Colloids Surf. A Physicochem. Eng. Asp. 246, 49 (2004). * (42) S. Ghosh and J. N. Coupland, Factors affecting the freeze–thaw stability of emulsions, Food Hydrocoll. 22, 105 (2008). * (43) D. Dedovets, C. Monteux, and S. Deville, Five-dimensional imaging of freezing emulsions with solute effects, Science 360, 303 (2018). * (44) P. Welander, Convective instability in a two-layer fluid heated uniformly from above, Tellus 16, 349 (1964). * (45) A. Wierschem, H. Linde, and M. G. Velarde, Internal waves excited by the Marangoni effect, Phys. Rev. E 62, 6522 (2000). * (46) A. Y. Rednikov, P. Colinet, M. G. Velarde, and J. C. Legros, Rayleigh–Marangoni oscillatory instability in a horizontal liquid layer heated from above: Coupling and mode mixing of internal and surface dilational waves, J. Fluid Mech. 405, 57 (2000). * (47) A. Nepomnyashchy, J. C. Legros, and I. Simanovskii, Interfacial convection in multilayer systems (Springer, New York, 2012). * (48) G. Oster, Density gradients, Sci. Am. 213, 70 (1965). * (49) D. Lin, J. R. Leger, M. Kunkel, and P. McCarthy, One-dimensional gradient-index metrology based on ray slope measurements using a bootstrap algorithm, Opt. Eng. 52, 112108 (2013). * (50) Y. Li, C. Diddens, A. Prosperetti, K. L. Chong, X. Zhang, and D. Lohse, Bouncing oil droplet in a stratified liquid and its sudden death, Phys. Rev. Lett. 122, 154502 (2019). * (51) N. O. Young, J. S. Goldstein, and M. J. Block, The motion of bubbles in a vertical temperature gradient, J. Fluid Mech. 6, 350 (1959). * (52) O. M. Phillips, On flows induced by diffusion in a stably stratified fluid, Deep Sea Res. Ocean. Abstr. 17, 435 (1970). * (53) C. Wunsch, On oceanic boundary mixing, Deep Sea Res. Ocean. Abstr. 17, 293 (1970). * (54) L. Ratke and A. Müller, On the destabilisation of fibrous growth in monotectic alloys, Scr. Mater. 54, 1217 (2006).
# KoreALBERT: Pretraining a Lite BERT Model for Korean Language Understanding Hyunjae Lee, Jaewoong Yoon, Bonggyu Hwang, Seongho Joe, Seungjai Min, Youngjune Gwon Samsung SDS ###### Abstract A Lite BERT (ALBERT) has been introduced to scale up deep bidirectional representation learning for natural languages. Due to the lack of pretrained ALBERT models for Korean language, the best available practice is the multilingual model or resorting back to the any other BERT-based model. In this paper, we develop and pretrain KoreALBERT, a monolingual ALBERT model specifically for Korean language understanding. We introduce a new training objective, namely Word Order Prediction (WOP), and use alongside the existing MLM and SOP criteria to the same architecture and model parameters. Despite having significantly fewer model parameters (thus, quicker to train), our pretrained KoreALBERT outperforms its BERT counterpart on 6 different NLU tasks. Consistent with the empirical results in English by Lan et al., KoreALBERT seems to improve downstream task performance involving multi- sentence encoding for Korean language. The pretrained KoreALBERT is publicly available to encourage research and application development for Korean NLP. ## I Introduction Pre-trained language models are becoming an essential component to build a modern natural language processing (NLP) application. Previously, recurrent neural nets such as LSTM have dominated sequence-to-sequence (seq2seq) [1] modeling for natural languages, upholding state-of-the-art performances for core language understanding tasks. Since the introduction of the Transformer [2], recurrent structures in a neural language model are reconsidered and opted for attention, a mechanism that relates different positions in a sequence to compute a representation of the sequence. Devlin _et al._ [3] have proposed Bidirectional Encoder Representations from Transformers (BERT) to improve on predominantly unidirectional training of a language model by using the masked language model (MLM) training objective. MLM is an old concept dating back to the 1950s [4]. By jointly conditioning on both left and right context in all layers, the MLM objective has made pre- training of the deep bidirectional language encoding possible. BERT uses an additional loss for pre-training known as next-sentence prediction (NSP). NSP is designed to learn high-level linguistic coherence by predicting whether or not given two text segments should appear consecutively as in the original text. NSP can improve performance on downstream NLP tasks such as natural language inference that would require reasoning about inter-sentence relations. A Lite BERT (ALBERT) uses parameter reduction techniques to alleviate scaling problems for BERT. ALBERT’s cross-layer parameter sharing can be thought as a form of regularization that helps stabilize the pre-training and generalize despite the substantially reduced number of model parameters. Also, the sentence order prediction (SOP) objective in ALBERT replaces the ineffective the next sentence prediction (NSP) loss in BERT for better inter-sentence coherence. Downstream tasks play critical measures for evaluating emerging language models and NLP applications today. Pre-trained language models are central to downstream task evaluations such as machine translation, text classification, and machine reading comprehension. At a high level, there are two approaches to use pre-trained language models. First, pre-trained models can provide additional feature representations for a downstream task. More importantly, pre-trained models can be a baseline upon which the downstream task is fine- tuned. By having an expensive, but shareable pre-training followed by much smaller fine-tuning, it is a powerful paradigm to focus on optimizing the performance of a downstream NLP task. Self-supervised learning with large corpora allows a suitable starting point for an outer task-specific layer being optimized from scratch while reusing the pre-trained model parameters. Since its introduction, BERT has achieved state-of-the-art accuracy performances for natural language understanding tasks such as GLUE [5], MultiNLI [6], SQuAD v1.1 [7] & SQuAD v2.0 [8], and CoNLL-2003 NER [9]. Despite having fewer parameters than BERT, ALBERT has been able to achieve new state- of-the-art results on the GLUE, RACE [10], and SQuAD benchmarks. It is important to remark that a large network is crucial in pushing state-of- the-art results for downstream tasks. While BERT gives a sound choice to build a general language model trained on large corpora, it is difficult to experiment with training large BERT models due to the memory limitations and computational constraints. Training BERT-large in fact is a lengthy process of consuming significant hardware resources. Besides, there are already a wide variety of languages pre-trained in BERT, which include the multilingual BERT and monolingual models pre-trained in 104 different languages. ALBERT, however, gives a much narrower choice in languages. Asserting an argument that having a better language model is roughly equivalent to pre-train a large model, all without imposing too much memory and computational requirements, we choose to go with ALBERT. In this paper, we develop and train KoreALBERT, a monolingual ALBERT model for Korean language understanding. Compared to a multilingual model, monolingual language models are known to optimize the performance for a specific language in every aspect, including downstream tasks critical to build modern NLP systems and applications. In addition to the original ALBERT MLM and SOP training objectives, we introduce a word order prediction (WOP) loss. WOP is fully compatible with the MLM and SOP losses and can be added gracefully in implementation. Our pre- trained KoreALBERT could outperform multilingual BERT and its BERT counterpart on a brief evaluation with KorQuAD 1.0 benchmark for machine reading comprehension. Consistent with the empirical results of ALBERT pre-trained in English reported by Lan _et al._ [11], KoreALBERT seems to improve supervised downstream task performances involving multiple Korean sentences. The rest of this paper is organized as follows. In Section II, we provide background on pre-trained neural language models. Section III presents KoreALBERT. In Section IV, we describe our implementation, pre-training, and empirical evaluation of KoreALBERT. Section V concludes the paper. Our pre- trained KoreALBERT is publicly available to encourage NLP research and application development for Korean language. ## II Background ### II-A Transformer, BERT, and ALBERT Transformer [2] is a sequence transduction model based solely on attention mechanism, skipping any recurrent and convolutional structures of a neural network. The transformer architecture includes multiple identical encoder and decoder blocks stacked on top of each other. While the encoder captures linguistic information of the input sequence and produces the contextual representations, the decoder generates output sequence corresponding to its pair of input. Thanks to multi-head self-attention layers in an encoder block, transformer can acquire varying attentions within a single sequence and alleviate inevitable dragging caused during the training of a recurrent neural network. BERT distinguishes itself from other language models that predict the next word given previous words by introducing new training methods. Instead of predicting the next token given only previous tokens, it has to predict replaced word by special token [MASK]. This training strategy gives BERT bidirectionality which means having an access to left and right context around the target word. Thus, BERT can produce deep bidirectional representation of input sequence. RoBERTa [12], ALBERT [11] and other variants [13, 14] utilize bidirectional context representation and established state-of-the-art results on a wide range of NLP tasks. BERT is trained with the masked language modeling (MLM) and the next sentence prediction (NSP) losses. NSP is a binary classification task to predict whether or not given two segments separated by another special token [SEP] follow each other in the original text. The task is intended to learn the relationship between two sentences in order to use on many downstream tasks of which input template consists of two sentences as in question answering (QA) and sentence entailment [3]. Recently, there is a criticism toward NSP that the NSP loss does not necessarily help improve the downstream task performances [11, 12, 15] for its loose inter-sentential coherence. Among them, ALBERT, whose architecture is derived from BERT, uses a sentence order prediction(SOP) task instead. In the SOP task, negative examples consist of a pair of sentences from the same document, but the sentence order is swapped, and the model should predict whether or not the order is swapped. With the improved SOP loss and other parameter reduction techniques, ALBERT significantly reduces the number of parameters– _i.e._ , 18x fewer for BERT-large, while achieving similar or better performance on downstream tasks [11]. KoreALBERT takes the unmodified ALBERT architecture as a baseline. We train KoreALBERT from scratch on large Korean corpora collected online. ### II-B Related Work Google has released BERT multilingual model (M-BERT) pre-trained using 104 different languages including the Korean. Karthikeyan _et al._ [16] show why and how well M-BERT works on many downstream NLP tasks without explicitly training with monolingual corpus. More recently, Facebook AI Research presented crosslingual model (XLM-R) [17] generally outperforming M-BERT. Recent literature argues that a monolingual language model is consistently superior to M-BERT. For French, FlauBERT [18] and CamemBERT [19] with the same approach as RoBERTa have been released. ALBERTo [20] focuses on Italian social network data. BERTje [21] for Dutch and FinBERT [22] for Finnish have been developed. They both have achieved superior results on the majority of downstream NLP tasks compared to M-BERT. Some previous work in the Korean language has focused on learning static representations by using language-specific properties [23]. More recently, SKT Brain has released BERT 111https://github.com/SKTBrain/KoBERT and GPT-2 pre- trained on large Korean corpora.222https://github.com/SKT-AI/KoGPT2 Korean Electronics and Telecommunications Research Institute (ETRI) has released two versions of BERT: the morpheme analytic based and the syllable based model.333http://aiopen.etri.re.kr/service_dataset.php These models are worthwhile to experiment with and provide good benchmark evaluations in Korean language model research. BART [24] features interesting denoising approaches for input text used in pre-training such as sentence permutation and text infilling. In the sentence permutation task, an input document is divided into sentences and shuffled in a random order. A combination of text infilling and sentence shuffling tasks has shown significant improvement of the performance over either applied separately. Inspired by BART, we have formulated word order prediction (WOP), a new pre-training loss used alongside the MLM and SOP losses for KoreALBERT. Differentiated from BART, which is essentially a sentence-level shuffling, WOP is an intra-sentence, token-level shuffling. ## III KoreALBERT: Training Korean Language Model Using ALBERT ### III-A Architecture KoreALBERT is a multi-layer bidirectional Transformer encoder with the same factorized embedding parameterization and cross-layer sharing as ALBERT. Inheriting ALBERT-base, KoreALBERT-base has 12 parameter sharing layers with an embedding size of 128 dimensions, 768 hidden units, 12 heads, and GELU nonlinearities [25]. The total number of parameters in KoreALBERT-base is 12 millions, and it increases to 18-million parameters for KoreALBERT-large having 1024 hidden dimensions. Lan _et al._ [11] argues that removing dropout has significantly helped pretraining with the masked language modeling (MLM) loss. For KoreALBERT, however, we have made an empirical decision to keep dropout after observing degraded downstream performances without dropout. ### III-B Training Objectives ALBERT pretrains on two objectives: masked language modeling (MLM) and sentence order prediction (SOP) losses. We keep both objectives for KoreALBERT and introduce an additional training objective called word order prediction (WOP). Word Order Prediction (WOP). Korean is an agglutinative language that a combination of affixes and word roots determines usage and meaning [26]. Decomposing a Korean word into several morphemes and shuffling its order can introduce grammatical errors and semantic altercations. We impose a word order prediction (WOP) loss for pretraining KoreALBERT. The WOP objective is a cross-entropy loss on predicting a correct order of shuffled tokens. WOP is fully compatible with the ALBERT MLM and SOP, and we expect to reinforce correct agglutination (or point out incorrect agglutinative usages) beyond simply checking intra-sentence word orderings. There is an interesting point of view about WOP mixed with MLM and SOP towards the problem of generating a full sentence from a small subset of permuted words. Our primary focus of this paper is on the empirical side of the design and pretraining of an ALBERT-based foreign language model rather than a formal analysis on training objectives. The pretraining of KoreALBERT is illustrated in Fig. 1. A randomly sampled subset of tokens in the input text are replaced with [MASK]. MLM computes a cross-entropy loss on prediction of the masked tokens. As with ALBERT-base, we uniformly choose 15% of the input tokens for possible masking, and the 80% of the chosen are actually replaced with [MASK], leaving 10% unchanged and the rest replaced with randomly selected tokens. SOP is known to focus on modeling inter-sentence coherence. The SOP loss uses two consecutive segments from the same text as a positive example and as a negative example if their order is swapped. We have found that if WOP is too difficult, it can crucially impact the KoreALBERT performance on downstream evaluations. We have experimentally determined WOP to inter-work with MLM and SOP and limited the shuffling rate up to 15%, which seemingly realizes the best empirical performance for our case. In addition, we have decided to include WOP into only specific portion of all batches. We revisit more detailed description of our experimental setup in Section 4. Like MLM, we choose a uniformly random set of tokens for WOP. The most crucial part of integrating WOP into pretraining is _not_ switching tokens across [MASK]. This constraint minimizes the corruption of contextual bidirectionality that acts as essential information in denoising the [MASK] tokens. Figure 1: Pre-training KoreALBERT with the MLM, SOP, and WOP objectives. The loss (on top) with respect to all three objectives is calculated for illustrative purposes. In our implementation, classification layer (highlighted gray) in the middle consisting of three identical heads produces a logit vector with respect to each label. ### III-C Optimization We use the LAMB optimizer [27] with a learning rate of $1.25\times 10^{-3}$ and a warm-up ratio $1.25\times 10^{-2}$. To speed up the pretraining, we maintain an input sequence length of 128 tokens despite the risk of suboptimal performance. Due to memory limitations, it is necessary to use gradient accumulation steps for a batch size of 2,048, which is comparable to BERT. We apply a dropout rate of 0.1 on all layers and attention weights. We use a GELU activation function [25]. ## IV Experiments ### IV-A Implementation We implement KoreALBERT based on Hugging Face’s transformer library [28] with almost an identical model configuration for ALBERT-base. We add another linear classifier on top of the encoder output for WOP task. The added layer is used to predict the probability of the original position of words in the sequence via softmax. Like the MLM objective, we take into account only switched tokens to compute the cross-entropy loss. We train our model using 4 NVidia V100 GPUs with half-precision floating-point weights. ### IV-B Data Many BERT-style language models include Wikipedia in the pre-training corpora for a wide coverage of topics in relatively high-quality writing. Korean Wikipedia currently ranks the 23rd by volume, and this is just 7.8% compared to English Wikipedia. To supplement training examples and the diversity of our corpus, we also use the text from NamuWiki444https://en.wikipedia.org/wiki/Namuwiki, which is another Korean online encyclopedia that contains more subjective opinions covering a variety of topics and writing styles. #### IV-B1 Pretraining corpora our pretraining corpora include the following. * • Web News: all articles from 8 major newspapers of Korea accross the topics including politics, social, economics, culture, IT, opinion, and sports from January 1, 2007 to December 31, 2019. * • Korean Wikipedia: 490,220 documents crawled in October, 2019. * • NamuWiki: 740,094 documents crawled in December, 2019. * • Book corpus: plots and editorial reviews about all Korean books published in 2010 to December 31, 2019555http://book.interpark.com/ #### IV-B2 Text preprocessing We have preprocessed our text data in the following manner. First, we remove all meta-tags such as the date of writing and name(s) of the author(s) in newspapers appearing in the beginning and at the end of each article. We think that the meta-tags do not contain any contextual or semantic information essential for NLU tasks. We also adjust the proportion of categories making up the news corpus in order to avoid topical bias of the examples. We tokenize the corpora into subwords using SentencePiece tokenizer [29] like ALBERT to construct vocabulary of a size 32k. We mask randomly sampled 15% of the words using the whole word masking strategy recently introduced by BERT. After cleaning and regularizing text, we obtain 43GB text with 325 million sentences, which are equivalent to 4.4 billion words or 18 billion characters. ### IV-C Compatibility of Word Order Prediction (WOP) We have performed ablation experiments with and without WOP to empirically observe its compatibility with the MLM and SOP objectives by pretraining for 125K steps, which is the half of the entire pre-training. A critical decision to introduce new noise via WOP is how many training examples should entail the additional noising process as well as how many tokens should be shuffled inside a sentence. We sample batches to contain re-ordered tokens proportionally from 30 to 100%. We have observed that about 30-50% shuffling achieves a good performance for most cases. Results are averaged over 10 different seeds and summarized in Table I. We set up three combinations of the pretraining objectives to compare against one another in the downstream evaluations to highlight the effect of WOP. We also observe the intrinsic performance of each objective. In the WOP and MLM combination, we configure the portion of corrupted examples to 30% for the WOP objective. From the result averaged over 10 different seeds in Table II, WOP hardly hurts the performance of MLM or SOP. The accuracy of MLM and WOP tasks has improved in case of leaving the SOP objective out. We believe that the best usage for WOP is not to disturb other intrinsic tasks for pretraining. WOP should be added by carefully observing the performance of other objectives on different WOP configurations. As expected, the deletion of SOP has caused a degradation more than 3% in the downstream performances of semantic textual similarity (8,628 examples) and paraphrase detection (7,576 examples). These two tasks are relatively small data experiments. Surprisingly, the performance of KorNLI is better without SOP because NLI tasks depend on inter-sentence coherence. Note that KorNLI is a much larger dataset (950,354 examples) compared to the semantic textual similarity and paraphrase detection datasets. Combining the two denoising objectives MLM and WOP seems to alleviate the performance degradation for a classification task with multi-sentence input. TABLE I: Experimental results on downstream tasks according to different portion of word order prediction tasks Portion of WOP | KorNLI | KorSTS | NSMC | PD | NER | KorQuAD1.0 ---|---|---|---|---|---|--- acc | spearman | acc | acc | acc | f1 100 % | 76.8 | 74.8 | 88.3 | 92.3 | 80.6 | 89.4 50 % | 76.4 | 76.6 | 88.3 | 92.7 | 81.2 | 89.3 30 % | 76.6 | 75.4 | 88.4 | 93.2 | 80.7 | 89.8 TABLE II: Experimental results on Downstream task performance comparing between different combination of pretraining objectives Objectives | MLM | SOP | WOP | KorNLI | KorSTS | NSMC | PD | NER | KorQuAD1.0 ---|---|---|---|---|---|---|---|---|--- acc | acc | acc | acc | spearman | acc | acc | acc | f1 MLM + SOP | 35.3 | 79.8 | - | 76.4 | 75.6 | 88.6 | 92.9 | 80.7 | 89.5 MLM + SOP + WOP | 35.1 | 79.1 | 80.7 | 76.9 | 76.6 | 88.4 | 93.2 | 81.2 | 89.8 MLM + WOP | 35.6 | - | 84.0 | 76.8 | 73.3 | 88.5 | 92.3 | 81.0 | 89.3 TABLE III: Experimental results on downstream tasks and model parameters Model | Params | Speedup | KorNLI | KorSTS | NSMC | PD | NER | KorQuAD1.0 | Avg. ---|---|---|---|---|---|---|---|---|--- acc | spearman | acc | acc | acc | f1 Multilingual BERT | 172M | 1.0 | 76.8 | 77.8 | 87.5 | 91.1 | 80.3 | 86.5 | 83.3 XLM-R | 270M | 0.5x | 80.0 | 79.4 | 90.1 | 92.6 | 83.9 | 92.3 | 86.4 KoBERT | 92M | 1.2x | 78.3 | 79.2 | 90.1 | 91.1 | 82.1 | 90.3 | 85.2 ETRI BERT | 110M | - | 79.5 | 80.5 | 88.8 | 93.9 | 82.5 | 94.1 | 86.6 KoreALBERT Base | 12M | 5.7x | 79.7 | 81.2 | 89.6 | 93.8 | 82.3 | 92.6 | 86.5 KoreALBERT Large | 18M | 1.3x | 81.1 | 82.1 | 89.7 | 94.1 | 83.7 | 94.5 | 87.5 ### IV-D Evaluation We fine-tune KoreALBERT for downstream performance evaluations. For comparison, we consider other pretrained BERT-base language models available off-the-shelf. #### IV-D1 Fine-tuning In addition to our KoreALBERT, we have downloaded pretrained models available online: multilingual BERT666https://github.com/google-research/bert, XLM-R from Facebook AI Research777https://github.com/facebookresearch/XLM, KoBERT888https://github.com/SKTBrain/KoBERT, and ETRI BERT999http://aiopen.etri.re.kr/service_dataset.php. We optimize respective hyperparameters for each pretrained model before measuring the best and average scores for each model. For all models, we use a batch size of 64 or 128 and from 3 to 5 epochs with a learning rate from $2.0\times 10^{-5}$ to $5.0\times 10^{-5}$ and a max-sequence length from 128 to 512. For NER task, we have found out that longer training epochs tend to work better and fine- tuned up to 7 epochs. #### IV-D2 Downstream Tasks We consider six downstream NLP tasks detailed below. * • KorNLI: Korean NLU Dataset includes two downstream tasks. In Korean Natural Language Inference (KorNLI) [30], the input is a pair of sentences, a premise and a hypothesis. The fine-tuned model should predict their relationship in one of the three possible labels: entailment, contradiction, and neutral. KorNLI has a total of 950,354 examples. * • KorSTS: the second task from Korean NLU is semantic textual similarity (STS) for Korean language. STS requires to predict how semantically similar the two input sentences are on a 0 (dissimilar) to 5 (equivalent) scale. There are 8,628 KorSTS examples in the Korean NLU dataset. * • Sentiment analysis: we use Naver Sentiment Movie Corpus,101010https://github.com/e9t/nsmc (NSMC) the biggest Korean movie review dataset, which is collected by the same method that the massive movie review dataset [31] proposes. NSMC consists of 200k reviews of shorter than 140 characters that are labeled with human annotations of sentiment. * • Paraphrase detection (PD): a PD model predicts whether or not a pair of sentences are semantically equivalent. The dataset we consider contains 7,576 examples from a publicly available github repository.111111https://github.com/songys/Question_pair * • Extractive machine reading comprehension (EMRC): EMRC takes in much longer text sequences as an input compared to other tasks. The EMRC model needs to extract the start and end indices inside a paragraph containing the answer of a question. KorQuAD 1.0 [32] is a Korean dataset for machine reading comprehension, which is similar to SQuAD 1.0 [7]. Having exactly the same format as SQuAD, KorQuAD 1.0 comprises 60,407 question-answer pairs. * • Named entity recognition (NER): NER distinguishes a real-world object such as a person, organization, and place (location) from documents. We use the NER corpus121212http://air.changwon.ac.kr/?page_id=10 constructed by Naver Corp. and Changwon University in South Korea. The corpus has 14 different types of entities with attached tags B/I/-, denoting multi- or single-word entities as described in Table IV. TABLE IV: Proportion of the type of entities of NER dataset. Category | Tag | Amount ---|---|--- NUMBER | NUM | 64,876 CIVILIZATION | CVL | 60,918 PERSON | PER | 48,321 ORGANIZATION | ORG | 45,550 DATE | DAT | 33,944 TERM | TRM | 22,070 LOCATION | LOC | 21,095 EVENT | EVT | 17,430 ANIMAL | ANM | 6,544 ARTIFACTS_WORKS | AFW | 6,069 TIME | TIM | 4,337 FIELD | FLD | 2,386 PLANT | PLT | 267 MATERIAL | MAT | 252 ### IV-E Discussion As indicated in Table III, KoreALBERT consistently outperforms M-BERT over all downstream NLU tasks considered. While KoreABLERT has the smallest number of model parameters among all monolingual and multilingual language models compared in this paper, it achieves better results in almost all downstream evaluations. The advantage of having fewer computations of KoreALBERT makes its base model about 5.7 faster than M-BERT and its large model 2.2 faster than XLM-R base at training time. In NSMC and NER, which are single-sentence classification tasks, KoreALBERT is subpar against XLM-R and KoBERT. For NSMC, KoreALBERT-large cannot produce more discriminnative result than the base model. We suspect the main reason for the performance drop being lack of covering the colloquial usage of words and phrases in our pretraining corpora that mostly consists of more formal style of writings such as news articles and wikipedia. Examples in NSMC seem to use much colloquialism. Also, XLM-R has shown a very good performance on the NER task. Such result is due to the fact that NER does not require much high-level language understanding like multi-sentence discourse coherence. ## V Conclusion We have introduced KoreALBERT, a pre-trained monolingual ALBERT model for Korean language understanding. We have described the details about training KoreALBERT. In particular, we have proposed a word order prediction loss, a new training objective, which is compatible with the original MLM and SOP objectives of ALBERT. KoreALBERT consistently outperforms multi and monolingual baselines on 6 downstream NLP tasks while having much fewer parameters. In our future work, we plan to experiment more comprehensively with the KoreALBERT WOP loss: i) replace token-level switching with word-level switching to improve the difficulty of label prediction; ii) use dynamic token shuffling with varying amount of tokens to be shuffled instead of fixed proportion. We also plan to investigate how well the proposed WOP loss works with non-agglutinative languages like English. ## References * [1] I. Sutskever, O. Vinyals, and Q. V. Le, “Sequence to sequence learning with neural networks,” _CoRR_ , vol. abs/1409.3215, 2014. [Online]. Available: http://arxiv.org/abs/1409.3215 * [2] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, and I. Polosukhin, “Attention is all you need,” in _Advances in Neural Information Processing Systems 30_ , I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, Eds. Curran Associates, Inc., 2017, pp. 5998–6008. [Online]. Available: http://papers.nips.cc/paper/7181-attention-is-all-you-need.pdf * [3] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training of Deep Bidirectional Transformers for Language Understanding,” _arXiv preprint arXiv:1810.04805_ , 2018. * [4] W. L. Taylor, “Cloze Procedure: a New Tool for Measuring Readability,” _Journalism Quarterly_ , vol. 30, no. 4, pp. 415–433, 1953. * [5] A. Wang, A. Singh, J. Michael, F. Hill, O. Levy, and S. Bowman, “Glue: A multi-task benchmark and analysis platform for natural language understanding,” _Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP_ , 2018. [Online]. Available: http://dx.doi.org/10.18653/v1/w18-5446 * [6] A. Williams, N. Nangia, and S. Bowman, “A broad-coverage challenge corpus for sentence understanding through inference,” _Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers)_ , 2018. [Online]. Available: http://dx.doi.org/10.18653/v1/N18-1101 * [7] P. Rajpurkar, J. Zhang, K. Lopyrev, and P. Liang, “Squad: 100, 000+ questions for machine comprehension of text,” _CoRR_ , vol. abs/1606.05250, 2016. [Online]. Available: http://arxiv.org/abs/1606.05250 * [8] P. Rajpurkar, R. Jia, and P. Liang, “Know what you don’t know: Unanswerable questions for squad,” _Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers)_ , 2018. [Online]. Available: http://dx.doi.org/10.18653/v1/P18-2124 * [9] E. F. Tjong Kim Sang and F. De Meulder, “Introduction to the conll-2003 shared task,” _Proceedings of the seventh conference on Natural language learning at HLT-NAACL 2003 -_ , 2003. [Online]. Available: http://dx.doi.org/10.3115/1119176.1119195 * [10] G. Lai, Q. Xie, H. Liu, Y. Yang, and E. Hovy, “RACE: Large-scale ReAding comprehension dataset from examinations,” in _Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing_. Copenhagen, Denmark: Association for Computational Linguistics, Sep. 2017, pp. 785–794. [Online]. Available: https://www.aclweb.org/anthology/D17-1082 * [11] Z. Lan, M. Chen, S. Goodman, K. Gimpel, P. Sharma, and R. Soricut, “Albert: A Lite BERT for Self-supervised Learning of Language Representations,” _arXiv preprint arXiv:1909.11942_ , 2019. * [12] Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, and V. Stoyanov, “Roberta: A robustly optimized BERT pretraining approach,” _CoRR_ , vol. abs/1907.11692, 2019. [Online]. Available: http://arxiv.org/abs/1907.11692 * [13] M. Joshi, D. Chen, Y. Liu, D. S. Weld, L. Zettlemoyer, and O. Levy, “Spanbert: Improving pre-training by representing and predicting spans,” _CoRR_ , vol. abs/1907.10529, 2019. [Online]. Available: http://arxiv.org/abs/1907.10529 * [14] Y. Cui, W. Che, T. Liu, B. Qin, Z. Yang, S. Wang, and G. Hu, “Pre-training with whole word masking for chinese BERT,” _CoRR_ , vol. abs/1906.08101, 2019. [Online]. Available: http://arxiv.org/abs/1906.08101 * [15] Z. Yang, Z. Dai, Y. Yang, J. G. Carbonell, R. Salakhutdinov, and Q. V. Le, “Xlnet: Generalized autoregressive pretraining for language understanding,” _CoRR_ , vol. abs/1906.08237, 2019. [Online]. Available: http://arxiv.org/abs/1906.08237 * [16] K. K, Z. Wang, S. Mayhew, and D. Roth, “Cross-lingual ability of multilingual bert: An empirical study,” 2019. * [17] A. Conneau, K. Khandelwal, N. Goyal, V. Chaudhary, G. Wenzek, F. Guzmán, E. Grave, M. Ott, L. Zettlemoyer, and V. Stoyanov, “Unsupervised cross-lingual representation learning at scale,” 2019. * [18] H. Le, L. Vial, J. Frej, V. Segonne, M. Coavoux, B. Lecouteux, A. Allauzen, B. Crabbé, L. Besacier, and D. Schwab, “Flaubert: Unsupervised language model pre-training for french,” 2019. * [19] L. Martin, B. Muller, P. J. O. Suárez, Y. Dupont, L. Romary, Éric Villemonte de la Clergerie, D. Seddah, and B. Sagot, “Camembert: a tasty french language model,” 2019. * [20] M. Polignano, P. Basile, M. de Gemmis, G. Semeraro, and V. Basile, “Alberto: Italian bert language understanding model for nlp challenging tasks based on tweets,” 11 2019. * [21] W. Vries, A. Cranenburgh, A. Bisazza, T. Caselli, G. van Noord, and M. Nissim, “Bertje: A dutch bert model,” 12 2019. * [22] A. Virtanen, J. Kanerva, R. Ilo, J. Luoma, J. Luotolahti, T. Salakoski, F. Ginter, and S. Pyysalo, “Multilingual is not enough: Bert for finnish,” 12 2019. * [23] S. Park, J. Byun, S. Baek, Y. Cho, and A. Oh, “Subword-level word vector representations for korean,” 01 2018, pp. 2429–2438. * [24] M. Lewis, Y. Liu, N. Goyal, M. Ghazvininejad, A. Mohamed, O. Levy, V. Stoyanov, and L. Zettlemoyer, “Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension,” 2019. * [25] D. Hendrycks and K. Gimpel, “Bridging nonlinearities and stochastic regularizers with gaussian error linear units,” _CoRR_ , vol. abs/1606.08415, 2016. [Online]. Available: http://arxiv.org/abs/1606.08415 * [26] J. J. Song, “The korean language:structure, use and context,” _Routledge_ , 2006. * [27] Y. You, J. Li, J. Hseu, X. Song, J. Demmel, and C. Hsieh, “Reducing BERT pre-training time from 3 days to 76 minutes,” _CoRR_ , vol. abs/1904.00962, 2019. [Online]. Available: http://arxiv.org/abs/1904.00962 * [28] T. Wolf, L. Debut, V. Sanh, J. Chaumond, C. Delangue, A. Moi, P. Cistac, T. Rault, R. Louf, M. Funtowicz, and J. Brew, “Huggingface’s transformers: State-of-the-art natural language processing,” _ArXiv_ , vol. abs/1910.03771, 2019. * [29] T. Kudo and J. Richardson, “Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing,” _CoRR_ , vol. abs/1808.06226, 2018. [Online]. Available: http://arxiv.org/abs/1808.06226 * [30] J. Ham, Y. J. Choe, K. Park, I. Choi, and H. Soh, “Kornli and korsts: New benchmark datasets for korean natural language understanding,” _arXiv preprint arXiv:2004.03289_ , 2020. * [31] A. L. Maas, R. E. Daly, P. T. Pham, D. Huang, A. Y. Ng, and C. Potts, “Learning word vectors for sentiment analysis,” in _Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies_. Portland, Oregon, USA: Association for Computational Linguistics, Jun. 2011, pp. 142–150. [Online]. Available: https://www.aclweb.org/anthology/P11-1015 * [32] M. K. L. Seungyoung Lim, “KorQuAD: Korean QA Dataset for Machine Comprehension,” _Journal of Computing Science and Engineering_ , vol. , pp. 539–541, 2018. [Online]. Available: http://www.dbpia.co.kr/journal/articleDetail?nodeId=NODE07613668
§ INTRODUCTION With the increase of multidimensional data availability and modern computing power, statistical models for spatial and spatio-temporal data are developing at a rapid pace. Hence, there is a need for stable and reliable, yet updated and efficient, software packages. In this section, we briefly discuss multidimensional data in climate and environmental studies as well as statistical software for space-time data. §.§ Multidimensional data Large multidimensional data sets often arise when climate and environmental phenomena are observed at the global scale over extended periods. In climate studies, relevant physical variables are observed on a three-dimensional (3D) spherical shell (the atmosphere) while time is the fourth dimension. For instance, measurements are obtained by radiosondes flying from ground level up to the stratosphere [Fassò et al., 2014], by interferometric sensors aboard satellites [Finazzi et al., 2018] or by laser-based methods, such as Light Detection and Ranging (LIDAR) [Negri et al., 2018]. In this context, statistical modelling of multidimensional data requires describing and exploiting the spatio-temporal correlation of the underlying phenomenon or data-generating process. This is done using explanatory variables and multidimensional latent variables with covariance functions defined over a convenient spatio-temporal support. When considering 3D$\times$T data (4D for brevity), covariance functions defined over the 4D support may be adopted. However, these covariance functions often have a complex form [Porcu et al., 2018]. Moreover, when estimating the model parameters or making inferences, very large covariance matrices (though they may be sparse) are implied. In large climate and environmental applications, 4D data are rarely collected at high frequency in all spatial and temporal dimensions. Often, only one dimension is sampled at high frequency while the remaining dimensions are sampled sparsely. Radiosonde data, for instance, are sparse over the Earth's sphere, but they are dense along the vertical dimension, providing atmospheric profiles. This suggests that handling all spatial dimensions equally (e.g. using a 3D covariance function) may not be the best option from a modelling or computational perspective, and a data reduction technique may be useful instead. In this paper, the functional data analysis (FDA) approach [Ramsay and Silverman, 2007] is adopted to model the relationship between measurements along the profile, while the remaining dimensions are handled following the classic spatio-temporal data modelling approach using only 2D spatial covariance functions. §.§ Statistical software Various software programmes are available for considering data on a plane or in a two-dimensional (2D) Euclidean space. The choice is more restricted when considering multidimensional or non-Euclidean spaces arising from atmospheric or remote sensing spatio-temporal data observed on the surface of a sphere and over time. For example, Figure <ref> depicts the spatial locations of measurements collected globally in a single day through radio sounding, as discussed in Section <ref>. Space is three-dimensional, and measurements are repeated over time at the same spatial locations over the Earth's surface but at different pressure values. Radio sounding data example. Each dot represents the spatial location of a measurement taken by a radiosonde. Dots of the same colour belong to the same radiosonde. (Pressure axis not in scale). The spBayes package [Finley et al., 2015] handles large spatio-temporal data sets, but space is only 2D. The documentation of the spacetime [Pebesma, 2012] and gstat [Pebesma and Heuvelink, 2016] packages does not explicitly address the multidimensional case, but, according to [Gasch et al., 2015], both packages have some capabilities to handle the 3D$\times$T. However, we want to avoid working with 3D spatial covariance functions or sample spatio-temporal variograms. Fixed rank kriging ([Cressie and Johannesson, 2008] [Cressie and Johannesson, 2008]) implemented in the R package FRK [Zammit-Mangion, 2018] handles spatial and spatio-temporal data both on the Euclidean plane and on the surface of the sphere. FRK implements a set of tools for data gridding and basis function computation, resulting in efficient dimension reduction, allowing it to handle large satellite data sets [Cressie, 2018]. It is based on a spatio-temporal random effects (SRE) model estimated by the expectation-maximisation (EM) algorithm. Recent extensions to FRK include the use of multi-resolution basis functions [Tzeng and Huang, 2018]. A second package based on SRE and the EM algorithm is D-STEM v1 [Finazzi and Fassò, 2014]. This package implements an efficient state-space approach for handling the temporal dimension and a heterotopic multivariate response approach that is useful when correlating heterogeneous networks [Fassò and Finazzi, 2011, Calculli et al., 2015]. D-STEM v1 has been successfully used in various medium-to-large applications, proving that the EM algorithm implementation, being mainly based on closed-form iterations, is quite stable. These applications include air quality assessment in the metropolitan areas of Milan, Teheran and Beijing [Fassò, 2013, Taghavi-Shahri et al., 2019, Wan et al., 2020]; multivariate spatio-temporal modelling at the country and continental levels in Europe [Finazzi et al., 2013, Fassò et al., 2016]; time series clustering [Finazzi et al., 2015]; emulators of atmospheric dispersion modelling systems [Finazzi et al., 2019]; and near real-time prediction of earthquake parameters [Finazzi, 2020]. A brief, non-exhaustive list of other models and/or software packages for advanced spatial data modelling is presented below, according to the principal technique, allowing the handling of large data sets. In general, these techniques aim at avoiding the Cholesky decomposition of large and dense covariance matrices. Some approaches, including FRK and D-STEM v1, leverage sparse variance–covariance matrices. Others exploit the sparsity of the precision matrix, thanks to a spatial Markovian assumption. This class includes the R packages LatticeKrig [Nychka et al., 2015, Nychka et al., 2016], INLA [Blangiardo et al., 2013, Lindgren and Rue, 2015, Bivand et al., 2015, Rue et al., 2014] and the multi-resolution approximation approach of [Katzfuss, 2017], which uses the predictive process and the state space representation [Jurek and Katzfuss, 2018] to model spatio-temporal data. Low-rank models are another popular approach used by spBayes. Finally, the R package laGP [Gramacy, 2016], based on a machine learning approach, implements an efficient nearest neighbour prediction-oriented method. [Heaton et al., 2018] develop an interesting spatial prediction competition considering a large data set and involving the above-mentioned approaches. We observe that, although some of the software packages mentioned above consider both space and time, to the best of our knowledge, none of them handles a spatio-temporal FDA approach for data sets of the kind discussed in <ref>. In this paper, we present D-STEM v2, which is a MATLAB package, extending D-STEM v1. The new version introduces modelling of functional data indexed over space and time. Moreover, new complexity reduction techniques have been added for both model estimation and dynamic mapping, which are especially useful for large data sets. The rest of the paper is organised as follows. Section <ref> introduces the methodology adopted in this paper and, in particular, the data modelling approach and the complexity-reduction techniques. Section <ref> describes the D-STEM v2 software in terms of the MATLAB classes used to define the data structure, model fitting and diagnostics and kriging. This is followed by an illustration of the software use through two case studies. The first one, discussed in Section <ref>, considers high-frequency spatio-temporal ozone data in Beijing. The second one, in Section <ref>, considers modelling of global atmospheric temperature profiles and exploits the complexity-reduction capabilities of the new package. Finally, concluding remarks are provided in Section <ref>. § METHODOLOGY This section discusses the methodology behind the modelling and the complexity-reduction techniques implemented in D-STEM v2 when dealing with functional space-time data sets. Moreover, model estimation, validation and dynamic kriging are briefly discussed. §.§ Model equations Let $\bm{s}=(s_{lat},s_{lon})^{ \top }$ be a generic spatial location on the Earth's sphere, $\mathbb{S}^{2}$, and $t\in \mathbb{N} $ a discrete time index. It is assumed that the function of interest, $f\left( \bm{s},h,t\right )$, with domain $\mathcal{H}=\left[ h_{1},h_{2}\right] \subset \mathbb{R}$, can be observed at any $\left( \bm{s},t\right) $ and $h\in\mathcal{H}$ through noisy measurements $y(\bm{s},h,t)$ according to the following model: \begin{align} y(\bm{s},h,t) & =f\left( \bm{s},h,t\right) +\varepsilon f\left( \bm{s},h,t\right) & =\bm{x}(\bm{s},h,t)^{ \top }\bm{\beta}\left( h\right) +\bm{\phi}(h)^{\top}\bm{z}(\bm{s}% \bm{z}(\bm{s},t) & =\bm{G}\bm{z}(\bm{s}% ,t-1)+\bm{\eta}(\bm{s},t). \label{eq:model_line3}% \end{align} This model is referred to as the functional hidden dynamic geostatistical model (f-HDGM). In Equation (<ref>), $\varepsilon$ is a zero-mean Gaussian measurement error independent in space and time with functional variance $\sigma_{\varepsilon}^{2}\left( h\right) $, implying that $\varepsilon$ is heteroskedastic across the domain $\mathcal{H}$. The variance is modelled as \[ \log(\sigma_{\varepsilon}^{2}\left( h\right) )=\bm{\phi}(h)^{\top \] where $\bm{\phi}(h)$ is a $p\times1$ vector of basis functions evaluated at $h$, while $\bm{c}_{\varepsilon}$ is a vector of coefficients to be estimated. In Equation (<ref>), $\bm{x}(\bm{s},h,t)$ is a $b \times 1$ vector of covariates while $\bm{\beta}\left( h\right) =(\beta_{1}(h),...,\beta_{b}(h))^{\top}$ is the vector of functional parameters modelled as \[ \beta_{j}(h)=\bm{\phi}(h)^{\top}\bm{c}_{\beta,j}, j=1,...,b, \] and $\bm{c}_{\beta}=\left( \bm{c}_{\beta,1}^{\top},...,\bm{c}_{\beta,b}^{\top}\right) ^{\top}$ is a $pb\times1$ vector of coefficients that needs to be estimated. Additionally, $\bm{z}(\bm{s},t)$ is a $p\times1$ latent space-time variable with Markovian dynamics given in Equation (<ref>). The matrix $\bm{G}$ is a diagonal transition matrix with diagonal elements in the $p \times 1$ vector $\bm{g}$. The innovation vector $\bm{\eta}$ is obtained from a multivariate Gaussian process that is independent in time but correlated across space with matrix spatial covariance function given by \[ \bm{\Gamma}(\bm{s},\bm{s}^{\prime};\bm{\theta})=diag\left( \rho(\bm{s},\bm{s}^{\prime};\bm{\theta}_{p})\right), \] where $\bm{v}=\left( v_{1},...,v_{p}\right) ^{\top}$ is a vector of variances and $\rho(\bm{s},\bm{s}^{\prime};\bm{\theta}_{j})$ is a valid spatial correlation function for locations $\bm{s},\bm{s}^{\prime}\in\mathbb{S}^{2}$, parametrised by $\bm{\theta}_{j}$, and $\bm{\theta}=(\bm{\theta}_{1},...,\bm{\theta}_{p})^{\top}$. The unknown model parameter vector is given by $\bm{\psi}=\left( \bm{c}_{\varepsilon}^{\top},\bm{c}_{\beta}^{\top},\bm{g}^{\top},\bm{v}^{\top},\bm{\theta}^{\top}\right) ^{\top} $. Note that, in order to ease the notation, the same $p$-dimensional basis functions $\bm{\phi}(h)$ are used to model $\sigma_{\varepsilon}^{2}$, $\beta_{j}$ and $\bm{\phi}(h)^{\top}\bm{z}(\bm{s},t)$ in Equations (<ref>)-(<ref>). In practice, D-STEM v2 allows one to specify a different number of basis functions for each model component. Also note that $\varepsilon$ is not a pure measurement error since it also accounts for model misspecification. Finally, the covariates $\bm{x}(\bm{s},h,t)$ are assumed to be known without error for any $\bm{s}$, $h$ and $t$, and thus they do not need a basis function representation. §.§ Basis function choice Choosing basis functions essentially means choosing the basis type and the number of basis functions. D-STEM v2 currently supports Fourier bases and B-spline bases. The former guarantee that the function is periodic in the domain $\mathcal{H}$, while the latter are not (in general) periodic but have higher flexibility in describing functions with a complex shape. Whichever basis function type is adopted, the number $p$ of basis functions must be fixed before model estimation. Usually, a high $p$ implies a better model $R^2$, but over-fitting may be an issue. Moreover, special care must be taken when choosing the number of basis functions for $\bm{\phi}(h)^{\top}\bm{z}(\bm{s},t)$. The classic FDA approach suggests fixing a high number of basis functions and adopting penalisation to avoid over-fitting. In our context, this is not viable since the covariance matrices involved in model estimation have dimension $n^3p^3 \times n^3p^3$. Since $n$ is usually large, a large $p$ would make model estimation unfeasible, especially if the number of time points $T$ is also high. When using B-spline basis, a small $p$ implies that the location of knots along the domain $\mathcal{H}$ also matters and may affect the model fitting performance. Ideally, $p$ and knot locations are chosen using a model validation technique (see <ref>) by trying different combinations of $p$ and knot locations. If, due to time constraints, this is not possible, equally spaced knots are a convenient option. §.§ Model estimation The estimation of $\bm{\psi}$ and the latent space-time variable $\bm{z}(\bm{s},t)$ is based on the maximum likelihood approach considering profile data observed at spatial locations $\mathcal{S}=\{\bm{s}_{i},i=1,...,n\}$ and time points $t=1,...,T$. At a specific location $\bm{s}_{i}$ and time $t$, $q_{i,t}$ measurements are taken at points $\bm{h}_{\bm{s}_{i},t}=\left( h_{i,1,t},...,h_{i,q_{i,t},t}\right) ^{\top}$ and collected in the vector \[ \bm{y}_{\bm{s}_{i},t}=(y(\bm{s}_{i},h_{i,1,t},t),...,y(\bm{s}_{i},h_{i,q_{i,t},t},t))^{\top}, \] here called the observed profile. Although D-STEM v2 allows for varying $q_{i,t}$, for ease of notation, it is assumed here that all profiles include exactly $q$ measurements, although $\bm{h}_{\bm{s}_{i},t}$ may be different across profiles. Profiles observed at time $t$ across spatial locations $\mathcal{S}$ are then stored in the $nq\times1$ vector $\bm{y}_{t}=(\bm{y}_{s_{1},t}^{\top},...,\bm{y}_{s_{n},t}^{\top})^{\top}$. Applying model (<ref>)-(<ref>) to the defined data above, we have the following matrix representation: \begin{align*} \bm{y}_{t} & =\tilde{\bm{X}}_{t}\bm{c}_{\bm{\beta}}+\bm{\Phi}_{\bm{z},t}\bm{z}% \bm{z}_{t} & =\tilde{\bm{G}}\bm{z}_{t-1}+\bm{\eta}_{t}, \end{align*} where $\tilde{\bm{X}}_{t}=\bm{X}_{t}\bm{\Phi}_{\bm{\beta},t}$ is a $nq\times bp$ matrix, with $\bm{X}_{t}$ the matrix of covariates and $\bm{\Phi}_{\bm{\beta},t}$ the basis matrix for $\bm{\beta}$. $\bm{\Phi}_{\bm{z},t}$ is the $nq\times np$ basis matrix for the latent $np\times1$ vector }(\bm{s}_{1},t)^{\top},...,\bm{\eta}(\bm{s}_{n},t)^{\top})^{\top}$ is the $np\times1$ innovation vector, while $\bm{\varepsilon}_{t}\ $ is the $nq\times1$ vector of measurement errors. Additionally, $\tilde{\bm{G}}= \bm{I}_{n} \otimes \bm{G}$ is the $np\times np$ diagonal transition matrix. The complete-data likelihood function $L(\bm{\psi};\bm{Y},\bm{Z})$ can be written as \[ \] where $\bm{Y}=\left( \bm{y}_{1},...,\bm{y}_{T}\right) $, $\bm{Z}=\left( \bm{z}_{0},\bm{z}_{1},...,\bm{z}_{T}\right) $, $\bm{\psi}_{\bm{z}}=\left( \bm{g}^{\top},\bm{v}^{\top},\bm{\theta}^{\top}\right) ^{\top} $, $\bm{\psi}_{\bm{y}}=\left( \bm{c}_{\varepsilon}^{\top},\bm{c}_{\beta}^{\top}\right) ^{\top} $, and $\bm{z}_{0}$ is the Gaussian initial vector with parameter $\bm{\psi}_{\bm{z}_{0}}$. Maximum likelihood estimation is based on an extension of the EM algorithm detailed in [Calculli et al., 2015]. The model parameter set $\bm{\psi}$ is initialised with starting values $\bm{\psi}^{\left\langle 0\right\rangle }$ and then updated at each iteration $\iota$ of the EM algorithm. The algorithm terminates if any of the following conditions is satisfied: \[ \max_{l}\left\vert \psi _{l}^{\left\langle \iota \right\rangle }-\psi _{l}^{\left\langle \iota -1\right\rangle }\right\vert /\left\vert \psi _{l}^{\left\langle \iota \right\rangle }\right\vert <\epsilon _{1} \] \[ \left\vert L(\bm{\psi}^{\left\langle \iota\right\rangle };\bm{Y}% )-L(\bm{\psi}^{\left\langle \iota-1\right\rangle };\bm{Y})\right\vert /\left\vert L(\bm{\psi}^{\left\langle \iota\right\rangle };\bm{Y})\right\vert \[ \iota>\iota^{\ast}, \] where $\psi _{l}^{\left\langle \iota \right\rangle }$ is the generic element of $\bm{\psi}^{\left\langle \iota \right\rangle }$ at the $\iota\text{-}th$ iteration, $L(\bm{\psi}^{\left\langle \iota\right\rangle };\bm{Y})$ is the observed-data likelihood function evaluated at $\bm{\psi}^{\left\langle \iota\right\rangle }$, $0<\epsilon_1\ll1$ and $0<\epsilon_2\ll1$ are small positive numbers (e.g. $10^{-4}% $), while $\iota^{\ast}$ is a user-defined positive integer number (e.g. $100$) to limit the iterations in the case of convergence failure of the EM algorithm. Note that $\mathcal{S}$ is not time-varying, which means that spatial locations are fixed. This could be a limit in applications where spatial locations change for each $t$. On the other hand, missing profiles are allowed; that is, $\bm{y}_{\bm{s}_{i},t}$ may be a vector of $q$ missing values at some $t$. In the extreme case, a given spatial location $\bm{s}_{i}$ has only one profile over the entire period (if all the profiles are missing, the spatial location can be dropped from the data set). <cit.> explains how the likelihood function of a state-space model changes in the case of a missing observation vector and how the EM estimation formulas are derived. Missing data handling in D-STEM v2 is based on the same approach. §.§ Partitioning At each iteration of the EM algorithm, the computational complexity of the E-step is $O\left( Tn^{3}p^{3}\right) $, which may be unfeasible if $n$ is large. When necessary, D-STEM v2 allows one to use a partitioning approach [Stein, 2013] for model The spatial locations $\mathcal{S}$ are divided into $k$ partitions, and $\bm{z}_{t}$ is partitioned conformably, namely, $\bm{z}_{t}=\left( \bm{z}_{t}^{(1)\top},...,\bm{z}% _{t}^{(k)\top}\right) ^{\top}$. Hence, the likelihood function becomes \[ L\left( \bm{\psi}_{\bm{y}};\bm{y}_{t}\mid\bm{z}_{t}\right) \cdot% L\left( \bm{\psi}_{\bm{z}_{0}};\bm{z}_{0}^{(j)}\right) \cdot% L\left( \bm{\psi}_{\bm{z}};\bm{z}_{t}^{(j)}\mid\bm{z}_{t-1}% \] From the EM algorithm point of view, this implies that the E-step is independently applied to each partition, possibly in parallel. When all partitions are equal in size, the computational complexity reduces to $\mathcal{O}\left( Tkr^{3}p^{3}\right)$, with $r$ as the partition size. Geographical partitioning, constructed aggregating proximal locations, is a natural choice for environmental applications. Given the number of partitions $k$, the k-means algorithm applied to spatial coordinates provides a geographical partitioning of $\mathcal{S}$. However, the number of points in each partition is not controlled, and a heterogeneous partitioning may arise. If some subsets are very large and others are small, the reduction in computational complexity given above is far from being achieved. This can easily happen, for example, when $\mathcal{S}$ is a global network constrained by continent shapes. For this reason, D-STEM v2 provides a heuristically modified k-means algorithm that encourages partitions with similar numbers of elements. The algorithm optimises the following objective function: \begin{equation} \sum_{j=1}^{k}\sum_{\bm{s}\in\mathcal{S}_{j}}d\left( \bm{s},\bm{c}_{j}\right) +\lambda\sum_{j=1}^{k}\left( r_{j}-\frac{n}{k}\right) ^{2}, \label{eq:k-means}% \end{equation} where $\lambda\ge0$, $\mathcal{S}_{j} \subset \mathcal{S}$ is the set of coordinates in the $j\text{-}th$ partition, $d$ is the geodesic distance on the sphere $\mathbb{S}^{2}$ and $\bm{c}_{j}$ and $r_{j}$ are the centroid and the number of elements in the $j\text{-}th$ partition, respectively. The second term in (<ref>) accounts for the variability of the partition sizes and acts as a penalisation for heterogeneous partitionings. Clearly, when $\lambda=0$, the above-mentioned objective function gives the classic k-means algorithm. For high values of $\lambda$, solutions with similarly sized partitions are favoured. Unfortunately, an optimality theory for this algorithm has not yet been developed, and the choice of $\lambda$ is left to the user. Nonetheless, it may be a useful tool to define a partitioning that is appropriate for the application at hand with regard to computing time and geographical properties. §.§ Variance-covariance matrix estimation The EM algorithm provides a point estimate of the parameter vector $\bm{\psi}$ but no uncertainty information. Building on <cit.>, D-STEM v2 estimates the variance–covariance matrix by means of the observed Fisher information matrix, $\mathbf{I}_{T}$, namely To understand its computational cost, note that the information matrix given above may be written as a sum: For large data sets, each matrix $\mathbf{i}_t$ may be expensive to compute, and the total computational cost is linear in $T$, provided missing data are evenly distributed in time. This results in a time-consuming task with a computational burden even higher than that for model estimation. For this reason, D-STEM v2 makes it possible to approximate $\hat{\Sigma}_{\bm{\psi},T}$ using a truncated information matrix, namely: \begin{equation} \tilde{\Sigma}_{\bm{\psi},t^*}=(\frac{T}{t^*}\mathbf{I}_{t^*})^{-1}, \label{eq:Fisher_approximated} \end{equation} which reduces the computational burden by a factor of $1-t^*/T$. $\tilde{\Sigma}_{\bm{\psi},t^*} \rightarrow \hat{\Sigma}_{\bm{\psi},T}$ for $t^* \rightarrow T$, the truncation time $t^*$ is chosen to control the approximation error in $\hat{\Sigma}_{\bm{\psi}}$. In particular, $t^*$ is the first integer such that \begin{equation} \frac{\left\Vert \tilde{\Sigma}_{\bm{\psi},t}-\tilde{\Sigma}_{\bm{\psi},t-1}% \right\Vert_{F} }{\left\Vert \tilde{\Sigma}_{\bm{\psi},t}\right\Vert_{F} }\leq \delta,\label{eq:varcov_approximated} \end{equation} where $\left\Vert { \cdot }\right\Vert_{F}$ is the Frobenius norm, and $\delta$ may be defined by the user. Generally speaking, the behaviour of $\hat{\Sigma}_{\bm{\psi},T}$ for large $T$ and, hence, the behaviour of $\tilde{\Sigma}_{\bm{\psi},t}$ relays on stationarity and ergodicity of the underlying stochastic process; see, for example, <cit.> and references therein. To have operative guidance for the user, let us assume first that no missing values are present, the information matrix is well-conditioned and the covariates have no isolated outliers or extreme trends. In this case, away from the borders $t\cong1$ and $t\cong T$, the observed conditional information $\mathbf{i}_t$ has a relatively smooth stochastic behaviour, and the approximation in (<ref>) is expected to be satisfactory at the level defined by $\delta$. Conversely, if some data are missing at time $t$, the information $\mathbf{i}_t$ is reduced accordingly. If the missing pattern is random over time, this is not an issue. But, in the unfavourable case with a high percentage of missing data mostly concentrated at the end the time series, $t\cong T$, the above approximation may over-estimate the information and under-estimate the variances of the parameter estimates. §.§ Dynamic kriging In this paper, dynamic kriging refers to evaluating the following \begin{align} \hat{f} \left( \bm{s},h,t\right) &= \mathbb{E}_{\hat{\bm{\psi}}}\left( f\left( \bm{s},h,t\right) \mid \bm{Y}\right), \label{eq:krig_exp}\\ \VAR\left( \hat{f} \left( \bm{s},h,t\right) \right) &= \mathbb{V}_{\hat{\bm{\psi}}}\left( f\left( \bm{s},h,t\right) \mid\bm{Y}\right), \label{eq:krig_var}% \end{align} for any $\bm{s}\in\mathbb{S}^{2}$, $h\in\mathcal{H}$ and $t=1,...,T$. A common approach is to map the kriging estimates on a regular pixelation $\mathcal{S}^{\ast}=\left\{ \bm{s}_{1}^{\ast},...,\bm{s}_{m}^{\ast}\right\} $. This may be a time-consuming task when $m$ and/or $n$ and/or $T$ are large. To tackle this problem, D-STEM v2 allows one to exploit a nearest-neighbour approach, where the conditioning term in Equations (<ref>) and (<ref>) is not $\bm{Y}$, but the data at the spatial locations $\mathcal{S}_{\sim j}$, where $\mathcal{S}_{\sim j}\subset\mathcal{S}$ is the set of the $\tilde{n}\ll n$ nearest spatial locations to $\bm{s}_{j}^{\ast}$. The use of the nearest-neighbour approach is justified by the so-called screening effect. Even when the spatial correlation function exhibits long-range dependence, it can subsequently be assumed that $y$ at spatial location $\bm{s}$ is nearly independent of spatially distant observations when conditioned on nearby observations <cit.>. For computational efficiency, D-STEM v2 performs kriging for blocks of pixels. To do this, $\mathcal{S}^{\ast}$ is partitioned in $u$ blocks $\mathcal{S}^{\ast}=\left\{\mathcal{S}_{1}^{\ast},...,\mathcal{S}_{u}^{\ast}\right\}$, and kriging is done on each block $\mathcal{S}_{l}^{\ast}$, $l=1,...,u$, with $u\ll m$ controlled by the user. For each target block $\mathcal{S}_{l}^{\ast}$, the conditioning term in Equations (<ref>) and (<ref>) is given by the data observed at $\mathcal{\tilde{S}}_{l}=\bigcup\nolimits_{j\in \mathcal{J}_{l}}\mathcal{S}_{\sim j},% \mathcal{J}_{l}=\left\{ j:s_{j}^{\ast }\in \mathcal{S}_{l}^{\ast}\right\}$. Note that, if $\mathcal{S}_{l}^{\ast}$ is dense and $\mathcal{S}$ is sparse (namely $n\ll m$), then $\mathcal{\tilde{S}}_{l}$ is not much larger than $\mathcal{S}_{\sim j}$ since most of the spatial locations in $\mathcal{S}_{l}^{\ast}$ tend to have the same neighbours $\mathcal{S}_{\sim j}$. §.§ Validation D-STEM v2 allows one to implement an out-of-sample validation by partitioning the original spatial locations $\mathcal{S}$ into subsets $\mathcal{S}_{est}$ and $\mathcal{S}_{val}$. Data at $\mathcal{S}% _{est}$ are used for model estimation while data at $\mathcal{S}% _{val}$ are used for validation. Once the model is estimated, the kriging formula in Equation (<ref>) is used to predict at $\mathcal{S}_{val}$ for all times $t$ and heights $\bm{h}$. The following validation mean squared errors are then computed \begin{align*} MSE_{t} & =\frac{1}{P_{1}}\sum_{\bm{s}\in\mathcal{S}_{val}}% \sum_{h\in\bm{h}_{\bm{s},t}}\left( y\left( \bm{s},h,t\right) -\hat{y}\left( \bm{s},h,t\right) \right) ^{2},\\ MSE_{\bm{s}} & =\frac{1}{P_{2}}\sum_{t=1}^{T}\sum_{h\in \bm{h}_{\bm{s},t}}\left( y\left( \bm{s},h,t\right) -\hat{y}\left( \bm{s},h,t\right) \right) ^{2},\\ MSE_{h} & =\frac{1}{P_{3}}\sum_{t=1}^{T}\sum_{\bm{s}\in\mathcal{S}_{val}}\left( y\left( \bm{s},h,t\right) -\hat{y}\left( \bm{s},h,t\right) \right) ^{2},% \end{align*} where $\hat{y}\left( \bm{s},h,t\right) $ is obtained from Equation (<ref>), while $P_{1}$, $P_{2}$ and $P_{3}$ are the number of terms in each sum. When $\bm{h}_{\bm{s},t}$ varies across the profiles, D-STEM v2 provides a binned MSE by splitting the continuous domain $\mathcal{H}$ into $B$ equally spaced intervals. Let $H^*_r$ be the set of observation points in the $r\text{-}th$ interval, let $n_r$ be the corresponding observation number and let $\bar{h}_r = \frac{1}{n_r} \sum_{h\in H^*_r} h$ be the mean of points in $b\text{-}th$ interval. Then, the $MSE_{\bar{h}_r}$ is computed by \begin{align*} MSE_{\bar{h}_r} & =\frac{1}{P_{4}} \sum_{h\in H^*_r} \sum_{t=1}^{T}\sum_{\bm{s}\in\mathcal{S}_{val}}\left( y\left( \bm{s},h,t\right) -\hat{y}\left( \bm{s},h,t\right) \right) ^{2},% \end{align*} where $P_{4}$ is the total number of observations in the $b\text{-}th$ interval. D-STEM v2 also provides the validation $R^2$ with respect to time \begin{align*} R^{2}_{t} & =1 - \frac{MSE_{t}}{\VAR\left( \{y\left( \bm{s},h,t\right), \bm{s}\in\mathcal{S}_{val}, h\in\bm{h}_{\bm{s},t} \} \right)}. \end{align*} and the analogous validation $R^2$ with respect to location $\bm{s}$ and $h_r$. § SOFTWARE This section starts by briefly describing the modelling capabilities of D-STEM v2 inherited by the previous version for dealing with spatio-temporal data sets. Then, it focuses on the D-STEM v2 classes and methods, which implement estimation, validation and dynamic mapping of the model presented in Section <ref>. Although some of the classes are already available in D-STEM v1, they are listed here for completeness. §.§ Software description D-STEM v1 implemented a substantial number of models. The dynamic coregionalisation model (DCM, [Finazzi and Fassò, 2014] [Finazzi and Fassò, 2014]) and the hidden dynamic geostatistical model (HDGM, [Calculli et al., 2015] [Calculli et al., 2015]) are suitable for modelling and mapping multivariate space-time data collected from unbalanced monitoring networks. Model-based clustering (MBC, [Finazzi et al., 2015] [Finazzi et al., 2015]) has been introduced for clustering time series, and it is suitable for large data sets with spatially registered time series. Moreover, the emulator model [Finazzi et al., 2019] is based on a Gaussian emulator, and it is exploited for modelling the multivariate output of a complex physical model. In addition, D-STEM v2 (available at <github.com/graspa-group/d-stem>) provides the functional version of HDGM, denoted by f-HDGM, which handles modelling and mapping of functional space-time data, following the methodology of Section <ref>. For implementing f-HDGM, D-STEM v2 relies on the MATLAB version of the fda package [Ramsay et al., 2018], which is automatically downloaded and installed by D-STEM v2. §.§ Data format Two data formats are available to define observations for the f-HDGM. One is the internal format used by the D-STEM v2 classes, and the other one is the user format based on the more user-friendly table data type implemented in recent versions of MATLAB. The latter permits storing measurement profiles, covariate profiles, coordinates, timestamps and units of measure in a single object. The internal format is not discussed here. Considering a table in the user format, each row includes the profiles collected at a given spatial location and time point. The column labels are defined as follows: columns Y and Y_name are used for the dependent variable $y$ and its name as a string field, respectively; the column with prefix X_h_ is used for the values of the domain $h$; eventually, columns with prefix X_beta_ are used for covariates $\bm{x}$. These tables have only one column for $y$ and only one column for $h$. Instead, we can have any number $b \geqslant 0$ of covariate columns. Additionally, the table has columns X_coordinate and Y_coordinate for spatial location $\bm{s}$ and column Time for the timestamp. Units of measure are stored in the Properties.VariableUnits property of the table columns and used in outputs and plots. Units for X_coordinate and Y_coordinate can be deg for degrees, m for meters and km for kilometres. Geodetic distance is used when the unit is deg; otherwise, the Euclidean distance is used. At the table row corresponding to location $\bm{s}_i$ and time $t$, the elements related to $y$ and $\bm{x}$ are vectors with $q_{i,t}$ elements. Vectors related to $y$ may include missing data (NaN). If $y$ is entirely missing for a given $\left( \bm{s},t\right) $, the row must be removed from the table. Since spatial locations $\mathcal{S}$ are fixed in time, and as their number $n$ is determined by the number of unique coordinates in the table, profiles observed at different time points but the same spatial location $\bm{s}$ must have the same coordinates. §.§ Software structure In D-STEM, a hierarchical structure of object classes and methods is used to handle data definition, model definition and estimation, validation, dynamic kriging and the related plotting capabilities. The structure is schematically given below. Further details on the use of each class are given within the two case studies in this paper, while class constructors, methods and property details can be obtained in MATLAB using the command doc <class_name>. §.§.§ Data handling The stem_data class allows the user to define the data used in f-HDGM models, mainly through the following objects and methods. * Objects of stem_data * stem_modeltype: model type (DCM, HDGM, MBC, Emulator or f-HDGM); note that model type is needed here because the data structure varies among the different models; * stem_fda: basis functions specification; * stem_validation (optional): definition of the learning and testing datasets for model validation. * Methods and Properties of stem_data * kmeans_partitioning: data partitioning for parallel EM computations of Section <ref>; this method is applied to a stem_data object, and its output is used by the EM_estimation method in the stem_model class below; * shape (optional): structure with geographical borders used for mapping. * Internal Objects of stem_data * stem_varset: observed data and covariates; * stem_gridlist: list of stem_grid objects * stem_grid: spatial locations coordinates; * stem_datestamp: temporal information. Interestingly, stem_misc.data_formatter is a helper method, which is useful for building stem_­varset objects starting from data tables. Its class, stem_misc, is a miscellanea static class implementing other methods for various intermediate tasks not discussed here for brevity. §.§.§ Model building The stem_model class is used to define, estimate, validate and output a f-HDGM, mainly through the following objects and methods. * Objects of stem_model * stem_data: defined above; * stem_par: model parameters; * stem_EM_result: container of the estimation output, after EM_estimate; * stem_validation_result (optional): container of validation output, available only if stem_data contains the stem_validation object; * stem_EM_options (optional): model estimation options; it is an input of the EM_estimate method below. * Methods of stem_model * EM_estimate: computation of parameter estimates; * set_varcov: computation of the estimated variance-covariance matrix; * plot_profile: plot of functional data; * print: print estimated model summary; * beta_Chi2_test: testing significance of covariates; * plot_par: plot functional parameter; * plot_validation: plot MSE validation. §.§.§ Kriging The kriging handling is implemented with two classes. The first is the stem_krig class, which implements the kriging spatial interpolation. * Objects of stem_krig * stem_krig_data: mesh data for kriging; * stem_krig_options: kriging options; * Methods of stem_krig * kriging: computation of kriging, the output is a stem_krig_result object. The second is the stem_krig_result class, which stores the kriging output and implements the methods for plotting the kriging output. * Methods of stem_krig_result * surface_plot: mapping of kriging estimate and their standard deviation for fixed $h$; * profile_plot: method for plotting the kriging function and the variance-covariance matrix for a fixed space and time. Although at first reading the user could prefer a single object for both input and output of the kriging, these objects may be quite large, making the current approach more flexible. § CASE STUDY ON OZONE DATA This section illustrates how to make inferences on an f-HDGM for ground-level high-frequency air quality data collected by a monitoring network. In particular, hourly ozone ($O_{3}$, in $\mu g/m^3$) measured in Beijing, China, is considered. §.§ Air quality data Ground-level $O_{3}$ is an increasing public concern due to its essential role in air pollution and climate change. In China, $O_{3}$ has become one of the most severe air pollutants in recent years [Wang et al., 2017]. In this case study, the aim is to model hourly $O_{3}$ concentrations from 2015 to 2017 with respect to temperature and ultraviolet radiation (UVB) across Beijing. Concentration and temperature data are available at twelve monitoring stations (Figure <ref>). Hourly UVB data are obtained from the ERA-Interim product of the European Centre for Medium-Range Weather Forecasts (ECMWF) at a grid size of $0.25^{\circ} \times 0.25^{\circ}$ over the city. locations of the twelve stations in Beijing [Kahle and Wickham, 2013]. To describe the diurnal cycle of $O_{3}$, which peaks in the afternoon and reaches a minimum at night-time, the 24 hours of the day are used as domain $\mathcal{H}$ of the basis functions, while the time index $t$ is on the daily scale. Moreover, due to the circularity of time, Fourier basis functions are adopted, which implies that $\beta_{j}\left( h\right) $, $\sigma_{\varepsilon}^{2}\left( h\right) $ are periodic functions. The measurement equation for $O_{3}$ is \begin{equation} _{temp}(h)+x_{uvb}\left( t\right) \beta_{uvb}(h)+\bm{\phi}(h)^{\top}\bm{z}(\bm{s},t)+\varepsilon(\bm{s},h,t), \label{eq:O3meaurement}% \end{equation} where $\bm{s}$ is the generic spatial location, $h\in\left[ 0,24\right) $ is the time within the day expressed in hours and $t=1,...,1096$ is the day index over the period 2015–2017. Based on a preliminary analysis, the number of basis functions for $\beta_{j}\left( h\right) $, $\sigma_{\varepsilon}^{2}\left( h\right) $ and $\bm{\phi}(h)^{\top}\bm{z}(\bm{s},t)$ is chosen to be $5$, $5$ and $7$, respectively. §.§ Software implementation This paragraph details the implementation of the D-STEM v2 in three aspects: model estimation, validation and kriging. Relevant scripts are demo_section4_model_estimate.m, demo_section4_validation.m and demo_section4_kriging.m, respectively, which are available in the supplementary material. All the scripts can be executed by choosing the option number from 1 to 3 in the demo_menu_user.m script. §.§.§ Model estimation This paragraph describes the demo_section4_model_estimate.m script devoted to the estimation of the model parameters and of their variance–covariance matrix. The data set needed to perform this case study is stored as a MATLAB table in the user format of Section <ref> and named Beijing_O3. It can be loaded from the corresponding file as follows: load ../Data/Beijing_O3.mat; In the Beijing_O3 table, each row refers to a fixed space-time point and gives a 24-element hourly ozone profile with the corresponding conformable covariates, which are: a constant, temperature and UVB. The following lines of code specify the model type and the basis functions, which are stored in an object of class stem_fda: o_modeltype = stem_modeltype('f-HDGM'); input_fda.spline_type = 'Fourier'; input_fda.spline_range = [0 24]; input_fda.spline_nbasis_z = 7; input_fda.spline_nbasis_beta = 5; input_fda.spline_nbasis_sigma = 5; o_fda = stem_fda(input_fda); When using a Fourier basis, spline_nbasis_z must be set to a positive odd Meanwhile, spline_nbasis_beta and/or spline_nbasis_sigma must be left empty, if $\bm{\beta}(h)\equiv\bm{\beta}$ and/or $\sigma _{\varepsilon}^{2}(h)\equiv\sigma_{\varepsilon}^{2}$ are constant functions. The next step is to define an object of class stem_data, which specifies the model type and contains the basis function object and the data from the Beijing_O3 table, transformed in the internal data format. This is done using the intermediate input_data structure: input_data.stem_modeltype = o_modeltype; input_data.data_table = Beijing_O3; input_data.stem_fda = o_fda; o_data = stem_data(input_data); Then, an object of class stem_model is created by using both information on data, stored in the o_data object, and on parametrisation, contained in the stem_par object named o_par: o_par = stem_par(o_data,'exponential'); o_model = stem_model(o_data, o_par); To facilitate visualisation, the method plot_profile of class stem_model shows the $O_3$ profile data at location (lat0, lon0), in the days between t_start and t_end (Figure <ref>): lat0 = 40; lon0 = 116; t_start = 880; t_end = 900; o_model.plot_profile(lat0, lon0, t_start, t_end); Before running the EM algorithm, the model parameters need to be initialised. This is done using the method get_beta0 of class stem_model, which provides the starting values for $\bm{\beta}$, and the method get_coe_log_sigma_eps0 for the case of a functional $\sigma_{\varepsilon}^{2}(h)$. Next, the method set_initial_values of the o_model object is called to complete the initialisation of model parameters: n_basis = o_fda.get_basis_number; o_par.beta = o_model.get_beta0(); o_par.sigma_eps = o_model.get_coe_log_sigma_eps0(); o_par.theta_z = ones(1, n_basis.z)*0.18; o_par.G = eye(n_basis.z)*0.5; o_par.v_z = eye(n_basis.z)*10; Note that the theta_z parameter must be provided in the same unit of measure as the spatial coordinates. $O_3$ concentrations at location 39.92 latitude and 116.19 longitude for 21 days beginning on 29 May 2017. Left: each dot is a concentration measurement. The colour of the dot depicts the concentration. Right: each graph is a daily concentration profile. Before model estimation, EM exiting conditions $\epsilon_1$ (exit_toll_par), $\epsilon_2$ (exit_toll_loglike) and $\iota^{\ast}$ (max_iterations) introduced in Section <ref> can be optionally defined as follows: o_EM_options = stem_EM_options(); o_EM_options.exit_toll_par = 0.0001; o_EM_options.exit_toll_loglike = 0.0001; o_EM_options.max_iterations = 200; Model estimation is started by calling the method EM_estimate of the o_model object, with the optional o_EM_options object passed as an input argument. After model estimation, the variance–covariance matrix of the estimated parameters is evaluated by calling the method set_varcov, with the optional approximation level $\delta$ of Equation (<ref>) passed as an input parameter. Finally, set_logL computes the observed data log-likelihood. delta = 0.001; All the relevant estimation results are found in the internal stem_EM_result object, which can be accessed as a property of the o_model object as follows: Figure <ref> is produced by calling the plot_par method and shows the estimated $\bm{\beta}_{0}(h)$, $\bm{\beta}_{temp}(h)$, $\bm{\beta}_{uvb}(h)$, and $\sigma_{\varepsilon}^{2}(h)$. Thanks to the use of a Fourier basis, the functions are periodic with a period of one day. In the plot of $\sigma_{\varepsilon}^{2}(h)$, the unexplained portion of $O_{3}$ variance, $\sigma_{\varepsilon}^{2}(h)$, is small during daylight hours, which is consistent with the results of [Dohan and Masschelein, 1987]. When the confidence bands of parplot contain zero, it may be useful to test the significance of the covariates. By calling the method beta_Chi2_test, the results of $\chi^{2}$ tests are obtained, and they are reported in Table <ref>. Although $\beta_{uvb}$ is close to $0$ in the morning, all fixed effects are highly significant overall. The model output is shown in the MATLAB command window by calling the print method. $\beta_{0}(h),~\beta_{temp}(h),~\beta_{uvb}(h)$ and $\sigma_{\epsilon}^{2}(h)$, with $90\%,~95\%,~99\%$ confidence bands, respectively, shown through the different shades. $\chi^{2}$ statistic $p$ value Constant 1r136.33 1r0 Temperature 1r14266.07 1r0 UVB 1r2094.34 1r0 $\chi^{2}$ tests for significance of covariates. §.§.§ Validation This paragraph describes the script demo_section4_validation.m, which implements validation. Compared to the code in demo_section4_model_estimate.m, it only differs in providing an object of class stem_validation. To create the object called o_validation, the name of the validation variable is needed as well as the indices of the validation stations. Moreover, if the size of the nearest neighbour set for each kriging site (nn_size) is not provided as the third input argument in the stem_validation class constructor, D-STEM v2 uses all the remaining stations. For example, a validation data set with three stations is constructed as follows: S_val = [1,7,10]; input_data.stem_validation = stem_validation('O3', S_val); The validation statistics, computed by EM_estimate, are saved in the internal object stem_­validation_result, which can be accessed as a property of the o_model object. The stem_validation_result object contains the estimated $O_{3}$ residuals for the above-mentioned validation stations as well as the validation mean square errors and $R^2$, as defined in Section <ref>. §.§.§ Kriging This paragraph describes the demo_section4_kriging.m script, which applies the approach of Section <ref> to the estimated model to map the $O_{3}$ concentrations over Beijing city. The first step is to create an object of class stem_grid, which collects the information about the regular grid of pixels $\mathcal{B}$ to be used for mapping. Then, an object of class stem_krig_data is created, where the o_krig_grid object is passed as an input argument: load ../Output/ozone_model; step_deg = 0.05; lat = 39.4:step_deg:41.1; lon = 115.4:step_deg:117.5; [lon_mat,lat_mat] = meshgrid(lon,lat); krig_coordinates = [lat_mat(:) lon_mat(:)]; o_krig_grid = stem_grid(krig_coordinates,'deg','regular','pixel',... o_krig_data = stem_krig_data(o_krig_grid); Two comments on the above lines follow. First, since the grid in the o_krig_grid object is regular, the dimensions of the grid (size(lat_mat), $35 \times 43$), must be provided as well as the shape of the pixels and the spatial resolution of the grid, which is $0.05^{\circ} \times 0.05 ^{\circ}$. Second, the above step using the stem_krig_data constructor may appear redundant at first glance. Indeed, it is needed for compatibility with other model types for which, in addition to the stem_grid object, other information is also necessary for the stem_krig_data constructor. Next, the stem_krig_options class provides some options for kriging. By default, the output is back-transformed in the original unit of measure if the observations have been log-transformed and/or standardised. The back_­transform property enables handling this. Moreover, the no_varcov property must be set to 1 to avoid the time-consuming computation of the kriging variance. Eventually, the block_size property is used to define the number of spatial locations in $\mathcal{S}% o_krig_options = stem_krig_options(); o_krig_options.back_transform = 0; o_krig_options.no_varcov = 0; o_krig_options.block_size = 30; After storing the map of Beijing boundaries into the o_model object, the latter is used with o_krig_data to create an object of class stem_krig. This and o_krig_options together contain all information for kriging, which is obtained by the corresponding kriging method: o_model.stem_data.shape = shaperead('../Maps/Beijing_adm1.shp'); o_krig = stem_krig(o_model, o_krig_data); o_krig_result = o_krig.kriging(o_krig_options); Note that this task may be time consuming for large grids. The kriging output saved in the o_krig_result object gives the latent process estimate $z_t$ and its variance. The surface_plot and profile_plot methods may be used to obtain and plot $\hat{f}(\bm{s},h,t)$ of Equation (<ref>). In this case, the user has to provide the corresponding covariate (X_beta) for the scale/vector h, time t or location $\bm{s}$ (lon0, lat0) of interest. Specifically, the surface_plot method is used to display the $O_{3}$ map using h, t, X_beta as input arguments. In the case of unavailable X_beta, the mapping concerns the component $\bm{\phi}(h)^{\top}\bm{z}(\bm{s},t)$. Loaded from the homonym file, the array X_beta_t_100 refers to time $t=100$ and hour $h=10.5$ and has the dimension $35 \times 43 \times 3$. Maps of $O_{3}$ concentrations and their standard deviation are shown in Figure <ref>. load ../Data/kriging/X_beta_t_100; t = 100; h = 10.5; [y_hat, diag_Var_y_hat] = o_krig_result.surface_plot(h, t, X_beta_t_100); $O_3$ concentrations and their standard deviation at 10:30 am ($h = 10.5$), on 10 April 2015, where 12 stations are marked with black stars. On the other hand, the profile_plot method is used to display the $O_{3}$ profile at a given spatial location $\bm{s}$ (lon0, lat0) and time t. Still, the profile plot concerns the component $\bm{\phi}(h)^{\top}\bm{z}(\bm{s},t)$ if X_beta is not provided. After loading the X_beta_h (dimension $25\times 3$) from the homonym file, this method represents the profile of $O_{3}$ concentrations and their variance–covariance matrix as in Figure <ref>: load ../Data/kriging/X_beta_h; h = 0:24; lon0 = 116.25; lat0 = 40.45; t = 880; [y_hat, diag_Var_y_hat] = o_krig_result.profile_plot(h, lon0, lat0, ... t, X_beta_h); Note that the prediction in Equation (<ref>) and the variance in Equation (<ref>) are stored in the output arguments y_hat, and diag_Var_y_hat, respectively. $O3$ concentrations with $90\%, 95\%, 99\%$ confidence bands (different shadings), and their variance–covariance at latitude 40.45, longitude 116.25 on 29 May 2017. § CASE STUDY ON CLIMATE DATA In order to show the complexity-reduction capabilities of D-STEM v2, a data set of temperature vertical profiles collected by the radiosondes of the Universal Radiosonde Observation Program (RAOB) is now considered. The profiles are observed over the Earth's sphere, and they are misaligned, that is, each profile differs in terms of the number of observations and altitude above the ground of each observation. Additionally, the computation burden is higher due to the higher number of spatial locations at which profiles are observed. §.§ RAOB data Radiosondes are routinely launched from stations all over the world to measure the state of the upper troposphere and lower stratosphere. Data collected by radio sounding have applications in weather prediction and climate studies. Temperature data from 200 globally distributed stations collected daily during January 2015 at 00:00 and 12:00 UTC are considered here. Each profile consists of a given number of measurements taken at different pressure levels. Since the weather balloon carrying the radiosonde usually explodes at an unpredictable altitude, the profile measurements are misaligned across the profiles and have different pressure ranges. A functional data approach is natural in this case since the underlying temperature profile can be seen as a continuous function sampled at some pressure levels. Figure <ref> depicts the spatial locations of temperature measurements taken on 1 January 2015 at 00:00 UTC. This demo data set, which only covers one month, includes around $10^5$ data points. When the full data set is used in climate studies, the number of data points grows to around $10^8$. In this case, a recent server machine with multiple CPUs with at least 256 GB of RAM is required for model estimation and kriging. The focus of the case study is on the difference between the radiosonde measurement and the output of the ERA-Interim global atmospheric reanalysis model provided by ECMWF. In particular, the aim is to study the spatial structure of the this difference in 4D space, where the dimensions are latitude, longitude, altitude and time. The model for temperature $y$ is as follows \[ y \left( \bm{s},h,t\right) =x_{ERA}\left( \bm{s},h,t\right) \beta_{ERA}\left( h\right) +\bm{\phi}\left( h\right) ^{\top}\bm{z}\left( \bm{s},t\right) +\varepsilon\left( \bm{s}% \] where $h\in\left[ 50, 925\right] $ $hPa$ is the pressure level, while $t=1,...,62$ is a discrete time index for January 2015. Figure <ref> shows the temperature measurements at a given station, where 50 and 925 $hPa$ correspond approximately to 25 and 1.3 km, respectively. Temperature at location 5.25 latitude and -3.93 longitude in January 2015. Left: each dot is a temperature measurement. The colour of the dot depicts the temperature. Right: each graph is a temperature vertical profile collected through radio sounding. §.§ Software implementation This section details the software implementation of the case study described above as in script demo_section5.m, which can be also executed in the demo_menu_user.m script. To avoid repetition, only the relevant parts of the script that differ from the case study of Section <ref> are reported and commented on here. In particular, data loading and instantiation of the stem_model object are not described. §.§.§ Model estimation The problem of vertical misalignment of the measurements is completely transparent to the user and is handled by the internal stem_misc.data_formatter method when creating the stem_data object. Note that the dimension of the matrices in o_varset depends on $q$, the maximum number of measurements in each profile. To prevent out-of-memory problems, it is advisable to avoid data sets in which only a few profiles have a large number of measurements, which could result in large matrices in o_varset, with most of the elements set to NaN. B-spline bases are used, since, in this application, vertical profiles are not periodic with respect to the pressure domain. The corresponding object of class stem_fda is created in the following way: spline_order = 2; rng_spline = [50,925]; knots_number = 5; knots = linspace(rng_spline(1),rng_spline(2),knots_number); input_fda.spline_type = 'Bspline'; input_fda.spline_order = spline_order; input_fda.spline_knots = knots; input_fda.spline_range = rng_spline; o_fda = stem_fda(input_fda); Note that the knots are equally spaced along the functional range. In general, however, non-equally spaced knots can be provided, and each model component (i.e. $\sigma_{\varepsilon}^{2}$, $\beta_{j}$ and $\bm{\phi}(h)^{\top}\bm{z}(\bm{s}% ,t)$) can have a different set of knots. This is obtained using spline_order and spline_knots with additional suffixes _sigma, _beta, _z. Although this data set is not large, the demo shows how to enable the partitioning discussed in Section <ref>. First, the spatial locations are partitioned using the modified k-means algorithm: k = 5; trials = 100; lambda = 5000; partitions = o_data.kmeans_partitioning(k, trials, lambda); where k is the number of partitions, trials is the number of times when the k-means algorithm is executed starting from randomised centroids and lambda is $\lambda$ in Equation (<ref>). At the end of the k-means algorithm, data are internally re-ordered for parallel computing. Model estimation is done after creating and setting an object of class stem_EM_options. To do this, the output of the kmeans_globe method is passed to the partitioning_block_size property of the o_EM_options object. Additionally, for parallel computing, the number of workers must be set to a value higher than 1. In general, this could be any number up to the number of cores available on the machine. o_EM_options = stem_EM_options(); o_EM_options.partitions = partitions; o_EM_options.workers = 2; The three validation MSEs defined in Section <ref> are shown in Figure <ref> and <ref>. To generate these figures, the method plot_validation is called with vertical = 1, which provides “atmospheric profile” plots with $h$ on the vertical axis: vertical = 1; (Left) Validation MSE with respect to the $\bar{h}$ coloured by the number of observations $n_b$, and (Right) the validation MSE with respect to time $t$. Validation MSE for the thirty-three stations, where the stations used for estimation are marked with blue stars. §.§.§ Kriging Interpolation across space and over time is done as in Section <ref>. However, complexity reduction is enabled by adopting the nearest neighbour approach detailed in Section <ref>. To do this, a class constructor is first called, where the block_size is used to define the number of spatial locations in $\mathcal{S}_{l}^{\ast}$, and then nn_size is used to define $\tilde{n}$. Additionally, setting o_krig_options.workers makes it possible to do the kriging over the $u$ blocks in parallel using up to the allocated number of workers: o_krig_options = stem_krig_options(); o_krig_options.block_size = 150; o_krig_options.nn_size = 10; o_krig_options.workers = 2; Finally, kriging predictions and standard errors are mapped for a given $h\in\mathcal{H}$ and time $t$: h = 875.3; t = 12; o_krig_result.surface_plot(h, t); Since covariates are not provided to the surface_plot method, the plots are on the component $\bm{\phi}(h)^{\top}\bm{\hat{z}}(\bm{s},t)$, namely, the difference between RAOB and ERA-Interim and its standard deviation. The output of the above code is depicted in Figures <ref> and <ref>. $\bm{\phi}(h)^{\top}\bm{\hat{z}}(\bm{s},t)$ at pressure $875.3$ $hPa$, and 12:00 am on 6 January 2015, where $200$ stations are shown as black stars. Standard deviation of $\bm{\phi}(h)^{\top}\bm{\hat{z}}(\bm{s},t)$ at pressure $875.3$ $hPa$, and 12:00 am on 06 January 2015, where $200$ stations are shown as black stars. § CONCLUDING REMARKS This paper introduced the package D-STEM v2 through two case studies of spatio-temporal modelling of functional data. It is shown that, in addition to maximum likelihood estimation, Hessian approximation and kriging for large data sets, D-STEM v2 also develops several data-handling capabilities, allows for automatic construction of relevant objects and provides graphical output. In particular, it provides high-quality global maps and two kinds of functional plotting: the traditional x–y plot and the vertical profile plot, which is popular, for example, in atmospheric data analysis. In this regard, model validation and kriging are straightforward. D-STEM v2 fills a gap in functional geostatistics. In fact, although statistical methods for georeferenced functional data have been recently developed (e.g. [Ignaccolo et al., 2014]), standard geostatistical packages do not consider functional data, especially in the spatio-temporal context. The successful use of D-STEM v1 in a number of applications proved that the EM algorithm implementation is quite stable. Now, due to improvements in computational efficiency, the new D-STEM v2 has the capability to handle large data sets. Moreover, thanks to the approximated variance–covariance matrix, it is possible to compute standard errors for all model parameters relatively fast and avoid the large number of iterations typically required by an MCMC approach for making inferences. However, a limit of the EM algorithm is its limited flexibility to changes in the model equations. Indeed, changes in parametrisation or latent variable structure usually require deriving new closed-form estimation formulas and changing the software accordingly. Moreover, changes in covariance functions are not easy to handle. Computationally, the main limit of D-STEM v2 is in the number $p$ of basis functions that can be handled. Even if partitioning is exploited in $k$ blocks of size $r$, computational complexity is $\mathcal{O}\left(Tkr^{3}p^{3}\right)$, meaning that $p$ cannot be large. Currently, the authors are working on a new version, which makes it possible to handle multivariate functional space-time data and user-defined spatial covariance functions, which will make D-STEM v2 a valid and comprehensive alternative to the Gaussian process regression models (fitrgp) implemented in the Statistics and Machine Learning Toolbox of MATLAB. § ACKNOWLEDGMENTS We would like to thank China's National Key Research Special Program Grant (No. 2016YFC0207701) and the Center for Statistical Science at Peking University. [Bivand et al., 2015] Bivand RS, Gómez-Rubio V, Rue H (2015). Spatial Data Analysis with R-INLA with Some Journal of Statistical Software, 63(20), 1–31. [Blangiardo et al., 2013] Blangiardo M, Cameletti M, Baio G, Rue H (2013). Spatial and Spatio-Temporal Models with Spatial and Spatio-Temporal Epidemiology, 4, 33–49. [Calculli et al., 2015] Calculli C, Fassò A, Finazzi F, Pollice A, Turnone A (2015). Maximum Likelihood Estimation of the Multivariate Hidden Dynamic Geostatistical Model with Application to Air Quality in Apulia, Environmetrics, 26(6), 406–417. [Cressie, 2018] Cressie N (2018). Mission CO2ntrol: A Statistical Scientist's Role in Remote Sensing of Atmospheric Carbon Dioxide. Journal of the American Statistical Association, 113(521), 152–168. [Cressie and Johannesson, 2008] Cressie N, Johannesson G (2008). Fixed Rank Kriging for Very Large Spatial Data Sets. Journal of the Royal Statistical Society B, 70(1), [Dohan and Masschelein, 1987] Dohan JM, Masschelein WJ (1987). The Photochemical Generation of Ozone: Present Ozone: Science $\&$ Engineering, 9(4), 315–334. [Fassò, 2013] Fassò A (2013). Statistical Assessment of Air Quality Interventions. Stochastic Environmental Research and Risk Assessment, 27(7), 1651–1660. <DOI: 10.1007/s00477-013-0702-5>. [Fassò and Finazzi, 2011] Fassò A, Finazzi F (2011). Maximum Likelihood Estimation of the Dynamic Coregionalization Model with Heterotopic Data. Environmetrics, 22(6), 735–748. [Fassò et al., 2016] Fassò A, Finazzi F, Ndongo F (2016). European Population Exposure to Airborne Pollutants Based on a Multivariate Spatio-Temporal Model. Journal of agricultural, biological, and environmental statistics, 21(3), 492–511. [Fassò et al., 2014] Fassò A, Ignaccolo R, Madonna F, Demoz B, Franco-Villoria M (2014). Statistical Modelling of Collocation Uncertainty in Atmospheric Thermodynamic Profiles. Atmospheric Measurement Techniques, 7(6), 1803–1816. [Finazzi, 2020] Finazzi F (2020). Fulfilling the Information Need after an Earthquake: Statistical Modelling of Citizen Science Seismic Reports for Predicting Earthquake Parameters in Near Realtime. Journal of the Royal Statistical Society: Series A, 183(3), 857–882. [Finazzi and Fassò, 2014] Finazzi F, Fassò A (2014). D-STEM: A Software for the Analysis and Mapping of Environmental Space-Time Variables. Journal of Statistical Software, 62(6), 1–29. [Finazzi et al., 2018] Finazzi F, Fassò A, Madonna F, Negri I, Sun B, Rosoldi M (2018). Statistical Harmonization and Uncertainty Assessment in the Comparison of Satellite and Radiosonde Climate Variables. arXiv preprint arXiv:1803.05835. [Finazzi et al., 2015] Finazzi F, Haggarty R, Miller C, Scott M, Fassò A (2015). A Comparison of Clustering Approaches for the Study of the Temporal Coherence of Multiple Time Series. Stochastic Environmental Research and Risk Assessment, 29(2), 463–475. [Finazzi et al., 2019] Finazzi F, Napier Y, Scott M, Hills A, Cameletti M (2019). A Statistical Emulator for Multivariate Model Outputs with Missing Values. Atmospheric Environment, 199, 415 – 422. [Finazzi et al., 2013] Finazzi F, Scott EM, Fassò A (2013). A Model-Based Framework for Air Quality Indices and Population Risk Evaluation, with an Application to the Analysis of Scottish Air Quality Data. Journal of the Royal Statistical Society: Series C, 62(2), 287–308. [Finley et al., 2015] Finley A, Banerjee S, Gelfand A (2015). spBayes for Large Univariate and Multivariate Point-Referenced Spatio-Temporal Data Models. Journal of Statistical Software, Articles, 63(13), ISSN 1548-7660. [Furrer et al., 2006] Furrer R, Genton MG, Nychka D (2006). Covariance Tapering for Interpolation of Large Spatial Journal of Computational and Graphical Statistics, 15(3), 502–523. [Gasch et al., 2015] Gasch CK, Hengl T, Gräler B, Meyer H, Magney TS, Brown DJ (2015). Spatio-Temporal Interpolation of Soil Water, Temperature, and Electrical Conductivity in 3D+T: The Cook Agronomy Farm Data Set. Spatial Statistics, 14, 70–90. [Gramacy, 2016] Gramacy RB (2016). laGP: Large-Scale Spatial Modeling via Local Approximate Gaussian Processes in R. Journal of Statistical Software, 72(1), 1–46. [Heaton et al., 2018] Heaton MJ, Datta A, Finley AO, Furrer R, Guinness J, Guhaniyogi R, Gerber F, Gramacy RB, Hammerling D, Katzfuss M, Lindgren F, Nychka DW, Sun F, Zammit-Mangion A (2018). A Case Study Competition among Methods for Analyzing Large Spatial Data. Journal of Agricultural, Biological and Environmental Statistics, pp. 1–28. [Ignaccolo et al., 2014] Ignaccolo R, Mateu J, Giraldo R (2014). Kriging with External Drift for Functional Data for Air Quality Monitoring. Stochastic Environmental Research and Risk Assessment, 28(5), 1171–1186. [Jurek and Katzfuss, 2018] Jurek M, Katzfuss M (2018). Multi-Resolution Filters for Massive Spatio-Temporal Data. arXiv preprint arXiv:1810.04200. [Kahle and Wickham, 2013] Kahle D, Wickham H (2013). ggmap: Spatial visualization with ggplot2. The R Journal, 5(1), 144–161. [Katzfuss, 2017] Katzfuss M (2017). A Multi-Resolution Approximation for Massive Spatial Journal of the American Statistical Association, 112(517), 201–214. [Lindgren and Rue, 2015] Lindgren F, Rue H (2015). Bayesian Spatial Modelling with R-INLA. Journal of Statistical Software, 63(19), 1–25. [Negri et al., 2018] Negri I, Fassò A, Mona L, Papagiannopoulos N, Madonna F (2018). Modeling Spatiotemporal Mismatch for Aerosol Profiles. In Quantitative Methods in Environmental and Climate Research, pp. 63–83. Springer-Verlag. [Nychka et al., 2015] Nychka D, Bandyopadhyay S, Hammerling D, Lindgren F, Sain S (2015). A Multiresolution Gaussian Process Model for the Analysis of Large Spatial Datasets. Journal of Computational and Graphical Statistics, 24(2), 579–599. [Nychka et al., 2016] Nychka D, Hammerling D, Sain S, Lenssen N (2016). LatticeKrig: Multiresolution Kriging Based on Markov Random Fields. University Corporation for Atmospheric Research, Boulder, CO, USA. R package version 7.0. [Pebesma, 2012] Pebesma E (2012). spacetime: Spatio-temporal Data in R. Journal of statistical software, 51(7), 1–30. [Pebesma and Heuvelink, 2016] Pebesma E, Heuvelink G (2016). Spatio-temporal Interpolation Using gstat. RFID Journal, 8(1), 204–218. [Porcu et al., 2018] Porcu E, Alegria A, Furrer R (2018). Modeling Temporally Evolving and Spatially Globally Dependent Data. International Statistical Review, 86(2), 344–377. [Ramsay and Silverman, 2007] Ramsay JO, Silverman BW (2007). Applied Functional Data Analysis: Methods and Case Studies. [Ramsay et al., 2018] Ramsay JO, Wickham H, Graves S, Hooker G (2018). fda: Functional Data Analysis. R package version 2.4.8, [Rue et al., 2014] Rue H, Martino S, Blangiardo FL, Simpson D, Riebler A, Krainski ET (2014). INLA: Functions which Allow to Perform Full Bayesian Analysis of Latent Gaussian Models using Integrated Nested Laplace R package version 0.0-1404466487, [Shumway and Stoffer, 2017] Shumway RH, Stoffer DS (2017). Time Series Analysis and Its Applications: with R [Stein, 2002] Stein ML (2002). The Screening Effect in Kriging. The Annals of Statistics, 30(1), 298–323. [Stein, 2013] Stein ML (2013). Statistical Properties of Covariance Tapers. Journal of Computational and Graphical Statistics, 22(4), 866–885. [Taghavi-Shahri et al., 2019] Taghavi-Shahri S, Fassò A, Mahaki B, Amin H (2019). Concurrent Spatiotemporal Daily Land Use Regression Modeling and Missing Data Imputation of Fine Particulate Matter Using Distributed Space-Time Expectation Maximization. Atmospheric Environment, 224, 1–11. [Tzeng and Huang, 2018] Tzeng S, Huang HC (2018). Resolution Adaptive Fixed Rank Kriging. Technometrics, 60(2), 198–208. [Wan et al., 2020] Wan Y, Xu M, Huang H, Chen S (2020). A Spatio-Temporal Model for the Analysis and Prediction of Fine Particulate Matter Concentration in Beijing. Environmetrics, accepted. [Wang et al., 2017] Wang T, Xue L, Brimblecombe P, Lam YF, Li L, Zhang L (2017). Ozone Pollution in China: A Review of Concentrations, Meteorological Influences, Chemical Precursors, and Effects. Science of the Total Environment, 575, 1582–1596. [Zammit-Mangion, 2018] Zammit-Mangion A (2018). FRK: Fixed Rank Kriging. R package version 0.2.2, get arXiv to do 4 passes: Label(s) may have changed. Rerun
11institutetext: INAF - Osservatorio Astronomico di Brera, via Brera 28, 20121 Milan, Italy 11email<EMAIL_ADDRESS>22institutetext: DiSAT - Università degli Studi dell’ Insubria, via Valleggio 11, 22100, Como, Italy 33institutetext: International Centre for Radio Astronomy Research, Curtin University, 1 Turner Avenue, Bentley, WA 6102, Australia # Radio detection of VIK J2318$-$3113, the most distant radio-loud quasar ($z$=6.44) L. Ighina 1122 S. Belladitta 1122 A. Caccianiga 11 J. W. Broderick 33 G. Drouart 33 A. Moretti 11 N. Seymour 33 We report the 888 MHz radio detection in the Rapid ASKAP Continuum Survey (RACS) of VIK J2318$-$3113, a $z$=6.44 quasar. Its radio luminosity (1.2 $\times 10^{26}$ W Hz-1 at 5 GHz) compared to the optical luminosity (1.8 $\times 10^{24}$ W Hz-1 at 4400 Å) makes it the most distant radio-loud quasar observed so far, with a radio loudness R$\sim$70 (R$=L_{\mathrm{{5GHz}}}/L_{\mathrm{{4400\AA}}}$). Moreover, the high bolometric luminosity of the source (Lbol=7.4 $\times 10^{46}$ erg s-1) suggests the presence of a supermassive black hole with a high mass ($\gtrsim$6 $\times$108 M⊙) at a time when the Universe was younger than a billion years. Combining the new radio data from RACS with previous ASKAP observations at the same frequency, we found that the flux density of the source may have varied by a factor of $\sim$2, which could suggest the presence of a relativistic jet oriented towards the line of sight, that is, a blazar nature. However, currently available radio data do not allow us to firmly characterise the orientation of the source. Further radio and X-ray observations are needed. ###### Key Words.: galaxies: active – galaxies: high-redshift – galaxies: jets – quasars: general – quasars individual: VIKING~J231818.3$-$311346 ## 1 Introduction In recent years, the exploitation of numerous optical and infrared (IR) wide- area surveys (e.g. the Panoramic Survey Telescope and Rapid Response System, Pan-STARRS, Chambers et al. 2016; the VISTA Kilo-degree Infrared Galaxy Survey, VIKING, Edge et al. 2013; the Dark Energy Survey, DES, Dark Energy Survey Collaboration et al. 2016, etc.) has led to the discovery of thousands of high-$z$ quasars (QSOs), with more than 200 sources discovered at $z$>6 (e.g. Mazzucchelli et al. 2017; Matsuoka et al. 2019; Fan et al. 2019; Wang et al. 2019; Andika et al. 2020), the three most distant of which are at z$\sim$7.5 (Bañados et al., 2018; Yang et al., 2020a; Wang et al., 2021). These sources have already proved to be very useful tools for investigating the intergalactic medium (IGM) at early cosmic times through the absorption of their optical spectra bluewards of Ly$\alpha$ (e.g. Kashikawa et al. 2006; Gaikwad et al. 2020). Moreover, the mere presence of such powerful and massive objects in the primordial Universe places strong constraints on theoretical models describing the evolution and the accretion rate of supermassive black holes (SMBHs; e.g. Volonteri et al. 2015; Wang et al. 2020). Decades of studies at low redshift have now established that radio-loud (RL111We considered a QSO to be radio loud when it has a radio loudness $R$>10, with $R$ defined as the ratio of the 5 GHz and 4400 Å rest-frame flux densities, $R=S_{\mathrm{5GHz}}/S_{\mathrm{{4400\AA}}}$ (Kellermann et al., 1989).) sources represent $\sim$10-15% of the total QSO population (e.g. Retana-Montenegro & Röttgering 2017), with no significant deviations until z$\sim$6 (e.g. Stern et al. 2000; Liu et al. 2021; Diana et al. in prep.). However, of all the $z$>6 QSOs, only a few have a radio detection, which means that there are far fewer confirmed high-$z$ RL QSOs. To date, only five have been found at $z$>6 (McGreer et al., 2006; Bañados et al., 2015; Belladitta et al., 2020; Liu et al., 2021), with the most distant being at $z$=6.21 (Willott et al. 2010). As described by Kellermann et al. (2016), the RL classification (R>10), as opposed to the radio quiet (RQ; R<10), should identify sources that produce the radio emission through a relativistic jet, which can significantly affect both the accretion process itself and the environment of the source (see Blandford et al. 2019 for a recent review). Identifying and characterising powerful RL sources at the highest redshifts therefore is of key importance for studying the role of relativistic jets in the primordial Universe. In this Letter we report the radio detection of the $z$=6.444$\pm$0.005 QSO VIKING J231818.35$-$311346.3 (hereafter VIK J2318$-$3113; Decarli et al. 2018). With a relatively bright radio flux density ($\sim$1.4 mJy at 888 MHz), this source is the most distant RL QSO observed to date. VIK J2318$-$3113 was discovered from the near-IR (NIR) VIKING survey with the dropout technique, and its redshift was confirmed with both X-Shooter in the NIR and the Atacama Large Millimetre/submillimetre Array (ALMA) in the submillimetre (Decarli et al., 2018; Yang et al., 2020b). In this Letter we present its radio properties using recent observations, and by combining them with archival data, we also compare VIK J2318$-$3113 with the small number of other high-$z$ RL QSOs. We use a flat $\Lambda$CDM cosmology with $H_{0}$=70 km s-1 Mpc-1, $\Omega_{m}$=0.3, and $\Omega_{\Lambda}$=0.7. Spectral indices are given assuming $S_{\nu}\propto\nu^{-\alpha}$ , and all errors are reported at 1$\sigma$ unless otherwise specified. ## 2 Radio observations ### 2.1 888 MHz ASKAP observations VIK J2318$-$3113 has been detected in the first data release of the Rapid ASKAP Continuum Survey (RACS; McConnell et al. 2020)222https://data.csiro.au/collections/collection/CIcsiro:46532. with a peak flux density of 1.43 mJy beam-1at 888 MHz, which considering the associated RMS (0.19 mJy beam-1), corresponds to a signal-to-noise ratio (S/N) >7 (values as reported in the catalogue released on 2020 December 17). The overall RACS survey is planned to cover the entire sky south of declination $+51^{\circ}$ (36656 deg2 in total) in three different radio bands centred at 888, 1296, and 1656 MHz, all with a bandwidth of 288 MHz. These observations are designed as a pilot project to prepare for the data calibration and handling of future deeper surveys (e.g. the evolutionary map of the Universe, EMU, Norris et al. 2011) with the Australian SKA Pathfinder (ASKAP; Johnston et al. 2008). In the first data release (December 2020), the sky south of declination $+41^{\circ}$ was covered in the lower frequency band (888 MHz) with a spatial resolution of $\sim$15′′. By cross-matching this first data release with the list of $z$>6 QSOs discovered to date in the same sky area (169 sources in total), we found the radio counterparts of three of them: VIK J2318$-$3113, and two other RL QSOs. For these last two objects a discussion of their radio properties has already been reported in the literature: FIRST J1427385+331241 ($z$=6.12; McGreer et al. 2006) and PSO J030947.49+271757.31 ($z$=6.10; Belladitta et al. 2020). Figure 1: 1′ $\times$ 1′ cutout of the $Y$-band VIKING image around VIK J2318$-$3113, overlaid with the 888 MHz radio contours from RACS (continuous red lines) and GAMA23 (dashed blue lines). In both cases the contours are spaced by $\sqrt{2}$ starting from three times the off-source RMS derived in our analysis, $\sim$0.20 mJy beam-1 for RACS and $\sim$0.04 mJy beam-1 for GAMA23. In the bottom left corner the beam sizes from the RACS (12.2′′ $\times$ 11.4′′) and GAMA23 (10.2′′ $\times$ 8.5′′) observations are shown. The radio source is located 1.6′′ from the optical/NIR counterpart of VIK J2318$-$3113, which is consistent with the positional error reported in the RACS catalogue ($\sim$4′′). Even considering typical uncertainties in interferometric radio positions ($\approx$$\frac{\Delta\theta}{2\times S/N}$$\sim$0.9′′, where $\Delta\theta$ is the size of the beam; Fomalont 1999) together with the typical astrometric precision of the survey ($\sim$0.8′′; McConnell et al. 2020), the observed offset is still consistent. Moreover, from the source density of the RACS survey ($\sim$80 sources deg-2, McConnell et al. 2020), we can also compute the probability of finding an unrelated radio source within a 1.6′′ radius from any given position (see e.g. eq. 4 in Condon et al. 1998). In this case, the probability is $\sim$5$\times$10-5, which means that the expected number of spurious associations of the 169 $z$>6 QSOs that we based the query on is <0.01. We can therefore conclude that the association between VIK J2318$-$3113 and the radio source is statistically significant and unlikely to be spurious. At the same time, VIK J2318$-$3113 also belongs to one of the Galaxy and Mass Assembly (GAMA; Driver et al. 2011) fields, GAMA23 (339 < R.A. [deg] < 351 and $-$35 < Dec. [deg] < $-$30). In particular, this region has recently (2019 March) been covered by a deeper ASKAP observation (RMS$\sim$0.04 mJy beam-1), again at 888 MHz, within an ASKAP/EMU early science project333https://data.csiro.au/collections/collection/CIcsiro:40262. and was reduced as described in Seymour et al. (2020). We report in Fig. 1 the 888 MHz radio contours from the RACS and GAMA23 observations, overlaid on the NIR VIKING image in the $Y$-band. Table 1: Results of the analysis of the 888 MHz ASKAP observations of VIK J2318$-$3113. Project: | RACS | GAMA23 ---|---|--- Total flux density (mJy): | 1.44$\pm$0.34∗ | 0.59$\pm$0.07 Peak surf. brightness (mJy/beam): | 1.48$\pm$0.20 | 0.59$\pm$0.04 Major axis∗∗ (arcsec): | 13.2$\pm$2.0 | 10.5$\pm$0.8 Minor axis∗∗ (arcsec): | 10.2$\pm$1.2 | 8.2$\pm$0.5 P.A. east of north (deg): | 45$\pm$18 | 105$\pm$10 Off-source RMS (mJy/beam): | 0.20 | 0.04 444 $*$$*$footnotetext: In the following we use the more conservative error of 0.60 mJy obtained from eq. 7 in McConnell et al. (2020). See section 2.1 for further details. $**$$**$footnotetext: Convolved with the beam of the instrument. In Tab. 1 we report the results of a single Gaussian fit performed on the RACS and GAMA23 images using the Common Astronomy Software Applications package (CASA; McMullin et al. 2007). In the GAMA23 observation the best-fit position is only 0.37′′ away from the NIR counterpart, thus providing further strong evidence for the radio association. Given the very similar angular resolution in both cases, the source is point-like and not resolved. However, the estimated flux density varies by a factor $\sim$2.4 in the two images, from 0.59$\pm$0.07 to 1.44$\pm$0.34 mJy. The time separation between the two observations is one year (2019 March – 2020 March), which in the source rest frame corresponds to $\sim$50 days (without taking possible relativistic effects into account). In order to verify whether the source variation between the GAMA23 and RACS observations is real or is only a systematic effect related to the calibration, we compared the integrated flux densities of the sources detected in the two images. In particular, as for VIK J2318$-$3113, we performed a single Gaussian fit with the CASA software on $\sim$70 sources with a flux density between 1 and 10 mJy and within one square degree from the QSO position555Although a primary beam correction was performed during the data reduction of the RACS survey and the GAMA23 images, we applied a search radius cutoff in order avoid any possible residual fluctuation of the flux calibration.. The distribution of the ratios of the flux densities measured in the two images is a Gaussian centred at one and with $\sigma$=0.16, consistent with the statistical errors on the flux densities and thus indicating that the observed difference for VIK J2318$-$3113 cannot be attributed to a systematic calibration offset in the two datasets. When we sum in quadrature the uncertainties related to the two flux density estimates, the significance of the variation observed in VIK J2318$-$3113 is $\sim$2.4$\sigma$. A large variation in a short period of time as observed in this case is usually associated with the presence of a relativistic jet oriented towards the line of sight, that is, a blazar nature (e.g. Hovatta et al. 2008). The uncertainty on the flux density ratios reported above ($\sigma$=0.16) was derived from the relative comparison of the RACS and GAMA23 images, that is, from datasets obtained from the same telescope. McConnell et al. (2020) have studied the uncertainties on the absolute flux density scale of RACS images by comparing sources with multiple independent RACS observations (i.e. on the overlapping edges of different tiles), also with other catalogues in the literature, finding $\Delta$Sν = 0.5 mJy + 0.07$\times$Sν (eq. 7 in their paper). In the particular case of VIK J2318$-$3113, the corresponding value is $\sim$0.60 mJy. We take this uncertainty into account in section 4 when we compute the quantities based on the RACS flux density (e.g. radio luminosity and radio loudness). ### 2.2 Archival radio observations Even though VIK J2318$-$3113 is not reported in any other public radio catalogue, we checked archival radio images at the NIR position of the source to search for the presence of a faint but significant (S/N>2.5) radio signal. We did not detect the source in the TIFR Giant Metrewave Radio Telescope Sky Survey (TGSS; Intema et al. 2017) at 150 MHz (image RMS$\sim$2.9 mJy beam-1), the Sydney University Molonglo Sky Survey (SUMSS; Mauch et al. 2003) at 843 MHz (image RMS$\sim$2.5 mJy beam-1), or in the NRAO Karl G. Jansky Very Large Array Sky Survey (NVSS; Condon et al. 1998) at 1.4 GHz (image RMS$\sim$0.45 mJy beam-1). In contrast, we did find a radio excess less than 0.6′′ away from the NIR position of the source in the first (2018 February) and second (2020 November) epochs of the Very Large Array Sky Survey (VLASS; Lacy et al. 2020) at 3 GHz. The peak flux density of the emission in the two epochs is 0.29$\pm$0.11 mJy beam-1 in the first and 0.40$\pm$0.13 mJy beam-1 in the second, which corresponds to a S/N of 2.6 and 3.0, respectively. Even though the two estimates are marginally consistent, we consider the average of the two and the overall range of uncertainty because of the possible intrinsic variability of the source: 0.35$\pm$0.18 mJy. In Tab. 2 we report the radio data and the 2.5$\sigma$ upper limits obtained from archival observations as described above. When we take currently available data with their uncertainties and the upper limits derived from non-detections into account, the spectral index of a single power law covering the observed frequency range is poorly constrained ($\alpha_{r}$=0–1.2). However, in addition to information on the flux density and the dimensions of the sources, the RACS catalogue also reports the spectral index computed within the 288 MHz band centred at 888 MHz. The spectral index reported for VIK J2318$-$3113 is $\alpha_{r}$=0.98, which is similar to what is typically observed in high-$z$ QSOs (e.g. Coppejans et al. 2017; Bañados et al. 2018). In the following, we consider this to be the best- fit value despite the relatively low S/N across the ASKAP band, even though a different assumption does not affect the results. A more detailed discussion of the broad-band radio properties of VIK J2318$-$3113 will be presented in a forthcoming work. Table 2: Estimates and 2.5$\sigma$ upper limits on the radio flux densities of VIK J2318$-$3113 from archival radio surveys. Survey: | TGSS | SUMSS | NVSS | VLASS ---|---|---|---|--- Obs. Freq. (GHz): | 0.15 | 0.843 | 1.4 | 3 Flux density (mJy/beam): | <7.3 | <6.3 | <1.1 | 0.35$\pm$0.18 ## 3 Optical/UV properties Given the high-redshift nature of VIK J2318$-$3113, the NIR photometric data from the VIKING survey (reported in Tab. 3) cover the UV/optical spectrum in its rest frame. Therefore we used these photometric points to estimate the bolometric luminosity (Lbol) of the source. In the following, we assume an optical spectral index given by the slope observed between the $K$ and $J$ bands, $\alpha_{{\mathrm{{o}}}}=0.54$, which is consistent with what is normally found in other QSOs (e.g. Vanden Berk et al. 2001). We started by computing the rest-frame monochromatic luminosities at 1350 and 3000 Å using the observed magnitudes in the filter with the closest corresponding rest- frame wavelength, that is, $Y$ ($\sim$1370 Å) and $K$ ($\sim$2860 Å), respectively. The bolometric luminosity can then be inferred using the correction factors derived in Shen et al. (2008) for 1350 Å ($L_{\mathrm{{bol}}}$= 8.0 $\pm$ 2.7 $\times$ 1046 erg s-1) and in Runnoe et al. (2012) for 3000 Å ($L_{\mathrm{{bol}}}$= 7.4 $\pm$ 0.3 $\times$ 1046 erg s-1). Averaging the two results with the corresponding variances as weights, we obtain $L_{\mathrm{{bol}}}$= 7.4 $\pm$ 0.3 $\times$ 1046 erg s-1. Assuming an Eddington-limited accretion, that is, Lbol $\leq$LEDD,666Where LEDD = 1.26 $\times$ 1038 (MBH/M⊙) erg s-1. this value of the bolometric luminosity implies that the SMBH mass must be higher than 6 $\times\leavevmode\nobreak\ 10^{8}$ M⊙. Table 3: NIR magnitudes of VIK J2318$-$3113 as measured in the VIKING survey (Vega system). Filter: | $Z$ | $Y$ | $J$ | $H$ | $K$ ---|---|---|---|---|--- $\lambda_{eff}$ ($\mu$m): | 0.878 | 1.021 | 1.254 | 1.646 | 2.149 magnitude: | 21.42 | 20.17 | 19.89 | 19.61 | 18.67 mag. error: | 0.11 | 0.08 | 0.11 | 0.18 | 0.14 ## 4 Radio loudness and comparison with high-$z$ RL QSOs In order to estimate the rest-frame monochromatic luminosity at 5 GHz, we considered the 888 MHz flux density obtained from the RACS observation (S888MHz= 1.44 mJy) with an uncertainty that takes the absolute calibration of the map into account (0.60 mJy, see previous section) and the spectral index reported in the RACS catalogue ($\alpha_{r}$=0.98). We also considered the GAMA23 flux density (S888MHz= 0.59$\pm$0.07 mJy) and a spectral index in the range $\alpha_{r}$=0–1.2 to estimate the associated uncertainty. The result, however, has little dependence on the $\alpha_{r}$ assumption because the observed frequency of 888 MHz corresponds to a rest-frame frequency of 6.6 GHz, which is very close to 5 GHz. The resulting radio luminosity is L5GHz= 1.2${}_{-0.9}^{+0.6}\,\times 10^{26}$ W Hz-1. Combining this estimate with the optical luminosity at 4400 Å (L4400Å= 1.8$\pm$0.1 $\times\leavevmode\nobreak\ 10^{31}$ erg s-1 Hz-1), computed from the observed $K$ magnitude, we obtain a radio loudness R= 66.3${}_{-46.7}^{+36.3}$. Adopting the typical value of R=10 as the threshold between RL and RQ sources, this makes VIK J2318$-$3113 the most distant RL QSO observed so far, at $z$=6.44. We note that this classification does not depend on the somewhat arbitrary criterion for separating the RL and RQ populations. Even if we consider a radio loudness as defined by Jiang et al. (2007)777In this case, the radio loudness is defined as the following rest-frame ratio: R=$S_{\mathrm{{5GHz}}}/S_{\mathrm{{2500\AA}}}$. or a single threshold in the radio luminosity (L5GHz>$10^{32.5}$ erg s-1 Hz-1; Jiang et al. 2007), the RL classification still holds. In RL QSOs, the radio emission is thought to be produced by relativistic jets and not by star-formation (SF) processes (e.g. Kellermann et al. 2016). VIK J2318$-$3113 was found to be very luminous in the far-IR (FIR) (log(LFIR/L⊙) in the range 11.89–12.46, between 42.5 and 122.5 $\mu$m; Decarli et al. 2018; Venemans et al. 2018, 2020), and this may imply that at least part of the observed radio emission is due to SF. However, considering the relation between radio and FIR luminosity observed in SF galaxies (Condon et al., 2002), we expect that only a few percent (<5%) of the observed radio emission can be produced by SF. This confirms that the high radio power observed in VIK J2318$-$3113 is likely produced by relativistic jets, as expected in RL sources. Interestingly, Venemans et al. (2020) also found that the FIR continuum and [C II] emissions extend up to $\sim$5 kpc (0.2′′) with an irregular morphology. Further radio observations at similar resolution would be fundamental for understanding the role of the different components at work in this complex QSO. Figure 2: Left: Rest-frame radio luminosity density at 5 GHz vs. the rest- frame optical luminosity density at 4400 Å for $z$>5.5 QSOs with a radio detection in the literature. Diagonal lines indicate constant radio-loudness values. Adapted from Bañados et al. (2015). Right: Radio loudness as a function of redshift for the $z$>5.5 confirmed RL QSOs compared to an optically selected sample of RL QSOs at lower redshift (orange points; Zhu et al. 2020) and all the RL QSOs at $z$>4 (yellow diamonds) known to date. The blue squares (circles) report $z$>5.5 (>6) sources in both graphs. The only confirmed $z$>5.5 blazar (Belladitta et al., 2020) is reported with a green triangle. At lower redshifts we did not distinguish this class because not all sources have a reliable classification. The red star represents VIK J2318$-$3113. Following Bañados et al. (2015), we report in Fig. 2 (left) the rest-frame radio luminosity (5 GHz) as a function of the rest-frame optical luminosity (4400 Å) for the updated list of $z$>5.5 QSOs with a radio observation and thus a firm RL/RQ classification888Data from Bañados et al. (2015, 2018), Belladitta et al. (2020), and Liu et al. (2021). Clearly, the radio loudness of VIK J2318$-$3113 is similar to that of the majority of $z$>5.5 RL QSOs, with 10<R<100\. Moreover, in Fig. 2 (right) we compare the confirmed RL QSOs at $z$>5.5 to the optically selected sample at lower redshift ($\sim$800 sources) discussed in Zhu et al. (2020) and to the $z$>4 RL QSOs discovered so far999These are all the RL QSOs published to date. To estimate their radio loudness, we considered the radio spectral index, if present; otherwise, we assumed $\alpha_{r}$=0.75 (Bañados et al., 2015) and considered a $\pm$0.25 variation to estimate the uncertainty. The full list of sources with the corresponding radio data and references will be presented in Belladitta et al. (in prep.).. Interestingly, only a small fraction of very radio-powerful high-$z$ sources (logR>2.5) has been found at $z$>5.5 compared to low redshifts. This may be a consequence of the fact that at these redshifts, QSOs have been selected mainly in the optical/UV, with only three radio-selected sources (which include the two radio-brightest sources at $z$>6). Nevertheless, we expect that upcoming and ongoing wide-area surveys such as RACS and the development of dedicated selection techniques in the radio band (e.g. Drouart et al. 2020) will find many more radio-powerful sources at $z$>6 (e.g. Amarantidis et al. 2019). ## 5 Conclusions We have presented the radio detection (at 888 MHz) of VIK J2318$-$3113, a $z$=6.44 QSO. Combining the new radio information from RACS with the archival data, we estimate a radio-loudness value of R$\sim$70, which means that this source is the most distant RL QSO observed to date. The radio association was made by cross-matching the first data release of the RACS survey and a list of the 169 previously discovered $z$>6 QSOs in the same area of the sky. As a result, we found radio counterparts for a total of three RL sources, VIK J2318$-$3113 included, which corresponds to a radio detection rate of $\sim$2% in the aforementioned list of $z$>6 QSOs. Because the RACS flux density limit is not deep enough to detect all the $z$>6 RL QSOs discovered so far, which have typical NIR magnitudes $\sim$22, this detection rate should be considered as a lower limit to the actual RL fraction at $z$>6\. We cannot fully characterise the radio spectral properties of VIK J2318$-$3113, and thus establish whether it is a flat, steep, or peaked source, with the currently available radio data. This is an important diagnostic for understanding the orientation of the relativistic jet with respect to the line of sight, that is, whether VIK J2318$-$3113 is a blazar. The possible presence of variability at 888 MHz, as found in the comparison of the RACS and GAMA23 observations, may suggest that the emission of this source is dominated by the relativistic beaming, which could mean that the jet is oriented at small angles from the line of sight. More data are required to confirm this result, however. Assuming an Eddington-limited accretion, the relatively high bolometric luminosity suggests the presence of a central SMBH with a mass $\gtrsim$6 $\times$ 108M⊙. This detection anticipates the discovery of many more RL high-$z$ sources in the next years when the new generation of all-sky radio surveys will be performed by the Square Kilometre Array and its precursors. ###### Acknowledgements. We thank the anonymous referee for the useful comments and suggestions. We acknowledge financial contribution from the agreement ASI-INAF n. I/037/12/0 and n.2017-14-H.0 and from INAF under PRIN SKA/CTA FORECaST. In this work we have used data from the ASKAP observatory. The Australian SKA Pathfinder is part of the Australia Telescope National Facility which is managed by CSIRO. Operation of ASKAP is funded by the Australian Government with support from the National Collaborative Research Infrastructure Strategy. ASKAP uses the resources of the Pawsey Supercomputing Centre. Establishment of ASKAP, the Murchison Radio-astronomy Observatory and the Pawsey Supercomputing Centre are initiatives of the Australian Government, with support from the Government of Western Australia and the Science and Industry Endowment Fund. We acknowledge the Wajarri Yamatji people as the traditional owners of the Observatory site. This paper includes archived data obtained through the CSIRO ASKAP Science Data Archive, CASDA (http://data.csiro.au). This research made use of Astropy (http://www.astropy.org) a community-developed core Python package for Astronomy (Astropy Collaboration et al., 2013, 2018). This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France (Wenger et al., 2000). ## References * Amarantidis et al. (2019) Amarantidis, S., Afonso, J., Messias, H., et al. 2019, MNRAS, 485, 2694 * Andika et al. (2020) Andika, I. T., Jahnke, K., Onoue, M., et al. 2020, ApJ, 903, 34 * Astropy Collaboration et al. (2018) Astropy Collaboration, Price-Whelan, A. M., Sipőcz, B. M., et al. 2018, AJ, 156, 123 * Astropy Collaboration et al. (2013) Astropy Collaboration, Robitaille, T. P., Tollerud, E. J., et al. 2013, A&A, 558, A33 * Bañados et al. (2018) Bañados, E., Venemans, B. P., Mazzucchelli, C., et al. 2018, Nature, 553, 473 * Bañados et al. (2015) Bañados, E., Venemans, B. P., Morganson, E., et al. 2015, ApJ, 804, 118 * Belladitta et al. (2020) Belladitta, S., Moretti, A., Caccianiga, A., et al. 2020, A&A, 635, L7 * Blandford et al. (2019) Blandford, R., Meier, D., & Readhead, A. 2019, ARA&A, 57, 467 * Chambers et al. (2016) Chambers, K. C., Magnier, E. A., Metcalfe, N., et al. 2016, eprint arXiv:1612.05560 [arXiv:1612.05560] * Condon et al. (2002) Condon, J. J., Cotton, W. D., & Broderick, J. J. 2002, AJ, 124, 675 * Condon et al. (1998) Condon, J. J., Cotton, W. D., Greisen, E. W., et al. 1998, AJ, 115, 1693 * Coppejans et al. (2017) Coppejans, R., van Velzen, S., Intema, H. T., et al. 2017, MNRAS, 467, 2039 * Dark Energy Survey Collaboration et al. (2016) Dark Energy Survey Collaboration, Abbott, T., Abdalla, F. B., et al. 2016, MNRAS, 460, 1270 * Decarli et al. (2018) Decarli, R., Walter, F., Venemans, B. P., et al. 2018, ApJ, 854, 97 * Driver et al. (2011) Driver, S. P., Hill, D. T., Kelvin, L. S., et al. 2011, MNRAS, 413, 971 * Drouart et al. (2020) Drouart, G., Seymour, N., Galvin, T. J., et al. 2020, PASA, 37, e026 * Edge et al. (2013) Edge, A., Sutherland, W., Kuijken, K., et al. 2013, The Messenger, 154, 32 * Fan et al. (2019) Fan, X., Wang, F., Yang, J., et al. 2019, ApJ, 870, L11 * Fomalont (1999) Fomalont, E. B. 1999, in Astronomical Society of the Pacific Conference Series, Vol. 180, Synthesis Imaging in Radio Astronomy II, ed. G. B. Taylor, C. L. Carilli, & R. A. Perley, 301 * Gaikwad et al. (2020) Gaikwad, P., Rauch, M., Haehnelt, M. G., et al. 2020, MNRAS, 494, 5091 * Hovatta et al. (2008) Hovatta, T., Nieppola, E., Tornikoski, M., et al. 2008, A&A, 485, 51 * Intema et al. (2017) Intema, H. T., Jagannathan, P., Mooley, K. P., & Frail, D. A. 2017, A&A, 598, A78 * Jiang et al. (2007) Jiang, L., Fan, X., Ivezić, Ž., et al. 2007, ApJ, 656, 680 * Johnston et al. (2008) Johnston, S., Taylor, R., Bailes, M., et al. 2008, Experimental Astronomy, 22, 151 * Kashikawa et al. (2006) Kashikawa, N., Shimasaku, K., Malkan, M. A., et al. 2006, ApJ, 648, 7 * Kellermann et al. (2016) Kellermann, K. I., Condon, J. J., Kimball, A. E., Perley, R. A., & Ivezić, Ž. 2016, ApJ, 831, 168 * Kellermann et al. (1989) Kellermann, K. I., Sramek, R., Schmidt, M., Shaffer, D. B., & Green, R. 1989, AJ, 98, 1195 * Lacy et al. (2020) Lacy, M., Baum, S. A., Chandler, C. J., et al. 2020, PASP, 132, 035001 * Liu et al. (2021) Liu, Y., Wang, R., Momjian, E., et al. 2021, ApJ, 908, 124 * Matsuoka et al. (2019) Matsuoka, Y., Onoue, M., Kashikawa, N., et al. 2019, ApJ, 872, L2 * Mauch et al. (2003) Mauch, T., Murphy, T., Buttery, H. J., et al. 2003, MNRAS, 342, 1117 * Mazzucchelli et al. (2017) Mazzucchelli, C., Bañados, E., Venemans, B. P., et al. 2017, ApJ, 849, 91 * McConnell et al. (2020) McConnell, D., Hale, C. L., Lenc, E., et al. 2020, PASA, 37, e048 * McGreer et al. (2006) McGreer, I. D., Becker, R. H., Helfand, D. J., & White, R. L. 2006, ApJ, 652, 157 * McMullin et al. (2007) McMullin, J. P., Waters, B., Schiebel, D., Young, W., & Golap, K. 2007, in Astronomical Society of the Pacific Conference Series, Vol. 376, Astronomical Data Analysis Software and Systems XVI, ed. R. A. Shaw, F. Hill, & D. J. Bell, 127 * Norris et al. (2011) Norris, R. P., Hopkins, A. M., Afonso, J., et al. 2011, PASA, 28, 215 * Retana-Montenegro & Röttgering (2017) Retana-Montenegro, E. & Röttgering, H. J. A. 2017, A&A, 600, A97 * Runnoe et al. (2012) Runnoe, J. C., Brotherton, M. S., & Shang, Z. 2012, MNRAS, 422, 478 * Seymour et al. (2020) Seymour, N., Huynh, M., Shabala, S. S., et al. 2020, PASA, 37, e013 * Shen et al. (2008) Shen, Y., Greene, J. E., Strauss, M. A., Richards, G. T., & Schneider, D. P. 2008, ApJ, 680, 169 * Stern et al. (2000) Stern, D., Djorgovski, S. G., Perley, R. A., de Carvalho, R. R., & Wall, J. V. 2000, AJ, 119, 1526 * Vanden Berk et al. (2001) Vanden Berk, D. E., Richards, G. T., Bauer, A., et al. 2001, AJ, 122, 549 * Venemans et al. (2018) Venemans, B. P., Decarli, R., Walter, F., et al. 2018, ApJ, 866, 159 * Venemans et al. (2020) Venemans, B. P., Walter, F., Neeleman, M., et al. 2020, ApJ, 904, 130 * Volonteri et al. (2015) Volonteri, M., Silk, J., & Dubus, G. 2015, ApJ, 804, 148 * Wang et al. (2020) Wang, F., Fan, X., Yang, J., et al. 2020, arXiv e-prints, arXiv:2011.12458 * Wang et al. (2021) Wang, F., Yang, J., Fan, X., et al. 2021, arXiv e-prints, arXiv:2101.03179 * Wang et al. (2019) Wang, F., Yang, J., Fan, X., et al. 2019, ApJ, 884, 30 * Wenger et al. (2000) Wenger, M., Ochsenbein, F., Egret, D., et al. 2000, A&AS, 143, 9 * Willott et al. (2010) Willott, C. J., Delorme, P., Reylé, C., et al. 2010, AJ, 139, 906 * Yang et al. (2020a) Yang, J., Wang, F., Fan, X., et al. 2020a, ApJ, 897, L14 * Yang et al. (2020b) Yang, J., Wang, F., Fan, X., et al. 2020b, ApJ, 904, 26 * Zhu et al. (2020) Zhu, S. F., Brandt, W. N., Luo, B., et al. 2020, MNRAS, 496, 245
# Photon-photon interactions in Rydberg-atom arrays Lida Zhang (gbsn张理达)1 Valentin Walther1,2 Klaus Mølmer1 Thomas Pohl1 1Center for Complex Quantum Systems, Department of Physics and Astronomy, Aarhus University, DK-8000 Aarhus C, Denmark 2ITAMP, Harvard-Smithsonian Center for Astrophysics, Cambridge, Massachusetts 02138, USA ###### Abstract We investigate the interaction of weak light fields with two-dimensional lattices of atoms with high lying atomic Rydberg states. This system features different interactions that act on disparate length scales, from zero-range defect scattering of atomic excitations and finite-range dipolar exchange processes to long-range Rydberg-state interactions, which span the entire array and can block multiple Rydberg excitations. Analyzing their interplay, we identify conditions that yield a nonlinear quantum mirror which coherently splits incident fields into correlated photon-pairs in a single transverse mode, while transmitting single photons unaffected. In particular, we find strong anti-bunching of the transmitted light with equal-time pair correlations that decrease exponentially with an increasing range of the Rydberg blockade. Such strong photon-photon interactions in the absence of photon losses open up promising avenues for the generation and manipulation of quantum light, and the exploration of many-body phenomena with interacting photons. Photons typically cross each other unimpeded. Yet, the scientific and technological prospects of effective photon interactions Chang _et al._ (2014) have motivated substantial research efforts into nonlinear optical processes at the ultimate quantum level. Here, optical resonators Birnbaum _et al._ (2005); Volz _et al._ (2014); Reiserer and Rempe (2015); Welte _et al._ (2018); O’Shea _et al._ (2013) and nano-scale photonic structures Goban _et al._ (2012); Thompson _et al._ (2013); Tiecke _et al._ (2014); Petersen _et al._ (2014); Lodahl _et al._ (2015); Chang _et al._ (2018); Noaman (2018); Yu _et al._ (2019); Prasad _et al._ (2020) have made it possible to couple photons to single saturable emitters, and strong interactions between highly excited atoms have been used to realize large optical nonlinearities in atomic ensembles Pritchard _et al._ (2010); Peyronel _et al._ (2012); Thompson _et al._ (2017); Paris-Mandoki _et al._ (2017); Murray and Pohl (2016); Firstenberg _et al._ (2016). The use of many-particle systems to reach a strong collective light-matter coupling present an attractive approach, and the exploitation of Rydberg-state interactions in atomic gases has indeed enabled recent breakthroughs that, for example, demonstrated single-photon switching Baur _et al._ (2014); Gorniaczyk _et al._ (2014); Tiarks _et al._ (2014) and photonic quantum gates Tiarks _et al._ (2019). Yet, photon losses that are intrinsic to such ensemble approaches Gorshkov _et al._ (2013); Murray _et al._ (2018) limit the performance of these applications Baur _et al._ (2014); Murray _et al._ (2016) and severely challenge the exploration of correlated quantum states Otterbach _et al._ (2013) beyond the few-photon regime Zeuthen _et al._ (2017); Bienias _et al._ (2020). At the same time, studies of ordered arrangements, instead of random atomic ensembles, have revealed a number of exciting linear optical properties Facchinetti _et al._ (2016); Perczel _et al._ (2017); Bettles _et al._ (2017); Guimond _et al._ (2019); Ballantine and Ruostekoski (2020) that arise from the many-body nature of light-matter interactions in these systems. In particular, the cooperative response of two-dimensional arrays facilitates strong photon coupling at greatly reduced losses Bettles _et al._ (2016); Shahmoon _et al._ (2017), as recently demonstrated experimentally with ultracold atoms in optical lattices Rui _et al._ (2020). Figure 1: (Color online) (a) A regular array of atoms interacts with weak coherent light and can induce nonclassical correlations in the transmitted probe field. (d) Its amplitude, $\mathcal{E}$, couples the ground states, $|g\rangle$, of the atoms to an excited state, $|e\rangle$, which is coupled to a high-lying Rydberg state $|s\rangle$ by an additional control field with a Rabi frequency $\Omega$. Atomic interactions, induced by the driven dipoles of the lower transition, lead to a collective energy shift, $\Delta_{c}$, and collective photon emission of the excited array with a rate $\Gamma_{c}$. For $\Omega=0$, this can result in near-perfect reflection [red line in (e)] when the probe-field detuning $\Delta$ matches $\Delta_{c}$, while EIT of the three-level system yields perfect transmission on two-photon resonance for a finite control-field coupling ($\Omega\neq 0$) [blue line in (e)]. (c) The van der Waals interaction between Rydberg atoms can be strong enough to inhibit the excitation of more than a single Rydberg state within a blockade area (blue circle) that may cover the entire array. In combination with EIT and the collective photon reflection of the array, this provides a nonlinear mechanism for strong coherent photon interactions that can generate highly non-classical states of light. This is shown by the strong pair correlations of transmitted ($\rightarrow$) and reflected ($\leftarrow$) photons in panel (b). In this work, we investigate the effects of strong Rydberg-state interactions in regular atomic arrays [Fig.1(a)], and analyse their _nonlinear cooperative_ response. Rydberg-state interactions can be used to couple a single atom to adjacent lattices Grankin _et al._ (2018); Bekenstein _et al._ (2020), and here we show that atomic interactions within Rydberg arrays can generate strong photon-photon interactions at greatly suppressed losses. This quantum optical nonlinearity emerges from the interplay of various atomic interactions and a narrow transmission feature [Fig.1(e)] that arises from three-level photon coupling [Fig.1(d)] under conditions of electromagnetically induced transparency (EIT), and can produce highly correlated states of light [see Fig.1(b)]. Compared to two-level systems, where few-photon nonlinearities can also arise in small arrays with very small distances, $\lesssim 100$nm Cidrim _et al._ (2020); Williamson _et al._ (2020), or be induced via Rydberg- dressing of low-lying states Moreno-Cardoner _et al._ (2021); Henkel _et al._ (2010), Rydberg-EIT in the present lattice-setting provides strong photon coupling and facilitates large nonlinearities under conditions of present experiments Rui _et al._ (2020); Zeiher _et al._ (2015). We consider a two dimensional regular array of closely spaced atoms at positions ${\bf r}_{j}$, as illustrated in Fig.1(a). A weak probe field with an amplitude $\mathcal{E}$ drives the transition between the ground state $|g\rangle$ and an intermediate state $|e\rangle$ at a frequency detuning $\Delta$, while a high-lying Rydberg state $|s\rangle$ is excited by an additional control field with a Rabi frequency $\Omega$ [Fig.1(d)]. The combined action of both light fields leads to two distinct types of atomic interactions that act on vastly different length scales. First, the Rydberg atoms feature van der Waals interactions that can be sufficiently strong to block the excitation of multiple Rydberg $|s\rangle$-states within distances of several micrometers Jaksch _et al._ (2000); Lukin _et al._ (2001). This Rydberg blockade has been explored for a range of applications Saffman _et al._ (2010); Adams _et al._ (2019). In particular, it has already been demonstrated in dense atomic lattices Zeiher _et al._ (2015), with an excitation blockade over distances of more than $\sim 10$ sites. We focus here on configurations in which the entire atomic array is covered by the blockade radius, and quantum states with more than a single Rydberg excitation are blocked by the strong atomic interaction [see Fig.1(c)]. Second, a small lattice constant $a\sim\lambda$, on the order of the $|g\rangle-|e\rangle$ transition wavelength $\lambda$, entails strong dipole- dipole interactions that arise from near-resonant photon exchange on the probe transition James (1993); Dung _et al._ (2002); Asenjo-Garcia _et al._ (2017), which leads to coherent exchange of atomic $|e\rangle$-excitations across the atomic array. This results in a collective optical response that can greatly suppress photon scattering and generate near-perfect coherent coupling to the single transverse mode of the incident field. One can find superradiant as well as subradiant states of a single de-localized $|e\rangle$-excitation, whose collective emission rate $\Gamma_{c}$ is respectively enhanced or suppressed relative to the single-atom decay rate $\Gamma$ Zoubi and Ritsch (2011); Sutherland and Robicheaux (2016); Facchinetti _et al._ (2016); Bettles _et al._ (2015); Guimond _et al._ (2019); Zhang and Mølmer (2019); Piñeiro Orioli and Rey (2019). For large extended arrays, the resulting collective level shift Glicenstein and Ferioli (2020), $\Delta_{c}$ of the $|g\rangle-|e\rangle$ transition marks the spectral position of reflection resonances at which an incoming photon is reflected perfectly Bettles _et al._ (2016); Shahmoon _et al._ (2017), without scattering into other transverse modes. Moreover, the control-field coupling to the Rydberg-state permits to control the optical response on the $|g\rangle-|e\rangle$ transition. In particular, on two-photon resonance, the three-level system features a dark eigenstate that does not contain the intermediate state $|e\rangle$ Fleischhauer and Lukin (2000); Arimondo (1996), and therefore enables lossless transmission of the incident light due to EIT Fleischhauer _et al._ (2005). As illustrated in Fig. 1(e), EIT is restricted to a narrow transparency window in the reflection spectrum of the array. Its width, $\Omega^{2}/\Gamma_{c}$, permits to control reflectivity by tuning the intensity of the classical control field Manzoni _et al._ (2018); Bekenstein _et al._ (2020). Figure 2: (Color online) (a) Coefficients for linear reflection ($R$), transmission ($T$) and loss ($L$) of the incident probe light for an array of two-level atoms ($\Omega=0$) with a circular boundary as shown in the inset. The depicted dependence on the waist, $w_{0}$, of the probe beam shows a maximum, virtually perfect reflection of $R\simeq 0.975$ at $w_{0}\simeq 2\lambda$ for an optimized lattice constant $a=0.75\lambda$ and probe detuning $\Delta=0.05\Gamma$. Panel (b) shows the average change of the linear response coefficients for identical parameters when adding a Rydberg defect in the form of an empty lattice site at ${\bf r}_{j}$ with a probability $p_{j}\propto|\mathcal{E}({\bf r}_{j})|^{2}$. A quantum mechanical switching mechanism can emerge from the strong interaction between the Rydberg states. Hereby, the Rydberg blockade of multiple atomic dark states exposes the reflective two-level response to multi-photon states, while single probe photons can pass the array unimpeded. Such a nonlinearity may yield effective photon-photon interactions that can operate at the level of single photons and greatly suppressed scattering losses. We have studied this behavior by solving the Master equation, $\partial_{t}\hat{\rho}(t)=-i[\hat{H},\hat{\rho}]+\mathcal{L}(\hat{\rho})$, for the density matrix, $\hat{\rho}$, of the atomic array. The Hamiltonian, $\hat{H}=\hat{H}_{\rm LA}+\hat{H}_{\rm dd}$ contains the light-atom coupling in the rotating wave approximation $\hat{H}_{\rm LA}=-\sum^{N}_{j=1}\left[g\mathcal{E}({\bf r}_{j})\hat{\sigma}^{(j)}_{eg}+\Omega\hat{\sigma}^{(j)}_{es}+{\rm h.c.}\right]+\Delta\hat{\sigma}^{(j)}_{ee},$ (1) where $\hat{\sigma}_{\alpha\beta}^{(j)}=|\alpha_{j}\rangle\langle\beta_{j}|$ denote the projection and transition operators for the $j$th atom, $g$ denotes the atom-photon coupling strength. The probe-field amplitude follows the paraxial wave equation $[4\pi i\partial_{z}+\lambda\nabla_{\perp}^{2}]\mathcal{E}=0$, and is normalized such that $|\mathcal{E}|^{2}$ yields a spatial photon density. The remaining photonic dynamics can be integrated out to obtain a Hamiltonian $\hat{H}_{\text{dd}}=-\sum_{i\neq j}J_{ij}\hat{\sigma}^{(i)}_{eg}\hat{\sigma}^{(j)}_{ge}$ (2) and Liouvillian $\mathcal{L}(\rho)=\sum^{N}_{i,j}\frac{1}{2}\Gamma_{ij}(2\hat{\sigma}^{(j)}_{ge}\rho\hat{\sigma}^{(i)}_{eg}-\hat{\sigma}^{(i)}_{eg}\hat{\sigma}^{(j)}_{ge}\rho-\rho\hat{\sigma}^{(i)}_{eg}\hat{\sigma}^{(j)}_{ge})$ (3) that describe the light-induced atomic interactions within the Born-Markov approximation Asenjo-Garcia _et al._ (2017). The interaction coefficients $J_{ij}$ and $\Gamma_{ij}$ for two atoms at positions ${\bf r}_{i}$ and ${\bf r}_{j}$ are determined by the Greens function tensor of the free-space electromagnetic field. Knowing the dipolar field from each atom, one can readily reconstruct the mean values and correlation functions of the photonic field from the solution, $\hat{\rho}$, of the driven atomic many-body dynamics. While this yields the entirety of the emitted light field, we focus here on its projection onto the single transverse mode of the driving field $\mathcal{E}$. This yields simple relations $\displaystyle\hat{a}_{\rightarrow}(t)=$ $\displaystyle\sqrt{\mathcal{P}}+i\frac{g}{c\sqrt{\mathcal{P}}}\sum_{j}\mathcal{E}^{*}({\bf r}_{j})\hat{\sigma}_{eg}^{(j)}(t),$ (4a) $\displaystyle\hat{a}_{\leftarrow}(t)=$ $\displaystyle i\frac{g}{c\sqrt{\mathcal{P}}}\sum_{j}\mathcal{E}^{*}({\bf r}_{j})\hat{\sigma}_{eg}^{(j)}(t)$ (4b) for the photon operators of the transmitted ($\hat{a}_{\rightarrow}$) and reflected ($\hat{a}_{\leftarrow}$) field modes, in terms of the atomic transition operators. Here, $\mathcal{P}=\int|\mathcal{E}({\bf r})|^{2}{\rm d}{\bf r}_{\perp}$ denotes the transverse integral over the input intensity profile and defines the probe beam power, which is conserved along the propagation. Figure 3: (Color online) (a) Equal-time two-photon correlation function of the transmitted light as a function of diameter of the array for $\sqrt{\mathcal{P}}=0.01\sqrt{\Gamma/c}$. Parameters are optimized to obtain a maximum linear reflection as shown in Fig.2(a). Panel (b) shows the correlation function as a function of the probe-beam waist for large arrays, $\ell\rightarrow\infty$. For reference, let us first consider the optical properties of the two-level array in the absence of Rydberg-state coupling ($\Omega=0$). While only an infinitely extended array can perfectly reflect an incident plane wave Shahmoon _et al._ (2017), finite arrays can yield high reflection for a judicious choice of the system parameters. This is illustrated in Fig. 2(a), where we show the steady-state reflectivity $R=\langle\hat{a}_{\leftarrow}^{\dagger}\hat{a}_{\leftarrow}\rangle/\mathcal{P}$ along with the transmission coeffcient $T=\langle\hat{a}_{\rightarrow}^{\dagger}\hat{a}_{\rightarrow}\rangle/\mathcal{P}$ and loss $L=1-T-R$ for circular disc-shaped arrays with a diameter of $\ell$ atoms and a Gaussian driving mode $|\mathcal{E}|=\sqrt{2\mathcal{P}/(\pi w^{2})}{\rm e}^{-r_{\perp}^{2}/w^{2}}$, whose width changes as $w^{2}=w_{0}^{2}+\lambda^{2}z^{2}/(\pi^{2}w_{0}^{2})$ along the propagation direction and has its waist centered at the mirror position (Fig.1). Already for an array with a $10$-atom diameter ($\ell=10$), one can reach near-unity reflection of $R\simeq 0.975$ for a beam waist of only $w_{0}\simeq 2\lambda$. The linear reflectivity for a finite Rydberg-state coupling, $\Omega$, is determined by the atomic dark state $|D\rangle\propto\Omega|G\rangle-g\sum_{j}\mathcal{E}({\bf r}_{j})\hat{\sigma}_{sg}^{(j)}|G\rangle$, where $|G\rangle$ denotes the $N$-atom ground state. Owing to the long Rydberg-atom lifetime, this state does not suffer from spontaneous emission and hence it facilitates a vanishing reflection and perfect transmission of the probe field. The nonlinear response arises from the intricate interplay between EIT and the different atomic interactions, from (i) local defect interactions with atomic excitations, and (ii) finite-range photon-mediated dipole-dipole interactions, to (iii) the long-range Rydberg interactions that extent across the array. We can estimate the first effect by sampling a single empty site of the array from the probability distribution $p_{j}\propto|\mathcal{E}({\bf r}_{j})|^{2}$ of generated Rydberg impurities. As shown in Fig.2(b), such a single de- localized Rydberg impurity can have significant consequences for the optical response, causing transverse photon scattering at the expense of the two-level reflection coefficient. We have performed stochastic wave function simulations Mølmer _et al._ (1993) to solve the $N$-body master equation determined by Eqs.(1)-(3). For sufficiently weak probe fields, one can truncate the many-body wave function of the atoms at maximally two $|e\rangle$-excitations. This describes the physics of two interacting photons, which can be analyzed via the two-photon densities $\rho_{{\begin{subarray}{c}\alpha\\\ \beta\end{subarray}}}(t,t^{\prime})=\langle\hat{a}_{\alpha}^{\dagger}(t)\hat{a}_{\beta}^{\dagger}(t^{\prime})\hat{a}_{\beta}(t^{\prime})\hat{a}_{\alpha}(t)\rangle,$ (5) and the associated correlation functions $g^{(2)}_{{\begin{subarray}{c}\alpha\\\ \beta\end{subarray}}}(|t-t^{\prime}|)\equiv\rho_{{\begin{subarray}{c}\alpha\\\ \beta\end{subarray}}}(t,t^{\prime})/(\langle\hat{a}_{\alpha}^{\dagger}(t)\hat{a}_{\alpha}(t)\rangle\langle\hat{a}_{\beta}^{\dagger}(t^{\prime})\hat{a}_{\beta}(t^{\prime})\rangle)$, where $\alpha,\beta=\rightarrow,\leftarrow$ labels the forward and backward propagating mode of emitted probe photons, as also indicated in Eqs.(4). The pair-correlation functions only depend on the time difference $\tau=|t-t^{\prime}|$ in the steady-state under cw-driving. Figure 4: (Color online) Pair correlation function of the transmitted (a) and reflected light (b) for different values of the control-field Rabi frequency $\Omega$, $\ell=10$, $w_{0}=1.7\lambda$, and otherwise identical parameters as in Fig. 3. For both correlation functions, all data approximately collapse onto a single curve upon scaling the time between successive photon detections by the EIT delay time $\tau_{d}$, as given in Eq.(6). In Fig.3(a) we show the equal-time two-photon correlations $g_{\rightrightarrows}^{(2)}(0)$ of the transmitted light for different sizes of the array, assuming parameters optimized to maximize the linear reflection. As can be seen, the Rydberg blockade can lead to strongly antibunched transmitted light. This is possible because the Rydberg-state component of the dark state, $|D\rangle$, generated by absorption of one photon, blocks excitation of further atoms into the dark state and therefore suppresses the simultaneous transmission of multiple photons. We find stronger anti-bunching for larger arrays, i.e. a rapid drop of $g_{\rightrightarrows}^{(2)}(0)$ with increasing size $\ell$ of the array. This behaviour arises from the effects of Rydberg impurities, discussed above and illustrated in Fig.2(b). A larger size of the array and illuminated area implies a lower density of the single Rydberg impurity, and therefore improves the efficiency of the nonlinear reflection. Simulations for a fixed beam waist and increasing lattice size show convergence to a finite value of $g_{\rightrightarrows}^{(2)}(0)$ for sufficiently large arrays. These asymptotic values are shown in Fig.3(b) and reveal a rapid exponential drop with increasing waist of the probe beam and yields strong anti-bunching with $g_{\rightrightarrows}^{(2)}(0)<0.1$ already for remarkably small values $w_{0}\sim 1.5\lambda$. These results demonstrate the strong suppression of multi-photon transmission, and the two-photon density depicted in Fig.1(b) shows how incident photons are rerouted by their nonlinear interaction with the Rydberg array. Here, the two- time density defined in Eq.(5) has been converted to a spatial steady-state correlation function, using the relation between spatial photon position and time, $z=ct$ by the speed of light, $c$. In particular, we see that the two- photon density, $\rho_{\rightleftarrows}(z,z^{\prime})=\rho_{\leftrightarrows}(z^{\prime},z)$ for counter-propagating photon pairs continuously connects to the density, $\rho_{\leftleftarrows}(z,z^{\prime})$, of simultaneously reflected photon pairs. This indicates that the two-photon component of the incident probe field is symmetrically rerouted into these two modes. We can understand this behavior as follows. In the linear limit and in the absence of EIT ($\Omega=0$), the weak probe field $\mathcal{E}$ only drives weak atomic excitations with small transition dipole moments determined by $\hat{\sigma}_{eg}^{(j)}$. At the reflection maxima, the field generated by these weak atomic dipoles just cancels the probe field in Eq.(4a) and therefore yields high reflection with $\hat{a}_{\leftarrow}\approx\sqrt{\mathcal{P}}$ according to Eq.(4b). In the opposite limit of perfect EIT, the atomic dipoles vanish entirely, leading to perfect transmission with $\hat{a}_{\rightarrow}=\sqrt{\mathcal{P}}$ and $\hat{a}_{\leftarrow}=0$, according to Eqs.(4). The nonlinear response, however, differs fundamentally, because the detection of the first reflected photon causes a projection of the $N$-atom wavefunction into a state with a definite de-localized excitation. This unit-probability, heralded excitation consequently contributes a much stronger emission that vastly overwhelms the incident probe field amplitude. Following Eq.(4), the subsequent conditioned photon emission becomes virtually symmetric, with $\hat{a}_{\rightarrow}\approx\hat{a}_{\leftarrow}$ and leads to the typical form of the correlated two-photon density shown in Fig.1(b). Fig.4 offers further insights into the dynamics of the nonlinear photon interaction. From the linear response of the array under EIT conditions, we find that a transmitted light pulse experiences a delay of $\tau_{d}=\frac{\Gamma_{c}}{2\Omega^{2}},$ (6) which coincides with the inverse width of the transparency window shown in Fig.1(e). This pulse delay emerges in analogy to slow-light propagation through an extended EIT medium Fleischhauer and Lukin (2000), and corresponds to the average time for which a transmitted photon is transferred to the de- localized Rydberg state $\sim\sum_{j}\mathcal{E}({\bf r}_{j})\hat{\sigma}_{sg}^{(j)}|G\rangle$ and blocks EIT for any other incident photons. The pair correlation functions and two-photon densities, depicted in Fig.4, accurately corroborate this picture, showing the same characteristic correlation time $\tau_{d}$ for bunched and anti-bunched photon states of the reflected and transmitted light for varying values of the control-field Rabi frequency $\Omega$. At the same time, we find that the outgoing photons maintain a high degree of coherence, as quantified by $g_{\alpha}^{(1)}(t)=\langle\hat{a}_{\alpha}^{\dagger}(t)\hat{a}_{\alpha}(0)\rangle/\langle\hat{a}_{\alpha}^{\dagger}(0)\hat{a}_{\alpha}(0)\rangle\sim 1$, on the scale of the characteristic correlation time of both fields, which reflects the suppression of photon loss and decoherence by the collective light-matter coupling of the ordered array. This combination of high coherence, low photon losses and strong photon-photon interactions offers a promising outlook for the generation and manipulation of non-classical light in optical-lattice experiments that have already demonstrated Rydberg blockade of more then $\sim 100$ atoms Zeiher _et al._ (2015) as well as efficient photon reflection by arrays with sub-wavelength lattice constants Rui _et al._ (2020). The demonstrated nonlinearity is akin to that of wave guide QED with single few-level emitters, whereby the eliminated scattering into other transverse modes effectively corresponds to a near-perfect coupling into a single guided mode. This limit of strong coherent photon coupling has thus far been difficult to reach in atomic systems Prasad _et al._ (2020); Stiesdal _et al._ (2021), but will enable a range of applications, from the generation of single narrow-bandwidth photons Parkins _et al._ (1993), and logic gates Ralph _et al._ (2015) to few-photon routing and sorting Witthaut _et al._ (2012). The Rydberg array can hereby be employed as an active or passive element under pulsed or cw operation, exploiting the additional temporal control provided by the control-field coupling. While we have focussed here on the few-photon domain in order to analyse the basic interaction mechanism, the multi-photon regime under strong- driving conditions provides an exciting perspectives for exploring quantum optical many-body phenomena. Similarly to cavity-QED with single emitters, the described nonlinearities may be further enhanced by positioning the array in front of mirrors or inside optical cavities Shahmoon _et al._ (2020). Arrangements of multiple Rydberg arrays, or more complex 3D configurations could be constructed with atoms in configurable optical tweezer arrays to form networks of quantum beam splitters and nonlinear resonators that also exploit multi-photon and multi-mode interference effects and may supplement proposals for quantum enhanced interferometry Demkowicz-Dobrzański _et al._ (2015). This work was supported by the Carlsberg Foundation through the ’Semper Ardens’ Research Project QCooL, by the NSF through a grant for ITAMP at Harvard University, by the DFG through the SPP1929, by the European Commission through the H2020-FETOPEN project ErBeStA (No. 800942), and by the Danish National Research Foundation through the Center of Excellence ”CCQ” (Grant agreement no.: DNRF156). _Note added_ : During completing of this manuscript we became aware of a related work Moreno-Cardoner _et al._ (2021) that describes nonlinearities in two-level arrays where finite-range interactions are induced by Rydberg dressing Henkel _et al._ (2010). ## References * Chang _et al._ (2014) D. E. Chang, V. Vuletić, and M. D. Lukin, Nature Photonics 8, 685 (2014). * Birnbaum _et al._ (2005) K. M. Birnbaum, A. Boca, R. Miller, A. D. Boozer, T. E. Northup, and H. J. Kimble, Nature 436, 87 (2005). * Volz _et al._ (2014) J. Volz, M. Scheucher, C. Junge, and A. Rauschenbeutel, Nature Photonics 8, 965 (2014). * Reiserer and Rempe (2015) A. Reiserer and G. Rempe, Rev. Mod. Phys. 87, 1379 (2015). * Welte _et al._ (2018) S. Welte, B. Hacker, S. Daiss, S. Ritter, and G. Rempe, Phys. Rev. X 8, 011018 (2018). * O’Shea _et al._ (2013) D. O’Shea, C. Junge, J. Volz, and A. Rauschenbeutel, Phys. Rev. Lett. 111, 193601 (2013). * Goban _et al._ (2012) A. Goban, K. S. Choi, D. J. Alton, D. Ding, C. Lacroûte, M. Pototschnig, T. Thiele, N. P. Stern, and H. J. Kimble, Phys. Rev. Lett. 109, 033603 (2012). * Thompson _et al._ (2013) J. D. Thompson, T. G. Tiecke, N. P. de Leon, J. Feist, A. V. Akimov, M. Gullans, A. S. Zibrov, V. Vuletić, and M. D. Lukin, Science 340, 1202 (2013). * Tiecke _et al._ (2014) T. G. Tiecke, J. D. Thompson, N. P. de Leon, L. R. Liu, V. Vuletić, and M. D. Lukin, Nature 508, 241 (2014). * Petersen _et al._ (2014) J. Petersen, J. Volz, and A. Rauschenbeutel, Science 346, 67 (2014). * Lodahl _et al._ (2015) P. Lodahl, S. Mahmoodian, and S. Stobbe, Rev. Mod. Phys. 87, 347 (2015). * Chang _et al._ (2018) D. E. Chang, J. S. Douglas, A. González-Tudela, C.-L. Hung, and H. J. Kimble, Rev. Mod. Phys. 90, 031002 (2018). * Noaman (2018) M. Noaman, M. Langbecker, and P. Windpassinger, Opt. Lett. 43, 3925 (2018). * Yu _et al._ (2019) S.-P. Yu, J. A. Muniz, C.-L. Hung, and H. J. Kimble, Proceedings of the National Academy of Sciences 116, 12743 (2019) . * Prasad _et al._ (2020) A. S. Prasad, J. Hinney, S. Mahmoodian, K. Hammerer, S. Rind, P. Schneeweiss, A. S. Sørensen, J. Volz, and A. Rauschenbeutel, Nature Photonics 14, 719 (2020). * Pritchard _et al._ (2010) J. D. Pritchard, D. Maxwell, A. Gauguet, K. J. Weatherill, M. P. A. Jones, and C. S. Adams, Phys. Rev. Lett. 105, 193603 (2010). * Peyronel _et al._ (2012) T. Peyronel, O. Firstenberg, Q.-Y. Liang, S. Hofferberth, A. V. Gorshkov, T. Pohl, M. D. Lukin, and V. Vuletić, Nature 488, 57 (2012). * Thompson _et al._ (2017) J. D. Thompson, T. L. Nicholson, Q.-Y. Liang, S. H. Cantu, A. V. Venkatramani, S. Choi, I. A. Fedorov, D. Viscor, T. Pohl, M. D. Lukin, and V. Vuletić, Nature 542, 206 (2017). * Paris-Mandoki _et al._ (2017) A. Paris-Mandoki, C. Braun, J. Kumlin, C. Tresp, I. Mirgorodskiy, F. Christaller, H. P. Büchler, and S. Hofferberth, Phys. Rev. X 7, 041010 (2017). * Murray and Pohl (2016) C. Murray and T. Pohl, Advances In Atomic, Molecular, and Optical Physics 65, 321 (2016). * Firstenberg _et al._ (2016) O. Firstenberg, C. S. Adams, and S. Hofferberth, J. Phys B 49, 152003 (2016). * Baur _et al._ (2014) S. Baur, D. Tiarks, G. Rempe, and S. Dürr, Phys. Rev. Lett. 112, 073901 (2014). * Gorniaczyk _et al._ (2014) H. Gorniaczyk, C. Tresp, J. Schmidt, H. Fedder, and S. Hofferberth, Phys. Rev. Lett. 113, 053601 (2014). * Tiarks _et al._ (2014) D. Tiarks, S. Baur, K. Schneider, S. Dürr, and G. Rempe, Phys. Rev. Lett. 113, 053602 (2014). * Tiarks _et al._ (2019) D. Tiarks, S. Schmidt-Eberle, T. Stolz, G. Rempe, and S. Dürr, Nature Physics 15, 124 (2019). * Gorshkov _et al._ (2013) A. V. Gorshkov, R. Nath, and T. Pohl, Phys. Rev. Lett. 110, 153601 (2013). * Murray _et al._ (2018) C. R. Murray, I. Mirgorodskiy, C. Tresp, C. Braun, A. Paris-Mandoki, A. V. Gorshkov, S. Hofferberth, and T. Pohl, Phys. Rev. Lett. 120, 113601 (2018). * Murray _et al._ (2016) C. R. Murray, A. V. Gorshkov, and T. Pohl, New Journal of Physics 18, 092001 (2016). * Otterbach _et al._ (2013) J. Otterbach, M. Moos, D. Muth, and M. Fleischhauer, Phys. Rev. Lett. 111, 113001 (2013). * Zeuthen _et al._ (2017) E. Zeuthen, M. J. Gullans, M. F. Maghrebi, and A. V. Gorshkov, Phys. Rev. Lett. 119, 043602 (2017). * Bienias _et al._ (2020) P. Bienias, J. Douglas, A. Paris-Mandoki, P. Titum, I. Mirgorodskiy, C. Tresp, E. Zeuthen, M. J. Gullans, M. Manzoni, S. Hofferberth, D. Chang, and A. V. Gorshkov, Phys. Rev. Research 2, 033049 (2020). * Facchinetti _et al._ (2016) G. Facchinetti, S. D. Jenkins, and J. Ruostekoski, Phys. Rev. Lett. 117, 243601 (2016). * Perczel _et al._ (2017) J. Perczel, J. Borregaard, D. E. Chang, H. Pichler, S. F. Yelin, P. Zoller, and M. D. Lukin, Phys. Rev. Lett. 119, 023603 (2017). * Bettles _et al._ (2017) R. J. Bettles, J. c. v. Minář, C. S. Adams, I. Lesanovsky, and B. Olmos, Phys. Rev. A 96, 041603 (2017). * Guimond _et al._ (2019) P.-O. Guimond, A. Grankin, D. V. Vasilyev, B. Vermersch, and P. Zoller, Phys. Rev. Lett. 122, 093601 (2019). * Ballantine and Ruostekoski (2020) K. E. Ballantine and J. Ruostekoski, Phys. Rev. Lett. 125, 143604 (2020). * Bettles _et al._ (2016) R. J. Bettles, S. A. Gardiner, and C. S. Adams, Phys. Rev. Lett. 116, 103602 (2016). * Shahmoon _et al._ (2017) E. Shahmoon, D. S. Wild, M. D. Lukin, and S. F. Yelin, Phys. Rev. Lett. 118, 113601 (2017). * Rui _et al._ (2020) J. Rui, D. Wei, A. Rubio-Abadal, S. Hollerith, J. Zeiher, D. M. Stamper-Kurn, C. Gross, and I. Bloch, Nature 583, 369 (2020). * Grankin _et al._ (2018) A. Grankin, P. O. Guimond, D. V. Vasilyev, B. Vermersch, and P. Zoller, Phys. Rev. A 98, 043825 (2018). * Bekenstein _et al._ (2020) R. Bekenstein, I. Pikovski, H. Pichler, E. Shahmoon, S. F. Yelin, and M. D. Lukin, Nature Physics 16, 676 (2020). * Cidrim _et al._ (2020) A. Cidrim, T. S. do Espirito Santo, J. Schachenmayer, R. Kaiser, and R. Bachelard, Phys. Rev. Lett. 125, 073601 (2020). * Williamson _et al._ (2020) L. A. Williamson, M. O. Borgh, and J. Ruostekoski, Phys. Rev. Lett. 125, 073602 (2020). * Moreno-Cardoner _et al._ (2021) M. Moreno-Cardoner, D. Goncalves, and D. E. Chang, “Quantum nonlinear optics based on two-dimensional rydberg atom arrays,” (2021), arXiv:2101.01936 [quant-ph] . * Henkel _et al._ (2010) N. Henkel, R. Nath, and T. Pohl, Phys. Rev. Lett. 104, 195302 (2010). * Zeiher _et al._ (2015) J. Zeiher, P. Schauß, S. Hild, T. Macrì, I. Bloch, and C. Gross, Phys. Rev. X 5, 031015 (2015). * Jaksch _et al._ (2000) D. Jaksch, J. I. Cirac, P. Zoller, S. L. Rolston, R. Côté, and M. D. Lukin, Phys. Rev. Lett. 85, 2208 (2000). * Lukin _et al._ (2001) M. D. Lukin, M. Fleischhauer, R. Cote, L. M. Duan, D. Jaksch, J. I. Cirac, and P. Zoller, Phys. Rev. Lett. 87, 037901 (2001). * Saffman _et al._ (2010) M. Saffman, T. G. Walker, and K. Mølmer, Rev. Mod. Phys. 82, 2313 (2010). * Adams _et al._ (2019) C. S. Adams, J. D. Pritchard, and J. P. Shaffer, Journal of Physics B: Atomic, Molecular and Optical Physics 53, 012002 (2019). * James (1993) D. F. V. James, Phys. Rev. A 47, 1336 (1993). * Dung _et al._ (2002) H. T. Dung, L. Knöll, and D.-G. Welsch, Phys. Rev. A 66, 063810 (2002). * Asenjo-Garcia _et al._ (2017) A. Asenjo-Garcia, J. D. Hood, D. E. Chang, and H. J. Kimble, Phys. Rev. A 95, 033818 (2017). * Zoubi and Ritsch (2011) H. Zoubi and H. Ritsch, Phys. Rev. A 83, 063831 (2011). * Sutherland and Robicheaux (2016) R. T. Sutherland and F. Robicheaux, Phys. Rev. A 94, 013847 (2016). * Bettles _et al._ (2015) R. J. Bettles, S. A. Gardiner, and C. S. Adams, Phys. Rev. A 92, 063822 (2015). * Zhang and Mølmer (2019) Y.-X. Zhang and K. Mølmer, Phys. Rev. Lett. 122, 203605 (2019). * Glicenstein and Ferioli (2020) A. Glicenstein, G. Ferioli, N. Sibalic, L. Brossard, I. Ferrier-Barbut, and A. Browaeys, Phys. Rev. Lett. 124, 253602 (2020). * Piñeiro Orioli and Rey (2019) A. Piñeiro Orioli and A. M. Rey, Phys. Rev. Lett. 123, 223601 (2019). * Fleischhauer and Lukin (2000) M. Fleischhauer and M. D. Lukin, Phys. Rev. Lett. 84, 5094 (2000). * Arimondo (1996) E. Arimondo (Elsevier, 1996) pp. 257 – 354. * Fleischhauer _et al._ (2005) M. Fleischhauer, A. Imamoglu, and J. P. Marangos, Rev. Mod. Phys. 77, 633 (2005). * Manzoni _et al._ (2018) M. T. Manzoni, M. Moreno-Cardoner, A. Asenjo-Garcia, J. V. Porto, A. V. Gorshkov, and D. E. Chang, New Journal of Physics 20, 083048 (2018). * Murray (2021) C. R. Murray, and T. Pohl, Phys. Rev. X 7, 031007 (2017). * Mølmer _et al._ (1993) K. Mølmer, Y. Castin, and J. Dalibard, J. Opt. Soc. Am. B 10, 524 (1993). * Stiesdal _et al._ (2021) N. Stiesdal, H. Busche, K. Kleinbeck, J. Kumlin, M. G. Hansen, H. P. Büchler, and S. Hofferberth, “Controlled multi-photon subtraction with cascaded Rydberg superatoms as single-photon absorbers,” (2021), arXiv:2103.15738 [quant-ph] . * Parkins _et al._ (1993) A. S. Parkins, P. Marte, P. Zoller, and H. J. Kimble, Phys. Rev. Lett. 71, 3095 (1993). * Ralph _et al._ (2015) T. C. Ralph, I. Söllner, S. Mahmoodian, A. G. White, and P. Lohdal, Phys. Rev. Lett. 114, 173603 (1993). * Witthaut _et al._ (2012) D. Witthaut, M. D. Lukin, and A. S. Sørensen, EPL 97, 50007 (2015). * Shahmoon _et al._ (2020) E. Shahmoon, D. S. Wild, M. D. Lukin, and S. F. Yelin, “Theory of cavity qed with 2d atomic arrays,” (2020), arXiv:2006.01972 [quant-ph] . * Demkowicz-Dobrzański _et al._ (2015) X. R. Demkowicz-Dobrzański, M. Jarzyna, and J. Kolodyński, Progress in Optics 60, 345 (2015).
# Learning Abstract Representations through Lossy Compression of Multi-Modal Signals Charles Wilmot, Gianluca Baldassarre, and Jochen Triesch C. Wilmot and J. Triesch are with the Frankfurt Institute for Advanced Studies, Ruth-Moufang- Str. 1, 60438 Frankfurt am Main, Germany. {wilmot,triesch}fias.uni- frankfurt.de G. Baldassarre is with the National Research Council, Institute of Cognitive Sciences and Techologies, Via S. Martino della Battaglia 44, I-00185 Rome, Italy<EMAIL_ADDRESS> Manuscript received September 30, 2020; revised September 31, 2020. ###### Abstract A key competence for open-ended learning is the formation of increasingly abstract representations useful for driving complex behavior. Abstract representations ignore specific details and facilitate generalization. Here we consider the learning of abstract representations in a multi-modal setting with two or more input modalities. We treat the problem as a lossy compression problem and show that generic lossy compression of multimodal sensory input naturally extracts abstract representations that tend to strip away modalitiy specific details and preferentially retain information that is shared across the different modalities. Specifically, we propose an architecture that is able to extract information common to different modalities based on the compression abilities of generic autoencoder neural networks. We test the architecture with two tasks that allow 1) the precise manipulation of the amount of information contained in and shared across different modalities and 2) testing the method on a simulated robot with visual and proprioceptive inputs. Our results show the validity of the proposed approach and demonstrate the applicability to embodied agents. ###### Index Terms: open-ended learning, abstraction, multimodality, lossy compression, autoencoder, intrinsic motivation. ## 1 Introduction Human intelligence rests on the ability to learn abstract representations. An abstract representation has stripped away many details of specific examples of a concept and retains what is common, thereby facilitating generalization and transfer of knowledge to new tasks [25]. A key challenge for natural and artificial developing agents is to learn such abstract representations. How can this be done? ### 1.1 Learning abstract representations In classic supervised learning settings, an external teacher provides the abstract concept by virtue of explicit labeling of training examples. For example, in neural network based image recognition explicit labels (cat, dog, etc.) are provided as a “one-hot” abstract code and the network learns to map input images to this abstract representation [1, 2, 3]. While such a learned mapping qualifies as an abstract representation that has “stripped away many details of specific examples of a concept and retains what is common” it needs to be provided by the teacher through typically millions of examples. This is clearly not how human infants learn and often leads to undesirable generalization behavior [4, 5]. Therefore unsupervised learning and reinforcement learning are more interesting settings for studying the autonomous formation of abstract representations. Reinforcement learning (RL) can also be viewed from the perspective of learning abstract representations. The essence of RL is to learn a policy, i.e., a mapping from states of the world to actions that an agent should take in order to maximize the sum of future rewards [6]. Often, this mapping is realized through neural networks. In deep Q-learning networks [7, 8], for example, a neural network learns to map the current state of the world (e.g., the current image of a computer game screen) onto the expected future rewards for performing different actions (e.g., joystick commands) in this particular state. This can be interpreted as the agent learning an abstract “concept” of the following kind: the set of all world states for which this particular joystick command will be the optimal choice of action. While these “concepts” are not provided explicitly by a teacher, they are provided implicitly through the definition of the RL problem (states, actions, rewards, environment dynamics). In fact, from an RL perspective, these “concepts” are the only ones the agent ever needs to know about. They suffice for behaving optimally in this particular environment. However, they may also become completely useless when the task changes, e.g., a different computer game should be played. This exemplifies the deep and unresolved problem of how abstract knowledge could be extracted in RL that is likely to transfer to new tasks. In the domain of unsupervised learning, on which we will focus in the remainder, a simple approach to learning somewhat abstracted representations is through clustering. For example, in $k$-means clustering [9] an input $x\in\mathbb{R}^{n}$ is represented by mapping it to one of $k$ cluster centers $c_{i}\in\mathbb{R}^{n},\,i\in\\{1,\ldots,k\\}$ based on a suitably defined distance metric in $\mathbb{R}^{n}$. Representing an input $x$ by the closest cluster center strips away information about the precise location of $x$ in $\mathbb{R}^{n}$, achieving a simple form of abstraction. However, the use of a predefined distance metric is limiting. A second set of approaches for learning more abstract representation through lossy compression are dimensionality reduction techniques. A classic example of such an approach is principal component analysis (PCA). PCA finds linear projections of the data such that the projected data has maximum variance, while orthogonal directions are discarded. Thus, the information that gets stripped away corresponds to linear projections of small variance. The central limitation of PCA is the restriction to linear projections. A popular and more powerful approach is the use of autoencoder neural networks [10], on which we will focus in the following. Like other dimensionality reduction techniques, autoencoders construct a more compact and abstract representation of the input domain by learning to map inputs $x\in\mathbb{R}^{n}$ onto a more compact latent representation $z\in\mathbb{R}^{m}$ with $m\ll n$ via a neural network. For this, the network has an encoder/decoder structure with a “bottleneck” in the middle. The $n$-dimensional input is mapped via several layers onto the $m$-dimensional bottleneck and from there via several layers to an $n$-dimensional output. The learning objective is to reconstruct the input at the output, but the trivial solution of a direct identity mapping is avoided, because the input needs to be “squeezed through” the central $m$-dimensional bottleneck. After training, the decoder is often discarded and only the encoder is retained, providing a mapping from the original $n$-dimensional input $x$ to a compressed lower-dimensional latent representation $z$. Due to the nonlinear nature of the neural network, it is difficult to characterize exactly what information will be stripped away in the encoding process. An exception is the special case of a linear network with a quadratic loss function. For this case, it can be shown that the network discovers the same subspace as linear PCA. In the following, we will consider the implications of a developing agent learning to encode an input $x$ that comprises multiple sensory modalities. ### 1.2 Multimodality, Abstraction, and Lossy Compression: An Information Theoretic Perspective Our different sensory modalities (vision, audition, proprioception, touch, smell, taste) provide us with different “views” of our physical environment. These views are not independent, but contain shared information about the underlying physical reality. Therefore, it must be possible to compress the information coming from the different modalities into a more compact code. As an example, consider viewing and touching your favorite coffee cup. Some information such as the color or the text or picture printed on the cup will only be accessible to the visual modality. Some information, such as the temperature or roughness of the surface, will only be accessible to the haptic modality. Some information, however, such as the 3-D shape of the cup, will be accessible to both modalities. This implies potential for compression. Let $X_{v}$ represent the visual input and $X_{h}$ the haptic input. We can quantify the amount of information using concepts from information theory. For now, we will ignore the fact that the amount of information is a function of our behavior, but we will return to this point in the Discussion. We can quantify the amount of information in $X_{v}$ and $X_{h}$ as: $H(X_{v},X_{h})=H(X_{v})+H(X_{h})-MI(X_{v};X_{h})\;,$ (1) where $H(X_{v}$) and $H(X_{h})$ are the individual entropies of the visual and haptic signals, respectively, $H(X_{v},X_{h})$ is their joint entropy, and $MI(X_{v};X_{h})$ is their mutual information, i.e., the amount of information that $X_{v}$ and $X_{h}$ have in common. The individual entropies $H(X_{v}$) and $H(X_{h})$ and the joint entropy $H(X_{v},X_{h})$ indicate, respectively, how many bits are required on average to encode $X_{v}$ and $X_{h}$ individually or jointly. The mutual information expresses how many bits can be “saved” by jointly encoding $X_{v}$ and $X_{h}$ compared to encoding them separately. If the visual and haptic inputs were statistically independent, then $MI(X_{v};X_{h})=0$ and no savings can be gained. If there are any statistical dependencies between $X_{v}$ and $X_{h}$, then $MI(X_{v};X_{h})>0$ and $H(X_{v},X_{h})<H(X_{v})+H(X_{h})$, i.e., the visual and haptic signals can be compressed into a more compact code. In principle, this compression can be achieved without any loss of information. However, it is the very nature of abstraction to “strip away” information about the details of a situation and only maintain a “coarser” and hopefully more generalizable representation. Consider again the situation of visuo- haptic perception of your favorite coffee cup. If we were to strip away the information that is only accessible to the visual modality (color and printed text/picture) and strip away the information that is accessible to only the haptic modality (temperature, surface roughness), then we are left with a much reduced and abstract representation that maintains essentially the 3-D shape of the cup and allows for many generalizations, e.g., how to grasp this particular cup versus many similarly shaped cups with virtually endless combinations of color, picture, text, surface roughness, and temperature. Thus, learning an abstract code by retaining information that is shared across modalities and stripping away information that is specific to only individual modalities may lead to very useful representations with high potential for generalization. How can this be done? Here we investigate a possible solution that relies on autoencoding the multimodal inputs into a sufficiently compact lossy code. The rationale for this approach is as follows. Consider an autoencoder that learns to map the concatenation of $X_{v}$ and $X_{h}$ onto a small latent vector $Z$ such that $X_{v}$ and $X_{h}$ can be reconstructed from $Z$ with minimal loss. What information should $Z$ encode? If the information coding capacity of $Z$ is at least $H(X_{v},X_{h})$, then it can simply encode all the information contained in the visual and haptic signals. If it is smaller, however, then some information must be discarded. So what information should be kept and what information should be discarded? In general, it appears best to keep the information that $X_{v}$ and $X_{h}$ have in common, because this information aids in reconstructing both of them, essentially killing two (metaphorical) birds with one stone. Information that is present in only either $X_{v}$ or $X_{h}$ cannot be helpful in reconstructing the other. Therefore, an autoencoder with limited capacity will learn a representation that tends to prioritize the information that $X_{v}$ and $X_{h}$ have in common and tends to strip away the information that is unique to either modality. This is exactly the kind of abstract representation that we would like to achieve. In the remainder of this article we make these ideas more concrete and study them in extensive computer simulations. We begin by a discussion of related work. We then propose a number of models to learn abstract representations from multimodal input via autoencoding and compare their behavior. In a first approach, we use synthetically generated inputs, since this allows us to precisely control the amount of information in individual sensory modalities and the amount of information that they share. In a second approach, we use visual and proprioceptive data from a robot simulation. We end by discussing broader implications of learning abstract representations through lossy compression of multimodal input for cognitive development. ## 2 Related Work Our work falls in the general area of unsupervised representation learning [24]. Representation learning aims to build machine learning algorithms that extract the explanatory variation factors of data. In so doing, the learned representations are also often compressed/abstracted in that they discard some information. Algorithms for representation learning can be grouped based on the general strategy they use to form representations. This strategy plays the role of the choice of generic priors on the process that is assumed to have generated the data. These priors might for example assume that there are indeed distinct factors capturing the variability of data (e.g., changing positions of objects and light sources relative to a camera giving rise to varying camera images), that observations close in space or time have similar values (e.g., as in natural images and videos), or that most of the probability mass of the data concentrates on manifolds with a dimensionality smaller than that of the perceptual space (e.g., as assumed in autoencoders). Information theoretic approaches to learning representations are of particular interest and have a long history. A complete review is beyond the scope of this article. Here we focus on relating our work to some classic approaches. In Efficient Coding [27, 23], the goal is to learn a representation for sensory signals that is “compact” and exploits redundancies in the signals to arrive at a more efficient code. Often this is formulated as maximizing the mutual information $I(X;Y)$ between an input $X$ and its representation $Y$ while putting additional constraints on $Y$. For example, in linear and non- linear independent component analysis (ICA) [28], one attempts to make the components of $Y$ statistically independent, corresponding to the prior of the data being generated by mixing independent information sources. Searching for codes that are factorial while retaining the maximum amount of information about the input leads to an efficient code. This is because if the extracted components $Y$ were not independent, then there would be potential for further compression of $Y$ to generate an even more efficient code. Sparse coding is a popular variant of efficient coding that is related to ICA and imposes a sparsity constraint (or prior) on $Y$ [29]. This has given rise to a large literature on learning sparse representations for sensory signals. Sparse coding models are frequently employed as models of learning sensory representations in the brain. More recently, such approaches have also been extended to active perception. In the Active Efficient Coding (AEC) framework, the sensory encoding is optimized by simultaneously optimizing the encoding of the sensory inputs and the movements of the sense organs that shape the statistics of these inputs [30, 31]. Predictive coding can be viewed as a special case of efficient coding that forms hierarchical representations, where higher levels predict the activity of lower levels and lower levels signal prediction errors to higher levels [32]. Note that in these approaches the goal is generally to retain as much information about the input as possible There is no notion of abstraction as we have defined it, i.e., of deliberately discarding information to arrive at a representation that generalizes more easily to new situations. An alternative information theoretic learning objective was proposed by Becker and Hinton [33]. In their IMAX approach, two distinct inputs are considered (in particular visual inputs from two neighboring locations of stereoscopic image pairs) and the objective is to extract information that is shared by these two input sources. Becker and Hinton demonstrate that this allows learning to extract disparity information from the stereoscopic images. While the specific texture projected to the two neighboring retinal locations may be different, the binocular disparity is typically the same, because it tends to vary smoothly across the image. Their approach can be viewed as an early attempt to extract more “abstract” information (disparity) by encoding information from multiple sources (binocular visual inputs from two neighboring locations) and only keeping the information that they have in common. IMAX can also be viewed as related to Canonical Correlation Analysis (CCA) [34]. In CCA one tries to find linear combinations of the components of a random vector $X$ and the components of a random vector $Y$ such that these linear combinations are maximally correlated. A high correlation between the two projections implies that they share a substantial amount of information (correlation implies statistical dependence and therefore non-zero mutual information while the reverse is generally not true). The information bottleneck method by Tishby and colleagues considers the objective of encoding a sensory input $X$ and only keeping the information that is useful for predicting a second “relevant” signal $Y$ [35]. Thus it also aims at keeping only information that is shared with (or: predictive of) another signal, while discarding all other information. However, in contrast to IMAX there is a clear asymmetry between the input $X$ and the signal $Y$. The information bottleneck objective can be expressed as follows: Let $T$ be a compressed version of $X$. $T$ should retain as little information from $X$ as possible, but at the same time keep as much information about $Y$ as possible. Thus $T$ functions as the information bottleneck. Formally, one is seeking to optimize the functional: $\min_{p(T=t|X=x)}I(X;T)-\beta I(T;Y),$ (2) where $p(T=t|X=x)$ describes the encoding of an input $x$ via its representation $t$, $I(X;T)$ is the mutual information between $X$ and $T$, $I(T;Y)$ is the mutual information between $T$ and $Y$, and $\beta$ is a parameter for balancing the two objectives. Unfortunately, as shown in the Appendix, it is not straightforward to derive a symmetric variant of the information bottleneck, i.e., to find an encoding $p(T=t|X=x,Y=y)$ that would keep only the information that $X$ and $Y$ share. Instead, in our approach we utilize the generic ability of (deep) autoencoders to perform lossy compression of sensory inputs. Autoencoders implicitly assume that much of the variability of the data can be accounted for by a smaller number of latent causes. Specifically, in our case we assume that inputs from different sensory modalities are related since they are consequences of the same underlying physical causes. For example, a human infant (or robot) hitting their hand on the table will observe the consequences in multiple modalities (feeling the contact, seeing the movement, hearing the sound). The choice of the size of the autoencoder’s bottleneck is analogous to the choice of $\beta$ in the information bottleneck. Note that many different types of autoencoders exist including denoising autoencoders [38], sparse autoencoders [36], contractive autoencoders [37], variational autoencoders [39], adversarial autoencoders [40], etc. In order to not distract from our main point about lossy compression of multimodal signals, we restrict our experiments to “generic” autoencoders. It should be understood that the results are expected to generalize to other types of autoencoders or, indeed, other lossy compression schemes. The novelty of the proposed solution resides in the fact that it exploits the multimodal organisation of the information available to an agent to form abstract representations. Furthermore, we propose a specific cross-modality prediction architecture to distill only the information that is shared across multiple modalities. Figure 1: Overview of the approaches, assuming only $2$ modalities ($n=2$). A baseline experiment, jointly encoding (JE) the dependent vectors $y_{i}$. B control experiment, jointly encoding the dependent vectors $y_{i}$ but reconstructing the original data $x_{i}$. C cross-modality prediction experiment (CM), jointly encoding the predicted vectors $\tilde{y}_{\smallsetminus i}$. In each schema, the red areas represent the random neural networks generating dependent vectors (section 3.1.1), the yellow areas represent the encoding and decoding networks (section 3.1.2), the green areas represent the readout networks (section 3.1.3), and the orange area represents the cross-modality prediction networks (section 3.1.4). ## 3 Methods Our general approach comprises three processing steps. The first step consists in generating multimodal sensory data. In order to have more control over the amount of mutual information among the different sensory modalities, we present a way to generate the latter from noise. We call this type of data synthetic as it has no real significance. For this we use random multi-layer neural networks that map independent information sources $x$ onto different “views” $y$ seen by different sensory modalities. This models the process how multimodal sensory information is generated from unobserved causes in the world. Our approach applied to the synthetic data will be explained in sub- section 3.1. We also test our approach on multimodal data from an actual robot simulation. This setup will be introduced in sub-section 3.2. In the second step, we train a neural network autoencoder with varying capacity, i.e., size of the central bottleneck, to learn a compressed (lossy) representation $z$ from the concatenation of the multimodal inputs $y$. This models the process of a developing agent learning an abstract representation from multimodal sensory information. In the third step we analyze the learned representation $z$ and measure how much information it retains that is unique to individual modalities versus shared among multiple modalities. For this we train a third set of neural networks to reconstruct the original information sources $x$ from the latent code $z$. The reconstruction error is a proxy for how much information has been lost during the encoding process. We now explain the three processing steps in detail starting with the synthetic setup. ### 3.1 Experiments with Synthetic Multimodal Input #### 3.1.1 Step 1: Generating Synthetic Multimodal Input We produce the multimodal sensory data in such a way that we can precisely control the amount of mutual information between the different sensory modalities. We first define a distribution $p_{m}$ from which we sample information that is shared by all sensory modalities. $p_{m}$ therefore represents independent information sources in the world that affect multiple sensory modalities. We define it as a $d_{m}$-dimensional distribution of independent standard normal distributions. We then define a second distribution $p_{e}$, from which the information exclusive to each modality is sampled. We define $p_{e}$ as a $d_{e}$-dimensional distribution of independent standard normal Gaussians. For a shared vector $x_{m}\sim p_{m}$ and $n$ vectors $x_{e,i}\sim p_{e}$, $i\in\mathbb{N}_{n-1}$ we can create the vectors $x_{i}=x_{e,i}\oplus x_{m}$ carrying the information of each modality, where $\oplus$ is the concatenation operation. Our sensory modalities do not sample the underlying causes directly (e.g., objects and light sources), but indirectly (e.g., images provided by the eyes). To mimic such an indirect sampling of the world without making any strong assumptions, we generate the sensory inputs $y$ that the learning agent perceives via random neural networks. Specifically, we define the input to modality $i$ as: $y_{i}=\frac{C\left(x_{i},\theta_{C,i}\right)-\mu_{i}}{\sigma_{i}}\;,$ (3) where $C$ and $\theta_{C,i}$ are the input construction network and its weights for the modality $i$ and $\mu_{i}$ and $\sigma_{i}$ are constants calculated to normalize the components of $y_{i}$ to zero mean and unit variance. Tuning the amount of mutual information between the vectors $y_{i}$ is done by changing the dimensionalities $d_{m}$ and $d_{e}$ of the vectors $x_{m}$ and $x_{e,i}$, respectively. The amount of information preserved from the vectors $x_{i}$ in the vectors $y_{i}$ depends on the dimension $d_{y}$ of the vectors $y_{i}$. We define $d_{y}$ to be proportional to the dimension $d_{x}=d_{m}+d_{e}$: $d_{y}=k\times d_{x}\;,$ (4) where $k\gg 1$. This ensures that the sensory inputs $y_{i}$ essentially retain all information from the sources $x_{m}$ and $x_{e,i}$. #### 3.1.2 Step 2: Learning an Abstract Representation of the Synthetic Multimodal Input via Autoencoding Taken together, the set of vectors $\\{y_{i}\\}_{i\in\mathbb{N}}$ carries once the information from each $x_{e,i}$ and $n$ times the information from the mutual vector $x_{m}$. To show that a lossy-compression algorithm achieves a better encoding when favoring the reconstruction of the repeated information, we train an autoencoder to jointly encode the set of the $y_{i}$. We therefore construct the concatenation $y=y_{0}\oplus\dots\oplus y_{n-1}$ to train the autoencoder: $\displaystyle z$ $\displaystyle=E\left(y,\theta_{E}\right)$ (5) $\displaystyle\tilde{y}$ $\displaystyle=D\left(z,\theta_{D}\right)\;,$ (6) where $E$ and $\theta_{E}$ are the encoding network and its weights and $D$ and $\theta_{D}$ are the decoding network and its weights. Tuning the dimension $d_{z}$ of the latent representation $z$ enables us to control the amount of information lost in the encoding process. The training loss for the weights $\theta_{E}$ and $\theta_{D}$ is the mean squared error between the data and its reconstruction, averaged over the component dimension and summed over the batch dimension: $\displaystyle L_{E,D}$ $\displaystyle=\sum_{\text{batch}}\frac{1}{nd_{y}}\left(y-\tilde{y}\right)^{2}\;.$ (7) #### 3.1.3 Step 3: Quantifying Independent and Shared Information in the Learned Latent Representation Finally, in order to measure what information is preserved in the encoding $z$, we train readout neural networks to reconstruct the original data $x_{m}$ and the vectors $x_{e,i}$: $\displaystyle\tilde{x}_{m}$ $\displaystyle=R_{m}\left(z,\theta_{m}\right)$ (8) $\displaystyle\tilde{x}_{e,i}$ $\displaystyle=R_{e}\left(z,\theta_{R,i}\right)\text{,}$ (9) where $R_{m}$ and $\theta_{m}$ are the mutual information readout network and its weights, and the $R_{e}$ and $\theta_{R,i}$ are the exclusive information readout networks and their weights. The losses for training the readout operations are the mean squared errors between the readout and the original data summed over the batch dimension and averaged over the component dimension: $\displaystyle L_{m}$ $\displaystyle=\frac{1}{d_{m}}\sum_{\text{batch}}\left(\tilde{x}_{m}-x_{m}\right)^{2}\;\text{and}$ (10) $\displaystyle L_{e,i}$ $\displaystyle=\frac{1}{d_{e}}\sum_{\text{batch}}\left(\tilde{x}_{e,i}-x_{e,i}\right)^{2}\;\text{.}$ (11) Finally, once the readout networks trained, we measure the average per data- point mean squared errors $\displaystyle r_{m}$ $\displaystyle=\frac{1}{d_{m}}\mathbb{E}\left[\left(\tilde{x}_{m}-x_{m}\right)^{2}\right]\;\text{and}$ (12) $\displaystyle r_{e}$ $\displaystyle=\frac{1}{d_{e}}\mathbb{E}\left[\left(\tilde{x}_{e,i}-x_{e,i}\right)^{2}\right]\;\text{,}$ (13) serving as a measure of the portion of the mutual and exclusive data retained in the encoding $z$. As a control condition, we also study a second encoding mechanism, where, instead of reconstructing the dependent vectors $y_{0}\oplus\dots\oplus y_{n-1}=y$, the decoder part reconstructs the original data $x_{0}\oplus\dots\oplus x_{n-1}=x$. The loss for training the encoder and decoder networks $E$ and $D$ from equation 7 then becomes: $L_{E,D}=\sum_{\text{batch}}\frac{1}{nd_{x}}\left(x-\tilde{x}\right)^{2}\;\text{.}$ (14) #### 3.1.4 An Alternative to Step 2: Isolating the Shared Information We also compare the previous approach to an alternative architecture, designed specifically to isolate the information shared between the modalities. Let $\left(A,B\right)$ be a pair of random variables defined over the space $\mathcal{A}\times\mathcal{B}$ with unknown joint distribution $P\left(A,B\right)$ and marginal distributions $P\left(A\right)$ and $P\left(B\right)$. Determining the mutual information between the variables $A$ and $B$ consists in finding either one of $P\left(A,B\right)$, $P\left(A|B\right)$ or $P\left(B|A\right)$. With no other assumptions, this process requires to sample many times from the joint distribution $P\left(A,B\right)$. We propose to make the strong assumption that the conditional probabilities are standard normal distributions with a fixed standard deviation $\sigma$ $\displaystyle P\left(B=b|A=a\right)=\mathcal{N}\left(\mu\left(a\right),\sigma,b\right)$ (15) We can then try to approximate the function $\mu\left(a\right)$ with a neural network $M\left(a,\theta_{M}\right)$ maximizing the probability $P\left(B=b|A=a\right)$. $\mu\left(a\right)$ thus represents the most likely $b$ associated with $a$, under the standard normal assumption. Training the network is done by minimizing the mean squared error loss $\displaystyle L_{M}$ $\displaystyle=-\mathbb{E}_{a,b\sim P\left(A,B\right)}\left[\log\left(\mathcal{N}\left(\mu,\sigma,b\right)\right)\right]$ (16) $\displaystyle=\mathbb{E}_{a,b\sim P\left(A,B\right)}\left[\left(\mu-b\right)^{2}\right]\cdot K_{1}+K_{2}$ (17) with $K_{1}$ and $K_{2}$ constants depending on $\sigma$. More concretely and using the notation from the first architecture, we define for each modality $i$ a neural network $M\left(y_{i},\theta_{M_{i}}\right)$ learning to predict all other modality vectors $y_{j},j\neq i$. The loss for the weights $\theta_{M_{i}}$ is defined $L_{M_{i}}=\sum_{\text{batch}}\frac{1}{\left(n-1\right)d_{y}}\left(y_{\smallsetminus i}-\tilde{y}_{\smallsetminus i}\right)^{2}$ (18) with $y_{\smallsetminus i}=\bigoplus_{j\neq i}y_{j}$ (19) the concatenation of all vectors $y_{j}$ for $j\neq i$ and $\tilde{y}_{\smallsetminus i}$ the output of the network. We then consider the concatenation of the $\tilde{y}_{\smallsetminus i}$ for all $i$ as a high- dimensional code of the shared information. This code is then compressed using an autoencoder, similarly to the description in Section 3.1.2. We vary the dimension of the encoder’s latent code. Finally, similarly to the first approach, we train readout networks from the compressed latent code to determine how mutual and exclusive information are preserved in the process. Overall, this way of processing the data is analogous to the baseline experiment in that the cross modality prediction networks and the subsequent auto-encoder, when considered together, form a network that transforms the vectors $y_{i}$ into themselves. Together, these two components can thus be considered as an auto-encoder, subject to a cross-modality prediction constraint. #### 3.1.5 Neural Network Training In the following, we compare three architectures against each other (compare Fig. 1): * • The baseline architecture (JE), simply auto-encoding the vectors $y_{i}$ jointly (cf. Fig. 1A). * • The control condition with a simpler encoding task, where the vectors $y_{i}$ are encoded into a latent code $z$, from which the decoder tries to reconstruct the original vectors $x_{i}$, from which the inputs $y_{i}$ were generated (cf. Fig. 1B). * • The alternative architecture (CM), where for each modality, a neural network tries to predict all other modalities and then all resulting predictions are jointly encoded, similarly to the baseline architecture (cf. Fig. 1C). We will now describe the training procedure and implementation details. In order to show the nature of the information preferably preserved by the encoding process, we measure the quality of the readouts obtained as we vary the dimension of the latent vector $d_{z}$. To this end, for each dimension $d_{z}\in\left[1;d_{z,max}\right]$, we successively train the cross modality prediction networks (experiment C only), the autoencoder weights $\theta_{E}$ and $\theta_{D}$ and the readout weights $\theta_{m}$ and $\theta_{R,i}$. Once training is completed, we measure the average mean squared error of the readouts $\tilde{x}_{m}$ and $\tilde{x}_{e,i}$. We choose the distributions of the vectors $x_{m}$ and $x_{e,i}$ to be multivariate standard normal with a zero mean and unit variance. Therefore, a random guess would score an average mean squared error of $1$. Each experiment is repeated $3$ times and results are averaged. The neural networks for the input construction $C$, cross-modality prediction $M$, encoding $E$, decoding $D$, mutual readout $R_{m}$, and exclusive readout $R_{e}$ all have three fully-connected layers. The two first layers always have a fixed dimension of $200$ and use a ReLU as non-linearity. The final layer is always linear, its dimension for each network is reported in Table I. Network | $M$ | $C$ | $E$ | $D$ | $R_{m}$ | $R_{e}$ ---|---|---|---|---|---|--- Dimension | $\left(n-1\right)\times d_{y}$ | $d_{y}$ | $d_{z}$ | $n\times d_{y}$ | $d_{m}$ | $d_{x}$ TABLE I: Dimension of the last layer of each neural network For each model architecture A, B, or C, we show the effect of varying the ratio between mutual and exclusive data and that of varying the number of modalities. The default experiment used $d_{e}=4$, $d_{m}=4$, $n=2$, $k=10$. We then varied $d_{e}\in\\{4,10,16,22\\}$ or $n\in\\{2,3,4,5\\}$, keeping all other parameters fixed. Each network is trained on $2500$ batches of data of size $128$ with a learning rate of $10^{-3}$ and using the Adam algorithm [11]. ### 3.2 Experiments with Multimodal Input from a Robot Simulation #### 3.2.1 Step 1: Generating Multimodal Input from the Robot Simulation Figure 2: High resolution image of the $2$ robot arms in the simulated environment. The images in the dataset have a resolution of $32$ by $64$ pixels only. In order to validate our approach in a more realistic setting, we applied it to data generated from a robot simulator, in which we placed two $7$-degrees- of-freedom robot arms side by side (see Fig. 2). We then generated a dataset comprised of pictures of the $2$ arms, representing the visual modality, and of the joint positions and speeds, representing the proprioceptive modality. It has a total size of $2.000.000$ samples. To generate the dataset, we sampled random target joints positions uniformly in the joints’ motion range, and we simulated $10$ iterations of $0.2$ seconds to let the agent reach the random target using its position PID controllers. At each iteration, a snapshots of the vision sensors and the joint sensors is recorded. Note how the position information is present both in the visual and proprioceptive modalities. However, as each data-sample is composed of a single image, the velocity information is present only in the proprioceptive modality. So as to also have at our disposal an information stream that is uniquely present in the visual modality, we decided to provide the encoding networks with only the proprioceptive information from one of the two arms (the right arm). Thus, the position information of the other arm (left arm) is only available through the visual information. Furthermore, by doing so, the velocity information about the left arm is present in neither of the two modalities and thus serves as a control factor. Finally, the dataset also contains records of the end effectors’ positions of the two arms. We consider the end effector position of the right arm as being implicitly part of the proprioceptive modality, as it can be accurately deduced from the position information, while not directly feeding it into the networks. A summary of the information available to each modality is provided in Fig. 3. In Sec. 3.1.1, we named the vectors representing the different modalities $y_{i}$. When dealing with the realistic data, we will use $y_{0}=y_{v}$ for the visual modality and $y_{1}=y_{p}$ for the proprioceptive one. The $y_{p}$ is $z$-scored, i.e., it has a $0$ mean and a standard deviation of $1$. The $y_{v}$ vector is normalized such that the pixel values are in $\left[-1,1\right]$. We will now redefine the steps $2$, $3$ and the alternative to step $2$ for this dataset. The main difference with the synthetic dataset lies in the fact that the visual information is processed with convolutional neural networks. Moreover, we propose to compare $2$ ways of jointly encoding the modalities, which we refer to as options. The default option is analogous to the way the synthetic data is encoded, with the difference that the visual information is processed by a convolutional neural network. The second option named Alternative Encoding Scheme (AES) consists in learning a latent representation of the visual information with a convolutional autoencoder prior to jointly encoding the latent visual code with the proprioceptive information. Figure 3: Schema representing the information available to each modalities for the realistic data dataset. $\varphi$ and ${\dot{\varphi}}$ denote the positions and velocities of the joints, respectively, and $X_{L}$ and $X_{R}$ the left / right parts of the visual information. Note that the positions and velocities of the left arm are not part of the proprioceptive modality. This way, the information about the position of the left arm is available only through the vision sensor. It also results that the velocity information of the left arm is present in neither of the two modalities, and thus serves as a control factor in our experiments. #### 3.2.2 Step 2: Learning an Abstract Representation of the Robot Data Similar to Sec. 3.1.2, the $y_{v}$ and $y_{p}$ vectors are jointly encoded and decoded with an autoencoder $\left(E,\theta_{E},D,\theta_{D}\right)$. This time, however, the encoding and decoding steps are divided into two parts: $\displaystyle z_{\text{pre}}$ $\displaystyle=E_{v}\left(y_{v},\theta_{E_{v}}\right)\oplus E_{p}\left(y_{p},\theta_{E_{p}}\right)$ (20) $\displaystyle z$ $\displaystyle=E_{\text{pre}}\left(z_{\text{pre}},\theta_{E_{\text{pre}}}\right)$ (21) for the encoder and $\displaystyle z_{\text{post}}$ $\displaystyle=D_{\text{post}}\left(z,\theta_{D_{\text{post}}}\right)$ (22) $\displaystyle\tilde{y}_{v}$ $\displaystyle=D_{v}\left(z_{{\text{post}},v},\theta_{D_{v}}\right)$ (23) $\displaystyle\tilde{y}_{p}$ $\displaystyle=D_{p}\left(z_{{\text{post}},p},\theta_{D_{p}}\right)$ (24) for the decoder, where $z_{{\text{post}},v}$ and $z_{{\text{post}},p}$ form a partition of $z_{\text{post}}$. The index within $z_{\text{post}}$ at which the split occurs is a hyper-parameter. The loss is then defined as: $\displaystyle L_{E,D}=\frac{1}{2d_{p}}\sum_{\text{batch}}\left(\tilde{y}_{p}-y_{p}\right)^{2}+\frac{1}{2d_{v}}\sum_{\text{batch}}\left(\tilde{y}_{v}-y_{v}\right)^{2}\;,$ (25) where $d_{p}$ and $d_{v}$ are the sizes of the proprioception and vision tensors, respectively. Dividing the encoding and decoding process in two parts enables to use convolutional and deconvolutional networks $E_{v}$ and $D_{v}$ to encode and decode the visual information. We did not find any difference when pre- encoding the proprioceptive information with a MLP compared to directly feeding it to the $E_{\text{pre}}$ network, we will therefore report the results for $E_{p}=\text{id}$ and $D_{p}=\text{id}$. In the case of the Alternative Encoding Scheme (AES) option, the $y_{v}$ is a compressed representation of the visual information and there is no need to process it with a convolutional neural network. In that case, we also set $E_{v}=\text{id}$ and $D_{v}=\text{id}$, meaning that the architecture of the networks in that case is the same as for the synthetic dataset. #### 3.2.3 Step 3: Deciphering the Latent Code Similarly to Sec. 3.1.3, we train readout neural networks to decipher the information contained in the encoding $z$. This time, however, since we do not have access to the original vectors $x$ which induced the vectors $y_{v}$ and $y_{p}$, the readout operation aims at reconstructing the proprioceptive information from both arms $y_{\text{target}}=y_{{\text{pos}},l}\oplus y_{{\text{vel}},l}\oplus y_{{\text{ee}},l}\oplus y_{{\text{pos}},r}\oplus y_{{\text{vel}},r}\oplus y_{{\text{ee}},r}$. The readout operation is written as: $\displaystyle y_{\text{readout}}=R\left(z,\theta_{\text{readout}}\right)$ (26) and its loss is: $\displaystyle L_{\text{readout}}=\frac{1}{d_{\text{target}}}\sum_{\text{batch}}\left(y_{\text{readout}}-y_{\text{target}}\right)^{2}\;.$ (27) #### 3.2.4 An Alternative to Step 2 for the Robot Data: Isolating the Shared Information Finally, similarly to Sec. 3.1.4, we propose an alternative to step number $2$ aiming at isolating only the information shared by both modalities. This is done by training two cross-modality prediction networks: $\displaystyle\tilde{y}_{\smallsetminus p}$ $\displaystyle=M_{v}\left(y_{p},\theta_{M_{v}}\right)\quad\text{and}$ (28) $\displaystyle\tilde{y}_{\smallsetminus v}$ $\displaystyle=M_{p}\left(y_{v},\theta_{M_{p}}\right)$ (29) with the losses $\displaystyle L_{M_{v}}$ $\displaystyle=\sum_{\text{batch}}\frac{1}{d_{v}}\left(\tilde{y}_{\smallsetminus p}-y_{v}\right)^{2}\quad\text{and}$ (30) $\displaystyle L_{M_{p}}$ $\displaystyle=\sum_{\text{batch}}\frac{1}{d_{p}}\left(\tilde{y}_{\smallsetminus v}-y_{p}\right)^{2}\;.$ (31) Again, like in Sec. 3.2.2, the representations $\tilde{y}_{\smallsetminus p}$ and $\tilde{y}_{\smallsetminus v}$ are encoded using the autoencoder networks $E_{v}$, $E_{p}$, $E_{\text{pre}}$, $D_{\text{post}}$, $D_{v}$ and $D_{p}$: $\displaystyle z_{\text{pre}}$ $\displaystyle=E_{v}\left(\tilde{y}_{\smallsetminus p},\theta_{E_{v}}\right)\oplus E_{p}\left(\tilde{y}_{\smallsetminus v},\theta_{E_{p}}\right)$ (32) $\displaystyle z$ $\displaystyle=E_{\text{pre}}\left(z_{\text{pre}},\theta_{E_{\text{pre}}}\right)$ (33) $\displaystyle z_{\text{post}}$ $\displaystyle=D_{\text{post}}\left(z,\theta_{D_{\text{post}}}\right)$ (34) $\displaystyle\tilde{y}_{v}$ $\displaystyle=D_{v}\left(z_{{\text{post}},v},\theta_{D_{v}}\right)$ (35) $\displaystyle\tilde{y}_{p}$ $\displaystyle=D_{p}\left(z_{{\text{post}},p},\theta_{D_{p}}\right)\;.$ (36) For the AES option, since $y_{v}$ is a one-dimensional vector, we set $E_{v}=D_{v}=E_{p}=D_{p}=\text{id}$. #### 3.2.5 Description of the Networks In the case of the default option, the network $E_{v}$ is a convolutional neural network composed of $2$ convolutional layers with kernel size $4$ and stride $2$ followed by a dense layer with output size $100$. The network $E_{\text{pre}}$ is a $3$-layered MLP where all layer sizes but the last are $200$. The last layer uses a linear activation function and has a size $d_{z}$. The network $D_{\text{post}}$ is a $3$-layered MLP where all layer sizes but the last are $200$. The last layer uses a linear activation function and has a size $100+d_{p}$. The network $D_{v}$ is a deconvolutional neural network composed of a dense layer of size $8192$ followed by two transposed convolutional layers with kernel sizes $4$ and strides $2$. Finally, the readout network is also a $3$-layered MLP where all layer sizes but the last are $200$. The last layer uses a linear activation function and has a size $d_{\text{target}}$. For the cross-modality prediction, the network $M_{v}$ is a deconvolutional neural network composed of a dense layer of size $8192$ followed by two transposed convolutional layers with kernel size $4$ and stride $2$ and the network $M_{p}$ is a convolutional neural network composed of two convolutional layers with kernel size $4$ and stride $2$ followed by a dense layer of size $d_{p}$. Finally, as stated above, in the case of the AES option, $y_{v}$ is a learned code of size $100$ representing the visual information. In this case we set $E_{v}=D_{v}=E_{p}=D_{p}=\text{id}$. The network learning the code is a convolutional autoencoder composed of $2$ convolutions, one dense layer of size $100$, one dense layer of size $8192$ and $2$ transposed convolutions. ## 4 Results ### 4.1 Lossy Compression of Multimodal Input Preferentially Encodes Information Shared Across Modalities Figure 4: Each plot represents the reconstruction error of the readout operation for the exclusive data $r_{e}$ in blue, and for the shared data $r_{m}$ in red, as a function of the auto-encoder latent dimension. The dotted vertical line indicates the latent dimension matching $nd_{e}+d_{m}$. The data point for a latent dimension of $0$ is theoretically inferred to be equal to $1.0$ (random guess). The four plots in one row correspond to different dimensions $d_{e}$ of the exclusive data. The results are presented for the three architectures A, B and C. Figure 4 shows the reconstruction errors for exclusive vs. shared information as a function of $d_{z}$, the size of the autencoder’s bottleneck, for the three different architectures. Each data point represents the mean of $3$ repetitions of the experiments, and the shaded region around the curves indicate the standard deviation. The grey dotted vertical line indicates the latent code dimension $d_{z}$ matching the number of different univariate gaussian distributions used for generating the correlated vectors $y_{i}$, $d_{min}=d_{m}+nd_{e}$. Assuming that each dimension in the latent vector can encode the information in one normally distributed data source, when $d_{z}=d_{min}$ both the exclusive data and the shared data can theoretically be encoded with minimal information loss. Knowing that random guesses would score a reconstruction error of $1.0$, we can augment the data with the theoretical values $r_{m}=1$ and $r_{e}=1$ for $d_{z}=0$. The results for the JE architecture (cf. Fig. 1A), jointly encoding the correlated vectors $y_{i}$, show that the data shared by all modalities is consistently better reconstructed by the autoencoder for all latent code sizes $d_{z}$. In particular, this is also true for over-complete codes when $d_{z}>d_{min}$. Information loss in that regime is due to imperfections of the function approximator. When the code dimension is bellow $d_{min}$, the information loss is greater, as not all of the data can pass through the antoencoder’s bottleneck. These results confirm our intuition from the Introduction that shared information should be preferentially encoded during lossy compression of multimodal inputs. This information filtering is a consequence of neural networks’ continuity, implying topological properties on the functions that they can learn. Indeed, while there exist non-continuous functions for which the dimensionality of the codomain is greater than that of the domain, the continuity property enforces that the dimension of the codomain is less or equal to that of the domain. As a consequence, the dimensionality of the codomain of the decoder network of an autoencoder is less than or equal to the dimensionality of the latent code. If the dimension of the latent code is itself lower than that of the data, as can be enforced by a bottleneck, it follows that the data and its reconstruction sit on manifolds of different dimensionality, implying information loss. In the under-complete regime, $d_{z}<d_{min}$, the autoencoder shows a stronger preference for retaining the shared data, partly filtering out the exclusive data. The chief reason for this is that the shared data is essentially counted $n$ times in the network’s reconstruction loss, while the exclusive data is counted only once. As the dimension of the exclusive data $d_{e}$ increases, we still observe the two training regimes for $d_{z}$ less or greater than $d_{min}$, even though the boundary between both tends to vanish as we reach the network’s capacity. The results for the second (control) architecture (cf. Fig. 1B), jointly encoding the correlated vectors $y_{i}$ by reconstructing the original data vectors $x_{i}$ rather than $y_{i}$, are similar in nature to those of the JE experiment. The main differences occur at low values of $d_{z}$. The readout quality of the exclusive data is overall higher and that of the shared data lower. Finally, results for the CM architecture (cf. Fig. 1C), encoding the cross- modality predictions, are significantly different and confirm that it is possible to isolate the mutual information between different data sources. Notice how for $d_{e}=4$, the readout quality of the exclusive data $r_{e}$ seems to improve slowly as the dimension $d_{z}$ increases. We verified that the values of $r_{e}$ remain high for high values of $d_{z}$, measuring reconstruction errors converging around $0.8$. Thus, this architecture is more effective in stripping away any exclusive information. This is because, by definition, exclusive information cannot be encoded during the initial cross- modality prediction (Fig. 1C, orange part). ### 4.2 Increasing the Number of Modalities Promotes Retention of Shared Information Figure 5: Similarly to Fig. 4, each plot represents the reconstruction error of the readout operation for the exclusive data $r_{e}$ in blue, and for the shared data $r_{m}$ in red, as a function of the auto-encoder latent dimension. The four plots correspond to a different number $n$ of modalities. The results are presented for the three architectures A, B and C. Figure 5 shows the results for varying the number of modalities. For the JE and control architectures, they show how increasing the number of modalities reinforces the retention of the shared data over the exclusive data. Note how the reconstruction errors for the shared information (red curves) decay more rapidly for higher numbers of modalities $n$. This is in contrast to the CM architecture, where results are very similar for different numbers of modalities. This is because the initial cross-modality prediction network (Fig. 1C, orange part) effectively removes all modality-specific information, leaving essentially the same encoding task for the subsequent autoencoder despite the different numbers of modalities $n$. ### 4.3 Results on the Robot Dataset Figure 6: Readout reconstruction errors for the JE and CM approaches as a function of the size of the bottleneck of the encoding. Blue and red curves correspond to right and left arms respectively, solid lines correspond to information present in both modalities, dashed lines to information present in one modality only, and the dotted line to information present in none of the modalities. Figure 6 shows the readout errors for the proprioceptive information from both arms, for the JE and CM architecture, as a function of the size of the latent code. In both cases, the velocity information about the left arm, which is present in neither of the $2$ modalities and thus serves as a control factor, is not recovered for any latent code size. This is indicated by a chance-level reconstruction quality of $1.0$. For the JE architecture, the information which is present in both modalities (i.e. the right arm’s joint positions and the right end-effector position) is well reconstructed. For the CM architecture, however, the right arm’s joint positions are recovered with a MSE of around at best $0.2$. The reason for this is that the position of some of the joints is not visible at all in the image frames (like for example the last joint in the arm rotating the wrist). Furthermore, in some positions occlusion effects occur. The information about these joints is therefore not present in the mutual information and is thus filtered out by the cross- modality prediction. The joint velocity information inherently has a low entropy, making it easier to compress. As a result, the JE approach is very good at recovering this information. The other approach, however, has properly filtered it out, even-though some of the information seems to have leaked out during the proprioception $\rightarrow$ vision cross-modality prediction. This is understandable given the big increase in dimensionality taking place in this operation. For the CM approach, the position of the left arm, which is present only in the visual modality, is properly filtered out with MSEs greater than $0.7$ for all latent sizes. In the first approach however, the left end-effector position is recovered only for latent sizes greater than $14$, which corresponds to the point where the proprioception of the right arm is fully represented in the latent code. Figure 7: Reconstruction error of the visual modality for the JE and CM approaches as a function of the size of the bottleneck of the encoding. The error is split in two parts corresponding to the left and right halves of the frames. The results show that the pixels which share information with the proprioceptive modality are better reconstructed. Figure 7 shows the reconstruction error of the visual information as a function of the latent code size for the JE and CM approaches. The reconstruction error is split into two parts corresponding to the left and right half of the frames. Note that the chance reconstruction error is around $0.027$ (MSE). The results show that in both approaches, the pixels corresponding to the right arm are better reconstructed. In the JE approach, when the latent code size allows it, the entirety of the frame is encoded while for the CM approach, the left half of the frame is never encoded. Figure 8: We show for each bottleneck sizes in $\\{1,10,20,64\\}$ an image and it’s reconstruction through the autoencoder, as well as the mean reconstruction error map (darker indicates lower error). The JE approach (A) encodes the visual information about both arms if the information bottleneck allows for it, while the CM approach (B) reconstructs only one of the two arm for any bottleneck size. Finally, Fig. 8 shows concrete reconstructions of images by the encoding process for bottleneck sizes taken from the set $\\{1,10,20,64\\}$ for the JE and CM approaches. The results clearly show that the JM approach (A) goes from a regime where neither the left nor the right part of the frame is reconstructed, to a regime where the right part is, but not the left, to a regime where the whole frame is correctly reconstructed. For the CM approach (B), the system only goes through the $2$ first regimes. ## 5 Discussion Forming abstract representations is critical for higher-level intelligence. We have argued that the essence of abstraction is the lossy compression of information — stripping away details to arrive at a representation that transcends the original context and more easily generalizes to new situations. In principle, this could be done in many different ways. The critical question is what information to strip away and what residual information to keep. Depending on the task, there may be different answers to this question. For example, in a supervised classification setting a system may learn to strip away any information that is irrelevant for determining the class label. Or in a reinforcement learning setting, an agent may attempt to strip away any information that is not helpful for predicting future rewards. Here we have focused on an unsupervised approach for the learning of abstract representations through lossy compression of multimodal information. Our key result is that lossy compression of multimodal inputs through autoencoding naturally favors the retention of information that is shared across the sensory modalities. Such shared information may be particularly useful for generalizing to new situations. We first demonstrated our approach using synthetic multimodal data and then validated it using a simulated embodied agent (two-armed robot). The results indicate that the approach scales well to a more realistic scenario with visual and proprioceptive information. However, to compensate for the vastly different dimensionality of visual and proprioceptive data, we used a more complex network architecture where visual information was first passed through several convolutional layers before being integrated with proprioceptive information. It is important to stress that different sensory modalities have evolved in biological organisms (and are used in robots) exactly because they provide different, complementary information about the world. Discarding information that is not shared among modalities but unique to a single modality therefore seems to undermine this modality’s raison d’être. Indeed, we are not arguing that such information is not useful and should be discarded altogether. What we are arguing is that such information may be less useful when the goal is to learn highly abstract and compressed representations of the physical world. One of the greatest challenges for a developing mind is to make sense of what William James called the “blooming, buzzing, confusion” of sensations provided by different modalities, which eventually must become “coalesced together into one and the same space” [41]. The challenge thus is to identify how the inputs provided by the different sensory modalities relate to one another, i.e., what information they share. As we have seen, our generic approach is able to distill this shared information from raw sensor data. In this work, we have focused on generic autoencoder networks, as they are popular tools for dimensionality reduction and learning compressed representations of sensory signals in many contexts. In deep reinforcement learning, for example, they are frequently used to learn a compact abstract representation of high-dimensional (e.g., visual) input. In the future, it will be interesting to consider extensions of the generic autoencoder framework such as sparse autoencoders [12, 13] or other forms of regularized autoencoders such as beta-variational autoencoders [14], or encoder networks that learn to simultaneously predict rewards to focus limited encoding resources on relevant aspects of the multimodal sensory inputs that are associated with rewards. Our approach to learning abstract representations from multimodal data can also be related to approaches that learn view-invariant visual object representations through temporal coherence [16, 15]. In such approaches, temporal input from a single modality, typically the visual one, is considered, and a code is learned that is slowly varying, for example through trace rules or slow feature analysis (SFA). This corresponds to a lossy compression across time: fast changing information is discarded and slowly changing information is retained. For example, the fast changing information may be the pose of the object and the slowly changing information may be the object’s identity. Both jointly determine the current image. By retaining information that is slowly varying, a viewpoint invariant representation of the object’s identity can be learned, which can be used to recognize the object in different poses. This form of temporal compression of information is complementary to the multimodal compression we have considered here. In fact, the simple information theoretic argument from the Introduction applies in the same way if we replace the visual and haptic inputs $X_{v}$ and $X_{h}$ with, say, successive inputs from a single sensory modality. Lossy compression of such data will naturally favor the retention of information that the successive inputs have in common, i.e., information that is helpful for predicting $X_{t+1}$ from $X_{t}$ or vice versa. In the future, it will be interesting to consider learning abstract representations by abstracting across time and sensory modalities in hierarchical cognitive architectures for agents that learn abstract dynamic models of their interactions with the world. Multimodal compression as discussed here may also be an effective driver of intrinsically motivated exploration in infants and robots, reviewed in [26]. Schmidhuber proposed using compression progress as an intrinsic motivation signal [18]. In our own work, we have proposed the Active Efficient Coding (AEC) framework, that uses an intrinsic motivation for coding efficiency [19, 20, 21, 22]. AEC is an extension of classic efficient coding ideas [23] to active perception. Next to learning an efficient code for sensory signals, it proposes to control behavior in a way to maximize the information coming from different sources, while reducing the reconstruction loss during lossy compression. This has been shown to lead to the fully autonomous self- calibration of active perception systems, e.g., active stereo vision or active motion vision. However, this approach has not been considered in a multimodal setting. For example, consider an infant (or a developing robot) moving her hand in front of her face. In this situation, visual and proprioceptive signals are coupled. As the hand is felt to move to the left, the visual sense indicates motion to the left. Thus, the signals can be jointly encoded more compactly or with less reconstruction loss. An intrinsic motivation trying to minimize such reconstruction loss will therefore promote behaviors, where signals from the different modalities are strongly coupled. For example, banging a toy on the table creates correlated sensations in visual, proprioceptive, haptic, and auditory modalities and affords strong compression when jointly encoding these signals. We conjecture that AEC-like intrinsic motivations may drive infants to engage in such behaviors and may be effective in guiding open-ended learning in robots who try to understand the world around them. ## Appendix A Futility of a Symmetric Information Bottleneck We start from the original Information Bottleneck objective: $\min_{p(t|x)}I(X;T)-\beta I(T;Y),$ (37) where $p(T=t|X=x)$ describes the encoding of an input $x$ via its representation $t$, $I(X;T)$ is the mutual information between $X$ and $T$, $I(T;Y)$ is the mutual information between $T$ and $Y$, and $\beta$ serves to balance the two objectives. We now consider the reverse problem of $T$ trying to extract information from $Y$ that is useful for predicting $X$. This leads to the following “reverse” objective: $\min_{p(t|y)}I(Y;T)-\beta I(T;X),$ (38) where the roles of $X$ and $Y$ have simply been swapped. A naive approach to derive a symmetric version of the information bottleneck is to consider an encoding where $T$ encodes $X$ and $Y$ jointly via a probability $p(t|x,y)$, while minimizíng both the forward and the reverse functionals: $\min_{p(t|x,y)}I(X;T)-\beta I(T;Y)+I(Y;T)-\beta I(T;X),$ (39) which simplifies to: $\min_{p(t|x,y)}(1-\beta)I(X;T)+(1-\beta)I(Y;T)$ (40) because of the symmetry of the mutual information (e.g. $I(X;T)=I(T;X)$). Unfortunately, however, for $\beta<1$ this amounts to minimizing the information that $T$ contains about $X$ and $Y$, while for $\beta>1$ this amounts to maximizing the information that $T$ contains about $X$ and $Y$. In neither case will $T$ contain only the information that is shared between $X$ and $Y$. ## Appendix B Results for the AES Option Figure 9: Readout reconstruction errors for the JE and CM approaches as a function of the size of the bottleneck of the encoding. Blue and red curves correspond to right and left arms respectively, solid lines correspond to information present in both modalities, dashed lines to information present in one modality only, and the dotted line to information present in none of the modalities. Figure 10: Reconstruction error of the visual modality for the JE and CM approaches as a function of the size of the bottleneck of the encoding. The error is split in two parts corresponding to the left and right halves of the frames. The results show that the pixels which share information with the proprioceptive modality are better reconstructed. Figure 11: We show for each bottleneck sizes in $\\{1,10,20,64\\}$ an image and it’s reconstruction through the autoencoder, as well as the mean reconstruction error map (darker indicates lower error). The JE approach (A) encodes the visual information about both arms if the information bottleneck allows for it, while the CM approach (B) reconstructs only one of the two arm for any bottleneck size. We here present the results for the AES option. Overall, the results for the default and the AES options are similar, thus showing that the underlying principle is independent of the implementation details. Compared to the default option, the information about the left arm seems to be marginally better filtered out in the AES option for the CM approach. Figure 9 shows the reconstruction error of the various components of the proprioceptive data of both arms as a function of the bottleneck size. Similarly to Fig. 6, the information contained in both the visual and the proprioceptive modalities is better reconstructed than the information present in only one of the modalities. Figure 10 shows the reconstruction error of the data from the vision sensor as a function of the bottleneck size. The two curves correspond to the left and right half of the frames. Similarly to the results presented in Fig. 7, the part of the image corresponding to the arm which is jointly encoded is better reconstructed than the other. In the CM approach, the other arm is not reconstructed at all. Finally, Fig. 11 shows concrete reconstructions obtained from the encoding process, and the mean reconstruction error map for bottleneck sizes in $\\{1,10,20,64\\}$. For Figs. 10 and 11, the frame reconstruction is $B\left(D_{\text{post}}\left(z,\theta_{D_{\text{post}}}\right),\theta_{B}\right)$ with $B$ the decoder part of the autoencoder learning the latent representation $y_{v}$. ## Acknowledgments This work was supported by the European Union’s Horizon 2020 Research and Innovation Program under Grant Agreement No 713010 (GOAL-Robots – Goal-based Open-ended Autonomous Learning Robots). JT acknowledges support from the Johanna Quandt foundation. ## References * [1] LeCun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11), 2278-2324. * [2] He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 770-778). * [3] Zhang, X., Zou, J., He, K., & Sun, J. (2015). Accelerating very deep convolutional networks for classification and detection. IEEE transactions on pattern analysis and machine intelligence, 38(10), 1943-1955. * [4] Su, J., Vargas, D. V., & Sakurai, K. (2019). One pixel attack for fooling deep neural networks. IEEE Transactions on Evolutionary Computation, 23(5), 828-841. * [5] Brown, T. B., Mané, D., Roy, A., Abadi, M., & Gilmer, J. (2017). Adversarial patch. arXiv preprint arXiv:1712.09665. * [6] Sutton, R. S., & Barto, A. G. (2018). Reinforcement learning: An introduction. MIT press. * [7] Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D., & Riedmiller, M. (2013). Playing atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602. * [8] Van Hasselt, H., Guez, A., & Silver, D. (2015). Deep reinforcement learning with double q-learning. arXiv preprint arXiv:1509.06461. * [9] Lloyd, S. (1982). Least squares quantization in PCM. IEEE transactions on information theory, 28(2), 129-137. * [10] Kramer, M. A. (1991). Nonlinear principal component analysis using autoassociative neural networks. AIChE journal, 37(2), 233-243. * [11] Kingma, D. P., & Ba, J. (2014). Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980. * [12] Makhzani, A., & Frey, B. (2013). K-sparse autoencoders. arXiv preprint arXiv:1312.5663. * [13] Ng, A. (2011). Sparse autoencoder. CS294A Lecture notes, 72(2011), 1-19. * [14] Higgins, I., Matthey, L., Pal, A., Burgess, C., Glorot, X., Botvinick, M. & Lerchner, A. (2016). beta-vae: Learning basic visual concepts with a constrained variational framework. * [15] Einhäuser, W., Hipp, J., Eggert, J., Körner, E., & König, P. (2005). Learning viewpoint invariant object representations using a temporal coherence principle. Biological Cybernetics, 93(1), 79-90. * [16] Földiák, P. (1991). Learning invariance from transformation sequences. Neural Computation, 3(2), 194-200. * [17] Berkes, P., & Wiskott, L. (2005). Slow feature analysis yields a rich repertoire of complex cell properties. Journal of vision, 5(6), 9-9. * [18] Schmidhuber, J. (2008, June). Driven by compression progress: A simple principle explains essential aspects of subjective beauty, novelty, surprise, interestingness, attention, curiosity, creativity, art, science, music, jokes. In Workshop on anticipatory behavior in adaptive learning systems (pp. 48-76). Springer, Berlin, Heidelberg. * [19] Zhao, Y., Rothkopf, C. A., Triesch, J., & Shi, B. E. (2012, November). A unified model of the joint development of disparity selectivity and vergence control. In 2012 IEEE International Conference on Development and Learning and Epigenetic Robotics (ICDL) (pp. 1-6). IEEE. * [20] Vikram, T. N., Teulière, C., Zhang, C., Shi, B. E., & Triesch, J. (2014, October). Autonomous learning of smooth pursuit and vergence through active efficient coding. In 4th International Conference on Development and Learning and on Epigenetic Robotics (pp. 448-453). IEEE. * [21] Eckmann, S., Klimmasch, L., Shi, B. E., & Triesch, J. (2020). Active efficient coding explains the development of binocular vision and its failure in amblyopia. Proceedings of the National Academy of Sciences, 117(11), 6156-6162. * [22] Wilmot, C., Shi, B. E., & Triesch, J. (2020, October). Self-Calibrating Active Binocular Vision via Active Efficient Coding with Deep Autoencoders. In International Conference on Development and Learning and on Epigenetic Robotics. IEEE. * [23] Barlow, H. B. (1961). Possible principles underlying the transformation of sensory messages. Sensory communication, 1, 217-234. * [24] Bengio, Y., Courville, A., & Vincent, P. (2013). Representation learning: A review and new perspectives. IEEE transactions on pattern analysis and machine intelligence, 35 (8), 1798–1828. * [25] Weiss, K., Khoshgoftaar, T. M., & Wang, D. (2016). A survey of transfer learning. Journal of Big Data, 3 (1), 9. * [26] Baldassarre, G., & Mirolli, M. (Eds.). (2013). Intrinsically motivated learning in natural and artificial systems (No. 907). Berlin: Springer. * [27] Attneave, F. (1954). Some informational aspects of visual perception. Psychological review , 61 (3), 183. * [28] Hyvärinen, A., & Oja, E. (2000). Independent component analysis: algorithms and applications. Neural networks, 13 (4-5), 411–430. * [29] Olshausen, B. A., & Field, D. J. (1997). Sparse coding with an overcomplete basis set: A strategy employed by v1? Vision research, 37 (23), 3311–3325. * [30] Zhao, Y., Rothkopf, C. A., Triesch, J., & Shi, B. E. (2012). A unified model of the joint development of disparity selectivity and vergence control. In 2012 ieee international conference on development and learning and epigenetic robotics (icdl) (pp. 1–6). * [31] Eckmann, S., Klimmasch, L., Shi, B. E., & Triesch, J. (2020). Active efficient coding explains the development of binocular vision and its failure in amblyopia. Proceedings of the National Academy of Sciences, 117 (11), 6156–6162. * [32] Rao, R. P., & Ballard, D. H. (1999). Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects. Nature neuroscience, 2 (1), 79–87. * [33] Becker, S., & Hinton, G. E. (1992). Self-organizing neural network that discovers surfaces in random-dot stereograms. Nature, 355 (6356), 161–163. * [34] Hotelling, H. (1936). Relations between two sets of variates. Biometrika, 28 (3–4), 321–377. Retrieved from https://doi.org/10.1093/biomet/28.3-4.321 doi: 10.1093/biomet/28.3-4.321 * [35] Tishby, N., Pereira, F. C., & Bialek, W. (1999). The information bottleneck method. Proceedings of The 37th Allerton Conference on Communication, Control, & Computing, Univ. of Illinois. * [36] Ng, A., et al. (2011). Sparse autoencoder. CS294A Lecture notes, 72 (2011), 1–19. * [37] Rifai, S., Vincent, P., Muller, X., Glorot, X., & Bengio, Y. (2011). Contractive auto-encoders: Explicit invariance during feature extraction. In ICML. * [38] Vincent, P., Larochelle, H., Bengio, Y., & Manzagol, P.-A. (2008). Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th international conference on machine learning (pp. 1096–1103). * [39] Kingma, D. P., & Welling, M. (2013). Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 . * [40] Makhzani, A., Shlens, J., Jaitly, N., Goodfellow, I., & Frey, B. (2015). Adversarial autoencoders. arXiv preprint arXiv:1511.05644 . * [41] James, W. (1890). The principles of psychology. Henry Holt and Company. | Charles Wilmot attended preparatory classes for France’s Grandes Ecoles at the University of Valenciennes and recieved the B.E. in maths in 2013, and received an engineering diploma and the M.E. in information technologies at the École nationale supérieure d’électronique, informatique, télécommunications, mathématiques et mécanique de Bordeaux in 2016. He then integrated the research team of Jochen Triesch in Frankfurt, where he studied developmental robotics, intrinsically motivated reinforcement learning and hierarchical reinforcement learning in the intent of obtaining the Ph.D. in 2021. ---|--- | Gianluca Baldassarre received the B.A. and M.A. degrees in economics from the Sapienza University of Rome, Italy, in 1998, the Diploma of the Specialization Course “Cognitive Psychology and Neural Networks” from the same University, in 1999, and the Ph.D. degree in Computer Science from the University of Essex, Colchester, U.K., in 2003. He was later a postdoc with the Italian Institute of Cognitive Sciences and Technologies, National Research Council (ISTC-CNR), Rome. Since 2006 he has been a researcher, now a Research Director, with ISTC-CNR where he founded and is Coordinator of the Research Group “Laboratory of Embodied Natural and Artificial Intelligence”. He was the Principal Investigator for the EU project “ICEA–Integrating Cognition Emotion and Autonomy” from 2006 to 2009, the Coordinator of the EU Integrated Project “IM-CLeVeR–Intrinsically Motivated Cumulative Learning Versatile Robots” from 2009 to 2013, and, since 2016, he has been the Coordinator of the EU FET-OPEN Project “GOALRobots–Goal-Based Open-ended Autonomous Learning Robots” ending in 2021. His research interests span open- ended learning of sensorimotor skills, driven by extrinsic and intrinsic motivations, in animals, humans, and robots. ---|--- | Jochen Triesch received his Diploma and Ph.D. degrees in Physics from the University of Bochum, Germany, in 1994 and 1999, respectively. After two years as a post-doctoral fellow at the Computer Science Department of the University of Rochester, NY, USA, he joined the faculty of the Cognitive Science Department at UC San Diego, USA as an Assistant Professor in 2001. In 2005 he became a Fellow of the Frankfurt Institute for Advanced Studies (FIAS), in Frankfurt am Main, Germany. In 2006 he received a Marie Curie Excellence Center Award of the European Union. Since 2007 he is the Johanna Quandt Research Professor for Theoretical Life Sciences at FIAS. He also holds professorships at the Department of Physics and the Department of Computer Science and Mathematics at the Goethe University in Frankfurt am Main, Germany. In 2019 he obtained a visiting professorship at the Université Clermont Auvergne, France. His research interests span Computational Neuroscience, Machine Learning, and Developmental Robotics. ---|---
# On the Pythagoras number of the simplest cubic fields Magdaléna Tinková Charles University, Faculty of Mathematics and Physics, Department of Algebra, Sokolovská 83, 18600 Praha 8, Czech Republic <EMAIL_ADDRESS> ###### Abstract. The simplest cubic fields $\mathbb{Q}(\rho)$ are generated by a root $\rho$ of the polynomial $x^{3}-ax^{2}-(a+3)x-1$ where $a\geq-1$. In this paper, we will show that the Pythagoras number of the order $\mathbb{Z}[\rho]$ is equal to $6$ for $a\geq 3$. ###### Key words and phrases: Pythagoras number, the simplest cubic fields, indecomposable integers ###### 2010 Mathematics Subject Classification: 11R16, 11R80, 11E25 The author was supported by Czech Science Foundation GAČR, grant 21-00420M, by projects PRIMUS/20/SCI/002, UNCE/SCI/022, GA UK 1298218 from Charles University, and by SVV-2020-260589. ## 1\. Introduction Let $\mathcal{O}$ be a commutative ring, and let $\sum\mathcal{O}^{2}$ and $\sum^{m}\mathcal{O}^{2}$ be the sets defined by $\sum\mathcal{O}^{2}=\Big{\\{}\sum_{i=1}^{n}\alpha_{i}^{2};\;\alpha_{i}\in\mathcal{O},n\in\mathbb{N}\Big{\\}},\qquad\sum^{m}\mathcal{O}^{2}=\Big{\\{}\sum_{i=1}^{m}\alpha_{i}^{2};\;\alpha_{i}\in\mathcal{O}\Big{\\}}.$ In this paper, we are concerned with the so-called Pythagoras number $\mathcal{P}(\mathcal{O})$ of the ring $\mathcal{O}$ given by $\mathcal{P}(\mathcal{O})=\min\Big{\\{}m\in\mathbb{N};\;\sum\mathcal{O}^{2}=\sum^{m}\mathcal{O}^{2}\Big{\\}}.$ Regarding some basic examples, $\mathcal{P}(\mathbb{R})=\mathcal{P}(\mathbb{C})=1$, and Lagrange’s famous four-square theorem implies $\mathcal{P}(\mathbb{Q})=4$. Moreover, it can be proved that $\mathcal{P}(K)\leq 4$ for every number field $K$ [14, 36]. Arguably the most important and classical cases are Pythagoras numbers of rings of algebraic integers $\mathcal{O}_{K}$ of totally real number fields $K$. The first result is, of course, Lagrange’s above-mentioned theorem giving $\mathcal{P}(\mathbb{Z})=4$ that led to the study of universal quadratic forms. Let $\mathcal{O}_{K}^{+}$ be the set of totally positive integers of $\mathcal{O}_{K}$ (by this, we mean those algebraic integers whose conjugates are all positive). Roughly speaking, universal quadratic form over $\mathcal{O}_{K}$ is a quadratic form which has coefficients from $\mathcal{O}_{K}$ and which represents all the elements in $\mathcal{O}_{K}^{+}$. For more details about universal quadratic forms, see also for example [2, 3, 7, 15, 17, 21, 22, 33]. Considering sums of squares, Maaß has shown that the sum of three squares is universal over $\mathcal{O}_{K}$ for $K=\mathbb{Q}(\sqrt{5})$, which implies $\mathcal{P}(\mathcal{O}_{K})=3$ in this case [27]. Nevertheless, the following result of Siegel says that a sum of any number of squares can be universal only in the fields $\mathbb{Q}$ and $\mathbb{Q}(\sqrt{5})$ [37]. It means that in the other totally real number fields, we cannot express all the elements of $\mathcal{O}_{K}^{+}$ as a sum of squares, and thus we must restrict to those which indeed lie in $\sum\mathcal{O}_{K}^{2}$. Let now $\mathcal{O}\subseteq\mathcal{O}_{K}$ be an order. Scharlau showed that the Pythagoras number of an order is always finite, although it can be arbitrarily large [34]. The case of quadratic orders was in great detail studied by Peters; he proved that except for a few cases, the Pythagoras number is always $5$ [30]. Moreover, he also characterized all the elements which are representable as a sum of squares. Considering the other cases, recently, Kala and Yatsyna [20] proved that $\mathcal{P}(\mathcal{O})\leq f(d)$ for every order $\mathcal{O}$ in totally real number $K$, where $f$ is a function depending only on the degree $d$ of the field $K$. Moreover, one can take $f(d)=d+3$ if $2\leq d\leq 5$. Note that their subsequent paper [19] studies sums of squares in certain subrings of $\mathcal{O}_{K}$. However, given the difficulty of studying $\mathcal{P}(\mathcal{O})$ for orders, most of the research so far focused on the situation over fields. In the case of non-formally real fields $K$ (i.e., in which $-1$ can be expressed as a sum of squares), the Pythagoras number is closely related to Stufe $s(K)$ of $K$, which is the minimal number of squares whose sum gives $-1$. We have here $s(K)\leq\mathcal{P}(K)\leq s(K)+1$. By the results of Pfister [31], the value of $s(K)$ can attain only the powers of $2$, which greatly limits the possibilities for the value of $\mathcal{P}(K)$. On the other hand, Hoffmann has shown that for every $n\in\mathbb{N}$ and formally real field $K_{0}$, there exists a formally real field $K$ over $K_{0}$ with $\mathcal{P}(K)=n$ [14]. Nevertheless, we can find many other results on the Pythagoras number of fields in the number of specific situations, for example, in relation with rational function fields, elliptic curves, Hasse number or Laurent series [5, 8, 16, 32]. In this paper, we will focus on orders in the so-called simplest cubic fields [9, 35]. They are generated by a root $\rho$ of the polynomial $x^{3}-ax^{2}-(a+3)x-1$ where $a\geq-1$, and were richly studied in many different contexts, see for example [1, 4, 11, 24, 25, 26, 38]. This is due to the fact that they have many useful properties: They contain units of all signatures, and every totally positive unit is a square [28]. Moreover, $\mathcal{O}_{K}=\mathbb{Z}[\rho]$ for infinitely many $a$ (for example, if the square root $a^{2}+3a+9$ of the discriminant is squarefree), and they are also cyclic. In this case, the result of Kala and Yatsyna gives the upper bound $6$ on $\mathcal{P}(\mathcal{O})$. We will show that this bound is attained in infinitely many cases by proving the following theorem: ###### Theorem 1.1. Let $\rho$ be a root of the polynomial $x^{3}-ax^{2}-(a+3)x-1$ where $a\geq 3$. Then $\mathcal{P}(\mathbb{Z}[\rho])=6$. To the best of our knowledge, there are no results on the Pythagoras number for orders in number fields of higher degrees similar to Peters’ results on quadratic orders. Thus, Theorem 1.1 represents the first breakthrough in this problem. Moreover, since $\mathcal{O}_{K}=\mathbb{Z}[\rho]$ in infinitely many cases of $a$, this conclusion also provides us a precise result for the maximal order $\mathcal{O}_{K}$. Having the upper bound from the result of Kala and Yatsyna, we will focus on the determination of the lower bound. To reach this aim, we will primarily rely on the idea of additively indecomposable integers in totally real algebraic fields. Probably the most studied case is when we consider totally positive elements. Let $\alpha\in\mathcal{O}_{K}^{+}$. We say that $\alpha$ is indecomposable in $\mathcal{O}_{K}$ if we cannot express it as $\alpha=\beta_{1}+\beta_{2}$ where $\beta_{1},\beta_{2}\in\mathcal{O}_{K}^{+}$. Note that under the name extremal elements, they can be found in the above-mentioned Siegel’s proof of the (non)-universality of sums of squares in number fields. However, in our proofs, we will need their extended definition for all the possible signatures (see Section 2). Regarding real quadratic fields, the indecomposable integers were fully described by Perron [29] and Dress and Scharlau [10], and their additive structure was studied in [13]. Some partial results for the biquadratic case can be found in the work of Čech, Lachman, Svoboda, Zemková and the present author [6], and in the following paper [23], which focuses on ternary quadratic forms in these fields. The cubic fields are in the center of interest of [18], where we also determined the full structure of indecomposable integers in the simplest cubic fields. The proof of Theorem 1.1 is based on this result. Nevertheless, so far, the indecomposable integers have been mainly used in the study of universal quadratic forms [2, 3, 6, 17, 18, 23, 37, 39] or the elements of small norms [18], thus this paper also provides a new application of this phenomenon. Moreover, some of the ideas introduced here can be also used for the determination of the Pythagoras number for other cubic orders. The proof of Theorem 1.1 is provided in Section 3. Moreover, in Section 4 we give some partial results on the Pythagoras number for the remaining cases $-1\leq a\leq 2$. ## 2\. Preliminaries Let $K=\mathbb{Q}(\rho)$ be a totally real cubic field, and let $\mathcal{O}_{K}$ be the ring of algebraic integers of $K$. Moreover, let $\rho^{\prime}$ and $\rho^{\prime\prime}$ be Galois conjugates of $\rho$. Then by signature $\sigma$ of $\alpha\in\mathbb{Q}(\rho)$, we mean the triple $(\text{sgn}(\alpha),\,\text{sgn}(\alpha^{\prime}),\,\text{sgn}(\alpha^{\prime\prime}))$ where sgn is the signum function, and $\alpha^{\prime}$ and $\alpha^{\prime\prime}$ are images of $\alpha$ under the $\mathbb{Q}$-isomorphism given by $\rho\mapsto\rho^{\prime}$, and $\rho\mapsto\rho^{\prime\prime}$, respectively. In the following, we will use symbols $+$ and $-$ instead of $\pm 1$, e.g., we will replace $(1,1,1)$ by $(+,+,+)$. Moreover, the norm of $\alpha$ is defined as $N(\alpha)=\alpha\alpha^{\prime}\alpha^{\prime\prime}$, and the trace of $\alpha$ as $\text{Tr}(\alpha)=\alpha+\alpha^{\prime}+\alpha^{\prime\prime}$. Let $\mathcal{O}\subseteq\mathcal{O}_{K}$ be an order in $K$, i.e., a subring of finite index in $\mathcal{O}_{K}$. By $\mathcal{O}^{\sigma}$, we will mean the set of those elements in $\mathcal{O}$ which have the signature $\sigma$. The element $\alpha\in\mathcal{O}^{\sigma}$ is $\sigma$-indecomposable in $\mathcal{O}$ if it cannot be written as $\alpha=\beta_{1}+\beta_{2}$ where $\beta_{1},\beta_{2}\in\mathcal{O}^{\sigma}$. Otherwise, we say that the element $\alpha$ is $\sigma$-decomposable in $\mathcal{O}$. Note that for example, all the units are $\sigma$-indecomposable for some signature $\sigma$. In particular, the totally positive elements, i.e., the elements with the signature $(+,+,+)$, were richly studied in the past, and for them, we will introduce some more notation. We will denote the subset of totally positive elements of $\mathcal{O}$ by $\mathcal{O}^{+}$. We say that $\alpha\in\mathcal{O}^{+}$ is totally greater than $\beta\in\mathcal{O}^{+}$ if $\alpha>\beta$, $\alpha^{\prime}>\beta^{\prime}$ and $\alpha^{\prime\prime}>\beta^{\prime\prime}$. We will denote it by $\alpha\succ\beta$. Sometimes, we will also use the symbol $\succeq$ when we want to include the case when $\alpha=\beta$. Note that, for example, all non- zero squares are totally positive. Let us now recall some facts about the simplest cubic fields, which we study in this paper. They are generated by a root of the polynomial $x^{3}-ax^{2}-(a+3)x-1$. Troughout this paper, we will denote the roots of this polynomial in the following way: $a+1<\rho$, $-2<\rho^{\prime}<-1$, and $-1<\rho^{\prime\prime}<0$. Nevertheless, if $a\geq 7$, we have more precise estimates on these roots, namely (2.1) $a+1<\rho<a+1+\frac{2}{a},\ \ -1-\frac{1}{a+1}<\rho^{\prime}<-1-\frac{1}{a+2},\text{ and }-\frac{1}{a+2}<\rho^{\prime\prime}<-\frac{1}{a+3}.$ Note that this result mostly comes from [24], only the original estimate for $\rho^{\prime}$ was too rough for the purposes of this paper, so we have stated its slightly improved form, which can be easily checked. We will use these estimates many times in the following proofs. Besides that, we will use the fact that we know the full structure of $\sigma$-indecomposable integers in $\mathbb{Z}[\rho]$. In particular, in [18], we have shown the following theorem: ###### Theorem 2.1 ([18, Theorem 1.2]). Let $K$ be the simplest cubic field with $a\in\mathbb{Z}_{\geq-1}$ such that $\mathcal{O}_{K}=\mathbb{Z}[\rho]$. The elements $1$, $1+\rho+\rho^{2}$, and $-v-w\rho+(v+1)\rho^{2}$ where $0\leq v\leq a$ and $v(a+2)+1\leq w\leq(v+1)(a+1)$ are, up to multiplication by totally positive units, all the totally positive indecomposable integers in $\mathbb{Q}(\rho)$. Note that in fact, Theorem 2.1 provides us all the totally positive indecomposable integers in the order $\mathbb{Z}[\rho]$ for every $a\geq-1$. Moreover, although this theorem considers only the totally positive indecomposable integers, it gives us also the complete information about $\sigma$-indecomposables for all the other signatures $\sigma$. These $\sigma$-indecomposables can be obtained as $\varepsilon\alpha$ where $\varepsilon$ runs over all the units with signature $\sigma$, and $\alpha$ runs over all the elements listed in Theorem 2.1. This property is given by the fact that $\mathbb{Q}(\rho)$ contains units of all signatures. Moreover, we can divide the totally positive indecomposable integers from Theorem 2.1 into three sets: units, the exceptional indecomposable integer $1+\rho+\rho^{2}$ and the “triangle” of indecomposables of the form $\blacktriangle=\blacktriangle(a)=\\{-v-w\rho+(v+1)\rho^{2},0\leq v\leq a\text{ and }v(a+2)+1\leq w\leq(v+1)(a+1)\\}.$ Nevertheless, except for $\alpha\in\blacktriangle$, the set $\blacktriangle$ also contains some specific unit multiples of conjugates of $\alpha$. Let $\alpha(v,W)=-v-(v(a+2)+1+W)\rho+(v+1)\rho^{2}\in\blacktriangle$ for some $0\leq W\leq a-v$, and let $a=3A+a_{0}$ where $a_{0}\in\\{0,1,2\\}$. Instead of $\blacktriangle$, we can consider its subset of the form $\blacktriangle_{0}=\blacktriangle_{0}(a)=\left\\{\begin{array}[]{ll}\left\\{\alpha(v,W);0\leq v\leq A\text{ and }v\leq W\leq a-2v-1\right\\}\text{ if }a_{0}\in\\{1,2\\},\\\ \\{\alpha(v,W);0\leq v\leq A-1\text{ and }v\leq W\leq a-2v-1\\}\cup\\{\alpha(A,A)\\}\text{ if }a_{0}=0.\end{array}\right.$ The excluded elements of $\blacktriangle$ are just these unit multiples of conjugates of $\blacktriangle_{0}$, and thus in some sense, covered by the elements in $\blacktriangle_{0}$. For more details, see [18]. In our proof, we will work with norms of these elements, and in particular, we will use the following lemma from [18], which partly compares norms of elements belonging to the set $\blacktriangle_{0}$. ###### Lemma 2.2 ([18, Lemma 6.4]). Let $a\geq 3$ and assume that $\alpha(v+1,W)\in\blacktriangle_{0}$. Then $N(\alpha(v,W))<N(\alpha(v+1,W))$. Note that for fixed $v$, the norm of $\alpha(v,W)$ firstly increases in $W$ and then it can start to decrease (in some cases, it increases in the whole interval for $W$ but one of these two cases always occurs). For more details, see [18]. As we will see below, we will also need to know more about units in $\mathbb{Q}(\rho)$. It was proved that the system of fundamental units of $\mathbb{Q}(\rho)$ (and also of $\mathbb{Z}[\rho]$) is formed by the pair $\rho$ and $\rho^{\prime}$ [12, 35]. Benefiting from this property, the authors of [18] prove the following lemma, which we will often use in this paper. ###### Lemma 2.3 ([18, Lemma 6.2]). Let $a\geq 7$ and let $\varepsilon$ be a unit such that $|\varepsilon|,|\varepsilon^{\prime}|,|\varepsilon^{\prime\prime}|<a$. Then $\varepsilon=1$. Especially, if $\varepsilon\neq 1$ is a totally positive unit (and thus a square), then by Lemma 2.3, some of its conjugates is greater than $a^{2}$. In some cases, we will also need the stronger result stated in the following lemma [18]. ###### Lemma 2.4 ([18, Lemma 6.3]). Let $a\geq 7$ and let $\varepsilon$ be a totally positive unit such that $\varepsilon>a^{2}$. If $\varepsilon\neq\rho^{2},\rho^{\prime\prime-2}$, then at least one of the following holds: 1. (1) $\varepsilon>a^{4}$, or 2. (2) $\varepsilon^{\prime}>a^{2}$, or 3. (3) $\varepsilon^{\prime\prime}>a^{2}$. ## 3\. Proof of Theorem 1.1 Now we will describe the method which we will use in the proof of Theorem 1.1. Recall that by the result of Kala and Yatsyna [19], the upper bound on the Pythagoras number in cubic orders is $6$. Thus, it suffices to prove that the lower bound is also $6$. To do that, it is enough to find an element $\gamma\in\mathbb{Z}[\rho]$ which can be written as a sum of six squares but not as a sum of five squares. Hence we will proceed as follows. We will suitably choose such an element $\gamma$ and find all the elements $\omega$ such that $\gamma\succeq\omega^{2}$. Every square decomposition of $\gamma$ can consist only of these elements. Then, using some combinatorics, we will show that none sum of five (or less) of these squares can give $\gamma$. In the determination of these squares, we will use the knowledge of $\sigma$-indecomposable integers in the simplest cubic fields originating from Theorem 2.1. Let $\omega$ be such that $\gamma\succeq\omega^{2}$. This element $\omega$ has some signature $\sigma$, and it can be thus expressed as $\omega=\sum_{i=1}^{n}\beta_{i}$ where $\beta_{i}$ are $\sigma$-indecomposable integers in $\mathbb{Z}[\rho]$, and $n\in\mathbb{N}$. Having this, we can see that $\gamma\succeq\sum_{i=1}^{n}\beta_{i}^{2}+2\sum_{\begin{subarray}{c}i,j=1\\\ i\neq j\end{subarray}}^{n}\beta_{i}\beta_{j}.$ Obviously, the squares $\beta_{i}^{2}$ are totally positive, as well as elements $\beta_{i}\beta_{j}$ for $i\neq j$ since $\beta_{i}$ and $\beta_{j}$ have the same signature $\sigma$. Thus, we can immediately conclude that $\gamma\succeq\beta_{i}^{2}$ for all $i=1,\ldots,n$. We will use this simple fact in the following way. First of all, we will find all the $\sigma$-indecomposable integers $\beta$ for all the signatures $\sigma$ such that $\gamma\succeq\beta^{2}$. Then, by summing these elements $\beta$ with the same signature $\sigma$, we will derive all the $\sigma$-decomposable integers $\omega$ satisfying $\gamma\succeq\omega^{2}$. Moreover, every of these $\sigma$-indecomposables $\beta$ can we rewritten as $\beta=\varepsilon\alpha$ where $\varepsilon$ is a unit and $\alpha$ is one of $1$, $1+\rho+\rho^{2}$ and elements of $\blacktriangle_{0}$, or one of their conjugates. Thus, we firstly detect the elements of this list whose squares have the norm smaller than $N(\gamma)$, and consequently use the results on units from Lemmas 2.3 and 2.4 to determine all the possible units $\varepsilon$ which indeed give $\beta^{2}=\varepsilon^{2}\alpha^{2}\preceq\gamma$. In the case of the simplest cubic fields, we can choose our element as $\gamma=a^{2}+a+8+(a^{2}-a+1)\rho+(2-a)\rho^{2}=1+1+1+4+\rho^{2}+(a+1+a\rho-\rho^{2})^{2}.$ As we see, we can write $\gamma$ as a sum of six squares. We will fix this choice of $\gamma$ and work with it for the rest of this paper. Moreover, we will show that except for a few cases of $a$, there exist only $8$ non-zero elements $\omega^{2}$ such that $\omega^{2}\preceq\gamma$, which is a great advantage of the choice of this element. Using estimates given in (2.1), we can easily deduce that for $a\geq 7$, $\displaystyle\gamma$ $\displaystyle<a^{2}+6a+9+\frac{2}{a},$ $\displaystyle\gamma^{\prime}$ $\displaystyle<10+\frac{4}{(a+2)^{2}}<11,$ $\displaystyle\gamma^{\prime\prime}$ $\displaystyle<a^{2}+11+\frac{a^{2}-8a-28}{(a+3)^{2}}.$ In particular, the conjugate $\gamma$ has the largest value. We can immediately see that, if $\omega$ is a rational integer, then necessarily $\omega^{2}\in\\{1,4,9\\}$. ### 3.1. Units Our first concern is to find all the totally positive units $\varepsilon$ satisfying $\gamma\succeq\varepsilon$. Recall that every such unit is a square, thus it can play a role in a square decomposition of the element $\gamma$. ###### Lemma 3.1. Let $a\geq 7$ and let $\varepsilon$ be a totally positive unit in $\mathbb{Z}[\rho]$. If $\gamma\succeq\varepsilon$, then $\varepsilon\in\\{1,\rho^{2}\\}$. ###### Proof. If $\varepsilon\neq 1$, Lemma 2.3 implies that one of the conjugates of $\varepsilon$ is greater than $a^{2}$. Without loss of generality, we can assume $\varepsilon>a^{2}$. Using the fundamental units, the unit $\varepsilon$ can be written as $\rho^{k}\rho^{\prime l}$ for some $k,l\in 2\mathbb{Z}$. As we can see from the estimates in (2.1), the value of $\varepsilon$ can be greater than $a^{2}$ only if $k\geq 2$. On the other hand, $\varepsilon^{\prime}=\rho^{\prime k}\rho^{\prime\prime l}\leq\gamma$ for $k\geq 2$ only if $l\geq-2$. This condition on $l$ also implies that $\varepsilon\leq\gamma$ only for $k=2$ and $a\geq 7$. The other cases are not possible as we have $\gamma>\gamma^{\prime},\gamma^{\prime\prime}$. Let us first focus on the case when $k=2$ and $l=-2$. Obviously, $\varepsilon^{\prime}>\gamma^{\prime\prime}$, thus the only conjugate of $\varepsilon$ which can be totally smaller $\gamma$ is $\varepsilon^{\prime}$. However, it can be directly verified that $\gamma-\rho^{\prime 2}\rho^{\prime\prime-2}$ is not totally positive for $a\geq 7$. Therefore, let $l\geq 0$. In these cases, clearly, $\varepsilon>\gamma^{\prime\prime}$ for $a\geq 7$, thus $\varepsilon$ is the only conjugate of $\varepsilon$ which can be totally smaller than $\gamma$. First of all, let us assume $l\geq 6$. In this case, we have $\varepsilon^{\prime\prime}=\rho^{\prime\prime 2}\rho^{l}>\frac{1}{(a+3)^{2}}(a+1)^{6}>\gamma^{\prime\prime}$ for $a\geq 7$, and thus we can exclude the cases with $l\geq 6$. Therefore, except for $\varepsilon=1$, we are left with the units $\rho^{2}$, $\rho^{2}\rho^{\prime 2}$ and $\rho^{2}\rho^{\prime 4}$. However, the last unit can be rewritten as $\rho^{\prime 2}\rho^{\prime\prime-2}$, which was excluded in the previous part. In the same manner, we can show that $\gamma-\rho^{2}\rho^{\prime 2}$ is not totally positive for $a\geq 7$. Thus, the only units which can be (and actually are) totally smaller than $\gamma$ are exactly $1$ and $\rho^{2}$. ∎ ### 3.2. Squares of $\sigma$-indecomposable integers In this part, we will find all non-unit $\sigma$-indecomposable integers $\beta$ such that $\gamma\succeq\beta^{2}$. Necessarily, in that case, $N(\gamma)\geq N(\beta^{2})$. It can be easily computed that $N(\gamma)=9a^{4}+22a^{3}+247a^{2}+258a+1493.$ In our investigation, we can use the knowledge of totally positive indecomposable integers given by Theorem 2.1, and the fact that every square of non-unit $\sigma$-indecomposable integer is a conjugate of some element of the form $\varepsilon\alpha^{2}$ where $\varepsilon$ is a totally positive unit and $\alpha\in\blacktriangle_{0}\cup\\{1+\rho+\rho^{2}\\}$. Thus, we firstly detect all the elements $\alpha$ for which $N(\alpha^{2})=N(\beta^{2})\leq N(\gamma)$. ###### Lemma 3.2. Let $a\geq 15$ and let $\alpha\in\blacktriangle_{0}\cup\\{1+\rho+\rho^{2}\\}$. If $N(\alpha^{2})\leq N(\gamma)$, then $\alpha\in\\{-w\rho+\rho^{2};1\leq w\leq a\\}\cup\\{1+\rho+\rho^{2},-1-(a+4)\rho+2\rho^{2}\\}.$ ###### Proof. It can be easily computed that $N((1+\rho+\rho^{2})^{2})<N(\gamma)$ for $a\geq-1$. Thus, let us now focus on $\alpha\in\blacktriangle_{0}$. In this case, we have $\alpha=\alpha(v,W)=-v-(a(v+2)+1+W)\rho+(v+1)\rho^{2}$ for some admissible values of $v,W$. In the following, we will use Lemma 2.2, which compares norms of elements belonging to $\blacktriangle_{0}$. Firstly, let us focus on the case when $v=1$. For $W=1$ (the smallest value of $W$ for $v=1$), we get $N(\alpha(1,1)^{2})=4a^{4}+24a^{3}-108a+81<N(\gamma)$ for $a\geq-1$, i.e., we obtain the element $-1-(a+4)\rho+2\rho^{2}$ listed in the statement of the lemma. On the other hand, $N(\alpha(1,2)^{2})=9a^{4}+54a^{3}-141a^{2}-666a+1369>N(\gamma)$ for $a\geq 15$. Recall that the norm of $\alpha(v,W)$ for fixed $v$ increases in $W$, and then it can start to decrease. Thus, to complete the proof for $v=1$, it suffices to check the norm for $W=a-3$ (the largest $W$ for $v=1$ and $a\geq 15$). Nevertheless, we obtain $N(\alpha(1,a-3)^{2})=16a^{4}-136a^{2}+289>N(\gamma)$ for $a\geq 10$. By Lemma 2.2 and using the previous part, the norms of $\alpha(v,W)^{2}$ for $v\geq 2$ are too large to be smaller than $N(\gamma)$. Thus, we are left with one element with $v=1$ and all the elements with $v=0$, which completes the proof. ∎ Therefore, we have determined all the representatives of the $\sigma$-indecomposable integers $\alpha$ with sufficiently small norms. Now we will find all the totally positive units $\varepsilon$ for which $\varepsilon\alpha^{2}$ or one of its conjugates is indeed totally smaller than $\gamma$. To reach this aim, we will use Lemmas 2.3 and 2.4, which state some useful results about units in the simplest cubic fields. ###### Lemma 3.3. Let $a\geq 15$ and let $\beta$ be a non-unit $\sigma$-indecomposable integer in $\mathbb{Z}[\rho]$ for some signature $\sigma$. If $\gamma\succeq\beta^{2}$, then $\beta^{2}$ is one of the following elements: 1. (1) $\rho^{\prime 2}\rho^{\prime\prime 2}(-\rho+\rho^{2})^{2}=1-2\rho+\rho^{2}$, 2. (2) $\rho^{\prime\prime 2}\rho^{2}(-\rho^{\prime}+\rho^{\prime 2})^{2}=a^{2}+a+1+(a^{2}-a+1)\rho-(a-1)\rho^{2}$, 3. (3) $\rho^{\prime\prime 2}\rho^{2}(-2\rho^{\prime}+\rho^{\prime 2})^{2}=a^{2}-a+(a^{2}-3a+1)\rho-(a-3)\rho^{2}$, 4. (4) $\rho^{2}\rho^{\prime 2}(-(a-1)\rho^{\prime\prime}+\rho^{\prime\prime 2})^{2}=a^{2}+a-1+(a^{2}-a-3)\rho-(a-2)\rho^{2}$. ###### Proof. We will proceed as follows. We will consider the elements $\alpha$ given by Lemma 3.2 and discuss whether some conjugate of $\varepsilon\alpha^{2}$ can be totally smaller than $\gamma$ for some totally positive unit $\varepsilon$. Let us start with $\alpha=1+\rho+\rho^{2}$. Using (2.1), we can see that $\alpha^{2}>a^{4}+6a^{3}+15a^{2}+18a+9>a^{2}+6a+9+\frac{2}{a}>\gamma>\gamma^{\prime},\gamma^{\prime\prime}$ for $a\geq 15$. Thus, we can immediately exclude all the conjugates of $(1+\rho+\rho^{2})^{2}$, i.e., when we multiple $\alpha^{2}$ by the totally positive unit $\varepsilon=1$. Let now $\varepsilon\neq 1$. By Lemma 2.3, without loss of generality, we can suppose $\varepsilon>a^{2}$. However, we can immediately exclude the conjugates of $\varepsilon(1+\rho+\rho^{2})^{2}$, and we are left with the elements of the form $\varepsilon(1+\rho^{\prime}+\rho^{\prime 2})^{2}$ and $\varepsilon(1+\rho^{\prime\prime}+\rho^{\prime\prime 2})^{2}$. Now, we will use Lemma 2.4. If $\varepsilon>a^{4}$, it can be easily verified that $\varepsilon(1+\rho^{\prime}+\rho^{\prime 2})^{2},\varepsilon(1+\rho^{\prime\prime}+\rho^{\prime\prime 2})^{2}>\gamma$, thus these cases are not possible. Suppose $\varepsilon\neq\rho^{2},\rho^{\prime\prime-2}$. In that case, our unit $\varepsilon$ has a conjugate greater than $a^{2}$. Since this conjugate cannot be paired with $1+\rho+\rho^{2}$, we can restrict to the elements $\varepsilon(1+\rho^{\prime}+\rho^{\prime 2})^{2}$ with $a^{2}<\varepsilon,\varepsilon^{\prime}<a^{4}$ (the case $\varepsilon(1+\rho^{\prime\prime}+\rho^{\prime\prime 2})^{2}$ is covered by that). However, using a similar method as in Lemma 3.1, we can show that under these conditions, $\varepsilon=\rho^{2}\rho^{\prime-2}$. Nevertheless, it can be easily checked that none conjugate of $\rho^{2}\rho^{\prime-2}(1+\rho^{\prime}+\rho^{\prime 2})^{2}$ is totally smaller than $\gamma$. Thus we are left with the elements of the form $\varepsilon(1+\rho^{\prime}+\rho^{\prime 2})^{2}$ and $\varepsilon(1+\rho^{\prime\prime}+\rho^{\prime\prime 2})^{2}$ where $\varepsilon=\rho^{2},\rho^{\prime\prime-2}$. However, for these cases, we can directly verify that none of them (or their conjugates) is totally smaller than $\gamma$. We will proceed with the elements from $\blacktriangle_{0}$. First of all, let us assume that $\alpha=-w\rho+\rho^{2}$ where $3\leq w\leq a-2$. In this case, as before, $\alpha^{2}>\gamma$ (as $\alpha^{2}>((a+1)^{2}-w(a+2))^{2}\geq 4a^{2}+20a+25>\gamma$), thus we can exclude $\varepsilon=1$. Let now $\varepsilon>a^{2}$. Obviously, $\varepsilon\alpha^{2}>\gamma$ , and $\varepsilon\alpha^{\prime 2}>\gamma$ as $\varepsilon\alpha^{\prime 2}>a^{2}(w+1)^{2}\geq 16a^{2}>\gamma$. Hence only the conjugates of $\varepsilon\alpha^{\prime\prime 2}$ can be totally smaller than $\gamma$. Similarly as before, using Lemma 2.4, we can exclude all the units with $\varepsilon>a^{4}$, $\varepsilon^{\prime}>a^{2}$ and $\varepsilon^{\prime\prime}>a^{2}$, and neither of $\rho^{2}$ or $\rho^{\prime\prime-2}=\rho^{2}\rho^{\prime 2}$ produce an element totally smaller than $\gamma$, which can checked by direct computations. Put $\alpha=-1-(a+4)\rho+2\rho^{2}$. Using Lemma 2.3 and basic estimates (2.1), we can easily find candidates on totally smaller integers, which are conjugates of $\varepsilon\alpha^{\prime\prime 2}$ where $\varepsilon\in\\{\rho^{2}\rho^{\prime 2},\rho^{4}\rho^{\prime 2},\rho^{4}\rho^{\prime 4}\\}$. Nevertheless, by comparing norms, traces, or the remaining coefficient of the minimal polynomial of $\gamma$ and $\alpha$, we can exclude all of these several concrete cases. Let now $\alpha=-a\rho+\rho^{2}$. In this case, $\alpha>a^{2}-2a+1$, $\alpha^{\prime}>(a+1)^{2}$ and $\alpha^{\prime\prime}>\frac{a^{2}}{(a+3)^{2}}$. Nevertheless, $\text{Tr}(\alpha)=2a^{2}+10a+18>2a^{2}+2a+36=\text{Tr}(\gamma)$, thus we can exclude $\varepsilon=1$. Regarding $\varepsilon\neq 1$, the only possible candidates are again conjugates of $\rho^{2}\alpha^{\prime\prime 2}$ and $\rho^{2}\rho^{\prime 2}\alpha^{\prime\prime 2}$. Nevertheless, by direct calculations, we can easily show that none of them is totally smaller than $\gamma$. We will proceed with $\alpha=-(a-1)\rho+\rho^{2}$. In this case, we have $\alpha^{2}>4a^{2}+8+\frac{4}{a^{2}}>\gamma$, thus we can exclude $\varepsilon=1$. For $\varepsilon\neq 1$, similarly, as before, we are left with $\rho^{2}\alpha^{\prime\prime 2}$ and $\rho^{2}\rho^{\prime 2}\alpha^{\prime\prime 2}$. In this case, we indeed get an element, which is totally smaller than $\gamma$, and it is equal to $\rho^{2}\rho^{\prime 2}\alpha^{\prime\prime 2}=a^{2}+a-1+(a^{2}-a-3)\rho-(a-2)\rho^{2}.$ Let now $\alpha=-2\rho+\rho^{2}$. We can conclude that only the conjugates of $\rho^{2}\rho^{\prime 2}\alpha^{\prime\prime 2}$ can be totally smaller than $\gamma$, from which only the one satisfies this condition, namely $\rho^{\prime\prime 2}\rho^{2}\alpha^{\prime 2}=a^{2}-a+(a^{2}-3a+1)\rho-(a-3)\rho^{2}.$ Therefore, it remains to consider the element $\alpha=-\rho+\rho^{2}$. Using Lemma 2.4 and estimates (2.1), we can easily derive that some conjugate of our element has to be of the form $\varepsilon\alpha^{\prime\prime 2}$ where $\varepsilon\in\\{\rho^{2}\rho^{\prime 2},\rho^{4}\rho^{\prime 2},\rho^{4}\rho^{\prime 4}\\}$. From these nine cases, only two are actually totally smaller than $\gamma$, specifically $\displaystyle\rho^{\prime 2}\rho^{\prime\prime 2}\alpha^{2}$ $\displaystyle=1-2\rho+\rho^{2},$ $\displaystyle\rho^{\prime\prime 2}\rho^{2}\alpha^{\prime 2}$ $\displaystyle=a^{2}+a+1+(a^{2}-a+1)\rho-(a-1)\rho^{2}.$ The others can be excluded by direct calculations, completing the proof. ∎ ### 3.3. Sums of $\sigma$-indecomposable integers In the previous subsections, we have found all the squares of $\sigma$-indecomposable integers $\beta$ for all signatures $\sigma$ for which we have $\gamma\succeq\beta^{2}$. Now we will consider possible $\sigma$-decomposable integers, which we can create from these $\sigma$-indecomposables and which are (possibly) totally smaller than $\gamma$. However, to do that, we must know the signatures of these elements, which can be found in Table 1. $\sigma$-indecomposable integer $\beta$ | Signature of $\beta$ | Signature of $-\beta$ ---|---|--- $1$ | $(+,+,+)$ | $(-,-,-)$ $\rho$ | $(+,-,-)$ | $(-,+,+)$ $\rho^{\prime}\rho^{\prime\prime}(-\rho+\rho^{2})$ | $(+,-,-)$ | $(-,+,+)$ $\rho^{\prime\prime}\rho(-\rho^{\prime}+\rho^{\prime 2})$ | $(-,-,+)$ | $(+,+,-)$ $\rho^{\prime\prime}\rho(-2\rho^{\prime}+\rho^{\prime 2})$ | $(-,-,+)$ | $(+,+,-)$ $\rho\rho^{\prime}(-(a-1)\rho^{\prime\prime}+\rho^{\prime\prime 2})$ | $(-,+,-)$ | $(+,-,+)$ Table 1. Signatures of $\sigma$-indecomposable integers whose squares are totally smaller than $\gamma$. ###### Lemma 3.4. Let $a\geq 15$ and $\gamma\succeq\omega^{2}$. If $\omega$ is $\sigma$-decomposable for some $\sigma$, then $\omega^{2}\in\\{4,9\\}$. ###### Proof. Now we will consider possible sums of our $\sigma$-indecomposable integers. Note that we can sum up only the elements with the same signatures. Moreover, the opposite signatures (i.e., which have all the signs opposite) behave in the same manner and give the same squares, and thus it suffices to study only one of each such a pair. 1. (1) Signature $(+,+,+)$ (respectively, $(-,-,-)$): Here we have only one element, namely $1$, which produces two $\sigma$-decomposable integers $2$ and $3$ whose squares $4$ and $9$ are totally smaller than $\gamma$. 2. (2) Signature $(+,-,-)$ (respectively, $(-,+,+)$): The set of $\sigma$-indecomposables for this signature consists of the elements $\rho$ and $\rho^{\prime}\rho^{\prime\prime}(-\rho+\rho^{2})$. However, we can easily compute that 1. (a) $(2\rho)^{2}>4(a+1)^{2}>\gamma$, 2. (b) $(\rho+\rho^{\prime}\rho^{\prime\prime}(-\rho+\rho^{2}))^{2}=1-4\rho+4\rho^{2}>4a^{2}+4a-3>\gamma$, 3. (c) $(2\rho^{\prime}\rho^{\prime\prime}(-\rho+\rho^{2}))^{2}=4-8\rho+4\rho^{2}>4a^{2}-8>\gamma$ for $a\geq 15$. Moreover, these results imply that our $\omega$ cannot also be a sum of more than two $\sigma$-indecomposable integers. Thus, in this signature, none square of $\sigma$-decomposable integer is totally smaller than $\gamma$. 3. (3) Signature $(-,-,+)$ (respectively, $(+,+,-)$): In this case, we consider exactly two $\sigma$-indecomposable integers $\rho^{\prime\prime}\rho(-\rho^{\prime}+\rho^{\prime 2})$ and $\rho^{\prime\prime}\rho(-2\rho^{\prime}+\rho^{\prime})$. However, we can easily show that 1. (a) $((2\rho^{\prime\prime}\rho(-\rho^{\prime}+\rho^{\prime 2}))^{2})^{\prime\prime}>\gamma^{\prime\prime}$, 2. (b) $(\rho^{\prime\prime}\rho(-\rho^{\prime}+\rho^{\prime 2})+\rho^{\prime\prime}\rho(-2\rho^{\prime}+\rho^{\prime 2}))^{2})^{\prime\prime}>\gamma^{\prime\prime}$, 3. (c) $((2\rho^{\prime\prime}\rho(-2\rho^{\prime}+\rho^{\prime 2}))^{2})^{\prime\prime}>\gamma^{\prime\prime}$ for $a\geq 15$. Thus we do not obtain any additional element. 4. (4) Signature $(-,+,-)$ (respectively, $(+,-,+)$): This case contains exactly one $\sigma$-indecomposable integer, namely $\rho\rho^{\prime}(-(a-1)\rho^{\prime\prime}+\rho^{\prime\prime 2})$. However, it can be easily computed that $((2\rho\rho^{\prime}(-(a-1)\rho^{\prime\prime}+\rho^{\prime\prime 2}))^{2})^{\prime\prime}>\gamma^{\prime\prime}$ for $a\geq 15$, thus this case does not produce more elements to consider. ∎ ### 3.4. Proof of Theorem 1.1 Using the results of Lemmas 3.1, 3.3 and 3.4, we can now prove Theorem 1.1 stated in the introduction. ###### Proof of Theorem 1.1. In Subsections 3.1, 3.2 and 3.3, we have found all the elements $\omega$ such that $\gamma\succeq\omega^{2}$ for $a\geq 15$. We have obtained the following squares: 1. (1) rational integers $1$, $4$ and $9$, 2. (2) squares of $\sigma$-indecomposable integers of the form 1. (a) $\rho^{2}$, 2. (b) $1-2\rho+\rho^{2}$, 3. (c) $a^{2}+a+1+(a^{2}-a+1)\rho-(a-1)\rho^{2}$, 4. (d) $a^{2}-a+(a^{2}-3a+1)\rho-(a-3)\rho^{2}$, 5. (e) $a^{2}+a-1+(a^{2}-a-3)\rho-(a-2)\rho^{2}$. Using a computer program (all the calculations were performed in Mathematica), we can show that we get the same elements (and none more) also for $5\leq a\leq 14$. For $a=3$, we get two additional elements $20+11\rho-3\rho^{2}$ and $1+2\rho+\rho^{2}$, and for $a=4$, we obtain $1+2\rho+\rho^{2}$. Nevertheless, using a similar procedure as below, we can prove that even in these cases, we need at least $6$ squares to express $\gamma$. Thus, in the following, we will suppose $a\geq 5$. Recall that $\gamma=a^{2}+a+8+(a^{2}-a+1)\rho+(2-a)\rho^{2}$. The coefficient before $\rho$ of $\gamma$ is clearly odd, thus in every square decomposition of $\gamma$, we need at least one element with this coefficient odd. Looking at the list, this is satisfied by the elements $\displaystyle a^{2}+a+1+(a^{2}-a+1)\rho-(a-1)\rho^{2},$ $\displaystyle a^{2}-a+(a^{2}-3a+1)\rho-(a-3)\rho^{2},$ $\displaystyle a^{2}+a-1+(a^{2}-a-3)\rho-(a-2)\rho^{2}.$ Note that these elements are also the only ones, which have positive coefficient before $\rho$. However, for $a^{2}-a+(a^{2}-3a+1)\rho-(a-3)\rho^{2}$ and $a^{2}+a-1+(a^{2}-a-3)\rho-(a-2)\rho^{2}$, the value of this coefficients is strictly smaller than $a^{2}-a+1$. Thus, if our square decomposition of $\gamma$ contained one of these two element, some other summand would have to be one of these three above-mentioned elements. Nevertheless, in that case, the coefficient before $1$ (we mean the coefficients in the basis $1,\rho,$ and $\rho^{2}$) is at least $2a^{2}-2a>a^{2}+a+8$ for $a\geq 5$. This is not possible since all the squares totally smaller than $\gamma$ have a non- negative coefficient before $1$. Hence no square decomposition of $\gamma$ can contain the elements $a^{2}-a+(a^{2}-3a+1)\rho-(a-3)\rho^{2}$ and $a^{2}+a-1+(a^{2}-a-3)\rho-(a-2)\rho^{2}$. It implies that one summand of our decomposition must be $a^{2}+a+1+(a^{2}-a+1)\rho-(a-1)\rho^{2}$, and we get $\gamma=a^{2}+a+1+(a^{2}-a+1)\rho-(a-1)\rho^{2}+\delta,$ where $\delta=7+\rho^{2}$. Obviously, every square decomposition of $\delta$ may consist of only the elements $1$, $4$, $\rho^{2}$ and $1-2\rho+\rho^{2}$ since the coefficient before $1$ of the other elements from the list is too large. Nevertheless, $1-2\rho+\rho^{2}$ cannot appear in this decomposition since its coefficient before $\rho$ is negative, and the remaining three integers have this coefficient equal to zero. Thus, only the elements $1$, $4$, and $\rho^{2}$ can appear in a square decomposition of $\delta$, and for that, we need at least $5$ of these elements. It implies that every square decomposition of $\gamma$ consists of at least $6$ non-zero squares, which together with the upper bound, gives $\mathcal{P}(\mathbb{Z}[\rho])=6$. ∎ ## 4\. The case $-1\leq a\leq 2$ We will now focus on the remaining cases of $a$, i.e., $-1\leq a\leq 2$. However, the situation is different here. Indeed, at least the element $\gamma$ can be expressed as a sum of less than $6$ squares for all of these cases, thus it cannot provide us the same lower bound as before. Moreover, based on computer experiments, we may propose that the Pythagoras number of $\mathbb{Z}[\rho]$ is even less than $6$. Nevertheless, our computer program searches for elements of small traces, and thus we cannot exclude that there exists an element that can be written as a sum of more squares and has a large trace. The lower bounds on $\mathcal{P}(\mathbb{Z}[\rho])$ for $-1\leq a\leq 2$ are provided in Table 2. We also show here an example of an element for which this lower bound is attained. $a$ | $\mathcal{P}(\mathbb{Z}[\rho])$ | Example of element ---|---|--- $-1$ | $\geq 4$ | $7$ $0$ | $\geq 5$ | $-8\rho+8\rho^{2}$ $1$ | $\geq 5$ | $4-3\rho+2\rho^{2}$ $2$ | $\geq 5$ | $7+\rho^{2}$ Table 2. The lower bound on $\mathcal{P}(\mathbb{Z}[\rho])$ for $-1\leq a\leq 2$ and an example of element for which this lower bound is attained. ## Acknowledgements The author is greatly indebted to Pavlo Yatsyna and Vítězslav Kala for their advice during the preparation of this paper. ## References * [1] S. Balady, Families of cyclic cubic fields, J. Number Theory 167, 394–406 (2016). * [2] V. Blomer and V. Kala: Number fields without $n$-ary universal quadratic forms, Math. Proc. Cambridge Philos. Soc. 159 (2), 239–252 (2015). * [3] V. Blomer and V. Kala, On the rank of universal quadratic forms over real quadratic fields, Doc. Math. 23, 15–34 (2018). * [4] D. Byeon, Class number 3 problem for the simplest cubic fields, Proc. Amer. Math. Soc. 128, 1319–1323 (2000). * [5] J. W. S. Cassels, W. J. Ellison and A. Pfister, On sums of squares and on elliptic curves over function fields, J. Number Theory 3, 125–149 (1971). * [6] M. Čech, D. Lachman, J. Svoboda, M. Tinková and K. Zemková, Universal quadratic forms and indecomposables over biquadratic fields, Math. Nachr. 292, 540–555 (2019). * [7] W. K. Chan, M.-H. Kim and S. Raghavan, Ternary universal integral quadratic forms over real quadratic fields, Japan. J. Math. 22, 263–273 (1996). * [8] M. D. Choi, Z. D. Dai, T. Y. Lam and B. Reznick. The Pythagoras number of some affine algebras and local algebras, J. Reine Angew. Math. 336, 45–82 (1982). * [9] H. Cohn, A device for generating fields of even class number, Proc. Amer. Math. Soc. 7, 595–598 (1956). * [10] A. Dress and R. Scharlau: Indecomposable totally positive numbers in real quadratic orders, J. Number Theory 14, 292–306 (1982). * [11] K. Foster, HT90 and “simplest” number fields, Illinois J. Math. 55, 1621–1655 (2011). * [12] H. J. Godwin, The determination of units in totally real cubic fields, Proc. Cambridge Philos. Soc. 56, 318–321 (1960). * [13] T. Hejda and V. Kala, Additive structure of totally positive quadratic integers, Manuscripta Math. 163, 263–278 (2020). * [14] D. W. Hoffmann, Pythagoras numbers of fields, J. Amer. Math. Soc. 12 (3), 839–848 (1999). * [15] J. S. Hsia, Y. Kitaoka and M. Kneser, Representations of positive definite quadratic forms, J. Reine Angew. Math. 301, 132–141 (1978). * [16] Y. Hu, The Pythagoras number and the $u$-invariant of Laurent series fields in several variables, J. Algebra 426, 243–258 (2015). * [17] V. Kala, Universal quadratic forms and elements of small norm in real quadratic fields, Bull. Aust. Math. Soc. 94, 7–14 (2016). * [18] V. Kala and M. Tinková, Universal quadratic forms, small norms and traces in families of number fields, preprint. https://arxiv.org/abs/2005.12312 * [19] V. Kala and P. Yatsyna, Sums of squares in S-integers, New York J. Math. 26, 1145–1154 (2020). * [20] V. Kala and P. Yatsyna, Lifting problem for universal quadratic forms, Adv. Math. 377, 107497 (2021). * [21] B. M. Kim, Finiteness of real quadratic fields which admit positive integral diagonal septenary universal forms, Manuscr. Math. 99, 181–184 (1999). * [22] B. M. Kim, Universal octonary diagonal forms over some real quadratic fields, Commentarii Math. Helv. 75, 410–414 (2000). * [23] J. Krásenský, M. Tinková and K. Zemková, There are no universal ternary quadratic forms over biquadratic fields, Proc. Edinb. Math. Soc. 63 (3), 861–912 (2020). * [24] F. Lemmermeyer and A. Pethö, Simplest Cubic Fields, Manuscripta Math. 88, 53–58 (1995). * [25] G. Lettl, A lower bound for the class number of certain cubic number fields, Math. Comp. 46, 659–666 (1986). * [26] S. Louboutin, Class-number problems for cubic number fields, Nagoya Math. J. 138, 199–208 (1995). * [27] H. Maaß, Über die Darstellung total positiver Zahlen des Körpers $R(\sqrt{5})$ als Summe von drei Quadraten, Abh. Math. Sem. Univ. Hamburg 14, 185–191 (1941). * [28] W. Narkiewicz, Elementary and analytic theory of algebraic numbers, 3rd Edition, Springer-Verlag, Berlin, 2004. * [29] O. Perron, Die Lehre von den Kettenbrüchen, B. G. Teubner, 1913. * [30] M. Peters, Summe von Quadraten in Zahlringen, J. Reine Angew. Math. 268/269, 318–323 (1974). * [31] A. Pfister,Quadratic forms with applications to algebraic geometry and topology, London Math. Soc. Lect. Notes 217, Cambridge University Press, 1995. * [32] A. Prestel, Remarks on the Pythagoras and Hasse number of real fields, J. Reine Angew. Math. 303/304, 284–294 (1978). * [33] H. Sasaki, Quaternary universal forms over $\mathbb{Q}[\sqrt{13}]$, Ramanujan J. 18, 73–80 (2009). * [34] R. Scharlau, On the Pythagoras number of orders in totally real number fields, J. Reine Angew. Math. 316, 208–210 (1980). * [35] D. Shanks, The simplest cubic number fields, Math. Comp. 28, 1137–1152 (1974). * [36] C. L. Siegel, Darstellung total positiver Zahlen durch Quadrate, Math. Z. 11, 246–275 (1921). * [37] C. L. Siegel, Sums of m-th powers of algebraic integers, Ann. of Math. 46, 313–339 (1945). * [38] L. Washington, Class numbers of the simplest cubic fields, Math. Comp. 48, 371–384 (1987). * [39] P. Yatsyna, A lower bound for the rank of a universal quadratic form with integer coeficients in a totally real field, Comment. Math. Helvet. 94, 221–239 (2019).
# Self-Calibrating Active Binocular Vision via Active Efficient Coding with Deep Autoencoders Charles Wilmot Frankfurt Institute for Advanced Studies Frankfurt, Germany Email<EMAIL_ADDRESS>Bertram E. Shi Hong Kong University of Science and Technology Clear Water Bay, Hong Kong Email<EMAIL_ADDRESS>Jochen Triesch Frankfurt Institute for Advanced Studies Frankfurt, Germany Email: <EMAIL_ADDRESS> ###### Abstract We present a model of the self-calibration of active binocular vision comprising the simultaneous learning of visual representations, vergence, and pursuit eye movements. The model follows the principle of Active Efficient Coding (AEC), a recent extension of the classic Efficient Coding Hypothesis to active perception. In contrast to previous AEC models, the present model uses deep autoencoders to learn sensory representations. We also propose a new formulation of the intrinsic motivation signal that guides the learning of behavior. We demonstrate the performance of the model in simulations. Keywords: active efficient coding, intrinsic motivation, binocular vision, vergence, pursuit, self-calibration, autonomous learning ## 1 Introduction Human vision is an active process and comprises a number of different types of eye movements. How the human visual system calibrates itself and learns the required sensory representations of the visual input signals is only poorly understood. A better understanding of this process might allow to build fully self-calibrating active vision systems that are robust to perturbations, e.g., [1], which could in turn form the basis for models describing the autonomous learning of object manipulation skills such as reaching or grasping, e.g., [2]. Here, we present a model of the self-calibration of active binocular vision that formulates the task as an intrinsically motivated reinforcement learning problem. In contrast to classic computer vision solutions to vergence control and object tracking, our model does not require pre-defined visual representations or kinematic models. Instead, it learns “from scratch” from raw images. Nevertheless, it achieves sub-pixel accuracy in its vergence and tracking movements demonstrating successful self-calibration. ### 1.1 Biological Background Humans and many other species have two forward facing eyes providing two largely overlapping views of the world. Initially, the visual signals are transformed to electrical signals in the retina. Different types of retinal ganglion cells are responsible for transmitting different kinds of information to the brain. In particular, the so-called magnocellular pathway conveys information at high temporal resolution required for motion vision [3], while the so-called parvo-cellular pathway has less temporal resolution but color sensitivity and higher spatial resolution [4]. From the retina, information is passed to the lateral geniculate nucleus, where information from both eyes is kept separate. Only in the primary visual cortex, the next processing stage, individual neurons receive information from both eyes. In particular, there are cells that detect small differences in local image structures at corresponding retinal locations in the left and right eye, so-called binocular disparities [5]. Furthermore, there are neurons that detect local image motion [6]. How the response properties of primary visual cortex cells develop has been the subject of a large body of research[7, 5, 8]. A widely accepted view is that these representations reflect an optimization of the visual system towards coding efficiency. ### 1.2 Efficient Coding in Perception In particular, inspired by information theory, Horace Barlow proposed a model of sensory coding postulating that neurons minimize the number of spikes needed for transmitting sensory information [9]. This would help to save energy, which is highly relevant, since the brain has high metabolic demands. It has been argued in [10], that retinal receptors can receive information at a rate of $10^{9}{\rm bit}/s$ [11], while the optic nerve can only transmit information at $10^{7}{\rm bit}/s$ [12]. This implies that the sensory information must be substantially compressed. Based on the idea of finding efficient codes for sensory signals, a large number of models have been proposed to explain the shapes of receptive fields in sensory brain areas for different modalities (vision, audition, olfaction, touch). More recently, it has been argued that adaptation of the organism’s behavior can also help to make the coding of sensory information more efficient. This theory is called Active Efficient Coding (AEC) [1, 13, 14]. It models the self-calibration of sensorimotor loops, where sensory input statistics shape the sensory representation, the sensory representation shapes the behavior and the behavior in turn shapes the input statistics. AEC has mostly been studied in the context of vision. There, it has been shown that AEC models can account for the self-calibration of active stereo vision, active motion-vision, and accommodation control or combinations thereof. These models have used only shallow neural network architectures to learn to encode the sensory signals. Here, we investigate potential benefits of deeper network architectures by utilizing deep autoencoders and formulate a new intrinsic reward signal to simultaneously learn the control of vergence and pursuit eye movements through reinforcement learning. Our results show that the model achieves sub-pixel accuracy in a simulated agent in a 3-D environment. ## 2 The Model Figure 1: Architecture of the model. Two regions are extracted at the center of the left / right camera images and encoded. The condensed representation is used to train the $Q$-function. The latter is trained to maximize the reward, which is proportional to the improvement of the reconstruction error of the encoder. ### 2.1 Sensory encoding via deep autoencoders When the two eyes verge on the same point, the foveal regions of the two retinal images become more and more similar. As a consequence, the mutual information between the left and right foveal image representations ${\rm MI}(I_{\rm L};I_{\rm R})$ for a given disparity $d$ increases as $d$ goes to $0$. It indicates how redundant the left and right images are and reflects the quality of the fixation. Similarly, tracking a moving object can be achieved by maximizing the information redundancy of the foveal image region across time by maximizing ${\rm MI}(I_{t};I_{t-1})$ (for one or more eyes). Here, we propose to measure the redundancy in the visual data via training an auto-encoder, hypothesizing that an information stream is better reconstructed when it is more redundant (see measurements in Figs. 2 and 3). The first advantage of this technique is that it is agnostic to the underlying data representation (for example the RGB channels in left and right data streams could be expressed in different bases or be non-linearly transformed, as only the data redundancy truly matters). Potentially, the algorithm could also exploit highly non-linear redundancies between different sensory modalities, given that the encoding network is sufficiently deep such that it captures these redundancies. The second advantage of this technique is that it learns a condensed representation of the sensory information which can be used by other learning components of the system. Such lossy compression of information may be essential for learning abstract representations at higher processing levels. We consider a binocular vision system with $3$ degrees of freedom: the pan and the tilt control conjugate horizontal and vertical movements of the gaze and the vergence controls inward and outward movement of the eyes with opposite sign across the two eyes. To properly fixate objects, the agent controlling the vergence must increase the redundancy in the left and right camera images, whereas the agents controlling the pan and tilt must increase the redundancy between consecutive images. We therefore used different inputs for the vergence, and for the pan and tilt agents. The visual sensory information for the vergence agent consists of the left and right images concatenated on the color dimension, whereas that of the pan and tilt joints consists of the concatenation of left and right images at time $t$ and $t-1$. The processing taking place on these inputs is the same for all agents. The different visual inputs for vergence vs. pan and tilt control reflect the distinction of two separate visual pathways discussed above: the magnocellular pathway with greater temporal resolution and lower spatial resolution compared to the parvocellular pathway. Let $v(t)$ denote one of the two visual sensory information streams (in the following we will drop the explicit time dependence for compactness of notation), and $E$ and $D$ be, respectively, the encoder and decoder parts of an auto-encoder, such that $\displaystyle s$ $\displaystyle=E\left(v;\theta_{E}\right)\;\text{and}$ (1) $\displaystyle\tilde{v}$ $\displaystyle=D\left(s;\theta_{D}\right)\;,$ (2) with $s$ representing the encoding and $\tilde{v}$ its reconstruction. We use the loss function $l=\frac{1}{3N_{\text{pixels}}}\sum_{ijk}(v_{ijk}-\tilde{v}_{ijk})^{2}$ for training the encoder and decoder weights $\theta_{E}$ and $\theta_{D}$, where $i$, $j$ and $k$ index the height, width, and color dimensions. Both eyes see the world from a slightly shifted perspective. The apparent shift of an object on the retinal image is called binocular disparity. In this paper we measure it in pixels. In order for the vergence control to work, the encoding / decoding component needs to learn a representation such that increasing binocular disparities induce a degradation of the reconstruction quality. While this is the case for all network architectures we tried, it is important to verify that the range of binocular disparities at which this is true matches the range of disparities at which we want the system to operate. For example, if we want the system to have a fixation accuracy better than $1$ pixel, we must check that the reconstruction error at $1$ pixel disparity is greater than the reconstruction error at $0$ pixel disparity. Similarly, if we want the agent to be capable of resolving disparities greater than $10$ pixels, we want the corresponding reconstruction error to be greater than for lower disparities. Since we want the model to operate over a wide range of disparities (and object velocities) we encode the visual input at different spatial scales, much like the retina samples the world at lower resolution towards the periphery. Details are given in Section 3. ### 2.2 Learning of the Behavior Component #### 2.2.1 Intrinsically motivated reinforcement learning formulation We consider the classical Markov decision process framework, where at discrete time $t=0,1,2,\ldots$ an agent observes sensory information $s_{t}=E\left(v_{t},\theta_{E}\right)$ and chooses action $a_{t}$ according to the distribution $\pi\left(a_{t}|s_{t}\right)$. After applying the action, the agent transitions to a new state according to a transition function $s_{t+1}=T(s_{t},a_{t})$, and receives a reward $r_{t}$. While reinforcement learning classically considers a reward provided by the agent’s environment through a potentially stochastic reward model, we here define an intrinsic reward based on the agent’s sensory encoding of its environment. Specifically, we define the reward $r^{\rm new}_{t}=C\left(l_{t}-l_{t+1}\right)\;,$ (3) where $C$ is a scaling factor. This reward signal measures the improvement of encoding quality, i.e., it favors movements that cause transitions from high to low reconstruction error of the autoencoder representing the visual input. We also compare the training speed obtained when training the agent with the simpler reward $r^{\rm old}_{t}=-Cl_{t+1}\;,$ (4) which has been used in previous AEC models and simply measures the quality of the encoding. The goal of reinforcement learning is to learn a policy function $\pi$ that maximizes the (discounted) sum of future rewards $R_{t}$ called return $R_{t}=\sum_{i=0}^{\infty}\gamma^{i}r_{t+i}$ (5) Where $\gamma$ is a discount factor in $\left[0,1\right]$ ensuring the convergence of the reward sum. In this particular application of reinforcement learning, the agent does not need to plan ahead his behaviour, consistent with our observation that the algorithm works best for $\gamma=0$. #### 2.2.2 RL-Algorithm We opted for an asynchronous version of the DQN algorithm [15, 16]. It consists of a $Q$-value function approximation $q=\begin{pmatrix}q_{0}\\\ \vdots\\\ q_{n}\end{pmatrix}=Q\left(s,\theta_{Q}\right)\;,$ (6) where $q_{j}$ represents an estimate of the return after performing discrete action $j$ in state $s$ and $n$ is the number of possible actions. The loss for training the $Q$-function is the Huber loss between the estimate and the return target [17]. Exploration during the training phase is performed via an $\epsilon$-greedy sampling. Table 1: Networks architecture Network | Architecture ---|--- Encoder | conv $96$ filters, size $8\times 8$ stride $4$ (vergence) | conv $24$ filters, size $1\times 1$ stride $1$ Decoder | conv $384$ filters, size $1\times 1$ stride $1$ (vergence) | Encoder | conv $192$ filters, size $8\times 8$ stride $4$ (pan, tilt) | conv $48$ filters, size $1\times 1$ stride $1$ Decoder | conv $768$ filters, size $1\times 1$ stride $1$ (pan, tilt) | Critic | conv $2\times 2$ stride $1$ | max-pooling $2\times 2$ stride $2$ | flatten | concatenate all scales | fully-connected $200$ | fully-connected $9$ Table 2: Parameters value Parameter | Value ---|--- Pan range | $\left[-10\degree,10\degree\right]$ Tilt range | $\left[-10\degree,10\degree\right]$ Vergence range | $\left[0\degree,20\degree\right]$ Discount factor $\gamma$ | $0$ Encoder learning rate | $5.10^{-4}$ Critic learning rate | $5.10^{-4}$ Episode length | $10$ iterations Buffer size | $1000$ transitions Batch size | $200$ transitions Epsilon $\epsilon$ | $0.05$ Reward scaling factor $C$ | $600$ ## 3 Experimental setup We conducted our experiments using the robot simulator CoppeliaSim (previously named V-REP) using the python API PyRep [18], within which a robot head composed of two cameras separated by $6.8$ cm was simulated. Each camera has a resolution of $240\times 320$ $\mathrm{p}\mathrm{x}$ (height $\times$ width) for a horizontal field of view of $90\degree$ (therefore $1$ pixel $=0.28\degree$). To make the results easier to interpret, we expressed all angles and angular velocities in $\mathrm{p}\mathrm{x}$ and $\mathrm{p}\mathrm{x}\mathrm{/}\mathrm{i}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}$, respectively. In the simulated environment, a screen was moving at uniform speeds varying from $0$ to $4$ $\mathrm{p}\mathrm{x}\mathrm{/}\mathrm{i}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}$ in front of the robot head, at distances between $0.5$ and $5$ meters. The screen displayed natural stimuli taken from the McGill Calibrated Color Image Database [19]. The two spatial scales of visual processing of the cameras are realized by extracting two centered $32\times 32$ pixel regions per camera (cf. Figure 1), respectively covering a field of view of $9\degree$ (fine scale) and $27\degree$ (coarse scale). The auto-encoder for each scale corresponds to a $3$-layered fully-connected network encoding patches of size $8\times 8$ pixels. This patch-wise autoencoder is implemented as a convolutional neural network with filter size $8\times 8$ in the first layer and $1\times 1$ in the following layers (see Tab. 1). Figure 3 shows the reconstruction error as a function of the binocular disparity for each scale after learning. The critic network $Q$ can be described in $2$ parts. The first part operates individually on each scale. It is composed of a convolutional layer followed by a pooling layer. The results are then flattened and concatenated before being processed by $2$ fully-connected layers in the second part (cf. Tab. 1). All networks are trained using the Adam algorithm [20] with a learning rate of $5.10^{-4}$. We use a value of $\epsilon=\;$0.05 for the $\epsilon$-greedy sampling. The training is divided into episodes of $10$ iterations. Each time an episode is simulated, all its transitions are placed in a replay buffer of size 1000 and a batch of data is then sampled uniformly at random from the buffer for training the networks. We use a batch size of $200$. The training is spread over multiple processes, each simulating one agent. Each process asynchronously pulls the current weight values from a server, uses them to compute the weight updates $\Delta\theta$, and sends them to the server, which is responsible for performing the updates. The robot has $3$ joints available to control the eyes. All joints used the same action discretization. However, the vergence joint operates in velocity control mode, while the pan and tilt actions are interpreted as accelerations. The action set we used for all joints is the following: $\left[-4,-2,-1,-\frac{1}{2},0,\frac{1}{2},1,2,4\right]$ $\mathrm{p}\mathrm{x}\mathrm{/}\mathrm{i}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}$ (vergence), or $\mathrm{p}\mathrm{x}\mathrm{/}\mathrm{i}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}^{2}$ (pan and tilt). The vergence angle of the robot’s eyes is constrained between $0\degree$ (parallel optical axes of the two eyes) and $20\degree$ (inward rotated eyes). The pan and tilt joints are constrained to remain in $\left[-10\degree,10\degree\right]$. At regular intervals, the training is paused, and the agents’ performances are measured. For evaluating the agents’ performances, we gather two sets of data at each testing step. One, the controlled-error set, gauges the performance of the agents under defined apparent disparities. The other, the behaviour set, measures how the policies of the agents recover from initial disparities. All measurements are repeated for $20$ stimuli displayed on the screen $2$ $\mathrm{m}$ away from the eyes of the robot. To construct the controlled- error set, we simulated various pan, tilt and vergence errors by manually setting the speed of the screen and the vergence angle of the eyes. Only one joint was tested at a time, meaning that the errors for the two others were set to $0$. We then recorded the reconstruction errors of the fine and coarse scales and the agents’ preferred actions. The behaviour set is the recording of $20$ iterations of the agents’ behaviour, starting from controlled initial pan, tilt, or vergence errors. ## 4 Results For successful learning, the pan, tilt and vergence errors must become reflected in the reconstruction errors of the encoders, as explained in Section 2.1. We start by analyzing the reconstruction errors of the encoders for every pan, tilt, or vergence error at the end of training with reward function $r^{\rm new}$. Figures 2 and 3 show that for each joint, the reconstruction error is minimal when the absolute joint error is minimal. In particular, Figure 2 shows the mean reconstruction error for each stimulus displayed on the screen (in blue) and the average for all stimuli (in red), while Figure 3 shows the mean reconstruction error for each scale separately. Repeating the same analysis using random weights for the encoder and decoder shows no difference in the reconstruction quality for low or high absolute errors (the mean error curve is flat instead of being V-shaped, with values around $0.27$, not shown). The characteristic V-shaped curves are a consequence of both learning a compact code of the visual input and adapting the behavior to shape the statistics of the visual input [21]. Figure 2: The autoencoders’ reconstruction errors averaged across the two scales as a function of the pan, tilt, and vergence error. Each blue curve corresponds to a different stimulus displayed on the screen. The red curve represents the mean. For each plot, the error for the two other joints is set to $0$. Figure 3: The autoencoders’ mean reconstruction errors plotted separately for the two scales as a function of the pan, tilt, and vergence error. For the generation of each plot, the error for the two other joints has been set to $0$. To show the precision of the learnt policies, we represent for each joint the probability of selecting each possible action in the action set as a function of that joint’s error in Figure 4. The diagonal shapes in the three policies indicate that the model has learned to accurately compensate for any vergence, pan, or tilt errors. Figure 4: Probability of choosing each action in the action sets as a function of the pan, tilt, and vergence error. For the generation of each plot, the error for the two other joints has been set to $0$. To show the speed at which the algorithm converges and compare the two reward functions $r^{\rm new}$ and $r^{\rm old}$, Figure 5 plots the average training error at the end of episodes (i.e. the joints’ absolute errors while following the $\epsilon$-greedy policies) as a function of training time for both rewards. The testing error is measured at regular intervals as the mean absolute joint error after $10$ iterations of following the greedy policy, starting from initial errors of $-4$, $-2$, $2$ and $4$ $\mathrm{p}\mathrm{x}$ (vergence) or $\mathrm{p}\mathrm{x}\mathrm{/}\mathrm{i}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}$ (pan and tilt). From the plots it is evident that the reward $r^{\rm new}$ measuring the improvement in encoding quality (Eq. 3), leads to faster convergence. Focusing on the improvement of the encoding quality helps the system to deal with very different levels of absolute difficulty for encoding different stimuli (cf. Fig. 2). Figure 5: Reduction of errors with training time for the $2$ different rewards. The blue curves correspond to the “improvement” reward $r^{\rm new}$ (Eq. 3) and the red curves to $r^{\rm old}$ (Eq. 4). The light curves show the pan, tilt, and vergence errors at the last iteration of an episode as a function of training time when following the $\epsilon$-greedy policies. The data is averaged over 5 independent runs and smoothed for clarity. The dark curves indicate the performance of the greedy policy after $10$ iterations, starting from initial (speed) errors in $\left[-4,4\right]$ $\mathrm{p}\mathrm{x}$ (vergence) or $\mathrm{p}\mathrm{x}\mathrm{/}\mathrm{i}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}$ (pan and tilt), with a step size of half a pixel. Vertical bars indicate the standard deviation across $5$ runs. The testing performance of the new reward is consistently below $1$ $\mathrm{p}\mathrm{x}$ (vergence) or $\mathrm{p}\mathrm{x}\mathrm{/}\mathrm{i}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}$ (pan and tilt) after $10,000$ episodes of training demonstrating sub-pixel accuracy. Finally, to show how quickly and accurately the algorithm fixates objects and tracks them, Figure 6 shows the mean accuracy and its standard deviation, during 20 consecutive iterations, starting from all errors in $\left[-4,4\right]$ $\mathrm{p}\mathrm{x}$ (vergence) or $\mathrm{p}\mathrm{x}\mathrm{/}\mathrm{i}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}$ (pan and tilt) with a step size of half a pixel. Subpixel error levels are typically reached in just one or two iterations despite the discrete action set. Note that an error of, e.g., 3 pixels cannot be reduced to zero in a single step because the two closest discrete actions of 2 pixels and 4 pixels would both lead to a remaining error of 1 pixel. Figure 6: Rapid object fixation and tracking. Pan speed error, tilt speed error, and vergence error are decreasing quickly during one episode and typically reach subpixel levels in one or two steps. The shaded region indicates one standard deviation. ## 5 Discussion Understanding the development of the human mind and replicating this development in artificial cognitive agents is a grand challenge for 21st century science. Here, we focused on a very early step in this development, which lays the foundation for most of what follows: the development of early visual representations and the ability to self-calibrate accurate eye movements. From learning to manipulate objects to interacting with social partners, vision is a key sensory modality. In this work, we focused on active stereo and motion vision, proposing a model for their completely autonomous self-calibration. Our work falls inside but also extends the recently proposed Active Efficient Coding (AEC) framework, which is itself an extension of Barlow’s classic efficient coding hypothesis[9] to active perception and therefore rooted in Shannon’s information theory. AEC postulates that visual representations and eye movements are jointly optimized to maximize the efficiency of the visual system to encode information. Along these lines, previous models have shown how AEC can explain the development of active stereo vision [21, 22], active motion vision[23, 13], as well as the control of torsional eye movements [24] and accommodation [14] and various combinations thereof, e.g., [25, 26]. The two key innovations of the present work are to consider “deeper” sensory representations compared to the shallow sparse coding approaches used earlier and to use a new intrinsic reward formulation. Regarding the first innovation, we have employed deep convolutional autoencoders for the sensory encoding stage and a Deep Q-Network (DQN) [15, 16] to map the learned representations onto behavior and obtained very good results. The model quickly achieves sub-pixel accuracy in all degrees of freedom. Regarding the second innovation, we have shown that an intrinsic reward for improvements in encoding quality leads to faster convergence. From the perspective of biological plausibility, this success comes at a price, however. The learning algorithms used to train the auto-encoder and the DQN rely on error back-propagation mechanisms, which are thought to be not biologically plausible by most researchers in the field. While this should be considered a weakness when evaluating this work as a model of biological mechanisms, it is not problematic from a robotics application perspective. Another limitation of the approach is that accurate performance requires that the object to be fixated and tracked must be sufficiently big. If the object fills only a fraction of the two regions defining the two spatial scales, then multiple disparities and velocities are present within these regions, because the rest is filled by background. In this case, the system will be “confused” and may decide to fixate and stabilize the background. To deal with this problem, a mechanism for foreground/background separation needs to be introduced. Our work follows the traditional structure of AEC models using a separation into two distinct learning modules — the first being the deep auto-encoders for unsupervised learning of a sensory representation and generation of reward signals and the second being the DQN for learning behavior via reinforcement learning. It should be questioned however, if this “hard” separation is strictly necessary. An alternative architecture might try to blend these functions into a single more homogeneous network. This topic is left for future work. Another interesting direction for future work is to consider other sensory modalities. Ongoing work (unpublished) is revealing that AEC can also be used to model the self-calibration of echolocation in bats in the auditory modality. More generally, the combination of different sensory modalities is an interesting topic for future research. Arguably, a key step in cognitive development is discovering and understanding the relationships between sensory signals from different modalities in order to arrive at a unified representation of the world. ## Funding This project has received funding from the European Union’s Horizon 2020 research and innovation program under grant 713010. J.Triesch is supported by the Johanna Quandt Foundation. ## References * [1] L. Lonini, S. Forestier, C. Teulière, Y. Zhao, B. Shi, and J. Triesch, “Robust active binocular vision through intrinsically motivated learning,” _Frontiers in Neurorobotics_ , vol. 7, p. 20, 2013. [Online]. Available: https://www.frontiersin.org/article/10.3389/fnbot.2013.00020 * [2] F. D. L. Bourdonnaye, C. Teulière, J. Triesch, and T. Chateau, “Stage-wise learning of reaching using little prior knowledge,” _Front. Robotics and AI_ , vol. 2018, 2018. * [3] P. J. Jeffries AM, Killian NJ, “Mapping the primate lateral geniculate nucleus: a review of experiments and methods.” _J Physiol Paris_ , 2014. * [4] X. Xu, J. M. Ichida, J. D. Allison, J. D. Boyd, A. Bonds, and V. A. Casagrande, “A comparison of koniocellular, magnocellular and parvocellular receptive field properties in the lateral geniculate nucleus of the owl monkey (aotus trivirgatus),” _The Journal of physiology_ , vol. 531, no. 1, pp. 203–218, 2001. * [5] N. Qian, “Binocular disparity and the perception of depth,” _Neuron_ , vol. 18, no. 3, pp. 359–368, 1997. * [6] A. Borst and M. Egelhaaf, “Principles of visual motion detection,” _Trends in neurosciences_ , vol. 12, no. 8, pp. 297–306, 1989. * [7] J. Y. Lettvin, H. R. Maturana, W. S. McCulloch, and W. H. Pitts, “What the frog’s eye tells the frog’s brain,” _Proceedings of the IRE_ , vol. 47, no. 11, pp. 1940–1951, 1959. * [8] B. Scholl, J. Burge, and N. J. Priebe, “Binocular integration and disparity selectivity in mouse primary visual cortex,” _Journal of neurophysiology_ , vol. 109, no. 12, pp. 3013–3024, 2013. * [9] H. Barlow, “Possible principles underlying the transformations of sensory messages,” _Sensory Communication_ , vol. 1, 01 1961. * [10] L. Zhaoping, “Theoretical understanding of the early visual processes by data compression and data selection,” _Network (Bristol, England)_ , vol. 17, pp. 301–34, 01 2007. * [11] D. H. Kelly, “Information capacity of a single retinal channel,” _IRE Trans. Information Theory_ , vol. 8, pp. 221–226, 1962. * [12] S. Nirenberg, S. M. Carcieri, A. L. Jacobs, and P. E. Latham, “Retinal ganglion cells act largely as independent encoders,” _Nature_ , vol. 411, no. 6838, pp. 698–701, 2001. [Online]. Available: https://doi.org/10.1038/35079612 * [13] T. N. Vikram, C. Teulière, C. Zhang, B. E. Shi, and J. Triesch, “Autonomous learning of smooth pursuit and vergence through active efficient coding,” in _4th International Conference on Development and Learning and on Epigenetic Robotics_ , 2014, pp. 448–453. * [14] S. Eckmann, L. Klimmasch, B. E. Shi, and J. Triesch, “Active efficient coding explains the development of binocular vision and its failure in amblyopia,” _Proceedings of the National Academy of Sciences_ , vol. 117, no. 11, pp. 6156–6162, 2020. [Online]. Available: https://www.pnas.org/content/117/11/6156 * [15] V. Mnih, K. Kavukcuoglu, D. Silver, A. Graves, I. Antonoglou, D. Wierstra, and M. Riedmiller, “Playing atari with deep reinforcement learning,” _arXiv preprint arXiv:1312.5602_ , 2013. * [16] V. Mnih, A. P. Badia, M. Mirza, A. Graves, T. Lillicrap, T. Harley, D. Silver, and K. Kavukcuoglu, “Asynchronous methods for deep reinforcement learning,” in _International conference on machine learning_ , 2016, pp. 1928–1937. * [17] P. J. Huber, “Robust estimation of a location parameter,” _Ann. Math. Statist._ , vol. 35, no. 1, pp. 73–101, 03 1964. [Online]. Available: https://doi.org/10.1214/aoms/1177703732 * [18] S. James, M. Freese, and A. J. Davison, “Pyrep: Bringing v-rep to deep robot learning,” _arXiv preprint arXiv:1906.11176_ , 2019. * [19] A. Olmos and F. A. A. Kingdom, “A biologically inspired algorithm for the recovery of shading and reflectance images,” _Perception_ , vol. 33, no. 12, pp. 1463–1473, 2004, pMID: 15729913. [Online]. Available: https://doi.org/10.1068/p5321 * [20] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” _arXiv preprint arXiv:1412.6980_ , 2014. * [21] Y. Zhao, C. A. Rothkopf, J. Triesch, and B. E. Shi, “A unified model of the joint development of disparity selectivity and vergence control,” in _2012 IEEE International Conference on Development and Learning and Epigenetic Robotics (ICDL)_. IEEE, 2012, pp. 1–6. * [22] L. Lonini, Y. Zhao, P. Chandrashekhariah, B. E. Shi, and J. Triesch, “Autonomous learning of active multi-scale binocular vision,” in _2013 IEEE Third Joint International Conference on Development and Learning and Epigenetic Robotics (ICDL)_ , 2013, pp. 1–6. * [23] Q. Zhu, J. Triesch, and B. E. Shi, “Autonomous, self-calibrating binocular vision based on learned attention and active efficient coding,” _2017 Joint IEEE International Conference on Development and Learning and Epigenetic Robotics (ICDL-EpiRob)_ , pp. 27–32, 2017. * [24] Q. Zhu, C. Zhang, J. Triesch, and B. E. Shi, “Autonomous learning of cyclovergence control based on active efficient coding,” in _2018 Joint IEEE 8th International Conference on Development and Learning and Epigenetic Robotics, ICDL-EpiRob 2018, Tokyo, Japan, September 17-20, 2018_. IEEE, 2018, pp. 251–256. [Online]. Available: https://doi.org/10.1109/DEVLRN.2018.8761033 * [25] A. Priamikov, V. Narayan, B. E. Shi, and J. Triesch, “The role of contrast sensitivity in the development of binocular vision: A computational study,” in _2015 Joint IEEE International Conference on Development and Learning and Epigenetic Robotics (ICDL-EpiRob)_ , 2015, pp. 33–38. * [26] A. Lelais, J. Mahn, V. Narayan, C. Zhang, B. E. Shi, and J. Triesch, “Autonomous development of active binocular and motion vision through active efficient coding,” _Frontiers in neurorobotics_ , vol. 13, p. 49, 2019.
# Effect of new jet substructure measurements on Pythia8 tunes Deepak Kar Pratixan Sarmah School of Physics, University of Witwatersrand Johannesburg, South Africa. Email<EMAIL_ADDRESS>Department of Physics, BITS Pilani Rajasthan, India. Email<EMAIL_ADDRESS> ###### Abstract This masters project used the recent ATLAS jet substructure measurements to see if any improvements can be made to the commonly used Pythia8 Monash and A14 tunes. ###### keywords: Pythia8, jet substructure, FSR, tune ## 1 Introduction The commonly used Pythia8 [1, 2] tunes, Monash [3] and A14 [4] are rather dated, and the latter was observed to have some tension with LEP measurements, primarily due to its lower Final State Radiation (FSR) $\alpha_{s}$ value. In last couple of years, a plethora of jet substructure [5, 6, 7, 8] measurements have been published by both ATLAS and CMS collaborations, utilising LHC Run 2 data. Here, we investigate the effect of four such ATLAS measurements on parameters sensitive to jet substructure observables. ## 2 Tuning setup The following ATLAS measurements were considered in this study (along with their Rivet identifiers): * 1. Soft-Drop Jet Mass [9](ATLAS_2017_I1637587) * 2. Jet substructure measurements in multijet events [10] (ATLAS_2019_I1724098) * 3. Soft-drop observables [11](ATLAS_2019_I1772062) * 4. Lund jet plane with charged particles [12] (ATLAS_2020_I1790256) The following parameters were considered in this tuning exercise, with the ranges stated in Table 1. Parameter | Lower value | Upper value ---|---|--- BeamRemnants:primordialKThard | 1.25 | 3 ColorReconncetion:range | 1.25 | 3 TimeShower:pTmin | 0.5 | 1.5 MultipartonInteractions:pT0Ref | 1.5 | 3 TimeShower:alphaSvalue | 0.118 | 0.145 Table 1: Sampling range of the parameters considered Weighted hardQCD events were generated with a PThatMin of 300 GeV. 100 Sampling runs were performed, each with 100000 events. Rivet3 [13] and Professor tuning system [14] were used. The goodness of sampling and the weight files used can be found in Appenix 5.2 and in Appenix 5.3. ## 3 Results The first step was to ascertain where we have a scope of improvement. While a detailed observable-by-observable determination was performed (see Appendix 5.1), here we highlight the most salient features: * 1. For Lund Jet Plane (LJP) distributions, we observed that the hard-wide angle emissions part is better modelled by the Monash tune whereas the region ranging from UE/MPI to Soft-collinear and Collinear limits are in general better modelled by the A14 tune. However, this distributions also offer the biggest scope of improved modelling. * 2. For the soft drop $\rho$ and $r_{g}$ observables, in general Monash tune performs somewhat better than A14. One deviation from this trend is when the jet construction is Cluster based, in which case the A14 tune performs better over a large range. * 3. Both the Jet Substructure and Soft drop jet mass distributions are somewhat better modelled by the A14 tune. Table 2 lists the parameter values of A14 and Monash along with our tuned values. A separate tune for LJP was performed as this analysis had the largest discrepancy. The LJP tune column shows the parameter values corresponding to the best tune for LJP and the Common Tune column shows the values of the best tune for all the analyses considered. Figures 1 and 2 show the tuned distributions for the one dimensional vertical slices of the LJP. Figure 3 shows the tuned distributions for the soft drop observables. Figure 4 shows the tuned distributions for soft drop mass. And lastly, Figure 5 shows the tuned distributions for the jet substructure observables. Parameters | A14 | Monash | LJP Tune | Common Tune ---|---|---|---|--- BeamRemnants:primordialKThard | 1.88 | 1.8 | 2.288 | 2.065 ColorReconnection:range | 1.71 | 1.8 | 2.73 | 1.69 TimeShower:pTmin | 0.40 | 0.50 | 1.288 | 0.775 MultipartonInteractions:pT0Ref | 2.09 | 2.28 | 2.766 | 2.91 TimeShower:alphaSvalue | 0.127 | 0.1365 | 0.1308 | 0.1309 Table 2: Comparison of tuned values with Monash and A14 Figure 1: Comparison of our tunes with A14 and Monash tunes for Lund Jet Plane distributions (vertical slices) Figure 2: Comparison of our tunes with A14 and Monash tunes for Lund Jet Plane distributions (horizontal slices) Figure 3: Comparison of our tunes with A14 and Monash tunes for soft drop jet mass distributions Figure 4: Comparison of our tunes with A14 and Monash tunes for soft drop jet mass distributions Figure 5: Comparison of our tunes with A14 and Monash tunes for jet Sub structure observable distribution for dijet selection ## 4 Summary The results obtained show small improvements of roughly 5-10% in the distributions of the Lund Jet Plane and Soft Drop Mass distributions from the previous A14 and Monash Tunes. As in Table 2, it can be seen that the parameter values of the tunes obtained are pulled up from the A14 and Monash Tunes. In the case of the LJP , we see that the A14 and Monash Tunes deviate most from the data near the peaks of the distributions. This is the region where soft collinear effects transitions to UE/MPI effects in the LJP. Since the tunes we obtained improve this region of the distributions, it can be inferred that higher values of these parameters facilitate more soft radiations in the final state. In the case of the soft drop observable distributions, there are regions that require generation of more mass to model the data better. These compete with the LJP values and decreases values of parameters: BeamRemnants:primordialKThard from 2.288 to 2.065, ColorReconnection:range from 2.73 to 1.69, TimeShower:pTmin from 1.288 to 0.775. For the other two parameters, MPI:pT0Ref and TimeShower:alphaSvalue, the values increased slightly. ## Acknowledgements DK is funded by National Research Foundation (NRF), South Africa through Competitive Programme for Rated Researchers (CPRR), Grant No: 118515. We thank Andy Buckley and Holger Schulz for technical assistance with Professor program, as well as for physics discussions. ## References * [1] T. Sjostrand, S. Mrenna, P. Z. Skands, A Brief Introduction to PYTHIA 8.1, Comput. Phys. Commun. 178 (2008) 852–867. arXiv:0710.3820, doi:10.1016/j.cpc.2008.01.036. * [2] T. Sjöstrand, S. Ask, J. R. Christiansen, R. Corke, N. Desai, P. Ilten, S. Mrenna, S. Prestel, C. O. Rasmussen, P. Z. Skands, An Introduction to PYTHIA 8.2, Comput. Phys. Commun. 191 (2015) 159–177. arXiv:1410.3012, doi:10.1016/j.cpc.2015.01.024. * [3] P. Skands, S. Carrazza, J. Rojo, Tuning PYTHIA 8.1: the Monash 2013 Tune, Eur. Phys. J. C74 (8) (2014) 3024. arXiv:1404.5630, doi:10.1140/epjc/s10052-014-3024-y. * [4] ATLAS Collaboration, ATLAS Pythia 8 tunes to $7\;\mbox{TeV}$ data, ATL-PHYS-PUB-2014-021 (2014). URL https://cds.cern.ch/record/1966419 * [5] A. Altheimer, et al., Jet Substructure at the Tevatron and LHC: New results, new tools, new benchmarks, J. Phys. G39 (2012) 063001. arXiv:1201.0008, doi:10.1088/0954-3899/39/6/063001. * [6] A. Altheimer, et al., Boosted objects and jet substructure at the LHC. Report of BOOST2012, held at IFIC Valencia, 23rd-27th of July 2012, Eur. Phys. J. C74 (3) (2014) 2792. arXiv:1311.2708, doi:10.1140/epjc/s10052-014-2792-8. * [7] S. Marzani, G. Soyez, M. Spannowsky, Looking inside jets: an introduction to jet substructure and boosted-object phenomenology, Vol. 958, Springer, 2019. arXiv:1901.10342, doi:10.1007/978-3-030-15709-8. * [8] R. Kogler, et al., Jet Substructure at the Large Hadron Collider: Experimental Review, Rev. Mod. Phys. 91 (4) (2019) 045003. arXiv:1803.06991, doi:10.1103/RevModPhys.91.045003. * [9] ATLAS Collaboration, Measurement of the Soft-Drop Jet Mass in pp Collisions at $\sqrt{s}=13$ TeV with the ATLAS Detector, Phys. Rev. Lett. 121 (9) (2018) 092001. arXiv:1711.08341, doi:10.1103/PhysRevLett.121.092001. * [10] ATLAS Collaboration, Measurement of jet-substructure observables in top quark, $W$ boson and light jet production in proton-proton collisions at $\sqrt{s}=13$ TeV with the ATLAS detector, JHEP 08 (2019) 033. arXiv:1903.02942, doi:10.1007/JHEP08(2019)033. * [11] ATLAS Collaboration, Measurement of soft-drop jet observables in $pp$ collisions with the ATLAS detector at $\sqrt{s}$ =13 TeV, Phys. Rev. D 101 (5) (2020) 052007. arXiv:1912.09837, doi:10.1103/PhysRevD.101.052007. * [12] ATLAS Collaboration, Measurement of the Lund Jet Plane Using Charged Particles in 13 TeV Proton-Proton Collisions with the ATLAS Detector, Phys. Rev. Lett. 124 (22) (2020) 222002. arXiv:2004.03540, doi:10.1103/PhysRevLett.124.222002. * [13] C. Bierlich, et al., Robust Independent Validation of Experiment and Theory: Rivet version 3, SciPost Phys. 8 (2020) 026. arXiv:1912.05451, doi:10.21468/SciPostPhys.8.2.026. * [14] A. Buckley, H. Hoeth, H. Lacker, H. Schulz, J. E. von Seggern, Systematic event generator tuning for the LHC, Eur. Phys. J. C 65 (2010) 331–357. arXiv:0907.2973, doi:10.1140/epjc/s10052-009-1196-7. ## 5 Appendix ### 5.1 Tune performance Plots | Observable | $\ln(R/\Delta R)$ or | Better Tune Performance regions | Better Tune ---|---|---|---|--- | | $\ln(1/z)$ Slice | A14 | Monash | (overall) d03-x01-y01 | $\ln(1/z)$ | 0.00-0.33 | 4.3-6 | 0.7-4.3 | Monash d04-x01-y01 | $\ln(1/z)$ | 0.33-0.67 | 3.4-4.2 | 0.7-3.4 , 4.4-6 | Monash d05-x01-y01 | $\ln(1/z)$ | 0.67-1.00 | 3-5 | 0.7-3 , 5-6 | Monash d06-x01-y01 | ln(1/z) | 1.00-1.33 | 3-4.2 | 0.7-3 , 4.2-6 | Monash d07-x01-y01 | $\ln(1/z)$ | 1.33-1.67 | 2.4-4.2 | 0.7-2.4 , 4.3-6 | Monash d08-x01-y01 | $\ln(1/z)$ | 1.67-2.00 | 2-4 , 5.2 , 5.8 | 0.7-2 , 4-5 , 5.5 | - d09-x01-y01 | $\ln(1/z)$ | 2.00-2.33 | 1.8-4 , 4.7, 5.2-6 | 0.7-1.8, 4-4.4 | A14 d10-x01-y01 | $\ln(1/z)$ | 2.33-2.67 | 1.6-3, 4.4-5.8 | 0.7-1.4, 3.2-4.2 | A14 d11-x01-y01 | $\ln(1/z)$ | 2.67-3.00 | 1.4-5 | 0.7-1.2, 5.2-5.8 | A14 d12-x01-y01 | $\ln(1/z)$ | 3.00-3.33 | 0.8-1.4, 3-4, 4.7, 5.8 | 1.6-3, 3.3, 4-4.5, 5-5.6 | - d13-x01-y01 | $\ln(1/z)$ | 3.33-3.67 | 0.7-3.4, 4.6-5 | 3.6-4.4, 5.1-6 | A14 d14-x01-y01 | $\ln(1/z)$ | 3.67-4.00 | 0-3.4 | 3.6-4.6 | A14 d15-x01-y01 | $\ln(1/z)$ | 4.00-4.33 | 0.7-2.6, 3.6, 4-4.6, 5, 5.57 | 2.7-3.4, 3.9, 4.7, 5.23 | A14 d16-x01-y01 | $\ln(R/\Delta R)$ | 0.69-0.97 | 3-4.5 | 0-3 | Monash d17-x01-y01 | $\ln(R/\Delta R)$ | 0.97-1.25 | 3-4 | 0-3 | Monash d18-x01-y01 | $\ln(R/\Delta R)$ | 1.25-1.52 | 2.7-3.2, 3.5-4 | 0-2.6 | Monash d19-x01-y01 | $\ln(R/\Delta R)$ | 1.52-1.80 | 2.4-4.5 | 0.5-2.3 | - d20-x01-y01 | $\ln(R/\Delta R)$ | 1.80-2.08 | 2-3, 3.3-4.5 | 0-2 | - d21-x01-y01 | $\ln(R/\Delta R)$ | 2.08-2.36 | 1.6-4.5 | 0-1.6, 2.2, 3.4-4 | - d22-x01-y01 | $\ln(R/\Delta R)$ | 2.36-2.63 | 1.4-3, 3.6-4.5 | 0-1.3, 3-3.6 | A14 d23-x01-y01 | $\ln(R/\Delta R)$ | 2.63-2.91 | 1.4-3, 3.5 | 0-1.3, 3-4.5 | - d24-x01-y01 | $\ln(R/\Delta R)$ | 2.91-3.19 | 0.6-4.5 | 0-0.5 | A14 d25-x01-y01 | $\ln(R/\Delta R)$ | 3.19-3.47 | 0.6-2.4, 3.4-4.5 | 0-0.5, 2.4-3.3 | A14 d26-x01-y01 | $\ln(R/\Delta R)$ | 3.47-3.74 | 0.5-4.5 | 0.2, 1.2, 2.5, 3.5 | A14 d27-x01-y01 | $\ln(R/\Delta R)$ | 3.74-4.02 | 0.3-4.5 | 0.2 | A14 d28-x01-y01 | $\ln(R/\Delta R)$ | 4.02-4.30 | 0-1.6, 3.7-4.5 | 1.7-3.6 | - d29-x01-y01 | $\ln(R/\Delta R)$ | 4.30-4.57 | 0.2, 0.5,1.2, 2.3-4.5, | 0.8, 1-2.2, 3.2, 3.8 | - d30-x01-y01 | $\ln(R/\Delta R)$ | 4.57-4.85 | 0.2, 1.6-3.3, 3.5 | 0.3-1.5, 3.7-4.5 | - d31-x01-y01 | $\ln(R/\Delta R)$ | 4.85-5.13 | 0-3, 3.4-4.5 | 3.2 | A14 d32-x01-y01 | $\ln(R/\Delta R)$ | 5.13-5.41 | 0.2, 1.7-4 | 0.4-1.7, 2.75 | A14 d33-x01-y01 | $\ln(R/\Delta R)$ | 5.41-5.68 | 0.2, 2-3 | 0.5-2, 3.2-4 | Monash d34-x01-y01 | $\ln(R/\Delta R)$ | 5.68-5.96 | 1.7-2.3 | 0-1.7, 2.5-3 | Monash Table 3: ATLAS_2020_I1790256(LJP) Plots | | $\beta$ | $z_{cut}$ | Observable | A14 | Monash | Better Tune ---|---|---|---|---|---|---|--- d01-x01-y01 | Calorimeter based | 0 | 0.1 | $\rho$ | - | all | Monash d02-x01-y01 | Track based | 0 | 0.1 | $\rho$ | - | all | Monash d03-x01-y01 | Cluster based | 1 | 0.1 | $\rho$ | [-4.5,-3.7], | [-3.5,-2.1], | A14/ | | | | | [-2,-1.3] | [-1,-0.5] | Monash d04-x01-y01 | Track based | 1 | 0.1 | $\rho$ | [-4.5,-3.7] | [-3.5,-0.5] | Monash d05-x01-y01 | Cluster based | 2 | 0.1 | $\rho$ | [-4.5,-1.1] | -0.7 | A14 d06-x01-y01 | Track based | 2 | 0.1 | $\rho$ | [-4.5,-3.7] | [-3.5,-0.5] | Monash d07-x01-y01 | Track based | 1 | 0.1 | $\rho$ | [-4.5,-3.7] | [-3.5,-0.5] | Monash d16-x01-y01 | Track based | 1 | 0.1 | $r_{g}$ | - | all | Monash d17-x01-y01 | Cluster based | 2 | 0.1 | $r_{g}$ | [-1.2,-0.2] | -0.15 | A14 d18-x01-y01 | Track based | 2 | 0.1 | $r_{g}$ | -1.1 | [-1,-0.1] | Monash d19-x01-y01 | Central jet/Calorimeter | 0 | 0.1 | $r_{g}$ | - | all | Monash d20-x01-y01 | Central jet/Track | 0 | 0.1 | $r_{g}$ | - | all | Monash d21-x01-y01 | Central jet/Cluster | 1 | 0.1 | $\rho$ | [-4.5,-1] | -0.7 | A14 d22-x01-y01 | Central jet/Track | 1 | 0.1 | $\rho$ | [-4.5,-3.7] | [-3.5,-0.5] | Monash d23-x01-y01 | Central jet/Cluster | 2 | 0.1 | $\rho$ | [-3.5,-0.9] | -0.7 | A14 d24-x01-y01 | Central jet/Track | 2 | 0.1 | $\rho$ | [-4.5,-3.7] | [-3.5,-0.7] | Monash d34-x01-y01 | Central jet/Track | 1 | 0.1 | $r_{g}$ | - | all | Monash d35-x01-y01 | Central jet/Cluster | 2 | 0.1 | $r_{g}$ | [-1.2,-0.4] | -0.5,-0.15 | A14 d36-x01-y01 | Central jet/Track | 2 | 0.1 | $r_{g}$ | -1.1 | [-1,-0.1] | Monash d37-x01-y01 | Forward jet/Calorimeter | 0 | 0.1 | $r_{g}$ | - | all | Monash d38-x01-y01 | Forward jet/Track | 0 | 0.1 | $\rho$ | - | all | Monash d39-x01-y01 | Forward jet/Cluster | 1 | 0.1 | $\rho$ | -4.3 | [-4,-0.5] | Monash d40-x01-y01 | Forward jet/Track | 1 | 0.1 | $\rho$ | [-4.5,-3.7] | [-3.5,-0.5] | Monash d41-x01-y01 | Forward jet/Cluster | 2 | 0.1 | $\rho$ | -3.9,[-3.1,-1] | -3.5,-0.7 | A14 d42-x01-y01 | Forward jet/Track | 2 | 0.1 | $\rho$ | [-4.5,-3.7] | [-3.5,-0.5] | Monash d49-x01-y01 | Forward jet/Track | 0 | 0.1 | $r_{g}$ | all | all | - d51-x01-y01 | Forward jet/Cluster | 1 | 0.1 | $r_{g}$ | [-0.8,-0.2] | [-1.2,-0.8],-0.1 | - d52-x01-y01 | Forward jet/Track | 1 | 0.1 | $r_{g}$ | - | all | Monash d53-x01-y01 | Forward jet/Cluster | 2 | 0.1 | $r_{g}$ | all | -0.15 | A14 d54-x01-y01 | Forward jet/Track | 2 | 0.1 | $r_{g}$ | -1.1 | [-1,-0.1] | Monash Table 4: ATLAS 2019 I1772062(Soft_Drop_Jet Observables) Plots | Observable | Better Tune Performance regions | Better Tune ---|---|---|--- | | A14 | Monash | (overall) d01-x01-y01 | Nsubjets | 0-10 | - | A14 d02-x01-y01 | $C_{2}$ | 0-0.86 | 0.36-0.42 , 0.64-0.72 | A14 d03-x01-y01 | $D_{2}$ | 0-0.5 | - | A14 d04-x01-y01 | LHA | 0,4.5 | - | A14 d05-x01-y01 | ECF${}_{2}^{norm}$ | 0-0.252 | 0.252-0.35 | A14 d06-x01-y01 | ECF${}_{3}^{norm}$ | 0-0.04 | - | A14 d23-x01-y01 | Nsubjets | 1-2 | 2-5 | Monash d24-x01-y01 | $C_{2}$ | 0-0.38 , 0.42-0.5 , 0.54-0.62 | 0.4 , 0.52 , 0.62-1.0 | A14 d25-x01-y01 | $D_{2}$ | 0-0.48 | - | A14 d26-x01-y01 | LHA | 0-1.4 | 1.4-9 | A14 d27-x01-y01 | ECF${}_{2}^{norm}$ | 0-0.23 | ECF${}_{2}^{norm}$ 0.23-0.35 | A14 d28-x01-y01 | ECF${}_{3}^{norm}$ | 0-0.02 | 0.02-0.32 | A14 Table 5: ATLAS 2019 I1724098(JSS, Dijet Selection) Plots | Observable | $\beta$ | Better Tune Performance regions | Better Tune ---|---|---|---|--- | ($p_{T}^{lead}>600$ GeV) | ($z_{cut}=0.1$) | A14 | Monash | (overall) d01-x01-y01 | $\log_{10}$[($m^{\textrm{soft drop}}/p_{T}^{\textrm{ungroomed}})^{2}$] | 0 | [-2.5,-0.5] | [-4,-2.5] | A14 d02-x01-y01 | $\log_{10}$[($m^{\textrm{soft drop}}/p_{T}^{\textrm{ungroomed}})^{2}$] | 1 | [-4.5,-0.5] | - | A14 d03-x01-y01 | $\log_{10}$[($m^{\textrm{soft drop}}/p_{T}^{\textrm{ungroomed}})^{2}$] | 2 | [-4.5,-0.5] | - | A14 Table 6: ATLAS 2017 I1637587(Soft Drop Mass) ### 5.2 Envelope plots Having decided the range of values for each parameter, we visualise the region of the distributions to check its utility and hence it is important before proceeding further. This can be done using the PROFESSOR tool by generating envelopes plots with the comand prof2-envelopes. The envelope plots show an area covering the distributions which indicates the bin values that the observables can take within the selected parameter ranges. These are shown in the following sub sections. #### 5.2.1 Soft Drop Mass Distributions Figure 6: Envelope plots for ATLAS_2017_I1637587 As can be seen in Figure 6, The envelopes cover the reference data in almost every bin and hence we can say that the range selected for the parameters are appropriate. #### 5.2.2 Lund Jet Plane Distributions Figure 7: Envelope plots for Lund Jet Plane As can be seen in Figure 7, the envelopes do not entirely cover the reference data. This is because we have reached a limit as to how much the distribution can be further fitted to the data with Pythia. Thus we consider this suitable for the purpose of this report and proceed with our set of parameter ranges. #### 5.2.3 Soft Drop Observables Distributions In the Figure 8, we see that the envelopes cover the data points in almost all the bins of our distributions of interest i.e soft drop jet mass from the soft drop jet observables analysis. Thus the parameter ranges are suitable for proceeding to tune the distributions. Figure 8: Envelope plots for Soft Drop Mass observable ### 5.3 Weight file The weight file assigns weights to the distributions to be tuned and hence is manually changed depending on our interests. For this project, a total of 16 distributions were assigned weights greater than 1 and 6 distributions were given no weight i.e 0 for obtaining the Common Tune. These are : Analysis | Distribution code | Weight ---|---|--- ATLAS_2020_I1790256 | d03-x01-y01 | 106 ATLAS_2020_I1790256 | d04-x01-y01 | 112 ATLAS_2020_I1790256 | d05-x01-y01 | 108 ATLAS_2020_I1790256 | d06-x01-y01 | 20 ATLAS_2020_I1790256 | d07-x01-y01 | 16.6 ATLAS_2020_I1790256 | d08-x01-y01 | 16 ATLAS_2020_I1790256 | d09-x01-y01 | 16 ATLAS_2019_I1772062 | d19-x01-y01 | 75 ATLAS_2019_I1772062 | d20-x01-y01 | 75 ATLAS_2019_I1772062 | d21-x01-y01 | 200 ATLAS_2019_I1772062 | d22-x01-y01 | 80 ATLAS_2019_I1772062 | d23-x01-y01 | 200 ATLAS_2019_I1772062 | d24-x01-y01 | 80 ATLAS_2017_I1637587 | d01-x01-y01 | 500 ATLAS_2017_I1637587 | d02-x01-y01 | 500 ATLAS_2017_I1637587 | d03-x01-y01 | 500 ATLAS_2019_I1772062 | d61-x01-y01 | 0 ATLAS_2019_I1772062 | d62-x01-y01 | 0 ATLAS_2019_I1772062 | d79-x01-y01 | 0 ATLAS_2019_I1772062 | d80-x01-y01 | 0 ATLAS_2019_I1772062 | d97-x01-y01 | 0 ATLAS_2019_I1772062 | d98-x01-y01 | 0 Table 7: Weights assigned to obtain the Common tune
# Some Estimates of the Generalized Beukers Integral with Techniques of Partial Fraction Decomposition Xiaowei Wang(Potsdam) This paper was written in June 2020 ###### Abstract In this paper we establish the generalized Beukers integral $I_{m}(a_{1},...,a_{n})$ with some methods of partial fraction decomposition. Thus one obtains an explicit expression of the generalized Beukers integral. Further, we estimate the rational denominator of $I$ and. In the second section of this paper, we provide some estimates of the upper and lower bound of the value $J_{3}$, which involves the generalized Beukers integral and is related to $\zeta(5)$. _K_ eywords generalized Beukers integral, zeta(5), partial fraction decomposition ## 1 The Lemmas ###### Lemma 1. (Homogeneous partial fraction decomposition) Let $a_{1},...,a_{n}$ be distinct complex number, $x\in\mathbb{C}\backslash\\{-a_{1},...,-a_{n}\\}$, then there exist $\lambda_{1},...,\lambda_{n}\in\mathbb{C}$ such that following identity is true, $\prod_{i=1}^{n}\frac{1}{a_{i}+x}=\sum_{i=1}^{n}\frac{\lambda_{i}}{a_{i}+x}$ (1) where $\lambda_{i}$ has explicit expression as following. They only depend on $a_{1},...,a_{n}$. $\lambda_{i}=\prod_{j=1,j\neq i}^{n}\frac{1}{a_{j}-a_{i}}$ Further, we have $\sum_{i=1}^{n}\lambda_{i}=0$ ###### Proof. In order to show (1), we multiply $\prod_{i=1}^{n}(a_{i}+x)$ on both side of (1). It becomes $\sum_{i=1}^{n}\lambda_{i}\prod_{j=1,j\neq i}^{n}(a_{j}+x)=1$ Now let $p(x)=\sum_{i=1}^{n}\lambda_{i}\prod_{j=1,j\neq i}^{n}(a_{j}+x)=\sum_{i=1}^{n}\prod_{j=1,j\neq i}^{n}\frac{a_{j}+x}{a_{j}-a_{i}}$ It’s easy to see that $p(x)$ is a polynomial with degree $n-1$ and satisfies that $p(-a_{i})=1$ for all $i=1,...,n$. On the one hand we already found $n$ zeros of $p(x)-1$, on the other hand by the fundamental theorem of algebra, $p(x)-1$ has $n-1$ zeros. Therefore it can only be $p(x)\equiv 1$. That is $\sum_{i=1}^{n}\lambda_{i}\prod_{j=1,j\neq i}^{n}(a_{j}+x)\equiv 1$ Comparing the coefficient of $x^{n-1}$ on both side, we obtain $\sum_{i=1}^{n}\lambda_{i}=0$ ∎ ###### Lemma 2. (Inhomogeneous partial fraction decomposition) Let $c_{1},...,c_{n}$ be distinct complex numbers, $b_{1},...,b_{n}$ be positive integers, then following decomposition is valid for $x\in\mathbb{C}\backslash\\{-c_{1},...,-c_{n}\\}$. $\prod_{i=1}^{n}\frac{1}{(c_{i}+x)^{b_{i}}}=\sum_{i=1}^{n}\sum_{j=1}^{b_{i}}\frac{\mu_{ij}}{(c_{i}+x)^{j}}$ The expression of $\mu_{ij}$ is given by $\mu_{ij}=\frac{(-1)^{j-1}}{(b_{i}-j)!}\frac{d^{b_{i}-j}}{dz^{b_{i}-j}}|_{z=c_{i}}(\prod_{\ell=1,\ell\neq i}^{n}\frac{1}{(c_{\ell}-z)^{b_{\ell}}})$ Note that if $b_{1}=...=b_{n}=1$, then $\mu_{i1}$ is exactly $\lambda_{i}$ in Lemma 1. Moreover, we have $\sum_{i=1}^{n}\mu_{i1}=0$ ###### Proof. Let $f,g$ are both functions of $z_{1},...,z_{n}$, namely $\displaystyle f(z_{1},...,z_{n})$ $\displaystyle:=\prod_{i=1}^{n}\frac{1}{(z_{i}+x)^{b_{i}}}$ $\displaystyle g(z_{1},...,z_{n})$ $\displaystyle:=\prod_{i=1}^{n}\frac{1}{z_{i}+x}$ According to Lemma 1, we have an equality for $x\in\mathbb{C}\backslash\\{-z_{1},...,-z_{n}\\}$ $\prod_{i=1}^{n}\frac{1}{z_{i}+x}=\sum_{i=1}^{n}\frac{\lambda_{i}}{z_{i}+x}$ (2) where $\lambda_{i}=\prod_{\ell=1,j\neq i}^{n}\frac{1}{z_{\ell}-z_{i}}$ holds for all $i$. Now we regard $\lambda_{i}$ as function of $z_{1},...,z_{n}$. Taking partial derivatives $\partial_{(b_{1}-1,...,b_{n}-1)}$ on both sides of (2), we obtain following. Here the notation $\partial_{(N_{1},...,N_{m})}$ means $\frac{\partial^{N_{1}+...+N_{m}}}{\partial z_{1}^{N_{1}}...\partial z_{m}^{N_{m}}}$, sometimes $\frac{\partial^{N_{1}+...+N_{m}}F}{\partial z_{1}^{N_{1}}...\partial z_{m}^{N_{m}}}$ is denote by $F^{(N_{1},...,N_{m})}$ for convenience. $\partial_{(b_{1}-1,...,b_{n}-1)}g=\prod_{i=1}^{n}\frac{(-1)^{b_{i}-1}(b_{i}-1)!}{(z_{i}+x)^{b_{i}}}=(\prod_{i=1}^{n}(-1)^{b_{i}-1}(b_{i}-1)!)f$ On the other hand, $\displaystyle\partial_{(b_{1}-1,...,b_{n}-1)}\sum_{i=1}^{n}\frac{\lambda_{i}}{z_{i}+x}$ $\displaystyle=\sum_{i=1}^{n}\partial_{(0,...,b_{i}-1,...,0)}\partial_{(b_{1}-1,...,b_{i-1}-1,0,b_{i+1}-1,...,b_{n}-1)}\frac{\lambda_{i}}{z_{i}+x}$ $\displaystyle=\sum_{i=1}^{n}\partial_{(0,...,b_{i}-1,...,0)}\frac{\lambda_{i}^{(b_{1}-1,...,b_{i-1}-1,0,b_{i+1}-1,...,b_{n}-1)}}{z_{i}+x}$ $\displaystyle=\sum_{i=1}^{n}\sum_{j=1}^{b_{i}}\binom{b_{i}-1}{j-1}\lambda_{i}^{(b_{1}-1,...,b_{i-1}-1,b_{i}-j,b_{i+1}-1,...,b_{n}-1)}(\frac{1}{z_{i}+x})^{(0,0,...,j-1,...,0)}$ $\displaystyle=\sum_{i=1}^{n}\sum_{j=1}^{b_{i}}\binom{b_{i}-1}{j-1}\lambda_{i}^{(b_{1}-1,...,b_{i-1}-1,b_{i}-j,b_{i+1}-1,...,b_{n}-1)}\frac{(-1)^{j-1}(j-1)!}{(z_{i}+x)^{j}}$ Supposed that $\prod_{i=1}^{n}\frac{1}{(z_{i}+x)^{b_{i}}}=\sum_{i=1}^{n}\sum_{j=1}^{b_{i}}\frac{\mu_{ij}}{(z_{i}+x)^{j}}$ by comparing the coefficients we obtain $\displaystyle\mu_{ij}$ $\displaystyle=\frac{(-1)^{j-1}(j-1)!}{\prod_{\ell=1}^{n}(-1)^{b_{\ell}-1}(b_{\ell}-1)!}\binom{b_{i}-1}{j-1}\lambda_{i}^{(b_{1}-1,...,b_{i-1}-1,b_{i}-j,b_{i+1}-1,...,b_{n}-1)}$ $\displaystyle=\frac{(-1)^{j-1}}{(b_{i}-j)!\prod_{\ell=1,\ell\neq i}^{n}(-1)^{b_{\ell}-1}(b_{\ell}-1)!}\lambda_{i}^{(b_{1}-1,...,b_{i-1}-1,b_{i}-j,b_{i+1}-1,...,b_{n}-1)}$ Finally, it remains to compute $\lambda_{i}^{(b_{1}-1,...,b_{i-1}-1,b_{i}-j,b_{i+1}-1,...,b_{n}-1)}$. $\displaystyle\partial_{(b_{1}-1,...,b_{i-1}-1,b_{i}-j,b_{i+1}-1,...,b_{n}-1)}\lambda_{i}$ $\displaystyle=$ $\displaystyle\partial_{(0,...,b_{i}-j,...,0)}\partial_{(b_{1}-1,...,b_{i-1}-1,0,b_{i+1}-1,...,b_{n}-1)}\prod_{\ell=1,\ell\neq i}^{n}\frac{1}{z_{\ell}-z_{i}}$ $\displaystyle=$ $\displaystyle\partial_{(0,...,b_{i}-j,...,0)}\prod_{\ell=1,\ell\neq i}^{n}\frac{(-1)^{b_{\ell}-1}(b_{\ell}-1)!}{(z_{\ell}-z_{i})^{b_{\ell}}}$ $\displaystyle=$ $\displaystyle(\prod_{\ell=1,\ell\neq i}^{n}(-1)^{b_{\ell}-1}(b_{\ell}-1)!)(\prod_{\ell=1,\ell\neq i}^{n}\frac{1}{(z_{\ell}-z_{i})^{b_{\ell}}})^{(0,...,b_{i}-j,...,0)}$ That is $\mu_{ij}=\frac{(-1)^{j-1}}{(b_{i}-j)!}\frac{d^{b_{i}-j}}{dz^{b_{i}-j}}|_{z=c_{i}}(\prod_{\ell=1,\ell\neq i}^{n}\frac{1}{(c_{\ell}-z)^{b_{\ell}}})$ In order to prove $\sum_{i=1}^{n}\mu_{i1}=0$ just need to multiply $\prod_{i=1}^{n}(x+c_{i})^{b_{i}}$ on both sides of $\prod_{i=1}^{n}\frac{1}{(x+c_{i})^{b_{i}}}\equiv\sum_{i=1}^{n}\sum_{j=1}^{b_{i}}\frac{\mu_{ij}}{(x+c_{i})^{j}}$ Then it becomes $1\equiv\sum_{i=1}^{n}\sum_{j=1}^{b_{i}}\mu_{ij}\frac{\prod_{k=1}^{n}(x+c_{k})^{b_{k}}}{(x+c_{i})^{j}}$ The right hand side of the equality is a polynomial of $x$ with the degree $b_{1}+...+b_{n}-1$. Since this polynomial is actually constant $1$, therefore the initial coefficient is $0$ and only $\mu_{i1}$ contributes to the coefficient of $x^{b_{1}+...+b_{n}-1}$. Consequently, we infer that $\sum_{i=1}^{n}\mu_{i1}=0$ ∎ ## 2 The First Attempt In this section, we discuss the more practical case $n=2$. Hadjicostas [1] called it the first generalization. The general cases are discussed in the next section. ###### Theorem 1. Suppose that $a,b,m$ are nonnegative integers. Define $I_{m}(a,b)=\frac{(-1)^{m}}{m!}\int_{(0,1)^{2}}\frac{\log^{m}(xy)x^{a}y^{b}}{1-xy}dxdy$ It’s easy to see $I_{m}(a,b)=I_{m}(b,a)$. Suppose that $a\leq b$, then without loss of generality, we have $I_{m}(a,b)=\begin{cases}&\frac{H_{m+1}(b)-H_{m+1}(a)}{b-a}\text{, if }a<b\\\ &(m+1)\zeta(m+2,a+1)\text{, if }a=b\\\ \end{cases}$ where $H_{m}(x)$ is the generalized harmonic number, which is given by $H_{m}(x)=\sum_{k=1}^{\lfloor x\rfloor}\frac{1}{k^{m}}$. ###### Proof. For $t\geq 0$, define $A(t,a,b):=\int_{(0,1)^{2}}\frac{x^{a+t}y^{b+t}}{1-xy}dxdy$ For $x,y\in(0,1)$, the series $\sum_{k=0}^{\infty}x^{a+t+k}y^{b+t+k}$ converges absolutely and uniformly for all $x,y\in(\varepsilon,1-\varepsilon)$ to $\frac{x^{a+t}y^{b+t}}{1-xy}$. Hence $\displaystyle A(t,a,b)$ $\displaystyle=\int_{(0,1)^{2}}\frac{x^{a+t}y^{b+t}}{1-xy}dxdy$ $\displaystyle=\sum_{k=0}^{\infty}\int_{0}^{1}x^{a+t+k}dx\int_{0}^{1}y^{b+t+k}dy$ $\displaystyle=\sum_{k=1}^{\infty}\frac{1}{(a+t+k)(b+t+k)}$ In following we consider taking $\frac{\partial^{m}}{\partial t^{m}}|_{t=0}$ on both sides of the equality. There are two cases: Case I. If $a=b$, $\sum_{k=1}^{\infty}\frac{1}{(a+t+k)(b+t+k)}=\sum_{k=1}^{\infty}\frac{1}{(a+t+k)^{2}}=\zeta(2,a+t+1)$ We have $\displaystyle\frac{(-1)^{m}}{m!}\frac{\partial^{m}}{\partial t^{m}}|_{t=0}A(t,a,b)$ $\displaystyle=$ $\displaystyle\frac{(-1)^{m}}{m!}\frac{\partial^{m}}{\partial t^{m}}|_{t=0}\zeta(2,a+t+1)$ $\displaystyle=$ $\displaystyle(m+1)\zeta(m+2,a+1)$ Case II. If $a<b$, then the decomposition $\frac{1}{(a+t+k)(b+t+k)}=\frac{1}{b-a}(\frac{1}{a+t+k}-\frac{1}{b+t+k})$ is true for all positive integer $k$ and all nonnegative real number $t$. This implies that $\sum_{k=1}^{\infty}\frac{1}{(a+t+k)(b+t+k)}=\frac{1}{b-a}\sum_{k=1}^{\infty}(\frac{1}{a+t+k}-\frac{1}{b+t+k})$ Therefore $\displaystyle\frac{(-1)^{m}}{m!}\frac{\partial^{m}}{\partial t^{m}}|_{t=0}A(t,a,b)$ $\displaystyle=$ $\displaystyle\frac{(-1)^{m}}{m!}\frac{\partial^{m}}{\partial t^{m}}|_{t=0}\sum_{k=1}^{\infty}\frac{1}{(a+t+k)(b+t+k)}$ $\displaystyle=$ $\displaystyle\frac{1}{b-a}\sum_{k=1}^{\infty}(\frac{1}{(a+k)^{m+1}}-\frac{1}{(b+k)^{m+1}})$ $\displaystyle=$ $\displaystyle\frac{1}{b-a}(\sum_{k=1}^{\infty}\frac{1}{(a+k)^{m+1}}-\sum_{k=b-a+1}^{\infty}\frac{1}{(a+k)^{m+1}})$ $\displaystyle=$ $\displaystyle\frac{H_{m+1}(b)-H_{m+1}(a)}{b-a}$ On the other hand, no matter in which case, from the above integral representation we have $\frac{(-1)^{m}}{m!}\frac{\partial^{m}}{\partial t^{m}}|_{t=0}A(t,a,b)=\frac{(-1)^{m}}{m!}\int_{(0,1)^{2}}\frac{\log^{m}(xy)x^{a}y^{b}}{1-xy}dxdy=I_{m}(a,b)$ The details about convergence and interchanging the order of integration, summation and derivatives are omitted here, one can see[1]. As a consequence, $I_{m}(a,b)=\begin{cases}&\frac{H_{m+1}(b)-H_{m+1}(a)}{b-a}\text{, if }a<b\\\ &(m+1)\zeta(m+2,a+1)\text{, if }a=b\\\ \end{cases}$ ∎ ## 3 The Generalized Beukers Integral In following we use the notation $\\{x_{1}^{(N{1})},...,x_{j}^{(N{j})}\\}$ to represent a finite multiset, where $N_{i}$ is the multiplicity of $x_{i}$, $i=1,...,j$. ###### Theorem 2. (Generalized Beukers Integral Representation) Assume that $n\geq 2$, and $m,a_{1},...,a_{n}$ be nonnegative integers. Define $I_{m}(a_{1},a_{2},\ldots,a_{n}):=\frac{(-1)^{m}}{m!}\int_{(0,1)^{n}}\frac{\log^{m}(\prod_{i=1}^{n}x_{i})\prod_{i=1}^{n}x_{i}^{a_{i}}}{1-\prod_{i=1}^{n}x_{i}}dx_{1}...dx_{n}$ Let $\\{c_{1}^{(b_{1})},...,c_{r}^{(b_{r})}\\}$ be the multiset of $a_{1},...,a_{n}$ with $c_{1}<...<c_{r}$ and $b_{1}+...+b_{r}=n$, then * • if $r=1$, $I_{m}(a_{1},a_{2},\ldots,a_{n})=\binom{m+n-1}{m}\zeta(n+m,c_{1}+1)$ * • if $1<r\leq n$, $I_{m}(a_{1},a_{2},\ldots,a_{n})=\sum_{i=1}^{r-1}\mu_{i1}(H_{m+1}(c_{r})-H_{m+1}(c_{i}))+\sum_{i=1}^{r}\sum_{j\geq 2}\binom{m+j-1}{m}\mu_{ij}\zeta(j+m,c_{i}+1)$ where $\lambda_{i}$ and $\mu_{ij}$ are defined by the homogeneous and inhomogeneous partial fraction decomposition (Lemma1 and Lemma2) as following respectively $\displaystyle\prod_{i=1}^{n}\frac{1}{a_{i}+x}=\sum_{i=1}^{n}\frac{\lambda_{i}}{a_{i}+x}$ $\displaystyle\prod_{i=1}^{r}\frac{1}{(c_{i}+x)^{b_{i}}}=\sum_{i=1}^{r}\sum_{j=1}^{b_{i}}\frac{\mu_{ij}}{(c_{i}+x)^{j}}$ ###### Proof. Assume that $t\geq 0$, define $A(t,a_{1},a_{2},...,a_{n}):=\int_{(0,1)^{n}}\frac{\prod_{i=1}^{n}x_{i}^{a_{i}+t}}{1-\prod_{i=1}^{n}x_{i}}dx_{1}...dx_{n}$ Since all $x_{1},...,x_{n}\in(0,1)$, it has a series expansion as $\frac{\prod_{i=1}^{n}x_{i}^{a_{i}+t}}{1-\prod_{i=1}^{n}x_{i}}=\sum_{k=0}^{\infty}\prod_{i=1}^{n}x_{i}^{a_{i}+k+t}$ Therefore $A(t,a_{1},a_{2},...,a_{n})=\int_{(0,1)^{n}}\frac{\prod_{i=1}^{n}x_{i}^{a_{i}+t}}{1-\prod_{i=1}^{n}x_{i}}dx_{1}...dx_{n}=\sum_{k=0}^{\infty}\prod_{i=1}^{n}\frac{1}{1+a_{i}+k+t}$ The series on the right hand side absolutely and uniformly converges on $x_{i}\in(\varepsilon,1-\varepsilon),i=1,2,...,n$. For the details, see[1]. Similar to the first attempt, the main idea is also taking the $m$-partial derivatives with respect to $t$ around $0$ on both sides of the equation. There are several different cases. Case I, $r=1$ and $b_{1}=n$, which means $a_{1}=a_{2}=...=a_{n}=a=c_{1}$. In this case $\sum_{k=0}^{\infty}\prod_{i=1}^{n}\frac{1}{1+a_{i}+k+t}$ becomes $\sum_{k=0}^{\infty}\frac{1}{(1+a+k+t)^{n}}$. Then $\displaystyle\frac{\partial^{m}}{\partial t^{m}}|_{t=0}A(t,a_{1},a_{2},...,a_{n})$ $\displaystyle=(-1)^{m}\frac{(n+m-1)!}{(n-1)!}\sum_{k=0}^{\infty}\frac{1}{(1+a+k)^{n+m}}$ $\displaystyle=(-1)^{m}\frac{(n+m-1)!}{(n-1)!}\zeta(n+m,a+1)$ $\displaystyle=(-1)^{m}m!\binom{m+n-1}{m}\zeta(n+m,a+1)$ Case II $r=n,b_{1}=...=b_{n}=1$, namely $a_{1}<a_{2}<...a_{n}$. At First, to decompose the product $\prod_{i=1}^{n}\frac{1}{1+a_{i}+k+t}$ as $\prod_{i=1}^{n}\frac{1}{1+a_{i}+k+t}=\sum_{i=1}^{n}\lambda_{i}\frac{1}{1+a_{i}+k+t}$ It follows from Lemma 1 that there exist $\lambda_{1},...,\lambda_{n}$ which are independent to $k,t$. At first obviously $\sum_{k=0}^{\infty}\prod_{i=1}^{n}\frac{1}{1+a_{i}+k}$ is convergent, hence $\displaystyle A(0,a_{1},a_{2},...,a_{n})$ $\displaystyle=\sum_{k=0}^{\infty}\sum_{i=1}^{n}\lambda_{i}\frac{1}{1+a_{i}+k}$ $\displaystyle=\lim_{N\rightarrow\infty}(\lambda_{1}\sum_{k=a_{1}+1}^{N}\frac{1}{k}+\ldots+\lambda_{n}\sum_{k=a_{n}+1}^{N}\frac{1}{k})$ $\displaystyle=\lambda_{1}\sum_{k=a_{1}+1}^{a_{n}}\frac{1}{k}+\ldots+\lambda_{n-1}\sum_{k=a_{n-1}+1}^{a_{n}}\frac{1}{k}+(\lambda_{1}+\ldots+\lambda_{n})\lim_{N\rightarrow\infty}\sum_{k=a_{n}+1}^{N}\frac{1}{k}$ recall that $\lambda_{1}+\ldots+\lambda_{n}=0$, therefore $A(0,a_{1},a_{2},...,a_{n})=\sum_{i=1}^{n-1}\lambda_{i}\sum_{k=a_{i}+1}^{a_{n}}\frac{1}{k}$ Now assume that $m\geq 1$, $\displaystyle\frac{\partial^{m}}{\partial t^{m}}|_{t=0}A(t,a_{1},a_{2},...,a_{n})$ $\displaystyle=$ $\displaystyle(-1)^{m}m!\sum_{k=0}^{\infty}\sum_{i=1}^{n}\lambda_{i}\frac{1}{(1+a_{i}+k)^{m+1}}$ $\displaystyle=$ $\displaystyle(-1)^{m}m!(\lambda_{1}\sum_{k=a_{1}+1}^{a_{n}}\frac{1}{k^{m+1}}+\ldots+\lambda_{n-1}\sum_{k=a_{n-1}+1}^{a_{n}}\frac{1}{k^{m+1}}+(\lambda_{1}+\ldots+\lambda_{n})\sum_{k=a_{n}+1}^{\infty}\frac{1}{k^{m+1}})$ $\displaystyle=$ $\displaystyle(-1)^{m}m!\sum_{i=1}^{n-1}\lambda_{i}\sum_{k=a_{i}+1}^{a_{n}}\frac{1}{k^{m+1}}$ In a nutshell, we have $\frac{\partial^{m}}{\partial t^{m}}|_{t=0}A(t,a_{1},a_{2},...,a_{n})=(-1)^{m}m!\sum_{i=1}^{n-1}\lambda_{i}\sum_{k=a_{i}+1}^{a_{n}}\frac{1}{k^{m+1}}$ Case III Some $a_{i}$ are the same. In this case $\\{a_{1},...,a_{n}\\}$ can be represented as multiset $\\{c_{1}^{(b_{1})},...,c_{r}^{(b_{r})}\\}$, where $c_{1}<...<c_{r}$, $b_{1}+...+b_{r}=n$. It follows from Lemma 2 that. $\prod_{i=1}^{n}\frac{1}{1+a_{i}+k+t}=\prod_{i=1}^{r}\frac{1}{(1+c_{i}+k+t)^{b_{i}}}=\sum_{i=1}^{r}\sum_{j=1}^{b_{i}}\frac{\mu_{ij}}{(1+c_{i}+k+t)^{j}}$ then $\displaystyle\frac{\partial^{m}}{\partial t^{m}}|_{t=0}A(t,a_{1},a_{2},...,a_{n})$ $\displaystyle=$ $\displaystyle\sum_{k=0}^{\infty}\sum_{i=1}^{r}\sum_{j=1}^{b_{i}}\frac{(-1)^{m}(m+j-1)!\mu_{ij}}{(j-1)!}\frac{1}{(1+c_{i}+k)^{j+m}}$ $\displaystyle=$ $\displaystyle(-1)^{m}(m)!\sum_{k=0}^{\infty}\sum_{i=1}^{r}\mu_{i1}\frac{1}{(1+c_{i}+k)^{1+m}}+\sum_{k=0}^{\infty}\sum_{i=1}^{r}\sum_{j=2}^{b_{i}}\frac{(-1)^{m}(m+j-1)!\mu_{ij}}{(j-1)!}\frac{1}{(1+c_{i}+k)^{j+m}}$ By the conclusion of Lemma 2, $\sum_{i=1}^{r}\mu_{i1}=0$, therefore $\displaystyle\sum_{k=0}^{\infty}\sum_{i=1}^{r}\mu_{i1}\frac{1}{(1+c_{i}+k)^{1+m}}$ $\displaystyle=\sum_{i=1}^{r}\sum_{k=c_{i}+1}^{\infty}\frac{\mu_{i1}}{k^{1+m}}$ $\displaystyle=\sum_{i=1}^{r-1}\sum_{k=c_{i}+1}^{c_{r}}\frac{\mu_{i1}}{k^{1+m}}+\sum_{i=1}^{r}\mu_{i1}\sum_{k=c_{r}+1}^{\infty}\frac{1}{k^{1+m}}$ $\displaystyle=\sum_{i=1}^{r-1}\sum_{k=c_{i}+1}^{c_{r}}\frac{\mu_{i1}}{k^{1+m}}$ This is a rational number. On the other hand, note that if $j\geq 2$, then $\sum_{k=0}^{\infty}\frac{1}{(1+c_{i}+k)^{j+m}}=\zeta(j+m)-\sum_{k=1}^{c_{i}}\frac{1}{k^{j+m}}$ Hence $\displaystyle\sum_{k=0}^{\infty}\sum_{i=1}^{r}\sum_{j=2}^{b_{i}}\frac{(-1)^{m}(m+j-1)!\mu_{ij}}{(j-1)!}\frac{1}{(1+c_{i}+k)^{j+m}}$ $\displaystyle=$ $\displaystyle(-1)^{m}\sum_{i=1}^{r}\sum_{j=2}^{b_{i}}\frac{(m+j-1)!\mu_{ij}}{(j-1)!}(\zeta(j+m)-\sum_{k=1}^{c_{i}}\frac{1}{k^{j+m}})$ $\displaystyle=$ $\displaystyle(-1)^{m}\sum_{i=1}^{r}\sum_{j=2}^{b_{i}}\frac{(m+j-1)!\mu_{ij}}{(j-1)!}\zeta(j+m,c_{i}+1)$ It turns out that $\displaystyle\frac{\partial^{m}}{\partial t^{m}}|_{t=0}A(t,a_{1},a_{2},...,a_{n})$ $\displaystyle=$ $\displaystyle(-1)^{m}m!\sum_{i=1}^{r-1}\sum_{k=c_{i}+1}^{c_{r}}\frac{\mu_{i1}}{k^{1+m}}+(-1)^{m}\sum_{i=1}^{r}\sum_{j=2}^{b_{i}}\frac{(m+j-1)!\mu_{ij}}{(j-1)!}\zeta(j+m,c_{i}+1)$ $\displaystyle=$ $\displaystyle(-1)^{m}m!\\{\sum_{i=1}^{r-1}\sum_{k=c_{i}+1}^{c_{r}}\frac{\mu_{i1}}{k^{1+m}}+\sum_{i=1}^{r}\sum_{j=2}^{b_{i}}\binom{m+j-1}{m}\mu_{ij}\zeta(j+m,c_{i}+1)\\}$ $\displaystyle=$ $\displaystyle(-1)^{m}m!\\{\sum_{i=1}^{r-1}\mu_{i1}(H_{m+1}(c_{r})-H_{m+1}(c_{i}))+\sum_{i=1}^{r}\sum_{j\geq 2}\binom{m+j-1}{m}\mu_{ij}\zeta(j+m,c_{i}+1)\\}$ If $r=n$, then $b_{1}=...=b_{n}=1$, that is $\lambda_{i}=\mu_{i1}$. Case II is in fact included in Case III. On the other hand, no matter in which case, since $A(t,a_{1},a_{2},...,a_{n})=\int_{(0,1)^{n}}\frac{\prod_{i=1}^{n}x_{i}^{t+a_{i}}}{1-\prod_{i=1}^{n}x_{i}}dx_{1}\ldots dx_{n}$ then $\frac{\partial^{m}}{\partial t^{m}}|_{t=0}A(t,a_{1},a_{2},...,a_{n})=\int_{(0,1)^{n}}\frac{\log^{m}(\prod_{i=1}^{n}x_{i})\prod_{i=1}^{n}x_{i}^{a_{i}}}{1-\prod_{i=1}^{n}x_{i}}dx_{1}\ldots dx_{n}$ Therefore as a consequence, if $r=1$ $I_{m}(a_{1},a_{2},\ldots,a_{n})==\binom{m+n-1}{m}\zeta(n+m,c_{1}+1)$ if $1<r\leq n$, $I_{m}(a_{1},a_{2},\ldots,a_{n})=\sum_{i=1}^{r-1}\mu_{i1}(H_{m+1}(c_{r})-H_{m+1}(c_{i}))+\sum_{i=1}^{r}\sum_{j\geq 2}\binom{m+j-1}{m}\mu_{ij}\zeta(j+m,c_{i}+1)$ The details about convergence and interchanging the order of integration, summation and derivatives are omitted here, one can see[1]. ∎ ###### Example 3. Let $I_{m}(a_{1},a_{2},a_{3})=\frac{(-1)^{m}}{m!}\int_{(0,1)^{3}}\frac{\log^{m}(xyz)x^{a_{1}}y^{a_{2}}z^{a_{3}}}{1-xyz}dxdydz$ where $a_{1},a_{2},a_{3}$ nonnegative integers, * • If $a_{1}=a_{2}=a_{3}=a$, then $I_{m}(a_{1},a_{2},a_{3})=\binom{m+2}{m}\zeta(m+3,a+1)=\frac{(m+1)(m+2)}{2}(\zeta(m+3)-H_{m+3}(a))$ * • If $a_{1}<a_{2}<a_{3}$, then $\displaystyle I_{m}(a_{1},a_{2},a_{3})$ $\displaystyle=$ $\displaystyle\frac{1}{(a_{2}-a_{1})(a_{3}-a_{1})}(H_{m+1}(a_{3})-H_{m+1}(a_{1}))+\frac{1}{(a_{1}-a_{2})(a_{3}-a_{2})}(H_{m+1}(a_{3})-H_{m+1}(a_{2}))$ * • If $c_{1}=a_{1}=a_{2}<a_{3}=c_{2}$, then $I_{m}(a_{1},a_{2},a_{3})=\mu_{11}(H_{m+1}(c_{2})-H_{m+1}(c_{1}))+(m+1)\mu_{12}\zeta(m+2,c_{1}+1)$ * • If $c_{1}=a_{1}<a_{2}=a_{3}=c_{2}$, then $I_{m}(a_{1},a_{2},a_{3})=\mu_{11}(H_{m+1}(c_{2})-H_{m+1}(c_{1}))+(m+1)\mu_{22}\zeta(m+2,c_{2}+1)$ ###### Example 4. As a special case of $I_{m}(a_{1},...,a_{n})$, let $n=1$, $a_{1}=a$, then $I_{m}(a)=\frac{(-1)^{m}}{m!}\int_{0}^{1}\frac{\log^{m}(x)x^{a}}{1-x}dx$ In fact this integral converges if $m\geq 1$. To see this, firstly consider $f_{N}(x)=x^{a+t}(1+x+...+x^{N-1})=\frac{x^{a+t}(1-x^{N})}{1-x}$ where $N$ is an integer sufficiently large. Observe the integral $\int_{0}^{1}f_{N}(x)dx=\sum_{k=1}^{N}\frac{1}{a+t+k}$ and taking $\frac{d^{m}}{dt^{m}}|_{t=0}$ on both sides, where $m\geq 1$, $m\in\mathbb{Z}$, we get $\int_{0}^{1}\frac{\log^{m}(x)x^{a}(1-x^{N})}{1-x}dx=\sum_{k=1}^{N}\frac{(-1)^{m}m!}{(a+k)^{m+1}}$ Let $N\rightarrow\infty$, then $x^{N}\rightarrow 0$ for all $x\in(0,1)$. That is $\int_{0}^{1}\frac{\log^{m}(x)x^{a}}{1-x}dx=\sum_{k=1}^{\infty}\frac{(-1)^{m}m!}{(a+k)^{m+1}}$ Therefore $I_{m}(a)=\frac{(-1)^{m}}{m!}\int_{0}^{1}\frac{\log^{m}(x)x^{a}}{1-x}dx=\sum_{k=1}^{\infty}\frac{1}{(a+k)^{m+1}}=\zeta(m+1,a+1)$ (3) It’s well defined if $m\geq 1,a\geq 0$, $a,m\in\mathbb{Z}$. In fact, recall the integral representation of Hurwitz zeta function $\zeta(m+1,a+1)=\frac{1}{\Gamma(m+1)}\int_{0}^{\infty}\frac{t^{m}e^{-(a+1)t}}{1-e^{-t}}dt$ for $\Re(m)>0,\Re(a)>-1$. To substitute $t$ by $-\log(x)$, by simple computation we obtain $\zeta(m+1,a+1)=\frac{1}{\Gamma(m+1)}\int_{0}^{1}\frac{(-\log(x))^{m}x^{a}}{1-x}dx$ It is exactly (3) formally, but here $a,m\in\mathbb{C}$ and $\Re(m)>0,\Re(a)>-1$. ###### Theorem 5. Assume that $n\geq 2$, and $m,a_{1},...,a_{n}$ be nonnegative integers, $\\{c_{1}^{(b_{1})},...,c_{r}^{(b_{r})}\\}$ be the multiset representation of $a_{1},...,a_{n}$ with $c_{1}<...<c_{r}$ and $b_{1}+...+b_{r}=n$, $b_{+}=\max\\{b_{1},...,b_{r}\\}$. According to Theorem 2, it follows that $I_{m}(a_{1},...,a_{n})=\frac{p_{1}+p_{2}\zeta(m+2)+...+p_{n}\zeta(m+b_{+})}{q}$ where $p_{1},...,p_{n},q\in\mathbb{Z}$ with $(p_{i},q)=1$ for all $i$, we have the following estimates of $q$. If $r=1$, then $q|lcm(1,...,a_{n})^{n+m}$ If $r>1$, then $q|(b_{+}-1)!\cdot lcm(1,...,c_{r})^{m+b_{+}}\prod_{1\leq s<t\leq r}(c_{t}-c_{s})^{n-1}$ Before showing the proof, we firstly recall some concepts and facts. Let $x\in\mathbb{Q}$ and $x\neq 0$, then there are always integers $p,q$ satisfying $x=p/q$ and $q>0$ with $(p,q)=1$. $q$ is called the reduced denominator of $x$, which is denoted by $\delta(x)$ in this paper. In fact, assume that $x\in\mathbb{Q},a\in\mathbb{Z}$, both $a,x\neq 0$, if $ax\in\mathbb{Z}$ then $\delta(x)|a$. The lowest common multiple of $x_{1},...,x_{n}$ is denoted by $lcm(x_{1},...,x_{n})$. A very simple fact is that, if $a,b\in\mathbb{Q}$ and $a,b\neq 0$, then $\delta(a+b)|lcm(\delta(a),\delta(b))$. This is due to $lcm(\delta(a),\delta(b))\cdot(a+b)\in\mathbb{Z}$. ###### Proof. Firstly reformulating the expression of $I_{m}(a_{1},...,a_{n})$, there are two cases Case I, if $r=1$, that is $c_{1}=a_{1}=a_{2}=...=a_{n}$. Follows from the result of preceding theorem, we have $\displaystyle I_{m}(a_{1},...,a_{n})$ $\displaystyle=\binom{m+n-1}{m}\zeta(n+m,c_{1}+1)$ $\displaystyle=\binom{m+n-1}{m}\zeta(n+m)-\binom{m+n-1}{m}H_{n+m}(c_{1})$ Since $\binom{m+n-1}{m}$ is always an integer, it’s sufficient to estimate the denominator of $H_{n+m}(c_{1})$. And since $H_{n+m}(c_{1})=\sum_{k=1}^{c_{1}}\frac{1}{k^{n+m}}$ the denominator of $H_{n+m}(c_{1})$ should be a divisor of $lcm(1,...,c_{1})^{n+m}$. Therefore if we represent $I_{m}(a_{1},...,a_{n})$ as $\frac{p_{1}+p_{2}\zeta(m+2)+...+p_{n}\zeta(m+n)}{q}$ under the condition of $c_{1}=a_{1}=a_{2}=...=a_{n}$, then $q|lcm(1,...,c_{1})^{n+m}$ Case II, if $1<r\leq n$, then it follows from the result of preceding theorem $I_{m}(a_{1},...,a_{n})=\sum_{i=1}^{r-1}\mu_{i1}(H_{m+1}(c_{r})-H_{m+1}(c_{i}))+\sum_{i=1}^{r}\sum_{j\geq 2}\binom{m+j-1}{m}\mu_{ij}\zeta(j+m,c_{i}+1)$ Reformulate $\zeta(j+m,c_{i}+1)$ as $\zeta(j+m)-H_{j+m}(c_{i})$, then we obtain $\displaystyle I_{m}(a_{1},...,a_{n})$ (4) $\displaystyle=$ $\displaystyle\sum_{i=1}^{r-1}\mu_{i1}(H_{m+1}(c_{r})-H_{m+1}(c_{i}))-\sum_{i=1}^{r}\sum_{j\geq 2}\binom{m+j-1}{m}\mu_{ij}H_{j+m}(c_{i})$ (5) $\displaystyle+\sum_{i=1}^{r}\sum_{j\geq 2}\binom{m+j-1}{m}\mu_{ij}\zeta(j+m)$ (6) $\displaystyle=$ $\displaystyle\sum_{i=1}^{r-1}\frac{N_{1,i}}{\delta(\mu_{i1})\delta(H_{m+1}(c_{r})-H_{m+1}(c_{i}))}-\sum_{i=1}^{r}\sum_{j\geq 2}\frac{N_{2,ij}}{\delta(\mu_{ij})\delta(H_{j+m}(c_{i}))}$ (7) $\displaystyle+\sum_{i=1}^{r}\sum_{j\geq 2}\frac{N_{3,ij}\zeta(j+m)}{\delta(\mu_{ij})}$ (8) where $N_{1,i},N_{2,ij},N_{3,ij}\in\mathbb{Z}$. In following we divide the proof in three steps: Firstly, to prove that there are integers $D_{i}$ such that both $\delta(\mu_{i1})$ and $\delta(\mu_{ij})$ are divisors of $D_{i}$. Secondly, to prove that there is an integer $D$ such that both $\delta((H_{m+1}(c_{r})-H_{m+1}(c_{i})))$ and $\delta(H_{j+m}(c_{i}))$ are divisors of of $D$. Finally, by showing that $q|D\cdot lcm(D_{1},...,D_{r})$ to find the estimate that we needed. STEP 1 Let $\prod_{i=1}^{r}\frac{1}{(c_{i}+x)^{b_{i}}}=\sum_{i=1}^{r}\sum_{j=1}^{b_{i}}\frac{\mu_{ij}}{(c_{i}+x)^{j}}$ By the Lemma2, we have the expression of $\mu_{ij}$ as follow $\mu_{ij}=\frac{(-1)^{j-1}}{(b_{i}-j)!}\frac{\partial^{b_{i}-j}}{\partial z^{b_{i}-j}}|_{z=c_{i}}\prod_{\ell=1,\ell\neq i}^{r}\frac{1}{(c_{\ell}-z)^{b_{\ell}}}$ For simplicity, we may let $A_{\ell}=\begin{cases}c_{\ell}\text{, if }\ell<i\\\ c_{\ell+1}\text{, if }\ell\geq i\end{cases}$ $B_{\ell}=\begin{cases}b_{\ell}\text{, if }\ell<i\\\ b_{\ell+1}\text{, if }\ell\geq i\end{cases}$ then $\prod_{\ell=1,\ell\neq i}^{r}\frac{1}{(c_{\ell}-z)^{b_{\ell}}}=\prod_{\ell=1}^{r-1}\frac{1}{(A_{\ell}-z)^{B_{\ell}}}$ Let $M=b_{i}-j$ and $0\leq M_{1},...,M_{r-1}\leq M$ be integers. If we denote $F(z)=\frac{\partial^{M}}{\partial z^{M}}\prod_{\ell=1,\ell\neq i}^{r-1}\frac{1}{(c_{\ell}-z)^{b_{\ell}}}$ $\displaystyle F(z)$ $\displaystyle=\sum_{M_{1}+...+M_{r-1}=M}\binom{M}{M_{1},...,M_{r-1}}\prod_{\ell=1}^{r-1}(\frac{1}{(A_{\ell}-z)^{B_{\ell}}})^{(M_{\ell})}$ $\displaystyle=\sum_{M_{1}+...+M_{r-1}=M}\binom{M}{M_{1},...,M_{r-1}}\prod_{\ell=1}^{r-1}\frac{(B_{\ell}+M_{\ell}-1)!}{(B_{\ell}-1)!}\frac{1}{(A_{\ell}-z)^{B_{\ell}+M_{\ell}}}$ Note that $\binom{M}{M_{1},...,M_{r-1}}\in\mathbb{Z}$, $\frac{(B_{\ell}+M_{\ell}-1)!}{(B_{\ell}-1)!}\in\mathbb{Z}$ for all $1\leq\ell\leq r-1$. then $F(z)\cdot\prod_{\ell=1}^{r-1}(A_{\ell}-z)^{B_{\ell}+M}$ should be a polynomial of $z$ with integer coefficients. This implies that the denominator of $F(c_{i})$ is a divisor of $\prod_{\ell=1}^{r-1}(A_{\ell}-c_{i})^{B_{\ell}+M}$. In other words $\delta(F(c_{i}))|\prod_{\ell=1,\ell\neq i}^{r}(c_{\ell}-c_{i})^{b_{\ell}+b_{i}-j}$ Because of $\mu_{ij}=\frac{(-1)^{j-1}}{(b_{i}-j)!}F(c_{i})$, therefore $\delta(\mu_{ij})|(b_{i}-j)!\prod_{\ell=1,\ell\neq i}^{r}(c_{\ell}-c_{i})^{b_{\ell}+b_{i}-j}$ As a special case, $\delta(\mu_{i1})|(b_{i}-1)!\prod_{\ell=1,\ell\neq i}^{r}(c_{\ell}-c_{i})^{b_{\ell}+b_{i}-1}$ It’s easy to check for $j\geq 1$ $\displaystyle(b_{i}-j)!|(b_{i}-1)!$ $\displaystyle(c_{\ell}-c_{i})^{b_{\ell}+b_{i}-j}|(c_{\ell}-c_{i})^{b_{\ell}+b_{i}-1}$ This implies that $\delta(\mu_{ij})|(b_{i}-1)!\prod_{\ell=1,\ell\neq i}^{r}(c_{\ell}-c_{i})^{b_{\ell}+b_{i}-1}$ Now denote $(b_{i}-1)!\prod_{\ell=1,\ell\neq i}^{r}(c_{\ell}-c_{i})^{b_{\ell}+b_{i}-1}$ by $D_{i}$, thus $\delta(\mu_{ij})|D_{i}$ for all $j$. STEP 2 By the expression of $H_{m+1}(x)$ it’s obvious to see that, $\delta(H_{m+1}(c_{r})-H_{m+1}(c_{i}))|lcm(c_{i}+1,...,c_{r})^{m+1}$ On the one hand, since $c_{1}<...<c_{r}$, this gives following is true for all $i\geq 1$ $lcm(c_{i}+1,...,c_{r})^{m+1}|lcm(c_{1}+1,...,c_{r})^{m+1}$ Hence $\delta(H_{m+1}(c_{r})-H_{m+1}(c_{i}))|lcm(c_{1}+1,...,c_{r})^{m+1}$ On the other hand, $\delta(H_{m+j}(c_{i}))|lcm(1,...,c_{i})^{m+j}$ and since $c_{1}<...<c_{r}$, this gives for all $i\geq 1$ $lcm(1,...,c_{i})^{m+j}|lcm(1,...,c_{r})^{m+j}$ Hence $\delta(H_{m+j}(c_{i}))|lcm(1,...,c_{r})^{m+j}$ Now let $D=lcm(1,...,c_{r})^{m+b_{+}}$, where $b_{+}=\max\\{b_{1},...,b_{r}\\}$, we have both $\delta(H_{m+1}(c_{r})-H_{m+1}(c_{i}))$ and $\delta(H_{m+j}(c_{i}))$ are divisors of $D$. STEP 3 Observe (4) and rewrite it as $\displaystyle I_{m}(a_{1},...,a_{n})$ $\displaystyle=$ $\displaystyle\sum_{i=1}^{r-1}\frac{N_{1,i}}{\delta(\mu_{i1})\delta(H_{m+1}(c_{r})-H_{m+1}(c_{i}))}-\sum_{i=1}^{r}\sum_{j\geq 2}\frac{N_{2,ij}}{\delta(\mu_{ij})\delta(H_{j+m}(c_{i}))}$ $\displaystyle+\sum_{i=1}^{r}\sum_{j\geq 2}\frac{N_{3,ij}\zeta(j+m)}{\delta(\mu_{ij})}$ By the result of Step 2, now multiplying $D=lcm(1,...,c_{r})^{m+b_{+}}$ on both sides, we have $D\cdot I_{m}(a_{1},...,a_{n})=\sum_{i=1}^{r-1}\frac{N^{\prime}_{1,i}}{\delta(\mu_{i1})}-\sum_{i=1}^{r}\sum_{j\geq 2}\frac{N^{\prime}_{2,ij}}{\delta(\mu_{ij})}+\sum_{i=1}^{r}\sum_{j\geq 2}\frac{N^{\prime}_{3,ij}\zeta(j+m)}{\delta(\mu_{ij})}$ Because $\delta(\mu_{ij})|D_{i}$ for all $i,j$, by multiplying $lcm(D_{1},...,D_{r})$ on both sides, we have $D\cdot lcm(D_{1},...,D_{r-1})I_{m}(a_{1},...,a_{n})=N^{\prime\prime}_{1}+N^{\prime\prime}_{2}\zeta(m+2)+...+N^{\prime\prime}_{b_{+}}\zeta(m+b_{+})$ That is $q|D\cdot lcm(D_{1},...,D_{r})$. Finally, let $D_{0}=(b_{+}-1)!\prod_{1\leq s<t\leq r}(c_{t}-c_{s})^{n-1}$ then $lcm(D_{1},...,D_{r})|D_{0}$. It’s easy to see $(b_{i}-1)!|(b_{+}-1)!$ And by $b_{1}+...+b_{r}=n$ we have $\prod_{\ell=1,\ell\neq i}^{r}(c_{\ell}-c_{i})^{b_{\ell}+b_{i}-1}|\prod_{1\leq s<t\leq r}(c_{t}-c_{s})^{n-1}$ Now we can give the estimate of $q$ as $q|(b_{+}-1)!\cdot lcm(1,...,c_{r})^{m+b_{+}}\prod_{1\leq s<t\leq r}(c_{t}-c_{s})^{n-1}$ That is what we need. ∎ ## 4 Estimates of the Rational Approximation of $\zeta(5)$ In order to prove $\zeta(3)$ is irrational, the key is to find a parametric representation of $\zeta(3)$ and to construct an effective rational approximation. This rational approximation is related to the Legendre-type polynomial. In the last section we have discussed the generalized Beukers integral. On the one hand, it provides a parametric representation of $\zeta(2n+1)$, on the other hand, such generalization makes it possible to construct rational approximation of $\zeta(2n+1)$. As a special case, by using the Legendre-type polynomials to find a approximation of $\zeta(5)$ is the most obvious way trying to prove the irrationality of $\zeta(5)$. But unfortunately, this approximation is not as effective as the case of $\zeta(3)$. In this section, we prove this result. Before showing the proof, we firstly give two lemmas. Through out this section, $1-xy$ is denoted by $f$, $1-s$ is denoted by $\overline{s}$, $1-r$ is denoted by $\overline{r}$ etc. More specifically, by theorem 1 we can construct a integral $I(a,b)$ for nonnegative integer $a,b$, such that $I(a,b)=\begin{cases}q_{0}\zeta(5)+q_{1}\text{, if }a=b\\\ q_{2}\text{, if }a\neq b\end{cases}$ where $q_{0},q_{1},q_{2}\in\mathbb{Q}$. It turns out that if we let $Q_{n}(x),Q_{n}(y)$ be polynomials of $x$ and $y$ respectively with integer coefficients and degree $n$, then $-\int_{(0,1)^{2}}\frac{\log^{3}(xy)Q_{n}(x)Q_{n}(y)}{1-xy}dxdy=\alpha_{n}\zeta(5)+\beta_{n}$ where $\alpha_{n},\beta_{n}\in\mathbb{Q}$. That is, we found a parametric representation of $\zeta(5)$. By letting $Q_{n}$ be the Legendre-type polynomial, which denoted by $P_{n}$ here, namely, $P_{n}(x):=\frac{1}{n!}\frac{d^{n}}{dx^{n}}(x(1-x))^{n}$, we are able to construct a rational approximation of $\zeta(5)$. Let $J_{3}(n):=-\int_{(0,1)^{2}}\frac{\log^{3}(xy)P_{n}(x)P_{n}(y)}{1-xy}dxdy$, then according to theorem 1, we have $J_{3}(n)=\frac{A_{n}\zeta(5)+B_{n}}{d_{n}^{5}}$, where $A_{n},B_{n}\in\mathbb{Z}$, $d_{n}=lcm(1,...,n)$. In following we prove that $\frac{6}{(n+1)^{4}}\leq J_{3}(n)\leq\frac{6\pi^{2}}{(n+\frac{1}{2})^{2}}$. Due to $d_{n}^{5}\frac{6}{(n+1)^{4}}>1$ for all sufficiently large $n$, we are not able to show the irrationality of $\zeta(5)$. ###### Lemma 3. For any integer $m\geq 2$, following inequality is true for all $x\in(0,+\infty)$. Moreover, the equations hold if and only if $x=1$. $m(1-\frac{1}{\sqrt[m]{x}})\leq\log(x)\leq m(\sqrt[m]{x}-1)$ ###### Proof. The proof is divided into two parts. I. Define $g(x):=\log(x)-m(\sqrt[m]{x}-1)$. Obviously $g(1)=1$ and $g^{\prime}(x)=\frac{1}{x}-\frac{x^{\frac{1}{m}}}{x}=\frac{1-x^{\frac{1}{m}}}{x}$ If $x\in(0,1)$, then $g^{\prime}(x)>0$. If $x\in(1,\infty)$ then $g^{\prime}(x)<0$. Therefore $g(x)$ is strictly monotonically increasing from negative number to $0$ on $(0,1)$, strictly monotonically decreasing from $0$ to negative number on $(1,+\infty)$. This shows $\log(x)\leq m(\sqrt[m]{x}-1)$. The two sides are equal if and only if $x=1$. II. Likewise we define $g(x):=\log(x)-m(1-x^{-\frac{1}{m}})$. Observe that $g(1)=1$ and $g^{\prime}(x)=\frac{1}{x}-\frac{1}{x^{1+\frac{1}{m}}}=\frac{x^{\frac{1}{m}}-1}{x^{1+\frac{1}{m}}}$ If $x\in(0,1)$, then $g^{\prime}(x)<0$. If $x\in(1,\infty)$ then $g^{\prime}(x)>0$. Therefore $g(x)$ is strictly monotonically decreasing from positive number to $0$ on $(0,1)$, strictly monotonically increasing from $0$ to positive number on $(1,+\infty)$. This shows $m(1-\frac{1}{\sqrt[m]{x}})\leq\log(x)$. The two sides are equal if and only if $x=1$. ∎ ###### Lemma 4. (Canonical transform) Define $L(a,b;n+1):=\int_{0}^{1}\frac{s^{a}\overline{s}^{b}}{(1-fs)^{n+1}}ds$ then the equality is valid $L(a,b;n+1)=(1-f)^{b-n}L(b,a;a+b+1-n)$ ###### Proof. Substitute $s$ by $\frac{1-r}{1-fr}$, then $1-s=\frac{r(1-f)}{1-fr}$, $1-fs=\frac{1-f}{1-fr}$ and $ds=-\frac{1-f}{(1-fr)^{2}}dr$. If $s=0$, then $r=1$, and if $s=1$, then $r=0$. Then $\displaystyle\int_{0}^{1}\frac{s^{a}\overline{s}^{b}}{(1-fs)^{n+1}}ds$ $\displaystyle=-\int_{1}^{0}(\frac{1-r}{1-fr})^{a}(\frac{r(1-f)}{1-fr})^{b}(\frac{1-f}{1-fr})^{-n-1}\frac{1-f}{(1-fr)^{2}}dr$ $\displaystyle=(1-f)^{b-n}\int_{0}^{1}\frac{r^{b}\overline{r}^{a}}{(1-fr)^{a+b+1-n}}dr$ This is what we need. For convenience, this transform is called the canonical transform. ∎ ###### Lemma 5. Assume that $\displaystyle J_{3}(n):=-\int_{(0,1)^{2}}\frac{\log^{3}(xy)P_{n}(x)P_{n}(y)}{1-xy}dxdy$ $\displaystyle R_{2}(n)=\int_{(0,1)^{4}}\frac{x^{n}\overline{x}^{n}y^{n}\overline{y}^{n}s^{n}\overline{u}^{n}}{(1-fs)^{n+1}}\frac{\log(\frac{s\overline{u}}{u\overline{s}})}{s-u}dxdydsdu$ then the equality $J_{3}(n)=6R_{2}(n)$ is valid for all $n\in\mathbb{Z}^{+}$. ###### Proof. Recall that $f:=1-xy$, since $-\frac{\log(1-f)}{f}=\int_{0}^{1}\frac{1}{1-fz}dz$, we can rewrite $J_{3}(n)$ as following, $\displaystyle J_{3}(n)$ $\displaystyle=-\int_{(0,1)^{2}}\frac{\log^{3}(xy)P_{n}(x)P_{n}(y)}{1-xy}dxdy$ $\displaystyle=-\int_{(0,1)^{2}}\frac{\log^{3}(1-f)}{f^{3}}f^{2}P_{n}(x)P_{n}(y)dxdy$ $\displaystyle=\int_{(0,1)^{5}}\frac{f^{2}P_{n}(x)P_{n}(y)}{(1-fz_{1})(1-fz_{2})(1-fz_{3})}dxdydz_{1}dz_{2}dz_{3}$ By the partial fraction decomposition $\displaystyle\frac{f^{2}}{(1-fz_{1})(1-fz_{2})(1-fz_{3})}$ $\displaystyle=$ $\displaystyle\frac{1}{(z_{2}-z_{i})(z_{3}-z_{1})}\frac{1}{1-fz_{1}}+\frac{1}{(z_{1}-z_{2})(z_{3}-z_{2})}\frac{1}{1-fz_{2}}+\frac{1}{(z_{1}-z_{3})(z_{2}-z_{3})}\frac{1}{1-fz_{3}}$ we obtain $J_{3}(n)=Q_{1}(n)+Q_{2}(n)+Q_{3}(n)$ where $\displaystyle Q_{1}(n)$ $\displaystyle=\int_{(0,1)^{5}}\frac{P_{n}(x)P_{n}(y)}{(1-fz_{1})(z_{2}-z_{1})(z_{3}-z_{1})}dxdydz_{1}dz_{2}dz_{3}$ $\displaystyle Q_{2}(n)$ $\displaystyle=\int_{(0,1)^{5}}\frac{P_{n}(x)P_{n}(y)}{(1-fz_{2})(z_{1}-z_{2})(z_{3}-z_{2})}dxdydz_{1}dz_{2}dz_{3}$ $\displaystyle Q_{3}(n)$ $\displaystyle=\int_{(0,1)^{5}}\frac{P_{n}(x)P_{n}(y)}{(1-fz_{3})(z_{1}-z_{3})(z_{2}-z_{3})}dxdydz_{1}dz_{2}dz_{3}$ It’s easy to see that $Q_{1}(n)=Q_{2}(n)=Q_{3}(n)$, namely $J_{3}(n)=3Q_{1}(n)$. Hence it’s sufficient to deal with $Q_{1}(n)$. For $Q_{1}(n)$, after taking n-fold partial integration with respect to $x$, we have $Q_{1}(n)=\int_{(0,1)^{5}}\frac{(xyz_{1})^{n}(1-x)^{n}P_{n}(y)}{(1-fz_{1})^{n+1}(z_{2}-z_{1})(z_{3}-z_{1})}dxdydz_{1}dz_{2}dz_{3}$ Now substitute $\frac{1-z_{i}}{1-fz_{i}}$ by $w_{i}$ for $i=1,2,3$ and by straightforward verification of following I, $z_{i}=\frac{1-w_{i}}{1-fw_{i}},\text{and }z_{i}=0\Leftrightarrow w_{i}=1,z_{i}=1\Leftrightarrow w_{i}=0$ II, $dz_{i}=\frac{f-1}{(1-fw_{i})^{2}}dw_{i}\\\ $ III, if $k=1,2,3$ and $k\neq i$, then $z_{k}-z_{i}=\frac{1-w_{k}}{1-fw_{k}}-\frac{1-w_{i}}{1-fw_{i}}=\frac{(f-1)(w_{k}-w_{i})}{(1-fw_{k})(1-fw_{i})}$ IV, $\frac{z_{1}^{n}}{(1-fz_{1})^{n+1}}=\frac{(1-fw_{1})(1-w_{1})^{n}}{(1-f)^{n+1}}$ we have $\displaystyle Q_{1}(n)$ $\displaystyle=\int_{(0,1)^{5}}\frac{(xyz_{1})^{n}(1-x)^{n}P_{n}(y)}{(1-fz_{1})^{n+1}(z_{2}-z_{1})(z_{3}-z_{1})}dxdydz_{1}dz_{2}dz_{3}$ $\displaystyle=\int_{(0,1)^{5}}\frac{x^{n}(1-x)^{n}y^{n}P_{n}(y)(1-fw_{1})(1-w_{1})^{n}}{(1-f)^{n}(1-fw_{2})(1-fw_{3})(w_{2}-w_{1})(w_{3}-w_{1})}dxdydw_{1}dw_{2}dw_{3}$ recall that $1-f=1-(1-xy)=xy$, thus $Q_{1}(n)=\int_{(0,1)^{5}}(1-x)^{n}(1-w_{1})^{n}P_{n}(y)\frac{(1-fw_{1})}{(1-fw_{2})(1-fw_{3})(w_{2}-w_{1})(w_{3}-w_{1})}dxdydw_{1}dw_{2}dw_{3}\\\ $ Once again using the partial fraction decomosition $\frac{(1-fw_{1})}{(1-fw_{2})(1-fw_{3})}=\frac{w_{1}-w_{2}}{w_{3}-w_{2}}\frac{1}{1-fw_{2}}+\frac{w_{3}-w_{1}}{w_{3}-w_{2}}\frac{1}{1-fw_{3}}$ Then $Q_{1}(n)=R_{1}(n)+R_{2}(n)$, where $\displaystyle R_{1}(n)$ $\displaystyle=-\int_{(0,1)^{5}}\frac{(1-x)^{n}(1-w_{1})^{n}P_{n}(y)}{(1-fw_{2})(w_{3}-w_{2})(w_{3}-w_{1})}dxdydw_{1}dw_{2}dw_{3}$ $\displaystyle R_{2}(n)$ $\displaystyle=\int_{(0,1)^{5}}\frac{(1-x)^{n}(1-w_{1})^{n}P_{n}(y)}{(1-fw_{3})(w_{3}-w_{2})(w_{2}-w_{1})}dxdydw_{1}dw_{2}dw_{3}$ Notice that actually $R_{1}(n)$ and $R_{2}(n)$ are the same, therefore $Q_{1}(n)=2R_{2}(n)$. It’s sufficient to compute $R_{2}(n)$. For convenience, substituting $w_{3},w_{2},w_{1}$ by $s,t,u$ respectively, i.e. $R_{2}(n)=\int_{(0,1)^{5}}\frac{(1-x)^{n}(1-u)^{n}P_{n}(y)}{(1-fs)^{n+1}(s-t)(t-u)}dxdydsdtdu\\\ $ After n-fold partial integration with respect to $y$ for $R_{2}(n)$, we have $R_{2}(n)=\int_{(0,1)^{5}}\frac{x^{n}(1-x)^{n}y^{n}(1-y)^{n}s^{n}(1-u)^{n}}{(1-fs)^{n+1}(s-t)(t-u)}dxdydsdtdu$ ∎ Note that if $s\neq u$, $\int_{0}^{1}\frac{1}{(s-t)(t-u)}dt=\frac{\log(\frac{s}{1-s})-\log(\frac{u}{1-u})}{s-u}=\frac{\log(\frac{s(1-u)}{u(1-s)})}{s-u}$ If $s>u$, then $\log(s)-\log(u)>0$ and $\log(1-u)>\log(1-s)$, therefore $\frac{\log(\frac{s(1-u)}{u(1-s)})}{s-u}>0$. If $u>s$, $\frac{\log(\frac{s(1-u)}{u(1-s)})}{s-u}=\frac{\log(\frac{u(1-s)}{s(1-u)})}{u-s}>0$. That is if $s\neq u$, $\frac{\log(\frac{s(1-u)}{u(1-s)})}{s-u}>0$. Now we can see $R_{2}(n)>0$, and $R_{2}(n)=\int_{(0,1)^{4}}\frac{x^{n}\overline{x}^{n}y^{n}\overline{y}^{n}s^{n}\overline{u}^{n}}{(1-fs)^{n+1}}\frac{\log(\frac{s\overline{u}}{u\overline{s}})}{s-u}dxdydsdu$ Since $J_{3}(n)=3Q_{1}(n)=6R_{2}(n)$. This is what we need to prove. ###### Theorem 6. For all integer $n\geq 1$, following inequalities are true. $\frac{6}{(n+1)^{4}}\leq J_{3}(n)\leq\frac{6\pi^{2}}{(n+\frac{1}{2})^{2}}$ ###### Proof. The proof is divided into two parts I. Firstly we give the upper bound of $J_{3}(n)$. In the preceding Lemma we proved that $J_{3}(n)=6R_{2}(n)$, where $R_{2}(n)=\int_{(0,1)^{4}}\frac{(x\overline{x}y\overline{y}s\overline{u})^{n}}{(1-fs)^{n+1}}\frac{\log(\frac{s\overline{u}}{\overline{s}u})}{s-u}dxdydsdu$ Now apply the Lemma 3 we obtain for any positive integer $m\geq 2$ $\displaystyle\frac{\log(\frac{s\overline{u}}{\overline{s}u})}{s-u}$ $\displaystyle\leq m(\sqrt[m]{\frac{s\overline{u}}{\overline{s}u}}-1)/(s-u)\leq m\frac{\sqrt[m]{s\overline{u}}-\sqrt[m]{\overline{s}u}}{\sqrt[m]{\overline{s}u}(s\overline{u}-\overline{s}u)}$ Note that $\frac{s\overline{u}-\overline{s}u}{(s\overline{u})^{\frac{1}{m}}-(\overline{s}u)^{\frac{1}{m}}}=\sum_{k=0}^{m-1}(s\overline{u})^{\frac{m-1-k}{m}}(\overline{s}u)^{\frac{k}{m}}$ we apply the inequality of arithmetic and geometric means, then $\sum_{k=0}^{m-1}(s\overline{u})^{\frac{m-1-k}{m}}(\overline{s}u)^{\frac{k}{m}}\geq m(s\overline{s}u\overline{u})^{\frac{m(m-1)}{2m}\frac{1}{m}}=m(s\overline{s}u\overline{u})^{\frac{(m-1)}{2m}}$ Therefore $\frac{\log(\frac{s\overline{u}}{\overline{s}u})}{s-u}\leq\frac{1}{(s\overline{u})^{\frac{1}{2}-\frac{1}{2m}}(\overline{s}u)^{\frac{1}{2}+\frac{1}{2m}}}$ It turns out that $R_{2}(n)\leq\int_{(0,1)^{4}}\frac{(x\overline{x}y\overline{y})^{n}}{(1-fs)^{n+1}}(s)^{n-\frac{1}{2}+\frac{1}{2m}}(\overline{s})^{-\frac{1}{2}-\frac{1}{2m}}(\overline{u})^{n-\frac{1}{2}+\frac{1}{2m}}(u)^{-\frac{1}{2}-\frac{1}{2m}}dxdydsdu$ On the one hand, $\int_{0}^{1}\overline{u}^{n-\frac{1}{2}+\frac{1}{2m}}u^{-\frac{1}{2}-\frac{1}{2m}}du=B(n+\frac{1}{2m}+\frac{1}{2},-\frac{1}{2m}+\frac{1}{2})$ On the other hand, by the canonical transform (Lemma 4) we obtain $\displaystyle\int_{(0,1)^{3}}\frac{(x\overline{x}y\overline{y})^{n}}{(1-fs)^{n+1}}(s)^{n-\frac{1}{2}+\frac{1}{2m}}(\overline{s})^{-\frac{1}{2}-\frac{1}{2m}}dxdyds$ $\displaystyle=$ $\displaystyle\int_{(0,1)^{3}}(1-f)^{-\frac{1}{2}-\frac{1}{2m}-n}(x\overline{x}y\overline{y})^{n}z^{-\frac{1}{2}-\frac{1}{2m}}\overline{z}^{n-\frac{1}{2}+\frac{1}{2m}}dxdydz$ $\displaystyle=$ $\displaystyle\int_{(0,1)^{3}}x^{-\frac{1}{2}-\frac{1}{2m}}\overline{x}^{n}y^{-\frac{1}{2}-\frac{1}{2m}}\overline{y}^{n}z^{-\frac{1}{2}-\frac{1}{2m}}\overline{z}^{n-\frac{1}{2}+\frac{1}{2m}}dxdydz$ $\displaystyle=$ $\displaystyle(B(n+1,\frac{1}{2}-\frac{1}{2m}))^{2}B(n+\frac{1}{2}+\frac{1}{2m},\frac{1}{2}-\frac{1}{2m})$ Therefore $R_{2}(n)\leq(B(n+1,\frac{1}{2}-\frac{1}{2m})B(n+\frac{1}{2}+\frac{1}{2m},\frac{1}{2}-\frac{1}{2m}))^{2}$. Moreover, both $B(n+1,\frac{1}{2}-\frac{1}{2m})$ and $B(n+\frac{1}{2}+\frac{1}{2m},\frac{1}{2}-\frac{1}{2m})$ are decreasing with $m$ increasing. Let $m\rightarrow\infty$, that is $\displaystyle\lim_{m\rightarrow\infty}(B(n+1,\frac{1}{2}-\frac{1}{2m})B(n+\frac{1}{2}+\frac{1}{2m},\frac{1}{2}-\frac{1}{2m}))^{2}$ $\displaystyle=$ $\displaystyle(B(n+1,\frac{1}{2})B(n+\frac{1}{2},\frac{1}{2}))^{2}$ $\displaystyle=$ $\displaystyle(\frac{\Gamma(n+1)\Gamma(\frac{1}{2})}{\Gamma(n+\frac{3}{2})}\frac{\Gamma(n+\frac{1}{2})\Gamma(\frac{1}{2})}{\Gamma(n+1)})^{2}$ $\displaystyle=$ $\displaystyle\frac{\pi^{2}}{(n+\frac{1}{2})^{2}}$ Therefore $J_{3}(n)\leq\frac{6\pi^{2}}{(n+\frac{1}{2})^{2}}$. II. In this part we give the lower bound of $J_{3}(n)$. By the Lemma 3, $\frac{\log(\frac{s\overline{u}}{\overline{s}u})}{s-u}\geq m\frac{1-(\frac{s\overline{u}}{\overline{s}u})^{-\frac{1}{m}}}{s\overline{u}-{\overline{s}u}}=m\frac{(s\overline{u})^{\frac{1}{m}}-(\overline{s}u)^{\frac{1}{m}}}{(s\overline{u})^{\frac{1}{m}}(s\overline{u}-{\overline{s}u})}$ Likewise $\frac{s\overline{u}-\overline{s}u}{(s\overline{u})^{\frac{1}{m}}-(\overline{s}u)^{\frac{1}{m}}}=\sum_{k=0}^{m-1}(s\overline{u})^{\frac{m-1-k}{m}}(\overline{s}u)^{\frac{k}{m}}$ If $s\overline{u}>\overline{s}u$, $\sum_{k=0}^{m-1}(s\overline{u})^{\frac{m-1-k}{m}}(\overline{s}u)^{\frac{k}{m}}<m(s\overline{u})^{\frac{m-1}{m}}<m$. If $s\overline{u}<\overline{s}u$, $\sum_{k=0}^{m-1}(s\overline{u})^{\frac{m-1-k}{m}}(\overline{s}u)^{\frac{k}{m}}<m(\overline{s}u)^{\frac{m-1}{m}}<m$ as well. So $\frac{\log(\frac{s\overline{u}}{\overline{s}u})}{s-u}>\frac{m}{m(s\overline{u})^{\frac{1}{m}}}=(s\overline{u})^{-\frac{1}{m}}$ Now come back to $R_{2}(n)$, $R_{2}(n)>\int_{(0,1)^{4}}\frac{(x\overline{x}y\overline{y})^{n}}{(1-fs)^{n+1}}(s)^{n-\frac{1}{m}}\overline{u}^{n-\frac{1}{m}}dxdydsdu$ Obviously on the one hand, $\int_{0}^{1}\overline{u}^{n-\frac{1}{m}}du=\frac{1}{n-\frac{1}{m}+1}$ On the other hand, with the canonical transform we obtain $\displaystyle\int_{(0,1)^{3}}\frac{(x\overline{x}y\overline{y})^{n}}{(1-fs)^{n+1}}s^{n-\frac{1}{m}}dxdyds$ $\displaystyle=$ $\displaystyle\int_{(0,1)^{3}}\frac{(x\overline{x}y\overline{y})^{n}\overline{z}^{n-\frac{1}{m}}}{(1-f)^{n}(1-fz)^{1-\frac{1}{m}}}dxdydz$ $\displaystyle\geq$ $\displaystyle\int_{(0,1)^{3}}\overline{x}^{n}\overline{y}^{n}\overline{z}^{n-\frac{1}{m}}dxdydz$ $\displaystyle=$ $\displaystyle\frac{1}{(n+1)^{2}}\frac{1}{(n-\frac{1}{m}+1)}$ Therefore $R_{2}(n)\geq\frac{1}{(n+1)^{2}(n-\frac{1}{m}+1)^{2}}$. Likewise, taking the limit we have $J_{3}(n)\geq\frac{6}{(n+1)^{4}}$. Therefore, finally $\frac{6}{(n+1)^{4}}\leq J_{3}(n)\leq\frac{6\pi^{2}}{(n+\frac{1}{2})^{2}}$ ∎ Hence we have the conclusion: as $n$ tends to infinity, although $J_{3}(n)\rightarrow 0$, $lcm(1,2,...,n)^{5}\cdot J_{3}(n)\rightarrow\infty$. That is, compare to the rational approximation of $\zeta(3)$ (see [2]), the approximation of $\zeta(5)$ by generalized Beukers’ method is too slow. One has to looking for another method to prove the irrationality of $\zeta(5)$. Following Table gives some numerical comparison. The numerical comparison of upper bound and lower bound of $J_{3}(n)$. $n$ | $\frac{6}{(n+1)^{4}}$ | $J_{3}(n)$ | $\frac{6\pi^{2}}{(n+\frac{1}{2})^{2}}$ ---|---|---|--- $1$ | $0.3750$ | $4.4313$ | $26.3189$ $2$ | $0.0741$ | $0.9474$ | $9.4748$ $3$ | $0.0234$ | $0.2996$ | $4.8341$ $4$ | $0.0096$ | $0.1237$ | $2.9243$ $5$ | $0.0046$ | $0.0605$ | $1.9576$ $6$ | $0.0025$ | $0.0332$ | $1.4016$ $7$ | $0.0015$ | $0.0198$ | $1.0528$ $8$ | $0.0009$ | $0.0182$ | $0.8196$ $9$ | $0.0006$ | $0.0126$ | $0.6562$ $10$ | $0.0004$ | $0.0058$ | $0.5371$ ## 5 Acknowledgement Ich möchte mich bei den Leuten bedanken, die im 2012/2013 mich online verleumdet hatten. Diese ungerechte Worte sind mir noch deutlich errinnerlish. Diese ungerechte Worte gaben mir Antrieb und machten mir ununterbrochen weiterkommen. ## References * [1] Some Generalizations of Beukers’ Integrals, KYUNGPOOK Math J 42(2002), 399-416 * [2] A Note on the Irrationality of $\zeta(2)$ and $\zeta(3)$ Bull. London Math. Soc 11(1979), 268-272 Xiaowei Wang(王骁威) Institut für Mathematik, Universität Potsdam, Potsdam OT Golm, Germany Email<EMAIL_ADDRESS>
# Remote multi-user control of the production of Bose-Einstein condensates for research and education J S Laustsen, R Heck, O Elíasson, J J Arlt, J F Sherson and C A Weidner Department of Physics and Astronomy, Aarhus University, 8000 Aarhus C, Denmark <EMAIL_ADDRESS> ###### Abstract Remote control of experimental systems allows for improved collaboration between research groups as well as unique remote educational opportunities accessible by students and citizen scientists. Here, we describe an experiment for the production and investigation of ultracold quantum gases capable of asynchronous remote control by multiple remote users. This is enabled by a queuing system coupled to an interface that can be modified to suit the user, e.g. a gamified interface for use by the general public or a scripted interface for an expert. To demonstrate this, the laboratory was opened to remote experts and the general public. During the available time, remote users were given the task of optimising the production of a Bose-Einstein condensate (BEC). This work thus provides a stepping stone towards the exploration and realisation of more advanced physical models by remote experts, students and citizen scientists alike. * Keywords: Remote experiment control, ultracold atoms, BEC ## 1 Introduction Ultracold quantum gases have become one of the prime platforms for simulating technologically relevant quantum systems within the last decades. In particular, extremely clean and pure quantum model systems can be realised that offer a high degree of controllability with respect to parameters such as the atoms’ interaction strength and temperature. This progress has led to increasingly poweful and complex experiments in lattice-based quantum simulation [1, 2], the simulation of strongly-correlated condensed matter systems [3], and quantum computing with Rydberg atoms [4], among others, rendering cold atoms a fantastic platform for the development of technologies that will drive the second quantum revolution [5]. Collaboration between experimental and theoretical groups is an essential part of developing and evaluating models applicable to quantum simulation experiments. To optimise experimental procedures, it may indeed be beneficial to use dedicated, automated protocols developed by theory groups. Opening the laboratory to direct remote control by collaborators may thus increase the efficiency of such collaborative efforts. Moreover, a remote control system opens up new possibilities regarding outreach to students and the general public. By allowing a broad audience of non-expert users to control some experimental parameters, one can imagine a number of scenarios geared towards public outreach and education. First, the public can take part in citizen science experiments, and, in particular, previous work using the system described here shows that valuable insight into cognitive science can be gained [6]. Secondly such platforms can be used to educate and engage non- experts in quantum physics, e.g. by allowing students access to cutting-edge research laboratories regardless of where the laboratory is physically located. In both cases this creates the need for an intuitive user interface which allows users to focus on the essential parts and hides the technical details. At the same time, the experimental system must also contain the infrastructure to handle the input from one of many users and return the relevant results to the correct user. A number of these open platforms already exist, including the IBM Quantum Experience [7], and the open availability of this platform has allowed for the production of a number of research articles (see, e.g., Refs. [8, 9, 10, 11, 12]), educational material [13], and games [14]. In principle, any experimental control program can be modified for remote control via the addition of a remote server and a suitable front-end for the user. In terms of experimental control programs, several publicly available systems for cold atom experiments have been published [15, 16, 17, 18, 19]. In addition, numerous commercial options are available, such as systems by ColdQuanta [20], MLab [21], ADwin [22] and IQ Technologies [23] that can be purchased together with suitable hardware. All of these control systems have sub-microsecond control of digital and analog channels and some allow for direct output of radio frequency (RF) signals. Additionally, they typically allow for communication with external hardware through different protocols or via implementation of appropriate drivers. These criteria define the typical minimum viable product for useful cold-atom experiment control. Software for camera control and analysis of the images enables some systems to optimise experimental performance in a closed loop optimisation of experimental parameters. Moreover, all of these systems are remotely-controllable either directly or via simple screen-sharing protocols. However, to our knowledge, none of these control programs had been used in a multi-user setting where several users simultaneously remotely controlled an experiment through the use of the aforementioned server and front-end, with the exception that, while preparing this manuscript, we became aware of the Albert system built by ColdQuanta that came online in late 2020 [24]. This work thus represents the growing commercial and academic interest in remote control of cold atom systems. Here we discuss the implementation of a remote controlled experiment usable by single expert user or multiple non-expert users accessing the experiment. Previously, we have documented the challenges that we provided to our users, as well as the main findings that arose from this work [6]. However, we have not explained the underlying system architecture and the overarching possibilities that this gives rise to in research and education. The general knowledge of these details is crucial for other groups to implement similar systems, and this is what we focus on in this work. In both of the use-cases considered here, there is a need for a queuing system for the input sequences and the return of the results. When considering multi-user access there is also the need to track the sender throughout the process of queuing, performing the experiment, analysis and reporting the results. The infrastructure of the experiment also allows for multiple expert users and this option will be explored in future work. For instance, one could imagine running multiple collaborative efforts simultaneously. This paper is organised as follows: In the first section the software enabling remote control is presented. Following this, the experimental sequence and its technical implementation is described. We then describe the two different implementations of remote user access that were used in previous work: single- and multi-user control [6]. The last section concludes and provides an outlook. ## 2 The control software The experimental control system is LabVIEW-based and capable of being expanded as new hardware is added to the experiment. A field-programmable gate array (FPGA, PCI-7813R) is used to control $70$ digital and $48$ analog channels through $4$ digital to analog converter modules (NI 9264, NI 9263). In addition, the system can communicate with hardware drivers to other hardware such as motion stages, piezoelectric devices and RF synthesisers. Thus, our control system meets the aforementioned criteria for usability in a cold atom experiment. The control program is based on a server/client architecture. The server controls all hardware, including the FPGA, and the client provides an interface for the user and compiles the programmed sequence. On the client side the sequence is built of waves which correspond to the output of a given digital or analog channel, a GPIB command or a command through a hardware driver. Regularly-used sequences of waves can be collected and collapsed into blocks, e.g., the commands required to load atoms into a trap or image an atom cloud. For each wave and block, externally-accessible variables can be declared, e.g. the frequencies of the RF tones applied to acousto-optic modulators (AOMs) or the duration of the RF pulse applied to the AOM. This allows the user to create sequences with an adaptive level of abstraction. For instance one can hide the exact technical implementation of experimental steps in a block but keep the essential control parameters accessible, which is useful for reducing the cognitive load of a remote user. An example of a block used for absorption imaging of ultracold atoms is shown in Fig. 1, where smaller blocks are incorporated. The waves and blocks are ordered in a tree structure that controls the timing of an experiment. The tasks are performed from top to bottom in such a tree. Any waves or blocks on indented branches are performed simultaneously, and delays can be defined within individual elements for more precise control of relative timing. Initialised outputs may be defined such that they either hold their last value or are reset to a default value after a given time. Thus the user need only handle the values of relevant outputs at any given point. Wave and block variables can be scanned individually or jointly in single- and multi- dimensional scans, respectively. Loops are also available where a subset of blocks is repeated while one or several variables are changed. For example, the user can loop the capture of atoms in a trap while changing a given parameter value during each loop iteration, effectively performing a parameter scan within a single realisation of the experiment. The novel aspects of the control system lie in its capability for communication with remote users. This includes loading sequences from a queue either created by a single user or multiple different users. After a remotely- requested sequence is performed, relevant results (e.g. atom number) are sent back to the user who designed the sequence. To make the remote control as flexible as possible, the control software does not provide any user interface for the remote user but communicates with stand-alone interfaces. Thus a remote user can easily set up closed-loop optimisation by linking the returned results into a script running a given optimisation algorithm that then generates the next desired sequence, as described in detail below. ## 3 The experiment To demonstrate the use of the control system and the communication necessary for multi-user operation, we conducted an experiment in which remote users create a Bose-Einstein Condensate (BEC). The experimental system is described in Refs. [25, 26] and only its main features are described here. The experimental sequence starts by precooling a cloud of Rb-87 atoms in a 3D MOT. Here the atoms are laser-cooled and trapped via a combination of light pressure and magnetic field gradients. Subsequently polarisation gradient cooling is performed and the atoms are optically pumped to the low-field- seeking $|F\,,\,m_{F}\rangle=|2,2\rangle$ state. The atoms are then trapped in a magnetic quadrupole trap generated by a pair of coils in an anti-Helmholtz configuration. These coils are mounted on a rail and are used to transport the atoms through a differential pumping stage to the final chamber. Here the atoms are evaporatively cooled by transferring the hottest atoms to a high- field-seeking sublevel. By the end of the evaporation sequence the atoms have a temperature of roughly $30\text{\,}\mathrm{\SIUnitSymbolMicro K}$. Subsequently, a crossed optical dipole trap (CDT) consisting of two laser beams (wavelength $\lambda=$1064\text{\,}\mathrm{nm}$$, $1/e^{2}$ waists of $45\text{\,}\mathrm{\SIUnitSymbolMicro m}$ and $85\text{\,}\mathrm{\SIUnitSymbolMicro m}$) is superimposed on the atoms. After the final evaporation stage, the atom cloud is released from the trap and an absorption image is recorded after a TOF. If the user-defined evaporation sequence is effective, the cloud is condensed and a BEC is visible in the image. Figure 1: An example of a block used take an absorption image at the end of an experimental sequence (cf. Sec. 3). The block contains individual analog (A) and digital (D) waves (W), as well as two embedded blocks (B) used to take the absorption and background images. The block runs from the top to bottom with indented elements running in parallel with the element above. In this block, the imaging shutter is opened while the frequency applied to the imaging AOM is set and the CDT AOM is turned off (thus turning off the CDT itself and dropping the atoms from the trap). Then, after a variable time-of-flight (TOF) the next block (blue squares, with zoom-in to the right of the main block) simultaneously pulses the imaging AOM and triggers the camera shutter to take the absorption image of the atoms. After 300 ms of camera processing time, the same block takes a background image without atoms. The imaging shutter is closed, and after an additional 300 ms, the camera is triggered again to take the dark image without any light present. Subtracting the absorption image from the background and dark images reveals the atom signal. For the remote experiments reported here, the control parameters available to the users are the laser powers of both laser beams forming the CDT and the current in the quadrupole coils as a function of time. This configuration allows the user to cool the atoms using forced evaporative cooling either in a pure CDT [27], in a so-called hybrid trap consisting of the quadrupole magnetic field and one of the dipole beams [28], or any combination of the two. The depth and geometry of the trap depends on these parameters in a non- trivial optimal way, providing an opportunity for external optimisation, the goal of which is to produce the largest possible BEC. For both expert and non- expert users a limitation of the available control space is necessary as only a small fraction of the full control landscape will yield a BEC. ## 4 Two cases of remote user control For a remotely-controllable system to be useful, appropriate user interfaces must be developed, and each user class has different requirements to optimally facilitate the interaction. For experts a scripted interface can be an advantage as complex algorithms can be directly implemented. A more visual interface of the control (for instance in a game-like setting) is needed for non-expert users. Importantly, a different program structure is needed when handling input from a single user or multiple users. In what follows, we describe two different implementations of our remote control geared towards single expert users and asynchronous use among the general public, respectively. In this section, we elaborate on each of these cases. Again, note that the data presented here is drawn from the same source as our initial work [6], and detailed research results can be found there. Here, we focus on more technical aspects of the experimental implementation and execution. ### 4.1 Single-user remote control In our first implementation of remote control, an expert user optimised the evaporation using the so-called dressed chopped random-basis (dCRAB) optimisation algorithm [29]. Note that the algorithm was implemented on the user side, so our implementation of remote expert control is algorithm agnostic. Here the user had access to the CDT laser powers and quadrupole coil currents as a function of time. Sequences of waveforms corresponding to the parameter values were created as text files and sent to the experimental control program through a folder on a cloud drive and placed in a queue. Even for a single user, a queue is necessary due to the relatively long ($30$ s) cycle time of the experiment. The queue operated on the first-in-first-out (FIFO) principle, allowing the user to submit several sequences simultaneously and easily keep track of the outputs; this is useful, e.g., when initialising the initial simplex for Nelder-Mead optimisation. For each user-accessible parameter, the parameter values can be defined at any desired time, while values between these times are linearly interpolated at the hardware level. Therefore the effective temporal resolution of the waveforms can be controlled by the user, and the total number of parameter/time pairs that can be used is ultimately limited by the memory of our FPGA. When a given sequence was ready to be run, the relevant experimental sequences were generated by reading in the waveforms the from text files generated by the expert user. The experiment was then run and the resulting image was analysed. From this image the BEC atom number was extracted and returned to the expert user through the same cloud drive, again as a text file. This atom number was read in by the expert user and served as the cost parameter closing the optimisation loop. ### 4.2 Multi-user remote control Figure 2: A screenshot of the interface used in the _Alice Challenge_ , showing (a) the spline editor used to create the ramps of the laser powers and coil current, shown on a logarithmic scale, (b) top score list, (c) latest executed sequences, and (d) the control buttons, including the estimated wait time until the submitted sequence is returned. In the second implementation, called the Alice Challenge, citizen scientists were given access to the experiment. This subsection details the architectural considerations required for the challenge as well as some statistics on user load in real time over the course of the challenge. This information is useful when considering the future implementation of similar systems. Citizen scientists were given access to the system via a gamified interface as shown in Fig. 2, and this is used to provide more intuitive access to the parameter space. The interface was designed to visualise the ramps of the laser powers and coil currents sent by the citizen scientists to the experiment. The control values were normalised and presented for ease of use on a logarithmic axis in a spline editor where the user could manipulate the curves by clicking-and-dragging points along the curve. When the user was done editing the curves, the sequence was submitted and subsequently realised in the experiment. This was done in the following manner: The user sequence (encoded as a JSON file with a unique user ID) was delivered to a web server. The web server then delivered the sequence to the cloud folder that served as the queue. When a sequence was ready to be evaluated, it was sent to the LabVIEW control system, where the JSON data was translated into waveform data identical to the type used in the single expert user configuration. This was done via a special _optimisation class_ defined in LabVIEW that was responsible for extracting the relevant parameters from the JSON file. Once a sequence was completed, the control program wrote the results to another JSON file, inserted the relevant user ID, and stored it in a separate folder on the cloud. The webserver then delivered the results, and the backend of the game interface scaled the BEC atom number to a _score_ which was displayed to the user. The score and corresponding sequence was also visible for other players who could then copy the sequence as inspiration when creating their own sequences. Figure 3: Schematic view of the data flow for the remote control of the experiment by multiple, asynchronous outside users during the _Alice Challenge_. Experimental sequences are submitted through a game-like user interface to a web server that subsequently sends them to the cloud-based queue in the order they are received (here, User A has submitted their sequence first). Each submission has a unique User ID that is tracked throughout the process. The control system reads the oldest files via the FIFO principle and runs the corresponding experiment. When image analysis is completed, the results are returned to the proper user via the UI. In contrast to the case of a single user, the web server was needed to track the run number and user ID if multiple users were running the experiment simultaneously. A schematic view of the multi-user data handling infrastructure used for the Alice Challenge is presented in Fig. 3. To ensure that the result of the experiment was linked to the right user sequence a check was made in the experimental control system such that the experimental sequence was repeated if no result was returned for a given run. Moreover, the state of the experiment was checked by inserting an established benchmark sequence in the queue every tenth run. This benchmark sequence was known to create a BEC under stable experimental conditions. In the case of a problem, such as a laser failure, the experiment was paused until the problem was solved. At the same time, the users were informed of the temporary delay caused by the disturbance. The benchmark sequence was also executed in case of an empty queue in order to keep the experiment in a stable condition. This also allows one to track overall experimental drifts, e.g. due to thermal fluctuations, which can be useful in advanced closed-loop optimisation schemes. Figure 4: The mean rate at which the experimental sequences were performed during the week in which the experiment was open to non-expert outside users. The different plateaus arise due to changes in the ramp time whereas the high run rates are an effect of synchronisation problems. The green shading denotes the time period depicted in Fig. 5. In the inset the accumulated number of unique citizen scientists that used the system throughout the week is shown. The date markers indicate midnight CET on a given day. The experiment was open to the public for a full week with only minor interruptions, resulting in total of $7577$ submitted and evaluated sequences. Figure 4 shows the rate of the experimental runs during this week. Over the course of the challenge, the preset duration of the evaporation ramps were varied. This allowed the citizen scientists to explore different optimisation landscapes, varying the challenge offered to them and keeping things interesting for returning users. The different evaporation durations create some variation of the rate at which experiments were performed over the course of the week. In addition, when experimental problems caused the experiment to be paused, the rate decreased. It should also be noted that the peaks of high run rates were caused by synchronisation problems between the web server and the control program. This problem was solved on the third day of the challenge, after which none of the larger peaks are visible. The inset shows the progress of the accumulated number of unique users through out the week. Throughout the day the number of active users varied as players from several parts of the world came online. Figure 5 shows the queue and number of active users on Friday evening (CET), one of the highest peaks in active users. Here we see that up to 15 unique users were active at any given time during the evening which created a wait time of above one hour. As the number of users declined, the length of the queue was slowly reduced. Since each user could submit several different sequences at a time, the correlation between the number of unique users and the queue length is nonlinear. Figure 5: The queue time and number of active users during the busiest period of the challenge. The blue trace shows the calculated waiting time from a submission of a sequence until a result is given and the green trace shows the number of active users at any given point. Figure 6 shows a histogram of how many times a given BEC number was achieved. We see that most sequences submitted to the experiment result in the creation of a BEC. This is despite the fact that citizen scientists had limited insight to the physical system they were controlling to create the condensates. Figure 6: A histogram of how many times a given BEC atom number was obtained by the sequences submitted by the users of the Alice challenge. Above $73\%$ of the submitted sequences created a BEC. This data is also presented in Ref. [6]. ## 5 Outlook In future work, remote controlled optimisation of a system may be advantageous, as remote optimisation allows for easy implementation of advanced optimisation algorithms. Several programs are available that can implement closed-loop optimisation of cold-atom experiments [15, 16, 17, 18, 19]. Students can also access such systems for educational purposes, as has already been done with quantum computers [13]. This allows students to explore complex, cutting-edge research systems that are not accessible in many educational learning laboratories. For remote users to be able to run optimise the experiment, the relevant experimental control parameters have to be easily controllable. Collaborative optimisation between several remote users requires a structure that includes a multi-user queue and tracking the ID of submitted sequences so that the results may be returned to the correct user. The control program presented here can be expanded to give remote access to larger parts of the control sequence or even the entire experiment. Thus, future work will give remote users expanded access, allowing them to tackle more advanced scientific problems in a research or educational setting. For example, with the new capabilities of the experiment to image single atoms using a quantum gas microscope [30] in combination with spin addressing [31] and arbitrary light potential generation techniques [32] the experiment can be used as an analog quantum simulator with remote control capability. Such advanced control will require a more complex user interface since the number of experimental parameters would increase, rendering algorithmic optimisation more difficult. However other groups have shown that optimisation of such systems with large parameter spaces is possible using genetic algorithms [33, 34, 35] or machine learning methods such as neural networks [36], Gaussian processes [37] or evolutionary optimisation [38]. ## 6 Acknowledgements The authors would like to acknowledge Aske Thorsen for the development of the LabVIEW control code used for the experiments described here. ## References ## References * [1] Gross C and Bloch I 2017 Science 357 995–1001 ISSN 0036-8075, 1095-9203 * [2] Schäfer F, Fukuhara T, Sugawa S, Takasu Y and Takahashi Y 2020 Nature Reviews Physics 2 411–425 ISSN 2522-5820 * [3] Hofstetter W and Qin T 2018 Journal of Physics B: Atomic, Molecular and Optical Physics 51 082001 ISSN 0953-4075, 1361-6455 * [4] Saffman M 2016 Journal of Physics B: Atomic, Molecular and Optical Physics 49 202001 ISSN 0953-4075, 1361-6455 * [5] Deutsch I H 2020 PRX Quantum 1 020101 ISSN 2691-3399 * [6] Heck R, Vuculescu O, Sørensen J J, Zoller J, Andreasen M G, Bason M G, Ejlertsen P, Elíasson O, Haikka P, Laustsen J S, Nielsen L L, Mao A, Müller R, Napolitano M, Pedersen M K, Thorsen A R, Bergenholtz C, Calarco T, Montangero S and Sherson J F 2018 Proceedings of the National Academy of Sciences 115 E11231–E11237 ISSN 0027-8424, 1091-6490 * [7] IBM Quantum Experience https://quantum-computing.ibm.com/ * [8] Berta M, Wehner S and Wilde M M 2016 New Journal of Physics 18 073004 ISSN 1367-2630 * [9] Dumitrescu E F, McCaskey A J, Hagen G, Jansen G R, Morris T D, Papenbrock T, Pooser R C, Dean D J and Lougovski P 2018 Physical Review Letters 120 210501 ISSN 0031-9007, 1079-7114 * [10] Wang Y, Li Y, Yin Z q and Zeng B 2018 npj Quantum Information 4 46 ISSN 2056-6387 * [11] Wootton J R and Loss D 2018 Physical Review A 97 052313 ISSN 2469-9926, 2469-9934 * [12] Zulehner A, Paler A and Wille R 2019 IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems 38 1226–1236 ISSN 0278-0070, 1937-4151 * [13] Wootton J R, Harkins F, Bronn N T, Vazquez A C, Phan A and Asfaw A T 2020 arXiv:2012.09629 [physics, physics:quant-ph] (Preprint 2012.09629) * [14] Wootton J R 2017 Proceedings of the Second Gamification and Serious Games Symposium 2 63–64 ISSN 2297-914X * [15] Owen S F and Hall D S 2004 Review of Scientific Instruments 75 259–265 ISSN 0034-6748, 1089-7623 * [16] Gaskell P E, Thorn J J, Alba S and Steck D A 2009 Review of Scientific Instruments 80 115103 ISSN 0034-6748, 1089-7623 * [17] Starkey P T, Billington C J, Johnstone S P, Jasperse M, Helmerson K, Turner L D and Anderson R P 2013 Review of Scientific Instruments 84 085111 ISSN 0034-6748, 1089-7623 * [18] Keshet A and Ketterle W 2013 Review of Scientific Instruments 84 015105 ISSN 0034-6748, 1089-7623 (Preprint 1208.2607) * [19] Perego E, Pomponio M, Detti A, Duca L, Sias C and Calosso C E 2018 Review of Scientific Instruments 89 113116 ISSN 0034-6748, 1089-7623 * [20] ColdQuanta control system, www.coldquanta.com * [21] MLabs ARTIQ https://m-labs.hk/ * [22] ADwin https://www.adwin.de/ * [23] IQ Technologies A Laboratory Control System for Cold Atom Experiments http://www.iq-technologies.net/projects/pc/024/ * [24] ColdQuanta Albert: Quantum matter on the cloud https://www.coldquanta.com/albertcloud/ * [25] Bason M G, Heck R, Napolitano M, Elíasson O, Müller R, Thorsen A, Zhang W Z, Arlt J J and Sherson J F 2018 Journal of Physics B: Atomic, Molecular and Optical Physics 51 175301 ISSN 0953-4075, 1361-6455 * [26] Elíasson O, Heck R, Laustsen J S, Napolitano M, Müller R, Bason M G, Arlt J J and Sherson J F 2019 Journal of Physics B: Atomic, Molecular and Optical Physics 52 075003 ISSN 0953-4075 * [27] Grimm R, Weidemüller M and Ovchinnikov Y B 2000 Optical Dipole Traps for Neutral Atoms Advances In Atomic, Molecular, and Optical Physics vol 42 (Elsevier) pp 95–170 ISBN 978-0-12-003842-8 * [28] Lin Y J, Perry A R, Compton R, Spielman I B and Porto J 2009 Physical Review A 79 063631 ISSN 1050-2947 * [29] Rach N, Müller M M, Calarco T and Montangero S 2015 Physical Review A 92 062343 * [30] Elíasson O, Laustsen J S, Heck R, Müller R, Weidner C A, Arlt J J and Sherson J F 2020 arXiv:1912.03079 [cond-mat, physics:quant-ph] (Preprint 1912.03079) * [31] Weitenberg C, Endres M, Sherson J F, Cheneau M, Schauß P, Fukuhara T, Bloch I and Kuhr S 2011 Nature 471 319–24 ISSN 1476-4687 * [32] Christie S Chiu, Geoffrey Ji, Anton Mazurenko, Daniel Greif and Markus Greiner 2018 120 243201 * [33] Rohringer W, Fischer D, Trupke M, Schmiedmayer J and Schumm T 2011 Stochastic Optimization of Bose-Einstein Condensation Using a Genetic Algorithm Stochastic Optimization - Seeing the Optimal for the Uncertain ed Dritsas I (InTech) ISBN 978-953-307-829-8 * [34] Lausch T, Hohmann M, Kindermann F, Mayer D, Schmidt F and Widera A 2016 Applied Physics B 122 * [35] Geisel I, Cordes K, Mahnke J, Jöllenbeck S, Ostermann J, Arlt J J, Ertmer W and Klempt C 2013 Applied Physics Letters 102 ISSN 00036951 * [36] Tranter A D, Slatyer H J, Hush M R, Leung A C, Everett J L, Paul K V, Vernaz-Gris P, Lam P K, Buchler B C and Campbell G T 2018 Nature Communications 9 4360 ISSN 2041-1723 * [37] Wigley P B, Everitt P J, van den Hengel A, Bastian J W, Sooriyabandara M A, McDonald G D, Hardman K S, Quinlivan C D, Manju P, Kuhn C C N, Petersen I R, Luiten A N, Hope J J, Robins N P and Hush M R 2016 Scientific Reports 6 25890 * [38] Barker A J, Style H, Luksch K, Sunami S, Garrick D, Hill F, Foot C J and Bentine E 2020 Machine Learning: Science and Technology 1 015007 ISSN 2632-2153
# The field theoretical ABC of epidemic dynamics Giacomo Cacciapaglia, Corentin Cot, Stefan Hohenegger, Shahram Vatani Institut de Physique des 2 Infinis (IP2I), CNRS/IN2P3, UMR5822, 69622 Villeurbanne, France Université de Lyon, Université Claude Bernard Lyon 1, 69001 Lyon, France Michele Della Morte IMADA & CP3-Origins. Univ. of Southern Denmark, Campusvej 55, DK-5230 Odense, Denmark Francesco Sannino Scuola Superiore Meridionale, Largo S. Marcellino, 10, 80138 Napoli NA, Italy CP3-Origins and D-IAS, Univ. of Southern Denmark, Campusvej 55, DK-5230 Odense, Denmark Dipartimento di Fisica, E. Pancini, Univ. di Napoli, Federico II and INFN sezione di Napoli Complesso Universitario di Monte S. Angelo Edificio 6, via Cintia, 80126 Napoli, Italy ###### Abstract Infectious diseases are a threat for human health with tremendous impact on our society at large. They are events that recur with a frequency that is growing with the exponential increase in the world population and growth of the human ecological footprint. The latter causes a frequent spillover of transmissible diseases from wildlife to humans. The recent COVID-19 pandemic, caused by the SARS-CoV-2, is the latest example of a highly infectious disease that, since late 2019, is ravaging the globe with a huge toll in terms of human lives and socio-economic impact. It is therefore imperative to develop efficient mathematical models, able to substantially curb the damages of a pandemic by unveiling disease spreading dynamics and symmetries. This will help inform (non)-pharmaceutical prevention strategies. It is for the reasons above that we decided to write this report. It goes at the heart of mathematical modelling of infectious disease diffusion by simultaneously investigating the underlying microscopic dynamics in terms of percolation models, effective description via compartmental models and the employment of temporal symmetries naturally encoded in the mathematical language of critical phenomena. Our report reviews these approaches and determines their common denominators, relevant for theoretical epidemiology and its link to important concepts in theoretical physics. We show that the different frameworks exhibit common features such as criticality and self-similarity under time rescaling. These features are naturally encoded within the unifying field theoretical approach. The latter leads to an efficient description of the time evolution of the disease via a framework in which (near) time-dilation invariance is explicitly realised. As important test of the relevance of symmetries we show how to mathematically account for observed phenomena such as multi-wave dynamics. Although we consider the COVID-19 pandemic as an explicit phenomenological application, the models presented here are of immediate relevance for different realms of scientific enquiry from medical applications to the understanding of human behaviour. Our review offers novel perspectives on how to model, capture, organise and understand epidemiological data and disease dynamics for modelling real-world phenomena, and helps devising public health and socio-economics strategies. ###### keywords: epidemiology , field theory ###### MSC: [2021] 92D30 ††journal: Physics Reports ###### Contents 1. 1 Introduction 1. 1.1 Historical Overview 2. 1.2 Current approaches to epidemiology 3. 1.3 Relating different scales in Field Theory 4. 1.4 Organisation of the Review 2. 2 Percolation Approach 1. 2.1 Lattice and Percolation Models 2. 2.2 Numerical Simulations and Criticality 1. 2.2.1 The principle 2. 2.2.2 Results 3. 2.3 Master Action and Field Theory 4. 2.4 Relation to Compartmental Models 3. 3 Compartmental Models 1. 3.1 SIR(S) Model, Basic Definitions 2. 3.2 Numerical Solutions and their Qualitative Properties 3. 3.3 From Lattice to SIR 4. 3.4 Parametric Solution of the Classical SIR Model 5. 3.5 Generalisations of the SIR Model 1. 3.5.1 Time Dependent Infection and Recovery Rates 2. 3.5.2 Spontaneous Creation and Multiple Waves 3. 3.5.3 Heterogeneous Transmission Rates and Superspreaders 6. 3.6 The SIR model as a set of Renormalisation Group Equations 1. 3.6.1 Beta Function 2. 3.6.2 Connection between SIR models and the eRG approach 7. 3.7 Analytic Solution during a Linear Growth Phase 1. 3.7.1 Simplified SIR Model with Constant New Infections 2. 3.7.2 Vanishing $\zeta$ and Constant $\epsilon$ 3. 3.7.3 Constant Active Number of Infectious Individuals 4. 4 Epidemic Renormalisation Group 1. 4.1 Beta Function and Asymptotic Fixed Points 1. 4.1.1 Generalisation to multiple regions 2. 4.2 Complex (fixed point) epidemic Renormalisation Group 3. 4.3 Modelling multi-wave patterns 5. 5 COVID-19 6. 6 Outlook and Conclusions ## 1 Introduction Infectious diseases that can efficiently spread across the human population and cause a pandemic have always been a threat to humanity. This menace has been growing with the increase in the population and the progressive destruction of the wild environment with its impact on wildlife. The last century has been affected by, at least, three major worldwide pandemics: the 1918 “Spanish” influenza of 1918-1920 [1], HIV/AIDS [2, 3] and the most recent COVID-19 that started at the end of 2019. Understanding in a mathematically consistent way the diffusion of a pandemic is of paramount importance in designing effective policies apt at curbing and limiting its diffusion and the impact on the life loss and economic damage. In this report we will review some crucial aspects of the mathematical modelling, ranging from the microscopic mechanisms encoded in diffusion models, to approaches based on symmetries. In this discussion, the application of field theory and other concepts borrowed from theoretical physics will play a crucial role. The dynamics of physical phenomena, from the fundamental laws of nature to quantum and ordinary matter phase transitions, even including protein behaviour, is well captured by effective descriptions in terms of fields and their interactions. Given the enormous success of the field theoretical interpretation of physical phenomena, it is highly interesting to review several main mathematical models employed to describe the diffusion of infectious diseases and show how the different approaches are related within the field theoretical framework. We will show that the models exhibit common features, such as criticality and self-similarity under time rescaling. These features are naturally encoded within the unifying field theoretical approach. The latter yields an efficient description of the time evolution of the disease via a framework in which (near) time-dilation invariance is explicitly realised. The models are extended to account for observed phenomena such as multi-wave dynamics. Because of the immediacy of the COVID-19 pandemic and the high quality data available, we use it as an explicit and relevant phenomenological test of the models and their effectiveness. It should be clear, however, that the methodologies presented here are relevant for any infectious disease, and can be extended to different realms of scientific enquiry, from medical applications to the understanding of human behaviour. We will complete this introduction with a historical overview of the mathematical modelling applied to infectious diseases, the contemporary applications and the role of field theory concepts, before offering a summary of the main body of the review. ### 1.1 Historical Overview The first application of mathematical modelling to an epidemiological process is the work of Daniel Bernoulli [4] on the effectiveness of an inoculation against smallpox in 1760. A more systematic application of mathematical methods to study the spread of infectious diseases occurred after the work of Robert Koch and Louis Pasteur, which showed that such diseases are caused by living organisms, triggering the question on how they are passed on from one individual to another. A related point in this regards is how (and why) outbreaks and epidemics end. As outlined in [5], there are two prevalent hypotheses: * 1. Farr’s hypothesis (mostly based on the work of W. Farr in 1866 [6]): epidemics stop because the potency of the microorganisms decreases with every new individual that is infected. * 2. Snow’s hypothesis (mostly based on the work of J. Snow in 1853 [7]): epidemics end due to a lack of sufficient available new individuals to infect (the disease runs out of “fuel”). In view of closer studies of actual data stemming from outbreaks of communicable diseases, Farr’s hypothesis was gradually dropped from the scientific discussion. Moreover, the focus of research shifted towards explaining regularities of observed _epidemic curves_. A first discovery along these lines can be found in the work of W.H. Hamer [8, 9, 10, 11], who described the biennial period of measles outbreaks in London and implicitly [5] introduced the concept of mass-action law into epidemiology. The latter was firmly established in the pioneering works of Sir R. Ross [12, 13, 14, 15] and A.G. McKendrick [16, 17, 18]. Specifically, in a model of discretised time (with time steps $\delta t$) such as in [12], the mass action law can be formulated as follows: $\displaystyle\text{number of cases at }t+\delta t\propto(\text{number of cases at }t)\times(\text{susceptibles at }t)\,.$ In models with a continuous time variable (as in later works of Ross and notably McKendrick), the model can be formulated in terms of differential equations, including additional contributions capturing the population dynamics due to recovery from the disease, birth, death, migration, _etc._. Credit for the so-called SIR model, still widely used today (and which we review in Section 3.1), is given to the work by W.O. Kermack and A.G. McKendrick in 1927 [19]. The basic idea behind models of this type is that the disease is passed on among individuals in the form of happenings or collisions, in analogy to how reactions work in chemistry. This led to numerous more refined models, see for example the reviews [20, 21, 22, 23, 24, 25, 26, 27, 28], including the historical overview in [29] . In the second half of the 20th century, progress in different disciplines influenced epidemiological investigations. It was understood that, to describe (and combat) large scale outbreaks such as HIV/AIDS, human behaviour plays a crucial role in modulating the spreading of the virus (_e.g._ [30, 31]). Thereby, mathematical modelling started going beyond models inspired by basic chemical reactions. The appearance of a large number of reviews and books on epidemiological modelling is testimony to the depth and complexity of the analysis as well as the interdisciplinary attention this topic has received, _e.g._ [32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57]. Besides influences and contributions from pure mathematics, chemistry and social sciences, certain features and symmetries of epidemic curves are similar to those found in particular physical systems. This has lead to novel approaches rooted in the physics of critical phenomena and phase transitions, such as _percolation models_ [58, 59] and their relation to _(scale invariant) field theories_. As we shall review in Section 2.1, there are various types of percolation models. Here we shall define them simply as collections of points in a given space, where some of which can be linked pairwise. The sets of points that are linked to each other are termed clusters and the spread of such clusters (following certain pre-determined rules) can be used to model the spread of a disease within a given population. In particular, the transition from finite sized clusters to the percolation phase (where all points are linked together) is a phase transition. This important feature allows percolation models to be organised in terms of universality classes and even to put them in correspondence to other physical systems. This property is useful to determine important physical quantities. The first attempts appeared in [60] and their relation to phase transitions was pointed out in [61], while excellent reviews on more complicated models can be found in [62, 63, 64, 65, 66, 67, 68, 69, 70, 71]. Direct formulations of percolation models in terms of field theories follow the approach of M. Doi and L. Peliti [72, 73, 74], which have been reviewed, for example, in [75]. Further work in this direction (notably the work by J. Cardy and P. Grassberger [76] and its relation to models in particle physics [77, 78]) shall be reviewed in Section 2.1. ### 1.2 Current approaches to epidemiology The aim of our work is to summarise, review and connect various current approaches to understand and model the time evolution of pandemics. From the brief historical analysis of the previous subsection it is clear that, over the course of almost a century, many different approaches have been developed. Classifying them is often difficult. From a mathematical perspective one can distinguish stochastic and deterministic approaches, based on how the basic fundamental (microscopic) processes of the transmission and development of the disease are modelled: all epidemiological models generally assume that new infected individuals can appear when an uninfected one (usually called a _susceptible_ individual) comes in contact with an _infectious_ individual such that the disease is passed on. After some time, infected individuals may turn non-infectious (at least temporarily) via recovering or dying from the disease or by some other means of _removal_ from the actively involved population. Mathematically speaking, these processes can be modelled in two different fashions: * 1. Stochastic approach: all (microscopic) processes between individuals are of a probabilistic nature. For instance, the contact between a susceptible and an infectious individual has a certain probability to lead to an infection of the former; infected individuals have a certain probability of removal after a certain time; _etc_. In these approaches, time is understood as a discrete variable and time-evolution is typically described in the form of differential-difference equations (called _master equations_). The solutions depend on a set of probabilities (_e.g._ the probability of a contact among individuals leading to an infection), geometric parameters (such as the number of ’neighbouring’ individuals that a single infectious individual can potentially infect) as well as the initial conditions. Furthermore, in order to make predictions or to compare with deterministic approaches, some sort of averaging process is required. * 2. Deterministic approach: the time evolution of the number of susceptible, infectious and removed individuals is understood as a fully predictable process and is typically described through systems of coupled, ordinary differential equations in time (the latter is understood as a continuous variable). Solutions of these systems are therefore determined by certain parameters (such as infection and recovery rates) as well as initial conditions (_e.g._ the number of infectious individuals at the outbreak of the disease). In this review, we prefer to think of this classification in a somewhat different (but equivalent) fashion, which (as we shall explain) is closer to the concept of (energy) _scale_ in particle physics. Indeed, we prefer to think of models as ranging from microscopic models, in which fundamental interactions (_i.e._ at the level of individuals) are explicitly modelled, to more and more macroscopic approaches, in which the microscopic interactions have been (at least partially) included into the interactions of new, effective degrees of freedom. A basic overview, with concrete models, is given in Figure 1: models in the left part of the diagram (red box) incorporate many details of how the disease spreads at a microscopic level, _i.e._ between single individuals. These models are mostly of a stochastic nature, using probabilistic means to simulate the spread of the disease. As we shall explain, many of them are inspired by chemical models, in which a random movement of molecules is considered, with collisions leading with a certain probability to a chemical reaction (and the creation of new molecules). The models further to the right of the diagram (blue box) are more macroscopic, in the sense that they no longer model individual interactions (_i.e._ the spread from one person to the next), but rather describe the time evolution of the disease in a larger population (_e.g._ an entire country). While, historically, the oldest models that have been developed to describe the spread of an infectious disease are in this category, many of them can be obtained from more microscopic approaches (_e.g._ percolation models) through a ’replacement’ of the degrees of freedom of the latter by more macroscopic ones. This can happen, for example, via a _mean field approximation_ or via certain averaging procedures or by describing the spread of the disease through suitable flow equations. The resulting models are mostly of a deterministic nature, but can retain stochastic elements. Besides the explicit models and approaches listed in Figure 1 (some of which we shall review in the main part of this article), there are also data- and computer-driven approaches [79, 80]. These generally use machine learning (also called statistical learning) tools to analyse existing data with the goal of finding patterns and predicting the future development of pandemics. On the one hand, these approaches use the large advances in computer technology (in particular the development of artificial intelligence). On the other hand, they are made viable in recent years due to the dramatical increase in the volume and quality of available data on the spread and development (_e.g._ its genetic mutations) of diseases in a large population. This allows data-driven approaches to be applied at any level, ranging from analysing microscopic interactions (see _e.g._ [81]) to more effective descriptions that only aim at predicting ’global’ key statistics of epidemics [82, 83]. Since the current review is aimed at studying field theoretic tools in epidemiology, we shall not discuss these methods here. However, we point out a number of excellent review articles [84] in the literature. Another class of models we will not discuss utilises complex networks to include the effect of human behaviour [85]. Microscopic approaches on the left spectrum of Figure 1 generally utilise first principles, however at the expense of a lack of symmetries (usually also entailing a large computational cost). Effective theories on the right side of the graph are, usually, less intuitive (since basic interactions of the disease enter into a less obvious manner). However, they incorporate basic symmetries that appear in the solutions of the microscopic models – in the sense of making them manifest – typically also leading to more streamlined and less expensive computations. Here is an incomplete list of the symmetries at the base of these approaches: microscopic models macroscopic models $\bullet$ lattice models$\bullet$ percolation models $\bullet$ random walks$\bullet$ diffusion models$\bullet$ (epidemic) field th.$\bullet$ network models $\bullet$ compartmental models$\bullet$ epidemic RenormalisationGroupeffective description microscopic degrees of freedom replaced by more appropriate effective ones: mean field approximations, averaging, beta-functions, flow equations,… \+ based on ’first principles’ \- symmetries not manifest input: basic properties of the disease and the way it spreads \+ based on manifest symmetries \+ computationally simpler \- modelling requires more intuition about the system and/or data input: ’effective’ properties of the disease in a specific population Figure 1: Schematic overview of different approaches to describe the time evolution of pandemics and their relation to field theoretical methods. 1. _(i)_ _Criticality:_ depending on the parameters of the model and the starting conditions, solutions of microscopic models feature either a quick eradication of the disease, where the total cumulative number of infected individuals remains relatively low, or a fast and widespread diffusion of the disease, leading to a much larger total number of infected. Which of these two classes of solutions is realised is usually governed by a single ordering parameter (_e.g._ the average number of susceptible individuals infected by a single infectious, also known as _reproduction number_ $R_{0}$), and the transition from one type to the other can be very sharp. 2. _(ii)_ _Self-similarity and waves:_ depending on the disease in question, solutions of microscopic models may exhibit distinct phases in their time evolution in the form of a wave pattern, where phases of exponential growth of the number of infected individuals are followed by intermediate periods of slow, approximately linear, growth. Each wave typically looks similar to the previous and following ones. Furthermore, certain classes of solutions may also exhibit spatial self-similarities, _i.e._ the solutions describing the temporal spread of the disease among individuals follow similar patterns as the spread among larger clusters (_e.g._ cities, countries _etc._). 3. _(iii)_ _Time-scale invariance:_ several microscopic models exhibit a (nearly) time- scale invariant behaviour, which is a symmetry under rescaling of the time variable and of the rates (infection, removal, _etc._). If the solution exhibits a wave-structure, these near-symmetric regions can appear in specific regimes, _e.g._ in between two periods of exponential growth. These properties are familiar from field theoretical models in physics, _e.g._ in solid state and high energy physics, which exhibit phase transitions. Indeed, over the years, it has been demonstrated that the various approaches mentioned above can be reformulated (or at least related to) field theoretical descriptions. The latter are typically no longer sensitive to microscopic details of the spread of the disease at the level of individuals, but instead capture _universal_ properties of their solutions. They are therefore an ideal arena to study properties of the dynamics of diseases and the mechanisms to counter their spread. ### 1.3 Relating different scales in Field Theory The dynamics of physical phenomena, ranging from the fundamental laws of nature to quantum and ordinary matter phase transitions including protein behaviour, is well captured by effective descriptions in terms of fields and their interactions. These fields are meant to capture the overall features of the phenomenon in question, describe the interaction between (elementary) constituents and even predict the evolution of the system. Once the field theoretical dynamics is married to underlying approximate or exact symmetries, it becomes an extremely powerful tool that, in a given range of energy, provides a faithful representation of the microscopic physics underlying many phenomena. Zooming in or out of the relevant physical scales involved in the dynamics of a given process generically requires a modification of the degrees of freedom needed to describe that specific process. This property is captured by the renormalisation group (RG) framework [86, 87]. Within this approach, in order to take into account the change in degrees of freedom, one modifies (renormalises) the interaction strengths and rescales the fields. In fact, the idea of scale transformations and scale invariance is ancient, dating back to the Pythagorean school. The concept was used in the work by Euclid and much later by Galileo. The idea received renewed popularity towards the end of the 19th century with the idea of enhanced viscosity of O. Reynolds to address turbulence in fluids [88, 89]. However, the seed-idea of the RG initially started in 1953 with the work of Ernst Stueckelberg and André Petermann [90]. They noted that the renormalisation procedure in quantum field theory exhibits a group of transformations, which acts on parameters that govern basic interactions of the system, _e.g._ changing the bare couplings in the Lagrangian by including (counter) terms needed to correct the theory. For example, the application to quantum electro-dynamics (QED) was elucidated by Murray Gell-Mann and Francis E. Low in 1954 [91]: this led to the renown determination of the variation of the electromagnetic coupling in QED with the energy of the physical processes. Hence, the basic idea at the heart of the RG approach stems from the property that, as the scale of the physical process varies, the theory displays a self- similar behaviour and any scale can be described by the same physics. In mathematical terms, this properties is reproduced by a group transformation acting on the interaction strengths of the theory. Thanks to Gell-Mann and Low a computational method based on a mathematical flow function of the interaction strength parameter was introduced. This function determines the differential change of the interaction strength with respect to a small change in the energy of the process through a differential equation known as the renormalisation group equation (RGE). Although mainly devised for particle physics, nowadays its applications extend to solid-state physics, fluid mechanics, physical cosmology, and even nanotechnology. In certain cases, such as in particle physics, the field theoretical description can be elevated to the ultimate description of fundamental interactions if short distance scale invariance occurs. Once scale invariance is married to relativity the group of invariance generically enlarges to the conformal group. ### 1.4 Organisation of the Review In the following we shall start by presenting examples of microscopic and effective (respectively deterministic and stochastic) approaches and show how they can be related to field theoretical models. We start in Section 2 with analysing the direct percolation approach, which is based on a microscopic stochastic description of the diffusion processes. We shall see that the approach, in the mean field approximation, naturally leads to compartmental models. The latter (as well as generalisations thereof) are reviewed in Section 3: we commence this investigation with a basic review of the SIR model and then investigate how to incorporate multi-wave epidemic dynamics paying particular attention to the inter-wave period. After highlighting further possible extensions of compartmental models, we finally provide a formulation of the SIR model in terms of flow equations, which resembles the $\beta$-function familiar from the RG approach to particle and high-energy physics. We use this last result to motivate the most recent approach to epidemic dynamics, _i.e._ the epidemiological renormalisation group (eRG) [92, 93] in Section 4. The latter is inspired by the Wilsonian renormalisation group approach [86, 87] and uses the approximate short and long time dilation invariance of the system to organise its description. For the COVID-19, the eRG has been shown to be very efficient when describing the epidemic and pandemic time evolution across the world [94] and in particular when predicting the emergence of new waves and the interplay across different regions of the world [95, 96]. The discussion in Sections 2, 3 and 4 is general in the sense that the methods apply to generic infectious diseases and populations. In Section 5 we consider particular features of the current ongoing COVID-19 pandemic, and discuss how the different approaches can be adapted to it. Several excellent reviews already exist in the literature [97, 85, 98, 32]. Our work complements and integrates them, adds to the literature on the field theoretical side and further incorporates more recent approaches. ## 2 Percolation Approach Executive Summary 1. We introduce percolation and lattice models as _stochastic_ approaches to directly simulate microscopic interactions down to the individual level. 2. The models are characterised through a set of probabilities (related for example to the infection and recovery rates of individuals) and the geometry of the system. Time is assumed to be a quantised variable. 3. The approach naturally models the spatial as well as the temporal evolution of a disease. 4. The models feature a (sharp) _phase transition_ in terms of the asymptotic infected fraction of the population. The latter is the order parameter of the system. 5. Compartmental models (see next Section) emerge as a mean field description of percolation models. ### 2.1 Lattice and Percolation Models Arguably the most direct way to (theoretically) study the spread of a communicable disease is via systems that simulate the process of infection at a microscopic level, _i.e._ at the level of individuals in a (finite) population. The most immediate such models are lattice simulations, in which the individuals are represented by the lattice sites on a spatial grid, some of which may be infected by the disease. These lattice sites can spread the disease with a certain probability to neighbouring sites, following an established set of rules. Lattice models, therefore, allow to track the spread of the disease in discretised time steps and, after taking the average of several simulations, allow to make statements about the time evolution (and asymptotic values) of the number of infected individuals. As we shall see in the following, even simple models of this type show particular time-scaling symmetries, as well as criticality (_i.e._ the fact that the asymptotic number of infected individuals changes rapidly, when a certain parameter of the model approaches a specific critical value). A larger class of models that work with a discrete number of individuals (as well as discretised time) consists of _percolation models_ , which broadly speaking consist of points (sites) scattered in space that can be connected by links. Depending on the specific details, one distinguishes [71]: * 1. _Bond percolation models_ : in this case the points are fixed and the links between them are created randomly. Examples of this type are (regular) lattices in various spatial dimensions with nearest neighbour sites being linked. * 2. _Site percolation models_ : in this case the position of the points is random, while the links between different points are created based on rules that depend on the positions of the points. More complex models can also incorporate both aspects. An important quantity to compute in any percolation model is the so-called _pair connectedness_ , _i.e._ the probability that two points are connected to each other (through a chain of links with other points). Assuming the system to extend infinitely (_i.e._ there are infinitely many sites), we can importantly distinguish whether it is made of only local clusters (in which finitely many sites are connected) or whether it is in a _percolating state_ (where infinitely many sites are connected). The probability of occurrence of these two situations usually depends on the value of a single parameter (typically related to the probability $p$ that a link exists between two ‘neighbouring’ sites), in such a way that the transition from local connectedness to percolation can be described as a _phase transition_ (see _e.g._ [61]). The system close to this critical value $p_{c}$ lies in the same universality class of several other models in molecular physics, solid state physics and epidemiology: this implies that the behaviour of certain quantities follows a characteristic power law behaviour that is the same for all the theories in the same universality class. For example, the probability $P(p)$ for a system to be in the percolating state (as a function of $p$) takes the form $\displaystyle\lim_{p\to p_{c}}P(p)\sim(p-p_{c})^{\nu}\,,$ (2.1) where $\nu$ is called _critical exponent_. Models within the same universality class share the same critical exponents despite the fact that the concrete details of the theory, in particular the concrete meaning of the quantity $P$ in Eq. (2.1), may be very different. This connection makes percolation models very versatile and many of them have been studied extensively (see [71] and references therein). In the following, we shall first present a simple lattice simulation model, which allows us to reveal important properties of the time evolution of the infection (notably criticality and time-rescaling symmetry). Furthermore, we shall discuss a percolation model that, near criticality, is in the same universality class as time-honoured epidemiological models, along with some of its extensions and generalisations. ### 2.2 Numerical Simulations and Criticality Lattice simulations of reaction-diffusion processes are a well established tool to study the epidemic spreading of a disease since the original work by P. Grassberger in [99]. In specific realisations the models have been studied to very high precision and the critical values of the parameters are known with an accuracy reaching the six digits, see for example Ref. [100] and references therein. Different geometries have been considered as well as different ranges of interactions, including random long-range couplings among sites, see [101, 102] for recent discussions. All of these follow a _Markov decision process[103, 104], _i.e._ the population is represented by a discrete lattice and the time evolution of the disease is organised in discretised time steps (so-called Markov iterations) between each of which the state of the lattice is changed based on a set of stochastic decisions._ Here we consider a synchronous algorithm (i.e., we update all the lattice sites in each Markov iteration), and isotropic interactions of range $r$ (in lattice units). #### 2.2.1 The principle For our purposes, the simplest and most direct way to study percolation models is to simulate the time evolution of the spread of a disease via stochastic processes on a finite dimensional lattice. The individuals, represented by each lattice site, can be in one and only one of the three given states: susceptible, infectious or removed. They are defined as follows: * 1. _Susceptible_ : these are individuals that are currently not infectious, but can contract the disease. We do not distinguish between individuals who have never been infected and those who have recovered from a previous infection, but are no longer immune. * 2. _Infectious_ : these are individuals who are currently infected by the disease and can actively transmit it to a susceptible individual. * 3. _Removed (recovered)_ : these are individuals who currently can neither be infected themselves, nor can infect susceptible individuals. This comprises individuals who have (temporary) immunity (either natural, or because they have recovered from a recent infection), but also all deceased individuals. The time evolution of the lattice configurations follows a set of rules, which implements the following two basic mechanisms into an algorithm that models the spread of the disease within a finite and isolated population in discretised time steps: * _i)_ the infection of susceptible individuals in the vicinity of an infectious one; * _ii)_ the removal (recovery) of an infectious individual (so that it can no longer infect other individuals). The infection process depends on the reach of an infectious site over potential nearby susceptible ones. This reach depends on the geometry of the lattice (here we always use square lattices) and on the range $r$. The removal instead depends on the site itself and on an intrinsic removal probability. Starting from the two principles above, there are two ways to let the lattice evolve and to define the elementary time steps, starting from a given initial spatial distribution of infectious and susceptible sites. On the one hand, we can randomly choose an infectious site and begin the infection process within its surrounding sites (_i.e._ determine how many susceptible neighbouring sites are turned infectious). Once the process is over, another infectious site is chosen randomly, defining the next time step. Such a sequence forbids multiple infections, as only one infected site is considered at each step of time. On the other hand, we could take into account all the possible infections at the same time and consider the susceptible sites that may become infected by them, according to the rules of the algorithm. The lattice is then updated with the new infected and thus the next time step begins. This process allows multiple infections to be considered, as susceptible sites can have multiple infected neighbours infecting them at a single time step. The first method is called “asynchronous” as opposed to the second “synchronous” algorithm. Having discussed the temporal structure of the simulation, we can turn now to the specific mechanism of the spread, which, in our setup, depends on three parameters: * 1. The _coordination radius_ $r\in\mathbb{R}_{+}$, which is a measure for the distance (on the lattice) over which direct infections between individuals can take place, _i.e._ only sites within a distance $r$ from the infectious one can be infected. We illustrate $r$ in Fig. 2 within a 2-dimensional squared lattice. $\vdots$$\vdots$$\cdots$$\cdots$$\mathbf{e}_{1}$$\mathbf{e}_{2}$$r$ Figure 2: Two-dimensional cubic lattice generated by the lattice vectors $(\mathbf{e}_{1},\mathbf{e}_{2})$. The blue circle of coordination radius $r$ ($r=2$ in the current example) contains all susceptible sites (blue) that may become infected by a single infectious one (red) at its centre. * 2. The _infection probability_ $\mathfrak{g}\in[0,1]$ for an infectious individual to infect a neighbour site. In practice, the probability of a single individual in the neighbourhood (defined in terms of the coordination radius) to be infected is equal to $\mathfrak{g}$ divided by the number of sites within a radius $r$ from the infectious one. This choice, as we shall see in Section 3.3, allows us to draw a more direct relation between $\mathfrak{g}$ and the infection rate parameter defined in other approaches. * 3. The _removal probability_ $\mathfrak{e}\in[0,1]$ for an infectious individual to become removed. In the following we shall highlight some of the key-features of this approach and study their dependence on the three parameters above. To do so, we consider a 2-dimensional lattice with periodic boundary conditions. We follow the “synchronous” algorithm with a slightly different path compared to the common approach in the literature [105]. Usually, in order to determine the time-evolution of the lattice configuration, one needs to go through all infectious sites and individually apply the infection algorithm to all susceptible sites within their coordination radius: each contact is simulated by the call of a randomly generated number $x$, between 0 and 1. If $x\leq\mathfrak{g}$, an infection occurs and the considered susceptible site will become infectious at the next time step. Else, nothing happens – the site will stay susceptible and the whole process is repeated for each of the sites surrounding a given infectious one. Instead of this infectious-site-centred procedure, we will consider an algorithm centred on the susceptible sites: for each susceptible site, we count the number $n$ of infectious sites within the coordination radius and calculate the cumulated probability of infection. One can show that, on average, the probability $\mathcal{P}\left(n\right)$ for this site to become infectious in the next Markov iteration is given by $\displaystyle\mathcal{P}\left(n\right)=1-(1-\mathfrak{g})^{n}$. We use this probability to determine the fate of each susceptible site. This improved procedure speeds-up the algorithm and reduces stochastic fluctuations, as it is equivalent to performing a local average at each time step. We turn now to the presentation of our results. #### 2.2.2 Results Figure 3: Number of infected individuals as a function of the discretised time for a lattice with $40401$ sites, $\mathfrak{g}=0.7$, $\mathfrak{e}=0.1$ and coordination radius $r=1$. A plot of the evolution of the cumulative number of infected sites as a function of the discretised time-steps is shown in Fig. 3 for a sample choice of the parameters $\mathfrak{g}=0.7$, $\mathfrak{e}=0.1$ and $r=1$ and for a square lattice with $201$ sites on each side (_i.e._ $40401$ sites in total). At large $t$, the cumulated number of infected approaches an asymptotic value, which, averaged over a sufficient number of simulations, is a function of the probabilities $(\mathfrak{g},\mathfrak{e})$ as well as of the coordination radius $r$. Varying these parameters leads to substantially different asymptotic values, as is shown in Fig. 4: in the four panels, we plot the asymptotic values as a function of the infection probability $\mathfrak{g}$. We use the same lattice as before and fix $\mathfrak{e}=0.1$. For each point, we repeat the process $50$ times to compute the shown mean and standard deviation. As expected, the larger $\mathfrak{g}$, the higher the number of infected sites at the end of the process. The plots also show the critical behaviour of the system, as the asymptotic value jumps from a very small value at small $\mathfrak{g}$ to a value of the same order of the total population (_i.e._ the number of sites in the lattice). For each value of $r$, one can define a critical value $\mathfrak{g}_{c}(r)$: increasing $r$ reduces the value of $\mathfrak{g}_{c}$. (a) $r=1$ (b) $r=2$ (c) $r=5$ (d) $r=50$ Figure 4: Evolution of the asymptotic number of infected sites as a function of the infection probability $\mathfrak{g}$ for different coordination radii $r$. The removal probability is fixed to $\mathfrak{e}=0.1$. In the simulations of Fig. 4 we use the same initial condition, where all the sites within a radius $5$ (in lattice units) from the centre of the lattice are set to the infectious state, thus having initially $81$ cases. Due to the stochastic nature of the process, the final number of infected cases does depend non-trivially on the initial state, especially for small coordination radius $r$. For $r=1$ and $\mathfrak{e}=0.1$, this dependence on the initial infected $N_{I}$ is shown in the left panel of Fig. 5, where we plot the asymptotic value of infected as a function of $N_{I}$, randomly distributed on the lattice. We plot the results for three different values of $\mathfrak{g}=0.4,\;0.5$ and $0.7$, where $\mathfrak{g}=0.5$ is close to the critical $\mathfrak{g_{c}}$. The critical behaviour described above seems to be also sensitive to $N_{I}$. This could be due to finite volume effects, as the evolution of the infection is expected to depend crucially on the density (rather than on the actual number) of initial infectious cases on the lattice as well as on their spatial distribution. The dependence on the initial state is consistent with the result obtained for the SIR compartmental models discussed in Sec. 3.2. This effect should disappear in the infinite volume limit. Especially near the critical value, we observe a large spread of results for the asymptotic numbers. This is particularly evident for small densities of initial infections, where stochastic effects become relevant. As an example, we show a bundle of 50 solutions near the critical value in the right panel of Fig. 5. Figure 5: Left panel: Evolution of the final number of infected sites as a function of the initial infectious ones. The mean and the standard deviation are computed over $50$ simulations for each point with $\mathfrak{e}=0.1$ and different values of $\mathfrak{g}$. Right panel: Time evolution of the infected cases for $50$ simulations with $\mathfrak{e}=0.1$, $\mathfrak{g}=0.5$, $N_{I}=2$ and $r=1$. ### 2.3 Master Action and Field Theory Here we briefly summarise the percolation approach and the derivation via field theory of the reaction diffusion processes. We follow G. Pruessner’s lectures [75] and borrow part of his notation. The overarching goal is to reproduce and extend the action given in the seminal work of J.L. Cardy and P. Grassberger [76]. We, therefore, consider a model of random walkers described by a field $W$ that diffuse through a lattice, reproduce themselves and drop some poison $P$ as they stroll around. The poison field $P$ does not diffuse but kills walkers if they hit a poisoned location. Interpreting the positions of the walkers as infected sites and those of the poison as simultaneously representing either the immune or removed individuals, the model effectively describes a disease diffusion process featuring infection and immunisation dynamics. The microscopic processes considered in [76] (see also [106, 107]) can be schematically summarised as follows: $\displaystyle W\,\rightarrow\,W+W\;,$ $\displaystyle{\rm with\;rate\;}\sigma\;,$ $\displaystyle W\,\rightarrow\,W+P\;,$ $\displaystyle{\rm with\;rate\;}\alpha\;,$ $\displaystyle W+P\,\rightarrow\,P\;,$ $\displaystyle{\rm with\;rate\;}\beta\;.$ (2.2) The first branching process corresponds to infection, while the last two processes describe immunisation. In addition we will consider a process of spontaneous creation, by which infected can appear at one site independently from the presence of other infected at neighbouring sites, with a rate $\xi$. $n^{W}_{\mathbf{x}},n^{P}_{\mathbf{x}}$$e_{1}$$e_{2}$ Figure 6: Schematic presentation of the state $\\{n_{\mathbf{x}}^{W},n_{\mathbf{x}}^{P}\\}$ with $e_{i}$ the basis vectors of $\Gamma$. The field theory is derived from a discretised version of the model, eventually taking the continuum limit. The starting point is a _Master Equation_ that directly leads to the action through a process of second- quantisation. Let $\Gamma\subset\mathbb{Z}^{d}$ be a $d$-dimensional hypercubic lattice with coordination number $q$, which is generated by a set of vectors $\mathbf{e}$. We denote by $\\{n^{W}_{\mathbf{x}},n^{P}_{\mathbf{x}}\\}$ a state with site $\mathbf{x}$ occupied by $n^{W}_{\mathbf{x}}$ and $n^{P}_{\mathbf{x}}$ particles of type $W$ and $P$ $\forall\mathbf{x}\in\Gamma$ (for a schematic representation see Fig. 6). The probability that such state is realised at time $t$ is denoted by $P(\\{n^{W}_{\mathbf{x}},n^{P}_{\mathbf{x}}\\};t)$. Configurations can change via the different mechanisms described above. The probability thus satisfies the first order differential equation (Master Equation): $\displaystyle\frac{dP(\\{{n}^{W}_{\mathbf{x}},n^{P}_{\mathbf{x}}\\};t)}{dt}=$ $\displaystyle\frac{H}{q}\,\sum_{\mathbf{y}\in\Gamma}\sum_{e\in\mathbf{e}}\left[(n^{W}_{\mathbf{y}+e}+1)P(\\{n^{W}_{\mathbf{y}}-1,n^{W}_{\mathbf{y}-e}+1,n^{P}_{\mathbf{x}}\\};t)-n^{W}_{\mathbf{y}}P(\\{n^{W}_{\mathbf{x}},n^{P}_{\mathbf{x}}\\};t)\right]$ $\displaystyle+\sigma\sum_{\mathbf{y}\in\Gamma}\left[(n^{W}_{\mathbf{y}}-1)P(\\{n^{W}_{\mathbf{y}}-1,n^{P}_{\mathbf{x}}\\};t)-n^{W}_{\mathbf{y}}P(\\{n^{W}_{\mathbf{x}},n^{P}_{\mathbf{x}}\\};t)\right]$ $\displaystyle+\alpha\sum_{\mathbf{y}\in\Gamma}\left[n^{W}_{\mathbf{y}}P(\\{n^{W}_{\mathbf{x}},n^{P}_{\mathbf{y}}-1\\};t)-n^{W}_{\mathbf{y}}P(\\{n^{W}_{\mathbf{x}},n^{P}_{\mathbf{x}}\\};t)\right]$ $\displaystyle+\beta\sum_{\mathbf{y}\in\Gamma}\left[(n^{W}_{\mathbf{y}}+1)n_{\mathbf{y}}^{P}P(\\{n^{W}_{\mathbf{y}}+1,n^{P}_{\mathbf{x}}\\};t)-n^{W}_{\mathbf{y}}n^{P}_{\mathbf{y}}P(\\{n^{W}_{\mathbf{x}},n^{P}_{\mathbf{x}}\\};t)\right]$ $\displaystyle+\xi\sum_{\mathbf{y}\in\Gamma}\left[P(\\{n^{W}_{\mathbf{y}}-1,n^{P}_{\mathbf{x}}\\};t)-P(\\{n^{W}_{\mathbf{x}},n^{P}_{\mathbf{x}}\\};t)\right]\,.$ (2.3) The first line describes diffusion of walkers from one lattice site to one of its $q$ nearest neighbours with frequency $H/q$. This process is schematically shown in Fig. 7. $n^{W}_{\mathbf{y}}-1,n^{P}_{\mathbf{y}}$$n^{W}_{\mathbf{y}+e}+1,n^{P}_{\mathbf{y}+e}$$e$$H/q$$n^{W}_{\mathbf{y}},n^{P}_{\mathbf{y}}$$n^{W}_{\mathbf{y}+e},n^{P}_{\mathbf{y}+e}$$e$ Figure 7: Schematic representation of the process leading to the first line of Eq.(2.3): a single walker moving to a neighbouring lattice site (with $n^{W}_{\mathbf{y}}\geq 1$ and $n^{P}_{\mathbf{y}}\,,n^{W}_{\mathbf{y}+e}\,,n^{P}_{\mathbf{y}+e}\geq 0$). There $\\{n^{W}_{\mathbf{y}}-1,n^{W}_{\mathbf{y}-e}+1,n^{P}_{\mathbf{x}}\\}$ denotes the state differing from $\\{n^{W}_{\mathbf{x}},n^{P}_{\mathbf{x}}\\}$ by having a walker less at $\mathbf{y}$ and a walker more at $\mathbf{y}-e$. The second and third lines produce the first two branching processes in Eq. (2.2) respectively and are schematically shown in Figs 8 and 9. $n^{W}_{\mathbf{y}}-1,n^{P}_{\mathbf{y}}$$\sigma$$n^{W}_{\mathbf{y}},n^{P}_{\mathbf{y}}$ Figure 8: Schematic representation of the branching process leading to the second line of (2.3): a single walker creating a copy of itself at the the site $\mathbf{y}$ (with $n^{W}_{\mathbf{y}}\geq 2$ and $n^{P}_{\mathbf{y}}\geq 0$). $n^{W}_{\mathbf{y}},n^{P}_{\mathbf{y}}-1$$\alpha$$n^{W}_{\mathbf{y}},n^{P}_{\mathbf{y}}$ Figure 9: Schematic representation of the branching process leading to the third line of (2.3): a walker ’drops’ poison at the lattice site $\mathbf{y}$ (with $n^{P}_{\mathbf{y}}\geq 1$ and $n^{W}_{\mathbf{y}}\geq 0$). The fourth line accounts for the third process there and is graphically represented in Fig. 10. $n^{W}_{\mathbf{y}}+1,n^{P}_{\mathbf{y}}$$\beta$$n^{W}_{\mathbf{y}},n^{P}_{\mathbf{y}}$ Figure 10: Schematic representation of the branching process leading to the fourth line of (2.3): a single walker ’dying’ from poison at the lattice site $\mathbf{y}$ (with $n^{P}_{\mathbf{y}}\,,n^{W}_{\mathbf{y}}\geq 0$). Finally, the last line gives the spontaneous creation of one walker at site $\mathbf{y}$ and is schematically shown in Fig. 11. $n^{W}_{\mathbf{y}}-1,n^{P}_{\mathbf{y}}$$\xi$$n^{W}_{\mathbf{y}},n^{P}_{\mathbf{y}}$ Figure 11: Schematic representation of the branching process leading to the fifth line of (2.3): a single walker is spontaneously created at the lattice site $\mathbf{y}$ (with $n^{W}_{\mathbf{y}}\geq 1$ and $n^{P}_{\mathbf{y}}\geq 0$). In view of a second quantisation, following the Doi-Peliti approach [72, 73, 74], it is natural to interpret the state $\\{n^{W}_{\mathbf{x}},n^{P}_{\mathbf{x}}\\}$ as obtained by the action of creation operators $a^{\dagger}(\mathbf{x})$ (for $W$) and $b^{\dagger}(\mathbf{x})$ (for $P$) on a vacuum state. One introduces also the corresponding annihilation operators, $a(\mathbf{x})$ and $b(\mathbf{x})$, such that $\displaystyle a^{\dagger}(\mathbf{x})|\\{n^{W}_{\mathbf{x}},n^{P}_{\mathbf{x}}\\}\rangle=|\\{n^{W}_{\mathbf{x}}+1,n^{P}_{\mathbf{x}}\\}\rangle\,,$ $\displaystyle b^{\dagger}(\mathbf{x})|\\{n^{W}_{\mathbf{x}},n^{P}_{\mathbf{x}}\\}\rangle=|\\{n^{W}_{\mathbf{x}},n^{P}_{\mathbf{x}}+1\\}\rangle\,,$ (2.4) $\displaystyle a(\mathbf{x})|\\{n^{W}_{\mathbf{x}},n^{P}_{\mathbf{x}}\\}\rangle=n^{W}_{\mathbf{x}}\,|\\{n^{W}_{\mathbf{x}}-1,n^{P}_{\mathbf{x}}\\}\rangle\,,$ $\displaystyle b(\mathbf{x})|\\{n^{W}_{\mathbf{x}},n^{P}_{\mathbf{x}}\\}\rangle=n^{P}_{\mathbf{x}}\,|\\{n^{W}_{\mathbf{x}},n^{P}_{\mathbf{x}}-1\\}\rangle\,,$ (2.5) $\displaystyle\left[a(\mathbf{x}),a^{\dagger}(\mathbf{y})\right]=\delta_{\mathbf{x},\mathbf{y}}\,,$ $\displaystyle\left[b(\mathbf{x}),b^{\dagger}(\mathbf{y})\right]=\delta_{\mathbf{x},\mathbf{y}}\,,$ (2.6) with all other possible commutators between $(a,a^{\dagger})$ and $(b,b^{\dagger})$ vanishing. The field theory is realised by considering the time-evolution of the state $|\Psi(t)\rangle=\sum_{\\{n_{\mathbf{x}}^{W},n_{\mathbf{x}}^{P}\\}}P(\\{n^{W}_{\mathbf{x}},n^{P}_{\mathbf{x}}\\};t)\,|\\{n_{\mathbf{x}}^{W},n_{\mathbf{x}}^{P}\\}\rangle\;,$ (2.7) which can be derived from the Master Equation (2.3). Upon mapping each operator to conjugate fields $\displaystyle a\rightarrow W\,$ , $\displaystyle\quad\tilde{a}=a^{\dagger}-1\rightarrow W^{+}\;,$ $\displaystyle b\rightarrow P\,$ , $\displaystyle\quad\tilde{b}=b^{\dagger}-1\rightarrow P^{+}\;,$ (2.8) where the tilded operators are known as Doi-shifted operators, one finds that the evolution is controlled by $\exp\\{-\int d^{d}xdt\,S(W^{+},W,P^{+},P)\\}$, with the action density $S$ given by $\displaystyle S=$ $\displaystyle W^{+}\partial_{t}W+P^{+}\partial_{t}P+D\nabla W^{+}\nabla W-\sigma(1+W^{+})W^{+}W$ $\displaystyle-\alpha(1+W^{+})P^{+}W+\beta(1+P^{+})W^{+}WP-\xi W^{+}\;,$ (2.9) where$D=\lim_{\mathsf{a}\to 0}H\mathsf{a}^{2}/q$ is the hopping rate in the continuum ($\mathsf{a}$ is the lattice spacing). The action in Eq.(2.9) corresponds to the result in [76] augmented here by the last source term due to spontaneous generation. This produces a background of infected and it is responsible in this approach for a ‘strolling’ dynamics, as we motivate in Section 3.5.2. The renormalisation group equations stemming from the action in Eq.(2.9), which follow closely those of other theories such as directed percolation models or reggeon field theory [77, 78], have been analysed in [76]. In particular, the Fourier transform of the correlation function of a field $W$ and a field $W^{+}$ was computed and shown to satisfy the following scaling law near criticality $\displaystyle\mathcal{F}\left(\langle W(\vec{x},t)\,W^{+}(0,0)\rangle\right)(\omega,\vec{k})=|\vec{k}|^{\eta-2}\,\Phi(\omega\,\Delta^{\nu_{t}}\,,\vec{k}\,\Delta^{\nu})\,,$ (2.10) for some function $\Phi$. Here $\Delta$ is a measure for the proximity to criticality (_i.e._ it is proportional to $p-p_{c}$ of Eq. (2.1) in the context of the percolation model) and $(\eta,\nu_{t},\nu)$ are critical exponents determining the universality class of the model.111In a dimensional regularisation scheme, they were found to be [76] $\displaystyle\eta=-\frac{\epsilon}{21}\,,$ $\displaystyle\nu_{t}=1+\frac{\epsilon}{28}\,,$ $\displaystyle\nu=\frac{1}{2}-\frac{5}{84}\,\epsilon\,,$ (2.11) where $\epsilon=6-d$. The quantity above is a measure for the probability of finding a walker at some generic time and position $(\vec{x},t)\in\mathbb{R}^{6}$ if there was one at the origin, where $d=6$ corresponds to the critical dimension of the system [76]. ### 2.4 Relation to Compartmental Models As mentioned before, the model described by the action in Eq.(2.9) is in the same universality class as numerous other models that are directly relevant for the study of epidemic processes. As shown in [76] the particular choice $\xi=0$, in fact, includes the SIR model, which is the most prominent representative of compartmental models. To make the connection more concrete, we return to studying the time evolution of a disease on a lattice $\Gamma$ and divide the individuals that are present at a given lattice site $\mathbf{x}\in\Gamma$ into three classes or compartments [99], as defined in Section 2.2.1. We shall denote ${n^{S}_{\mathbf{x}},n^{I}_{\mathbf{x}},n^{R}_{\mathbf{x}}}$ the number of susceptible, infectious and removed individuals at $\mathbf{x}$, respectively. 222The occupation numbers $(n^{S}_{\mathbf{x}},n^{I}_{\mathbf{x}},n^{R}_{\mathbf{x}})$ are denoted $(X(\mathbf{x}),Y(\mathbf{x}),Z(\mathbf{x}))$ respectively in [99]. Concretely, for $\xi=0$, the model in [99] is very suitable for numerical Markovian simulations and can be connected to the SIR model. The processes of the model in [99] are $\displaystyle n^{S}_{\mathbf{x}}+n^{I}_{\mathbf{x}^{\prime}}$ $\displaystyle\rightarrow n^{I}_{\mathbf{x}}+n^{I}_{\mathbf{x}^{\prime}}\;,$ $\displaystyle{\rm infection\;with\;rate\;}\hat{\gamma}\;,$ $\displaystyle n^{I}_{\mathbf{x}}$ $\displaystyle\rightarrow n^{R}_{\mathbf{x}}\;,$ $\displaystyle{\rm recovery\;with\;rate\;}\hat{\epsilon}\;,$ (2.12) where $\mathbf{x}$ and $\mathbf{x}^{\prime}$ are nearest neighbour sites on $\Gamma$ (_i.e._ $\mathbf{x}^{\prime}=\mathbf{x}+e$ for some basis vector $e\in\mathbf{e}$). As discussed in [99], treating the process as deterministic (in particular, interpreting $(n^{S}_{\mathbf{x}},n^{I}_{\mathbf{x}},n^{R}_{\mathbf{x}})$ as continuous functions of time) one obtains the following equations of motion $\displaystyle\frac{dn^{S}_{\mathbf{x}}}{dt}(t)$ $\displaystyle=$ $\displaystyle-\hat{\gamma}\,n^{S}_{\mathbf{x}}(t)\sum_{e\in\mathbf{e}}n^{I}_{\mathbf{x}+e}(t)\;,$ $\displaystyle\frac{dn^{I}_{\mathbf{x}}}{dt}(t)$ $\displaystyle=$ $\displaystyle\hat{\gamma}\,n^{S}_{\mathbf{x}}(t)\sum_{e\in\mathbf{e}}n^{I}_{\mathbf{x}+e}(t)-\hat{\epsilon}\,n^{I}_{\mathbf{x}}(t)\;,$ $\displaystyle\frac{dn^{R}_{\mathbf{x}}}{dt}(t)$ $\displaystyle=$ $\displaystyle\hat{\epsilon}\,n^{I}_{\mathbf{x}}(t)\;,$ (2.13) where the sums on the right hand side extend over the nearest neighbours of $\mathbf{x}$. Since the sum of all three equations in (2.13) implies $\frac{d}{dt}(n^{S}_{\mathbf{x}}+n^{I}_{\mathbf{x}}+n^{R}_{\mathbf{x}})(t)=0$, the total number of individuals is conserved and we denote its value by $\displaystyle N=\sum_{\mathbf{x}\in\Gamma}(n^{S}_{\mathbf{x}}(t)+n^{I}_{\mathbf{x}}(t)+n^{R}_{\mathbf{x}}(t))\,.$ (2.14) Furthermore, we introduce the relative number of susceptible, infectious and removed individuals respectively $\displaystyle S(t)=\frac{1}{N}\,\sum_{\mathbf{x}\in\Gamma}n^{S}_{\mathbf{x}}(t)\,,$ $\displaystyle I(t)=\frac{1}{N}\,\sum_{\mathbf{x}\in\Gamma}n^{I}_{\mathbf{x}}(t)\,,$ $\displaystyle R(t)=\frac{1}{N}\,\sum_{\mathbf{x}\in\Gamma}n^{R}_{\mathbf{x}}(t)\,,$ (2.15) which satisfy $\displaystyle S(t)+I(t)+R(t)=1\,.$ (2.16) Finally, by taking a _mean-field approximation_ for the infected field in Eq.(2.13) (_i.e._ replacing $n^{I}_{\mathbf{x}}$ by $I(t)$ $\forall\mathbf{x}\in\Gamma$, such that the sums $\sum_{e\in\mathbf{e}}n^{I}_{\mathbf{x}+e}$ in Eq.(2.13) are replaced by $\frac{q}{N}\sum_{\mathbf{x}\in\Gamma}n^{I}_{\mathbf{x}}=qI(t)$) and summing over all $\mathbf{x}\in\Gamma$, one obtains the following coupled first order differential equations: $\displaystyle\frac{dS}{dt}(t)=-q\,\hat{\gamma}\,S(t)\,I(t)\,,$ $\displaystyle\frac{dI}{dt}(t)=q\,\hat{\gamma}\,S(t)\,I(t)-\hat{\epsilon}\,I(t)\,,$ $\displaystyle\frac{dR}{dt}(t)=\hat{\epsilon}\,I(t)\;,$ (2.17) where $q$ is the coordination number, _i.e._ , the number of nearest neighbours for each site ($4$ in a two-dimensional rectangular lattice). As we shall discuss in the next section, this system of differential equations, which has to be solved under the constraint in Eq.(2.16) and with suitable initial conditions, is structurally of the same form as the SIR model [19], one of the oldest deterministic models to describe the spread of a communicable disease. Spontaneous generation can be included in Eq.(2.17) as an additional process $\displaystyle n^{S}_{\mathbf{x}}\rightarrow n^{I}_{\mathbf{x}}\,,$ $\displaystyle\text{ with rate }\hat{\xi}\,.$ (2.18) In the deterministic and mean-field equations, this amounts to a term $-\hat{\xi}S(t)$ in the first equation of (2.17), and the corresponding one with opposite sign in the second equation, as we shall discuss in the context of the SIR model in Section 3.5.2. ## 3 Compartmental Models Executive Summary 1. We introduce compartmental models as _deterministic_ approaches that describe the diffusion of infectious diseases through coupled differential equations in time. 2. These models are characterised through a set of diffusion rates and initial conditions. Time is assumed to be a continuous variable. 3. We discuss mathematical aspects of different models capturing the asymptotic behaviour of their dynamics. 4. We show how to extract several epidemiologically relevant phenomena from compartmental models, including the endemic behaviour of diseases, the impact of superspreaders, the possibility of re-infection and multi wave patterns. 5. We show that compartmental models can be naturally related to the renormalisation group framework. ### 3.1 SIR(S) Model, Basic Definitions Independently of percolation models and epidemic field theory descriptions, the differential equations (2.17) have been proposed as early as 1927 to describe the dynamic spread of infectious diseases in an isolated population of total size $N\gg 1$. As reviewed in the Historical Overview (Section 1.1), a major breakthrough in the systematic study of the time evolution of infectious diseases was the application of the mass-action law [8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19], stating roughly that the rate at which individuals of two different types meet is proportional to the product of their total numbers. 333In fact H. Heesterbeek [5] remarks that ’In short, mass-action turned epidemiology into a science.’ This law underlies many chemical reactions, in which different agents mix. In the context of epidemiology, it leads to a class of deterministic approaches that are called _compartmental models_ , whose hallmark is to divide the population into several distinct classes. Each class or compartment is comprised of individuals that have peculiar behaviour in the context of the disease. The simplest of this class of models is called SIR and, as the name indicates, it includes three compartments, as already described in detail in Section 2.4: * 1. _S usceptible:_ the total number of susceptible individuals at time $t$ shall be denoted $N\,S(t)$. * 2. _I nfectious:_ the total number of infectious individuals at time $t$ shall be denoted $N\,I(t)$. * 3. _R emoved (recovered):_ the total number of removed individuals at time $t$ shall be denoted $N\,R(t)$. Depending on the type of disease under consideration, other compartments can be included to make the model more realistic, _e.g._ [98, 32] * 1. _Passively immune ( $M$):_ infants that have been born with (temporary) passive immunity ($M$ stands for maternally-derived immunity). * 2. _E xposed ($E$):_ individuals in the latent period, who are infected but not yet infectious. * 3. _D eceased ($D$):_ individuals who have died from the disease (in some models $D$ is considered to be part of $R$). * 4. _C arrier ($C$):_ individuals in a state where they have not completely recovered and still carry the disease, but do not suffer from it (examples of diseases for which this compartment is of relevance are tuberculosis and typhoid fever, see _e.g._ [108]). * 5. _Q uarantine ($Q$):_ individuals who have been put under quarantine or lockdown measures (see _e.g._ [109]). * 6. _V accinated ($V$):_ individuals who are vaccinated against the disease, thus acquiring partial or total immunity. Models are usually named/classified according to the compartments they contain, _e.g._ SIR, MESIR, _etc._ Repetition of labels indicates that individuals may return into a given compartment several times, _e.g._ SIS denotes a model in which infectious individuals may become susceptible again after an infection. Furthermore, each compartment can be generalised to include dependences on biological (_e.g._ age, gender, _etc._) and/or geometric parameters (_e.g._ parameters measuring geographic mobility, _etc._). To better model social and behavioural particularities among the population (but also to simulate different variants of a given disease), models can include multiple copies of a compartment with slightly different properties (see _e.g._ [48], which includes several different classes of susceptible, each of which with a different infection rate, to model the spread of gonorrhoea). Finally, depending on the duration of the epidemic, the birth and death dynamic needs to be taken into account [110, 111]. That is, into each class, new individuals may be born, or individuals of each compartment can die from causes other than the disease. In the following, for simplicity, we shall only consider models including the compartments S, I and R. We assume that the total size of the population remains constant, _i.e._ we impose the algebraic relation $\displaystyle 1=S(t)+I(t)+R(t)\,,$ $\displaystyle\forall t\in\mathbb{R}_{+}\,,$ (3.1) where, without restriction of generality, we assume that the outbreak of the epidemic starts at $t=0$. We shall also refer to $S$, $I$ and $R$ as the _relative_ number of susceptible, infectious and removed individuals, respectively. Furthermore, we assume that $N$ is sufficiently large such that we can treat $S$, $I$ and $R$ as continuous functions of time: $\displaystyle S\,,I\,,R\,:\hskip 28.45274pt\mathbb{R}_{+}\longrightarrow[0,1]\,.$ (3.2) While in Section 2.4 the differential equations (2.17) were a consequence of the basic microscopic processes in Eq.(2.12) on the lattice $\Gamma$, within compartmental models they are independently argued on the basis of dynamical mechanisms that change $(S,I,R)$ as functions of time: * 1. Infectious individuals can infect susceptible individuals, turning the latter into infectious individuals themselves. We call an ‘infectious contact’ any type of contact that results in the transmission of the disease between an infectious and a susceptible and we denote the average number of such contacts per infectious individual per unit of time by $\gamma$. In the original SIR model [19], $\gamma$ is considered to be constant (_i.e._ it does not change over time), however, in the following sections we shall not always limit ourselves to this restriction. The total number of susceptible individuals that are infected per unit of time (and thus become infectious themselves) is thus $\gamma\,N\,S\,I$. * 2. Infectious individuals can be removed by recovering (and thus gaining temporary immunity) or by being given immunity (_e.g_ via vaccinations), by death or via any other form of removal. We shall denote $\epsilon$ the rate at which infected individuals become removed. As before, we consider $\epsilon$ as a function that may change with time. * 3. Removed individuals may become susceptible again after some time or, conversely, susceptible individuals may become directly removed. In both cases we shall denote the respective rate by $\zeta$, which may be positive or negative. If removed individuals are only temporarily immune against the disease, they can become susceptible again. In this case $\zeta>0$, which corresponds to the rate at which removed individuals become susceptible again. Susceptible individuals may become immunised against the disease (_e.g._ through vaccinations). In this case $\zeta<0$. We remark that this is not the only way to implement vaccinations to compartmental models, as the most direct way is to add a specific compartment. The flow among susceptible, infectious and removed is schematically shown in Fig. 12. The dynamics of the system is also crucially determined by the initial conditions in each compartment. As already mentioned, we consider $t=0$ as the start of the epidemic diffusion, where a non-zero number of infectious individuals is needed for the diffusion to start. Without loss of generality, we start with zero removed at the initial time. Hence, the initial conditions are given by $\displaystyle S(t=0)=S_{0}\,,$ $\displaystyle I(t=0)=I_{0}\,,$ $\displaystyle R(t=0)=0\,,$ (3.3) where $S_{0},I_{0}\in[0,1]$ are constants that satisfy $S_{0}+I_{0}=1$. With this notation, the time dependence of $S$, $I$ and $R$ is described by the following set of coupled first order differential equations 444These equations coincide with Eq.(2.17) upon identifying $q\hat{\gamma}\equiv\gamma$, $\hat{\epsilon}\equiv\epsilon$, and for $\zeta=0$. Spontaneous generation of infectious individuals can be added straightforwardly. $\gamma\,N\,I\,S$$N\,S$$N\,I$$\epsilon N\,I$$N\,R$$\zeta\,N\,R$ Figure 12: Flow between susceptible, infectious and removed individuals. $\displaystyle\frac{dS}{dt}=-\gamma\,I\,S+\zeta\,R\,,$ $\displaystyle\frac{dI}{dt}=\gamma\,I\,S-\epsilon\,I\,,$ $\displaystyle\frac{dR}{dt}=\epsilon\,I-\zeta\,R\,,$ (3.4) together with the initial conditions (3.3). Notice that $\frac{d}{dt}(S(t)+I(t)+R(t))=0$ such that the initial conditions (3.3) with $S_{0}=I_{0}=1$ guarantee the algebraic relation (3.1). For $\zeta=0$, the system of equations (3.4) is indeed the same model as described in Section 2.4, which is called the SIR-model [19]. For $\zeta>0$, this model is sometimes referred to as the SIRS model, since it holds the possibility that recovered individuals may become susceptible again. ### 3.2 Numerical Solutions and their Qualitative Properties Figure 13: Numerical solution of the differential equations (3.4) for $S_{0}=0.92$, $\gamma=0.1$ and $\zeta=0$ for two different choices of $\epsilon$: $\epsilon=0.1001$ such that $R_{e,0}=0.919$ (left) and $\epsilon=0.05$ such that $R_{e,0}=1.84$ (right). The Eqs (3.4) can be solved analytically for $\zeta=0$, as we will discuss in the next subsection. First, we shall present some qualitative remarks that can be deduced by considering numerical solutions, which we obtained by using a simple forward Euler method (see _e.g._ [112, 113]). We first consider $\zeta=0$, for which the temporal evolution of $(S,I,R)$ is illustrated in Fig. 13 in two qualitatively different scenarios, depending on the value of the _initial effective reproduction number_ $R_{e,0}$, that we define as [114] (see also [115, 116, 117, 118, 119, 120] for further discussion of the effective reproduction number) $\displaystyle R_{e,0}=S_{0}\,\sigma\,,\qquad\sigma=\frac{\gamma}{\epsilon}\,.$ (3.5) The quantity $\sigma$, often called _basic reproduction number_ ($R_{0}$), can be interpreted as the average number of infectious contacts of a single infectious individual during the entire period they remain infectious. In other words, $\sigma$ is the average number of susceptible individuals infected by a single infectious one. In the left panel of Fig. 13, $(\gamma,\epsilon,S_{0})$ have been chosen such that $R_{e,0}<1$: in this case, even though at initial time a significant fraction of the population ($8\%$) is infectious, the function $I(t)$ decreases continuously, leading to a relatively quick eradication of the disease. This is also visible directly from Eqs (3.4): since (for $\zeta=0$) $S(t)$ is a monotonically decreasing function (_i.e._ $S(t)\leq S_{0}$ $\forall t>0$), then $\frac{dI}{dt}(t)<0$ $\forall t>0$ such that the number of infectious individuals is continuously decreasing. In the right panel of Fig. 13, we chose $R_{e,0}>1$: the number of infectious cases grows to a maximum and starts decreasing once only a small number of susceptible individuals remain available. This maximum is reached once $S(t)=\frac{1}{\sigma}$ such that $\frac{dI}{dt}=0$. This behaviour is more clearly visible in the asymptotic number of susceptible (_i.e._ $S(\infty)=\lim_{t\to\infty}S(t)$) or (equivalently) the cumulative number of individuals that have become infected throughout the entire epidemic. Both quantities are a measure of how far the disease has spread among the population. For later use, we define the function $I_{\text{c}}(t):\,[0,\infty)\mapsto[0,N]$ as $\displaystyle I_{\text{c}}(t)=N\,I_{0}+\int_{0}^{t}dt^{\prime}\,\gamma\,N\,I(t^{\prime})\,S(t^{\prime})\,.$ (3.6) It quantifies the cumulative total number of individuals who have been infected by the disease up to time $t$. The definition (3.6) can be used for generic $\zeta$ as a function of time. For $\zeta=0$, using Eqs (3.4), we obtain the identity $\gamma\,I\,S=\frac{d}{dt}(I+R)$ that allows to simplify Eq.(3.6) to: $\displaystyle I_{\text{c}}(t)=N(I(t)+R(t))=N(1-S(t))\,,$ for $\displaystyle\zeta=0\,.$ (3.7) For $\zeta=0$, we also have that $\lim_{t\to\infty}I(t)\to 0$, thus we find the following relations at infinite time: $\displaystyle I_{\text{c}}(\infty)=\lim_{t\to\infty}I_{\text{c}}(t)=1-S(\infty)=R(\infty)=\lim_{t\to\infty}R(t)\,.$ (3.8) Figure 14: Asymptotic number of susceptible and cumulative number of infectious as a function of $R_{e,0}$ for $S_{0}=1-10^{-6}$. The limit $S(\infty)$ can be computed analytically, by realising that $\displaystyle G(t)=S(t)\,e^{\sigma\,R(t)}\,,$ (3.9) is conserved, _i.e._ $\frac{dG}{dt}(t)=0$ $\forall t\in\mathbb{R}$. This implies $\displaystyle S(t)=S_{0}\,e^{-\sigma(1-I(t)-S(t))}\,.$ (3.10) With $\lim_{t\to\infty}I(t)=0$, this equation can be solved for the asymptotic number of susceptible in the limit $t\to\infty$, giving $\displaystyle S(\infty)=-\frac{S_{0}}{R_{e,0}}\,W(-R_{e,0}\,e^{-\frac{R_{e,0}}{S_{0}}})\,,$ (3.11) where $W$ is the Lambert function. The limiting values $S(\infty)$ and $I_{\text{c}}(\infty)/N$ are shown in Fig. 14 as functions of $R_{e,0}$ for the initial conditions of $S_{0}=1-10^{-6}$, _i.e._ a starting configuration with one infectious individual per million. A kink seems to appear for $R_{e,0}=1$, however both functions are smooth (continuous and differentiable) for $S_{0}<1$, as highlighted in the subplots. In the limit $S_{0}\to 1$, the solutions discontinuously jump to constants, as the absence of initial infectious individuals prevents the spread of the disease. Qualitatively, this plot shows that for $R_{e,0}<1$, the disease becomes eradicated before a significant fraction of the population can be infected. However for $R_{e,0}>1$ the cumulative number of infected grows rapidly. For $\zeta\neq 0$, we can distinguish two different cases, depending on the sign: * 1. Re-infection $\zeta>0$: a positive $\zeta$ implies that removed individuals become susceptible again after some time. This can be interpreted to mean that recovery from the disease only grants temporary immunity, such that a re- infection at some later time is possible. At large times $t\to\infty$, the system enters into an equilibrium state, such that $(S(t)\,,I(t)\,,R(t))$ approach constant values $(S(\infty)\,,I(\infty)\,,R(\infty))$. To find the latter, we impose the equilibrium conditions $\displaystyle\lim_{t\to\infty}\frac{d^{n}S}{dt^{n}}(t)=\lim_{t\to\infty}\frac{d^{n}I}{dt^{n}}(t)=\lim_{t\to\infty}\frac{d^{n}R}{dt^{n}}(t)=0\,,$ $\displaystyle\forall n\in\mathbb{N}\,,$ (3.12) which have as solution $\displaystyle(S(\infty),I(\infty),R(\infty))=\left\\{\begin{array}[]{lcl}(1,0,0)&\text{if}&\sigma\leq 1\text{ or }S_{0}=1\,,\\\\[10.0pt] \left(\frac{\epsilon}{\gamma}\,,\frac{(\gamma-\epsilon)\zeta}{\gamma(\epsilon+\zeta)}\,,\frac{(\gamma-\epsilon)\epsilon}{\gamma(\epsilon+\zeta)}\right)&\text{if}&\sigma>1\,,\end{array}\right.$ for $\displaystyle\zeta>0\,.$ (3.15) Here we have used that $0\leq(S(t)\,,I(t)\,,R(t))\leq 1$ (in particular that $(S(t)\,,I(t)\,,R(t))$ cannot become negative) as well as the fact that the equilibrium point $(1,0,0)$ cannot be reached for $S_{0}<1$ and $\gamma>\epsilon$: indeed, this would require $\displaystyle S(t)>\frac{\epsilon}{\gamma}\,,$ and $\displaystyle\frac{dI}{dt}(t)<0\,,$ (3.16) which are not compatible with Eqs (3.4). 555Furthermore, the only solutions of the conditions $\frac{d^{2}S}{dt^{2}}(t)=\frac{dI}{dt}(t)=\frac{d^{2}R}{dt^{2}}(t)=0$ are in fact the two equilibrium points (3.15) (where in fact all derivatives of $(S\,,I\,,R)$ vanish). This therefore suggests that there are no solutions that are continuous oscillations with non-decreasing amplitudes and the system indeed reaches an equilibrium at $t\to\infty$. The numerical solutions in Fig. 15 comply with this expectation. The two qualitatively different solutions of Eqs (3.4) that lead to the asymptotic equilibria (3.15) are plotted in Fig. 15: for $\sigma<1$ (left panel), the disease is eradicated and the individuals that have been infected eventually move back to be susceptible; for $\sigma>1$ (right panel), after some oscillations, an equilibrium is reached between the infections and the end of immunity and the number of infectious individuals tends to the non-zero constant given in Eq.(3.15) (this corresponds to an endemic state of the disease). The distinction between eradication of the disease and the endemic phase does not depend on $S_{0}$ (except for the trivial initial condition $S_{0}=1$) but only on the basic reproduction number $\sigma$. This fact can be intuitively understood as the rate $\zeta$ dynamically increases the number of susceptible individuals, thus the regime becomes independent of the initial condition. Figure 15: Numerical solution of the differential equations (3.4) for $S_{0}=0.92$, $\gamma=0.1$ and $\zeta=0.01$ for two different choices of $\epsilon$: $\epsilon=0.2$ implying $\sigma=0.5$ (left) and $\epsilon=0.05$ implying $\sigma=2$ (right). * 2. Direct immunisation $\zeta<0$: a negative $\zeta$ implies the possibility that over time susceptible individuals can become removed and thus immune to the disease, proportionally to the number of removed individuals. Schematically, different solutions are shown in Fig. 16. For $\zeta<0$ the dynamics always leads to the asymptotic values $(S(\infty)\,,I(\infty)\,,R(\infty))=(0,0,1)$ at large $t\to\infty$. Figure 16: Numerical solution of the differential equations (3.4) for $S_{0}=0.92$, $\gamma=0.1$ and $\zeta=-0.01$ for two different choices of $\epsilon$: $\epsilon=0.2$ implying $\sigma=0.5$ (left) and $\epsilon=0.05$ implying $\sigma=2$ (right). ### 3.3 From Lattice to SIR The relation between Compartmental Models and Percolation Field Theory has already been established in Section 2.4. However it is also possible to link the numerical simulations to the SIR model directly, as the microscopic processes in the lattice simulations are in one-to-one correspondence with the transfer mechanisms among compartments in the SIR model. To visualise this we used the results in Fig. 4, where the lattice is of size $201\times 201$ (_i.e._ a population of $40401$) and the recovery probability is fixed to $0.1$. Once the recovery rate and the initial number of susceptible individuals $S_{0}$ is fixed, in the SIR model the value of the infection rate completely determines the asymptotic number of total infected via Eq.(3.11). For each coordination radius, we look for the best rescaling of the infection probability that could reproduce the behaviour in Fig. 4, _i.e._ we compute the optimal $\rho$ such that changing $\mathfrak{g}\longrightarrow\rho\mathfrak{g}$ gives the best fit of the numerical results. We show the solution in Fig. 17. (a) $r=1$ (b) $r=2$ (c) $r=5$ (d) $r=50$ Figure 17: Evolution of the final number of infected cases as a function of the infection probability for different coordination radii $r$, compared to the asymptotic solution of the SIR model. The optimal factor found for the cases (a),(b),(c) and (d) are respectively: $\rho=0.27,\;0.42,\;0.50,\;0.99$. The results clearly show that increasing the coordination radius improves the match between the lattice and the SIR model results. The reason for this is simple: for maximal coordination radius, the mean-field approximation applied to Eq. (2.13) leads directly to the SIR equations. The reason is that any infectious site can infect any susceptible site on the lattice with equal probability. Numerical lattice simulations of compartmental models, and in particular of the SIR type, have been widely used in the literature (see _e.g._ [121, 122, 123, 124]). ### 3.4 Parametric Solution of the Classical SIR Model Apart from the numerical solutions, we can also gain insight into analytical aspects by discussing a parametric solution of the classical SIR model [30]. For simplicity, we assume $\zeta=0$, such that the system in Eqs (3.4), (3.1) and (3.3) reduces to $\displaystyle\begin{array}[]{l}\frac{dS}{dt}(t)=-\gamma\,I(t)\,S(t)\,,\\\\[2.0pt] \frac{dI}{dt}(t)=\gamma\,I(t)\,S(t)-\epsilon\,I(t)\,,\\\\[2.0pt] \frac{dR}{dt}(t)=\epsilon\,I(t)\,,\end{array}$ with $\displaystyle(S+I+R)(t)=1$ and $\displaystyle\begin{array}[]{l}S(t=0)=S_{0}>0\,,\\\\[2.0pt] I(t=0)=I_{0}>0\,,\\\ R(t=0)=0\,.\end{array}$ (3.23) Since the constraint in Eq.(3.1) allows to remove one function, _e.g._ $R(t)=1-S(t)-I(t)$, it is sufficient to consider the differential equations for $S$ and $I$. Dividing the latter by the former, we obtain a differential equation for $I$ as a function of $S$ $\displaystyle\frac{dI}{dS}=-1+\frac{1}{\sigma\,S}\,,$ (3.24) which can be integrated to $\displaystyle I(S)=-S+\frac{1}{\sigma}\,\ln S+c\,,$ for $\displaystyle c\in\mathbb{R}\,.$ (3.25) The parameter $\sigma$ is defined in Eq.(3.5) and the constant $c$ appearing in Eq.(3.25) can be fixed by the initial conditions at $t=0$ and gives $c=I_{0}+S_{0}-\frac{1}{\sigma}\,\ln S_{0}$, such that $\displaystyle I(S)=1-S+\frac{1}{\sigma}\,\ln\frac{S}{S_{0}}\,.$ (3.26) A plot of this function in the allowed region $\displaystyle\mathbb{P}=\\{(S,I)\in[0,1]\times[0,1]|S+I\leq 1\\}\,,$ (3.27) for different initial conditions and $\sigma=0.9$ (left) and $\sigma=3$ (right) is shown in Fig. 18. Figure 18: Relative number of infectious $I$ as a function of the relative number of susceptible $S$ for $S_{0}\in\\{0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9\\}$ and $\sigma=0.9$ (left) as well as $\sigma=3$ (right). Curves with a local maximum are drawn in orange while curves that are monotonically growing within $\mathbb{P}$ are drawn in blue. These plots once more highlight the qualitatively different solutions: the solution $I(S)$ in Eq.(3.26) has a maximum at $I_{\text{max}}=1-\frac{1}{\sigma}\left(1+\ln(\sigma S_{0})\right)$, which lies inside of $\mathbb{P}$ only if the initial effective reproduction number defined in Eq. (3.5) is $R_{e,0}\equiv\sigma S_{0}\geq 1$. Since $S(t)$ is a monotonically decreasing function of time, as demonstrated in [30], this implies that: * 1. If $R_{e,0}\leq 1$, then $I(t)$ tends to $0$ monotonically for $t\to\infty$, as already established before. * 2. If $R_{e,0}>1$, $I(t)$ first increases to a maximum equal to $1-\frac{1}{\sigma}\left(1+\ln(\sigma S_{0})\right)$ and then decreases to zero for $t\to\infty$. The limit $S(\infty)=\lim_{t\to\infty}S(t)$ is the unique root of $\displaystyle 1-S(\infty)+\frac{1}{\sigma}\,\ln\left(\frac{S(\infty)}{S_{0}}\right)=0\,,$ (3.28) in the interval $[0,\tfrac{1}{\sigma}]$, which is explicitly given in terms of the Lambert function in Eq.(3.11). Furthermore, inserting the solution (3.26) into Eq.(3.23), we obtain the following non-linear, first order differential equation for $S$ (as a function of time) $\displaystyle\frac{dS}{dt}=\gamma\,S(S-1)-\gamma\,\frac{S}{\sigma}\,\ln\frac{S}{S_{0}}\,.$ (3.29) The latter can be solved numerically using various methods. ### 3.5 Generalisations of the SIR Model The SIR model, with 3 compartments $(S,I,R)$ and constant rates $\gamma$, $\epsilon$ and $\zeta$, provides a simple, but rather crude, description of the time evolution of an epidemic in an isolated population. This description can be refined and extended in various fashions. The most common way consists in adding more compartments, with more refined properties, giving birth to models like SIRD (including Deceased separately), SEIR (including Exposed individuals, in presence of a substantial incubation period), SIRV [125, 126] (see also [127]) (including vaccinated individuals), an so on [32]. Here, as an illustration, we shall discuss some generalisations of the SIR model that do not introduce fundamentally new compartments: in Section 3.5.1 we shall allow for time-dependent infection and recovery rates, in Section 3.5.2 we shall include new terms in the differential equations (3.4) that simulate the spontaneous appearance of new infectious (_e.g._ from outside of the population), while in Section 3.5.3 we allow for multiple different types of infectious individuals in an attempt to model inhomogeneous spreading of the disease among the population. While these variations add new compartments to the system, these are not of a completely new nature but simply copy an already existing compartment. In all cases we shall motivate how these modifications can be used to describe specific features of certain diseases. For more general compartmental models (notably with the addition of completely new compartments) we refer the reader to the above mentioned literature (see _e.g._ [32] for an overview). Another generalisation is the inclusion of the spatial evolution of the disease. This generally leads to coupled differential equations which are of first order in the time variable and of second order in the spatial variable. We shall not discuss these approaches in any detail in this review. #### 3.5.1 Time Dependent Infection and Recovery Rates In the SIR model of Eqs (3.4), the rates $(\gamma,\epsilon,\zeta)$ are considered to be constant in time. This assumption is difficult to justify, in particular for epidemics that last over an extended period of time: many diseases show (natural) seasonal effects [128, 129] related to the weather dependence of the effectiveness of transmission vectors or the behaviour of hosts (_e.g._ it can be argued that the rate of child infections is linked to the cycle of school holidays [130]). Furthermore, even in the absence of an effective vaccine, populations may take measures to prevent the spread of the disease by imposing social distancing rules or quarantine procedures, thus changing the (effective) infection rate $\gamma$. Pathogen mutations and various forms of immunisations (including vaccines) can also increase or reduce the value of $\gamma$ over time. With a prolonged duration of an epidemic, more data about the disease can be collected, leading to better ways to fight it on a biological and medical level, thus changing the recovery rate $\epsilon$. Similarly, the disease may mutate and bypass previous immunisation strategies, thus changing the rate $\zeta$ at which removed individuals may become susceptible again. Modelling such effects and gauging their impact on the time evolution of an epidemics requires $(\gamma,\epsilon,\zeta)$ to change over time. In practice, this can be achieved by either interpreting them as explicit functions of $t\in\mathbb{R}$, _i.e._ $(\gamma(t),\epsilon(t),\zeta(t))$, or by considering them to be functions of the relative number of susceptible and/or infectious individuals, _i.e._ $(\gamma(S,I),\epsilon(S,I),\zeta(S,I))$. Since $(S,I)$ themselves are functions of time, the latter possibility induces an implicit dependence on $t$. For example, periodic and seasonal models in which these rates are assumed to be smoothly varying functions in $t$ have been developed for HIV [131], tuberculosis [132] or cutaneous leishmaniasis [133], while models for pulse-vaccinations have been proposed in [134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144] (a model which in addition takes into account seasonal effects was presented in [145]). The functional dependence can furthermore be used, for example, to model population-wide lockdowns, _i.e._ quarantine measures that are imposed if the relative number of infectious individuals exceeds a certain value. In the following we shall provide a simple (numerical) example of how the time dependence of different Figure 19: Numerical solution of the SIR equations (3.4) for the time- dependent infection rate (3.30) with $S_{0}=0.99$, $\epsilon=0.05$, $\zeta=0$, $\gamma_{0}=0.1$, $w=0.1$ and $\Delta I=0.05$. parameters affects the time-evolution of the pandemic. We start by a simple model that can be used to qualitatively assess the efficiency of lockdown measures. To this end, we assume a ‘base’ infection rate $\gamma_{0}=$const., but assume that the population takes measures (social distancing, lockdowns, _etc._) to ensure that the actual infection rate $\gamma(t)$ is reduced by a percentage $w$ if the number of (active) infectious individuals exceeds a certain value $\Delta I$. To model such social distancing measures in a very simplistic fashion, we introduce the following implicit time-dependence: $\displaystyle\gamma(I)=\gamma_{0}\,\left[1-w\,\theta(I(t)-\Delta I)\right]\,,$ (3.30) where $\theta$ is the Heaviside theta-function. 666To be mathematically rigorous, since $\theta$ is not a continuous function, using this infection rate in Eqs (3.4) would require to interpret $(S(t),I(t),R(t))$ as distributions. This can be circumvented by replacing $\theta(I(t)-\Delta I)$ by $1+\tanh(\kappa_{0}(I(t)-\Delta I))$ with $\kappa_{0}$ a parameter that ‘smoothens’ the step function. For the following discussion, however, this point shall not be relevant. We hasten to add that Eq.(3.30) offers a very crude depiction of lockdown and quarantine measures taken by societies in the real-world: indeed, decisions on whether or not to impose a lockdown (or other social distancing measures) are usually based on numerous indicators which would (at least) require a more complicated dependence of $\gamma$ on $I$ (_e.g._ its derivatives or averages of $I$ over a certain period of time prior to $t$). Furthermore, the conditions when a lockdown is lifted are typically independent of those when it is imposed. An exemplary numerical solution of Eqs (3.4) for the particular $\gamma$ in Eq.(3.30) is shown in Fig. 19. For better comparison we have also plotted $I_{\text{no-q}}(t)$, which is the solution for $I(t)$ in the case of constant $\gamma=\gamma_{0}=\mbox{const.}$ (_i.e._ with no reduction of the infection rate) and all remaining parameters chosen the same. Despite its simplicity and shortcomings, the model allows to make a few basic observations: the plot shows that the time-dependent infection rate leads to a reduction of the maximum of infectious individuals (‘flattening of the curve’). Moreover, this simple model allows to compare the effectiveness of the quarantine measures as a function of $w$ and $\Delta I$. To gauge this effectiveness, we consider the cumulative number of infected individuals, which is plotted for different values of $w$ and $\Delta I$ in Fig. 20. These plots confirm the intuitive expectation that lockdown measures are the more effective the stronger the reduction of the infection rate is and the earlier they are introduced. However, due to its simplicity, the model also misses certain aspects compared to the time evolution of real-world communicable diseases in the presence of measures to prevent its spread: for example, possibly due to non-zero incubation time of most infectious diseases, the effect of quarantine measures on the number of infectious individuals can be detected only a certain time after the measures have been imposed (see [146, 147, 148, 149] where this has been established for the COVID-19 pandemic). To include the latter would require a refinement of the model. Figure 20: Numerical solution of the SIR equations (3.4) for the time- dependent infection rate (3.30), with $S_{0}=0.99$, $\epsilon=0.05$, $\zeta=0$, $\gamma_{0}=0.1$ and different choices of $(w,\Delta I)$: $w\in\\{0.05\,,0.1\,,0.5\\}$ and $\Delta I=0.05$ (left) and $w=0.25$ and $\Delta I\in\\{0.01\,,0.05\,,0.1\,,1\\}$ (right). Figure 21: Numerical solution of the differential equations (3.31) for $S_{0}=0.99$, $\gamma=0.055$ and $\zeta=0.045$ for two different choices of $\xi$: $\xi=0$ (left) and $\xi=0.002$ (right). #### 3.5.2 Spontaneous Creation and Multiple Waves In Section 2.3, in the context of percolation models, we have discussed microscopic processes that correspond to the spontaneous creation of infected individuals. Such processes can simulate, for example, the infection of individuals through external sources (_e.g._ pathogen sources, contaminated food sources, wildlife, _etc_.), but may also be used to model the infection of susceptible individuals through asymptomatic infectious individuals or the appearance of infectious individuals from outside of the population through travel. How to introduce this process in SIR-type models has been discussed at the end of Section 2.4. Mathematically, the SIR equations (3.4) can be extended to $\displaystyle\frac{dS}{dt}=-\gamma\,I\,S+\zeta\,R-\xi\,S\,,$ $\displaystyle\frac{dI}{dt}=\gamma\,I\,S-\epsilon\,I+\xi\,S\,,$ $\displaystyle\frac{dR}{dt}=\epsilon\,I-\zeta\,R\,,$ (3.31) This is schematically shown in Fig. 21, where we show the solutions for $\xi=0$ (left panel) compared to Figure 22: Numerical solution of the differential equations (3.31) for $S_{0}=0.99$, $\gamma=0.055$, $\zeta=0.045$ and $\xi=0.002\,\left|\sin\left(\tfrac{2\pi t}{200}\right)\right|$. where the rate $\xi=\hat{\xi}$ of Section 2.4. The system still needs to be solved with the initial conditions (3.3). Here $\xi\in\mathbb{R}_{+}$ is a constant that governs the rate at which new infectious individuals appear in the population, corresponding to a qualitative change in the basic infection mechanism: since susceptible individuals can contract the disease even if there are no infectious individuals present in the population, the epidemic can not be stopped before the entire population becomes infected. As a consequence, the cumulative number of infected tends to $N$ for $t\to\infty$. the solution for $\xi\neq 0$ (right panel). In the former case, the number of cumulative infected tends to a finite value, while in the latter case, $\lim_{t\to\infty}S(t)\to 0$. Following the discussion in Section 3.5.1, we can also analyse the effect of a time-dependent rate $\xi(t)$. This can be used to model a time-dependent rate of the spontaneous creation of new infectious individuals, _e.g._ induced by quarantine measures or geographical restrictions of the population. As a simple example, we have plotted the numerical solution for a periodic function $\xi$ in Fig. 22. Since $\xi$ does not remain zero after finite time, the relative number of susceptible tends to $0$ (indicating that the entire population is infected for $t\to\infty$). Moreover, the solution features oscillations in time, which could be interpreted as different waves of the epidemic spreading in the population. #### 3.5.3 Heterogeneous Transmission Rates and Superspreaders As another generalisation of compartmental models, we consider adding multiple versions of the compartments $S$, $I$, $R$ [150, 151] to model the heterogeneity of social interactions and their impact on the spread of a disease: indeed, indications for _superspreaders_ (_i.e._ individuals who transmit the disease with a significantly higher rate than average) have been found in many diseases (_e.g._ influenza [152, 153], rubella [154]) and for certain diseases it has in fact been suggested that only a small fraction of the population is responsible for most infections (see _e.g._ [155, 156] for a study of COVID-19). Similarly, the gender of individuals plays an important role in the modelling of sexually transmitted diseases (see _e.g._ [157, 158, 159, 160] for the study of gonorrhoea, which also suggests the necessity of an extended range of contact rates [150]). To account for these modified contact rates, modifications of the SIR model (as described above) have been suggested, which consist in adding multiple compartments of infectious individuals, _i.e._ new subgroups that allow to refine the study of the disease spread in a not-so-uniform population. These additional compartments can, $N\,S$$N\,I_{2}$$N\,I_{1}$$N\,R$$\beta(\gamma_{1}I_{1}+\gamma_{2}I_{2})\,N\,S$$(1-\beta)(\gamma_{1}I_{1}+\gamma_{2}I_{2})\,N\,S$$\epsilon\,N\,I_{1}$$\epsilon\,N\,I_{2}$$\zeta\,N\,R$ Figure 23: Flow between susceptible, 2 compartments of infectious and removed individuals. therefore, distinguish individuals based on biological/medical indicators (_e.g._ gender, age, preexistent medical conditions, etc.), geographic distribution, social behaviour and/or may be used to introduce additional stages in the progression of the disease, such as latency periods or different stages of symptoms. Inclusion of more compartments naturally renders the relevant set of differential equations more complicated and is more demanding in terms of computational costs (see [161] as an example). Furthermore, the increase in the number of parameters (rates) leads to a loss of predictive power compared to simpler models. In the following we shall present one simple example that includes one additional class of infectious individuals. This model is useful in characterising different (social) behaviours among individuals. Indeed, in general, the infection rate $\gamma$ is not homogeneous throughout the entire population, since it depends on various factors such as geographical mobility, social behaviour _etc._ , which may vary considerably. A particular effect in this regard is the existence of so-called _superspreaders_. These are individuals who are capable of transmitting the disease to susceptible individuals at a rate that significantly exceeds the average. The presence of superspreaders can be described by introducing two groups of infectious individuals $I_{1,2}$, with different infection rates $\gamma_{1,2}$ and appearing with a relative ratio $\beta\in[0,1]$. Extending Fig. 12, the new flow among compartments is shown in Fig. 23 (for $\zeta=0$), and can be described by the following differential equations [150]: $\displaystyle\frac{dS}{dt}=-(\gamma_{1}\,I_{1}+\gamma_{2}\,I_{2})\,S\,,$ $\displaystyle\frac{dI_{1}}{dt}=\beta(\gamma_{1}\,I_{1}+\gamma_{2}\,I_{2})\,S-\epsilon\,I_{1}\,,$ $\displaystyle\frac{dI_{2}}{dt}=(1-\beta)(\gamma_{1}\,I_{1}+\gamma_{2}\,I_{2})\,S-\epsilon\,I_{2}\,,$ $\displaystyle\frac{dR}{dt}=\epsilon(I_{1}+I_{2})\,,$ (3.32) together with the initial conditions $\displaystyle S(t=0)=S_{0}\,,$ $\displaystyle I_{1}(t=0)=I_{0,1}\,,$ $\displaystyle I_{2}(t=0)=I_{0,2}\,,$ $\displaystyle R(t=0)=0\,,$ (3.33) with $\displaystyle 0\leq S_{0},I_{0,1},I_{0,2}\leq 1\,,$ $\displaystyle 1=S_{0}+I_{0,1}+I_{0,2}\,.$ (3.34) In [150] the parameters $\beta$, $\gamma_{1,2}$, and $\epsilon$ were assumed to be constant in time. By defining an effective infectious population $J=(\gamma_{1}\,I_{1}+\gamma_{2}\,I_{2})/\lambda$, we can extract the following differential equations for $(S,J)$ 777Note that our definition of $J$ differs from the definition of the infective potential $J=\gamma_{1}\,I_{1}+\gamma_{2}\,I_{2}$ in [150] by a constant normalisation. $\displaystyle\frac{dS}{dt}=-\lambda\,J\,S\,,$ $\displaystyle\frac{dJ}{dt}=\lambda\,J\,S-\epsilon\,J\,,$ with $\displaystyle\lambda=\gamma_{1}\,\beta+(1-\beta)\,\gamma_{2}\,.$ (3.35) Thus, for $S$ and $J$ we obtain the same equations as in the classical SIR model, which can be solved along the lines of Section 3.4: we extract the following non-linear first-order equation for $S$: $\displaystyle\frac{dS}{dt}=\lambda\,S^{2}-\epsilon\,S\,\ln S+\mathfrak{c}_{0}\,S\,,$ with $\displaystyle\mathfrak{c}_{0}=\epsilon\,\ln S_{0}-\lambda\,S_{0}-(\gamma_{1}I_{0,1}+\gamma_{2}I_{0,2})\,.$ (3.36) which leads to the asymptotic number of susceptible $S(\infty)$ implicitly given by $\displaystyle 0=\lambda\,S(\infty)-\epsilon\,\ln S(\infty)+\mathfrak{c}_{0}\,.$ (3.37) As was pointed out in [150], the SIR model with superspreaders leads to the same dynamics as the classical SIR models, albeit with a larger-than-average infection rate $\lambda$, due to the contribution of superspreaders. With constant infection and recovery rates and monotonically diminishing number of susceptible (_i.e._ for $\zeta=0$), the impact of superspreaders is conceptually not detectable. Nevertheless, from the perspective of the total number of infected, superspreaders may have a significant impact in driving the epidemics. In Fig. 24 (left) we have plotted the time evolution of a typical solution, which indeed follows the same pattern as the usual SIR model. However, as visible from Fig. 24 (right), even the presence of a relatively small number of superspreaders can have a strong impact on the cumulative number of infected. Figure 24: Numerical solution of the SIR equations in the presence of superspreaders, Eqs (3.32): time evolution for $S_{0}=0.99$, $I_{0,1}=0.01$, $I_{0,2}=0$, $\gamma_{1}=0.04$, $\gamma_{2}=1$, $\epsilon=0.05$ and $\beta=0.95$ (left) and comparison of the cumulative number of infected with the ‘usual’ SIR model without superspreaders (_i.e._ $\beta=1$) (right). Finally, it was argued in [150] that in situations in which the number of susceptible individuals is no longer a monotonical function (which can for example be achieved by allowing for a non-trivial $\zeta$), the time evolution of the SIR model looks qualitatively different in the presence of superspreaders. ### 3.6 The SIR model as a set of Renormalisation Group Equations As we have seen from simple numerical studies in Section 3.2, solutions $(S(t),I(t),R(t))$ of the classical SIR equations (3.4) exhibit interesting properties as functions of time, which structurally remain valid for many of the generalisations discussed in Section 3.5. In particular, the solutions show a qualitatively different behaviour when a key parameter (in the classical SIR model, the initial effective reproduction number $R_{e,0}=S_{0}\sigma$) exceeds a critical value. This seems to play a similar role to an ordering parameter in physical systems undergoing a phase transition. A further related observation is the fact that Eqs (3.4) are invariant under a re-scaling of the time-variable, if simultaneously all the rates are also re-scaled: $\displaystyle t\rightarrow\frac{1}{\mu}\,t\,,$ $\displaystyle\gamma\to\mu\,\gamma\,,$ $\displaystyle\epsilon\to\mu\,\epsilon\,,$ $\displaystyle\zeta\to\mu\,\zeta\,,$ $\displaystyle\forall\mu\in\mathbb{R}\setminus\\{0\\}\,.$ (3.38) This rescaling of the time-variable is structurally not unlike the change of the energy scale in quantum field theories that is used to describe the _Wilsonian renormalisation_ of the couplings among elementary particles [86, 87]. The renormalisation flow can also feature similar symmetries to the ones of the solutions of the SIR equations. Compartmental models can be formulated in a way that is structurally similar to Renormalisation Group Equations (RGEs) [92, 162], and this analogy lead to the formulation of an effective description called _epidemiological Renormalisation Group_ [92, 93], which we will introduce in the next section. To understand the analogy, we recall that most (perturbative) quantum field theories are effective models: they are typically based on an action that encodes fundamental interactions of certain ‘bare’ fundamental fields, whose strength is described by a set of coupling constants $\\{\lambda_{i}\\}$ (where $i$ takes values in a suitable set $\\{\mathcal{S}\\}$). Each effective description, however, is generally well adapted only at a certain energy scale, beyond which new degrees of freedom are more appropriate and new interactions may become important. In practice, one introduces a cut-off parameter (or some other regularisation form), beyond which the effective description is no longer valid. The effective theory can thus be interpreted as encoding all effective interactions, after having integrated out all interactions at energy scales higher than the cut-off. From this perspective it is clear that changing the energy scale (and thus the cut-off) will lead to different interactions being integrated out and thus has a strong impact on the theory, along with the fundamental degrees of freedom and the couplings used to describe it. The process of arriving at the new effective theory is called _renormalisation_. To describe it, we study universal quantities that are invariant under the renormalisation, first and foremost the partition function $\mathcal{Z}(\\{\lambda_{i}\\})$, which encodes the statistical properties of the quantum system and depends on the set of coupling constants mentioned before. For the purpose of this review, we can think of $\mathcal{Z}$ as a mathematical function that encodes all the physical properties of the system and its symmetries, independently on its explicit definition. One of the symmetries is, as already mentioned, the invariance under renormalisation, _i.e._ the change in the energy scale of the physical interactions. If $\\{\lambda^{\prime}_{a}\\}$ (with $a$ taking values in a new set $\\{\mathcal{S}^{\prime}\\}$) is the new set of renormalised couplings and $\mathcal{Z}^{\prime}$ the partition function of the renormalised theory, invariance of the partition function implies $\displaystyle\mathcal{Z}(\\{\lambda_{i}\\})=\mathcal{Z}^{\prime}(\\{\lambda^{\prime}_{a}\\})\,.$ (3.39) Hence, by continuously changing the energy scale, the theory sweeps out a trajectory in the space of all possible effective theories, called the _renormalisation group flow_ , which is governed by the invariance Eq.(3.39). From the perspective of the interactions, the theory sweeps out a trajectory in the space of all couplings $\lambda_{i}$. This is governed by the _beta- functions_ $\beta_{i}(\lambda_{k})$, defined as the derivatives of the couplings $\lambda_{i}$ with respect to the logarithm of the cut-off parameter, and are functions of the couplings $\lambda_{i}$ themselves. The flow is thus described in terms of a system of differential equations, like the SIR model does, whose fixed points (_i.e._ zeros of the beta functions) denote critical (_i.e._ scale invariant) points of the theory. Before making the connection to epidemiology, we remark that physical theories in general allow for field redefinitions, which means that they can be equivalently formulated using different bare fields. This implies that the coupling set $\\{\lambda_{i}\\}$ is not uniquely determined, but should rather be thought of as a (local) choice of basis in the space of couplings. A specific choice of a set of $\\{\lambda_{i}\\}$ is called a renormalisation _scheme_. While a priori the specific form of the beta-functions depend on the scheme (in particular their perturbative expansions as functions of the $\\{\lambda_{i}\\}$), a change of scheme can be understood as an analytic transformation in the space of couplings. In [92], and subsequent works [93, 94, 163], it was suggested to interpret the time evolution of the spread of a disease (specifically COVID-19) within the framework of the Wilsonian renormalisation group equation. We shall explain this description in more detail in Section 4. In the following, however, we shall show how such a description can at least qualitatively be obtained from the SIR equations by allowing time-dependent infection and removal rates, as first pointed out in [93]. #### 3.6.1 Beta Function In preparation to Section 4, we notice that the SIR model (with $\zeta=0$, but time-dependent infection and recovery rates $\gamma(t)$ and $\epsilon(t)$) can be written in a form which is strongly reminiscent of a RGE. To this end, we return to Eqs (3.23) and repeat the same steps as in Section 3.4, except for allowing $\sigma:\,[0,1]\to\mathbb{R}_{+}$ to be a priori a function of $S$. Thus, we can integrate Eq.(3.24) in the following form $\displaystyle I(S)=1-S+\int_{S_{0}}^{S}\frac{du}{u\,\sigma(u)}\,,$ (3.40) which is compatible with the initial conditions in Eq.(3.23) at $t=0$. Inserting this relation into the first equation of (3.4), for $\zeta=0$ it yields $\displaystyle\frac{dS}{dt}=-\gamma(t)\,S(t)\,\left[1-S+\int_{S_{0}}^{S}\frac{du}{u\,\sigma(u)}\right]\,.$ (3.41) Instead of the relative number of susceptible, this equation can be re-written in terms of the cumulative number of infected individuals $I_{\text{c}}$, as defined in Eq. (3.6). Thus, Eq.(3.41) can be rewritten as $\displaystyle\frac{dI_{\text{c}}}{dt}=N\,\gamma\,\left(1-\frac{I_{\text{c}}}{N}\right)\left[\frac{I_{\text{c}}}{N}+\int_{S_{0}}^{1-\frac{I_{\text{c}}}{N}}\frac{du}{u\,\sigma(u)}\right]\,.$ (3.42) Next, generalising what was proposed in [92, 163], we define an _epidemic coupling_ $\alpha(t)$ as a function of the cumulative number of infected individuals: $\displaystyle\alpha(t)=\phi(I_{\text{c}}(t))\,,$ (3.43) where $\phi:\,[0,N]\rightarrow\mathbb{R}$ is a strictly monotonically growing, continuously differentiable function with non-vanishing first derivative. A priori, $\phi$ could also explicitly depend on $t$ (not only through $I_{\text{c}}(t)$), but in the following we shall not explore this possibility. In [92], in the context of the COVID-19 pandemic, $\phi$ was chosen to be the natural logarithm, while in [163, 164] $\phi(x)=x$ was chosen. For the moment, we shall leave $\phi$ arbitrary, which mimics the liberty to choose different renormalisation schemes in the framework of the Wilsonian approach. Upon defining formally the $\beta$-function as $\displaystyle\beta(I_{\text{c}}(t))=-\frac{d\alpha}{dt}\,,$ (3.44) Eq. (3.42) can be re-formulated as $\displaystyle-\beta=\left(\frac{d\phi}{dI_{\text{c}}}\right)\frac{dI_{\text{c}}}{dt}=\left(\frac{d\phi}{dI_{\text{c}}}\right)\,N\,\gamma\,\left(1-\frac{I_{\text{c}}}{N}\right)\left[\frac{I_{\text{c}}}{N}+\int_{S_{0}}^{1-\frac{I_{\text{c}}}{N}}\frac{du}{u\,\sigma(u)}\right]\,.$ (3.45) An explicit example that is designed to make contact with the work in [163] is discussed in Section 3.6.2. Eq.(3.45), at least structurally, resembles a RGE and has several intriguing properties to support this interpretation. Note that with Eq.(3.6), we can also write $\displaystyle\beta(t)=-\left(\frac{d\phi}{dI_{\text{c}}}\right)\,\frac{dI_{\text{c}}}{dt}=-\left(\frac{d\phi}{dI_{\text{c}}}\right)\,N\,\gamma(t)\,I(t)\,S(t)\,,$ (3.46) which vanishes when: * 1. the infection rate vanishes $\gamma(t)=0$, * 2. or there are no susceptible individuals left $S(t)=0$, * 3. or the number of active infected vanishes $I(t)=0$ and the disease is eradicated. Further (structural) evidence can be given by considering concrete solutions: an example for the interplay between the beta-function and $\sigma$ is provided in the following Section 3.6.2. Furthermore, independently of its connection to compartmental models, a renormalisation group approach can be used to model and describe the dynamics of an epidemic, as we discuss in Section 4. #### 3.6.2 Connection between SIR models and the eRG approach We now discuss via a concrete example how to formulate a SIR model (with time- dependent $\sigma(t)$) in a way that reproduces the eRG framework, which will be discussed in more detail in the next section. This relation has been first discussed in [93, 94]. Following the logic outlined above, we will highlight the similarities between the SIR equations and RGEs. In particular, we show how a particular beta-function can be obtained from a time-dependent $\sigma$, starting from Eq. (3.45). Concretely, we shall make contact with the following: $\displaystyle-\beta_{0}(I_{\text{c}})={\lambda}\,I_{\text{c}}\left[\left(1-\frac{I_{\text{c}}}{{A}}\right)^{2}-\delta\right]^{p}\,,$ (3.47) where $\phi(I_{\text{c}})=I_{\text{c}}$, and $p,\delta,A$ are constant. The form of the beta-function (3.47) will be motivated and discussed in more detail in Section 4.2 and is used to study a single wave followed by an endemic period characterised by a quasi linear growth, which can be precursor to a next wave. We shall return on this linear period in Section 3.7. As a starting point, we shall consider a SIR model where, for simplicity, $\epsilon$ is constant, _i.e._ the rate of recovery remains constant throughout the pandemic888$\epsilon$ depends on biological properties of the virus as well medical and pharmaceutical means of the population to cure it. Since these are difficult to change without significant effort, the value of $\epsilon$ is difficult to change., while $\gamma$ and $\sigma=\frac{\gamma}{\epsilon}$ are continuous functions of $S$. Finally, to make contact with Eq.(3.47), we shall consider the asymptotic limit $S_{0}\rightarrow 1$. Identifying the function $\beta(t)$ in Eq.(3.46) with $\beta_{0}$ leads to an integral equation that, for constant $\epsilon$, can be turned into a differential equation for $\sigma(t)$ (recall that $S=1-\frac{I_{\text{c}}}{N}$): $\displaystyle\frac{d}{dI_{\text{c}}}\left[\frac{\beta_{0}(I_{\text{c}})}{\epsilon\,\sigma\left(1-\frac{I_{\text{c}}}{N}\right)}\right]=1-\frac{1}{\left(1-\frac{I_{\text{c}}}{N}\right)\,\sigma\left(1-\frac{I_{\text{c}}}{N}\right)}\,.$ (3.48) The equation above can be brought into the form $\displaystyle 0=\sigma^{\prime}(S)+g_{1}(S)\,\sigma(S)+g_{2}(S)\,\sigma^{2}(S)\,,$ with $\displaystyle\begin{array}[]{l}g_{1}(S)=\frac{1}{S}-\frac{N}{\beta_{0}(N(1-S))}\,\left(\epsilon-\beta^{\prime}_{0}(N(1-S))\right)\,,\\\\[4.0pt] g_{2}(S)=\frac{N\epsilon S}{\beta_{0}(N(1-S))}\,.\end{array}$ (3.51) In the above and following equations, the prime indicates a derivative with respect to the argument of the function. The general solution of this first order, non-linear differential equation is $\displaystyle\sigma(S)=\frac{D(S)}{\frac{1}{\sigma_{0}}+\int_{S_{0}}^{S}dx\,D(x)\,g_{2}(x)}\,,$ with $\displaystyle D(S)=\text{exp}\left[-\int_{S_{0}}^{S}g_{1}(x)\,dx\right]\,.$ (3.52) Here $\sigma_{0}$ is an integration constant, which can be determined by comparing the first derivative of $\beta_{0}$ and $\beta$ at $S=S_{0}\rightarrow 1$ (_i.e._ at $I_{\text{c}}=N(1-S_{0})=0$). In fact, $\beta^{\prime}_{0}(0)=\beta^{\prime}(0)$ implies $\displaystyle\sigma(1)=\sigma_{0}=1-\frac{1}{\epsilon}\,\beta_{0}^{\prime}(0)=1+\frac{\lambda}{\epsilon}(1-\delta)^{p}\,.$ (3.53) With $\beta_{0}$ given in Eq.(3.47), the integral over $g_{1}$ can be performed analytically (involving an Appell hypergeometric function). However, using this result to insert $D(S)$ into the first expression in Eq.(3.52), the integral in the denominator is more involved and we could only find analytic solutions for generic999We remark in passing that we were able compute analytic solutions for other combinations of $(p,\delta)$ for specific combinations of $(\lambda,\epsilon)$, _i.e._ for certain fixed ratios $\frac{\lambda}{\epsilon}$. $\lambda,\epsilon$ for $(p=\tfrac{1}{4},\delta=0)$ and $(p=\tfrac{1}{2},\delta=0)$, whose limit $S_{0}\to 1$ is $\displaystyle\lim_{S_{0}\to 1}\sigma(1-\tfrac{I_{\text{c}}}{N})\bigg{|}_{{p=\frac{1}{4}}\atop{\delta=0}}$ $\displaystyle=\frac{\frac{\lambda N}{\epsilon(N-I_{\text{c}})}\sqrt{1-\frac{I_{\text{c}}}{A}}}{1+\frac{2^{1-\frac{\epsilon}{\lambda}}A\epsilon}{I_{\text{c}}(\lambda+\epsilon)}\left(\sqrt{1-\frac{I_{\text{c}}}{A}}-1\right)\left(\sqrt{1-\frac{I_{\text{c}}}{A}}+1\right)^{\frac{\epsilon}{\lambda}}\,_{2}F_{1}\left(\frac{\epsilon}{\lambda},\frac{\lambda+\epsilon}{\lambda};\frac{\epsilon}{\lambda}+2;\frac{1-\sqrt{1-\frac{I_{\text{c}}}{A}}}{2}\right)}\,,$ $\displaystyle\lim_{S_{0}\to 1}\sigma(1-\tfrac{I_{\text{c}}}{N})\bigg{|}_{{p=\frac{1}{2}}\atop{\delta=0}}$ $\displaystyle=\frac{N(A-I_{\text{c}})(\lambda+\epsilon)\left(1-\frac{I_{\text{c}}}{A}\right)^{-\frac{\epsilon}{\lambda}}}{A\epsilon(N-I_{\text{c}})\,_{2}F_{1}\left(\frac{\epsilon}{\lambda},\frac{\lambda+\epsilon}{\lambda};2+\frac{\epsilon}{\lambda};\frac{I_{\text{c}}}{A}\right)}\,.$ (3.54) However, the integration can be performed numerically, and for different values of $(p,\delta)$, $\sigma$ as a function of $I_{\text{c}}$ is shown in Fig. 25. Figure 25: $\sigma$ as a function of $I_{\text{c}}$ for different values of $p$ and $\delta=0$ in the limit $S_{0}\to 1$ with $N=1.000.000$, $A=50.000$, $\lambda=0.5$ and $\epsilon\in\\{0.1\,,0.3\,,0.5\,,0.7\,,0.9\,,1.1\,,1.3\\}$. We note that for $p\leq 1/2$, $\text{Im}(\sigma)\neq 0$ for $I_{\text{c}}>A$, thus indicating that the solution does not extend beyond the maximal number of cumulative infected $I_{\text{c}}=A$ (see Fig. 26). Similar plots for $\delta\neq 0$ are shown in Fig. 27. Figure 26: Numerical computation of the imaginary part of $\sigma(1-\tfrac{I_{\text{c}}}{N})$ for $\lambda=0.5$, $\epsilon=0.7$, $N=1.000.000$ and $A=50.000$ in the limit $S_{0}\to 1$ for various values of $p$ and $\delta=0$. Finally, we also remark that the numerical integration allows us to include $\delta<0$ and can even be generalised to more general classes of $\beta$-functions proposed in [163] $\displaystyle-\beta_{0}(I_{\text{c}})={\lambda}\,I_{\text{c}}\left[\left(1-\frac{I_{\text{c}}}{{A}}\right)^{2}-\delta\right]^{p}(1-\zeta I_{\text{c}})\,,$ (3.55) as shown in Fig. 27. In the case $\zeta>0$ we remark that $\text{Im}(\sigma)\neq 0$ for $I_{\text{c}}>\zeta^{-1}$, indicating as above the breakdown of the assumptions. Figure 27: $\sigma$ as a function of $I_{\text{c}}$ for different values of $(p,\delta,\zeta)$ in the limit $S_{0}\to 1$ with $N=1.000.000$, $A=50.000$, $\lambda=0.5$ and $\epsilon\in\\{0.1\,,0.3\,,0.5\,,0.7\,,0.9\,,1.1\,,1.3\\}$. ### 3.7 Analytic Solution during a Linear Growth Phase Many epidemics generated by an infectious disease feature a multi-wave pattern, with periods in between waves where an approximately linear growth of the number of infected is observed. As an example, COVID-19 data show this period very clearly in most of the countries, thanks to the large amount of data collected (see Section 5). This phase of the epidemic, which links two consecutive waves, has found a natural explanation in the eRG framework [163, 164], which we will review in Section 4.2. Here we attempt to describe this linear phase from the perspective of compartmental models. In fact, we have seen from the explicit solutions in Section 2 that such a behaviour is not found in simple percolation models in which, notably, the probability or rate of infection remains constant throughout the entire pandemic. Similarly, this type of solutions is absent in compartmental models. However, more general approaches and extensions of these simple models might exhibit such linear growth phases. Since the phenomenon is seen in the cumulative number of infected (which is a ‘global’ key figure pertaining to the entire population), we shall in the following analyse it from the perspective of a SIR model, with time-dependent infection and recovery rates. #### 3.7.1 Simplified SIR Model with Constant New Infections We consider a SIR model described by the equations (3.4) and the initial conditions (3.3) with time-dependent $\gamma$, $\epsilon$ and $\zeta$ (see Section 3.5.1). We define a linear growth regime as a period in time $[t_{1},t_{2}]$ during which the cumulative number of infected $I_{\text{c}}$, defined in Eq.(3.6) as $\displaystyle I_{\text{c}}(t)=N\,I_{0}+\int_{0}^{t}dt^{\prime}\,\gamma(t^{\prime})\,N\,I(t^{\prime})\,S(t^{\prime})\,,$ (3.56) is a linear function of time. In other words, $\displaystyle\frac{d}{dt}\,I_{\text{c}}(t)=N\,f=\text{const.}$ $\displaystyle\forall t\in[t_{1},t_{2}]\,,$ (3.57) while $0\leq S(t),I(t),R(t)\leq 1$, with $f\in\mathbb{R}_{+}$. This implies $\displaystyle\gamma(t)\,I(t)\,S(t)=f$ $\displaystyle\forall t\in[t_{1},t_{2}]\,.$ (3.58) The condition above allows to analytically solve the SIR equations (3.4) $\forall t\in[t_{1},t_{2}]$ with the initial conditions at the beginning of the linear growth $\displaystyle S(t=t_{1})=S_{s}\,,$ $\displaystyle I(t=t_{1})=I_{s}\,,$ $\displaystyle R(t=t_{1})=R_{s}\,,$ with $\displaystyle\begin{array}[]{l}0\leq S_{s},I_{s},R_{s}\leq 1\,,\\\ S_{s}+I_{s}+R_{s}=1\,.\end{array}$ (3.61) To see this, we define $\displaystyle D_{\epsilon}(t)=e^{\int_{t_{1}}^{t}\epsilon(t^{\prime})dt^{\prime}}\,,$ and $\displaystyle D_{\zeta}(t)=e^{\int_{t_{1}}^{t}\zeta(t^{\prime})dt^{\prime}}\,,$ (3.62) which have the properties $\displaystyle\frac{dD_{\epsilon}}{dt}(t)=\epsilon(t)\,D_{\epsilon}(t)\,,$ $\displaystyle\frac{dD_{\zeta}}{dt}(t)=\zeta(t)\,D_{\zeta}(t)\,,$ $\displaystyle D_{\epsilon}(t=t_{1})=1=D_{\zeta}(t=t_{1})\,.$ (3.63) Next, we insert the constraint in Eq.(3.58) into Eqs (3.4) to obtain $\displaystyle\frac{dI}{dt}=-\epsilon\,I+f\,,$ $\displaystyle\forall t\in[t_{1},t_{2}]\,.$ (3.64) This differential equation only contains $I$ (hence, it is decoupled from $S$ and $R$). Multiplying by $D_{\epsilon}(t)$, we find $\left[\frac{dI}{dt}+\epsilon\,I\right]\,D_{\epsilon}(t)=f\,D_{\epsilon}(t)\qquad\qquad\Rightarrow\qquad\qquad\frac{d}{dt}\left[I(t)\,D_{\epsilon}(t)\right]=f\,D_{\epsilon}(t)\,,$ (3.65) which can be directly integrated, with the initial conditions (3.61), as: $\displaystyle I(t)=\frac{1}{D_{\epsilon}(t)}\left[f\,\int_{t_{1}}^{t}\,D_{\epsilon}(t^{\prime})\,dt^{\prime}+I_{s}\right]\,,$ $\displaystyle\forall t\in[t_{1},t_{2}]\,.$ (3.66) For the relative number of recovered, $R$, we can integrate the last equation of (3.4) $\displaystyle\frac{dR}{dt}(t)+\zeta(t)\,R=\epsilon(t)\,I(t)\,,$ (3.67) where, inserting the solution for $I(t)$ in Eq.(3.66), the right hand side can be understood as an inhomogeneity. Multiplying by $D_{\zeta}$ we obtain, as before, $\displaystyle\frac{d}{dt}\left[R(t)\,D_{\zeta}(t)\right]=\epsilon(t)\,I(t)\,D_{\zeta}(t)\,,$ (3.68) which can be directly integrated, with the initial conditions (3.61), to give $\displaystyle R(t)=\frac{R_{s}}{D_{\zeta}(t)}+I_{s}\,\int_{t_{1}}^{t}dt^{\prime}\,\frac{\epsilon(t^{\prime})}{D_{\epsilon}(t^{\prime})}\,\frac{D_{\zeta}(t^{\prime})}{D_{\zeta}(t)}+f\,\int_{t_{1}}^{t}dt^{\prime}\int_{t_{1}}^{t^{\prime}}dt^{\prime\prime}\,\epsilon(t^{\prime})\,\frac{D_{\epsilon}(t^{\prime\prime})}{D_{\epsilon}(t^{\prime})}\,\frac{D_{\zeta}(t^{\prime})}{D_{\zeta}(t)}\,,$ $\displaystyle\forall t\in[t_{1},t_{2}]\,.$ (3.69) Finally, $S(t)$ is obtained through the constraint (3.1): $S(t)=1-I(t)-R(t)$. Notice that the solutions (3.66) and (3.69) remain valid as long as $0\leq S(t),I(t),R(t)\leq 1$. #### 3.7.2 Vanishing $\zeta$ and Constant $\epsilon$ To simplify the solutions found above, we can adapt the functions $\zeta$ and $\epsilon$ to reflect more closely the COVID-19 pandemic: since currently only very few cases of patients contracting COVID-19 twice within a short time, _i.e._ a single epidemic wave, are known in the medical literature [165] we can set $\zeta(t)=0$ to simplify the solutions (3.66) and (3.69). Since $\zeta=0$ also implies $D_{\zeta}(t)=1$, we find for these solutions $\displaystyle S(t)$ $\displaystyle=S_{s}-f(t-t_{1})\,,$ $\displaystyle I(t)$ $\displaystyle=\frac{I_{s}}{D_{\epsilon}(t)}+f\,\int_{t_{1}}^{t}\,\frac{D_{\epsilon}(t^{\prime})}{D_{\epsilon}(t)}\,dt^{\prime}\,,$ $\displaystyle R(t)$ $\displaystyle=R_{s}+I_{s}\,\int_{t_{1}}^{t}dt^{\prime}\,\frac{\epsilon(t^{\prime})}{D_{\epsilon}(t^{\prime})}+f\,\int_{t_{1}}^{t}dt^{\prime}\int_{t_{1}}^{t^{\prime}}dt^{\prime\prime}\,\epsilon(t^{\prime})\,\frac{D_{\epsilon}(t^{\prime\prime})}{D_{\epsilon}(t^{\prime})}\,,\hskip 51.21504pt\forall t\in[t_{1},t_{2}]\,.$ (3.70) We have explicitly verified that this is indeed a solution of Eqs (3.4) that satisfies the correct initial conditions. Furthermore, since the recovery rate in most cases depends on the disease in question and/or the state of medical advancement to cure it, $\epsilon$ is difficult to change throughout a pandemic without significant effort. For simplicity, we therefore consider it in the following to be constant, _i.e._ $\epsilon=$ const. (in addition to $\zeta=0$), such that $D_{\epsilon}(t)=e^{\epsilon(t-t_{1})}$. In this case, we can perform the integrations in Eq.(3.70) to obtain $\displaystyle I(t)$ $\displaystyle=e^{-\epsilon(t-t_{1})}\,\left[f\int_{t_{1}}^{t}dt^{\prime}\,e^{\epsilon(t^{\prime}-t_{1})}+I_{s}\right]=e^{-\epsilon(t-t_{1})}\,I_{s}+\frac{f}{\epsilon}\left(1-e^{-\epsilon(t-t_{1})}\right)\,,$ $\displaystyle\forall t\in[t_{1},t_{2}]\,,$ (3.71) as well as the relative number of removed $\displaystyle R(t)$ $\displaystyle=R_{s}+I_{s}\,\epsilon\,\int_{t_{1}}^{t}\,dt^{\prime}\,e^{-\epsilon(t^{\prime}-t_{1})}+\epsilon f\int_{t_{1}}^{t}dt^{\prime}\,e^{-\epsilon t^{\prime}}\int_{t_{1}}^{t^{\prime}}dt^{\prime\prime}\,e^{\epsilon t^{\prime\prime}}$ $\displaystyle=R_{s}+f(t-t_{1})+\left(I_{s}-\frac{f}{\epsilon}\right)\left(1-e^{-\epsilon(t-t_{1})}\right)\,,$ $\displaystyle\forall t\in[t_{1},t_{2}]\,.$ (3.72) One can directly verify that these expressions satisfy Eqs (3.4) along with $\displaystyle S(t)+I(t)+R(t)=S_{s}+I_{s}+R_{s}\,,$ $\displaystyle\forall t\in[t_{1},t_{2}]\,.$ (3.73) For some (random) values of $\epsilon$, $f$, $S_{s}$, $I_{s}$ and $R_{s}$, the functions $S(t)$, $I(t)$ and $R(t)$ (for the region where $0\leq S(t),I(t),R(t)\leq 1$) are plotted in the left panel of Fig. 28, while the associated $\gamma(t)=\frac{f}{S(t)\,I(t)}$ is plotted in the right panel. Figure 28: Solutions (3.70) and $\gamma(t)$ for $\epsilon=0.05$, $f=0.002$, $S_{s}=0.9$, $I_{s}=0.1$, $R_{s}=0$ and $t_{1}=0$ as a function of time $t$. #### 3.7.3 Constant Active Number of Infectious Individuals During the linear growth periods, the COVID-19 data also shows that the number of active infectious individuals remains constant. Intriguingly, this feature is also observed in the solutions in the left panel of Fig. 28. In this section, we will seek a solution of the SIR model with this property, _i.e._ $\displaystyle I(t)=\mathsf{f}=\text{const.}$ $\displaystyle\forall t\in[t_{1},t_{2}]\,,$ (3.74) for some $\mathsf{f}\in[0,1]$, which in particular implies $\displaystyle\frac{d}{dt}\,I(t)=0\,,$ $\displaystyle\forall t\in[t_{1},t_{2}]\,.$ (3.75) Injecting this into Eqs (3.4), we obtain under the assumption $I(t)\neq 0$ $\forall t\in[t_{1},t_{2}]$ $\displaystyle S=\frac{\epsilon}{\gamma}\,,$ $\displaystyle\forall t\in[t_{1},t_{2}]\,,$ (3.76) and thus for $\zeta\neq 0$ $\displaystyle\frac{d}{dt}\left(\frac{\epsilon}{\gamma}\right)=-\epsilon\,\mathsf{f}\,+\zeta\,R\,,$ $\displaystyle\Longrightarrow$ $\displaystyle R=\frac{1}{\zeta}\left[\frac{d}{dt}\left(\frac{\epsilon}{\gamma}\right)+\epsilon\mathsf{f}\right]\,,$ $\displaystyle\forall t\in[t_{1},t_{2}]\,.$ (3.77) For $\zeta=0$ we obtain the following constraint for the infection and recovery rate $\displaystyle\frac{d}{dt}\left(\frac{\epsilon}{\gamma}\right)=-\epsilon\,\mathsf{f}\,,$ $\displaystyle\forall t\in[t_{1},t_{2}]\,.$ (3.78) For the classical SIR model (for which $\epsilon$ and $\gamma$ are time- independent $\forall t$ and $\zeta=0$), assuming that $\gamma\neq 0$, the constraint (3.78) implies that either * 1. $\mathsf{f}=0$, which however is excluded since $I\neq 0$; * 2. or $\epsilon=0$, in which case $\frac{dR}{dt}=0$ $\forall t$ (_i.e._ not just $t\in[t_{1},t_{2}]$). However, with the initial conditions (3.3) this implies $R(t)=0$ and thus $\displaystyle\frac{d}{dt}S(t)=-\gamma\,\mathsf{f}\,S$ $\displaystyle\Longrightarrow$ $\displaystyle S=c\,e^{-\gamma\mathsf{f}\,t}\,,$ $\displaystyle\forall t\in[t_{1},t_{2}]\,,$ (3.79) for $c\in[0,1]$. On the other hand the relation (3.1) implies that $\frac{dS}{dt}=0$ and thus (with $\gamma\neq 0$ and $\mathsf{f}\neq 0$) $S=0$ (consistent with Eq.(3.76)), in which case $I=\mathsf{f}=1$ and the entire population is infected (and stays infected for all times). Thus, within the classical SIR model, the only solution with $I(t)=\mathsf{f}\neq 0$ constant is $\epsilon=0$ (_i.e._ instead of the SIR model we only consider the SI model) and $I=1$. This corresponds to the late phase of the SI model, where the entire population is infected. We, therefore, see that the traditional SIR model cannot account for the linear growth of the cumulative number of infected related to Eq.(3.75) and observed in the COVID-19 data. ## 4 Epidemic Renormalisation Group Executive Summary 1. We introduce the epidemic renormalisation group approach that efficiently captures asymptotic time symmetries of the diffusion process. 2. The framework is based on flow equations characterised by fixed point dynamics. 3. We show the power of the approach by studying single wave epidemics, which can be naturally generalised to describe multi wave patterns via complex fixed points 4. We demonstrate how the spreading of diseases across different regions of the world can be efficiently described and predicted As anticipated in the previous section, it has been proposed in [92, 94] to study the spread of a communicable infectious disease within the framework of the Wilsonian renormalisation group [86, 87]. We have already pointed out in Section 3 that the SIR model, defined by the differential equations (3.4), can be formulated in a fashion that is structurally similar to a set of RGEs (see [162]). In this section we review the new framework, first proposed in [92, 94], dubbed _epidemic Renormalisation Group (eRG)_. The eRG framework consists, effectively, in a single differential equation that captures the time evolution of the disease diffusion, after the microscopic degrees of freedom and interactions have been ‘integrated out’ and all the detailed effects (virulence of the disease, social measures, _etc._) are taken into account by the few parameters in the equation. This leads to a much more economical description in terms of calculation complexity as compared to microscopic or compartmental models. At this stage, the main relation with the renormalisation group is the fact that symmetries can be explicitly included in the formalism. In the case of the eRG, the symmetries are related to time scale invariance, _i.e._ the presence of phases where the disease diffusion is (nearly) stable in time. In the Wilsonian renormalisation group, which describes the energy dependence of physical charges (for instance, the interaction strength among fundamental particles), the symmetries involved are related to scale invariance of distances and energies. A RGE, therefore, describes the energy flow of a charge from the Ultra-Violet (UV) regime at high energies to the Infra-Red (IR) regime at low energies. The eRG also describes a flow, however in time instead of in energy, as we will see shortly. The physical charge is replaced by an _epidemiological charge_ , which is defined as a monotonic, differentiable function of the cumulative number of individuals infected by the disease as a function of time. This discussion has already been anticipated in Section 3.6.1. The economy of this approach in terms of free parameters and computing time needed to solve the flow of the disease makes it an ideal tool to study the diffusion of an infectious disease at different scales, from small regions to a global level. The eRG framework has been first used to characterise a single epidemic wave, _i.e._ a single episode of exponential increase in the number of infections followed by an attenuation [92], and extended to study the inter-region propagation of the disease [94], with validation on the COVID-19 data in Europe [95] and in the US states [96]. Mobility data have also been used to study the effect of non-pharmaceutical interventions [149] as well as the role played by passenger flights in the US [96]. As we also review in this section, the framework can be extended to include the multi-wave pattern [163, 164] that emerges in many communicable diseases, like the 1918 influenza pandemic, the seasonal flu and the COVID-19 pandemic of 2019. Finally, preliminary work on the inclusion of vaccinations [96] and virus mutations [166, 167] are present in the literature, however we will not cover them in this review. ### 4.1 Beta Function and Asymptotic Fixed Points The main motivation behind the eRG approach to epidemiology stems from the observation that a single epidemic wave starts with a very small number of infected individuals and ends when the cumulative number of infections reaches a constant, hence no new infections are detected. This dynamics is characteristic of a system that flows from a fixed point at $t=-\infty$, when no infections are present, to a new fixed point at $t=\infty$, when the number of cumulative infections reaches a constant value again. The dynamical flow between the two fixed points can be described by the following differential equation: $\displaystyle-\beta(\alpha)=\frac{d\alpha}{dt}(t)=\lambda\,\alpha\left(1-\frac{\alpha}{A}\right)\,,$ (4.1) where $\alpha$ is a function of the number of infected, hence a function of time. The precise form of this equation is an ansatz, for now, and we will establish the precise relation between $\alpha$ and the number of infections later. The main feature to stress is the presence of two zeros, corresponding to the fixed points of the system: if $\alpha(t_{0})=0$ or $\alpha(t_{0})=A$ at any time $t_{0}$, the system will remain in this state at all times. The zeros can be characterised through the so-called _scaling exponents_ : $\displaystyle\vartheta=\frac{\partial\beta}{\partial\alpha}\bigg{|}_{\alpha^{\ast}}=\left\\{\begin{array}[]{lcl}-\lambda&\text{for}&\alpha_{1}^{*}=0\,,\\\\[4.0pt] \lambda&\text{for}&\alpha_{2}^{*}=A\,,\end{array}\right.$ (4.4) where $\alpha^{\ast}$ is the epidemic coupling constant at the fixed point. A negative (positive) scaling exponent corresponds to a repulsive (attractive) fixed point. Thus, a system in the repulsive fixed point at $\alpha^{\ast}=0$, once perturbed (by a small initial number of infected individuals) will flow towards the attractive fixed point at $\alpha^{\ast}=A$. As such, $A$ is a function of the cumulative number of individuals infected during the epidemic wave. Figure 29: The logistic function schematically representing the cumulative number of infected as a function of time. With regards to (4.5) we have $A=20.000$, $B=1.000.000$ and $\kappa=0.2$. The solution of Eq.(4.1) is a _logistic function_ (sigmoid) of the form: $\displaystyle\alpha\,:\mathbb{R}$ $\displaystyle\longrightarrow[0,A]$ $\displaystyle t$ $\displaystyle\longmapsto\alpha(t)=\frac{A}{1+B\,e^{-\lambda t}}\,,$ (4.5) where $A,B,\lambda\in\mathbb{R}_{+}\setminus\\{0\\}$. This function shows a characteristic ‘S’-shape (see Fig. 29 for a schematic representation) and has the following asymptotic values $\displaystyle\lim_{t\to-\infty}\alpha(t)=0\,,$ $\displaystyle\lim_{t\to\infty}\alpha(t)=A\,,$ (4.6) corresponding to the zeros of the derivative $\frac{d\alpha}{dt}=0$. As already mentioned, the parameter $A$ corresponds to (a function of) the asymptotic number of infected cases during the epidemic wave. The second parameter in Eq.(4.1), $\lambda$, which has dimension of a rate, measures how fast the number of infections increases, while $B$ is an integration constant that corresponds to a shift of the entire curve in time and determines the beginning of the infection increase. More details about the properties of this function and its epidemiological interpretation can be found in [92] and will not be repeated here. It is, however, important to notice that the parameters $\lambda$ and $A$ can be removed from the differential equation by a simple rescaling of the function and of the time variable: $\displaystyle\frac{d\tilde{\alpha}}{d\tau}=\tilde{\alpha}(\tau)\,(1-\tilde{\alpha}(\tau))\,,\qquad\tau=\lambda t\,,\quad\tilde{\alpha}(\tau)=\frac{\alpha(\tau/\lambda)}{A}\,.$ (4.7) Thus, while $A$ is a mere normalisation, $\lambda$ can be thought of as a ‘time dilation’ parameter. Once the normalised solutions are shown in the ‘local time’ $\tau$, therefore, all epidemic waves should reveal the same universal temporal shape. This universality property has been first pointed out in [92] from data of the Hong Kong (HK) Sars-2003 outbreak as well as the COVID-19 pandemic during the spring of 2020. It has been shown that the time dependence of the cumulative total number of infected cases in various regions of the world shows the same characteristic behaviour. In [92], the epidemic coupling has been defined as the logarithm of the cumulative infected, $\alpha(t)=\ln I_{c}(t)$, however other choices, like $\alpha(t)=I_{c}(t)$, can also reproduce the data. The same framework can also be applied to the number of hospitalisations or the number of deceased individuals. This feature of the epidemiological data shows that the dynamics encoded in Eq.(4.1) provides an accurate description of the diffusion of an infectious diseases in terms of a flow equation. In [92, 94, 163] the following dictionary between the spread of an epidemic and the Wilsonian renormalisation group was proposed: * 1. The time variable is identified with the (negative) logarithm of the energy scale $\mu$ $\displaystyle\frac{t}{t_{0}}=-\ln\left(\frac{\mu}{\mu_{0}}\right)\,,$ (4.8) where $t_{0}$/$\mu_{0}$ set the scale for the time and energy (for simplicity, and without loss of generality, we will fix $t_{0}=1$). With this identification, Eq.(4.1) is similar to the RGE for the gauge coupling in a theory that features a Banks-Zaks type fixed point [168], _i.e._ an interactive fixed point at low energies (in the Infra-Red). * 2. The solution can be associated to a coupling constant in the high energy physics RGEs, $\alpha$. The epidemic coupling strength is defined as a monotonic, differentiable and bijective, function $\phi$ of the cumulative number of infected cases $\displaystyle\alpha(t)=\phi(I_{\text{c}}(t))\,.$ (4.9) In [92, 163] $\phi$ was chosen as the natural logarithm $\phi(x)=\ln(x)$, while in [163, 164] it was chosen $\phi(x)=x$. The choice was justified by a better fit to the actual data of the COVID-19 pandemic, while from the perspective of the Wilsonian renormalisation group, the difference corresponds to a different choice of scheme. * 3. The _beta function_ is defined as the (negative) time-derivative of the epidemic coupling strength $\displaystyle\beta\equiv\frac{d\alpha}{d\ln\left(\frac{\mu}{\mu_{0}}\right)}=-\frac{d\alpha}{dt}=-\frac{d\phi}{dI_{\text{c}}}\,\frac{dI_{\text{c}}}{dt}(t)\,.$ (4.10) In order to better model the respective data of various countries during the COVID-19 pandemic, it was furthermore proposed in [163, 164] to consider the more general beta-function $\displaystyle-\beta(\alpha)=\frac{d\alpha}{dt}(t)=\lambda\,\alpha\left(1-\frac{\alpha}{A}\right)^{2p}\,,$ (4.11) for $p\in[1/2,\infty]$ and $\lambda,A\in\mathbb{R}_{+}$. The role of the exponent $p$ is to smoothen the ‘S’-shape of the solution when it approaches the attractive fixed point at $\alpha^{\ast}=A$. #### 4.1.1 Generalisation to multiple regions The approach discussed so far assumes an isolated population of sufficient size. However, the simplicity of the eRG approach allows for a simple generalisation to study the interaction between various regions of the world [94] via the mobility of individuals. For $M$ separated populations (labelled by $i=1,\ldots,M$) of size $N_{i}$ whose cumulative number of infected is denoted by $I_{\text{c},i}$, it was proposed in [94] that infections can be transmitted between these populations by travellers. Hence, the epidemic diffusion can be described by $M$ coupled differential equations, in the form of Eq.(4.11) for each population, with the addition of an interaction term: $\displaystyle-\beta(\alpha_{i})=\lambda\,\alpha_{i}\left(1-\frac{\alpha_{i}}{A}\right)^{2p}+\frac{d\phi}{dI_{\text{c},i}}\,\sum_{j=1}^{M}\frac{k_{ij}}{N_{i}}\left(I_{\text{c},j}(t)-I_{\text{c},i}(t)\right)\,,$ (4.12) where $k_{ij}\in\mathbb{R}$ is a measure for the number of travellers between populations $i$ and $j$. The contribution to the beta function can be obtained by replacing $I_{\text{c},i}\to\phi^{-1}(\alpha_{i})$, where $\alpha_{i}$ is the epidemic coupling in each population. For more details, see Ref. [94]. Figure 30: Schematic representation of the flow in a two-region coupled eRG framework of Eq. (4.12). In this fictitious example we fix $\lambda_{1}=0.7$, $\lambda_{2}=0.9$, $N_{1}=200000$, $N_{2}=300000$, $A_{1}=\log\left(\frac{1}{40}N_{1}\right)$, $A_{2}=\log\left(\frac{N_{2}}{10}\right)$ and $p_{1}=p_{2}=\frac{1}{2}$. For the matrix of couplings $k_{ij}$, we use $k_{12}=k_{21}=10^{-3}$ and $k_{11}=k_{22}=0$. The two-component vectors are given by $(-\beta(\alpha_{1}),-\beta(\alpha_{2}))$, with $I_{c}\in\mathbb{C}$ with the overall length represented by the colour-coding. The function $\phi$ was chosen $\phi(x)=x$ $\forall x\in\mathbb{R}$. These coupled differential equations can be thought of flow equations, in the spirit of the Wilsonian renormalisation, with the second term representing a coupling between the different regions. A graphical representation of the coupled $\beta$-functions in Eq.(4.12) can be given in the form of flow in an $M$-dimensional space. In Fig. 30, we provide a numerical (fictitious) example for $M=2$: choosing the scheme $\alpha_{i}(I_{c,i})=\ln(I_{c,i})$ for $i=1,2$, the arrows indicate the vector field $\left(\begin{array}[]{c}-\beta(\alpha_{1})\\\ -\beta(\alpha_{2})\end{array}\right)$ with the colour representing the length $\sqrt{\beta(\alpha_{1})^{2}+\beta(\alpha_{2})^{2}}$ at any point in the $(I_{c,1},I_{c,2})$-plane. The black dots are the actual trajectory of the system calculated as the numerical solution of the coupled differential equations (4.12). As it can be seen, the former flows along the arrows from a repulsive fixed point at $(I_{c,1},I_{c,2})=(0,0)$ (all arrows point away from it), which represents the absence of the disease in both countries, to an attractive fixed point (all arrows point towards it) which corresponds to the eradication of the disease. The coupled eRG framework in Eq.(4.12) has been used to explain the diffusion of the COVID-19 pandemic across different regions of the world. This is one of the main mechanisms that can generate multiple waves across a geographic region, while a second one will be discussed in the next section. The method has been used to predict the arrival of a second COVID-19 wave, which has hit Europe in the fall of 2020 [95]: the new infections originate from a seed region, which can be interpreted as inflow from outside Europe or the effect of hotspots and clusters, while the number of travellers, i.e. the entries of $k_{ij}$, were generated randomly. In Ref. [96], the same framework was used to explain the geographical wave patterns observed in the United States, with the aid of open-source flight data to estimate the couplings. ### 4.2 Complex (fixed point) epidemic Renormalisation Group Although the beta-function in Eq.(4.1) is relatively simple and contains only two parameters, it describes the time evolution of short-time epidemics (such as HK SARS-2003 and each wave of COVID-19) quite efficiently, as the flow from a repulsive to an attractive fixed point (or from an UV to an IR fixed point in the language of high-energy physics). However, this beta-function is too simple to describe correctly longer lasting pandemics with a more intricate time-evolution, such as subsequent waves of COVID-19: the attractive fixed point at $t\to\infty$ corresponds to a complete eradication of the disease and Eq.(4.1) describes outbreaks that follow a single wave. We have already discussed the role of passenger mobility in generating further epidemiological waves. However, data from COVID-19 has unveiled a second potential mechanism that may be at the origin of multiple-waves: in fact, after the end of each wave, a period of linear growth has been observed in all regions of the world (except those where the virus has been locally eradicated thanks to aggressive isolation policies). This is characterised by a nearly-constant number of new infected cases, and it can be seen as an endemic phase of the pandemic, where the virus circulates within the local population, without an exponential increase. Figure 31: Right: solutions of the CeRG equation, normalised to $A=1$ and with time in units of $\lambda$, for $-\delta=0,10^{-4},10^{-3},10^{-2}$ and $\delta_{\rm max}$, for $p=0.55$. Left: Estimated duration of the linear growth phase, in units of $\lambda$, as a function of $-\delta$ for $p=0.5$, $0.6$, $0.7$, $0.8$, $0.9$ and $1$. The lines end for $\delta=-\delta_{\rm max}$. In [163] it was proposed that this linear phase is evidence for a near time- scale invariance symmetry in the dynamics governing the diffusion of the virus. In practice, the system does not reach the second fixed point of Eq.(4.1), instead it hits an instability that drives the system to a new exponential phase after a given amount of time. The time-evolution of pandemics can still be described within the framework of a RGE, however with a more complicated beta-function that features a richer structure of (complex) fixed points. The new framework was called the _Complex epidemic Renormalisation Group (CeRG)_. In the CeRG approach, the beta function of Eq.(4.11) is modified as follows: $\displaystyle-\beta(I_{\text{c}})=\frac{dI_{\text{c}}}{dt}=\lambda\,I_{\text{c}}\left[\left(1-\frac{I_{\text{c}}}{{A}}\right)^{2}-\delta\right]^{p}=\lambda\,I_{\text{c}}\left(\frac{I_{\text{c}}}{{A}}-1+\sqrt{\delta}\right)^{p}\left(\frac{I_{\text{c}}}{{A}}-1-\sqrt{\delta}\right)^{p}\,,$ (4.13) where the additional parameter $\delta\in\mathbb{R}_{-}$, _i.e._ $\delta=-|\delta|$. While this equation can be written for any epidemic coupling $\alpha$, here we commit to the case $\alpha(t)=I_{\text{c}}(t)$ for reasons that will be clear in the next Section. The eRG equation (4.11) can be recovered for $\delta\to 0$. For non-vanishing $\delta$, instead of only two asymptotic fixed points, this functions has three fixed points $\displaystyle I_{\text{c},0}=0\,,$ $\displaystyle I_{\text{c},\pm}=A\left(1\pm i\sqrt{|\delta|}\right)\,,$ (4.14) with complex $I_{\text{c},\pm}\in\mathbb{C}$. Besides the repulsive fixed point at $I_{\text{c}}^{\ast}=0$, which remains, the attractive fixed point splits into two complex fixed points. Since the (cumulative) number of infected individuals is a strictly real number, the system cannot actually reach the complex fixed points and thus cannot exactly enter into a time-scale invariant regime at infinite time. Instead, for small $|\delta|$, when the solution approaches the would-be fixed point at $I_{\text{c}}\approx A$, the time evolution will be strongly slowed down due to the effect of the nearby complex fixed points. This results in a near-linear behaviour of the solution, as shown in the left panel of Fig. 31. Thus, the new beta function (4.13) realises an approximate time-scale symmetry in the solution. Concretely, the precise form of the flow in the vicinity of these complex fixed points depends on $|\delta|$: Figure 32: Schematic flow diagrams representing the $\beta$-function (4.13) with $A=1$, $\lambda=0.05$, $\delta=-0.003$ and $p=1/2$: the two-component vectors are given by $(\text{Re}(-\beta(I_{c})),\text{Im}(-\beta(I_{c})))$, with $I_{c}\in\mathbb{C}$ with the overall length represented by the colour- coding. Left panel: trajectories of the flow in the complex plane with initial conditions $I_{c}(t=0)=I_{c,0}$, with $\text{Im}(I_{c,0})$ specified in the figure. Right panel: close-up on the complex fixed point $I_{c,+}$. The dashed line represents a branch cut, which needs to be chosen such that it does not intersect the real axis. * 1. For $|\delta|<\delta_{\text{max}}=\frac{p^{2}}{1+2p}$, the beta-function has a local maximum and $I_{\text{c}}$ enters into a regime of near linear growth characterised by $\displaystyle\frac{dI_{\text{c}}}{dt}(t)\sim\text{const.}$ (4.15) In the context of epidemics, the linear growth phase can be associated to an endemic phase of the disease, when the virus keeps diffusing within the population without an exponential growth in the number of new infected (this corresponds to a situation with reproduction number $R_{0}=1$, which keeps the number of infectious cases constant). A connection of this regime with compartmental models of the SIR type has been presented in Section 3.7. * 2. In the CeRG, the linear growth is only an intermediate phase, which preludes to a new exponential increase in the number of infections. The duration depends on $|\delta|$, and can can be estimated as [163] $\displaystyle\Delta t_{\rm endemic}=-2\int_{A}^{\infty}\frac{dI_{\text{c}}}{\beta(I_{\text{c}})}\ .$ (4.16) This time is plotted for different values of $p$ as a function of $\delta$ in the right panel of Fig. 31. * 3. For $|\delta|\geq\delta_{\text{max}}$ the beta-function no longer has a local maximum and $I_{\text{c}}$ keeps growing exponentially, without a linear growing phase. In Fig. 32 we represent the dynamics encoded in Eq.(4.13) as a flow in the complex space of $I_{c}$. We clearly see that the system starts from the unstable fixed point at $I_{c}=0$, and moves towards the approximate one at $\text{Re}I_{c}\approx A$, where the evolution slows down. This is represented by the closeness of the data-points, which are calculated at equal intervals of time. We also show flows in the complex plane, which are unrealistic as $I_{c}$ remains a real number when describing a pandemic. Anyhow, all the solutions feature a slowing down of the infection growth near the complex fixed points, which reproduced the endemic phase of linear growth. The endemic linear-growing phase, therefore, is the prelude of a new wave of the epidemic diffusion. The CeRG approach can describe this endemic phase and the beginning of the next wave, however the number of infections would continue to grow indefinitely. In the following section we will further extend the approach to take into account the multi-wave pattern. ### 4.3 Modelling multi-wave patterns Pandemics like the 1918 Spanish flu [1] and COVID-19 have shown the appearance of multiple consecutive waves of exponential increase in the number of infections. In the case of COVID-19, the data support the fact that an endemic linearly-growing phase is always present in between two consecutive waves [163]. The CeRG model can be extended to take into account this structure, in a way that reproduces nicely the current data [164]. The multi-wave beta function, for an epidemic with $w$ consecutive waves, can be written as: $\displaystyle-\beta_{\rm multi-waves}(I_{\text{c}})=\lambda I_{\text{c}}\;\prod_{\rho=1}^{w}\left[\left(1-\zeta_{\rho}\,\frac{I_{\text{c}}}{A}\right)^{2}-\delta_{\rho}\right]^{p_{\rho}}\,,$ (4.17) with $\zeta_{\rho}\leq 1$, $|\delta_{\rho}|\ll 1$ and $p_{\rho}>0$ for $\rho\in\\{1,\ldots,w\\}$. The normalisation $A$ can be fixed to match the first wave, so that $0<\zeta_{w}<\dots<\zeta_{2}<\zeta_{1}=1\,.$ (4.18) Besides the repulsive fixed point at $I_{\text{c}}^{\ast}=0$, the equation has a series of complex fixed points ruled by the parameters $\delta_{\rho}$. Without loss of generality, we can fix $\delta_{w}=0$ so that the disease is extinguished after the last wave, and the total number of infections during the whole epidemic is given by $\lim_{t\to\infty}I_{\text{c}}(t)=A/\zeta_{w}$. This description, however, only works for $\alpha(t)\propto I_{\text{c}}(t)$, for which the value of the various fixed points are well separated [164], but not for $\alpha(t)\propto\ln I_{\text{c}}(t)$. Figure 33: Left panel: Schematic flow diagram representing the $\beta$-function (4.17) and trajectories for different initial conditions with $A=1/2$, $\zeta_{1}=1$, $\zeta_{2}=1/2$, $\lambda=0.05$, $\delta_{1}=-0.003$, $p_{1}=1/2$ and $p_{2}=1$. The trajectories of the flow start from the initial conditions $I_{c}(t=0)=I_{c,0}$, with $\text{Im}(I_{c,0})$ specified in the figure. Right panel: Cumulative number of infected $I_{c}$ for the trajectory with $\text{Im}(I_{c,0})=0$, _i.e._ the light blue dots in the left panel. In Fig. 33 we show the flow in the complex plane for Eq.(4.17) with two waves ($\omega=2$). After leaving the unstable fixed point at $I_{c}=0$, the system slows down near the complex fixed points, hence generating the linear endemic phase like in the CeRG approach, before entering a second wave. The latter ends at the final attractive fixed point. In the right panel we show the time evolution of $I_{c}(t)$ for this fictitious example, clearly showing two exponential episodes. As for the CeRG, the time delay between the two waves is controlled by the number of new cases in the endemic phase, i.e. by the parameter $\delta$ in Eq.(4.17). Hence, this model highlights the importance of imposing some measures to limit the circulation of the virus after the end of an epidemic wave in order to tame and control the emergence of the next one. Note, finally, that this formalism can also be used for studying the diffusion in between different regions, by adding a coupling term like the second term in Eq.(4.12). ## 5 COVID-19 The approaches that we have discussed in the previous sections are applicable to a large variety of infectious diseases. The main differences are in certain key parameters, such as the method of transmission of the pathogen, the incubation time, the infection and removal (mortality) rate, _etc._. They influence the resulting time evolution of the epidemic and lead, for example, to a different total duration of the epidemic, total number of infected and fatalities, _etc_. In this section, as a study case, we present data for the cumulative number of infected individuals in various countries during the COVID-19 pandemic, which started at the end of 2019 and is still ravaging the world. Since large scale testing is at the heart of many countries strategies to combat this pandemic, there is a large amount of publicly available data documenting the spread of the SARS-CoV-2 across the globe. Here we use data from public repositories [169] for the time period of 15/02/2020 until 17/08/2021. We use these data to highlight peculiarities of the time evolution of the spread of the disease, namely the previously mentioned distinct multi-wave structure of repeated phases of exponential growth in the number of infected individuals interspersed with phases of (quasi-)linear growth: Figure 34 shows examples of the first of such waves in countries taken from all around the globe. The plots show the cumulative number of infected individuals as well as the cumulative number of deaths. These plots provide examples of the epidemiological dynamics under very different conditions not only with regards to geographical (_e.g._ size of the population, population density, level of urbanisation), climatic, economical (_e.g._ the gross national product of each country), socio-cultural and political factors (_e.g._ the level of medical care the population has access to), but also different strategies the countries have deployed to combat the epidemic. While this has lead to a different dynamics in each country (with regards for example to the total number of cases or the duration of the wave), the infection numbers all follow a similar shape. Indeed, as the solid lines in Fig. 34 shows, in each case the data can be fitted with a a logistic function of the form of Eq.(4.5) and the differences only lie with the different numerical parameters, as reported in Table 1. Country | $A$ | $\lambda$ | $B$ ---|---|---|--- Australia | $6854\pm 24$ | $0.2095\pm 0.0051$ | $7509\pm 1615$ Azerbaijan | $37703\pm 136$ | $0.0547\pm 0.0004$ | $2013\pm 110$ Brazil | $5624661\pm 26343$ | $0.0333\pm 0.0003$ | $314\pm 14$ Canada | $101987\pm 544$ | $0.0716\pm 0.0012$ | $222\pm 18$ Germany | $177112\pm 910$ | $0.1192\pm 0.003$ | $399\pm 58$ Kenya | $39469\pm 236$ | $0.0571\pm 0.0006$ | $13234\pm 1225$ New Zealand | $1491\pm 3$ | $0.2133\pm 0.0032$ | $24095\pm 3636$ South Africa | $686350\pm 1586$ | $0.0637\pm 0.0007$ | $19449\pm 1910$ Table 1: Parameters of the logistic function in Eq.(4.5) obtained by fitting the epidemiological curves (cumulative number of infected) shown in Fig. 34. Furthermore, Fig. 34 also shows the cumulative number of deaths in each country, which can also be fitted with a logistic function of the same form. As discussed in Section 4, the fact that epidemiological curves in very different regions of the world, under very different circumstances, can be fitted with a single class of functions is due to a self-similarity structure of the corresponding dynamics. Indeed, the underlying symmetry principle that organises the spread of the disease around (near) fixed points of flow equations is at the heart of the eRG approach. Figure 34: Cumulative number of individuals infected with SARS-Cov-2 and cumulative number of deaths during the first wave in countries across all continents. The dots represent the data reported at [169] (averaged over a week) and the coloured lines fits with logistic functions of the form (4.5). After the first wave of COVID-19, most countries have entered into an endemic phase, where the cumulative number of infected individuals has grown linearly, followed by further waves. As an example, Figs 35–38 show the infection numbers in 48 European countries for the entire duration of the pandemic so far. We have highlighted in each of them individual waves and have fitted them with a logistic function as a solution of the eRG approach that we have reviewed in Section 4.101010For ease of visibility, we have focused on the larger such epidemiological episodes. As can be seen from the quality of the fit, although these functions only have three parameters $(A,B,\lambda)$, they capture correctly the cumulative number of cases despite the fact that the data (even for different waves within the same country) represent very different epidemiological situations: * 1. the countries show large geographical, climatic as well as socio-cultural differences; * 2. the waves occur during different seasons under different meteorological conditions; * 3. during each wave the governments of these countries have imposed different non-pharmaceutical interventions to reduce the spread of the virus; * 4. since the beginning of 2021 all countries have started vaccination campaigns which have lead to a rate of roughly 60% of all adults across Europe being fully vaccinated by summer 2021; * 5. since the beginning of the pandemic, the SARS-CoV-2 virus has mutated multiple times and several different variants (with different infection and mortality rates as well as different efficacy for the vaccines) have dominated certain periods of the epidemiological dynamics. As it is visible from the plots in Figs 35–38, despite all of these differences, the cumulative number of infected can still be organised by a self-similarity principle, which is characterised by logistic functions. Finally, in Figs 35–38 we have restricted ourselves to fit waves that occurred before the summer of 2021. Many countries, however, show in late summer/early fall of 2021 once more a tendency of growing infection numbers, which (despite the vaccination efforts), may indicate the onset of new waves. Figure 35: Cumulative number of individuals infected with SARS-CoV-2 from 15/02/2020 until 17/08/2021 in different countries of Europe. The red dots represent the data reported at [169] and the coloured lines fits with logistic functions of the form (4.5). The coloured regions indicate the time frame over which the data were fitted for a single wave. Figure 36: Cumulative number of individuals infected with SARS-CoV-2 (contd.) Figure 37: Cumulative number of individuals infected with SARS-CoV-2 (contd.) Figure 38: Cumulative number of individuals infected with SARS-CoV-2 (contd.) Figure 39: Cumulative number of individuals infected with SARS-CoV-2 (contd.) Figure 40: Cumulative number of individuals infected with SARS-CoV-2 (contd.) ## 6 Outlook and Conclusions The study of the time evolution of infectious diseases is a long standing subject: the impact of pandemics on human society cannot be overstated (as the recent devastating case of COVID-19 has highlighted). Consequently, over the course of more than a century, numerous approaches and mathematical models have been proposed with the aim to predict the spread of diseases among a population, devise tools to estimate their biological, social and economical impact and develop strategies to mitigate the harm done to society as a whole. In this report we give a review of this endeavour that is inspired by theoretical physics, in particular the study of phase transitions and critical phenomena, encompassed by the framework of field theory. Indeed, we organise mathematical models ranging from ‘microscopic’ models, in which the spread of the disease is modelled at the individual level, to ‘effective’ models, in which these microscopic interactions have been ‘summed up’ and replaced by the description of the time evolution of suitable macroscopic degrees of freedom. We give concrete examples in each case and show how they are related to one another. We also show how to extend the models to account for observed phenomena, like multi-wave dynamics and the emergence of time-dependent symmetries such as approximate time-dilation invariance. We start with lattice and percolation models in Section 2. These are among the most ‘microscopic’ models and allow to simulate the spread of a disease at the level of individuals, therefore permitting to easily incorporate biological and social peculiarities related to the transfer of the disease from an infected individual to a susceptible one. Typically at great computational cost, these models provide insight into how these details influence the time evolution of the disease at larger scales and can highlight emerging patterns and symmetries. Indeed, via numerical analyses of simple models, we show in Section 2 the emergence of critical behaviour: as a function of some key parameters, the system undergoes a phase transition from a state where only a small fraction of the population gets infected over time to a state where a significant portion of individuals is affected. Near the critical point, this behaviour can be cast into a field theoretical description for which we review an action formalism. We next argue that mean field and averaging procedures of percolation models naturally lead to compartmental models. The latter are among the oldest descriptions of epidemiological processes (the SIR-model dating back almost a century) and are ubiquitous in the modern study of infectious diseases. As we review in Section 3, following our classification of approaches, compartmental models are effective descriptions: rather than describing the spread of a disease among individuals of the population, they comprise (first order) differential equations that yield (among others) the total number of infectious individuals in the population. The microscopic details of the spread of the disease have been ‘averaged’ and enter into the details of the equations. The seeming loss of control over the microscopic details of the infectious dynamics comes at the benefit of a more ‘global’ description of the disease (and typically a reduced computational cost). In Section 3, we provide an in-depth review of SIR-like compartmental models that, from a theoretical vantage point, elucidates their mechanics and dynamics. We analyse, review and extend the models to take into account single-wave dynamics, multi-wave patterns and even superspreaders, thus highlighting the flexibility of the approach as a whole. Finally, we also discuss that these models can be re- organised in a fashion to make efficient use of time-scaling symmetries of the epidemiological dynamics and which emphasises the role of fixed points. In Section 4 we develop these ideas further and discuss the epidemic Renormalisation Group framework, which is in fact organised around the symmetry principle of time-scale invariance of the diffusion solutions. Using intuition from particle physics, the epidemiological process is described through flow equations (called beta-functions), which govern the trajectories of the system that connect different fixed points. The latter correspond to stationary solutions of the dynamics, in which either no disease is present in the first place or it has been completely eliminated. By invoking an even richer structure of fixed points, an extended eRG approach allows to model multi-wave pandemics. The approaches and models outlined in this review can be adapted to a large range of different situations and cases: in Section 5 we have presented results related to the COVID-19 pandemic. We highlight how the multi-wave dynamics, as well as the impact of non-pharmaceutical interventions, vaccines and the geographical mobility of the (a portion of the) population can be modelled by the approaches outlined in the previous sections. In order to keep the discussion as simple as possible and to focus on the underlying ‘physics’, we have illustrated the ideas in this review by rather simple models. The latter can be much more refined and, for example, take into account other aspects and phenomena related to the spread of diseases. These range, for example, from developing strategies for protecting the population by implementing efficient vaccination campaigns and concrete strategies on the use of non-pharmaceutical interventions (such as lockdowns, social distancing measures and travel restrictions), to gauging the impact of mutations and adaptation mechanisms of pathogens [170, 171]. We have refrained from working out the latter in detail, but instead refer the reader to more specialised literature. Furthermore, due to their obvious applications, the tools developed in this review have been exclusively focused on the description of infectious diseases among a human population. While many of them have in fact been inspired by other systems (notably chemical reactions), they can be applied to other fields as well with equal ease and success: apart from the immediate applicability to other species (_e.g._ the spread of diseases among livestock), the ideas underlying the concrete models discussed here can be applied to a much larger range of problems. In fact, similar problems to the ones tackled here can be found in other complex systems as well, ranging from applications in network systems (_e.g._ the spread of computer malware in a decentralised system) to human behaviour [172, 173] as well as social engineering and media science (_i.e._ the spread of ideas and information in a network/society). ## References ## References * [1] J. K. Taubenberger, D. M. Morens, 1918 influenza: The mother of all pandemics, Rev Biomed 17(1) (2006) 69–79. * [2] R. A. Weiss, How does hiv cause aids?, Science 260 (1993) 1273–9. doi:https://doi.org/doi:10.1126/science.8493571. * [3] D. C. Douec, M. Roeder, R. A. Koup, Emerging concepts in the immunopathogenesis of aids, Annual Review of Medicine 60 (2009) 471–84. doi:https://doi.org/doi:10.1146/annurev.med.60.041807.123549. * [4] D. Bernoulli, Essai d’une nouvelle analyse de la mortalité causée par la petite vt́role et des avantages de l’inoculation pour la prévenir, Mémoires de Mathématiques et de Physique, Académie Royale des Sciences, Paris (1760) 1–45. * [5] H. Heesterbeek, The law of mass-action in epidemiology: a historical perspective, in: K. Cuddington, B. Beisner (Eds), Epidemiological Paradigms Lost: Routes of Theory Change, p.81-106, Academic Press, 2005. * [6] W. Farr, On the cattle plague, J. Soc. Sci. March 20. * [7] J. Snow, On Continuous Molecular Changes, More Particularly in Their Relation to Epidemic Diseases, J.Churchill, London. * [8] W. Hamer, Age-incidence in relation with cycles of disease prevalence, Trans. Epidem. Soc. London 15 (1896) 64–77. * [9] W. Hamer, Epidemic disease in England: The evidence of variability and of persistency of type; Lecture 1, Lancet (1906) 569–574. * [10] W. Hamer, Epidemic disease in England: The evidence of variability and of persistency of type; Lecture 2, Lancet (1906) 655–662. * [11] W. Hamer, Epidemic disease in England: The evidence of variability and of persistency of type; Lecture 3, Lancet (1906) 733–739. * [12] R. Ross, The Prevention of Malaria, second edition, John Murray, London. * [13] R. Ross, An application of the theory of probabilities to the study of _a priori_ pathometry: Part I, Proc. Roy. Soc. Lond. A 92 (1916) 204–230. * [14] R. Ross, H. Hudson, An application of the theory of probabilities to the study of _a priori_ pathometry: Part II, Proc. Roy. Soc. Lond. A 93 (1916) 212–225. * [15] R. Ross, H. Hudson, An application of the theory of probabilities to the study of _a priori_ pathometry: Part III, Proc. Roy. Soc. Lond. A 93 (1916) 225–240. * [16] A. McKendrick, The rise and fall of epidemics, Paludism (Transactions of the Committee for the Study of Malaria in India) 1 (1912) 54–66. * [17] A. McKendrick, Studies on the theory of continuous probabilities, with special reference to its bearing on natural phenomena of a progressive nature, Proceedings of the London Mathematical Society 13 (1914) 401–416. * [18] A. McKendrick, Applications of mathematics to medical problems, Proc. Edinburgh Math. Soc. 44 (1926) 98–130. * [19] W. O. Kermack, A. McKendrick, G. T. Walker, A contribution to the mathematical theory of epidemics, Proceedings of the Royal Society A 115 (1927) 700–721. doi:https://doi.org/10.1098/rspa.1927.0118. URL https://royalsocietypublishing.org/doi/10.1098/rspa.1927.0118 * [20] N. Bailey, The Mathematical Theory of Infectious Diseases, 2nd ed., Hafner, New York, 1975. * [21] N. Becker, The use of epidemic models, Biometrics 35 (1978) 295–305. * [22] e. C. Castillo-Chavez, Mathematical and Statistical Approaches to AIDS Epidemiology, Lecture Notes in Biomath, Springer-Verlag, Berlin 83. * [23] K. Dietz, Epidemics and rumours: A survey, J. Roy. Statist. Soc. Ser. A 130 (1967) 505–528. * [24] K. Dietz, Density dependence in parasite transmission dynamics, Parasit. Today 4 (1988) 91–97. * [25] H. Hethcote, A thousand and one epidemic models, in Frontiers in Theoretical Biology, S.A. Levin, ed., Lecture Notes in Biomath., Springer-Verlag, Berlin 100 (1994) 504–515. * [26] H. Hethcote, S. Levin, Periodicity in epidemiological models, in Applied Mathematical Ecology, L. Gross, T.G. Hallam, and S.A. Levin, eds., Springer-Verlag, Berlin (1989) 193–211. * [27] H. S. H.W. Hethcote, P. V. D. Driessche, Periodicity and stability in epidemic models: A survey, in Differential Equations and Applications in Ecology, Epidemics and Population Problems, S. N. Busenberg and K.L. Cooke, eds. , Academic Press, New York (1981) 65–82. * [28] K. Wickwire, Mathematical models for the control of pests and infectious diseases: A survey, Theoret. Population Biol. 11 (1977) 182–238. * [29] K. Dietz, D. Schenzle, Mathematical models for infectious disease statistics, in: A. Atkinson (Ed.), A Celebration of Statistics, Springer (1985) 167–204. * [30] H. W. Hethcote, Qualitative analyses of communicable disease models, Mathematical Biosciences 28 (3) (1976) 335 – 356. doi:https://doi.org/10.1016/0025-5564(76)90132-2. URL http://www.sciencedirect.com/science/article/pii/0025556476901322 * [31] E. R.M. Anderson, R.M. May, Infectious Diseases of Humans: Dynamics and Control, Oxford University Press, UK, 1991. * [32] H. W. Hethcote, The mathematics of infectious diseases, SIAM Review 42 (4). doi:https://doi.org/10.1137/S0036144500371907. * [33] e. R.M. Anderson, Population of Infectious Diseases, Chapman and Hall, London, 1982\. * [34] R. Anderson, E. R.M. May, Population Biology of Infectious Diseases, Springer Verlag, Berlin, Heidelberg, New York, 1982. * [35] R. Anderson, R. May, Population biology of infectious diseases I, Nature 180 (1979) 361–367. * [36] N. Bailey, The Biomathematics of Malaria, Charles Griffin, London, 1982. * [37] M. Bartlett, Stochastic Population Models in Ecology and Epidemiology, Methuen, London, 1960. * [38] N. Becker, Analysis of Infectious Disease Data, Chapman and Hall, New York, 1989\. * [39] S. Busenberg, K. Cooke, Vertically Transmitted Diseases, Biomathematics 23, Springer-Verlag, Berlin, 1993. * [40] V. Capasso, Mathematical Structures of Epidemic Systems, Lecture Notes in Biomath. 97, Springer-Verlag, Berlin, 1993. * [41] A. Cliff, P. Haggett, Atlas of Disease Distributions: Analytic Approaches to Epidemiological Data, Blackwell, London, 1988. * [42] D. Daley, J. Gani, Epidemic Modelling: An Introduction, Cambridge University Press, Cambridge, UK, 1999. * [43] O. Diekmann, J. Heesterbeek, Mathematical Epidemiology of Infectious Diseases, Wiley, New York, 2000. * [44] J. Frauenthal, Mathematical Modelling in Epidemiology, Springer-Verlag Universitext Berlin, 1980. * [45] C. L. J.P. Gabriel, P.Picard, Stochastic Processes in Epidemic Theory, Springer-Verlag, Berlin, 1990. * [46] B. Grenfell, E. A.P. Dobson, Ecology of Infectious Diseases in Natural Populations, Cambridge University Press, Cambridge UK, 1995. * [47] H. Hethcote, J. V. Ark, Modeling HIV Transmission and AIDS in the United States, Lecture Notes in Biomath. 95, Springer-Verlag, Berlin, 1992. * [48] H. Hethcote, J. Yorke, Gonorrhea Transmission Dynamics and Control, Lecture Notes in Biomath. 56, Springer-Verlag, Berlin, 1984. * [49] V. Isham, E. G. Medley, Models for Infectious Human Diseases, Cambridge University Press, Cambridge UK, 1996. * [50] E. J. Kranz, Epidemics of Plant Diseases: Mathematical Analysis and Modelling, Springer-Verlag, Berlin, 1990. * [51] H. Lauwerier, Mathematical Models of Epidemics, Mathematisch Centrum, Amsterdam, 1981. * [52] D. Ludwig, E. K.L. Cooke, Epidemiology, SIMS Utah Conference Proceedings, SIMS, Philadelphia, 1975. * [53] D. Mollison, Epidemic Models: Their Structure and Relation to Data, Cambridge University Press, Cambridge UK, 1996. * [54] I. Nasell, Hybrid Models of Tropical Infections, Springer-Verlag, Berlin, 1985. * [55] M. Scott, E. G. Smith, Parasitic and Infectious Diseases, Academic Press, San Diego, 1994. * [56] J. Vanderplank, Plant Diseases: Epidemics and Control, Academic Press, New York, 1963. * [57] P. Waltman, Deterministic Threshold Models in the Theory of Epidemics, Lecture Notes in Biomath. 1, Springer-Verlag, Berlin, 1974. * [58] P. J. Flory, Molecular size distribution in three dimensional polymers i. gelation, J. Am. Chem. Soc. 63 (1941) 3083. * [59] W. H. Stockmayer, Theory of molecular size distribution and gel formation in branched polymers ii. general cross linking, Journal of Chemical Physics 12,4 (1944) 125. * [60] S. Broadbent, J. Hammersley, Percolation Processes, Proc. Camb. Phil. Soc. 53 (1957) 629–41. * [61] C. Domb, Fluctuation phenomena and stochastic processes, Nature 184 (1959) 509–12. doi:https://doi.org/10.1038/184509a0. * [62] H. Frisch, J. Hammersley, Percolation Processes and Related Topics, J. Soc. Indust. Appl. Math. 11 (1963) 894–917. * [63] J. Essam, Graph theory and statistical physics, Discrete Math. 1 (1971) 83–112. * [64] V. Shante, S. Kirkpatrick, An introduction to percolation theory, Adv. Phys. 20 (1971) 325–57. * [65] J. Essam, Phase Transitions and Critical Phenomena vol 2, chap 6, ed. C. Domb and M.S. Green, pp.197-270, New York: Academic, 1972. * [66] S. Kirkpatrick, The nature of percolation ‘channels’, Solid St. Commun. 12 (1973) 1279–83. * [67] S. Kirkpatrick, Percolation and Conduction, Rev. Mod. Phys. 45 (1973) 574–88. * [68] D. Welsh, Percolation and Related Topics, Sci. Prog. Oxford. 64 (1977) 65–83. * [69] F. Wu, Studies in Foundations and Combinatorics, Adv. Math. Suppl. Studies. vol 1 (1978) 151–66. * [70] D. Stauffer, Scaling theory of percolation clusters, Phys. Rep. 54 (1979) 1–74. * [71] J. W. Essam, Percolation theory, Rep. Prog. Phys. 43 (1980) 833. URL https://iopscience.iop.org/article/10.1088/0034-4885/43/7/001/pdf * [72] M. Doi, Second quantization representation for classical many-particle system, J. Phys. A: Math. Gen. 9 (1976) 1465. URL https://iopscience.iop.org/article/10.1088/0305-4470/9/9/008 * [73] M. Doi, Stochastic theory of diffusion-controlled reaction, J. Phys. A: Math. Gen. 9 (1976) 1479. doi:https://doi.org/10.1016/S0378-4371(03)00458-8. * [74] L. Peliti, Path integral approach to birth-death processes on a lattice, J. Phys. France (Paris) 46 (1985) 1469–1483. doi:htpps://doi.org/10.1051/jphys:019850046090146900. * [75] G. Pruessner, Field theory notes, chapter 6, wwwf.imperial.ac.uk/$\sim$pruess/publications/Gunnar_Pruessner_field_theory_notes.pdfdoi:http://wwwf.imperial.ac.uk/~pruess/publications/Gunnar_Pruessner_field_theory_notes.pdf. * [76] J. L. Cardy, P. Grassberger, Epidemic models and percolation, Journal of Physics A: Mathematical and General 18 (6) (1985) L267–L271. doi:10.1088/0305-4470/18/6/001. URL https://doi.org/10.1088/0305-4470/18/6/001 * [77] H. Abarbanel, J. Bronzan, Structure of the pomeranchuk singularity in reggeon field theory, Phys. Rev. D 9 (1974) 2397. doi:https://doi.org/10.1103/PhysRevD.9.2397. * [78] J. Cardy, R. Sugar, Directed percolation and reggeon field theory, J. Phys. A: Math. Gen. 13 (1980) L423. URL https://iopscience.iop.org/article/10.1088/0305-4470/13/12/002/meta * [79] L. Breiman, Statistical Modeling: The Two Cultures (with comments and a rejoinder by the author), Statistical Science 16 (3) (2001) 199 – 231. doi:10.1214/ss/1009213726. URL https://doi.org/10.1214/ss/1009213726 * [80] A. Flaxman, T. Vos, Machine learning in population health: Opportunities and threats, PLOS Medicine 15 (2018) e1002702. doi:10.1371/journal.pmed.1002702. * [81] T. Wiemken, R. Kelley, R. Fernández-Botrán, W. Mattingly, F. Arnold, S. Furmanek, M. Restrepo, J. Chalmers, P. Peyrani, J. Bordón, S. Aliberti, J. Ramírez, Using cluster analysis of cytokines to identify patterns of inflammation in hospitalized patients with community-acquired pneumonia: a pilot study, Journal of Respiratory Infections 1. doi:10.18297/jri/vol1/iss1/1/. * [82] A. Motsinger-Reif, S. Dudek, L. Hahn, M. Ritchie, Comparison of approaches for machine-learning optimization of neural networks for detecting gene-gene interactions in genetic epidemiology, Genetic epidemiology 32 (2008) 325–40. doi:10.1002/gepi.20307. * [83] A. Ramasubramanian, R. Muckom, C. Sugnaux, C. Fuentes, B. L. Ekerdt, D. S. Clark, K. E. Healy, D. V. Schaffer, High-throughput discovery of targeted, minimally complex peptide surfaces for human pluripotent stem cell culture, ACS Biomaterials Science & Engineering 7 (4) (2021) 1344–1360, pMID: 33750112. arXiv:https://doi.org/10.1021/acsbiomaterials.0c01462, doi:10.1021/acsbiomaterials.0c01462. URL https://doi.org/10.1021/acsbiomaterials.0c01462 * [84] T. Wiemken, R. Kelley, Machine learning in epidemiology and health outcomes research, Annual Review of Public Health 41 (2020) 21–36. * [85] Z. Wang, M. A. Andrews, Z.-X. Wu, L. Wang, C. T. Bauch, Coupled disease–behavior dynamics on complex networks: A review, Physics of Life Reviews 15 (2015) 1 – 29. doi:https://doi.org/10.1016/j.plrev.2015.07.006. URL http://www.sciencedirect.com/science/article/pii/S1571064515001372 * [86] K. G. Wilson, Renormalization group and critical phenomena. 1. Renormalization group and the Kadanoff scaling picture, Phys. Rev. B 4 (1971) 3174–3183. doi:https://doi.org/10.1103/PhysRevB.4.3174. * [87] K. G. Wilson, Renormalization group and critical phenomena. 2. Phase space cell analysis of critical behavior, Phys. Rev. B 4 (1971) 3184–3205. doi:https://doi.org/10.1103/PhysRevB.4.3184. * [88] O. Reynolds, An experimental investigation of the circumstances which determine whether the motion of water shall be direct or sinuous, and of the law of resistance in parallel channels, Phil. Trans. R. Soc. Lond. 174 (1883) 935–982. URL http://rstl.royalsocietypublishing.org/con-tent/174/935 * [89] O. Reynolds, On the dynamical theory of incompressible viscous fluids and the determination of the criterion, Phil. Trans. R. Soc. Lond. 186 (1895) 123–164. URL http://rsta.royalsocietypublishing.org/content/186/123 * [90] E. Stueckelberg, A. Petermann, La renormalisation des constants dans la theorie de quanta, Helv. Phys. Acta 26. * [91] M. Gell-Mann, F. E. Low, Quantum electrodynamics at small distances, Phys. Rev. 95 (1954) 1300–1312. doi:10.1103/PhysRev.95.1300. URL https://link.aps.org/doi/10.1103/PhysRev.95.1300 * [92] M. Della Morte, D. Orlando, F. Sannino, Renormalization Group Approach to Pandemics: The COVID-19 Case, Front. in Phys. 8 (2020) 144. doi:https://doi.org/10.3389/fphy.2020.00144. * [93] M. Della Morte, F. Sannino, Renormalisation Group approach to pandemics as a time-dependent SIR model, Front. in Phys. 8 (2021) 583. doi:https://doi.org/10.3389/fphy.2020.591876. * [94] G. Cacciapaglia, F. Sannino, Interplay of social distancing and border restrictions for pandemics (COVID-19) via the epidemic Renormalisation Group framework, Sci Rep 10 (2020) 15828. arXiv:2005.04956, doi:https://doi.org/10.1038/s41598-020-72175-4. * [95] G. Cacciapaglia, C. Cot, F. Sannino, Second wave covid-19 pandemics in europe: A temporal playbook, Sci Rep 10 (2020) 15514. arXiv:2007.13100, doi:https://doi.org/10.1038/s41598-020-72611-5. * [96] G. Cacciapaglia, C. Cot, A. S. Islind, M. Óskarsdóttir, F. Sannino, Impact of us vaccination strategy on covid-19 wave dynamics, Sci Rep 11(1) (2021) 1–11. arXiv:2012.12004. * [97] M. Perc, J. J. Jordan, D. G. Rand, Z. Wang, S. Boccaletti, A. Szolnoki, Statistical physics of human cooperation, Physics Reports 687 (2017) 1 – 51. doi:https://doi.org/10.1016/j.physrep.2017.05.004. URL http://www.sciencedirect.com/science/article/pii/S0370157317301424 * [98] Z. Wang, C. T. Bauch, S. Bhattacharyya, A. d’Onofrio, P. Manfredi, M. Perc, N. Perra, M. Salathé, D. Zhao, Statistical physics of vaccination, Physics Reports 664 (2016) 1 – 113. doi:https://doi.org/10.1016/j.physrep.2016.10.006. URL http://www.sciencedirect.com/science/article/pii/S0370157316303349 * [99] P. Grassberger, On the critical behavior of the general epidemic process and dynamical percolation, Mathematical Biosciences 63 (2) (1983) 157 – 172. doi:https://doi.org/10.1016/0025-5564(82)90036-0. URL http://www.sciencedirect.com/science/article/pii/0025556482900360 * [100] T. Tomé, R. M. Ziff, Critical behavior of the susceptible-infected-recovered model on a square lattice, Phys. Rev. E 82 (2010) 051921. doi:10.1103/PhysRevE.82.051921. URL https://link.aps.org/doi/10.1103/PhysRevE.82.051921 * [101] G. Santos, T. Alves, G. Alves, A. Macedo-Filho, R. Ferreira, Epidemic outbreaks on two-dimensional quasiperiodic lattices, Physics Letters A 384 (2) (2020) 126063. doi:https://doi.org/10.1016/j.physleta.2019.126063. URL https://www.sciencedirect.com/science/article/pii/S0375960119309533 * [102] R. I. Mukhamadiarov, S. Deng, S. R. Serrao, Priyanka, R. Nandi, L. H. Yao, U. C. Täuber, Social distancing and epidemic resurgence in agent-based susceptible-infectious-recovered models, Scientific Reports 11 (1). doi:10.1038/s41598-020-80162-y. URL http://dx.doi.org/10.1038/s41598-020-80162-y * [103] R. Bellman, A markovian decision process, Journal of Mathematics and Mechanics 6 (5) (1957) 679–684. URL http://www.jstor.org/stable/24900506 * [104] A. Howard, Dynamic Programming and Markov Processes, M.I.T. Press, 1960. * [105] E. Arashiro, T. Tomé, The threshold of coexistence and critical behaviour of a predator–prey cellular automaton, Journal of Physics A: Mathematical and Theoretical 40 (5) (2007) 887–900. doi:10.1088/1751-8113/40/5/002. URL http://dx.doi.org/10.1088/1751-8113/40/5/002 * [106] D. Mollison, Spatial contact models for ecological and epidemic spread, J. Roy. Statist. Soc. Ser. B 39 (1977) 283. doi:https://doi.org/10.1111/j.2517-6161.1977.tb01627.x. * [107] N. T. J. Bailey, The mathematical theory of infectious diseases, Griffin, London. * [108] A. Menon, N. Rajendran, A. Chandrachud, G. Setlur, Modelling and simulation of covid-19 propagation in a large population with specific reference to indiadoi:10.1101/2020.04.30.20086306. * [109] X. Liu, G. Hewings, S. Wang, M. Qin, X. Xiang, S. Zheng, X. Li, Modelling the situation of covid-19 and effects of different containment strategies in china with dynamic differential equations and parameters estimation, medRxivarXiv:https://www.medrxiv.org/content/early/2020/03/13/2020.03.09.20033498.full.pdf, doi:10.1101/2020.03.09.20033498. URL https://www.medrxiv.org/content/early/2020/03/13/2020.03.09.20033498 * [110] A. Omran, The epidemiological transition: A theory of the epidemiology of population change, The Milbank Quarterly 83 (4) (1971) 731–57. doi:doi:10.1111/j.1468-0009.2005.00398.x. URL https://doi.org/10.1111/j.1468-0009.2005.00398.x * [111] A. Santosa, S. Wall, E. Fottrell, U. H ogsberg, P. Byass, The development and experience of epidemiological transition theory over four decades: a systematic review, Global Health Action 7 (2014) 23574. doi:doi:10.3402/gha.v7.23574. URL https://doi.org/10.3402/gha.v7.23574 * [112] W. H. Press, B. P. Flannery, S. A. Teukolsky, W. T. Vetterling, Numerical Recipes in FORTRAN: The Art of Scientific Computing, 2nd ed., Cambridge University Press, 1992. * [113] J. C. Butcher, Numerical Methods for Ordinary Differential Equations, New York: John Wiley & Sons, 2003. * [114] P. L. Delamater, E. J. Street, T. F. Leslie, Y. Yang, K. H. Jacobsen, Complexity of the Basic Reproduction Number (R0), Emerg Infect Dis. 25(1) (2019) 1–4. doi:https://dx.doi.org/10.3201/eid2501.171901. URL https://dx.doi.org/10.3201/eid2501.171901 * [115] J. A. P. Heesterbeek, K. Dietz, The concept of ro in epidemic theory, Statistica Neerlandica 50 (1) (1996) 89–110. arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.1111/j.1467-9574.1996.tb01482.x, doi:https://doi.org/10.1111/j.1467-9574.1996.tb01482.x. URL https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1467-9574.1996.tb01482.x * [116] M. Keeling, B. T. Grenfell, Individual-based perspectives on r0, Journal of Theoretical Biology 203 (1) (2000) 51–61. doi:https://doi.org/10.1006/jtbi.1999.1064. URL https://www.sciencedirect.com/science/article/pii/S0022519399910640 * [117] J. Heesterbeek, A brief history of $r_{0}$ and a recipe for its calculation., Acta Biotheoretica 50 (2002) 189–204. doi:10.1023/A:1016599411804. * [118] J. Heffernan, R. Smith, L. Wahl, Perspectives on the basic reproductive ratio, Journal of The Royal Society Interface 2 (4) (2005) 281–293. arXiv:https://royalsocietypublishing.org/doi/pdf/10.1098/rsif.2005.0042, doi:10.1098/rsif.2005.0042. URL https://royalsocietypublishing.org/doi/abs/10.1098/rsif.2005.0042 * [119] M. Roberts, The pluses and minuses of r0, Journal of the Royal Society, Interface 4 (16) (2007) 949—961. doi:10.1098/rsif.2007.1031. URL https://europepmc.org/articles/PMC2075534 * [120] L. Pellis, F. Ball, P. Trapman, Reproduction numbers for epidemic models with households and other social structures. i. definition and calculation of r0, Mathematical Biosciences 235 (1) (2012) 85–97. doi:https://doi.org/10.1016/j.mbs.2011.10.009. URL https://www.sciencedirect.com/science/article/pii/S0025556411001556 * [121] D. R. de Souza, T. Tomé, Stochastic lattice gas model describing the dynamics of the sirs epidemic process, Physica A: Statistical Mechanics and its Applications 389 (5) (2010) 1142?1150. doi:10.1016/j.physa.2009.10.039. URL http://dx.doi.org/10.1016/j.physa.2009.10.039 * [122] T. Tomé, R. M. Ziff, Critical behavior of the susceptible-infected-recovered model on a square lattice, Physical Review E 82 (5). doi:10.1103/physreve.82.051921. URL http://dx.doi.org/10.1103/PhysRevE.82.051921 * [123] G. Santos, T. Alves, G. Alves, A. Macedo-Filho, R. Ferreira, Epidemic outbreaks on two-dimensional quasiperiodic lattices, Physics Letters A 384 (2) (2020) 126063\. doi:10.1016/j.physleta.2019.126063. URL http://dx.doi.org/10.1016/j.physleta.2019.126063 * [124] T. F. A. Alves, G. A. Alves, A. Macedo-Filho, R. S. Ferreira, Epidemic outbreaks on random delaunay triangulations (2019). arXiv:1901.03029. * [125] R. Ghostine, M. E. Gharamti, S. Hassrouny, I. Hoteit, An extended seir model with vaccination for forecasting the covid-19 pandemic in saudi arabia using an ensemble kalman filter, Mathematics (9) (2021) 636. doi:10.3390/math9060636. * [126] X. Meng, Z. Cai, H. Dui, H. Cao, Vaccination strategy analysis with SIRV epidemic model based on scale-free networks with tunable clustering, IOP Conference Series: Materials Science and Engineering 1043 (3) (2021) 032012. doi:10.1088/1757-899x/1043/3/032012. URL https://doi.org/10.1088/1757-899x/1043/3/032012 * [127] S. Gao, Z. Teng, J. J. Nieto, A. Torres, Analysis of an SIR epidemic model with pulse vaccination and distributed time delay, J Biomed Biotechnol. 2007 (2007) 64870. doi:https://doi.org/10.1155/2007/64870. * [128] N. Grassly, C. Fraser, Seasonal infectious disease epidemiology, Proceedings. Biological sciences / The Royal Society 273 (2006) 2541–50. doi:10.1098/rspb.2006.3604. * [129] S. Dowell, Seasonal variation in host susceptibility and cycles of certain infectious diseases., Emerging Infectious Diseases 7 (2001) 369 – 374. * [130] M. Keeling, P. Rohani, B. Pourbohloul, Modeling infectious diseases in humans and animals, Clinical infectious diseases : an official publication of the Infectious Diseases Society of America 47 (2008) 864–865. doi:10.1086/591197. * [131] Y. Wang, Y. Zhou, Mathematical modeling and dynamics of hiv progression and treatment, Chinese Journal of Engineering Mathematics 27. * [132] L. Liu, X.-Q. Zhao, Y. Zhou, A tuberculosis model with seasonality, Bulletin of mathematical biology 72 (4) (2010) 931—952. doi:10.1007/s11538-009-9477-8. URL https://doi.org/10.1007/s11538-009-9477-8 * [133] N. Bacaër, S. Guernaoui, The epidemic threshold of vector-borne diseases with seasonality: The case of cutaneous leishmaniasis in chichaoua, morocco, Journal of mathematical biology 53 (2006) 421–36. doi:10.1007/s00285-006-0015-0. * [134] Z. Agur, L. Cojocaru, G. Mazor, R. Anderson, Y. Danon, Pulse mass measles vaccination across age cohorts, Proceedings of the National Academy of Sciences of the United States of America 90 (1994) 11698–702. doi:10.1073/pnas.90.24.11698. * [135] L. Stone, B. Shulgin, Z. Agur, Theoretical examination of the pulse vaccination policy in the sir epidemic model, Mathematical and Computer Modelling 31 (2000) 207–215. * [136] B. Shulgin, L. Stone, Z. Agur, Pulse vaccination strategy in the sir epidemic model, Bulletin of Mathematical Biology 60 (6) (1998) 1123–1148. doi:https://doi.org/10.1016/S0092-8240(98)90005-2. URL https://www.sciencedirect.com/science/article/pii/S0092824098900052 * [137] S. Gao, L. Chen, J. J. Nieto, A. Torres, Analysis of a delayed epidemic model with pulse vaccination and saturation incidence, Vaccine 24 (35-36) (2006) 6037—6045. doi:10.1016/j.vaccine.2006.05.018. URL https://doi.org/10.1016/j.vaccine.2006.05.018 * [138] S. Gao, L. Chen, Z. Teng, Pulse vaccination of an seir epidemic model with time delay, Nonlinear Analysis: Real World Applications 9 (2008) 599–607. doi:10.1016/j.nonrwa.2006.12.004. * [139] S. Liu, Y. Pei, C. Li, L. Chen, Three kinds of tvs in a sir epidemic model with saturated infectious force and vertical transmission, Applied Mathematical Modelling 33 (4) (2009) 1923–1932. doi:https://doi.org/10.1016/j.apm.2008.05.001. URL https://www.sciencedirect.com/science/article/pii/S0307904X08001066 * [140] X. Liu, P. Stechlinski, Infectious disease models with time-varying parameters and general nonlinear incidence rate, Applied Mathematical Modelling 36 (5) (2012) 1974–1994. doi:https://doi.org/10.1016/j.apm.2011.08.019. URL https://www.sciencedirect.com/science/article/pii/S0307904X11005191 * [141] X. Meng, L. Chen, The dynamics of a new sir epidemic model concerning pulse vaccination strategy, Applied Mathematics and Computation 197 (2008) 582–597. doi:10.1016/j.amc.2007.07.083. * [142] A. D’Onofrio, Pulse vaccination strategy in the sir epidemic model: Global asymptotic stable eradication in presence of vaccine failures, Mathematical and Computer Modelling 36 (2002) 473–489. doi:10.1016/S0895-7177(02)00177-2. * [143] W. Chunjin, C. Lansun, A delayed epidemic model with pulse vaccination, Discrete Dynamics in Nature and Society 2008. doi:10.1155/2008/746951. * [144] Y. Zhou, H. Liu, Stability of periodic solutions for an sis model with pulse vaccination, Mathematical and Computer Modelling 38 (2003) 299–308. * [145] Y. He, S. Gao, D. Xie, An sir epidemic model with time-varying pulse control schemes and saturated infectious force, Applied Mathematical Modelling 37 (16) (2013) 8131–8140. doi:https://doi.org/10.1016/j.apm.2013.03.035. URL https://www.sciencedirect.com/science/article/pii/S0307904X13001947 * [146] S. Lai, N. W. Ruktanonchai, L. Zhou, O. Prosper, W. Luo, J. R. Floyd, A. Wesolowski, M. Santillana, C. Zhang, X. Du, H. Yu, A. J. Tatem, Effect of non-pharmaceutical interventions for containing the covid-19 outbreak in china, Naturedoi:https://doi.org/10.1038/s41586-020-2405-7. * [147] P. Liautaud, P. Huybers, M. Santillana, Fever and mobility data indicate social distancing has reduced incidence of communicable disease in the united statesarXiv:2004.09911. * [148] X. Huang, Z. Li, Y. Jiang, X. Ye, C. Deng, J. Zhang, X. Li, The characteristics of multi-source mobility datasets and how they reveal the luxury nature of social distancing in the u.s. during the covid-19 pandemic, medRxivdoi:http://doi.org/10.1101/2020.07.31.20143016. * [149] C. Cot, G. Cacciapaglia, F. Sannino, Mining google and apple mobility data: temporal anatomy for covid-19 social distancing, Scientific Reports 11 (1) (2021) 4150. doi:10.1038/s41598-021-83441-4. URL https://doi.org/10.1038/s41598-021-83441-4 * [150] J. T. Kemper, On the identification of superspreaders for infectious disease, Mathematical Biosciences 48 (1) (1980) 111 – 127. doi:https://doi.org/10.1016/0025-5564(80)90018-8. URL http://www.sciencedirect.com/science/article/pii/0025556480900188 * [151] I. Szapudi, Heterogeneity in sir epidemics modeling: superspreaders, medRxivarXiv:https://www.medrxiv.org/content/early/2020/07/06/2020.07.02.20145490.full.pdf, doi:10.1101/2020.07.02.20145490. URL https://www.medrxiv.org/content/early/2020/07/06/2020.07.02.20145490 * [152] J. Fox, E. Kilbourne, Epidemiology of influenza – summary of influenza workshop iv, J. Infectious Disease 128 (1973) 361–386. * [153] I. Elveback, J. Fox, E. Ackerman, A. Langworthy, M. Boyd, I. Gatewood, An influenza simulation model for immunisation studies, Amer. J. Epidemiology 103 (1976) 152–165. * [154] R. Hattis, K. e. a. Halstead, S.B. Herrman, Rubella is an immunised island population, JAMA 223 (1973) 1019–1021. * [155] D. Adam, P. Wu, J. Y. Wong, E. Lau, T. Tsang, S. Cauchemez, G. Leung, B. Cowling, Clustering and superspreading potential of severe acute respiratory syndrome coronavirus 2 (sars-cov-2) infections in hong kongdoi:10.21203/rs.3.rs-29548/v1. * [156] A. Schuchat, Public health response to the initiation and spread of pandemic covid-19 in the united states, february 24–april 21, 2020, MMWR. Morbidity and Mortality Weekly Report 69. doi:10.15585/mmwr.mm6918e2. * [157] R. R. Wilcox, The essence of gonorrhea control i, Acta Dermata-Venereologica 45 (1965) 302–308. * [158] R. R. Wilcox, The essence of gonorrhea control ii, Acta Dermata-Venereologica 46 (1966) 95–100. * [159] R. R. Wilcox, The essence of gonorrhea control iii, Acta Dermata-Venereologica 46 (1966) 250–256. * [160] R. R. Wilcox, The essence of gonorrhea control iv, Acta Dermata-Venereologica 46 (1966) 460–465. * [161] H. H. A. Yorke, A. Nold, Dynamics and control of the transmission of gonorrhea, Sexual $\&$ Transmitted Diseases 5(2) (1978) 51 – 56. * [162] M. McGuigan, Pandemic modeling and the renormalization group equations: Effect of contact matrices, fixed points and nonspecific vaccine waningarXiv:2008.02149. * [163] G. Cacciapaglia, F. Sannino, Evidence for complex fixed points in pandemic data, Front. Appl. Math. Stat. 7 (2021) 659580. arXiv:2009.08861, doi:https://doi.org/10.3389/fams.2021.659580. * [164] G. Cacciapaglia, C. Cot, F. Sannino, Multiwave pandemic dynamics explained: How to tame the next wave of infectious diseases, Sci Rep 11 (2021) 6638. arXiv:2011.12846, doi:https://doi.org/10.1038/s41598-021-85875-2. * [165] C. Stokel-Walker, What we know about covid-19 reinfection so far, BMJ 372. arXiv:https://www.bmj.com/content/372/bmj.n99.full.pdf, doi:10.1136/bmj.n99. URL https://www.bmj.com/content/372/bmj.n99 * [166] G. Cacciapaglia, C. Cot, A. de Hoffer, S. Hohenegger, F. Sannino, S. Vatani, Epidemiological theory of virus variants (2021). arXiv:2106.14982. * [167] A. de Hoffer, S. Vatani, C. Cot, G. Cacciapaglia, F. Conventi, A. Giannini, S. Hohenegger, F. Sannino, Variant-driven multi-wave pattern of covid-19 via machine learning clustering of spike protein mutations (2021). arXiv:2107.10115. * [168] T. Banks, A. Zaks, On the Phase Structure of Vector-Like Gauge Theories with Massless Fermions, Nucl. Phys. B 196 (1982) 189–204. doi:10.1016/0550-3213(82)90035-9. * [169] Worldometer, Coronavirus cases, https://www.worldometers.info/coronavirus/ (2021). URL https://www.worldometers.info/coronavirus/ * [170] Z. Wang, F. Schmidt, Y. Weisblum, F. Muecksch, C. O. Barnes, S. Finkin, D. Schaefer-Babajew, M. Cipolla, C. Gaebler, J. A. Lieberman, Z. Yang, M. E. Abernathy, K. E. Huey-Tubman, A. Hurley, M. Turroja, K. A. West, K. Gordon, K. G. Millard, V. Ramos, J. Da Silva, J. Xu, R. A. Colbert, R. Patel, J. Dizon, C. Unson-O’Brien, I. Shimeliovich, A. Gazumyan, M. Caskey, P. J. Bjorkman, R. Casellas, T. Hatziioannou, P. D. Bieniasz, M. C. Nussenzweig, mrna vaccine-elicited antibodies to sars-cov-2 and circulating variants, bioRxivarXiv:https://www.biorxiv.org/content/early/2021/01/19/2021.01.15.426911.full.pdf, doi:10.1101/2021.01.15.426911. URL https://www.biorxiv.org/content/early/2021/01/19/2021.01.15.426911 * [171] M. McCallum, A. D. Marco, F. Lempp, M. A. Tortorici, D. Pinto, A. C. Walls, M. Beltramello, A. Chen, Z. Liu, F. Zatta, S. Zepeda, J. di Iulio, J. E. Bowen, M. Montiel-Ruiz, J. Zhou, L. E. Rosen, S. Bianchi, B. Guarino, C. S. Fregni, R. Abdelnabi, S.-Y. Caroline Foo, P. W. Rothlauf, L.-M. Bloyet, F. Benigni, E. Cameroni, J. Neyts, A. Riva, G. Snell, A. Telenti, S. P. Whelan, H. W. Virgin, D. Corti, M. S. Pizzuto, D. Veesler, N-terminal domain antigenic mapping reveals a site of vulnerability for sars-cov-2, bioRxivarXiv:https://www.biorxiv.org/content/early/2021/01/14/2021.01.14.426475.full.pdf, doi:10.1101/2021.01.14.426475. URL https://www.biorxiv.org/content/early/2021/01/14/2021.01.14.426475 * [172] H. Wang, Q. Xia, Z. Xiong, Z. Li, W. Xiang, Y. Yuan, Y. Liu, Z. Li, The psychological distress and coping styles in the early stages of the 2019 coronavirus disease (covid-19) epidemic in the general mainland chinese population: A web-based survey, PLOS ONE 15 (5) (2020) 1–10. doi:10.1371/journal.pone.0233410. URL https://doi.org/10.1371/journal.pone.0233410 * [173] D. Sakan, D. Zuljevic, N. Rokvic, The role of basic psychological needs in well-being during the covid-19 outbreak: A self-determination theory perspective, Frontiers in Public Health 8 (2020) 713\. doi:10.3389/fpubh.2020.583181. URL https://www.frontiersin.org/article/10.3389/fpubh.2020.583181
See copyright ###### Abstract Polynomial multiplication is a bottleneck in most of the public-key cryptography protocols, including Elliptic-curve cryptography and several of the post-quantum cryptography algorithms presently being studied. In this paper, we present a library of various large integer polynomial multipliers to be used in hardware cryptocores. Our library contains both digitized and non- digitized multiplier flavours for circuit designers to choose from. The library is supported by a C++ generator that automatically produces the multipliers’ logic in Verilog HDL that is amenable for FPGA and ASIC designs. Moreover, for ASICs, it also generates configurable and parameterizable synthesis scripts. The features of the generator allow for a quick generation and assessment of several architectures at the same time, thus allowing a designer to easily explore the (complex) optimization search space of polynomial multiplication. ###### Index Terms: schoolbook multiplier, karatsuba multiplier, toom cook multiplier, digitized polynomial multiplication, Large integer polynomial multipliers ††publicationid: pubid: 978-1-5386-5541-2/18/$31.00 ©2021 IEEE ## I Introduction Polynomial multiplication (i.e., $c(x)=a(x)\times b(x)$) is a fundamental building block for cryptographic hardware and is often identified as the bottleneck in implementing efficient circuits. The most widely deployed public key crypto systems (e.g., RSA and ECC) need polynomial multiplications [1]. Many of the post-quantum cryptography (PQC) algorithms (e.g., NTRU-Prime, FrodoKEM, Saber, etc.) also require large integer multipliers for multiplying polynomial coefficients utilized to perform key-encapsulations and digital signatures [2]. Another application is in fully homomorphic encryption, a specific branch of cryptography that requires large integer multipliers to enable multi-party and secure-by-construction on the cloud computations [3]. There is a clear demand for large integer multipliers to perform multiplication over polynomial coefficients. To our knowledge, today, no widely available repository of open source multiplier architectures exists. This is the gap that our library addresses. There are several multiplication methods employed to perform multiplication over polynomial coefficients, including the schoolbook method (SBM), Karatsuba, Toom-Cook, Montgomery, and number theoretic transformation (NTT). A quick scan of the PQC algorithms involved in the NIST standardization effort [4] reveals that many reference implementations suggest the use of these multipliers: SBM is suggested by the authors of NTRU-Prime and FrodoKEM, Karatsuba and Toom-Cook methods are used in Saber and NTRU, a combination of NTT and SBM is suggested for CRYSTALS-Kyber, SBM and Montgomery are considered in Falcon. Examples of recent works employing non-digitized and digitized polynomial multiplication methods are given in [5, 6, 7, 8, 9, 10, 11] and [12, 13, 14], respectively. In [5], for different polynomial sizes, an architectural evaluation of different multiplication methods (SBM, comba, Karatsuba, Toom- Cook, Montgomery, and NTT) is performed over a Virtex-7 FPGA platform. An improved Montgomery polynomial multiplier is presented in [7] for a polynomial size of 1024 bits over a Virtex-6 FPGA. A run-time configurable and highly parallelized NTT-based polynomial multiplication architecture over Virtex-7 is discussed in [8]. A systolic based digit serial multiplier wrapper on an Intel Altera Stratix-V FPGA is described in [12], where digit sizes of 22 and 30 bits are considered for operand lengths 233 and 409 bits, respectively. A digit serial Montgomery based wrapper is provided in [13], where a digit size of 64 is selected for the operand length 571 bits, on a Virtex-6. Similarly, a digit serial modular multiplication based wrapper on Virtex-7 is shown in [14], where digit sizes of 2, 4 and 8 bits are preferred for an operand length of 2048 bits. ASIC implementations, while less frequent, also explore the polynomial multiplication design space. In [6], different polynomial multipliers with different operand lengths are considered for area and power evaluations on a 65nm technology. On similar technology, a bit level parallel-in-parallel-out (BL-PIPO) multiplier architecture and a modified interleaved modular reduction multiplication with bit-serial sequential architecture is proposed in [10, 9], respectively. Using a 65nm commercial node, for an operand length of 409 bits. For fully homomorphic encryption schemes, an optimized multi-million bit multiplier based on the Schonhage Strassen multiplication algorithm is described in [11] on 60nm technology node. Although there are several reported implementations of different multiplication methods [5, 6, 7, 8, 9, 10, 11, 12, 13, 14], these implementations tend to be specifically tailored for a given operand size and for a given target (e.g., high speed or low area). The matter is that this trade-off space is rather complicated to navigate without automation. Consequently, a common approach to assess (several) multiplication methods is required. In order to tackle the aforementioned limitations of the available literature and the need for automation, we develop an open-source library of multipliers which we name TTech-LIB. Our library is supported by a C++ generator utility that produces – following user specifications – hardware description of four selected multiplication methods: (a) SBM, (b) 2-way Karatsuba, (c) 3-way Toom- Cook, and (d) 4-way Toom-Cook. For selected multiplication methods, our library also offers a digitized solution: a single parameterized digit-serial wrapper to multiply polynomial coefficients. By default, the wrapper instantiates a singular SBM multiplier, but it can be replaced by any other multiplier method since the interfaces are identical between all methods. Finally, FPGA and ASIC designers can select their own multiplication method, size of the input operands, and digit size (only for the digitized wrapper, naturally). Moreover, for ASIC designers, there is the possibility to generate synthesis scripts for one of two synthesis tools, either Cadence Genus or Synopsys Design Compiler (DC). The user is not restricted to generating a single architecture at a time, the generator will produce multiple solutions if asked to do so, which will appear as separate Verilog (.v) files. The remainder of this work is structured as follows: The mathematical background for selected multiplication methods is described in Section II. The generator architecture and the structure of proposed TTech-LIB is provided in Section III. Section IV shows the experimental results and provide comparisons of non-digitized and digitized flavours of multiplication methods. Finally, Section V concludes the paper. ## II Mathematical background In this section, we present the mathematical formulations behind polynomial multiplication. We assume the inputs are two $m$-bit polynomials and the output is a polynomial of size $2m-1$. ### II-A Non-digitized multiplication The SBM is the traditional way to multiply two input polynomials $a(x)\times b(x)$, as shown in Eq. 1. To produce resultant polynomial $c(x)$ by performing bit by bit operations, it requires $2\times m$ clock cycles, $m^{2}$ multiplications and $(m-1)^{2}$ additions. $\displaystyle c(x)=\sum_{i=0}^{m-1}\sum_{j=0}^{m-1}a\textsubscript{i}b\textsubscript{j}x^{i+j}$ (1) Other approaches such as the 2-way Karatsuba, 3-way Toom-Cook, and 4-way Toom- Cook are more time efficient since they split the polynomials into $n$ equal parts, as shown in Eq. 2. The value of $n$ for 2-way Karatsuba, 3-way Toom- Cook and 4-way Toom-Cook multipliers is 2, 3 and 4, respectively and as the name implies. In Eq. 2, the variable $k$ determines the index of the split input polynomial. For example, for a 4-way Toom-Cook multiplier, the values of $k$ are {3, 2, 1, 0}, meaning the input polynomial $a(x)$ becomes $a_{3}(x)$, $a_{2}(x)$, $a_{1}(x)$, and $a_{0}(x)$. (2) In Eq. 3, the expanded version of Eq. 2 is presented for the case of 2-way split of input polynomials. The straightforward computation would require four multiplications: (1) one for the computation of inner product resulting polynomial $c_{1}(x)$, two multiplications for the computation of $c_{2}(x)$, and finally one multiplication for the computation of $c_{0}(x)$. However, $c_{2}(x)$ could be alternatively calculated with only one multiplication, as shown in Eq. 4. This is the Karatsuba observation. To generate the final resultant polynomial $c(x)$, addition of inner products is required, as presented in Eq. 5. Similarly, when considering the 3-way and 4-way Toom-Cook multipliers, the expanded versions of Eq. 2 produce nine and sixteen multiplications, respectively. These multiplications are then reduced to five and seven using a process similar to the 2-way Karatsuba, respectively. We omit the equations for Toom-Cook multipliers for the sake of brevity. (3) (4) (5) Now, let us assume that the polynomials involved in the multiplications above remain relatively large in size even after split. Thus, SBM multipliers can be employed to resolve the partial products. For a 2-way Karatsuba multiplier of $m$-bit input polynomials, there will be 3 SBM multipliers and each will take two polynomials of size $\frac{m}{2}$ as inputs. Each multiplier requires $\frac{m}{2}$ clock cycles to be completed. If all multipliers operate in parallel, the overall computation also takes $\frac{m}{2}$ cycles. For 3-way and 4-way splits, the number of clock cycles is $\frac{m}{3}$ and $\frac{m}{4}$, respectively. Since our library is aimed at large polynomials, the 2-way Karatsuba, 3-way Toom-Cook, and 4-way Toom-Cook codes available in it actually implement the parallel SBM strategy discussed above. In fact, our non-digitized multipliers are hybrid multipliers. ### II-B Digitized multiplication The digit serial wrapper in TTech-LIB takes two $m$-bit polynomials $a(x)$ and $b(x)$ as an input and produces $c(x)$ as an output. Digits are created for polynomial $b(x)$ with different sizes which are user-defined as follows: $d=\frac{m}{n}$, where $d$ determines the total number of digits, $m$ denotes the size of input polynomial $b(x)$, and $n$ is the size of each digit. Then, the multiplication of each created digit is performed serially with the input polynomial $a(x)$, while the final resultant polynomial $c(x)$ is produced using shift and add operations. The main difference here is that our digitized solution is serial, while the 2-, 3-, and 4-way multipliers are parallel. The required computational cost (in clock cycles) to perform one digit multiplication is $n$. Since there are $d$ digits, the overall computation takes $d\times n$ clock cycles. It is important to mention that users/designers can choose any multiplication method inside the described digit serial wrapper as per their application requirements. We have used an SBM multiplication method as default. ## III How to access TTech-LIB The complete project files (written in C++) are freely available to everyone on our GitHub repository [15]. A sample of pre-generated multipliers is also included in the repository. As shown in Fig. 1, the user settings can be customized by using a configuration file (config.xml). The structure of the library is rather simple and includes five directories: (1) bin, (2) run, (3) src, (4) synth, and (5) vlog. After running the generator binary, the produced synthesis scripts are put in the synth directory while the generated multipliers are put in the vlog folder. All generated multipliers have the same interface (i.e., inputs are $clk$, $rst$, $a$, and $b$; the output is $c$). Figure 1: Generator architecture and file structure of TTech-LIB ## IV Experimental Results and Comparisons ### IV-A Implementation results and evaluations The experimental results for non-digitized and digitized polynomial multiplication methods over NIST defined field lengths [16] on 65nm technology node using Genus, Cadence is provided in Table I and Table II, respectively. Moreover, the implementation results for various digit sizes of digitized SBM multiplication method over an Artix-7 FPGA device is given in Table III. In tables I–II, clock frequency (MHz), area (in $\mu m^{2}$), and power (mW) values are achieved after synthesis using Cadence Genus. Similarly, in Table III, clock frequency (MHz), look-up-tables (LUTs), utilized registers (Regs) and power (mW) values are achieved after synthesis using Vivado design tool. Finally, latency for both digitized and non-digitized multipliers (in tables I–III) is calculated using Eq. 6: (6) TABLE I: Results of non-digitized multipliers for NIST recommended Elliptic curves over prime and binary fields Multiplier | m | Freq (MHz) | latency ($\mu s$) | Area ($\mu m^{2}$) | Power (mW) ---|---|---|---|---|--- Schoolbook | P-192 | 500 | 0.382 | 32011.2 | 13.8 P-224 | 486 | 0.458 | 38048.0 | 17.1 P-256 | 480 | 0.531 | 48726.7 | 16.9 P-384 | 444 | 0.862 | 67861.8 | 27.1 P-521 | 434 | 1.198 | 100242.0 | 28.0 B-163 | 500 | 0.324 | 29341.4 | 12.9 B-233 | 476 | 0.487 | 39321.4 | 16.0 B-283 | 454 | 0.621 | 50603.4 | 17.8 B-409 | 442 | 0.923 | 73587.6 | 28.2 B-571 | 413 | 1.380 | 89993.2 | 29.1 2-way Karatsuba | P-192 | 473 | 0.202 | 41379.5 | 8.2 P-224 | 469 | 0.238 | 49514.4 | 9.6 P-256 | 467 | 0.274 | 59532.1 | 11.8 P-384 | 420 | 0.457 | 74844.0 | 15.2 P-521 | 408 | 0.639 | 105059.5 | 20.8 B-163 | 487 | 0.168 | 35060.0 | 7.7 B-233 | 478 | 0.244 | 52328.2 | 10.0 B-283 | 455 | 0.312 | 64743.8 | 12.6 B-409 | 432 | 0.474 | 84778.6 | 17.2 B-571 | 418 | 0.684 | 120374.3 | 21.7 3-way Toom-Cook | P-192 | 909 | 0.070 | 96498.4 | 44.4 P-224 | 869 | 0.086 | 102470.8 | 46.9 P-256 | 826 | 0.104 | 104820.9 | 49.4 P-384 | 689 | 0.185 | 139375.1 | 57.2 P-521 | 680 | 0.255 | 201341.2 | 80.0 B-163 | 934 | 0.058 | 75085.6 | 36.0 B-233 | 877 | 0.088 | 106357.7 | 49.5 B-283 | 800 | 0.118 | 115188.1 | 54.5 B-409 | 775 | 0.176 | 170509.0 | 78.4 B-571 | 766 | 0.249 | 256604.4 | 115.9 4-way Toom-Cook | P-192 | 900 | 0.053 | 105679.1 | 56.9 P-224 | 847 | 0.066 | 125124.1 | 62.0 P-256 | 826 | 0.077 | 122298.1 | 63.6 P-384 | 793 | 0.121 | 241893.7 | 98.2 P-521 | 767 | 0.170 | 332534.9 | 139.4 B-163 | 925 | 0.044 | 94834.1 | 49.9 B-233 | 892 | 0.066 | 132080.0 | 64.2 B-283 | 826 | 0.085 | 145709.3 | 70.6 B-409 | 769 | 0.133 | 236989.4 | 99.0 B-571 | 746 | 0.191 | 340750.8 | 148.2 * • m determines the field size or length of the inputs (in bits), where ‘P’ stands for Prime and ‘B’ stands for Binary TABLE II: Results of digitized multipliers for NIST recommended Elliptic curves over prime and binary fields m | digit size (n) | total digits (d) | Freq (MHz) | latency ($\mu s$) | Area ($\mu m^{2}$) | Power (mW) ---|---|---|---|---|---|--- 521$\times$521 | 32 | 17 | 505 | 1.07 | 106956.7 | 30.9 41 | 13 | 377 | 1.41 | 101538.7 | 26.1 53 | 10 | 340 | 1.55 | 94752.7 | 20.0 81 | 7 | 336 | 1.68 | 84321.0 | 15.4 571$\times$571 | 32 | 18 | 487 | 1.18 | 114999.8 | 36.7 41 | 14 | 369 | 1.55 | 116010.3 | 28.9 53 | 11 | 312 | 1.86 | 91393.9 | 18.1 81 | 8 | 291 | 2.22 | 76146.8 | 14.1 1024$\times$1024 | 2 | 512 | 363 | 2.82 | 196131.2 | 38.0 4 | 256 | 357 | 2.86 | 178581.2 | 35.1 8 | 128 | 353 | 2.90 | 167536.4 | 31.5 16 | 64 | 343 | 2.98 | 166533.1 | 30.2 32 | 32 | 313 | 3.27 | 148489.5 | 23.0 64 | 16 | 285 | 3.59 | 122257.8 | 20.8 128 | 8 | 268 | 3.82 | 123164.6 | 19.9 256 | 4 | 263 | 3.89 | 129542.4 | 19.5 512 | 2 | 261 | 3.92 | 136292.4 | 23.1 1024 | 1 | 259 | 3.95 | 177834.2 | 24.1 TABLE III: FPGA based results of digitized 1024$\times$1024 SBM multiplier for different digit sizes (Artix-7) m | digit size (n) | total digits (d) | Freq (MHz) | latency ($\mu s$) | LUTs | Regs | Carry | Power (mW) ---|---|---|---|---|---|---|---|--- 521$\times$521 | 32 | 17 | 33.11 | 16.43 | 6369 | 1692 | 408 | 184 41 | 13 | 29.15 | 18.28 | 7995 | 1681 | 416 | 192 53 | 10 | 28.32 | 22.72 | 8079 | 1732 | 417 | 191 64 | 9 | 34.48 | 15.12 | 6095 | 1758 | 408 | 220 81 | 8 | 30.30 | 21.38 | 8207 | 1795 | 415 | 247 | 128 | 5 | 34.84 | 14.95 | 5964 | 1881 | 424 | 220 571$\times$571 | 32 | 17 | 30.12 | 18.06 | 6397 | 1847 | 447 | 194 41 | 13 | 27.17 | 19.62 | 8750 | 1834 | 455 | 192 53 | 10 | 26.04 | 20.35 | 9053 | 1880 | 449 | 187 81 | 8 | 28.01 | 23.13 | 8958 | 1951 | 452 | 226 1024$\times$1024 | 2 | 512 | 14.22 | 72.11 | 10993 | 3634 | 1085 | 173 4 | 256 | 15.89 | 64.48 | 10824 | 3384 | 928 | 172 8 | 128 | 16.86 | 60.66 | 11074 | 3261 | 849 | 180 16 | 64 | 17.51 | 58.48 | 10634 | 3248 | 811 | 185 32 | 32 | 17.89 | 57.28 | 11371 | 3267 | 791 | 190 64 | 16 | 17.95 | 57.04 | 11947 | 3330 | 792 | 195 128 | 8 | 18.57 | 55.14 | 12207 | 3450 | 800 | 221 256 | 4 | 18.93 | 54.09 | 11367 | 3740 | 832 | 247 512 | 2 | 19.12 | 53.55 | 10385 | 4295 | 896 | 226 1024 | 1 | 18.46 | 55.50 | 11462 | 5303 | 1024 | 235 #### IV-A1 ASIC non-digitized multipliers Our results consider NIST-defined prime (P-192 to P-521) and binary (B-163 to B-571) fields utilized in ECC-based public key cryptosystems. As the operand sizes increase, the corresponding clock frequency decreases, as shown in column three of Table I. The decrease in frequency leads to an increase in latency, as presented in column four of Table I. In addition to latency, the corresponding area and power values also increase with the increase in size of multiplier operands (see columns five and six of Table I). It is evident from these results that the SBM multiplier requires less hardware resources than 2-way Karatsuba, 3-way Toom-Cook, and 4-way Toom-Cook multipliers. Moreover, the 2-way Karatsuba achieves lower power values as compared to other selected multipliers. This is explained by the datapath and the composition of the different multipliers. SBM requires $2m+2m$ bit adder, 2-way Karatsuba requires $m+m+m$ bit adder/subtracter for generating final polynomial, 3-way Toom-Cook requires fifteen $\frac{m}{3}$ bit incrementers, and 4-way Toom-Cook requires sixteen $\frac{m}{4}$ bit incrementers. There is always a trade-off between various design parameters such as area, latency, power etc. Consequently, the SBM multiplier is more useful for area constrained applications. For better latency, other multipliers are more convenient. #### IV-A2 ASIC digitized multipliers For digitizing, we have selected 521, 571, and 1024 as the lengths of the input operands, as shown in column one of Table II. Moreover, for input lengths of 521 and 571, digit sizes of 32, 41, 53 and 81 have been adopted. For an input length of 1024 bits, digit sizes are given in powers of two, for $n$ = $2,\ldots,1024$. Digit size $n$ and total digits $d$ are listed in columns two and three of Table II, respectively. It is noteworthy that the increase in digit size results in a decrease in clock frequency, as presented in column four of Table II. Moreover, it also translates to an increase in latency, as shown in column five of Table II. For the $1024\times 1024$ multiplier, the obtained values for area and power show behavior similar to a parabolic curve with respect to digit size, as given in the last two columns of Table II. This is intuitive, as in the extreme cases of too small or too large digits, the wrapper logic becomes inefficient and may even become the bottleneck for timing. In summary, for an application that requires high clock frequency, shorter digits are preferred; however, this brings a significant cost in area and power. #### IV-A3 FPGA digitized multipliers Alike ASIC demonstrations (presented in Sec. IV-A2), we have chosen similar lengths of the input operands (521, 571, and 1024) for the evaluation on an Artix-7 FPGA platform, as shown in column one of Table III.We have used Xilinx Viviado Desig Suite for the FPGA based experiments. Furthermore, for input lengths of 521 and 571, digit sizes of 32, 41, 53 and 81 have been considered. For an input length of 1024 bits, digit sizes are adopted in powers of two, for $n$ = $2,\ldots,1024$. Digit size $n$ and total digits $d$ are listed in columns two and three of Table III, respectively. The synthesis results (clock frequency, latency, area in terms of LUTs and Regs, and power) achieved for FPGA are totally distinct when compared to ASIC values as the implementation platforms are quite contrasting. It is important to note that the frequency of the multiplier architecture increases with the increase in digit size (shown in column four of Table III). This phenomenon keeps on-going until it reaches a saturation point (i.e., best possible performance in terms of clock frequency with respect to $n$). Once it reaches a saturation point, then there is a decrease in the clock frequency. Moreover, the saturation occurs at any digit size between 0 to $n$ (in this work and for this experiment, the saturation occurs when the value for $n$ = $512$). The saturation point also varies with the change in operand size of the multiplier as given in Table III. For other reported parameters, i.e., latency, LUTs and power, the saturation point is not possible to show as there is a non-linear behavior (see columns five, six and nine of Table III). It is noteworthy that we have considered the worst case scenario by excluding the DSP (Digital Signal Processing) blocks during synthesis. The performance of multiplier architectures will be higher by considering the conventional synthesis flow with DSPs. #### IV-A4 Figure-of-Merit (FoM) for digitized SBM multiplier A FoM is defined to perform a comparison while taking into account different design characteristics at the same time. A FoM to evaluate the latency and area parameters for both ASIC and FPGA platforms is defined using Eq. 7. The higher the FoM values, the better. Similarly, a ratio for latency and power characteristics are calculated considering Eq. 8. $FoM=\frac{1}{latency\,(\mu s)\times area}$ (7) $FoM=\frac{1}{latency\,(\mu s)\times power\,(mW)}$ (8) The calculated values of defined FoMs for ASIC are given in figures 2 and 3, where various digit sizes were considered for a $1024\times 1024$ multiplier. Figure 2: Area and latency FoM for various digit sizes of a $1024\times 1024$ multiplier Figure 3: Power and latency FoM for various digit sizes of a $1024\times 1024$ multiplier For both FoMs (shown in figures 2 and 3), it becomes clear that the extreme cases lead to suboptimal results. For the studied 1024 $\times$ 1024 multiplier, the variant with $n=64$ and $d=16$ presents an optimal solution. Other similar values, such as $n=32$ and $n=128$, also give very close to optimal solutions. Likewise ASICs, the calculated values of defined FoM (from Eq. 7) for FPGA is given in Fig. 4, where various digit sizes were considered for a 1024$\times$1024 multiplier. To calculate FPGA area utilizations, the slices flip-flops, LUTs and carry units are the basic building-blocks. Therefore, the FoM in Eq. 7 can be calculated by employing different metrics-of-interest (e.g., slices, LUTs, registers and carry blocks). Note that we have used an FPGA slices as area in Eq. 7. Fig. 4 reveals that the FoM value for $n=512$ and $d=2$ results an optimal solution. Figure 4: Slices and latency FoM for various digit sizes of a $1024\times 1024$ multiplier The combined relation between frequency, latency and power for different values of $n$ is illustrated in Fig. 5. Therefore, it is noted from Fig. 5 that the value of latency decreases, frequency increases with the increase in $n$. The increase in frequency and decrease in latency keeps on-going until saturation point occurs (when $n=512$). Figure 5: Frequency, latency and power analysis for various digit sizes of a $1024\times 1024$ multiplier ### IV-B Comparison to the state of the art To perform a fair comparison with existing state-of-the-art modular multiplier architectures, we have used similar operand lengths, digit sizes and implementation technologies (for FPGA and ASIC) as used in the corresponding solutions, shown in Table IV. In state-of-the-art solutions, multiplication results are given for different operands length. However, we have provided comparison of our results with only the larger operands. Moreover, we have used symbol ‘N/A’ in Table IV where the values for design parameters (Freq, latency and area) are not given. TABLE IV: Area and latency comparisons of non-digitized and digitized multipliers with state of the art Ref | Multiplier | Device | m | Freq (MHz) | latency ($\mu s$) | Area ($\mu m^{2}$)/LUTs ---|---|---|---|---|---|--- [5] | 2-way KM | V7 | 128 | 104.3 | 0.61 | 3499 2-way KM | V7 | 256 | 74.5 | 1.71 | 7452 2-way KM | V7 | 512 | 51.6 | 4.96 | 20474 [9] | BL-PIPO | 65nm | 163 | N/A | N/A | 5328 GE [13] | DSM (ds=64) | V6 | 571 | 258.5 | 0.03 | 10983 [14] | DSMM (ds=2) | V7 | 2048 | N/A | N/A | 18067 DSMM (ds=4) | V7 | 2048 | N/A | N/A | 33734 DSMM (ds=8) | V7 | 2048 | N/A | N/A | 62023 TW | SBM | 65nm | 163 | N/A | N/A | 11727 GE 2-way KM | V7 | 128 | 167.4 | 0.38 | 2110 2-way KM | V7 | 256 | 119.9 | 1.06 | 4318 2-way KM | V7 | 512 | 63.8 | 4.01 | 9582 SBM (ds=2) | V7 | 2048 | 15.03 | 69760 | 25559 SBM (ds=4) | V7 | 2048 | 16.6 | 15790 | 22040 SBM (ds=8) | V7 | 2048 | 17.4 | 3760 | 23315 SBM (ds=64) | V6 | 571 | 46.4 | 1.74 | 6181 * • V7: Xilinx Virtex-7, V6: Xilinx Virtex-6, ds: digit size, TW: this work, DSM: Digit Serial Montgomery multiplier based wrapper, BL-PIPO: Bit level parallel in parallel out multiplier using SBM multiplication method, GE: gate equivalents Concerning only the non-digitized multipliers for comparison, the 2-way Karatsuba multiplier of [5] over Virtex-7 FPGA for operand sizes of 128, 256 and 512 bits presents 38%, 39% and 20% higher latency when compared to 2-way Karatsuba multiplier generated by TTech-LIB, as shown in Table IV. Moreover, the generated multiplier utilizes lower hardware resources in terms of LUTs (see column seven in Table IV) as compared to resources (LUTs) utilized in [5]. On 65nm node, the BL-PIPO multiplier of [9] utilizes 55% lower hardware resources in terms of gate counts as compared to our SBM multiplier generated by TTech-LIB. When digitized flavor of polynomials multiplication is considered for comparison over different digit sizes, the digit serial Montgomery multiplier based wrapper of [13] results 83% higher clock frequency and requires 58% less computational time as compared to our SBM based digit serial wrapper generated by TTech-LIB. On the other hand, the SBM based digit serial wrapper results 56% lower hardware resources over Virtex-6 FPGA. There is always a trade-off between performance and area parameters. Another digit serial modular multiplication based wrapper of [14] results 14% (for ds=2) lower FPGA LUTs while for remaining digit sizes of 4 and 8, it utilizes 35% and 63% higher FPGA LUTs as compared to SBM wrapper generated by TTech-LIB. The frequency and latency parameters cannot be compared as these are not given. The comparisons and discussion above show that the multipliers generated by TTech-LIB provide a realistic and reasonable comparison to state-of-the-art multiplier solutions [5, 9, 13, 14]. Hence, not only can users explore various design parameters within our library, they can also benefit from implementations that are competitive with respect to the existing literature. ## V Conclusion This work has presented an open-source library for large integer polynomial multipliers. The library contains digitized and non-digitized flavors of polynomial coefficient multipliers. For non-digitized multipliers, based on the values for various design parameters, users/designers can select amongst several studied multipliers according to needs of their targeted application. Furthermore, we have shown that for digitized multipliers, the evaluation of individual design parameters may not be comprehensive, and figures of merit are better suited to capture the characteristics of a circuit. Furthermore, we believe the results enabled by TTech-LIB will guide hardware designers to select an appropriate digit size that reaches an acceptable performance according to application requirements. This is achieved with the aid of TTech- LIB’s generator, which helps a designer to quickly explore the complex design space of polynomial multipliers. ## References * [1] H. Eberle, N. Gura, S. Shantz, V. Gupta, L. Rarick, and S. Sundaram, “A public-key cryptographic processor for rsa and ecc.” IEEE, 2004, pp. 98–110. * [2] NIST, “Computer security resource centre: Pqc standardization process, third round candidate announcement,” 2020. [Online]. Available: https://csrc.nist.gov/news/2020/pqc-third-round-candidate-announcement * [3] A. López-Alt, E. Tromer, and V. Vaikuntanathan, “On-the-fly multiparty computation on the cloud via multikey fully homomorphic encryption,” in _Proceedings of the Forty-Fourth Annual ACM Symposium on Theory of Computing_ , ser. STOC ’12. New York, NY, USA: Association for Computing Machinery, 2012, p. 1219–1234. * [4] NIST, “Computer security resource centre: Post-quantum cryptography, round 2 submissions,” 2020. [Online]. Available: https://csrc.nist.gov/projects/post-quantum-cryptography/round-2-submissions * [5] C. Rafferty, M. O’Neill, and N. Hanley, “Evaluation of large integer multiplication methods on hardware,” _IEEE Transactions on Computers_ , vol. 66, no. 8, pp. 1369–1382, 2017. * [6] M. Imran, Z. U. Abideen, and S. Pagliarini, “An experimental study of building blocks of lattice-based nist post-quantum cryptographic algorithms,” _Electronics_ , vol. 9, no. 11, p. 1953, Nov 2020. * [7] A. A. Abd-Elkader, M. Rashdan, E.-S. A. Hasaneen, and H. F. Hamed, “Advanced implementation of montgomery modular multiplier,” _Microelectronics Journal_ , vol. 106, p. 104927, 2020. * [8] A. C. Mert, E. Öztürk, and E. Savaş, “FPGA implementation of a run-time configurable ntt-based polynomial multiplication hardware,” _Microprocessors and Microsystems_ , vol. 78, p. 103219, 2020. * [9] R. Azarderakhsh, K. U. Järvinen, and M. Mozaffari-Kermani, “Efficient algorithm and architecture for elliptic curve cryptography for extremely constrained secure applications,” _IEEE Transactions on Circuits and Systems I: Regular Papers_ , vol. 61, no. 4, pp. 1144–1155, 2014. * [10] S. R. Pillutla and L. Boppana, “An area-efficient bit-serial sequential polynomial basis finite field gf(2m) multiplier,” _AEU - International Journal of Electronics and Communications_ , vol. 114, p. 153017, 2020. [Online]. Available: http://www.sciencedirect.com/science/article/pii/S1434841119318485 * [11] Y. Doröz, E. Öztürk, and B. Sunar, “Accelerating fully homomorphic encryption in hardware,” _IEEE Transactions on Computers_ , vol. 64, no. 6, pp. 1509–1521, 2015. * [12] J. Xie, P. K. Meher, X. Zhou, and C. Lee, “Low register-complexity systolic digit-serial multiplier over $gf(2^{m})$ based on trinomials,” _IEEE Transactions on Multi-Scale Computing Systems_ , vol. 4, no. 4, pp. 773–783, 2018. * [13] M. Morales-Sandoval, C. Feregrino-Uribe, P. Kitsos, and R. Cumplido, “Area/performance trade-off analysis of an fpga digit-serial gf(2m) montgomery multiplier based on lfsr,” _Computers & Electrical Engineering_, vol. 39, no. 2, pp. 542 – 549, 2013. * [14] J. Pan, P. Song, and C. Yang, “Efficient digit-serial modular multiplication algorithm on fpga,” _IET Circuits, Devices Systems_ , vol. 12, no. 5, pp. 662–668, 2018. * [15] M. Imran, Z. U. Abideen, and S. Pagliarini, “TTech-LIB: Center for hardware security,” 2020. [Online]. Available: https://github.com/Centre-for-Hardware-Security/TTech-LIB * [16] C. Lily, M. Dustin, R. Andrew, and R. Karen, “Recommendations for discrete logarithm-based cryptography: Elliptic curve domain parameters,” 2020. [Online]. Available: https://nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP.800-186-draft.pdf
Research Article Daniel Lemire, DOT-Lab Research Center, Université du Québec (TELUQ), Montreal, Quebec, H2S 3L5, Canada<EMAIL_ADDRESS>Natural Sciences and Engineering Research Council of Canada, Grant Number: RGPIN-2017-03910 # Number Parsing at a Gigabyte per Second Daniel Lemire ###### Abstract With disks and networks providing gigabytes per second, parsing decimal numbers from strings becomes a bottleneck. We consider the problem of parsing decimal numbers to the nearest binary floating-point value. The general problem requires variable-precision arithmetic. However, we need at most 17 digits to represent 64-bit standard floating-point numbers (IEEE 754). Thus we can represent the decimal significand with a single 64-bit word. By combining the significand and precomputed tables, we can compute the nearest floating- point number using as few as one or two 64-bit multiplications. Our implementation can be several times faster than conventional functions present in standard C libraries on modern 64-bit systems (Intel, AMD, ARM and POWER9). Our work is available as open source software used by major systems such as Apache Arrow and Yandex ClickHouse. The Go standard library has adopted a version of our approach. ###### keywords: Parsing, Software Performance, IEEE-754, Floating-Point Numbers ## 1 Introduction Computers approximate real numbers as binary IEEE-754 floating-point numbers: an integer $m$ (the _significand_ 111The use of the term _mantissa_ is discouraged by IEEE [1, 2] and by Knuth [3].) multiplied by 2 raised to an integer exponent $p$: $m\times 2^{p}$. Most programming languages have a corresponding 64-bit data type and commodity processors provide the corresponding instructions. In several mainstream programming languages (C, C++, Swift, Rust, Julia, C#, Go), floating-point numbers adopt the 64-bit floating-point type by default. In JavaScript, all numbers are represented using a 64-bit floating-point type, including integers—except maybe for the large integer type BigInt. There are other number types beyond the standard binary IEEE-754 number types. For example, Gustafson [4] has proposed Unums types, Microsoft promotes its Microsoft Floating Point (MSFP) types [5] and many programming languages support decimal-number data types [6]. However, they are not as ubiquitous. Numbers are frequently serialized on disk or over a network as ASCII strings representing the value in decimal form (e.g., 3.1416, 1.0e10, 0.1). It is generally impossible to find a binary IEEE-754 floating-point number that matches exactly a decimal number. For example, the number 0.2 corresponding to $1/5$ can never be represented exactly as a binary floating-point number: its binary representation requires an infinite number of digits. Thus we must find the nearest available binary floating-point number. The nearest approximation to 0.2 using a standard 64-bit floating-point value is $$7\,205\,759\,403\,792\,794$\times 2^{-55}$ or approximately $0.200\,000\,000\,000\,000\,011\,10$. The second nearest floating-point value is $$7\,205\,759\,403\,792\,793$\times 2^{-55}$ or approximately $0.199\,999\,999\,999\,999\,983\,35$. In rare cases, the decimal value would be exactly between two floating-point values. In such cases, the convention is that we _round ties to even_ : of the two nearest floating-point values, we choose the one with an even _significand_. Thus, since $9\,000\,000\,000\,000\,000.5$ falls at equal distance from $$9\,000\,000\,000\,000\,000$\times 2^{0}$ and $$9\,000\,000\,000\,000\,001$\times 2^{0}$, we round it to $9\,000\,000\,000\,000\,000$. Meanwhile we round $9\,000\,000\,000\,000\,001.5$ and $9\,000\,000\,000\,000\,002.5$ to $9\,000\,000\,000\,000\,002$ and so forth. Finding the binary floating-point value that is closest to a decimal string can be computationally challenging. Widely used number parsers fail to reach $200\text{\,}\mathrm{MiB}\text{/}\mathrm{s}$ on commodity processors (see Fig. 2) whereas our disks and networks are capable of transmitting data at gigabytes per second (more than 5 times as fast). If we write a 64-bit floating-point number as a string using a decimal significand with 17 significant digits, we can always parse it back exactly: starting from the ASCII string representing the number with a correctly rounded 17-digit decimal significand and picking the nearest floating-point number, we retrieve our original number. Programmers and data analysts use no more than 17 digits in practice since there is no benefit to the superfluous digits if the number is originally represented as a standard binary floating- point number. There are only some exceptional cases where more digits could be expected: e.g., 1. 1. when the value has been entered by a human being, 2. 2. when the value was originally computed using higher-accuracy arithmetic, 3. 3. when the original value was in a different number type, 4. 4. when a system was poorly designed. We have that 64-bit unsigned integers can represent all 19-digit non-negative integers since $10^{19}<2^{64}$. Given an ASCII string (e.g., 2.2250738585072019e-308), we can parse the decimal significand as a 64-bit integer and represent the number as $22250738585072019\times 10^{-308-16}$. It remains to convert it to a binary floating-point number. In this instance, we must divide $22250738585072019$ by $10^{308+16}$ and round the result: we show that we can solve such problems using an efficient algorithm (§ 5). Further, when more than 19 digits are found, we may often be able to determine the nearest floating-point value from the most significant 19 digits (§ 11). term | notation ---|--- $m$ | binary significand; $m$ is a non-negative integer (often in $[2^{52},2^{53})$) $p$ | binary exponent; $p$ is an integer, it can be negative non-negative floating-point number | $m\times 2^{p}$ $w$ | decimal significand; $w$ is a non-negative integer (often in $[0,2^{64})$) $q$ | decimal exponent; $q$ is an integer, it can be negative non-negative decimal number | $w\times 10^{q}$ rounding | $\operatorname{round}(x)$ is an integer value nearest to $x$, with ties broken by rounding to the nearest even integer ceiling | given $x\geq 0$, $\operatorname{ceiling}(x)$ is the smallest integer no smaller than $x$ floor | given $x\geq 0$, $\operatorname{floor}(x)$ is the largest integer no larger than $x$ integer quotient | given non-negative integers $n,m$, $n\div m=\operatorname{floor}(n/m)$ integer remainder | $\operatorname{remainder}(n,m)=n-m\times\operatorname{floor}(n/m)$ trailing zeros | the positive integer $x$ has $k\in\mathbb{N}$ trailing zeros if and only if $x$ is divisible by $2^{k}$ Table 1: Notational conventions Our main contribution is to show that we can reach high parsing speeds (e.g., $1\text{\,}\mathrm{GiB}\text{/}\mathrm{s}$) on current 64-bit processors without sacrificing accuracy by focusing on optimizing the common number- parsing scenario where we have no more than 19 digits. We make all of our software freely available. ## 2 IEEE-754 Binary Floating-Point Numbers Many of the most popular programming languages and many of the most common processors support 64-bit and 32-bit IEEE-754 binary floating-point numbers. See Table 2. The IEEE-754 standard defines other binary floating-point types (binary16, binary128 and binary256) but they are less common. According to the IEEE-754 standard, a positive _normal_ double-precision floating-point number is a binary floating-point number where the 53-bit integer $m$ (the _significand_) is in the interval $[2^{52},2^{53})$ while being interpreted as a number in $[1,2)$ by virtually dividing it by $2^{52}$, and where the 11-bit exponent $p$ ranges from $-1022$ to $1023$ [7]. Such a double-precision number can represent all values between $2^{-1022}$ and up to but not including $2^{1024}$; these are the positive _normal_ values. Some values smaller than $2^{-1022}$ can be represented and are called _subnormal_ values: they use a special exponent code which has the value $2^{-1022}$ and the significand is then interpreted as a value in $[0,1)$. A sign bit distinguishes negative and positive values. A double-precision floating-point number uses 64 bits and is often called binary64. The binary64 format can represent exactly all decimal numbers made of a 15-digit significand from $\approx-1.8\times 10^{308}$ to $\approx 1.8\times 10^{308}$. Importantly, the reverse is not true: it is not sufficient to have 15 digits of precision to distinguish any two floating-point numbers: we may need up to 17 digits. The single-precision floating-point numbers are similar but span 32 bits (binary32). They are binary floating-point numbers where the 24-bit significand $m$ is in the interval $[2^{23},2^{24})$—considered as value in $[1,2)$ after virtually dividing it by $2^{23}$—and where the 8-bit exponent $p$ ranges from $-126$ to $127$. We can represent all numbers between $2^{-126}$ up to, but not including, $2^{128}$; with special handling for some numbers smaller than $2^{-126}$ (subnormals). The binary32 type can represent exactly all decimal numbers made of a 6-digit significand. If we serialize a 32-bit number using 9 digits, we can always parse it back exactly. name | exponent bits | significand (stored) | decimal digits (exact) ---|---|---|--- binary64 | 11 bits | 53 bits (52 bits) | 15 (17) binary32 | 8 bits | 24 bits (23 bits) | 6 (9) Table 2: Common IEEE-754 binary floating-point numbers: 64 bits (binary64) and 32 bits (binary32). A single bit is reserved for the sign in all cases. ## 3 Related Work Clinger [8, 9] describes accurate decimal to binary conversion; he proposes a fast path using the fact that small powers of 10 can be represented exactly as floats. Indeed, if we seek to convert the decimal number $1245\times 10^{14}$ to a binary floating-point number, we observe that the number $10^{14}$ can be represented exactly as a 64-bit floating-point number because $10^{14}=5^{14}\times 2^{14}$, and $5^{14}<2^{53}$. Of course, the significand ($1245$) can also be exactly represented. Thus if the value $10^{14}$ is precomputed as an exact floating-point value, it remains to compute the product of $1245\times 10^{14}$. The IEEE-754 specification requires that the result of an elementary arithmetic operation is correctly rounded.222Mainstream commodity processors (e.g., x64 and 64-bit ARM) have fast floating-point instructions with correct 64-bit and 32-bit rounding. In this manner, we can immediately convert a decimal number to a 64-bit floating- point number when it can be written as $w\times 10^{q}$ with $-22\leq q\leq 22$ and $w\leq 2^{53}$. We can extend Clinger’s fast approach to 32-bit floating-point numbers with the conditions $-10\leq q\leq 10$ and $w\leq 2^{24}$. Gay [10] improves upon Clinger’s work in many respects. He provides a secondary fast path, for slightly larger powers. Indeed if we need to compute $w\times 10^{q}$ for some small integer $w$ for some integer $q>22$, we may decompose the problem as $(w\times 10^{q-22})\times 10^{22}$. If $q$ is sufficiently small so that $w\times 10^{q-22}$ is less than $2^{53}$, then the computation is still exact, and thus $(w\times 10^{q-22})\times 10^{22}$ is also exact. Unfortunately, this approach is limited to decimal exponents $q\in(22,22+16)$ since $10^{16}>2^{53}$. Gay contributes a fast general decimal-to-binary implementation that is still in wide use: we benchmark against its most recent implementation in § 12. The general strategy for decimal-to-binary conversion involves first finding quickly a close approximation, that is within a few floating-point numbers of the accurate value, and then to refine it using one or two more steps involving exact big- integer arithmetic. Though there has been many practical attempts at optimizing number parsing [11], we are not aware of improved follow-up work to Gay’s approach in the scientific literature. A tangential problem is the conversion of binary floating-point numbers to decimal strings; the inverse of the problem that we are considering. The binary-to-decimal problem has received much attention [12, 13, 14, 15, 16, 17]. Among other problems is the one of representing a floating-point number using as few digits as possible so that the exact original value can be retrieved [13]: we need between 1 and 17 digits. ## 4 Parsing the String A floating-point value may be encoded in different manners as a string. For example, 1e+1, 10, 10.0, 10., 1.e1, +1e1, 10E-01 all represent the same value (10). There are different conventions and rules. For example, in JSON [18], the following strings are invalid numbers: +1, 01, 1.e1. Furthermore, there are locale-specific conventions. When parsing decimal numbers, our first step is to convert the string into a significand and an exponent. Though details differ depending on the requirement, the general strategy we propose is as follows: 1. 1. The string may be explicitly delimited—we have the end point or a string length—or we may use a sentinel such as the null character. The parser must not access characters outside the string range to avoid memory errors and security issues. 2. 2. It may be necessary to skip all leading white-space characters. In general, what constitutes a white-space character is locale-specific. 3. 3. The number may begin with the ‘+’ or the ‘-’ character. Some formats may disallow the leading ‘+’ character. 4. 4. The significand is a sequence of digits (0,…,9) containing optionally the decimal separator: the period character ‘.’ or a locale-specific equivalent. We must check that at least one digit was encountered: we thus forbid length- zero significands and significands made solely of the decimal separator. Some formats like JSON disallow a leading zero (only the zero value may begin with a zero) or an empty integer component (there must be digits before the decimal separator) or an empty fractional component (there must be digits after the decimal separator). To compute the significand, we may use a 64-bit unsigned integer $w$. We also record the beginning of the significand (e.g., as a pointer). We compute the digit value from the character using integer arithmetic: the digits have consecutive code point values in the most popular character encodings (Unicode and ASCII define values from 48 to 57). We can similarly use the fact that the digits occupy consecutive code points to quickly check whether a character is a digit. With each digit encountered, we can compute the running significand with a multiplication by ten followed by an addition ($w=10\times d+v$ where $v$ is the digit value). The multiplication by ten is often optimized by compilers into efficient sequences of instructions. If there are too many digits, the significand may overflow which we can guard against by counting the number of processed digits: it is not necessary to guard each addition and multiplication against overflows. We can either be optimistic and later check whether an overflow was possible, or else we may check our position in the string, making sure that we never parse more than 19 digits, after omitting leading zeros. When the decimal separator is encountered, we record its position, but we otherwise continue computing the running significand. It is common to encounter many digits after the decimal separator. Instead of processing the digits one by one, we may check all at once whether a sequence of 8 digits is available and then update a single time the running significand—using a technique called SIMD within a register (SWAR) [19]. See Appendix D. If we find a sequence of eight digits, it can be beneficial to check again whether eight more digits (for a total of 16) can be found. 5. 5. If there is a decimal separator, we must record the number of fractional digits, which we compute from the position of the decimal separator and the end of the significand. E.g., if there are 12 digits after the decimal separator, then the exponent is $10^{-12}$. If there is no decimal separator, then the exponent is implicitly zero ($10^{0}$). 6. 6. The significand may be followed by the letter ‘e’ or the letter ‘E’ in which case we need to parse the exponent if the scientific notation is allowed. Conversely, if the scientific notation is prescribed, we might fail if the exponent character is not detected. The parsing of the exponent proceeds much like the parsing of the significand except that no decimal separator is allowed. An exceptional condition may occur if the exponent character is not followed by digits, accounting for the possible ‘+’ and ‘-’ characters. We may either fail, if the scientific notation is required, or we may decide to truncate the string right before the exponent character. To avoid overflow with the exponent, we may update it only if its absolute value is under some threshold: it makes no difference whether the exponent is $-1000$ or $-10\,000$; whether it is $1000$ or $10\,000$. The explicit exponent must be added to the exponent computed from the decimal separator. 7. 7. If the number of digits used to express the significand is less than 19, then we know that the significand cannot overflow. If the number of digits is more than 19, we may count the significant digits by omitting leading zeros (e.g., $0.000\,123$ has only three significant digits). Finally, if an overflow cannot be dismissed, we may need to parse using a higher-precision code path (§ 11). There are instances when we can quickly terminate the computation after decoding the decimal significand and its exponent. If the significand is zero or the exponent is very small then the number must be zero. If the significand is non-zero but the exponent is very large then we have an infinite value ($\pm\infty$). ## 5 Fast Algorithm A fast algorithm to parse floating-point numbers might start by processing the ASCII string (see § 4) to find a decimal significand and a decimal exponent. If the number of digits in the significand is less than 19, then our approach is applicable. (see § 11). However, before we apply our algorithm, we use Clinger’s fast path [8, 9], see § 3. Even though it adds an additional branch at the beginning, it is an inexpensive code path when it is applicable, implying a single floating-point multiplication or division. We can implement it efficiently. We check whether the decimal power is within the allowable interval ($q\in[-22,22]$ in the 64-bit case, $q\in[-10,10]$ in the 32-bit case) and whether the absolute value of the decimal significand is in the allowable interval ($[0,2^{53}]$ in the 64-bit case or $[0,2^{24}]$ in the 32-bit case). When these conditions are encountered, we losslessly convert the decimal significand to a floating-point value, we lookup the precomputed power of ten $10^{|q|}$ and we multiply (when $q\geq 0$) or divide ($q<0$) the converted significand.333The case with negative exponents where a division is needed requires some care on systems where the division of two floating-point numbers is not guaranteed to round the nearest floating-point value: when such a system is detected, we may limit the fast path to positive decimal powers. Gay [10] proposes an extended fast path that covers a broader range of decimal exponents, but with more stringent conditions on the significand. We do not make use of this secondary fast path. It adds additional branching and complexity for relatively little gain in our context. In particular, Clinger’s fast path covers all integer values in $[0,2^{53}]$ (64-bit case). We could also add an additional fast path specifically for integers. We can readily identify such cases because the decimal exponent is zero: $w\times 10^{0}$. We can rely on the fact that the IEEE standard specifies that conversion between integer and floating-point be correctly rounded [7]. Thus a cast from an integer value to a floating-point value is often all that is needed. It may often require nearly just a single instruction (e.g., cvtsi2sd under x64 and ucvtf under ARM). However, we choose to disregard this potential optimization because the gains are modest while it increases the complexity of the code. We must then handle the general case, after the application of Clinger’s fast path. We formalize our approach with Algorithm 1. This concise algorithm can handle rounding, including ties to even, subnormal numbers and infinite values. We specialize the code for positive numbers, but negative numbers are handled by flipping the sign bit in the result. As the pseudo-code suggests, it can be implemented in a few lines of code. The algorithm always succeeds unless we have large or small decimal exponents ($q\notin[-27,55]$) in which case we may need to fall back on a higher-precision approach in uncommon instances. The algorithm relies on a precomputed table of 128-bit values $T[q]$ for decimal exponents $q\in[-342,308]$ (see Appendix B). * • In lines 3 and 4, we check for very large or very small decimal exponents as well as for zero decimal significands. In such cases, the result is always either zero or infinity. * • In lines 5 and 6, we normalize the decimal significand $w$ by shifting it so that $w\in[2^{63},2^{64})$. * • We must convert the decimal significand $w$ into the binary significand $m$. We have that $w\times 10^{q}=w\times 5^{q}\times 2^{q}\approx m\times 2^{p}$ so we must estimate $w\times 5^{q}$. At line 7, we multiply the normalized significand $w$ by the 128-bit value $T[q]$ using one or two 64-bit multiplications. Intuitively, the product $w\times T[q]$ approximates $w\times 5^{q}$ after shifting the result. We describe this step in § 7 and § 8 for positive decimal exponents ($q\geq 0$), and in § 9 for negative decimal exponents. We have a 128-bit result $z$. * • At line 8, we check for failure, requiring the software to fall back on a higher-precision approach. It corresponds to the case where we failed to provably approximate $w\times 5^{q}$ to a sufficient degree. In practice, it is unlikely and only ever possible if $q\notin[-27,55]$. * • At line 9, we compute the expected binary significand with one extra bit of precision (for rounding) from the product $z$. * • At lines 10 and 11, we compute the expected binary exponent. We justify this step in § 10. * • At line 12, we check whether the binary exponent is too small. When it is too small, the result is zero. * • At line 13, we check whether we have a subnormal value when the binary exponent is too small. See § 9.3. * • At line 18, we handle the case where we might have a value that is exactly between two binary floating-point numbers. We describe this step generally in § 6 where we show that subnormal values cannot require rounding ties. We describe it specifically in § 8.1 for the positive-exponent case ($q\geq 0$) and in § 9.1 in the negative-exponent case. Intuitively, we identify ties when the product $z$ from which we extracted the binary significand $m$ contains many trailing zeroes after ignoring the least significant bit. We need to be concerned when we would (at line 21) round up from an even value: we adjust the value to prevent rounding up. * • At line 21, we round the binary significand. At line 22, we handle the case where rounding up caused an overflow, in which case we need to increment the binary exponent. At line 23, we handle the case where the binary exponent is too large and we have an infinite value. We show that the algorithm is correct by examining each step in the following sections. We assess our algorithm experimentally in § 12. 1:an integer $w\in[0,10^{19}]$ and an integer exponent $q$ 2:a table $T$ containing 128-bit reciprocals and truncated powers of five for all powers from $-342$ to $308$ (see Appendix B) 3:if $w=0$ or $q<-342$ then Return 0 end if 4:if $q>308$ then Return $\infty$ end if 5:$l\leftarrow$ the number of leading zeros of $w$ as a 64-bit (unsigned) word 6:$w\leftarrow 2^{l}\times w$ $\triangleright$ Normalize the decimal significand 7: Compute the 128-bit truncated product $z\leftarrow(T[q]\times w)\div 2^{64}$, stopping after one 64-bit multiplication if the most significant 55 bits (64-bit) or 26 bits (32-bit) are provably exact. 8:if $z\bmod 2^{64}=2^{64}-1$ and $q\notin[-27,55]$ then Abort end if 9:$m\leftarrow$ the most significant 54 bits (64-bit) or 25 bits (32-bit) of the product $z$, not counting the eventual leading zero bit 10:$u\leftarrow z\div 2^{127}$ value of the most significant bit of $z$ 11:$p\leftarrow((217706\times q)\div 2^{16})+63-l+u$ $\triangleright$ Expected binary exponent 12:if $p\leq-1022-64$ (64-bit) or $p\leq-126-64$ (32-bit) then Return 0 end if 13:if $p\leq-1022$ (64-bit) or $p\leq-126$ (32-bit) then$\triangleright$ Subnormals 14: $s\leftarrow-1022-p+1$ (64-bit) or $s\leftarrow-126-p+1$ (32-bit) 15: $m\leftarrow m\div 2^{s}$ and $m\leftarrow m+1$ if $m$ is odd (round up), and $m\leftarrow m\div 2$ 16: Return $m\times 2^{p}\times 2^{-52}$ (64-bit) or $m\times 2^{p}\times 2^{-23}$ (32-bit case) 17:end if 18:if $z\bmod 2^{64}\leq 1$ and $m$ is odd and $m\div 2$ is even and ($q\in[-4,23]$ (64-bit) or $q\in[-17,10]$ (32-bit)) then $\triangleright$ Round ties to even 19: if $(z\div 2^{64})/m$ is a power of two then $m\leftarrow m-1$ $\triangleright$ Will not round up 20:end if 21: $m\leftarrow m+1$ if $m$ is odd; followed by $m\leftarrow m\div 2$ $\triangleright$ Round the binary signficand 22:if $m=2^{54}$ (64-bit) or $m=2^{25}$ (32-bit) then $m=m\div 2$; $p\leftarrow p+1$ end if 23:if $p>1023$ (64-bit) or $p>127$ (32-bit) then Return $\infty$ end if 24:Return $m\times 2^{p}\times 2^{-52}$ (64-bit) or $m\times 2^{p}\times 2^{-23}$ (32-bit case) Algorithm 1 Algorithm to compute the binary floating-point number nearest to a decimal floating-point number $w\times 10^{q}$. We give just one algorithm for both the 32-bit and 64-bit cases. For negative integers, we need to negate the result. ## 6 Exact Numbers and Ties We seek to approximate a decimal floating-point number of the form $w\times 10^{q}$ using a binary decimal floating-point number of the form $m\times 2^{p}$. Sometimes, there is no need to approximate since an exact representation is possible. That is, we have that $w\times 10^{q}=m\times 2^{p}$ or, equivalently, $w\times 5^{q}\times 2^{q}=m\times 2^{p}$. In our context, we refer to these numbers as _exact numbers_. We seek to better identify when they can occur. * • When $q\geq 0$, we have $m=w\times 5^{q}\times 2^{q}\times 2^{-p}$ so that $m$ is divisible by $5^{q}$. In the 64-bit case, we have that $m<2^{53}$; and in the 32-bit case, we have that $m<2^{24}$. Thus we have, respectively, $5^{q}<2^{53}$ and $5^{q}<2^{24}$. These inequalities become $q\leq 22$ and $q\leq 10$. For example, we have that $1\times 10^{22}$ is an exact 64-bit number while $1\times 10^{23}$ is not. * • When $q<0$, we have that $w=5^{-q}\times 2^{-q}\times 2^{p}\times m$. We have that $5^{-q}$ divides $w$. If we assume that $w<2^{64}$ then we have that $5^{-q}<2^{64}$ or $q\geq-27$. For example, the number $7450580596923828125\times 10^{-27}$ is the smallest exact 64-bit number. It follows that no exact number is sufficiently small to qualify as a subnormal value: the largest subnormal number has a small decimal power (e.g., $\approx 10^{-38}$ in the 32-bit case). Thus we have that exact numbers must be of the form $w\times 10^{q}$ with $q\in[-27,22]$ (64-bit case) or $q\in[-27,10]$ (32-bit case) subject to the constraint that the decimal significand can be stored in a 64-bit value. Yet floating-point numbers occupy a much wider range (e.g., from $4.9\times 10^{-324}$ to $1.8\times 10^{308}$). In other words, exact numbers are only possible when the decimal exponent is near zero. To find the nearest floating-point number when parsing, it is almost always sufficient to round to the nearest value without an exact computation. However, when the number we are parsing might fall exactly between two numbers, more care is needed. The IEEE-754 standard recommends that we round to even. We may need an exact computation to apply the round-ties-to-even strategy.444We focus solely on rounding ties to even, as it is ubiquitous. However, our approach could be extended to other rounding modes. The sign can be ignored when rounding ties to even: if a value is exactly between the two nearest floating-point numbers and they have different signs, then the midpoint value must be zero, by symmetry. It may seem that we could generate many cases where we fall exactly between two floating-point numbers. Indeed, it suffices to take any floating-point number that is not the largest one, and then take the next largest floating- point number. From these two numbers, we pick a number that is right in- between and we have a number that requires rounding to even. However, for such a half-way number to be a concern to us, it must be represented exactly in decimal form using a small number of decimal digits. In particular, the decimal significand must be divisible by $5^{-q}$ and yet must be no larger than $2^{64}$. It implies that the decimal exponent cannot be too small ($q\geq-27$). It also implies that the nearby binary floating-point values are normal numbers. Let us formalize the analysis. A mid-point between two floating-point numbers, $m\times 2^{p}$ and $(m+1)\times 2^{p}$, can be written as $(2m+1)\times 2^{p-1}$. Assume that both numbers $m\times 2^{p}$ and $(m+1)\times 2^{p}$ can be represented exactly using a standard floating-point type: for 64-bit floating point numbers, it implies that $m+1<2^{53}$; for 32-bit numbers it implies $m+1<2^{24}$. We only need rounding if $(2m+1)\times 2^{p-1}$ cannot be represented. There are two reasons that might explain why a number cannot be represented. Either it requires a power of two that is too small or too large, or else its significand requires too many bits. Because the value is exactly between two numbers that can be represented, we know that it is not outside the bounds of the power of two. Thus the significand must require too many bits. Furthermore, both $m\times 2^{p}$ and $(m+1)\times 2^{p}$ must be normal numbers. Hence the fact that $2m+1$ has too many bits implies that $2m+1\in(2^{53},2^{54}]$ for 64-bit floating point numbers and that $2m+1\in(2^{24},2^{25}]$ for 32-bit numbers. * • When $q\geq 0$, we have that $5^{q}\leq 2m+1$. In the 64-bit case, we have $5^{q}\leq 2m+1\leq 2^{54}$ or $q\leq 23$. In the 32-bit case, we have $5^{q}\leq 2m+1\leq 2^{25}$ or $q\leq 10$. * • When $q<0$, we have $w\geq(2m+1)\times 5^{-q}$. We must have that $w<2^{64}$ so $(2m+1)\times 5^{-q}<2^{64}$. We have that $2m+1>2^{53}$ (64-bit case) or $2m+1>2^{24}$ (32-bit case). Hence, we must have $2^{53}\times 5^{-q}<2^{64}$ (64-bit) and $2^{24}\times 5^{-q}<2^{64}$ (32-bit). Hence we have $5^{-q}<2^{11}$ or $q\geq-4$ (64-bit case) and $5^{-q}<2^{40}$ or $q\geq-17$ (32-bit case). Thus we have that we only need to round ties to even when we have that $q\in[-4,23]$ (in the 64-bit case) or $q\in[-17,10]$ (in the 32-bit case). In both cases, the power of five ($5^{|q|}$) fits in a 64-bit word. ## 7 Most Significant Bits of a Product When converting decimal values to binary values, we may need to multiply or divide by large powers of ten (e.g., $10^{300}=5^{300}2^{300}$). Mainstream processors compute the 128-bit product of two 64-bit integers using one or two fast instructions: e.g., with the single instruction imul (x64 processors) or two instructions umulh and mul (aarch64 processors). However, we cannot represent an integer like $5^{300}$ using a single 64-bit integer. We may represent such large integers using multiple 64-bit words, henceforth a _multiword integer_. We may compute the product between two multiword integers starting from the least significant bits. Thus if we are multiplying an integer that requires a single machine word $w$ with an integer that requires $n$ machine words, we can use $n$ 64-bit multiplications starting with a multiplication between the word $w$ and the least significant word of the other integer, going up to the most significant words. See Algorithm 2. 1:an integer $w\in(0,2^{64})$ 2:a positive integer $b$ represented as $n$ words ($n>0$) $b_{0},b_{1},\ldots,b_{n-1}$ such that $b=\sum_{i=0}^{n-1}b_{i}2^{64i}$. 3:Allocate $n+1$ words $u_{0},u_{1},\ldots,u_{n}$ 4:$p\leftarrow w\times b_{0}$ 5:$u_{0}\leftarrow p\bmod 2^{64}$ 6:$r\leftarrow p\div 2^{64}$ 7:for $i=1,\ldots,n-1$ do 8: $p\leftarrow w\times b_{i}$ $\triangleright$ $p\leq(2^{64}-1)^{2}$ 9: $p\leftarrow p+r$ $\triangleright$ $p\leq 2^{128}-2^{64}+1$ 10: $u_{i}\leftarrow p\bmod 2^{64}$ 11: $r\leftarrow p\div 2^{64}$ 12:end for 13:$u_{n}\leftarrow r$ 14:Return: The result of the multiplication as an $n+1$-word $u$ such that $u=\sum_{i=0}^{n}u_{i}2^{64i}$. Algorithm 2 Conventional algorithm to compute the product of a single-word integer and a multiple-word integer. Such a conventional algorithm is inefficient when we only need to approximate the product. For example, maybe we only want the most significant word of the product and we would like to do the computation using only one or two multiplications. Thankfully it is often possible in practice. Such partial multiplications are sometimes called _truncated_ multiplications [20] and the result is sometimes called a _short_ product [21, 22, 23]. 1:an integer $w\in(0,2^{64})$ 2:a positive integer $b$ represented as $n$ words ($n>0$) $b_{0},b_{1},\ldots,b_{n-1}$ such that $b=\sum_{i=0}^{n-1}b_{i}2^{64i}$. 3:a desired number of exact words $m\in(0,n+1]$ 4:Allocate $n+1$ words $u_{0},u_{1},\ldots,u_{n}$ 5:$p\leftarrow w\times b_{n-1}$ 6:$u_{n-1}\leftarrow p\bmod 2^{64}$ 7:$u_{n}\leftarrow p\div 2^{64}$ 8:if $m=1$ and $u_{n-1}<2^{64}-w$ then 9: Return:$u_{n}$ $\triangleright$ Stopping condition 10:end if 11:for $i=n-2,n-1,\ldots,0$ do 12: $p\leftarrow w\times b_{i}$ 13: $u_{i}\leftarrow p\bmod 2^{64}$ 14: if $u_{i+1}+(p\div 2^{64})\geq 2^{64}$ then 15: add 1 to $u_{i+2}$, if it exceeds $2^{64}$, set it to zero and add 1 to $u_{i+3}$ and so forth up to $u_{n}$ potentially 16: end if 17: $u_{i+1}\leftarrow(u_{i+1}+(p\div 2^{64}))\bmod 2^{64}$ 18: if $m\leq n-i$ and $u_{i}<2^{64}-w$ then 19: Return:$u_{u-m+1},\ldots,u_{n}$ $\triangleright$ Stopping condition 20: end if 21: if $m<n-i$ and $u_{i}<2^{64}-1$ then 22: Return:$u_{n-m+1},\ldots,u_{n}$ $\triangleright$ Stopping condition 23: end if 24:end for 25:Return:$u_{n-m+1},\ldots,u_{n}$. Algorithm 3 Algorithm to compute the $m$ most significant words of the product of a single-word integer and a multiple-word integer. Suppose that we have computed the product of the single-word integer ($w$) with the $k$ most significant words of the multiword integer: we have computed the $k+1$ words of the product $w\times(\sum_{i=n-k}^{n-1}b_{i}2^{64i})\div 2^{64(n-k)}$. Compared with the $k+1$ most significant words of the full product $(w\times(\sum_{i=0}^{n-1}b_{i}2^{64i}))\div 2^{64(n-k)}$, we are possibly underestimating because we omit the contribution of the product between the word $w$ and the least $n-k$ significant words. These less $n-k$ significant words have maximal value $2^{64(n-k)}-1$. Their product with the word $w$ is thus at most $2^{64(n-k)}w-w$ and their contribution to the most significant words is at most $(2^{64(n-k)}w-w)\div 2^{64(n-k)}=w-1$. Hence if the least significant computed word that is no larger than $2^{64}-w+1$, then all computed words are exact except maybe for that least significant one. Our short product matches the full product. Thus we have _stopping condition_ : if we only want the $k$ most significant words of the product, we can compute the $k+1$ most significant words from the $k$ most significant words of the multiword integers, and stop if the least significant word of the product is no larger than $2^{64}-w+1$. This stopping condition is most useful if $w$ is small ($w\ll 2^{64}$). We also have another stopping condition that is more generally useful. Even if the least significant word is larger than $2^{64}-w+1$, then the second least significant word needs to be incremented by one in the worst case. If the second least significant word is not $2^{64}-1$, then all other more significant words are exact. That is, we have $k-2$ exact most significant words if the second last of our $k+1$ most significant words is not $2^{64}-1$. By combining these two conditions, we rarely have to compute more than 3 words using two multiplications to get the exact value of the most significant word. Algorithm 3 presents a general algorithm. We can stop maybe even earlier if we need even less than the most significant word, say $t$ bits. Indeed, unless all of the less significant bits in the computed most significant word have value 1, then an overflow (+1) does not affect the most significant bits of the most significant word. Thus the condition $u_{n-1}<2^{64}-w$ (line 8 in Algorithm 3) can be replaced by ($u_{n-1}<2^{64}-w$ or $u_{n}\bmod 2^{64-t}\not=2^{64-t}-1$) if we only need $t$ exact bits of the product. ## 8 Multiplication by Positive Powers of Five When parsing a decimal number, we follow the general strategy of first identifying the non-negative decimal significand $w$ and its corresponding exponent $q$. We then seek to convert $w\times 10^{q}$ to a binary floating- point number. For example, given the mass of the Earth in kilograms as $5.972\times 10^{24}$, we might parse it first as $5972\times 10^{27}$. Our goal is to represent it as a nearest binary floating-point number such as $5561858415603638\times 2^{30}$. The sign bit is handled separately. The largest integer we can represent with a 64-bit floating-point number is $\approx 1.8\times 10^{308}$. Thus, when processing numbers of the form $w\times 10^{q}$ for non-negative powers of $q$, we only have to worry about $q\in[0,308]$. If any larger value of $q$ is found and the decimal significand is non-zero ($w>0$), the result is an infinity value (either $+\infty$ or $-\infty$). Figure 1: Number of 64-bit words necessary to represent $5^{q}$ exactly for $q\in[0,308]$. Our first step is to expand the exponent: $w\times 10^{q}=w\times 2^{q}\times 5^{q}$. Thus we seek to compute $w\times 5^{q}$. Though the number $5^{308}$ may appear large, it only requires twelve 64-bit words since $5^{308}<2^{64\times 12}$. Storing all of these words require about $15\text{\,}\mathrm{KiB}\text{/}$: we need between 1 and 12 words per power (see Fig. 1). Thus it could be practical to memoize all of the exponents $5^{q}$ for $q=1,\ldots,308$. In § 7, we compute the most significant word of a product with few words of the multiword integer: it follows that even if we need, in the worst case $15\text{\,}\mathrm{KiB}\text{/}$ of storage, an actual implementation may touch only a fraction of that memory. Thus it may be more efficient to only store two 64-bit words per power, as long as we can fall back on a higher-precision approach. We then use only $5\text{\,}\mathrm{KiB}\text{/}$ of storage (three times less). Effectively, given a large power of five, we store 128 bits of precision. We truncate the result, discarding its less significant words. It could be that the most significant word of the product $w\times 5^{q}$ contains many leading zeros. Yet we desire a normalized result, without leading zeros for efficiency and simplicity reasons. * • We store the powers of five in a format where the most significant bit of the most significant word is 1: hence, we shift the power of five adequately: e.g., instead of storing $5^{2}$, we store $5^{2}\times 2^{59}$. See Table 3. * • Further, we shift $w$ by an adequate power of two so that the 63${}^{\textrm{rd}}$ bit has value 1. It is always possible as $w$ is non- zero; the case when $w=0$ is handled separately. Normalizing numbers such that they have no leading zeros is inexpensive: the number of leading zeros can be computed using a single instruction on popular processors (clz on ARM, lzcnt on x64). $q$ | first word | $q$ | first word | second word ---|---|---|---|--- 0 | 8000000000000000 | 28 | 813f3978f8940984 | 4000000000000000 1 | a000000000000000 | 29 | a18f07d736b90be5 | 5000000000000000 2 | c800000000000000 | 30 | c9f2c9cd04674ede | a400000000000000 3 | fa00000000000000 | 31 | fc6f7c4045812296 | 4d00000000000000 4 | 9c40000000000000 | 32 | 9dc5ada82b70b59d | f020000000000000 5 | c350000000000000 | 33 | c5371912364ce305 | 6c28000000000000 6 | f424000000000000 | 34 | f684df56c3e01bc6 | c732000000000000 7 | 9896800000000000 | 35 | 9a130b963a6c115c | 3c7f400000000000 8 | bebc200000000000 | 36 | c097ce7bc90715b3 | 4b9f100000000000 9 | ee6b280000000000 | 37 | f0bdc21abb48db20 | 1e86d40000000000 10 | 9502f90000000000 | 38 | 96769950b50d88f4 | 1314448000000000 11 | ba43b74000000000 | 39 | bc143fa4e250eb31 | 17d955a000000000 12 | e8d4a51000000000 | 40 | eb194f8e1ae525fd | 5dcfab0800000000 13 | 9184e72a00000000 | 41 | 92efd1b8d0cf37be | 5aa1cae500000000 14 | b5e620f480000000 | 42 | b7abc627050305ad | f14a3d9e40000000 15 | e35fa931a0000000 | 43 | e596b7b0c643c719 | 6d9ccd05d0000000 16 | 8e1bc9bf04000000 | 44 | 8f7e32ce7bea5c6f | e4820023a2000000 17 | b1a2bc2ec5000000 | 45 | b35dbf821ae4f38b | dda2802c8a800000 18 | de0b6b3a76400000 | 46 | e0352f62a19e306e | d50b2037ad200000 19 | 8ac7230489e80000 | 47 | 8c213d9da502de45 | 4526f422cc340000 20 | ad78ebc5ac620000 | 48 | af298d050e4395d6 | 9670b12b7f410000 21 | d8d726b7177a8000 | 49 | daf3f04651d47b4c | 3c0cdd765f114000 22 | 878678326eac9000 | 50 | 88d8762bf324cd0f | a5880a69fb6ac800 23 | a968163f0a57b400 | 51 | ab0e93b6efee0053 | 8eea0d047a457a00 24 | d3c21bcecceda100 | 52 | d5d238a4abe98068 | 72a4904598d6d880 25 | 84595161401484a0 | 53 | 85a36366eb71f041 | 47a6da2b7f864750 26 | a56fa5b99019a5c8 | 54 | a70c3c40a64e6c51 | 999090b65f67d924 27 | cecb8f27f4200f3a | 55 | d0cf4b50cfe20765 | fff4b4e3f741cf6d Table 3: Most significant bits in hexadecimal form of the powers $5^{q}$ for some exponents $q$. The values are normalized by multiplication with a power of two so that the most significant bit is always 1. For each power, the first word represent the most significant 64 bits, the second word the next most significant 64 bits. For $q\leq 27$, the second word is made of zeros and omitted. In practice, we need to materialize this table for up to $q=308$ to cover all of the relevant powers of five for 64-bit numbers. Hence, because of our normalization, the computed product is at least as large as $2^{63}\times 2^{63}=2^{126}$. That is, as a 128-bit value, it has at most one leading 0. We can check for this case and shift the result by one bit accordingly. To compute the product of the normalized decimal significand with normalized words representing the power of five (see in Table 3), we use up to two multiplications. * • When $5^{q}<2^{64}$, then a single multiplication always provide an exact result. In particular, it implies that whenever we need to round ties to even, we have the exact product (see § 8.1). * • We can always do just one multiplication as long as it provides the number of significant bits of the floating-point standard (53 bits for 64-bit numbers and 24 bits for 32-bit numbers) plus one extra bit to determine the rounding direction, and one more bit to handle the case where the computed product has a leading zero. It suffices to check the least significant bits of the most significant 64 bits and verify that there is at least one non-zero bit out of 9 bits for 64-bit numbers and out of 38 bits for 32-bit numbers. * • When that is not the case, we execute a second multiplication. This second multiplication is always sufficient to compute the product exactly as long as $q\leq 55$ since $5^{55}<2^{128}$. We have that the largest 32-bit floating- point number is $\approx 3.4\times 10^{38}$ so that exponents greater than $55$ are irrelevant in the sense that they result in infinite values as long as the significand is non-zero. However, 64-bit floating-point numbers can be larger. For larger values of $q$, we have that the most significant 64 bits of the truncated product are exact when the second most significant word is not filled with 1-bits ($2^{64}-1$). When it is filled with ones then the computation of a third (or subsequent) multiplication could add one to this second word which would overflow into the most significant word, adding a value of one again. It could maybe result into a value that was seemingly just slightly under the midpoint between two floating-point values (and would thus be rounded down) to appear to switch to just over the midpoint between two floating-point values (and to be thus rounded up). In this unlikely case, when the least significant 64 bits of the most significant 128 bits of the computed product are all ones, we choose to fall back on a higher-precision approach. ###### Technical Remark 1 To be more precise, before we fall back, we could check that out of the most significant 128 bits, all but the leading 55 bits (64-bit case) or leading 26 bits are ones, instead of merely checking the least significant 64 bits. However, we can show that checking that the least significant 64 bits of the truncated 128-bit product are all ones is sufficient. We sketch the proof. Before the computation of the second product, the most significant 64 bits of the product have trailing ones (e.g., 9 bits for 64-bit numbers), otherwise we would not have needed a second multiplication. If the second multiplication does not overflow into the most significant 64 bits, then the final result still has trailing ones in its most significant 64 bits. Suppose it is not the case: there was an overflow in the most significant 64 bits following the second multiplication. When the second product overflows into the most significant 64 bits (turning these ones into zeros), the second 64-bit word of the product is at most $2\times(2^{64}-1)\bmod 2^{64}=2^{64}-2$. Thus when the second most significant 64 bits are all ones then there was no overflow into the most significant 64-bit word by the second multiplication and the least significant bits of the first 64-bit word of the product are also ones. Once we have determined that we have sufficient accuracy, we either need to check for round-to-even cases ($q\leq 23$), see § 8.1, or else we proceed with the general case. In the general case, we consider the most significant 54 bits (64-bit numbers) or 24 bits (32-bit numbers), not counting the possible leading zero, and then we round up or down based on the least significant bit. E.g., we round down to 53 bits in the 64-bit cases when the least significant bit out of 54 bits is zero, otherwise we round up. If all of the bits are ones, rounding up overflows into a more significant bit and we shift the result by one bit so we have the desired number of bits (53 bits or 24 bits). In § 10, we show how to compute the binary exponent efficiently. Whenever we find that the resulting binary floating-point number $m\times 2^{p}$ is too large, we return the infinite value. The upper limits are $2^{1024}$ ($\approx 1.7976934\times 10^{308}$) for 64-bit numbers and $2^{128}$ ($3.4028236\times 10^{38}$) for 32-bit numbers. The infinite values are signed ($+\infty,-\infty$). ###### Example 1 Consider the string 2440254496e57. We get $w=2440254496$ and $q=57$. We normalize the decimal significand $w$ by shifting it by 32 bits so that the most significant bit is set (considering $w$ as a 64-bit word). We look up the 64 most significant bits of $5^{57}$ which are given by the word value 11754943508222875079 (or $5^{57}\div 2^{133-64}$). The most significant 64 bits of the product are, in hexadecimal notation, 0x5cafb867790ea3ff. All of the least significant 9 bits are set. Thus we need to execute a second multiplication. In practice, it is a rare instance. We load the next most significant 64 bits of the product (0xaff72d52192b6a0d) and compute the second product. After updating our product, we get that the most significant word is 0x5cafb867790ea400 whereas the second most significant word is 0x1974b67f5e78017. The most significant 55 bits of the product have been computed exactly. The most significant bit of 0x5cafb867790ea400 is zero. We select the most significant 55 bits of the word: 13044452201105234. It is an even value so we do not round up. Finally, we shift the result to get the binary significand $m=6522226100552617\in[2^{52},2^{53})$. ### 8.1 Round Ties to Even with Positive Powers To handle rounding ties to even when the decimal power is positive ($q\geq 0$), we only have to be concerned when the exponent $q$ is sufficient small: $q\leq 23$ assuming that we want to support both 64-bit numbers (which require $q\leq 23$) and 32-bit numbers (which require $q\leq 10$) due to the fact that we require the decimal significand to be smaller than $2^{64}$ (see § 6). In such cases, the product corresponding to $m\times 5^{q}$ is always exact after one multiplication (since $5^{q}<2^{64}$). We need to round-to-even when the resulting binary significand uses one extra significant bit for a total of 54 bits in the 64-bit case and 25 bits in the 32-bit case. We can check for the occurrence of a round-to-even case as follows: * • The most significant 64-bit word has exactly 10 trailing zeros (64-bit case) or exactly 39 trailing zeros (32-bit case).555We assume that the most significant word of the product contains a leading 1-bit, maybe following a shift. If the word has a leading zero, we must subtract one from these numbers: 9 for the 64-bit case and 38 for the 32-bit case. * • The second 64-bit word of the product containing the least significant bits equals zero. In such cases, we get a 54-bit (64-bit case) or a 25-bit (32-bit case) word that ends with a 1-bit. We need to round to a 53-bit or 24-bit value. We can either round up or round down. To implement the round to even, we check the second last significant bit. If it is zero, we round down; if it is one, we round up. When rounding up, if we overflow because we only have 1-bits, we shift the result to get the desired number of bits (53 bits or 24 bits). When the round-to-even case is not encountered, we either round up or down as in the general case. Given that we have computed the exact product ($q\leq 23$), there can be no ambiguity. ###### Technical Remark 2 When $q\leq 23$, the second most significant 64-bit word must have some trailing zeros, whether we are in a round-to-even case or not. Indeed, because we always have $5^{q}\leq 2^{54}$ and because we store the powers of five shifted so that they have a leading 1-bit in their respective 64-bit words (e.g., $5^{23}$ is loaded as $5^{23}\times 2^{10}$), we have at least 10 trailing zeros, irrespective of the word $w$ as long as $q\leq 23$. Hence we can replace a comparison with zero with a check that it is no larger than a small value like 1, if it is convenient. It turns out to be useful to us for simplifying Algorithm 1, see line 18. ###### Example 2 Consider the input string 9007199254740993. It is equal to $2^{53}+1$, a number that cannot be represented as a standard 64-bit floating-point number. After parsing the string, we get $w=9007199254740993$ and $q=0$. We normalize $w$ to 9223372036854776832 ($2^{63}+2^{10}$). We multiply this shifted $w$ by $5^{q}$ which, in this case, is just 1 ($5^{q}=1$). The precomputed $5^{q}$ has been normalized to the 64-bit value $2^{63}$. Hence we get the product $2^{63}w$. We then select the most significant 54 bits of the result after skipping the leading zero bit ($2^{53}+1$). Because this result has its least significant bit set, and because all the bits that were not selected are zero, we know that we need to round to even. The second least significant bit is zero, and we know that we need to round down. We therefore end up with a significand of one and an exponent of $2^{53}$. ## 9 Division by Powers of Five We turn our attention to negative decimal exponents ($q<0$). Consider the decimal number $9.109\times 10^{-31}$ corresponding to the mass of the electron. We can write it as $9109\times 10^{-34}$ or $9109\times 2^{-34}\times 5^{-34}$. Though we can represent the integer $5^{34}$ using few binary words, the value $5^{-34}$ could only be approximated in binary form. E.g., it is approximately equal to $4676805239458889\times 2^{-131}$. To convert it to a binary floating-point number, we must find $m$ and $p$ such that $9109\times 2^{-34}\times 5^{-34}\approx m\times 2^{p}$. Consider the general case, replacing $9109\times 2^{-34}$ by $w\times 10^{q}$ where $0<w<2^{64}$ and $q<0$. We need to approximate the decimal floating- point number $w\times 10^{q}$ as closely as possible with the binary floating- point number $m\times 2^{p}$. We formalize $m\times 2^{p}\approx w\times 10^{q}$ as the equation $m\times 2^{p}+\epsilon 2^{p}=w\times 10^{q}$ where $\epsilon$ is the approximation error. If the binary significand $m$ is chosen optimally, then the error must be as small as possible: $|\epsilon|\leq 1/2$. Dividing throughout, we get $m+\epsilon=w\times 2^{q-p}/5^{-q}$ or $m=\operatorname{round}(w\times 2^{q-p}/5^{-q})$. Thus the problem is reduced to a division by an integer power of five ($5^{-q}$) of an integer $w$ that fits in a 64-bit word multiplied by a power of two. The correct value of $p$ is such that $m$ should fit in 53 bits (64-bit case) or 24 bits (32-bit case). In practice, we can derive the correct value $m$ after rounding if we compute the division $w\times 2^{b}/5^{-q}$ for a sufficiently large power of two $2^{b}$. Dividing a large integer by another large integer could be expensive. Thankfully, we can compute the quotient and remainder of such a division by the divisor $d=5^{-q}$ using a multiplication followed by a right shift when the divisor $d$ is known ahead of time. Many optimizing compilers have been using such a strategy for decades. We apply the following result derived from Warren [24, 25]. ###### Theorem 9.1. Consider an integer divisor $d>0$ and a range of integer numerators $n\in[0,N]$ where $N\geq d$ is an integer. We have that $\displaystyle n\div d=\operatorname{floor}(c\times n/t)$ for all integer numerators $n$ in the range if and only if $\displaystyle 1/d\leq c/t<\left(1+\frac{1}{N-\operatorname{remainder}(N+1,d)}\right)1/d.$ Intuitively, if $c/t$ is close to $1/d$, we can replace $n/d$ by $n\times c/t$. We apply Theorem 9.1 by picking a power of two for $t$ so that the computation of $\operatorname{floor}(c\times n/t)$ can be implemented as a multiplication followed by a shift. Given $t$, the smallest constant $c$ such that $1/d\leq c/t$ holds is $c=\operatorname{ceiling}(t/d)$. It remains to check that $c=\operatorname{ceiling}(t/d)$ is sufficiently close to $t/d$: $\displaystyle\operatorname{ceiling}(t/d)\times d<\left(1+\frac{1}{N-\operatorname{remainder}(N+1,d)}\right)t.$ We have that $\operatorname{ceiling}(t/d)\times d\leq t-1+d$. By substitution, we get the necessary condition $d-1<t/(N-\operatorname{remainder}(N+1,d))\leq t/N$. And so we have the convenient sufficient condition $t>(d-1)\times N$. We have shown Corollary 9.2. ###### Corollary 9.2. Consider an integer divisor $d>0$ and a range of integer numerators $n\in[0,N]$ where $N\geq d$ is an integer. We have that $\displaystyle n\div d=\operatorname{floor}\left(\operatorname{ceiling}\left(\frac{t}{d}\right)\frac{n}{t}\right)$ for all integer numerators $n$ if $t>(d-1)\times N$. Observe how we multiply the numerator by a reciprocal ($\frac{t}{d}$) that is rounded up. It suggests that we need to store precomputed reciprocals while rounding up. In contrast, when processing positive decimal exponents, we would merely truncate the power, thus effectively rounding it down. We consider two scenarios. Firstly, the case with negative powers with exponents near zero requires more care because we may need to round to the nearest even (§ 9.1). Secondly, we consider the general case ($5^{-q}\geq 2^{64}$) (§ 9.2). ### 9.1 Negative Exponents Near Zero ($q\geq-27$) We are interested in identifying cases when rounding to even is needed because of a tie (e.g., $2^{-1022}+2^{-1074}+2^{-1075}$ is one such number). Round-to- even only happens for negative values of $q$ when $q\geq-4$ in the 64-bit case and when $q\geq-17$ in the 32-bit case (see § 6). In either cases, we have that $5^{-q}$ fits in a 64-bit word (i.e., $5^{-q}<2^{64}$). Theorem 9.1 tells us how to quickly compute a quotient given a precomputed reciprocal, but for values of $q$ near zero, we need to be able to detect when the remainder is zero, so we know when to round ties to even. We use the following technical lemma which allows us to verify quickly whether our significand is divisible by the power of five. ###### Lemma 9.3. Given an integer divisor $d>0$, pick an integer $K>0$ such that $2^{K}\geq d$. Given an integer $w>0$, then $(w\times 2^{K})\div d$ is divisible by $2^{K}$ if and only if $w$ is divisible by $d$. ###### Proof 9.4. If $w$ is divisible by $d$ then $(w\times 2^{K})\div d=(w\div d)\times 2^{K}$ and thus $(w\times 2^{K})\div d$ is divisible by $2^{K}$. Suppose that $(w\times 2^{K})\div d$ is divisible by $2^{K}$, then we can write $w\times 2^{K}=d\times 2^{K}\times\gamma+\rho$ where $\gamma,\rho$ are non-negative integers with $\rho<d$. But because $\rho<2^{K}$, we must have $\rho=0$ and so $w\times 2^{K}=d\times 2^{K}\times\gamma$ or $w=d\times\gamma$ which shows that $w$ is divisible by $d$. To detect a zero remainder, we can consider the decimal significand $w$ as a 128-bit integer (made of two 64-bit words). Within the most significant 64-bit word, we shift $w$ by an adequate power of two so that the 63${}^{\textrm{rd}}$ bit has value 1: let us call the result $w^{\prime}$. The second, least significant, word is virtual and assumed to be zero. In effect, we consider the 128-bit value $2^{64}\times w^{\prime}$. We are going to treat mathematically this 128-bit value as a 127-bit value equal to $w^{\prime}\times 2^{63}$, ignoring the least significant bit. We could ignore more than one bit, but it is convenient to ignore the least significant bit. As long as $5^{-q}<2^{63}$, we have that $(w^{\prime}\times 2^{63})\div 5^{-q}\geq 2^{63}$ and we have more than enough accuracy to compute the desired binary significand. Hence, we can go as low as $q=-27$.666Though we go as low as $q=-27$, values of $q$ smaller than $-17$ cannot lead to a tie-to- even scenario using 64-bit or 32-bit floating-point numbers. We apply Theorem 9.1 to divide this value ($w^{\prime}\times 2^{63}$) no larger than $N=2^{127}$ by the power of five $5^{-q}$: we set $t=2^{b}$ with $b=127+\operatorname{ceiling}(\log_{2}(5^{-q}))$ and the reciprocal $c=\operatorname{ceiling}(t/d)=\operatorname{ceiling}(2^{b}/5^{-q})$. We can check that $c\in[2^{127},2^{128})$. We can verify the condition ($t>(d-1)\times N$) of Corollary 9.2. The values of the 128-bit reciprocals are given in Table 4. $q$ | reciprocal $\div 2^{64}$ | reciprocal $\bmod 2^{64}$ ---|---|--- -1 | cccccccccccccccc | cccccccccccccccd -2 | a3d70a3d70a3d70a | 3d70a3d70a3d70a4 -3 | 83126e978d4fdf3b | 645a1cac083126ea -4 | d1b71758e219652b | d3c36113404ea4a9 -5 | a7c5ac471b478423 | 0fcf80dc33721d54 -6 | 8637bd05af6c69b5 | a63f9a49c2c1b110 -7 | d6bf94d5e57a42bc | 3d32907604691b4d -8 | abcc77118461cefc | fdc20d2b36ba7c3e -9 | 89705f4136b4a597 | 31680a88f8953031 -10 | dbe6fecebdedd5be | b573440e5a884d1c -11 | afebff0bcb24aafe | f78f69a51539d749 -12 | 8cbccc096f5088cb | f93f87b7442e45d4 -13 | e12e13424bb40e13 | 2865a5f206b06fba -14 | b424dc35095cd80f | 538484c19ef38c95 -15 | 901d7cf73ab0acd9 | 0f9d37014bf60a11 -16 | e69594bec44de15b | 4c2ebe687989a9b4 -17 | b877aa3236a4b449 | 09befeb9fad487c3 -18 | 9392ee8e921d5d07 | 3aff322e62439fd0 -19 | ec1e4a7db69561a5 | 2b31e9e3d06c32e6 -20 | bce5086492111aea | 88f4bb1ca6bcf585 -21 | 971da05074da7bee | d3f6fc16ebca5e04 -22 | f1c90080baf72cb1 | 5324c68b12dd6339 -23 | c16d9a0095928a27 | 75b7053c0f178294 -24 | 9abe14cd44753b52 | c4926a9672793543 -25 | f79687aed3eec551 | 3a83ddbd83f52205 -26 | c612062576589dda | 95364afe032a819e -27 | 9e74d1b791e07e48 | 775ea264cf55347e Table 4: Values of the 128-bit reciprocals in hexadecimal form for powers with negative exponents near zero as two 64-bit words. The reciprocal is given by $\operatorname{ceiling}(\frac{2^{b}}{5^{-q}})$ with $b=127+\operatorname{ceiling}(\log_{2}(5^{-q}))$. By Theorem 9.1, we have that $(2^{63}\times w^{\prime})\div 5^{-q}=((2^{63}\times w^{\prime})\times c)\div 2^{b}$. The 128-bit constant $c$ is precomputed for all relevant powers of $q=-1,-2,\ldots,-27$. Computationally, we do not need to store the product of the shifted 128-bit $w$ with the 128-bit reciprocal $c$ as a 256-bit product: we compute only the first two words (i.e., the most significant 128-bit). And then we select only the most significant 127 bits. Thus checking that the second (least significant) word of the computed product is zero except maybe for the least significant bit is enough to determine that our word ($w$) is divisible by five. By dividing (the shifted) $w$ (that is $w^{\prime}$) by $5^{-q}$, we effectively compute a binary significand $2m+1$ as per the equation $(2m+1)\times 2^{p-1}=w\times 10^{q}$. We need to round to even whenever $(2m+1)\in[2^{53},2^{54})$ for 64-bit floating-point numbers and whenever $(2m+1)\in[2^{24},2^{53})$ for 32-bit floating-point numbers. Otherwise, we round the result normally to the nearest 53-bit word (64-bit numbers) or 24-bit word (32-bit numbers). We need up to two 64-bit multiplications. We stop after the first multiplication if and only if the least significant bits of the most significant words are not all ones. We want that the most significant bits of the most significant 64-bit word are exact: 55 bits for 64-bit numbers and 26 bits for 32-bit numbers. We have that $64-55$ is 9 and $64-26$ is 38. Hence, we check the least significant 9 bits for 64-bit numbers and the least significant 38 bits for 32-bit numbers. As long as we are not in a round-to- even case, we round up or down based on the least significant selected bit. If all bits have the value 1, then rounding up overflows into a more significant bit and we must shift by one bit. We need to ensure that we correctly identify all ties requiring the round-to- even strategy. Specifically, we need to never incorrectly classify a number as a tie, and we need to never miss a tie. * • We need to be concerned about a false round-to-even scenario when, after stopping with just one multiplication, we end up with a misleading result that could pass as a round-to-even case. Indeed, we can stop after one multiplication when the least significant bits of the most significant words are all zeros. However, a round-to-even case cannot occur after a single multiplication: 1. 1. It could happen if the least significant $64+9$ bits of the product are zeros. The 128-bit product of two 64-bit words may only have as many trailing zeros as the sum of the number of trailing zeros of the first 64-bit word with the number of trailing zeros of the second word. To get a total of $64+9$ trailing zeros, assuming that both words are non-zero, we have the necessary conditions that both words must have at least 10 trailing zeros. Thus, for this problem to occur, we need for the most significant 64-bit word of the reciprocal $c$ to have at least 10 trailing zeros. We can check that it does not happen: there are only 17 powers to examine. We find at most 2 trailing zeros. See Table 5 (third column). 2. 2. A false round-to-even may also happen if all the least significant $64+9$ bits of the product are zeros, except for the least significant bit. However, for the 128-bit product of two 64-bit words to have its least significant bit be 1, we need for both of the 64-bit words to have their least significant bits set to 1 (they are odd). Given an odd 64-bit integer, there is only one other 64-bit integer such as the least significant 64 bits of the product is 1. Indeed suppose that $a\times b_{1}=1\bmod 2^{64}$ and $a\times b_{2}=1\bmod 2^{64}$ for numbers in $[0,2^{64})$ then $a\times(b_{1}-b_{2})=0\bmod 2^{64}$ which implies that $b_{1}=b_{2}$. They are effectively multiplicative inverses (modulo $2^{64}$). We can thus compute the multiplicative inverses (see Fig. 4) and check the full 128-bit product. Again, we only need to examine 17 powers. We find the powers that have an odd integer in their most significant 64 bits, we compute the multiplicative inverse and we compute the full product. Looking at the most significant 64 bits of the resulting product, we find that they have at most 5 trailing zeros. See Table 5 (last column). $q$ | reciprocal $\div 2^{64}$ | inverse | product | 0s ---|---|---|---|--- -3 | 83126e978d4fdf3b | c687d6343eb1a1f3 | 65a5cdedb181dc22 | 1 -4 | d1b71758e219652b | 6978533007ec3183 | 5666aa8c1bca175b | 0 -5 | a7c5ac471b478423 | b464ceec1a874b8b | 76390df51733b898 | 3 -6 | 8637bd05af6c69b5 | 2d28ff519dc1fc9d | 17ad4acbd85ad372 | 1 -9 | 89705f4136b4a597 | 47a5ffb53d302a27 | 26774920b7634d5b | 0 -12 | 8cbccc096f5088cb | ccda17e7d0519ce3 | 709e5881abf430de | 1 -13 | e12e13424bb40e13 | a976a8f009f3ec1b | 950fca8d051f7f36 | 1 -14 | b424dc35095cd80f | 4776114e932f16ef | 32494e3df377fbda | 1 -15 | 901d7cf73ab0acd9 | b7d434f9093d1369 | 677c8a9266f5159b | 0 -16 | e69594bec44de15b | 30fad280461f66d3 | 2c1df79145125a20 | 5 -17 | b877aa3236a4b449 | 89ee897ef59d7df9 | 6363ec689fe3979b | 0 Table 5: Values of the most significant 64 bits of the 128-bit reciprocals in hexadecimal form for powers of negative exponents near zero and the multiplicative inverse modulo $2^{64}$ of the reciprocal for odd reciprocals ($q=-1,-2,-7,-8,-10,-11$ are omitted since their reciprocals are even). We compute the most significant bits of the 128-bit product between the reciprocal and its inverse. We indicate the number of trailing zeros for the most significant bits of the product. * • We need to be concerned with the reverse scenario where, after a single multiplication, we stop the computation and fail to detect an actual round-to- even case. If we stop after one multiplication, then at least one of the least significant bits (9 bits for 64-bit numbers, 38 bits for 32-bit numbers) of the most significant 64 bits is zero. In such a case, the 128 most significant bits of the full (exact) product must end with a long stream of zeros, except maybe for the least significant bit. We know that the most significant 64 bits are exact after a single product, except maybe for the need to increment by 1. The most significant 64 bits cannot be exact after one multiplication if we have a round-to-even case. So we must increment them by 1 following the second multiplication, and then the final result contains at least one non-zero bit in the least significant bits (9 bits for 64-bit numbers, 38 bits for 32-bit numbers) of the most significant 64 bits. It contradicts the fact that we had an actual round-to-even case. Hence, we cannot fail to detect an actual round- to-even case by stopping the computation after one multiplication. Thus we can identify accurately the round-to-even cases. In these cases, we proceed as in § 8.1. After discarding a potential leading zero-bit, we have 54 bits (64-bit case) and 25 bits (32-bit case). The least significant bit is always a 1-bit. We round down when the second least significant bit is zero, otherwise we round up. When rounding up, we might overflow into an additional bit if we only have ones, in such a case we shift the result. ### 9.2 Other Negative Powers ($q<-27$) Consider the case where the decimal exponent is far from zero ($q<-27$). In such cases, the decimal number can never be exactly in-between two floating- point numbers: thus with a single extra bit of accuracy, we can safely either round up or down. The smallest positive value that can be represented using a 64-bit floating- point number is $2^{-1074}$. For 32-bit numbers, we have the larger value $2^{-149}$. Because we have that $w\times 10^{-343}<2^{-1074}$ for all $w<2^{64}$, it follows that we never have to be concerned with overly small decimal exponents: when $q<-342$, then the number is assuredly zero. From the decimal number $m\times 10^{q}$, we seek the binary significand $m=\operatorname{round}(w\times 2^{q-p}/5^{-q})$ where the binary power $p$ is chosen such that $m$ is within the range of the floating-point numbers (e.g., $m\in[2^{52},2^{53})$). It is enough to compute $m^{\prime}=\operatorname{floor}(w\times 2^{b}/5^{-q})$ with $b$ large enough that $m^{\prime}\geq 2^{53}$ so that we can compute $m=\operatorname{round}(w\times 2^{q-p}/5^{-q})$ accurately by selecting the most significant 53 bits (64-bit numbers) or 24 bits (32-bit numbers) of the wider value $m^{\prime}$ and then round it up (or down) based on the $54^{\mathrm{th}}$ or $25^{\mathrm{th}}$ bit value. We can pick $b=64+\operatorname{ceiling}(\log_{2}5^{-q})$. We apply Corollary 9.2 with $t=2^{2b}$, $d=5^{-q}$, and $N=(2^{64}-1)2^{b}$. We precompute $c=\operatorname{ceiling}(t/d)=\operatorname{ceiling}(2^{2b}/5^{-q})$ for all relevant powers of $q\geq-342$. See Table 6. We only store the most significant 128 bits of $c$, and rely on a truncated multiplication. Because there is no concern with rounding to even, we can safely round up from the most significant bits of the computed quotient. We do just one multiplication if it provides the number of significant bits of the floating-point standard (53 bits for 64-bit numbers and 24 bits for 32-bit numbers) plus one additional bit to determine the rounding direction, and yet one more bit to handle the scenario where the computed product has a leading zero. We always stop after this second multiplication when we have a truncated product with the second most significant word not filled with ones ($2^{64}-1$). Otherwise, we fall back on a higher-precision approach, an unlikely event. After possibly omitting the leading zero of the resulting product, we select the most significant bits (54 bits in the 64-bit case, 25 bits in the 32-bit case). We then round up or down based on the least significant bit to 53 bits (64-bit case) or to 24 bits (32-bit case). When rounding up, we might overflow to an additional bit if we have all ones: in such case we shift to get back 53 bits (64-bit case) or 24 bits (32-bit case). $q$ | reciprocal (64 msb) | reciprocal (next 64 msb) ---|---|--- -40 | 8b61313bbabce2c6 | 2323ac4b3b3da015 -39 | ae397d8aa96c1b77 | abec975e0a0d081a -38 | d9c7dced53c72255 | 96e7bd358c904a21 -37 | 881cea14545c7575 | 7e50d64177da2e54 -36 | aa242499697392d2 | dde50bd1d5d0b9e9 -35 | d4ad2dbfc3d07787 | 955e4ec64b44e864 -34 | 84ec3c97da624ab4 | bd5af13bef0b113e -33 | a6274bbdd0fadd61 | ecb1ad8aeacdd58e -32 | cfb11ead453994ba | 67de18eda5814af2 -31 | 81ceb32c4b43fcf4 | 80eacf948770ced7 -30 | a2425ff75e14fc31 | a1258379a94d028d -29 | cad2f7f5359a3b3e | 96ee45813a04330 -28 | fd87b5f28300ca0d | 8bca9d6e188853fc Table 6: Values of the 128-bit reciprocals in hexadecimal form for negative exponents as two 64-bit words. The reciprocal is given by $\operatorname{ceiling}(\frac{2^{2b}}{5^{-q}})$ with $b=64+\operatorname{ceiling}(\log_{2}5^{-q})$. ###### Example 3 Consider the case of the string 9.109e-31. We parse it as $9109\times 10^{-34}$. We load up the most significant 64 bits of the reciprocal corresponding to $q=-34$ which is 0x84ec3c97da624ab4 in hexadecimal form (see Table 4). We normalize 9109 so that, as a 64-bit word, its most significant bit is 1: $9109\times 2^{50}$. We multiply the two words to get that the most significant 64 bits of the product are 49e6a7201cf62db0 whereas the next most significant 64 bits are 0x5b10000000000000. We stop the computation since the second word is not filled with ones. The most significant bit of the product contains a 0. We shift the most significant 64 bits by 9 bits to get 10400639386286870. The least significant bit is zero so we round down to 5200319693143435 or 0x1279a9c8073d8b in hexadecimal form. We get that $9109\times 10^{-34}$ is the floating-point number 0x1.279a9c8073d8bp-100. See Example 4 in § 10 to learn how we determine that the binary exponent is -100. ### 9.3 Subnormals To represent values that are too small, the floating-point standard uses special values called subnormals. Whenever we end up with a value $m\times 2^{p}$ with $m\in[2^{52},2^{53})$ (64-bit case) or $m\in[2^{23},2^{24})$ (32-bit) but with $p$ too small, smaller than $-1022-52$ in the 64-bit case or smaller than $-126-23$ in the 32-bit case, we fall back on the subnormal representation. It uses a small value for the exponent to represent values in the range $[2^{-1022-52},2^{-1022})$ (64-bit case) or in the range $[2^{-126-23},2^{-126})$ (32-bit case). The values are given by $m\times 2^{-1022-52}$ (64-bit) or $m\times 2^{-126-23}$ (32-bit) while allowing $m$ to be any positive value no larger than $2^{52}$ or $2^{23}$. To construct the subnormal value, we take the original binary significand $m$ and we divide it by a power of two, with rounding. Thus, for example, if we are given the 64-bit value $(2^{53}-1)\times 2^{-1022-54}$, we observe that the power of two is too small ($-1022-53<-1022-52$) by exactly two. Thus we take the binary significand $2^{53}-1$ and divide it by four, with rounding: we get $2^{51}$ and so we get the subnormal floating-point number $2^{51}\times 2^{-1022-52}$. Thankfully, rounding is relatively easy since we never need to handle the round-to-even case with subnormals, because it only occurs with powers of exponents near zero. We should be mindful that, in exceptional cases, the rounding process can lead us to find that we do not have a subnormal. Indeed, consider the value $(2^{53}-1)\times 2^{-1022-53}$, its power of two is too small ($-1022-53<-1022-52$) by exactly one. We take the binary significand $2^{53}-1$ and divide it by two, with rounding, getting $2^{52}$ and so we end up with the normal number $2^{52}\times 2^{-1022-52}$. ## 10 Computing the Binary Exponent Efficiently We are approximating a decimal floating-point number $w\times 10^{q}$ with a binary floating-point number $m\times 2^{p}$. We must compute the binary exponent $p$. Starting from the power of ten $10^{q}$, we want write it as a value in $[1,2)$, as prescribed by the floating-point standard, multiplied by a power of two. We have two distinct cases depending on the sign of $q$: * • when $q\geq 0$, we have $10^{q}=2^{q}\times 5^{q}=\frac{5^{q}}{2^{\operatorname{floor}(\log_{2}5^{q})}}\times 2^{q+\operatorname{floor}(\log_{2}5^{q})}$, * • when $q<0$, we have $10^{q}=2^{q}\times 5^{q}=\frac{2^{\operatorname{ceiling}(\log_{2}5^{-q})}}{5^{-q}}\times 2^{q-\operatorname{ceiling}(\log_{2}5^{-q})}$. We can verify that both constraints are satisfied: $5^{q}/2^{\operatorname{floor}(\log_{2}5^{q})}\in[1,2)$ and $2^{\operatorname{ceiling}(\log_{2}5^{-q})}/5^{-q}\in[1,2)$. Hence we have that the binary powers corresponding to the powers of ten are given by $q+\operatorname{floor}(\log_{2}(5^{q}))=q-\operatorname{ceiling}(\log_{2}5^{-q})$. For example, we have that $10^{5}=5^{5}/2^{11}\times 2^{16}=1.52587890625\times 2^{16}$ since $\operatorname{floor}(\log_{2}5^{5})=11$. Computing $q+\log_{2}(5^{q})$ could require an expensive iterative process. The decimal exponent $q$ is in limited range of values, say $q\in(-400,350)$. We have that $q+\log_{2}(5^{q})=q+q\log_{2}(5)=q(1+\log_{2}(5))$ and $1+\log_{2}(5)\approx 217706/2^{16}$. We can check that over the interval $q\in(-400,350)$, we have that $q+\operatorname{floor}(\log_{2}(5^{q}))=(217706\times q)\div 2^{16}$ (exactly) as one can verify numerically. The division (by $2^{16}$) can be implemented as a logical shift. Thus we only require a multiplication followed by a shift. We initially derived this efficient formula using a satisfiability-modulo-theories (SMT) solver [26]. In our algorithm, we normalize the decimal significand so that it is in $[2^{63},2^{64})$. That is, given the string 1e12, we first parse it as the decimal significand $w=1$ and the decimal exponent $q=12$. We then normalize $w=1$ to $w^{\prime}=2^{63}$ (shifting it by 63 bits) and we proceed with the computation of the binary significand. Had we started with $w=2^{4}$ (say), then we would have shifted by only $63-4$ bits and then the binary exponent must be incremented by 4. For example, using the input string 16e12 instead of 1e12, we would have used the decimal significand $w=16$ but still ended up with the normalized significand $w^{\prime}=2^{63}$. Yet the binary exponent of 16e12 is clearly 4 more than the binary exponent of 1e12. In other words, we need to take into account the number of leading zeroes of the decimal significand. Thus we increment the binary exponent by $63-l$ where $l$ is the number of leading zeros of the original decimal significand $w$ as a 64-bit word. For powers of ten, the product of the normalized significand with either the power of five or its reciprocal has a leading zero since $2^{63}\times(2^{64}-1)<2^{127}$. When the product is larger and it overflows in the most significant bit, then the binary exponent must be incremented by one. Thus we finally have the following formula $\left(\left(217706\times q\right)\div 2^{16}\right)+63-l+u$ where $u$ is the value of the most significant bit of the product (0 or 1) and where $l$ is the number of leading zeros of $w$. Furthermore, when we round up the resulting significand, it may sometimes overflow: e.g., if the most significant bits of the product are all ones, we overflow to a more significant bit and we need to shift the result. In such cases, we increment the binary exponent by one. When serializing the exponent in the IEEE binary format, we need to add either 1023 (64-bit) or 127 (32-bit) to the exponent; these constants (1023 and 127) are sometimes called _exponent biases_. For example, the 64-bit binary exponent value from $-1022$ to $1023$ are stored as the unsigned integer values from $1$ to $2046$. The serialized exponent value 0 is reserved for subnormal values while the serialized exponent value $2047$ is reserved for non-finite values. ###### Example 4 Consider again Example 3. We start from $9109\times 10^{-34}$. Because $q<0$, we compute $q-\operatorname{ceiling}(\log_{2}5^{-q})$ and get -113. We have that 9109 has 50 leading zeros as a 64-bit word and we normalize it as $9109\times 2^{50}$. Thus we have $I=50$ and so we need to increment the binary exponent by $63-I$ or 13. We get a binary exponent of -100. We verify that the product has a leading zero bit so we have that the binary exponent must be -100. ## 11 Processing Long Numbers Quickly In some uncommon instances, we may have a decimal significand that exceeds 19 digits. Unfortunately, if we are given a value with superfluous digits, we cannot truncate the digits: it may be necessary to read tens or even hundreds of digits (up to 768 digits in the worst case). Indeed, consider the second smallest 64-bit normal floating-point value: $2^{-1022}+2^{-1074}$ ($\approx 2.2250738585072019\times 10^{-308}$) and the next smallest value $2^{-1022}+2^{-1073}$ ($\approx 2.2250738585072024\times 10^{-308}$). If we pick a value that is exactly in-between ($2^{-1022}+2^{-1074}+2^{-1075}$), we need to break the tie by rounding to even (to the larger value $2.2250738585072024\times 10^{-308}$ in this case). Yet any truncation of the value would be slightly closer to the lower value ($\approx 2.2250738585072019\times 10^{-308}$). We can write $2^{-1022}+2^{-1074}+2^{-1075}$ exactly as a decimal floating-point value $w\times 10^{q}$ for integers $w$ and $q$, but the significand requires 768 digits. We can show that it is the worst case. When there are too many digits, we could immediately fall back on a higher- precision approach. However, if we just use the most significant 19 digits, and truncate any subsequent digits, we might be able to uniquely identify the exact number. It is trivially the case if the truncated digits are all zeros, in which case we can safely dismiss the zeros. Otherwise, if $w$ is the truncated significand, then the exact value is in the interval $(w\times 10^{q},(w+1)\times 10^{q})$. Thus we may apply our algorithm to both $w\times 10^{q}$ and $(w+1)\times 10^{q}$. If they both round to the same binary floating-point number, then this floating-point number has to match exactly the true decimal value. If $w$ is limited to 19 digits, then $w+1\leq 10^{19}<2^{64}$ so we do not have to worry about possible overflows. To assess the effectiveness of this approach, we can try a numerical experiment. We generate random 19-digit significands and append an exponent (e.g., 1383425612993491676e-298 and 1383425612993491677e-298). We find that for such randomly generated values, about 99.8% of the successive values map to the same 64-bit floating-point number, over a range of exponents (e.g., from $-300$ to $300$). We can also generate random 64-bit numbers in the unit interval $[0,1]$, serialize them to 19 digits and add one to the last digit. We get that in about 99.7% of all cases, changing the last digit does not affect the value. In other words, we often can determine exactly a floating- point value after truncating to 19 digits in most cases. When it fails, we can fall back on a higher-precision approach. In our software implementation (see § 5), we adapted a general implementation used as part of the Go standard library. Given that it should be rarely needed, its performance is secondary. However, it has to be exact. ## 12 Experiments We implemented our algorithm and published it as an open source software library.777https://github.com/fastfloat/fast_float It closely follows the C++17 standard for the std::from_chars functions, supporting both 64-bit and 32-bit floating-point numbers. It has been thoroughly tested. Though our code is written using generally efficient C++ patterns, we have not micro-optimized it. Our implementation requires a C++11-compliant compiler. It does not allocate memory on the heap and it does not throw exceptions. To implement our algorithm, we use a precomputed table of powers of five and reciprocals, see Appendix B. Though it uses $10\text{\,}\mathrm{KiB}\text{/}$, we should compare it with the original Gay’s implementation of strtod in C which uses $160\text{\,}\mathrm{KiB}\text{/}$ and compiles to tens of kilobytes. Our table is used for parsing both 64-bit and 32-bit numbers. There are many libraries that support number parsing. For our purposes, we limit ourselves to C++ production-quality libraries. We only consider libraries that offer exact parsing. See Table 7. We choose to omit libraries written in other programming languages (Java, D, Rust, etc.) since direct comparisons between programming languages are error prone—see Appendix E for benchmarks of a Rust version of our algorithm.888The release notes for Go version 1.16, which makes use of our approach, state that “ParseFloat now uses the [new] algorithm, improving performance by up to a factor of 2.”, https://golang.org/doc/go1.16. Our C++ code was also ported to C#, https://github.com/CarlVerret/csFastFloat, and Java, https://github.com/wrandelshofer/FastDoubleParser with good results. We also include in our benchmarks the system’s C function strtod, configured with the default locale. Though the standard Linux C++ library supports the C++17 standard, it does not yet provide an implementation of the std::from_chars functions for floating-point numbers. To ensure reproducibility, we publish our full benchmarking software.999https://github.com/lemire/simple_fastfloat_benchmark, git tag v0.1.0 Our benchmarking routine takes as input a long array of strings that are parsed in sequence. Somewhat arbitrarily, we seek to compute the minimum of all encountered numbers. Such a running-minimum function carries minimal overhead compared to number parsing. Hence, we effectively measure the throughput of number parsing. We are also careful to use datasets containing thousands of numbers for two reasons: * • On the one hand, all measures have a small bounded error: by using large sequence of tests, we amortize such errors. * • On the other hand, the performance of modern processors is often closely related to its ability to predict branches. A single mispredicted branch can waste between 10 to 20 cycles of computations. When in a repeating loop, some recent processors can learn to predict with high accuracy a few thousands of branches [seznec2011new]. We repeat all experiments 100 times. We avoid memory allocations throughout the process. On such a computational benchmark, timings follow a distribution resembling a log-normal distribution with a long tail associated with noise (interrupts, cache competition, context switches, etc.) and a non-zero minimum. The median is located between the minimum and the average. Using a common convention [28], we compute both the minimum time and the average time: the difference between the two is our margin of error. If the minimum time and the average time are close, our measures are reliable. We find that the error margin is consistently less than 5% on all platforms—often under 1%. On Linux platforms, we can _instrument_ our benchmark so that we can programmatically track the number of cycles and number of instructions retired using CPU performance counters from within our own software. Such instrumentation is precise (i.e., not the result of sampling) and does not add overhead to the execution of the code. Typically, the number of instructions retired by a given routine varies little from run to run and may be considered exact, especially given that we ensure a stable number of branch mispredictions. One benefit of instrumented code is that we can measure the effective CPU clock frequency during the benchmarked code: modern processors adjust their frequency dynamically based on load, power usage and heat. Our Linux systems are configured for performance and we observe the expected CPU frequencies. Our benchmarks exclude disk access or memory allocation: strings are preallocated once. To ensure a consistent and reproducible system configuration, we run our benchmark under a privileged docker environment based on a Ubuntu 20.10 image.101010https://github.com/lemire/docker_programming_station, git tag v0.1.0 According to our tests, the docker overhead for purely computational tasks when the host is itself Linux, is negligible. For our benchmarks, we use the GNU GCC 10.2 compiler with full optimization (-O3 -DNDEBUG) under Linux. Our benchmark programs are single binaries applying the different parsing functions to the same strings. Though we access megabytes of memory, most of the data remains in the last-level CPU cache. We are not limited by cache or memory performance. We rely a realistic data source that is used by the Go developers to benchmark the standard library: the canada dataset comes from a JSON file commonly used for benchmarking [28]. It contains 111k 64-bit floating-point numbers serialized as strings. The canada number strings are part of geographic coordinates: e.g., 83.109421000000111. We also include synthetic datasets containing 100k numbers each. The uniform dataset is made of 64-bit random numbers in the unit interval $[0,1]$. The integer data set is made of randomly generated 32-bit integers. Though it is inefficient to use a floating-point number parser for integer values, we believe that it might be an interesting test case. It is an instance where our code fails to show large benefits. For the synthetic dataset, we considered two subcases: the floating-point number can either be serialized using a fixed decimal significand (17 digits) or using a minimal decimal significand as 64-bit numbers (using at most 17 digits [16]). We found relatively little difference in performance (no more than 10%) on a per-float basis between these two cases. In both cases, the serialization is exact: an exact 64-bit parser should recover exactly the original floating- point value. We present our results with the concise serialization. Table 7: Production-quality number parsing C++ libraries. Both double-conversion and abseil have been authored by Google engineers. Processor | snapshot | link ---|---|--- Gay’s strtod (netlib) | 2001 | www.netlib.org/fp/ double-conversion | version 3.1.5 | github.com/google/double-conversion.git abseil | 20200225.2 | github.com/abseil/abseil-cpp To better assess our algorithm, we tested it on a wide range of Linux-based systems which include x64 processors, an ARM server processor and an IBM POWER9 processor. See § 8. We report the effective frequency, that is, the CPU frequency measured during the execution of our code. Our experiments are single-threaded: the Ampere system contains 32 ARM cores and would normally be competitive against the other systems if all cores were used. However, on a single-core basis, it is not expected to match the other processors. Table 8: Systems tested Processor | Effective Frequency | Microarchitecture | Compiler ---|---|---|--- Intel i7-6700 | $3.7\text{\,}\mathrm{GHz}\text{/}$ | Skylake (x64, 2015) | GCC 10.2 AMD EPYC 7262 | $3.39\text{\,}\mathrm{GHz}\text{/}$ | Zen 2 (x64, 2019) | GCC 10.2 Ampere | $3.2\text{\,}\mathrm{GHz}\text{/}$ | ARM Skylark (aarch64, 2018) | GCC 10.2 IBM | $3.77\text{\,}\mathrm{GHz}\text{/}$ | POWER9 (ppc64le, 2018) | GCC 10.2 We report the speed in millions of numbers per second for our different datasets and different processors in Table 9. We find that the from_chars function in the abseil library is often superior to Gay’s implementation of strtod (labeled as netlib) which is itself superior to both double-conversion and the strtod function including the GNU standard library. The implementation notes of the abseil library [11] indicate that it relies on a general strategy which is not fundamentally different from our own.111111The abseil library does not rely on Clinger’s fast path when parsing numbers. It also uses less accurate product computation. Even so, our approach is generally twice as fast as the abseil library and up to five times faster than what the standard library offers. We find that for the integer test, netlib is superior to all other alternatives (including abseil) except for our own. The gap between our approach and netlib when parsing integers is modest (about 20%). Overall, our proposed approach is three to five times faster than the strtod function available in the GNU standard library. And it is often more than twice as fast as the state-of-the-art abseil library. Table 9: Millions of 64-bit floating-point numbers parsed per second under different processor architectures | canada | uniform | integer ---|---|---|--- netlib | 9.6 | 10 | 48 d.-conversion | 9.4 | 10 | 18 strtod | 9.0 | 9.4 | 20 abseil | 18 | 19 | 27 our parser | 45 | 45 | 61 (a) Intel Skylake (x64) canada | uniform | integer ---|---|--- 10 | 11 | 57 9.0 | 9.9 | 24 9.3 | 9.9 | 18 21 | 21 | 30 51 | 52 | 70 (b) AMD Zen 2 (x64) | canada | uniform | integer ---|---|---|--- netlib | 8.1 | 8.7 | 23 d.-conversion | 5.4 | 5.8 | 12 strtod | 3.9 | 4.2 | 8.7 abseil | 9.1 | 9.4 | 13 our parser | 22 | 21 | 26 (c) Ampere Skylark (ARM, aarch64) canada | uniform | integer ---|---|--- 9.0 | 10 | 39 5.8 | 6.4 | 18 4.8 | 5.3 | 12 12 | 12 | 17 42 | 39 | 46 (d) IBM POWER 9 To understand our good results, we look and the number of instructions and cycles per number for one representative dataset (uniform) and for the AMD Zen 2 processor. See Table 10. As expected, we use half as many instructions on average as the abseil library. We find interesting that we use only about three times fewer instructions than the strtod function, but 5.6 times fewer cycles. Our approach causes almost no branch mispredictions, in contrast with Gay’s netlib library. Similarly, while we retire 4.2 instructions per cycle, Gay’s netlib library is limited at 2.2 instructions per cycle. To summarize, our approach uses fewer instructions, generates fewer branch mispredictions and retires more instructions per cycle. To identify our bottleneck, we run the parsing routine while skipping the conversion from a decimal significand and exponent to the standard decimal form. Instead, we sum the decimal significand and the exponent and return the result as a simulated floating-point value. We find that we save only about a quarter of the number of instructions and a quarter of the time (cycles). In other words, our decimal-to-binary routine is so efficient that it only uses about a quarter of our computational time. Most of the time goes into parsing the input string and converting it to a decimal significand and exponent. Table 10: Instructions, mispredicted branches and cycles per 64-bit floating-point number in the uniform model on the AMD Zen 2 processor. We also provide the number of instructions per cycle. The “just string” row corresponds to our parser but without the final decimal to binary conversion. | Instructions | mispredictions | cycles | instructions/cycle ---|---|---|---|--- netlib | 740 | 4.1 | 330 | 2.2 double-conversion | 1100 | 1.7 | 380 | 3.0 strtod | 1100 | 0.7 | 370 | 3.0 abseil | 600 | 0.5 | 160 | 3.8 our parser | 280 | 0.01 | 66 | 4.2 (just string) | 215 | 0.00 | 46 | 4.7 Our results using 32-bit numbers are similar. To ease comparison, we produce exactly the same numbers strings as in the 64-bit case. We replace the strtod function with the equivalent strtof function. We present the result in Table 11. It suggests that there is little speed benefit in reading numbers as 32-bit floating-point numbers instead of 64-bit floating-point numbers given the same input strings. The result does not surprise us given that we rely on the same algorithm. Table 11: Instructions, mispredicted branches and cycles per 32-bit floating-point number in the uniform model on the AMD Zen 2 processor. We also provide the number of instructions per cycle. | Instructions | mispredictions | cycles | instructions/cycle ---|---|---|---|--- strtof | 1100 | 0.7 | 350 | 3.1 abseil | 600 | 0.5 | 170 | 3.6 our parser | 280 | 0.00 | 64 | 4.3 We find it interesting to represent the parsing speed in terms of bytes per second. On the canada dataset using the AMD Zen2 system, our parser exceeds $1\text{\,}\mathrm{GiB}\text{/}\mathrm{s}$ ($1080\text{\,}\mathrm{MiB}\text{/}\mathrm{s}$). It is $2.5$ times faster than the fastest competitor (abseil) and $5$ times faster than the other parser. See Fig. 2. For the synthetic dataset, we use the concise number serialization instead of relying on a fixed number of digits, to avoid overestimating the parsing speed. Our parser runs at almost over $900\text{\,}\mathrm{MiB}\text{/}\mathrm{s}$ compared to less than $200\text{\,}\mathrm{MiB}\text{/}\mathrm{s}$ for the strtod function. If we serialize the numbers so that they use a fixed number of digits (17), we reach higher speeds: our parser exceeds $1\text{\,}\mathrm{GiB}\text{/}\mathrm{s}$ (not shown). netlibd.-conv.strtodabseilthis paper$0$$200$$400$$600$$800$$1{,}000$throughput ($\mathrm{MiB}\text{/}\mathrm{s}$) (a) canada netlibd.-conv.strtodabseilthis paper$0$$200$$400$$600$$800$$1{,}000$throughput ($\mathrm{MiB}\text{/}\mathrm{s}$) (b) uniform Figure 2: Parsing speed for the canada dataset and for random 64-bit floating-point number in the uniform model, serialized concisely, on the AMD Zen 2 processor. In Table 12, we provide statistics regarding which code paths are used by different datasets. The integer dataset is entirely covered by Clinger’s fast path. It explains why our performance on this dataset is similar to the netlib approach, since we rely on essentially the same algorithm. For both the canada and uniform dataset, most of the processing falls on our parser as opposed to Clinger’s fast path. After initially parsing the input string, our fast algorithm begins with one or two multiplications between the decimal significand and looked up table values. We observe that a single multiplication is all that is necessary in most cases. In our experiments, we never need to fall back on a higher-precision approach. Table 12: Code path frequencies for different datasets using our parser. The percentages are relative to the number of input number strings. | canada | uniform | integer ---|---|---|--- Clinger’s fast path | 8.8% | 0% | 100% our path | 91.2% | 100% | 0% two multiplications | 0.6% | 0.66% | 0% #### Many digits We designed our algorithm for the scenario where numbers are serialized to strings using no more than 17 digits. However, we can not always ensure that such a reasonable limit is respected. To test the case where we have many more than 17 digits, we create big integer values by serializing three randomly selected 64-bit integers in sequence. On the AMD Zen 2 system, we find that our parser exceeds $1100\text{\,}\mathrm{MiB}\text{/}\mathrm{s}$. The abseil library achieves similar speeds $910\text{\,}\mathrm{MiB}\text{/}\mathrm{s}$ which is more than twice as fast as Gay’s netlib $390\text{\,}\mathrm{MiB}\text{/}\mathrm{s}$. The strtod function is limited to $110\text{\,}\mathrm{MiB}\text{/}\mathrm{s}$. #### Visual Studio Unfortunately, we are not aware of a standard implementation of the from_chars function under Linux. However, Microsoft provides one such fast function as part of its Visual Studio 2019 system. We use the latest available Microsoft C++ compiler (19.26.28806 for x64). We compile in release mode with the flags /O2 /Ob2 /DNDEBUG. These results under Windows are generally comparable to our Linux results. See Table 13. Microsoft’s from_chars function is faster than its strtod function. However, our parser is several times faster than Microsoft’s from_chars function. Table 13: Millions of 64-bit floating-point numbers parsed per second under a $4.2\text{\,}\mathrm{GHz}\text{/}$ Intel 7700K processor using Visual Studio 2019 | canada | uniform | integer ---|---|---|--- netlib | 20 | 18 | 48 d.-conversion | 10 | 10 | 18 strtod | 6.0 | 5.8 | 15 from_chars | 6.7 | 7.2 | 22 abseil | 16 | 15 | 22 our parser | 37 | 48 | 60 #### Apple M1 Processor In November 2020, Apple released laptops with a novel ARM 3.2 GHz processor (M1). The M1 processor has 8 instruction decoders compared to only 4 decoders on most x64 processors. Though we would normally avoid benchmarking on a laptop due to potential frequency throttling, we found consistent run-to-run results (within 1%) and a low margin of error (within 1%). We compiled our benchmark software on such a laptop using Apple’s LLVM clang compiler (Apple clang version 12.0.0 using the flags -O3 -DNDEBUG). We present our throughput results in Fig. 3. Our parser reaches $1.5\text{\,}\mathrm{GiB}\text{/}\mathrm{s}$ on the uniform dataset. On the Apple platform, the strtod function is several times slower than any other number parser. Other parsers (netlib, double-conversion and abseil) are about three times slower in these tests. netlibd.-conv.strtodabseilthis paper$0$$500$$1{,}000$$1{,}500$throughput ($\mathrm{MiB}\text{/}\mathrm{s}$) (a) canada netlibd.-conv.strtodabseilthis paper$0$$500$$1{,}000$$1{,}500$throughput ($\mathrm{MiB}\text{/}\mathrm{s}$) (b) uniform Figure 3: Parsing speed for the canada dataset and for random 64-bit floating-point number in the uniform model, serialized concisely, on the Apple M1 processor. ## 13 Conclusion Parsing floating-point numbers from strings is a fundamental operation supported by the standard library of almost all programming languages. Our results suggest that widely used implementations might be several times slower than needed on modern 64-bit processors. When the input strings are retrieved from disks or networks with gigabytes per second in bandwidth, a faster approach should be beneficial. We expect that more gains are possible mostly in how we parse the input strings into a decimal significand and exponent. For example, we could use advanced processor instructions such as SIMD instructions [28]. It also be possible to accelerate the processing by relaxing correctness conditions: e.g., the parsing could be only exact up to an error in the last digit. However, we should be mindful of the potential problems that arise when different software components parse the same numbers to different binary values. Floating-point numbers may be stored in binary form and accessed directly without parsing. However, some engineers prefer to rely on text formats. Hexadecimal floating-point numbers (Appendix C) may provide a convenient alternative for greater speed in such cases. ## Acknowledgements Our work benefited especially from exchanges with M. Eisel who motivated the original research with his key insights. We thank N. Tao who provided invaluable feedback and who contributed an earlier and simpler version of this algorithm to the Go standard library. Our fallback implementation includes code adapted from Google Wuffs, a memory-safe programming language, which was published under the Apache 2.0 license. To our knowledge, the fast path for long numbers was first implemented by R. Oudompheng for the Go standard library. We thank A. Milovidov for his feedback regarding benchmarking. We are grateful to W. Muła for his thorough review of an early manuscript: his comments helped us improve the document significantly. We thank I. Smirnov for his feedback on benchmarking statistics. We thank P. Cawley for his feedback on the manuscript. ## References * IEEE [2000] IEEE Standard for Modeling and Simulation (M&S) High Level Architecture (HLA) – Framework and Rules. IEEE Std 1516-2000 2000;p. 1–28. * Grützmacher et al. [2020] Grützmacher T, Cojean T, Flegar G, Göbel F, Anzt H. A customized precision format based on mantissa segmentation for accelerating sparse linear algebra. Concurrency and Computation: Practice and Experience 2020;32(15):e5418. https://onlinelibrary.wiley.com/doi/abs/10.1002/cpe.5418, e5418 cpe.5418. * Knuth [2014] Knuth DE. Art of computer programming, volume 2: Seminumerical algorithms. Addison-Wesley Professional; 2014. * Gustafson [2017] Gustafson JL. The End of Error: Unum Computing. CRC Press; 2017. * Darvish Rouhani et al. [2020] Darvish Rouhani B, Lo D, Zhao R, Liu M, Fowers J, Ovtcharov K, et al. Pushing the Limits of Narrow Precision Inferencing at Cloud Scale with Microsoft Floating Point. Advances in Neural Information Processing Systems 2020;33. * Cowlishaw et al. [2001] Cowlishaw MF, Schwarz EM, Smith RM, Webb CF. A decimal floating-point specification. In: Proceedings 15th IEEE Symposium on Computer Arithmetic. ARITH-15 2001 IEEE; 2001. p. 147–154. * Goldberg [1991] Goldberg D. What Every Computer Scientist Should Know about Floating-Point Arithmetic. ACM Comput Surv 1991 Mar;23(1):5–48. https://doi.org/10.1145/103162.103163. * Clinger [1990] Clinger WD. How to Read Floating Point Numbers Accurately. SIGPLAN Not 1990 Jun;25(6):92–101. https://doi.org/10.1145/93548.93557. * Clinger [2004] Clinger WD. How to Read Floating Point Numbers Accurately. SIGPLAN Not 2004 Apr;39(4):360–371. https://doi.org/10.1145/989393.989430. * Gay [1990] Gay DM, Correctly rounded binary-decimal and decimal-binary conversions; 1990. AT&T Bell Laboratories Numerical Analysis Manuscript 90-10. * Abseil [2020] Abseil, charconv Design Notes; 2020. https://abseil.io/about/design/charconv [last checked November 2020]. * Adams [2018] Adams U. Ryū: Fast Float-to-String Conversion. In: Proceedings of the 39th ACM SIGPLAN Conference on Programming Language Design and Implementation PLDI 2018, New York, NY, USA: Association for Computing Machinery; 2018. p. 270–282. https://doi.org/10.1145/3192366.3192369. * Adams [2019] Adams U. Ryu Revisited: Printf Floating Point Conversion. Proc ACM Program Lang 2019 Oct;3(OOPSLA). https://doi.org/10.1145/3360595. * Andrysco et al. [2016] Andrysco M, Jhala R, Lerner S. Printing Floating-Point Numbers: A Faster, Always Correct Method. In: Proceedings of the 43rd Annual ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages POPL ’16, New York, NY, USA: Association for Computing Machinery; 2016. p. 555–567. https://doi.org/10.1145/2837614.2837654. * Burger and Dybvig [1996] Burger RG, Dybvig RK. Printing Floating-Point Numbers Quickly and Accurately. SIGPLAN Not 1996 May;31(5):108–116. https://doi.org/10.1145/249069.231397. * Loitsch [2010] Loitsch F. Printing Floating-Point Numbers Quickly and Accurately with Integers. In: Proceedings of the 31st ACM SIGPLAN Conference on Programming Language Design and Implementation PLDI ’10, New York, NY, USA: Association for Computing Machinery; 2010. p. 233–243. https://doi.org/10.1145/1806596.1806623. * Steele and White [2004] Steele GL, White JL. How to Print Floating-Point Numbers Accurately. SIGPLAN Not 2004 Apr;39(4):372–389. https://doi.org/10.1145/989393.989431. * Bray [2017] Bray T, The JavaScript Object Notation (JSON) Data Interchange Format; 2017. Internet Engineering Task Force, Request for Comments: 8259. https://tools.ietf.org/html/rfc8259. * Fisher and Dietz [1998] Fisher RJ, Dietz HG. Compiling for SIMD within a register. In: International Workshop on Languages and Compilers for Parallel Computing Springer; 1998. p. 290–305. * Hars [2006] Hars L. Applications of fast truncated multiplication in cryptography. EURASIP Journal on Embedded Systems 2006;2007(1):061721. * Fousse et al. [2007] Fousse L, Hanrot G, Lefèvre V, Pélissier P, Zimmermann P. MPFR: A multiple-precision binary floating-point library with correct rounding. ACM Transactions on Mathematical Software (TOMS) 2007;33(2):13–es. * Krandick and Johnson [1993] Krandick W, Johnson JR. Efficient multiprecision floating point multiplication with optimal directional rounding. In: Proceedings of IEEE 11th Symposium on Computer Arithmetic IEEE; 1993\. p. 228–233. * Mulders [2000] Mulders T. On short multiplications and divisions. Applicable Algebra in Engineering, Communication and Computing 2000;11(1):69–88. * Lemire et al. [2019] Lemire D, Kaser O, Kurz N. Faster remainder by direct computation: Applications to compilers and software libraries. Software: Practice and Experience 2019;49(6):953–970. * Warren [2013] Warren HS Jr. Hacker’s Delight. 2nd ed. Boston: Addison-Wesley; 2013. * Dutertre [2014] Dutertre B. Yices 2.2. In: International Conference on Computer Aided Verification Springer; 2014\. p. 737–744. * Lemire [2020] Lemire D, Making Your Code Faster by Taming Branches; 2020. https://www.infoq.com/articles/making-code-faster-taming-branches/ [last checked November 2020]. * Langdale and Lemire [2019] Langdale G, Lemire D. Parsing gigabytes of JSON per second. The VLDB Journal 2019;28(6):941–960. * Dumas [2013] Dumas JG. On newton–raphson iteration for multiplicative inverses modulo prime powers. IEEE Transactions on Computers 2013;63(8):2106–2109. ## Appendix A Multiplicative Inverses Given an odd 64-bit integer $x$, there is a unique integer $y$ such that $x\times y\bmod 2^{64}=1$. We refer to $y$ as the _multiplicative inverse_ of $x$ [29]. Fig. 4 presents an efficient C++ function to compute the multiplicative inverse of 64-bit odd integers. It relies on five successive calls to a function involving two integer multiplications. ⬇ uint64_t f64(uint64_t x, uint64_t y) { return y * ( 2 - y * x ); } uint64_t findInverse64(uint64_t x) { uint64_t y = x; y = f64(x,y); y = f64(x,y); y = f64(x,y); y = f64(x,y); y = f64(x,y); return y; } Figure 4: C++ function (findInverse64) to compute the multiplicative inverse of an odd 64-bit integer using Newton’s method ## Appendix B Table Generation Script Fig. 5 provides a convenient Python script to general all relevant reciprocal and normalized powers of five. In practice, each 128-bit value may be stored as two 64-bit words. ⬇ for q in range(-342,-27): power5 = 5**-q z = 0 while( (1<<z) < power5) : z += 1 b = 2 * z + 2 * 64 c = 2 ** b // power5 + 1 while(c >= (1<<128)): c //= 2 print(c) for q in range(-27,0): power5 = 5**-q z = 0 while( (1<<z) < power5) : z += 1 b = z + 127 c = 2 ** b // power5 + 1 print(c) for q in range(0,308+1): power5 = 5**q while(power5 < (1<<127)) : power5 *= 2 while(power5 >= (1<<128)): power5 //= 2 print(power5) Figure 5: Python script to print out all 128-bit reciprocals ($q\in[-342,0)$) and all 128-bit truncated powers of five ($q\in[0,308)$). ## Appendix C Hexadecimal Floating-Point Numbers It could be convenient to represent floating-point numbers using the hexadecimal floating-point notation. The hexadecimal notation may provide an exact ASCII string representation of the binary floating-point number. It makes it relatively easy to provide an unambiguous string that should always be parsed to the same binary value. Furthermore, the parsing and serialization speeds could be much higher. The main downsides are that human beings may find such strings harder to understand and that they are not natively supported in all mainstream programming languages. The hexadecimal floating-point notation is supported in the C (C99), C++ (C++17), Swift, Java, Julia and Go programming languages. As in the usual hexadecimal notation for integers, we start the string with 0x followed by the significand in hexadecimal form. Each hexadecimal character (0–9, A–F) represents 4 bits (a _nibble_). Instead of writing the exponential part in full (e.g., $\times 2^{4}$ or $\times 2^{-4}$), we append the suffix p followed by the exponent (e.g., p4 or p-4). Optionally, we can add an hexadecimal point in the significand. With a decimal point, we interpret the decimal fraction by dividing it by the appropriate power of ten. E.g, we write $1.45=145/10^{2}$. The hexadecimal point works similarly. Thus 0x1.FCp17 means $\mathtt{0x1FC}/4^{2}\times 2^{17}$ or $3.968\,75$ where we divide $\mathtt{0x1FC}$ by $4^{2}$ because there are two nibbles after the binary point. When the value is a normal 64-bit floating-point number, the significand can be expressed as a most significant 1 followed by up to 52 bits, or 13 hexadecimal character. Thus $9\,000\,000\,000\,000\,000$ can be written as 0x1.ff973cafa8p+52. The mass of the Earth in kilogram ($5.972\times 10^{24}$) is 0x1.3c27b13272fb6p+82. ## Appendix D String-Parsing Functions in C++ Fig. 6 illustrates the computation of the decimal significand with pseudo-C++ code. We omit the code necessary to check whether there are leading spaces and sign characters (+ or -) and other error checks. We must further parse an eventual exponent preceded by the characters e or E. Moreover, we must also check whether we had more than 19 digits in the decimal significand. Thus our actual code is slightly more complex. Fig. 7 presents SWAR [19] functions to check all at once whether a sequence of 8 digits is available and to compute the corresponding decimal integer. ⬇ const char *p = // points at the beginning of the string const char *pend = // points at the end of the string int64_t exponent = 0; // exponent uint64_t i = 0; // significand while ((p != pend) && is_integer(*p)) { i = 10 * i + uint64_t(*p - ’0’); ++p; } if ((p != pend) && (*p == ’.’)) { ++p; const char *first_after_period = p; if ((p + 8 <= pend) && is_made_of_eight_digits(p)) { i = i * 100000000 + parse_eight_digits(p); p += 8; if ((p + 8 <= pend) && is_made_of_eight_digits(p)) { i = i * 100000000 + parse_eight_digits(p); p += 8; } } while ((p != pend) && is_integer(*p)) { uint8_t digit = uint8_t(*p - ’0’); ++p; i = i * 10 + digit; } exponent = first_after_period - p; } Figure 6: Simplified pseudo-C++ code to compute the decimal significand from an ASCII string ⬇ bool is_made_of_eight_digits(const char *chars) { uint64_t val; memcpy(&val, chars, 8); return !((((val + 0x4646464646464646) | (val - 0x3030303030303030)) & 0x8080808080808080)); } uint32_t parse_eight_digits(const char *chars) { uint64_t val; memcpy(&val, chars, sizeof(uint64_t)); val = (val & 0x0F0F0F0F0F0F0F0F) * 2561 >> 8; val = (val & 0x00FF00FF00FF00FF) * 6553601 >> 16; return uint32_t((val & 0x0000FFFF0000FFFF) * 42949672960001 >> 32); } Figure 7: C++ functions to check whether 8 ASCII characters are made of digits, and to convert them to an integer value under a little-endian system ## Appendix E Benchmarks in Rust Our C++ implementation and benchmarks have been ported to Rust by I. Smirnov.121212https://github.com/aldanor/fast-float-rust Unlike our C++ implementation, it does not attempt to skip leading white spaces, but there are otherwise few differences. This Rust port allows us to compare against a popular Rust number processing library (lexical131313https://docs.rs/lexical/5.2.0/lexical/—v5.2.0) as well as the standard Rust library (from_str). Using Rust 1.49 on our AMD Rome (Zen 2) system, we get the following results on the canada dataset: the standard Rust library is limited to $92\text{\,}\mathrm{MiB}\text{/}\mathrm{s}$, lexical library achieves $280\text{\,}\mathrm{MiB}\text{/}\mathrm{s}$ while the Rust port of our library achieves $670\text{\,}\mathrm{MiB}\text{/}\mathrm{s}$. On the Apple M1 system, we get $130\text{\,}\mathrm{MiB}\text{/}\mathrm{s}$ (standard library), $370\text{\,}\mathrm{MiB}\text{/}\mathrm{s}$ (lexical) and $1200\text{\,}\mathrm{MiB}\text{/}\mathrm{s}$ (Rust port). The tests are repeated 1000 times and the difference between the best speed and the median speed is low on our test systems (less than 1%).
# Reproducing kernel Hilbert $C^{*}$-module and kernel mean embeddings Yuka Hashimoto<EMAIL_ADDRESS> NTT Network Service Systems Laboratories, NTT Corporation 3-9-11, Midori-cho, Musashinoshi, Tokyo, 180-8585, Japan / Graduate School of Science and Technology, Keio University 3-14-1, Hiyoshi, Kohoku, Yokohama, Kanagawa, 223-8522, Japan Isao Ishikawa <EMAIL_ADDRESS> Center for Data Science, Ehime University 2-5, Bunkyo-cho, Matsuyama, Ehime, 790-8577, Japan / Center for Advanced Intelligence Project, RIKEN 1-4-1, Nihonbashi, Chuo-ku, Tokyo 103-0027, Japan Masahiro Ikeda <EMAIL_ADDRESS> Center for Advanced Intelligence Project, RIKEN 1-4-1, Nihonbashi, Chuo-ku, Tokyo 103-0027, Japan / Faculty of Science and Technology, Keio University 3-14-1, Hiyoshi, Kohoku, Yokohama, Kanagawa, 223-8522, Japan Fuyuta Komura <EMAIL_ADDRESS> Faculty of Science and Technology, Keio University 3-14-1, Hiyoshi, Kohoku, Yokohama, Kanagawa, 223-8522, Japan / Center for Advanced Intelligence Project, RIKEN 1-4-1, Nihonbashi, Chuo-ku, Tokyo 103-0027, Japan Takeshi Katsura <EMAIL_ADDRESS> Faculty of Science and Technology, Keio University 3-14-1, Hiyoshi, Kohoku, Yokohama, Kanagawa, 223-8522, Japan / Center for Advanced Intelligence Project, RIKEN 1-4-1, Nihonbashi, Chuo-ku, Tokyo 103-0027, Japan Yoshinobu Kawahara <EMAIL_ADDRESS> Institute of Mathematics for Industry, Kyushu University 744, Motooka, Nishi-ku, Fukuoka, 819-0395, Japan / Center for Advanced Intelligence Project, RIKEN 1-4-1, Nihonbashi, Chuo-ku, Tokyo 103-0027, Japan ###### Abstract Kernel methods have been among the most popular techniques in machine learning, where learning tasks are solved using the property of reproducing kernel Hilbert space (RKHS). In this paper, we propose a novel data analysis framework with reproducing kernel Hilbert $C^{*}$-module (RKHM) and kernel mean embedding (KME) in RKHM. Since RKHM contains richer information than RKHS or vector-valued RKHS (vvRKHS), analysis with RKHM enables us to capture and extract structural properties in such as functional data. We show a branch of theories for RKHM to apply to data analysis, including the representer theorem, and the injectivity and universality of the proposed KME. We also show RKHM generalizes RKHS and vvRKHS. Then, we provide concrete procedures for employing RKHM and the proposed KME to data analysis. Keywords: reproducing kernel Hilbert $C^{*}$-module, kernel mean embedding, structured data, kernel PCA, interaction effects ## 1 Introduction Kernel methods have been among the most popular techniques in machine learning (Schölkopf and Smola, 2001), where learning tasks are solved using the property of reproducing kernel Hilbert space (RKHS). RKHS is the space of complex-valued functions equipped with an inner product determined by a positive-definite kernel. One of the important tools with RKHS is kernel mean embedding (KME). In KME, a probability distribution (or measure) is embedded as a function in an RKHS (Smola et al., 2007; Muandet et al., 2017; Sriperumbudur et al., 2011), which enables us to analyze distributions in RKHSs. Whereas much of the classical literature on RKHS approaches has focused on complex-valued functions, RKHSs of vector-valued functions, i.e., vector- valued RKHSs (vvRKHSs), have also been proposed (Micchelli and Pontil, 2005; Álvarez et al., 2012; Lim et al., 2015; Minh et al., 2016; Kadri et al., 2016). This allows us to learn vector-valued functions rather than complex- valued functions. In this paper, we develop a branch of theories on reproducing kernel Hilbert $C^{*}$-module (RKHM) and propose a generic framework for data analysis with RKHM. RKHM is a generalization of RKHS and vvRKHS in terms of $C^{*}$-algebra, and we show that RKHM is a powerful tool to analyze structural properties in such as functional data. An RKHM is constructed by a $C^{*}$-algebra-valued positive definite kernel and characterized by a $C^{*}$-algebra-valued inner product (see Definition 2.21). The theory of $C^{*}$-algebra has been discussed in mathematics, especially in operator algebra theory. An important example of $C^{*}$-algebra is $L^{\infty}(\Omega)$, where $\Omega$ is a compact measure space. Another important example is $\mathcal{B}(\mathcal{W})$, which denotes the space of bounded linear operators on a Hilbert space $\mathcal{W}$. Note that $\mathcal{B}(\mathcal{W})$ coincides with the space of matrices $\mathbb{C}^{m\times m}$ if the Hilbert space $\mathcal{W}$ is finite dimensional. Although there are several advantages for studying RKHM compared with RKHS and vvRKHS, those can be summarized into two points as follows: First, an RKHM is a “Hilbert $C^{*}$-module”, which is mathematically more general than a “Hilbert space”. The inner product in an RKHM is $C^{*}$-algebra-valued, which captures more information than the complex-valued one in an RKHS or vvRKHS and enables us to extract richer information. For example, if we set $L^{\infty}(\Omega)$ as a $C^{*}$-algebra, we can control and extract features of functional data such as derivatives, total variation, and frequency components. Also, if we set $\mathcal{B}(\mathcal{W})$ as a $C^{*}$-algebra and the inner product is described by integral operators, we can control and extract features of continuous relationships between pairs of functional data. This cannot be achieved, in principle, by RKHSs and vv-RKHSs. This is because their inner products are complex-valued, where such information degenerates into one complex value or is lost by discretizations of function into complex values. Therefore, we cannot reconstruct the information from a vector in an RKHS or vvRKHS. Second, RKHM generalizes RKHS and vvRKHS, that is, it can be shown that we can reconstruct RKHSs and vvRKHSs from RKHMs. This implies that existing algorithms with RKHSs and vvRKHSs are reconstructed by using the framework of RKHM. The theory of RKHM has been studied in mathematical physics and pure mathematics (Itoh, 1990; Heo, 2008; Szafraniec, 2010). On the other hand, to the best of our knowledge, as for the application of RKHM to data analysis, we can find the only literature by Ye (2017), where only the case of setting the space of matrices as a $C^{*}$-algebra is discussed. In this paper, we develop a branch of theories on RKHM and propose a generic framework for data analysis with RKHM. We show a theoretical property on minimization with respect to orthogonal projections and give a representer theorem in RKHMs. These properties are fundamental for data analysis that have been investigated and applied in the cases of RKHS and vvRKHS, which has made RKHS and vvRKHS widely-accepted tools for data analysis (Schölkopf et al., 2001). Moreover, we define a KME in an RKHM, and provide theoretical results about the injectivity of the proposed KME and the connection with universality of RKHM. Note that, as is well known for RKHSs, these two properties have been actively studied to theoretically guarantee the validity of kernel-based algorithms (Steinwart, 2001; Gretton et al., 2006; Fukumizu et al., 2007; Sriperumbudur et al., 2011). Then, we apply the developed theories to generalize kernel PCA (Schölkopf and Smola, 2001), analyze time-series data with the theory of dynamical system, and analyze interaction effects for infinite dimensional data. The remainder of this paper is organized as follows. First, in Section 2, we briefly review RKHS, vvRKHS, and the definition of RKHM. In Section 3, we provide an overview of the motivation of studying RKHM for data analysis. In Section 4, we show general properties of RKHM for data analysis and the connection of RKHMs with RKHSs and vvRKHSs. In Sections 5, we propose a KME in RKHMs, and show the connection between the injectivity of the KME and the universality of RKHM. Then, in Section 6, we discuss applications of the developed results to kernel PCA, time-series data analysis, and the analysis of interaction effects in finite or infinite dimensional data. Finally, in Section 7, we discuss the connection of RKHMs and the proposed KME with the existing notions, and conclude the paper in Section 8. ##### Notations Lowercase letters denote $\mathcal{A}$-valued coefficients (often by $a,b,c,d$), vectors in a Hilbert $C^{*}$-module $\mathcal{M}$ (often by $p,q,u,v$), or vectors in a Hilbert space $\mathcal{W}$ (often by $w,h$). Lowercase Greek letters denote measures (often by $\mu,\nu,\lambda$) or complex-valued coefficients (often by $\alpha,\beta$). Calligraphic capital letters denote sets. And, bold lowercase letters denote vectors in $\mathcal{A}^{n}$ for $n\in\mathbb{N}$ (a finite dimensional Hilbert $C^{*}$-module). Also, we use $\sim$ for objects related to RKHSs. Moreover, an inner product, an absolute value, and a norm in a space or a module $\mathcal{S}$ (see Definitions 2.12 and 2.13) are denoted as $\left\langle\cdot,\cdot\right\rangle_{\mathcal{S}}$, $|\cdot|_{\mathcal{S}}$, and $\|\cdot\|_{\mathcal{S}}$, respectively. The typical notations in this paper are listed in Table 1. Table 1: Notation table $\mathcal{A}$ | A $C^{*}$-algebra ---|--- $1_{\mathcal{A}}$ | The multiplicative identity in $\mathcal{A}$ $\mathcal{A}_{+}$ | The subset of $\mathcal{A}$ composed of all positive elements in $\mathcal{A}$ $\leq_{\mathcal{A}}$ | For $c,d\in\mathcal{A}$, $c\leq_{\mathcal{A}}d$ means $d-c$ is positive. $<_{\mathcal{A}}$ | For $c,d\in\mathcal{A}$, $c<d$ means $d-c$ is strictly positive, i.e., $d-c$ is positive and invertible. $L^{\infty}(\Omega)$ | The space of complex-valued $L^{\infty}$ functions on a measure space $\Omega$ $\mathcal{B}(\mathcal{W})$ | The space of bounded linear operators on a Hilbert space $\mathcal{W}$ $\mathbb{C}^{m\times m}$ | A set of all complex-valued $m\times m$ matrix $\mathcal{M}$ | A Hilbert $\mathcal{A}$-module $\mathcal{X}$ | A nonempty set for data $C(\mathcal{X},\mathcal{Y})$ | The space of $\mathcal{Y}$-valued continuous functions on $\mathcal{X}$ for topological spaces $\mathcal{X}$ and $\mathcal{Y}$ $n$ | A natural number that represents the number of samples $k$ | An $\mathcal{A}$-valued positive definite kernel $\phi$ | The feature map endowed with $k$ $\mathcal{M}_{k}$ | The RKHM associated with $k$ $\mathcal{S}^{\mathcal{X}}$ | The set of all functions from a set $\mathcal{X}$ to a space $\mathcal{S}$ $\tilde{k}$ | A complex-valued positive definite kernel $\tilde{\phi}$ | The feature map endowed with $\tilde{k}$ $\mathcal{H}_{\tilde{k}}$ | The RKHS associated with $\tilde{k}$ $\mathcal{H}_{k}^{\operatorname{v}}$ | The vvRKHS associated with $k$ $\mathcal{D}(\mathcal{X},\mathcal{A})$ | The set of all $\mathcal{A}$-valued finite regular Borel measures $\Phi$ | The proposed KME in an RKHM $\delta_{x}$ | The $\mathcal{A}$-valued Dirac measure defined as $\delta_{x}(E)=1_{\mathcal{A}}$ for $x\in E$ and $\delta_{x}(E)=0$ for $x\notin E$ $\tilde{\delta}_{x}$ | The complex-valued Dirac measure defined as $\tilde{\delta}_{x}(E)=1$ for $x\in E$ and $\tilde{\delta}_{x}(E)=0$ for $x\notin E$ $\chi_{E}$ | The indicator function of a Borel set $E$ on $\mathcal{X}$ ${C}_{0}(\mathcal{X},\mathcal{A})$ | The space of all continuous $\mathcal{A}$-valued functions on $\mathcal{X}$ vanishing at infinity $\mathbf{G}$ | The $\mathcal{A}$-valued Gram matrix defined as $\mathbf{G}_{i,j}=k(x_{i},x_{j})$ for given samples $x_{1},\ldots,x_{n}\in\mathcal{X}$ $p_{j}$ | The $j$-th principal axis generated by kernel PCA with an RKHM $r$ | A natural number that represents the number of principal axes $Df_{\mathbf{c}}$ | The Gâteaux derivative of a function $f:\mathcal{M}\to\mathcal{A}$ at $\mathbf{c}\in\mathcal{M}$ $\nabla f_{\mathbf{c}}$ | The gradient of a function $f:\mathcal{M}\to\mathcal{A}$ at $\mathbf{c}\in\mathcal{M}$ ## 2 Background We briefly review RKHS and vvRKHS in Subsections 2.1 and 2.2, respectively. Then, we review $C^{*}$-algebra and $C^{*}$-module in Subsection 2.3, Hilbert $C^{*}$-module in Subsection 2.4, and RKHM in Subsection 2.5. ### 2.1 Reproducing kernel Hilbert space (RKHS) We review the theory of RKHS. An RKHS is a Hilbert space and useful for extracting nonlinearity or higher-order moments of data (Schölkopf and Smola, 2001; Saitoh and Sawano, 2016). We begin by introducing positive definite kernels. Let $\mathcal{X}$ be a non- empty set for data, and $\tilde{k}$ be a positive definite kernel, which is defined as follows: ###### Definition 2.1 (Positive definite kernel) A map $\tilde{k}:\mathcal{X}\times\mathcal{X}\to\mathbb{C}$ is called a positive definite kernel if it satisfies the following conditions: 1. 1. $\tilde{k}(x,y)=\overline{\tilde{k}(y,x)}$ for $x,y\in\mathcal{X}$, 2. 2. $\sum_{i,j=1}^{n}\overline{\alpha}_{i}\alpha_{j}\tilde{k}(x_{i},x_{j})\geq 0$ for $n\in\mathbb{N}$, $\alpha_{i}\in\mathbb{C}$, $x_{i}\in\mathcal{X}$. Let $\tilde{\phi}:\mathcal{X}\to\mathbb{C}^{\mathcal{X}}$ be a map defined as $\tilde{\phi}(x)=\tilde{k}(\cdot,x)$. With $\tilde{\phi}$, the following space as a subset of $\mathbb{C}^{\mathcal{X}}$ is constructed: $\mathcal{H}_{\tilde{k},0}:=\bigg{\\{}\sum_{i=1}^{n}\alpha_{i}\tilde{\phi}(x_{i})\bigg{|}\ n\in\mathbb{N},\ \alpha_{i}\in\mathbb{C},\ x_{i}\in\mathcal{X}\bigg{\\}}.$ Then, a map $\left\langle\cdot,\cdot\right\rangle_{\mathcal{H}_{\tilde{k}}}:\mathcal{H}_{\tilde{k},0}\times\mathcal{H}_{\tilde{k},0}\to\mathbb{C}$ is defined as follows: $\bigg{\langle}\sum_{i=1}^{n}\alpha_{i}\tilde{\phi}(x_{i}),\sum_{j=1}^{l}\beta_{j}\tilde{\phi}(y_{j})\bigg{\rangle}_{\mathcal{H}_{\tilde{k}}}:=\sum_{i=1}^{n}\sum_{j=1}^{l}\overline{\alpha_{i}}\beta_{j}\tilde{k}(x_{i},y_{j}).$ By the properties in Definition 2.1 of $\tilde{k}$, $\left\langle\cdot,\cdot\right\rangle_{\mathcal{H}_{\tilde{k}}}$ is well- defined, satisfies the axiom of inner products, and has the reproducing property, that is, $\langle\tilde{\phi}(x),v\rangle_{\mathcal{H}_{\tilde{k}}}=v(x)$ for $v\in\mathcal{H}_{\tilde{k},0}$ and $x\in\mathcal{X}$. The completion of $\mathcal{H}_{\tilde{k},0}$ is called the RKHS associated with $\tilde{k}$ and denoted as $\mathcal{H}_{\tilde{k}}$. It can be shown that $\left\langle\cdot,\cdot\right\rangle_{\mathcal{H}_{\tilde{k}}}$ is extended continuously to $\mathcal{H}_{\tilde{k}}$ and the map $\mathcal{H}_{\tilde{k}}\ni v\mapsto(x\mapsto\langle\tilde{\phi}(x),v\rangle_{\mathcal{H}_{\tilde{k}}})\in\mathbb{C}^{\mathcal{X}}$ is injective. Thus, $\mathcal{H}_{\tilde{k}}$ is regarded to be a subset of $\mathbb{C}^{\mathcal{X}}$ and has the reproducing property. Also, $\mathcal{H}_{\tilde{k}}$ is determined uniquely. The map $\tilde{\phi}$ maps data into $\mathcal{H}_{\tilde{k}}$ and is called the feature map. Since the dimension of $\mathcal{H}_{\tilde{k}}$ is higher (often infinite dimensional) than that of $\mathcal{X}$, complicated behaviors of data in $\mathcal{X}$ are expected to be transformed into simple ones in $\mathcal{H}_{\tilde{k}}$ (Schölkopf and Smola, 2001). ### 2.2 Vector-valued RKHS (vvRKHS) We review the theory of vvRKHS. Complex-valued functions in RKHSs are generalized to vector-valued functions in vvRKHSs. Similar to the case of RKHS, we begin by introducing positive definite kernels. Let $\mathcal{X}$ be a non-empty set for data and $\mathcal{W}$ be a Hilbert space. In addition, let $k$ be an operator-valued positive definite kernel, which is defined as follows: ###### Definition 2.2 (Operator-valued positive definite kernel) A map $k:\mathcal{X}\times\mathcal{X}\to\mathcal{B}(\mathcal{W})$ is called an operator-valued positive definite kernel if it satisfies the following conditions: 1. 1. $k(x,y)=k(y,x)^{*}$ for $x,y\in\mathcal{X}$, 2. 2. $\sum_{i,j=1}^{n}\left\langle w_{i},k(x_{i},x_{j})w_{j}\right\rangle_{\mathcal{W}}\geq 0$ for $n\in\mathbb{N}$, $w_{i}\in\mathcal{W}$, $x_{i}\in\mathcal{X}$. Here, ∗ represents the adjoint. Let $\phi:\mathcal{X}\to\mathcal{B}(\mathcal{W})^{\mathcal{X}}$ be a map defined as $\phi(x)=k(\cdot,x)$. With $\phi$, the following space as a subset of $\mathcal{W}^{\mathcal{X}}$ is constructed: $\mathcal{H}_{k,0}^{\operatorname{v}}:=\bigg{\\{}\sum_{i=1}^{n}\phi(x_{i})w_{i}\bigg{|}\ n\in\mathbb{N},\ w_{i}\in\mathcal{W},\ x_{i}\in\mathcal{X}\bigg{\\}}.$ Then, a map $\left\langle\cdot,\cdot\right\rangle_{\mathcal{H}_{k}^{\operatorname{v}}}:\mathcal{H}_{k,0}^{\operatorname{v}}\times\mathcal{H}_{k,0}^{\operatorname{v}}\to\mathbb{C}$ is defined as follows: $\displaystyle\bigg{\langle}\sum_{i=1}^{n}\phi(x_{i})w_{i},\sum_{j=1}^{l}\phi(y_{j})h_{j}\bigg{\rangle}_{\mathcal{H}_{k}^{\operatorname{v}}}:=\sum_{i=1}^{n}\sum_{j=1}^{l}\left\langle w_{i},k(x_{i},y_{j})h_{j}\right\rangle_{\mathcal{W}}.$ By the properties in Definition 2.2 of $k$, $\left\langle\cdot,\cdot\right\rangle_{\mathcal{H}_{k}^{\operatorname{v}}}$ is well-defined, satisfies the axiom of inner products, and has the reproducing property, that is, $\left\langle\phi(x)w,u\right\rangle_{\mathcal{H}_{k}^{\operatorname{v}}}=\left\langle w,u(x)\right\rangle_{\mathcal{W}}$ (1) for $u\in\mathcal{H}_{k,0}^{{\operatorname{v}}}$, $x\in\mathcal{X}$, and $w\in\mathcal{W}$. The completion of $\mathcal{H}_{k,0}^{\operatorname{v}}$ is called the vvRKHS associated with $k$ and denoted as $\mathcal{H}_{k}^{\operatorname{v}}$. Note that since an inner product in $\mathcal{H}_{k}^{\operatorname{v}}$ is defined with the complex-valued inner product in $\mathcal{W}$, it is complex-valued. ### 2.3 $C^{*}$-algebra and Hilbert $C^{*}$-module A $C^{*}$-algebra and a $C^{*}$-module are generalizations of the space of complex numbers $\mathbb{C}$ and a vector space, respectively. In this paper, we denote a $C^{*}$-algebra by $\mathcal{A}$ and a $C^{*}$-module by $\mathcal{M}$, respectively. As we see below, many complex-valued notions can be generalized to $\mathcal{A}$-valued. A $C^{*}$-algebra is defined as a Banach space equipped with a product structure, an involution $(\cdot)^{*}:\mathcal{A}\rightarrow\mathcal{A}$, and additional properties. We denote the norm of $\mathcal{A}$ by $\|\cdot\|_{\mathcal{A}}$. ###### Definition 2.3 (Algebra) A set $\mathcal{A}$ is called an algebra on a filed $\mathbb{F}$ if it is a vector space equipped with an operation $\cdot:\mathcal{A}\times\mathcal{A}\to\mathcal{A}$ which satisfies the following conditions for $b,c,d\in\mathcal{A}$ and $\alpha\in\mathbb{F}$: $\bullet$ $(b+c)\cdot d={\color[rgb]{0,0,0}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@gray@stroke{0}\pgfsys@color@gray@fill{0}b}\cdot d+c\cdot d$, $\bullet$ $b\cdot(c+d)=b\cdot c+b\cdot d$, $\bullet$ $(\alpha c)\cdot d=\alpha(c\cdot d)=c\cdot(\alpha d)$. The symbol $\cdot$ is omitted when it does not cause confusion. ###### Definition 2.4 ($C^{*}$-algebra) A set $\mathcal{A}$ is called a $C^{*}$-algebra if it satisfies the following conditions: 1. 1. $\mathcal{A}$ is an algebra over $\mathbb{C}$ and equipped with a bijection $(\cdot)^{*}:\mathcal{A}\to\mathcal{A}$ that satisfies the following conditions for $\alpha,\beta\in\mathbb{C}$ and $c,d\in\mathcal{A}$: $\bullet$ $(\alpha c+\beta d)^{*}=\overline{\alpha}c^{*}+\overline{\beta}d^{*}$, $\bullet$ $(cd)^{*}=d^{*}c^{*}$, $\bullet$ $(c^{*})^{*}=c$. 2. 2. $\mathcal{A}$ is a normed space with $\|\cdot\|_{\mathcal{A}}$, and for $c,d\in\mathcal{A}$, $\|cd\|_{\mathcal{A}}\leq\|c\|_{\mathcal{A}}\|d\|_{\mathcal{A}}$ holds. In addition, $\mathcal{A}$ is complete with respect to $\|\cdot\|_{\mathcal{A}}$. 3. 3. For $c\in\mathcal{A}$, $\|c^{*}c\|_{\mathcal{A}}=\|c\|_{\mathcal{A}}^{2}$ holds. ###### Definition 2.5 (Multiplicative identity and unital $C^{*}$-algebra) The multiplicative identity of $\mathcal{A}$ is the element $a\in\mathcal{A}$ which satisfies $ac=ca=c$ for any $c\in\mathcal{A}$. We denote by $1_{\mathcal{A}}$ the multiplicative identity of $\mathcal{A}$. If a $C^{*}$-algebra $\mathcal{A}$ has the multiplicative identity, then it is called a unital $C^{*}$-algebra. ###### Example 2.6 Important examples of (unital) $C^{*}$-algebras are $L^{\infty}(\Omega)$ and $\mathcal{B}(\mathcal{W})$, i.e., the space of complex-valued $L^{\infty}$ functions on a $\sigma$-finite measure space $\Omega$ and the space of bounded linear operators on a Hilbert space $\mathcal{W}$, respectively. 1. 1. For $\mathcal{A}=L^{\infty}(\Omega)$, the product of two functions $c,d\in\mathcal{A}$ is defined as $(cd)(t)=c(t)d(t)$ for any $t\in\Omega$, the involution is defined as $c(t)=\overline{c(t)}$, the norm is the $L^{\infty}$-norm, and the multiplicative identity is the constant function whose value is $1$ at almost everywhere $t\in\Omega$. 2. 2. For $\mathcal{A}=\mathcal{B}(\mathcal{W})$, the product structure is the product (the composition) of operators, the involution is the adjoint, the norm $\|\cdot\|_{\mathcal{A}}$ is the operator norm, and the multiplicative identity is the identity map. In fact, by the Gelfand–Naimark theorem (see, for example, Murphy (1990)), any $C^{*}$-algebra can be regarded as a subalgebra of $\mathcal{B}(\mathcal{W})$ for some Hilbert space $\mathcal{W}$. Therefore, considering the case of $\mathcal{A}=\mathcal{B}(\mathcal{W})$ is sufficient for applications. The positiveness is also important in $C^{*}$-algebras. ###### Definition 2.7 (Positive) An element $c$ of $\mathcal{A}$ is called positive if there exists $d\in\mathcal{A}$ such that $c=d^{*}d$ holds. For a unital $C^{*}$-algebra $\mathcal{A}$, if a positive element $c\in\mathcal{A}$ is invertible, i.e., there exists $d\in\mathcal{A}$ such that $cd=dc=1_{\mathcal{A}}$, then $c$ is called strictly positive. For $c,d\in\mathcal{A}$, we denote $c\leq_{\mathcal{A}}d$ if $d-c$ is positive and $c<_{\mathcal{A}}d$ if $d-c$ is strictly positive. We denote by $\mathcal{A}_{+}$ the subset of $\mathcal{A}$ composed of all positive elements in $\mathcal{A}$. ###### Example 2.8 1. 1. For $\mathcal{A}=L^{\infty}(\Omega)$, a function $c\in\mathcal{A}$ is positive if and only if $c(t)\geq 0$ for almost everywhere $t\in\Omega$, and strictly positive if and only if $c(t)>0$ for almost everywhere $t\in\Omega$. 2. 2. For $\mathcal{A}=\mathcal{B}(\mathcal{W})$, the positiveness is equivalent to the positive semi-definiteness of self-adjoint operators and the strictly positiveness is equivalent to the positive definiteness of self-adjoint operators. The positiveness provides us the (pre) order in $\mathcal{A}$ and, thus, enables us to consider optimization problems in $\mathcal{A}$. ###### Definition 2.9 (Supremum and infimum) 1. 1. For a subset $\mathcal{S}$ of $\mathcal{A}$, $a\in\mathcal{A}$ is said to be an upper bound with respect to the order $\leq_{\mathcal{A}}$, if $d\leq_{\mathcal{A}}a$ for any $d\in\mathcal{S}$. Then, $c\in\mathcal{A}$ is said to be a supremum of $\mathcal{S}$, if $c\leq_{\mathcal{A}}a$ for any upper bound $a$ of $\mathcal{S}$. 2. 2. For a subset $\mathcal{S}$ of $\mathcal{A}$, $a\in\mathcal{A}$ is said to be a lower bound with respect to the order $\leq_{\mathcal{A}}$, if $a\leq_{\mathcal{A}}d$ for any $d\in\mathcal{S}$. Then, $c\in\mathcal{A}$ is said to be a infimum of $\mathcal{S}$, if $a\leq_{\mathcal{A}}c$ for any lower bound $a$ of $\mathcal{S}$. We now introduce a $C^{*}$-module over $\mathcal{A}$, which is a generalization of the vector space. ###### Definition 2.10 (Right multiplication) Let $\mathcal{M}$ be an abelian group with operation $+$. For $c,d\in\mathcal{A}$ and $u,v\in\mathcal{M}$, if an operation $\cdot:\mathcal{M}\times\mathcal{A}\to\mathcal{M}$ satisfies 1. 1. $(u+v)\cdot c=u\cdot c+v\cdot c$, 2. 2. $u\cdot(c+d)=u\cdot c+u\cdot d$, 3. 3. $u\cdot(cd)=(u\cdot{\color[rgb]{0,0,0}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@gray@stroke{0}\pgfsys@color@gray@fill{0}c})\cdot{\color[rgb]{0,0,0}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@gray@stroke{0}\pgfsys@color@gray@fill{0}d}$, 4. 4. $u\cdot 1_{\mathcal{A}}=u$ if $\mathcal{A}$ is unital, then, $\cdot$ is called a (right) $\mathcal{A}$-multiplication. The multiplication $u\cdot c$ is usually denoted as $uc$. ###### Definition 2.11 ($C^{*}$-module) Let $\mathcal{M}$ be an abelian group with operation $+$. If $\mathcal{M}$ is equipped with a (right) $\mathcal{A}$-multiplication, $\mathcal{M}$ is called a (right) $C^{*}$-module over $\mathcal{A}$. In this paper, we consider column vectors rather than row vectors for representing $\mathcal{A}$-valued coefficients, and column vectors act on the right. Therefore, we consider right multiplications. However, considering row vectors and left multiplications instead of column vectors and right multiplications is also possible. ### 2.4 Hilbert $C^{*}$-module A Hilbert $C^{*}$-module is a generalization of a Hilbert space. We first consider an $\mathcal{A}$-valued inner product, which is a generalization of a complex-valued inner product, and then, introduce the definition of a Hilbert $C^{*}$-module. ###### Definition 2.12 ($\mathcal{A}$-valued inner product) A $\mathbb{C}$-linear map with respect to the second variable $\left\langle\cdot,\cdot\right\rangle_{\mathcal{M}}:\mathcal{M}\times\mathcal{M}\to\mathcal{A}$ is called an $\mathcal{A}$-valued inner product if it satisfies the following properties for $u,v,p\in\mathcal{M}$ and $c,d\in\mathcal{A}$: 1. 1. $\left\langle u,vc+pd\right\rangle_{\mathcal{M}}=\left\langle u,v\right\rangle_{\mathcal{M}}c+\left\langle u,p\right\rangle_{\mathcal{M}}d$, 2. 2. $\left\langle v,u\right\rangle_{\mathcal{M}}=\left\langle u,v\right\rangle_{\mathcal{M}}^{*}$, 3. 3. $\left\langle u,u\right\rangle_{\mathcal{M}}\geq_{\mathcal{A}}0$, 4. 4. If $\left\langle u,u\right\rangle_{\mathcal{M}}=0$ then $u=0$. ###### Definition 2.13 ($\mathcal{A}$-valued absolute value and norm) For $u\in\mathcal{M}$, the $\mathcal{A}$-valued absolute value $|u|_{\mathcal{M}}$ on $\mathcal{M}$ is defined by the positive element $|u|_{\mathcal{M}}$ of $\mathcal{A}$ such that $|u|_{\mathcal{M}}^{2}=\left\langle u,u\right\rangle_{\mathcal{M}}$. The (real-valued) norm $\|\cdot\|_{\mathcal{M}}$ on $\mathcal{M}$ is defined by $\|u\|_{\mathcal{M}}=\big{\|}|u|_{\mathcal{M}}\big{\|}_{\mathcal{A}}$. Since the absolute value $|\cdot|_{\mathcal{M}}$ takes values in $\mathcal{A}$, it behaves more complicatedly. For example, the triangle inequality does not hold for the absolute value. However, it provides us with more information than the norm $\|\cdot\|_{\mathcal{M}}$ (which is real- valued). For example, let $\mathcal{M}={\mathcal{A}=}\mathbb{C}^{m\times m}$, $c=\operatorname{diag}\\{\alpha,0,\ldots,0\\}$, and $d=\operatorname{diag}\\{\alpha,\ldots,\alpha\\}$, where $\alpha\in\mathbb{C}$. Then, $\|c\|_{\mathcal{M}}=\|d\|_{\mathcal{M}}$, but $|c|_{\mathcal{M}}\neq|d|_{\mathcal{M}}$. For a self-adjoint matrix, the absolute value describes the whole spectrum of it, but the norm only describes the largest eigenvalue. ###### Definition 2.14 (Hilbert $C^{*}$-module) Let $\mathcal{M}$ be a (right) $C^{*}$-module over $\mathcal{A}$ equipped with an $\mathcal{A}$-valued inner product defined in Definition 2.12. If $\mathcal{M}$ is complete with respect to the norm $\|\cdot\|_{\mathcal{M}}$, it is called a Hilbert $C^{*}$-module over $\mathcal{A}$ or Hilbert $\mathcal{A}$-module. ###### Example 2.15 A simple example of Hilbert $C^{*}$ modules over $\mathcal{A}$ is $\mathcal{A}^{n}$ for a natural number $n$. The $\mathcal{A}$-valued inner product between $\mathbf{c}=[c_{1},\ldots,c_{n}]^{T}$ and $\mathbf{d}=[d_{1},\ldots,d_{n}]^{T}$ is defined as $\left\langle\mathbf{c},\mathbf{d}\right\rangle_{\mathcal{A}^{n}}=\sum_{i=1}^{n}c_{i}^{*}d_{i}$. The absolute value and norm in $\mathcal{A}^{n}$ are given as $|\mathbf{c}|_{\mathcal{A}^{n}}^{2}=(\sum_{i=1}^{n}c_{i}^{*}c_{i})$ and $\|\mathbf{c}\|_{\mathcal{A}^{n}}=\|\sum_{i=1}^{n}c_{i}^{*}c_{i}\|_{\mathcal{A}}^{1/2}$, respectively. Similar to the case of Hilbert spaces, the following Cauchy–Schwarz inequality for $\mathcal{A}$-valued inner products is available (Lance, 1995, Proposition 1.1). ###### Lemma 2.16 (Cauchy–Schwarz inequality) For $u,v\in\mathcal{M}$, the following inequality holds: $|\left\langle u,v\right\rangle_{\mathcal{M}}|_{\mathcal{A}}^{2}\;\leq_{\mathcal{A}}\|u\|_{\mathcal{M}}^{2}\left\langle v,v\right\rangle_{\mathcal{M}}.$ An important property associated with an inner product is the orthonormality. The orthonormality plays an important role in data analysis. For example, an orthonormal basis constructs orthogonal projections and an orthogonally projected vector minimizes the deviation from its original vector in the projected space. Therefore, we also introduce the orthonormality in Hilbert $C^{*}$-module. See, for example, Definition 1.2 in (Bakić and Guljaš, 2001) for more details. ###### Definition 2.17 (Normalized) A vector $q\in\mathcal{M}$ is normalized if $0\neq\left\langle q,q\right\rangle_{\mathcal{M}}=\left\langle q,q\right\rangle_{\mathcal{M}}^{2}$. Note that in the case of a general $C^{*}$-valued inner product, for a normalized vector $q$, $\left\langle q,q\right\rangle_{\mathcal{M}}$ is not always equal to the identity of $\mathcal{A}$ in contrast to the case of a complex-valued inner product. ###### Definition 2.18 (Orthonormal system and basis) Let $\mathcal{I}$ be an index set. A set $\mathcal{S}=\\{q_{i}\\}_{i\in\mathcal{I}}\subseteq\mathcal{M}$ is called an orthonormal system (ONS) of $\mathcal{M}$ if $q_{i}$ is normalized for any $i\in\mathcal{I}$ and $\left\langle q_{i},q_{j}\right\rangle_{\mathcal{M}}=0$ for $i\neq j$. We call $\mathcal{S}$ an orthonormal basis (ONB) if the module generated by $\mathcal{S}$ is an ONS and dense in $\mathcal{M}$. In Hilbert $C^{*}$-modules, $\mathcal{A}$-linear is often used instead of $\mathbb{C}$-linear. ###### Definition 2.19 ($\mathcal{A}$-linear operator) Let $\mathcal{M}_{1},\mathcal{M}_{2}$ be Hilbert $\mathcal{A}$-modules. A linear map $L:\mathcal{M}_{1}\to\mathcal{M}_{2}$ is referred to as $\mathcal{A}$-linear if it satisfies $L(uc)=(Lu)c\;$ for any $u\in\mathcal{M}$ and $c\in\mathcal{A}$. ###### Definition 2.20 ($\mathcal{A}$-linearly independent) The set $\mathcal{S}$ of $\mathcal{M}$ is said to be $\mathcal{A}$-linearly independent if it satisfies the following condition: For any finite subset $\\{v_{1},\ldots,v_{n}\\}$ of $\mathcal{S}$, if $\sum_{i=1}^{n}v_{i}c_{i}=0$ for $c_{i}\in\mathcal{A}$, then $c_{i}=0$ for $i=1,\ldots,n$. For further details about $C^{*}$-algebra, $C^{*}$-module, and Hilbert $C^{*}$-module, refer to Murphy (1990); Lance (1995). ### 2.5 Reproducing kernel Hilbert $C^{*}$-module (RKHM) We summarize the theory of RKHM, which is discussed, for example, in Heo (2008). Similar to the case of RKHS, we begin by introducing an $\mathcal{A}$-valued generalization of a positive definite kernel on a non-empty set $\mathcal{X}$ for data. ###### Definition 2.21 ($\mathcal{A}$-valued positive definite kernel) An $\mathcal{A}$-valued map $k:\ \mathcal{X}\times\mathcal{X}\to\mathcal{A}$ is called a positive definite kernel if it satisfies the following conditions: 1. 1. $k(x,y)=k(y,x)^{*}$ for $x,y\in\mathcal{X}$, 2. 2. $\sum_{i,j=1}^{n}c_{i}^{*}k(x_{i},x_{j})c_{j}\geq_{\mathcal{A}}0$ for $n\in\mathbb{N}$, $c_{i}\in\mathcal{A}$, $x_{i}\in\mathcal{X}$. ###### Example 2.22 1. 1. Let $\mathcal{X}=C([0,1]^{m})$. Let $\mathcal{A}=L^{\infty}([0,1])$ and let $k:\mathcal{X}\times\mathcal{X}\to\mathcal{A}$ be defined as $k(x,y)(t)=\int_{[0,1]^{m}}{\color[rgb]{0,0,0}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@gray@stroke{0}\pgfsys@color@gray@fill{0}\overline{(t-x(s))}}(t-y(s))ds$ for $t\in[0,1]$. Then, for $x_{1},\ldots,x_{n}\in\mathcal{X}$, $c_{1},\ldots,c_{n}\in\mathcal{A}$ and $t\in[0,1]$, we have $\displaystyle\sum_{i,j=1}^{n}c_{i}^{*}(t)k(x_{i},x_{j})(t)c_{j}(t)$ $\displaystyle=\int_{[0,1]^{m}}\sum_{i,j=1}^{n}{\color[rgb]{0,0,0}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@gray@stroke{0}\pgfsys@color@gray@fill{0}\overline{c_{i}(t)(t-x_{i}(s))}}(t-x_{j}(s))c_{j}(t)ds$ $\displaystyle=\int_{[0,1]^{m}}\sum_{i=1}^{n}{\color[rgb]{0,0,0}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@gray@stroke{0}\pgfsys@color@gray@fill{0}\overline{c_{i}(t)(t-x_{i}(s))}}\sum_{j=1}^{n}(t-x_{j}(s))c_{j}(t)ds\geq 0$ for $t\in[0,1]$. Thus, $k$ is an $\mathcal{A}$-valued positive definite kernel. 2. 2. Let $\mathcal{A}=L^{\infty}([0,1])$ and $k:\mathcal{X}\times\mathcal{X}\to\mathcal{A}$ be defined such that $k(x,y)(t)$ is a complex-valued positive definite kernel for any $t\in[0,1]$. Then, $k$ is an $\mathcal{A}$-valued positive definite kernel. 3. 3. Let $\mathcal{W}$ be a separable Hilbert space and let $\\{e_{i}\\}_{i=1}^{\infty}$ be an orthonormal basis of $\mathcal{W}$. Let $\mathcal{A}=\mathcal{B}(\mathcal{W})$. Let $k_{i}:\mathcal{X}\times\mathcal{X}\to\mathbb{C}$ be a complex-valued positive definite kernel for any $i=1,2,\ldots$. Assume for any $x\in\mathcal{X}$, there exists $C>0$ such that for any $i=1,2,\ldots$, $|k_{i}(x,x)|\leq C$ holds. Let $k:\mathcal{X}\times\mathcal{X}\to\mathcal{A}$ be defined as $k(x,y)e_{i}=k_{i}(x,y)e_{i}$. Then, for $x_{1},\ldots,x_{n}\in\mathcal{X}$, $c_{1},\ldots,c_{n}\in\mathcal{A}$ and $w\in\mathcal{W}$, we have $\displaystyle\bigg{\langle}w,\bigg{(}\sum_{i,j=1}^{n}c_{i}^{*}k(x_{i},x_{j})c_{j}\bigg{)}w\bigg{\rangle}_{\mathcal{W}}$ $\displaystyle=\sum_{i,j=1}^{n}\sum_{l=1}^{\infty}\left\langle\alpha_{i,l}e_{l},k(x_{i},x_{j})\alpha_{j,l}e_{l}\right\rangle_{\mathcal{W}}$ $\displaystyle=\sum_{l=1}^{\infty}\sum_{i,j=1}^{n}\overline{\alpha_{i,l}}\alpha_{j,l}\tilde{k}_{l}(x_{i},x_{j})\geq 0,$ where $c_{i}w=\sum_{l=1}^{\infty}\alpha_{i,l}e_{l}$ is the expansion with respect to $\\{e_{i}\\}_{i=1}^{\infty}$. Thus, $k$ is an $\mathcal{A}$-valued positive definite kernel. 4. 4. Let $\mathcal{X}=C(\Omega,\mathcal{Y})$ and $\mathcal{W}=L^{2}(\Omega)$ for a topological space $\Omega$ with a finite Borel measure and a topological space $\mathcal{Y}$. Let $\mathcal{A}=\mathcal{B}(\mathcal{W})$, and $\tilde{k}:\mathcal{Y}\times\mathcal{Y}\to\mathbb{C}$ be a complex-valued bounded continuous positive definite kernel. Moreover, let $k:\mathcal{X}\times\mathcal{X}\to\mathcal{A}$ be defined as $(k(x,y)w)(s)=\int_{t\in\Omega}\tilde{k}(x(s),y(t))w(t)dt$. Then, for $x_{1},\ldots,x_{n}\in\mathcal{X}$, $c_{1},\ldots,c_{n}\in\mathcal{A}$ and $w\in\mathcal{W}$, we have $\displaystyle\bigg{\langle}w,\bigg{(}\sum_{i,j=1}^{n}c_{i}^{*}k(x_{i},x_{j})c_{j}\bigg{)}w\bigg{\rangle}_{\mathcal{W}}$ $\displaystyle=\int_{t\in\Omega}\int_{s\in\Omega}\sum_{i,j=1}^{n}\overline{d_{i}(s)}\tilde{k}(x_{i}(s),x_{j}(t))d_{j}(t)dsdt\geq 0,$ where $d_{i}=c_{i}w$. Thus, $k$ is an $\mathcal{A}$-valued positive definite kernel. Let $\phi:\mathcal{X}\to\mathcal{A}^{\mathcal{X}}$ be the feature map associated with $k$, which is defined as $\phi(x)=k(\cdot,x)$ for $x\in\mathcal{X}$. Similar to the case of RKHS, we construct the following $C^{*}$-module composed of $\mathcal{A}$-valued functions by means of $\phi$: $\mathcal{M}_{k,0}:=\bigg{\\{}\sum_{i=1}^{n}\phi(x_{i})c_{i}\bigg{|}\ n\in\mathbb{N},\ c_{i}\in\mathcal{A},\ x_{i}\in\mathcal{X}\bigg{\\}}.$ An $\mathcal{A}$-valued map $\left\langle\cdot,\cdot\right\rangle_{\mathcal{M}_{k}}:\mathcal{M}_{k,0}\times\mathcal{M}_{k,0}\to\mathcal{A}$ is defined as follows: $\bigg{\langle}\sum_{i=1}^{n}\phi(x_{i})c_{i},\sum_{j=1}^{l}\phi(y_{j})d_{j}\bigg{\rangle}_{\mathcal{M}_{k}}:=\sum_{i=1}^{n}\sum_{j=1}^{l}c_{i}^{*}k(x_{i},y_{j})d_{j}.$ By the properties in Definition 2.21 of $k$, $\left\langle\cdot,\cdot\right\rangle_{\mathcal{M}_{k}}$ is well-defined and has the reproducing property $\left\langle\phi(x),v\right\rangle_{\mathcal{M}_{k}}=v(x)$ for $v\in\mathcal{M}_{k,0}$ and $x\in\mathcal{X}$. Also, it satisfies the properties in Definition 2.12. As a result, $\left\langle\cdot,\cdot\right\rangle_{\mathcal{M}_{k}}$ is shown to be an $\mathcal{A}$-valued inner product. The reproducing kernel Hilbert $\mathcal{A}$-module (RKHM) associated with $k$ is defined as the completion of $\mathcal{M}_{k,0}$. We denote by $\mathcal{M}_{k}$ the RKHM associated with $k$. Heo (2008) focused on the case where a group acts on $\mathcal{X}$ and investigated corresponding actions on RKHMs. Moreover, he considered the space of operators on Hilbert $\mathcal{A}$-module and proved that for each operator-valued positive definite kernel associated with a group and cocycle, there is a corresponding representation on the Hilbert $C^{*}$-module associated with the positive definite kernel. ## 3 Application of RKHM to functional data In this section, we provide an overview of the motivation for studying RKHM for data analysis. We especially focus on the application of RKHM to functional data. Analyzing functional data has been researched to take advantage of the additional information implied by the smoothness of functions underlying data (Ramsay and Silverman, 2005; Levitin et al., 2007; Wang et al., 2016). By describing data as functions, we obtain information as functions such as derivatives. Applying kernel methods to functional data is also proposed (Kadri et al., 2016). In these frameworks, the functions are assumed to be vectors in a Hilbert space such as $L^{2}(\Omega)$ for a measure space $\Omega$, or they are embedded in an RKHS or vvRKHS. Then, analyses are addressed in these Hilbert spaces. However, since functional data itself is infinite-dimensional data, Hilbert spaces are not always sufficient for extracting its continuous behavior. This is because the inner products in Hilbert spaces are complex-valued, degenerating or failing to capture the continuous behavior of the functional data. We compare algorithms in Hilbert spaces and those in Hilbert $C^{*}$-modules and show advantages of algorithms in Hilbert $C^{*}$-modules over those in Hilbert spaces, which are summarized in Figure 1. We first consider algorithms in Hilbert spaces for analyzing functional data $x_{1},x_{2},\ldots\in C(\Omega,\mathcal{X})$, where $\Omega$ is a $\sigma$-finite measure space and $\mathcal{X}$ is a Hilbert space. There are two possible typical patterns of algorithms in Hilbert spaces. The first pattern (Pattern 1 in Fig. 1) is regarding each function $x_{i}$ as a vector in a Hilbert space $\mathcal{H}$ containing $C(\Omega,\mathcal{X})$. In this case, the inner product $\left\langle x_{i},x_{j}\right\rangle_{\mathcal{H}}$ between two functions $x_{i}$ and $x_{j}$ is single complex-valued although $x_{i}$ and $x_{j}$ are functions. Therefore, information of the value of functions at each point degenerates into a complex value. The second pattern (Pattern 2 in Fig. 1) is discretizing each function $x_{i}$ as $x_{i}(t_{0}),x_{i}(t_{1}),\ldots$ for $t_{0},t_{1},\ldots\in\Omega$ and regarding each discretized value $x_{i}(t_{l})$ as a vector in the Hilbert space $\mathcal{X}$. In this case, we obtain the complex-valued inner product $\left\langle x_{i}(t_{l}),x_{j}(t_{l})\right\rangle_{\mathcal{X}}$ at each point $t_{l}\in\Omega$. However, because of the discretization, continuous behaviors, for example, derivatives, total variation, and frequency components, of the function $x_{i}$ are lost. Algorithms of both patterns in the Hilbert spaces proceed by using the computed complex-valued inner products. As a result, capturing features of functions with the algorithms in the Hilbert spaces is difficult. On the other hand, if we regard each function $x_{i}$ as a vector in a Hilbert $C^{*}$-module $\mathcal{M}$ (the rightmost picture in Fig. 1), then the inner product $\left\langle x_{i},x_{j}\right\rangle_{\mathcal{M}}$ between two functions $x_{i}$ and $x_{j}$ in the Hilbert $C^{*}$-module is $C^{*}$-algebra-valued. Thus, if we set the $C^{*}$-algebra as a function space such as $L^{\infty}(\Omega)$, the inner product $\left\langle x_{i},x_{j}\right\rangle_{\mathcal{M}}$ is function-valued. Therefore, algorithms in Hilbert $C^{*}$-modules enable us to capture and extract continuous behaviors of functions. Moreover, in the case of the outputs are functions, we can control the outputs according to the features of the functions. Since RKHM is a generalization of RKHS and vvRKHS (see Subsection 4.2 for further details), the framework of RKHMs (Hilbert $C^{*}$-modules) allows us to generalize kernel methods in RKHSs and vvRKHSs (Hilbert spaces) to those in Hilbert $C^{*}$-modules. Therefore, by using RKHM, we can capture and extract features of functions in kernel methods. The remainder of this paper is devoted to developing the theory of applying RKHMs to data analysis and showing examples of practical applications of data analysis in RKHMs (PCA, time-series data analysis, and analysis of interaction effects). $x_{2}(t)$$t$$x_{1}(t)$$t$Compute theinner product$\left\langle x_{1},x_{2}\right\rangle_{\mathcal{H}}\in\mathbb{C}$$c_{i}=\left\langle x_{1}(t_{i}),x_{2}(t_{i})\right\rangle\in\mathbb{C}$Degenerates informationalong $t$aa$x_{1}$,$x_{2}$ : Functional data$x_{1},x_{2}\in\mathcal{H}$Algorithms in Hilbert spaces(e.g. RKHSs and vvRKHSs)(Pattern 1) $x_{2}(t)$$t$$x_{1}(t)$$t$$t_{0}$$t_{1}$$t_{2}$$t_{3}$$t_{4}$$t_{5}$$t_{6}$$t_{7}$$c_{0}$$c_{1}$$c_{2}$$c_{3}$$c_{4}$$c_{5}$$c_{6}$$c_{7}$$c_{i}=\left\langle x_{1}(t_{i}),x_{2}(t_{i})\right\rangle_{\mathcal{X}}\in\mathbb{C}$Fails to capturecontinuous behaviors(derivatives, total variation,frequency components,…)Functional dataAlgorithms(e.g. RKHSs)(Pattern 2)$x_{1}(t),x_{2}(t)\in\mathcal{X}$ $x_{2}(t)$$t$$x_{1}(t)$$t$$\left\langle x_{1},x_{2}\right\rangle(t)$$t$$\left\langle x_{1},x_{2}\right\rangle_{\mathcal{M}}\in\mathcal{A}$Capture and controlcontinuous behaviorsFunctional dataAlgorithms inHilbert $C^{*}$-modules(e.g. RKHMs)$x_{1},x_{2}\in\mathcal{M}$ Figure 1: Advantages of algorithms in Hilbert $C^{*}$-modules over those in Hilbert spaces ## 4 RKHM for data analysis As we mentioned in Section 1, RKHM has been studied in mathematical physics and pure mathematics. In existing studies, mathematical properties of RKHM such as the relationship between group actions and RKHMs (see the last paragraph of Subsection 2.5) have been discussed. However, these studies have not been focused on data and algorithms for analyzing it. Therefore, we fill the gaps between the existing theory of RKHM and its application to data analysis in this section. We develop theories for the validity to applying it to data analysis in Subsection 4.1. Also, we investigate the connection of RKHM with RKHS and vvRKHS in Subsection 4.2. Generalizations of theories of Hilbert space and RKHS are quite nonobvious for general $C^{*}$-algebras since fundamental properties in Hilbert spaces such as the Riesz representation theorem and orthogonal complementedness are not always obtained in Hilbert $C^{*}$-modules. Therefore, we consider limiting $C^{*}$-algebras to an appropriate class of $C^{*}$-algebras. In fact, von Neumann-algebras satisfy desired properties. ###### Definition 4.1 (von Neumann-algebra) A C*-algebra $\mathcal{A}$ is called a von Neumann-algebra if $\mathcal{A}$ is isomorphic to the dual Banach space of some Banach space. The following propositions are fundamental for deriving useful properties for data analysis in Hilbert $C^{*}$-modules and RKHMs (Skeide, 2000, Theorem 4.16), (Manuilov and Troitsky, 2000, Proposition 2.3.3). ###### Proposition 4.2 (The Riesz representation theorem for Hilbert $\mathcal{A}$-modules) Let $\mathcal{A}{\color[rgb]{0,0,0}\definecolor[named]{pgfstrokecolor}{rgb}{0,0,0}\pgfsys@color@gray@stroke{0}\pgfsys@color@gray@fill{0}\subseteq\mathcal{B}(\mathcal{W})}$ be a von Neumann-algebra and let $\mathcal{M}$ be a Hilbert $\mathcal{A}$-module. Let $\mathcal{H}=\mathcal{M}\otimes_{\mathcal{A}}\mathcal{W}$ (see Definition 4.12 for the definition of the product $\mathcal{M}\otimes_{\mathcal{A}}\mathcal{W}$). Then, every $v\in\mathcal{M}$ can be regarded as an operator in $\mathcal{B}(\mathcal{W},\mathcal{H})$, the set of bounded linear operators from $\mathcal{W}$ to $\mathcal{H}$. If $\mathcal{M}\subseteq\mathcal{B}(\mathcal{W},\mathcal{H})$ is strongly closed (in this case, we say that $\mathcal{M}$ is a von Neumann $\mathcal{A}$-module), then for a bounded $\mathcal{A}$-linear map $L:\mathcal{M}\to\mathcal{A}$ (see Definition 2.19), there exists a unique $u\in\mathcal{M}$ such that $Lv=\left\langle u,v\right\rangle_{\mathcal{M}}$ for all $v\in\mathcal{M}$. Let $\mathcal{A}$ be a von Neumann-algebra. We remark that the Hilbert $\mathcal{A}$-module $\mathcal{A}^{n}$ for some $n\in\mathbb{N}$ is a von Neumann $\mathcal{A}$-module. Moreover, for an $\mathcal{A}$-valued positive definite kernel defined as $\tilde{k}1_{\mathcal{A}}$, where $\tilde{k}$ is a (standard) positive definite kernel, the RKHM $\mathcal{M}_{k}$ is a von Neumann $\mathcal{A}$-module. (Generally, the Hilbert $\mathcal{A}$-module represented as $\mathcal{H}\otimes\mathcal{A}$ for a Hilbert space $\mathcal{H}$ is a von Neumann $\mathcal{A}$-module. Here, $\otimes$ represents the tensor product of a Hilbert space and $C^{*}$-module. See Lance (1995, p.6) for further details about the tensor product.) ###### Proposition 4.3 (Orthogonal complementedness in Hilbert $\mathcal{A}$-modules) Let $\mathcal{A}$ be a unital $C^{*}$-algebra and let $\mathcal{M}$ be a Hilbert $\mathcal{A}$-module. Let $\mathcal{V}$ be a finitely (algebraically) generated closed submodule of $\mathcal{M}$. Then, any $u\in\mathcal{M}$ is decomposed into $u=u_{1}+u_{2}$ where $u_{1}\in\mathcal{V}$ and $u_{2}\in\mathcal{V}^{\perp}$. Here, $\mathcal{V}^{\perp}$ is the orthogonal complement of $\mathcal{V}$ defined as $\\{u\in\mathcal{M}\mid\ \left\langle u,v\right\rangle_{\mathcal{M}}=0\\}$. Let $\mathcal{A}$ be a unital $C^{*}$-algebra, let $\mathcal{M}$ be a Hilbert $\mathcal{A}$-module, and let $\\{q_{1},\ldots,q_{n}\\}$ be an ONS of $\mathcal{M}$. Then, the submodule $\mathcal{V}$ generated by $\\{q_{1},\ldots,q_{n}\\}$ is isomorphic to $\bigoplus_{i=1}^{n}\mathcal{V}_{i}$, where $\mathcal{V}_{i}=\\{\left\langle q_{i},q_{i}\right\rangle_{\mathcal{M}}c\mid\ c\in\mathcal{A}\\}$ is a closed submodule of $\mathcal{A}$. Thus, we have $\mathcal{M}=\mathcal{V}\oplus\mathcal{V}^{\perp}$. Therefore, we set $\mathcal{A}$ as a von Neumann-algebra to derive useful properties of RKHM for data analysis. Note that every von Neumann-algebra is unital (see Definition 2.5). ###### Assumption 4.4 We assume $\mathcal{A}$ is a von Neumann-algebra throughout this paper. $C^{*}$-algebras in Example 2.6 are also von Neumann-algebras. As we noted after Example 2.6, any $C^{*}$-algebra can be regarded as a subalgebra of $\mathcal{B}(\mathcal{W})$. Thus, this fact implies setting the range of the positive definite kernel as $\mathcal{B}(\mathcal{W})$ rather than general $C^{*}$-algebras is effective for data analysis. ### 4.1 General properties of RKHM for data analysis #### 4.1.1 Fundamental properties of RKHM Similar to the cases of RKHSs, we show RKHMs constructed by $\mathcal{A}$-valued positive definite kernels have the reproducing property. Also, we show that the RKHM associated with an $\mathcal{A}$-valued positive definite kernel $k$ is uniquely determined. ###### Proposition 4.5 The map $\left\langle\cdot,\cdot\right\rangle_{\mathcal{M}_{k}}$ defined on $\mathcal{M}_{k,0}$ is extended continuously to $\mathcal{M}_{k}$ and the map $\mathcal{M}_{k}\ni v\mapsto(x\mapsto\left\langle\phi(x),v\right\rangle_{\mathcal{M}_{k}})\in\mathcal{A}^{\mathcal{X}}$ is injective. Thus, $\mathcal{M}_{k}$ is regarded to be the subset of $\mathcal{A}^{\mathcal{X}}$ and has the reproducing property. ###### Proposition 4.6 Assume a Hilbert $C^{*}$-module $\mathcal{M}$ over $\mathcal{A}$ and a map $\psi:\mathcal{X}\to\mathcal{M}$ satisfy the following conditions: 1. 1. ${}^{\forall}x,y\in\mathcal{X}$, $\left\langle\psi(x),\psi(y)\right\rangle_{\mathcal{M}}=k(x,y)$ 2. 2. $\overline{\\{\sum_{i=1}^{n}\psi(x_{i})c_{i}\mid\ x_{i}\in\mathcal{X},\ c_{i}\in\mathcal{A}\\}}=\mathcal{M}$ Then, there exists a unique $\mathcal{A}$-linear bijection map $\Psi:\mathcal{M}_{k}\to\mathcal{M}$ that preserves the inner product and satisfies the following commutative diagram: $\textstyle{\mathcal{M}_{k}\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{\Psi}$$\textstyle{\mathcal{M}}$$\textstyle{\mathcal{X}\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces\ignorespaces}$$\scriptstyle{\phi}$$\scriptstyle{\psi}$$\scriptstyle{\circlearrowright}$ We give the proofs for the above propositions in Appendix A. #### 4.1.2 Minimization property and representer theorem in RKHMs We now develop some theories for the validity to apply RKHM to data analysis. First, we show a minimization property of orthogonal projection operators, which is a fundamental property in Hilbert spaces, is also available in Hilbert $C^{*}$-modules. ###### Theorem 4.7 (Minimization property of orthogonal projection operators) Let $\mathcal{I}$ be a finite index set. Let $\\{q_{i}\\}_{i\in\mathcal{I}}$ be an ONS of $\mathcal{M}$ and $\mathcal{V}$ be the submodule of $\mathcal{M}$ spanned by $\\{q_{i}\\}_{i\in\mathcal{I}}$. For $u\in\mathcal{M}_{k}$, let $P:\mathcal{M}\to\mathcal{V}$ be the projection operator defined as $Pu:=\sum_{i\in\mathcal{I}}q_{i}\left\langle q_{i},u\right\rangle_{\mathcal{M}}$. Then $Pu$ is the unique solution of the following minimization problem, where the minimum is taken with respect to a (pre) order in $\mathcal{A}$ (see Definition 2.9): $\min_{v\in\mathcal{V}}|u-v|_{\mathcal{M}}^{2}.$ (2) Proof By Proposition 4.3, $u\in\mathcal{M}$ is decomposed into $u=u_{1}+u_{2}$, where $u_{1}=Pu\in\mathcal{V}$ and $u_{2}=u-u_{1}\in\mathcal{V}^{\perp}$. Let $v\in\mathcal{V}$. Since $u_{1}-v\in\mathcal{V}$, the identity $\left\langle u_{2},u_{1}-v\right\rangle_{\mathcal{M}}=0$ holds. Therefore, we have $|u-v|_{\mathcal{M}}^{2}=|u_{2}+(u_{1}-v)|_{\mathcal{M}}^{2}=|u_{2}|_{\mathcal{M}}^{2}+|u_{1}-v|_{\mathcal{M}}^{2},$ (3) which implies $|u-v|_{\mathcal{M}}^{2}-|u-u_{1}|_{\mathcal{M}}^{2}\geq_{\mathcal{A}}0$. Since $v\in\mathcal{V}$ is arbitrary, $u_{1}$ is a solution of $\min_{v\in\mathcal{V}}|u-v|_{\mathcal{M}}$. Moreover, if there exists $u^{\prime}\in\mathcal{V}$ such that $|u-u_{1}|_{\mathcal{M}}^{2}=|u-u^{\prime}|_{\mathcal{M}}^{2}$, then letting $v=u^{\prime}$ in Eq. (3) derives $|u-u^{\prime}|_{\mathcal{M}}^{2}=|u_{2}|_{\mathcal{M}}^{2}+|u_{1}-u^{\prime}|_{\mathcal{M}}^{2}$, which implies $|u_{1}-u^{\prime}|_{\mathcal{M}}^{2}=0$. As a result, $u_{1}=u^{\prime}$ holds and the uniqueness of $u_{1}$ has been proved. Proposition 4.7 shows the orthogonally projected vector uniquely minimizes the deviation from an original vector in $\mathcal{V}$. Thus, we can generalize methods related to orthogonal projections in Hilbert spaces to Hilbert $C^{*}$-modules. Next, we show the representer theorem in RKHMs. ###### Theorem 4.8 (Representer theorem) Let $x_{1},\ldots,x_{n}\in\mathcal{X}$ and $a_{1},\ldots,a_{n}\in\mathcal{A}$. Let $h:\mathcal{X}\times\mathcal{A}^{2}\to\mathcal{A}_{+}$ be an error function and let $g:\mathcal{A}_{+}\to\mathcal{A}_{+}$ satisfy $g(c)\leq_{\mathcal{A}}g(d)$ for $c\leq_{\mathcal{A}}d$. Assume the module spanned by $\\{\phi(x_{i})\\}_{i=1}^{n}$ is closed. Then, any $u\in\mathcal{M}_{k}$ minimizing $\sum_{i=1}^{n}h(x_{i},a_{i},u(x_{i}))+g(|u|_{\mathcal{M}_{k}})$ admits a representation of the form $\sum_{i=1}^{n}\phi(x_{i})c_{i}$ for some $c_{1},\ldots,c_{n}\in\mathcal{A}$. Proof Let $\mathcal{V}$ be the module spanned by $\\{\phi(x_{i})\\}_{i=1}^{n}$. By Proposition 4.3, $u\in\mathcal{M}_{k}$ is decomposed into $u=u_{1}+u_{2}$, where $u_{1}\in\mathcal{V}$, $u_{2}\in\mathcal{V}^{\perp}$. By the reproducing property of $\mathcal{M}_{k}$, the following equalities are derived for $i=1,\ldots,n$: $\displaystyle u(x_{i})=\left\langle\phi(x_{i}),u\right\rangle_{\mathcal{M}_{k}}=\left\langle\phi(x_{i}),u_{1}+u_{2}\right\rangle_{\mathcal{M}_{k}}=\left\langle\phi(x_{i}),u_{1}\right\rangle_{\mathcal{M}_{k}}.$ Thus, $\sum_{i=1}^{n}h(x_{i},a_{i},u(x_{i}))$ is independent of $u_{2}$. As for the term $g(|u|_{\mathcal{M}_{k}})$, since $g$ satisfies $g(c)\leq_{\mathcal{A}}g(d)$ for $c\leq_{\mathcal{A}}d$, we have $\displaystyle g(|u|_{\mathcal{M}_{k}})=g(|u_{1}+u_{2}|_{\mathcal{M}_{k}})=g\Big{(}\big{(}|u_{1}|_{\mathcal{M}_{k}}^{2}+|u_{2}|_{\mathcal{M}_{k}}^{2}\big{)}^{1/2}\Big{)}\geq_{\mathcal{A}}g(|u_{1}|_{\mathcal{M}_{k}}).$ Therefore, setting $u_{2}=0$ does not affect the term $\sum_{i=1}^{n}h(x_{i},a_{i},u(x_{i}))$, while strictly reducing the term $g(|u|_{\mathcal{M}_{k}})$, which implies any minimizer must have $u_{2}=0$. As a result, any minimizer takes the form $\sum_{i=1}^{n}\phi(x_{i})c_{i}$. ### 4.2 Connection with RKHSs and vvRKHSs We show that the framework of RKHM is more general than those of RKHS and vvRKHS. Let $\tilde{k}$ be a complex-valued positive definite kernel and let $\mathcal{H}_{\tilde{k}}$ be the RKHS associated with $\tilde{k}$. In addition, let $k$ be an $\mathcal{A}$-valued positive definite kernel and $\mathcal{M}_{k}$ be the RKHM associated with $k$. The following proposition is derived by the definitions of RKHSs and RKHMs. ###### Proposition 4.9 (Connection between RKHMs with RKHSs) If $\mathcal{A}=\mathbb{C}$ and $k=\tilde{k}$, then $\mathcal{H}_{\tilde{k}}=\mathcal{M}_{k}$. As for the connection between vvRKHSs and RKHMs, we first remark that in the case of $\mathcal{A}=\mathcal{B}(\mathcal{W})$, Definition 2.21 is equivalent to the operator valued positive definite kernel (Definition 2.2) for the theory of vv-RKHSs. ###### Lemma 4.10 (Connection between Definition 2.21 and Definition 2.2) If $\mathcal{A}=\mathcal{B}(\mathcal{W})$, then, the $\mathcal{A}$-valued positive definite kernel defined in Definition 2.21 is equivalent to the operator valued positive definite kernel defined in Definition 2.2. The proof for Lemma 4.10 is given in Appendix A. Let $\mathcal{A}=\mathcal{B}(\mathcal{W})$ and let $\mathcal{H}_{k}^{\operatorname{v}}$ be the vvRKHS associated with $k$. To investigate further connections between vvRKHSs and RKHMs, we introduce the notion of interior tensor (Lance, 1995, Chapter 4). ###### Proposition 4.11 Let $\mathcal{M}$ be a Hilbert $\mathcal{B}(\mathcal{W})$-module and let $\mathcal{M}\otimes\mathcal{W}$ be the tensor product of $\mathcal{M}$ and $\mathcal{W}$ as vector spaces. The map $\left\langle\cdot,\cdot\right\rangle_{\mathcal{M}\otimes\mathcal{W}}:\mathcal{M}\otimes\mathcal{W}\;\times\;\mathcal{M}\otimes\mathcal{W}\to\mathbb{C}$ defined as $\left\langle v\otimes w,u\otimes h\right\rangle_{\mathcal{M}\otimes\mathcal{W}}=\left\langle w,\left\langle v,u\right\rangle_{\mathcal{M}}h\right\rangle_{\mathcal{W}}$ is a complex-valued pre inner product on $\mathcal{M}\otimes\mathcal{W}$. ###### Definition 4.12 (Interior tensor) The completion of $\mathcal{M}\otimes\mathcal{W}$ with respect to the pre inner product $\left\langle\cdot,\cdot\right\rangle_{\mathcal{M}\otimes\mathcal{W}}$ is referred to as the interior tensor between $\mathcal{M}$ and $\mathcal{W}$, and denoted as $\mathcal{M}\otimes_{\mathcal{B}(\mathcal{W})}\mathcal{W}$. Note that $\mathcal{M}\otimes_{\mathcal{B}(\mathcal{W})}\mathcal{W}$ is a Hilbert space. We now show vvRKHSs are reconstructed by the interior tensor between RKHMs and $\mathcal{W}$. ###### Theorem 4.13 (Connection between RKHMs and vvRKHSs) If $\mathcal{A}=\mathcal{B}(\mathcal{W})$, then two Hilbert spaces $\mathcal{H}_{k}^{\operatorname{v}}$ and $\mathcal{M}\otimes_{\mathcal{B}(\mathcal{W})}\mathcal{W}$ are isomorphic. Theorem 4.13 is derived by the following lemma. ###### Lemma 4.14 There exists a unique unitary map $U\colon\mathcal{M}_{k}\otimes_{\mathcal{B}(\mathcal{W})}\mathcal{W}\to\mathcal{H}_{k}^{\operatorname{v}}$ such that $U(\phi(x)c\otimes w)=\phi(x)(cw)$ holds for all $x\in\mathcal{X}$, $c\in\mathcal{B}(\mathcal{W})$ and $w\in\mathcal{W}$. Proof First, we show that $\displaystyle\bigg{\langle}\sum_{i=1}^{n}\phi({x_{i}})c_{i}\otimes w_{i},\sum_{j=1}^{l}\phi({y_{j}})d_{j}\otimes h_{j}\bigg{\rangle}_{\mathcal{M}_{k}\otimes\mathcal{W}}=\bigg{\langle}\sum_{i=1}^{n}\phi({x_{i}})(c_{i}w_{i}),\sum_{j=1}^{l}\phi({y_{j}})(d_{j}h_{j})\bigg{\rangle}_{\mathcal{H}_{k}^{\operatorname{v}}}$ holds for all $\sum_{i=1}^{n}\phi({x_{i}})c_{i}\otimes w_{i},\sum_{j=1}^{l}\phi({y_{j}})d_{j}\otimes h_{j}\in\mathcal{M}_{k}\otimes_{\mathcal{B}(\mathcal{W})}\mathcal{W}$. This follows from the straightforward calculation. Indeed, we have $\displaystyle\bigg{\langle}\sum_{i=1}^{n}\phi({x_{i}})c_{i}\otimes w_{i},\sum_{j=1}^{l}\phi({y_{j}})d_{j}\otimes h_{j}\bigg{\rangle}_{\mathcal{M}_{k}\otimes\mathcal{W}}=\sum_{i=1}^{n}\sum_{j=1}^{l}\left\langle w_{i},\left\langle\phi(x_{i})c_{i},\phi(y_{j})d_{j}\right\rangle_{k}h_{j}\right\rangle_{\mathcal{W}}$ $\displaystyle\qquad=\sum_{i=1}^{n}\sum_{j=1}^{l}\left\langle w_{i},c_{i}^{*}k(x_{i},y_{j})d_{j}h_{j}\right\rangle_{\mathcal{W}}=\sum_{i=1}^{n}\sum_{j=1}^{l}\left\langle c_{i}w_{i},k(x_{i},y_{j})d_{j}h_{j}\right\rangle_{\mathcal{W}}$ $\displaystyle\qquad=\bigg{\langle}\sum_{i=1}^{n}\phi({x_{i}})(c_{i}w_{i}),\sum_{j=1}^{l}\phi({y_{j}})(d_{j}h_{j})\bigg{\rangle}_{\mathcal{H}_{k}^{\operatorname{v}}}.$ Therefore, by the standard functional analysis argument, it turns out that there exists an isometry $U\colon\mathcal{M}_{k}\otimes_{\mathcal{B}(\mathcal{W})}\mathcal{W}\to\mathcal{H}_{k}^{\operatorname{v}}$ such that $U(\phi(x)c\otimes w)=\phi(x)(cw)$ holds for all $x\in\mathcal{X}$, $c\in\mathcal{B}(\mathcal{W})$ and $w\in\mathcal{W}$. Since the image of $U$ is closed and dense in $\mathcal{H}_{k}^{\operatorname{v}}$, $U$ is surjective. Thus $U$ is a unitary map. ## 5 Kernel mean embedding in RKHM We generalize KME in RKHSs, which is widely used in analyzing distributions, to RKHMs. By using the framework of RKHM, we can embed $\mathcal{A}$-valued measures instead of probability measures (more generally, complex-valued measures). We provide a brief review of $\mathcal{A}$-valued measures and the integral with respect to $\mathcal{A}$-valued measures in Appendix B. We define a KME in RKHMs in Subsection 5.1 and show its theoretical properties in Subsection 5.2. To define a KME by using $\mathcal{A}$-valued measures and integrals, we first define $c_{0}$-kernels. ###### Definition 5.1 (Function space ${C}_{0}(\mathcal{X},\mathcal{A})$) For a locally compact Hausdorff space $\mathcal{X}$, the set of all $\mathcal{A}$-valued continuous functions on $\mathcal{X}$ vanishing at infinity is denoted as $C_{0}(\mathcal{X},\mathcal{A})$. Here, an $\mathcal{A}$-valued continuous function $u$ is said to vanish at infinity if the set $\\{x\in\mathcal{X}\mid\ \|u(x)\|_{\mathcal{A}}\geq\epsilon\\}$ is compact for any $\epsilon>0$. The space ${C}_{0}(\mathcal{X},\mathcal{A})$ is a Banach $\mathcal{A}$-module with respect to the sup norm. Note that if $\mathcal{X}$ is compact, any continuous function is contained in ${C}_{0}(\mathcal{X},\mathcal{A})$. ###### Definition 5.2 ($c_{0}$-kernel) Let $\mathcal{X}$ be a locally compact Hausdorff space. An $\mathcal{A}$-valued positive definite kernel $k:\mathcal{X}\times\mathcal{X}\to\mathcal{A}$ is referred to as a $c_{0}$-kernel if $k$ is bounded and $\phi(x)=k(\cdot,x)\in{C}_{0}(\mathcal{X},\mathcal{A})$ for any $x\in\mathcal{X}$. In this section, we impose the following assumption. ###### Assumption 5.3 We assume $\mathcal{X}$ is a locally compact Hausdorff space and $k$ is an $\mathcal{A}$-valued $c_{0}$-positive definite kernel. In addition, we assume $\mathcal{M}_{k}$ is a von Neumann $\mathcal{A}$-module (see Proposition 4.2). For example, we often consider $\mathcal{X}=\mathbb{R}^{d}$ in practical situations. Also, we provide examples of $c_{0}$-kernels as follows. ###### Example 5.4 1. 1. Let $\mathcal{A}=L^{\infty}([0,1])$ and $k$ is an $\mathcal{A}$-valued positive definite kernel defined such that $k(x,y)(t)$ is a complex-valued positive definite kernel for $t\in[0,1]$ (see Example 2.22.2). Assume there exists a complex-valued $c_{0}$-positive definite kernel $\tilde{k}$ such that for any $t\in[0,1]$, $|k(x,y)(t)|\leq|\tilde{k}(x,y)|$ holds. If $\|k(x,y)\|_{\mathcal{A}}$ is continuous with respect to $y$ for any $x\in\mathcal{X}$, then the inclusion $\\{y\in\mathcal{X}\mid\ \|k(x,y)\|_{\mathcal{A}}\geq\epsilon\\}\subseteq\\{y\in\mathcal{X}\mid\ |\tilde{k}(x,y)|\geq\epsilon\\}$ holds for $x\in\mathcal{X}$ and $\epsilon>0$. Since $\tilde{k}$ is a $c_{0}$-kernel, the set $\\{y\in\mathcal{X}\mid\ |\tilde{k}(x,y)|\geq\epsilon\\}$ is compact (see Definition 5.1). Thus, $\\{y\in\mathcal{X}\mid\ \|k(x,y)\|_{\mathcal{A}}\geq\epsilon\\}$ is also compact and $k$ is an $\mathcal{A}$-valued $c_{0}$-positive definite kernel. Examples of complex-valued $c_{0}$-positive definite kernels are Gaussian, Laplacian and $B_{2n+1}$-spline kernels. 2. 2. Let $\mathcal{W}$ be a separable Hilbert space and let $\\{e_{i}\\}_{i=1}^{\infty}$ be an orthonormal basis of $\mathcal{W}$. Let $\mathcal{A}=\mathcal{B}(\mathcal{W})$ and let $k:\mathcal{X}\times\mathcal{X}\to\mathcal{A}$ be defined as $k(x,y)e_{i}=k_{i}(x,y)e_{i}$, where $k_{i}:\mathcal{X}\times\mathcal{X}\to\mathbb{C}$ is a complex-valued positive definite kernel for any $i=1,2,\ldots$ (see Example 2.22.3). Assume there exists a complex-valued $c_{0}$-positive definite kernel $\tilde{k}$ such that for any $i=1,2,\ldots$, $|k_{i}(x,y)|\leq|\tilde{k}(x,y)|$ holds. If $\|k(x,y)\|_{\mathcal{A}}$ is continuous with respect to $y$ for any $x\in\mathcal{X}$, then $k$ is shown to be an $\mathcal{A}$-valued $c_{0}$-positive definite kernel in the same manner as the above example. We introduce $\mathcal{A}$-valued measure and integral in preparation for defining a KME in RKHMs. They are special cases of vector measure and integral (Dinculeanu, 1967, 2000), respectively. We review vector measure and integral as $\mathcal{A}$-valued ones in Appendix B. The notions of measure and the Lebesgue integral are generalized to $\mathcal{A}$-valued. ### 5.1 Kernel mean embedding of $C^{*}$-algebra-valued measures We now define a KME in RKHMs. ###### Definition 5.5 (KME in RKHMs) Let $\mathcal{D}(\mathcal{X},\mathcal{A})$ be the set of all $\mathcal{A}$-valued finite regular Borel measures. A kernel mean embedding in an RKHM $\mathcal{M}_{k}$ is a map $\Phi:\mathcal{D}(\mathcal{X},\mathcal{A})\rightarrow\mathcal{M}_{k}$ defined by $\Phi(\mu):=\int_{x\in\mathcal{X}}\phi(x)d\mu(x).$ (4) We emphasize that the well-definedness of $\Phi$ is not trivial, and von Neumann $\mathcal{A}$-module is adequate to show it. More precisely, the following theorem derives the well-definedness. ###### Theorem 5.6 (Well-definedness for the KME in RKHMs) Let $\mu\in\mathcal{D}(\mathcal{X},\mathcal{A})$. Then, $\Phi(\mu)\in\mathcal{M}_{k}$. In addition, the following equality holds for any $v\in\mathcal{M}_{k}$: $\left\langle\Phi(\mu),v\right\rangle_{\mathcal{M}_{k}}=\int_{x\in\mathcal{X}}d\mu^{*}(x)v(x).$ (5) To show Theorem 5.6, we use the Riesz representation theorem for Hilbert $\mathcal{A}$-modules (Proposition 4.2). Proof Let $L_{\mu}:\mathcal{M}_{k}\to\mathcal{A}$ be an $\mathcal{A}$-linear map defined as $L_{\mu}v:=\int_{x\in\mathcal{X}}d\mu^{*}(x)v(x)$. The following inequalities are derived by the reproducing property and the Cauchy–Schwarz inequality (Lemma 2.16): $\displaystyle\|L_{\mu}v\|_{\mathcal{A}}$ $\displaystyle\leq\int_{x\in\mathcal{X}}\|v(x)\|_{\mathcal{A}}d|\mu|(x)=\int_{x\in\mathcal{X}}\|\left\langle\phi(x),v\right\rangle_{\mathcal{M}_{k}}\|_{\mathcal{A}}d|\mu|(x)$ $\displaystyle\leq\|v\|_{\mathcal{M}_{k}}\int_{x\in\mathcal{X}}\|\phi(x)\|_{\mathcal{M}_{k}}d|\mu|(x)\leq|\mu|(\mathcal{X})\|v\|_{\mathcal{M}_{k}}\sup_{x\in\mathcal{X}}\|\phi(x)\|_{\mathcal{M}_{k}},$ (6) where the first inequality is easily checked for a step function $s(x):=\sum_{i=1}^{n}c_{i}\chi_{E_{i}}(x)$ as follows: $\displaystyle\bigg{\|}\int_{x\in\mathcal{X}}d\mu^{*}(x)s(x)\bigg{\|}_{\mathcal{A}}$ $\displaystyle=\bigg{\|}\sum_{i=1}^{n}\mu(E_{i})^{*}c_{i}\bigg{\|}_{\mathcal{A}}\leq\sum_{i=1}^{n}\|\mu(E_{i})\|_{\mathcal{A}}\|c_{i}\|_{\mathcal{A}}$ $\displaystyle\leq\sum_{i=1}^{n}|\mu|(E_{i})\|c_{i}\|_{\mathcal{A}}=\int_{x\in\mathcal{X}}\|s(x)\|_{\mathcal{A}}d|\mu|(x).$ Thus, it holds for any totally measurable functions. Since both $|{\mu}|(\mathcal{X})$ and $\sup_{x\in\mathcal{X}}\|\phi(x)\|_{\mathcal{M}_{k}}$ are finite, inequality (6) means $L_{\mu}$ is bounded. Thus, by the Riesz representation theorem for Hilbert $\mathcal{A}$-modules (Proposition 4.2), there exists $u_{\mu}\in\mathcal{M}_{k}$ such that $L_{\mu}v=\left\langle u_{\mu},v\right\rangle_{\mathcal{M}_{k}}$. By setting $v=\phi(y)$, for $y\in\mathcal{X}$, we have $u_{\mu}(y)=L_{\mu}\phi(y)^{*}=\int_{x\in\mathcal{X}}k(y,x)d\mu(x)$. Therefore, $\Phi(\mu)=u_{\mu}\in\mathcal{M}_{k}$ and $\left\langle\Phi(\mu),v\right\rangle_{\mathcal{M}_{k}}=\int_{x\in\mathcal{X}}d\mu^{*}(x)v(x)$. ###### Corollary 5.7 For $\mu,\nu\in\mathcal{D}(\mathcal{X},\mathcal{A})$, the inner product between $\Phi(\mu)$ and $\Phi(\nu)$ is given as follows: $\left\langle\Phi(\mu),\Phi(\nu)\right\rangle_{\mathcal{M}_{k}}=\int_{x\in\mathcal{X}}\int_{y\in\mathcal{X}}d\mu^{*}(x)k(x,y)d\nu(y).$ Moreover, many basic properties for the existing KME in RKHS are generalized to the proposed KME as follows. ###### Proposition 5.8 (Basic properties of the KME $\Phi$) For $\mu,\nu\in\mathcal{D}(\mathcal{X},\mathcal{A})$ and $c\in\mathcal{A}$, $\Phi(\mu+\nu)=\Phi(\mu)+\Phi(\nu)$ and $\Phi(\mu c)=\Phi(\mu)c$ (i.e., $\Phi$ is $\mathcal{A}$-linear, see Definition 2.19) hold. In addition, for $x\in\mathcal{X}$, $\Phi(\delta_{x})=\phi(x)$ (see Definition B.2 for the definition of the $\mathcal{A}$-valued Dirac measure $\delta_{x}$). This is derived from Eqs. (4) and (5). Note that if $\mathcal{A}=\mathbb{C}$, then the proposed KME (4) is equivalent to the existing KME in RKHS considered in Sriperumbudur et al. (2011). ### 5.2 Injectivity and universality Here, we show the connection between the injectivity of the KME and the universality of RKHM. The proofs of the propositions in this subsection are given in Appendix C. #### 5.2.1 injectivity In practice, the injectivity of $\Phi$ is important to transform problems in $\mathcal{D}(\mathcal{X},\mathcal{A})$ into those in $\mathcal{M}_{k}$. This is because if a KME $\Phi$ in an RKHM is injective, then $\mathcal{A}$-valued measures are embedded into $\mathcal{M}_{k}$ through $\Phi$ without loss of information. Note that, for probability measures, the injectivity of the existing KME is also referred to as the “characteristic” property. The injectivity of the existing KME in RKHS has been discussed in, for example, Fukumizu et al. (2007); Sriperumbudur et al. (2010, 2011). These studies give criteria for the injectivity of the KMEs associated with important complex- valued kernels such as transition invariant kernels and radial kernels. Typical examples of these kernels are Gaussian, Laplacian, and inverse multiquadratic kernels. Here, we define the transition invariant kernels and radial kernels for $\mathcal{A}$-valued measures, and generalize their criteria to RKHMs associated with $\mathcal{A}$-valued kernels. To characterize transition invariant kernels, we first define a Fourier transform and support of an $\mathcal{A}$-valued measure. ###### Definition 5.9 (Fourier transform and support of an $\mathcal{A}$-valued measure) For an $\mathcal{A}$-valued measure $\lambda$ on $\mathbb{R}^{d}$, the Fourier transform of $\lambda$, denoted as $\hat{\lambda}$, is defined as $\hat{\lambda}(x)=\int_{\omega\in\mathbb{R}^{d}}e^{-\sqrt{-1}x^{T}\omega}d\lambda(\omega).$ In addition, the support of $\lambda$ is defined as $\operatorname{supp}(\lambda)=\\{x\in\mathbb{R}^{d}\mid\ \lambda(\mathcal{U})>_{\mathcal{A}}0\mbox{ for any open set $\mathcal{U}$ such that }x\in\mathcal{U}\\}.$ ###### Definition 5.10 (Transition invariant kernel and radial kernel) 1. 1. An $\mathcal{A}$-valued positive definite kernel $k:\mathbb{R}^{d}\times\mathbb{R}^{d}\to\mathcal{A}$ is called a transition invariant kernel if it is represented as $k(x,y)=\hat{\lambda}(y-x)$ for a positive $\mathcal{A}$-valued measure $\lambda$. 2. 2. An $\mathcal{A}$-valued positive definite kernel $k:\mathbb{R}^{d}\times\mathbb{R}^{d}\to\mathcal{A}$ is called a radial kernel if it is represented as $k(x,y)=\int_{[0,\infty)}e^{-t\|x-y\|^{2}}d\eta(t)$ for a positive $\mathcal{A}$-valued measure $\eta$. Here, an $\mathcal{A}$-valued measure $\mu$ is said to be positive if $\mu(E)\geq_{\mathcal{A}}0$ for any Borel set $E$. We show transition invariant kernels and radial kernels induce injective KMEs. ###### Proposition 5.11 (The injectivity for transition invariant kernels) Let $\mathcal{A}=\mathbb{C}^{m\times m}$ and $\mathcal{X}=\mathbb{R}^{d}$. Assume $k:\mathcal{X}\times\mathcal{X}\to\mathcal{A}$ is a transition invariant kernel with a positive $\mathcal{A}$-valued measure $\lambda$ that satisfies $\operatorname{supp}(\lambda)=\mathcal{X}$. Then, KME $\Phi:\mathcal{D}(\mathcal{X},\mathcal{A})\to\mathcal{M}_{k}$ defined as Eq. (4) is injective. ###### Proposition 5.12 (The injectivity for radial kernels) Let $\mathcal{A}=\mathbb{C}^{m\times m}$ and $\mathcal{X}=\mathbb{R}^{d}$. Assume $k:\mathcal{X}\times\mathcal{X}\to\mathcal{A}$ is a radial kernel with a positive definite $\mathcal{A}$-valued measure $\eta$ that satisfies $\operatorname{supp}(\eta)\neq\\{0\\}$. Then, KME $\Phi:\mathcal{D}(\mathcal{X},\mathcal{A})\to\mathcal{M}_{k}$ defined as Eq. (4) is injective. ###### Example 5.13 1. 1. If $k:\mathbb{R}^{d}\times\mathbb{R}^{d}\to\mathbb{C}^{m\times m}$ is a matrix-valued kernel whose diagonal elements are Gaussian, Laplacian, or $B_{2n+1}$-spline and nondiagonal elements are $0$, then $k$ is a $c_{0}$-kernel (See Example 2.22.1). There exists a matrix-valued measure $\lambda$ that satisfies $k(x,y)=\hat{\lambda}(y-x)$ and whose diagonal elements are nonnegative and supported by $\mathbb{R}^{d}$ (c.f. Table 2 in Sriperumbudur et al. (2010)) and nondiagonal elements are $0$. Thus, by Proposition 5.11, $\Phi$ is injective. 2. 2. If $k$ is a matrix-valued kernel whose diagonal elements are inverse multiquadratic and nondiagonal elements are $0$, then $k$ is a $c_{0}$-kernel. There exists a matrix-valued measure $\eta$ that satisfies $k(x,y)=\int_{[0,\infty)}e^{-t\|x-y\|^{2}}d\eta(t)$, and whose diagonal elements are nonnegative and $\operatorname{supp}(\eta)\neq\\{0\\}$ and nondiagonal elements are $0$ (c.f. Theorem 7.15 in Wendland (2004)). Thus, by Proposition 5.12, $\Phi$ is injective. #### 5.2.2 Connection with universality Another important property for kernel methods is universality, which ensures that kernel-based algorithms approximate each continuous target function arbitrarily well. For RKHS, Sriperumbudur et al. (2011) showed the equivalence of the injectivity of the existing KME in RKHSs and universality of RKHSs. We define a universality of RKHMs as follows. ###### Definition 5.14 (Universality) An RKHM is said to be universal if it is dense in ${C}_{0}(\mathcal{X},\mathcal{A})$. We show the above equivalence holds also for RKHM in the case of $\mathcal{A}=\mathbb{C}^{m\times m}$. ###### Proposition 5.15 (Equivalence of the injectivity and universality for $\mathcal{A}=\mathbb{C}^{m\times m}$) Let $\mathcal{A}=\mathbb{C}^{m\times m}$. Then, $\Phi:\mathcal{D}(\mathcal{X},\mathcal{A})\to\mathcal{M}_{k}$ is injective if and only if $\mathcal{M}_{k}$ is dense in ${C}_{0}(\mathcal{X},\mathcal{A})$. By Proposition 5.15, if $k$ satisfies the condition in Proposition 5.11 or 5.12, then $\mathcal{M}_{k}$ is universal. For the case where $\mathcal{A}$ is infinite dimensional, the universality of $\mathcal{M}_{k}$ in ${C}_{0}(\mathcal{X},\mathcal{A})$ is a sufficient condition for the injectivity of the proposed KME. ###### Theorem 5.16 (Connection between the injectivity and universality for general $\mathcal{A}$) If $\mathcal{M}_{k}$ is dense in ${C}_{0}(\mathcal{X},\mathcal{A})$, then $\Phi:\mathcal{D}(\mathcal{X},\mathcal{A})\to\mathcal{M}_{k}$ is injective. However, the equivalence of the injectivity and universality, and the injectivity for transition invariant kernels and radial kernels are open problems. This is because their proofs strongly depend on the Hahn–Banach theorem and Riesz–Markov representation theorem, and generalizations of these theorems to $\mathcal{A}$-valued functions and measures are challenging problems due to the situation peculiar to the infinite dimensional spaces. Further details of the proofs of propositions in this section are given in Appendix C. ## 6 Applications We apply the framework of RKHM described in Sections 4 and 5 to problems in data analysis. We propose kernel PCA in RKHMs in Subsection 6.1, time-series data analysis in RKHMs in Subsection 6.2, and analysis of interaction effects in finite or infinite dimensional data with the proposed KME in RKHMs in Subsection 6.3. Then, we discuss further applications in Subsection 6.4. ### 6.1 PCA in RKHMs Principal component analysis (PCA) is a fundamental tool for describing data in a low dimensional space. Its implementation in RKHSs has also been proposed (c.f. Schölkopf and Smola (2001)). It enables us to deal with the nonlinearlity of data by virtue of the high expressive power of RKHSs. Here, we generalize the PCA in RKHSs to capture more information in data, such as multivariate data and functional data, by using the framework of RKHM. ##### Applying RKHM to PCA In the existing framework of PCA in Hilbert spaces, the following reconstruction error is minimized with respect to vectors $p_{1},\ldots,p_{r}$: $\sum_{i=1}^{n}\bigg{\|}x_{i}-\sum_{j=1}^{r}p_{j}\left\langle p_{j},x_{i}\right\rangle\bigg{\|}^{2},$ (7) where $x_{1},\ldots,x_{n}$ are given samples in a Hilbert space and $p_{1},\ldots,p_{r}$ are called principal axes. Here, the complex-valued inner product $\left\langle p_{j},x_{i}\right\rangle$ is the weight with respect to the principal axis $p_{j}$ for representing the sample $x_{i}$. PCA for functional data (functional PCA) has also investigated (Ramsay and Silverman, 2005). For example, in standard functional PCA settings, we set the Hilbert space as $L^{2}(\Omega)$ for a $\sigma$-finite measure space $\Omega$. However, if samples $x_{1},\ldots,x_{n}$ are finite dimensional vectors or functions, Eq. (7) fails to describe their element wise or continuous dependencies on the principal axes. For $d$-dimensional (finite dimensional) vectors, we can just split $x_{i}=[x_{i,1},\ldots,x_{i,d}]$ into $d$ vectors $[x_{i,1},0,\ldots,0],\ldots,[0,\ldots,0,x_{i,d}]$. Then, we can understand which element is dominant for representing $x_{i}$ by using the principal axis $p_{j}$. On the other hand, for functional data, the situation is completely different. For example, assume samples are in $L^{2}(\Omega)$. Since delta functions are not contained in $L^{2}(\Omega)$, we cannot split a sample $x_{i}=x_{i}(t)$ into discrete functions. In this case, how can we understand the continuous dependencies on the principal axes with respect to the variable $t\in\Omega$? One possible way to answer this question is to employ Hilbert $C^{*}$-modules instead of Hilbert spaces. We consider the same type of reconstruction error as Eq. (7) in Hilbert $C^{*}$-modules. In this case, the inner product $\left\langle p_{j},x_{i}\right\rangle_{\mathcal{W}}$ is $C^{*}$-algebra-valued, which allows us to provide more information than the complex-valued one. If we set the $C^{*}$-algebra as the function space on $\Omega$ such as $L^{\infty}(\Omega)$ and define a $C^{*}$-algebra-valued inner product which depends on $t\in\Omega$, then, the weight $\left\langle p_{j},x_{i}\right\rangle_{\mathcal{W}}$ depends on $t$. As a result, we can extract continuous dependencies of samples on the principal axes. More generally, PCA is often considered in an RKHS $\mathcal{H}_{\tilde{k}}$. In this case, $x_{i}$ in Eq. (7) is replaced with $\tilde{\phi}(x_{i})$, where $\tilde{\phi}$ is the feature map, and the inner product and norm are replaced with those in the RKHS. We can extract continuous dependencies of samples on the principal axes by generalizing RKHS to RKHM. #### 6.1.1 Generalization of the PCA in RKHSs to RKHMs Let $x_{1},\ldots,x_{n}\in\mathcal{X}$ be given samples. Let $k:\mathcal{X}\times\mathcal{X}\to\mathcal{A}$ be an $\mathcal{A}$-valued positive definite kernel on $\mathcal{X}$ and let $\mathcal{M}_{k}$ be the RKHM associated with $k$. We explore a useful set of axes $p_{1},\ldots,p_{r}$ in $\mathcal{M}_{k}$, which are referred to as principal axes, to describe the feature of given samples $x_{1},\ldots,x_{n}$. The corresponding components $p_{j}\left\langle p_{j},\phi(x_{i})\right\rangle_{\mathcal{M}_{k}}$ are referred to as principal components. We emphasize our proposed PCA in RKHM provides weights of principal components contained in $\mathcal{A}$, not in complex numbers. This is a remarkable difference between our method and existing PCAs. When samples have some structures such as among variables or in functional data, $\mathcal{A}$-valued weights provide us richer information than complex-valued ones. For example, if $\mathcal{X}$ is the space of functions of multi-variables and if we set $\mathcal{A}$ as $L^{\infty}([0,1])$, then we can reduce multi-variable functional data to functions in $L^{\infty}([0,1])$, functions of single variable (as illustrated in Section 6.1.4). To obtain $\mathcal{A}$-valued weights of principal components, we consider the following minimization problem regarding the following reconstruction error (see Definition 2.18 for the definition of ONS): $\inf_{\\{p_{j}\\}_{j=1}^{r}\subseteq\mathcal{M}_{k}\mbox{\footnotesize: ONS}}\;\sum_{i=1}^{n}\bigg{|}\phi(x_{i})-\sum_{j=1}^{r}p_{j}\left\langle p_{j},\phi(x_{i})\right\rangle_{\mathcal{M}_{k}}\bigg{|}_{\mathcal{M}_{k}}^{2},$ (8) where the infimum is taken with respect to a (pre) order in $\mathcal{A}$ (see Definition 2.9). Since the identity $|\phi(x_{i})-\sum_{j=1}^{r}p_{j}\left\langle p_{j},\phi(x_{i})\right\rangle_{\mathcal{M}_{k}}|_{\mathcal{M}_{k}}^{2}=k(x_{i},x_{i})-\sum_{j=1}^{r}\left\langle\phi(x_{i}),p_{j}\right\rangle_{\mathcal{M}_{k}}\left\langle p_{j},\phi(x_{i})\right\rangle_{\mathcal{M}_{k}}$ holds and $\left\langle\phi(x_{i}),p_{j}\right\rangle_{\mathcal{M}_{k}}$ is represented as $p_{j}(x_{i})$ by the reproducing property, the problem (8) can be reduced to the minimization problem $\inf_{\\{p_{j}\\}_{j=1}^{r}\subseteq\mathcal{M}_{k}\mbox{\footnotesize: ONS}}\;\sum_{i=1}^{n}\sum_{j=1}^{r}-p_{j}(x_{i})p_{j}(x_{i})^{*}.$ (9) In the case of RKHS, i.e., $\mathcal{A}=\mathbb{C}$, the solution of the problem (9) is obtained by computing eigenvalues and eigenvectors of Gram matrices (see, for example, Schölkopf and Smola (2001)). Unfortunately, we cannot extend their procedure to RKHM straightforwardly. Therefore, we develop two methods to obtain approximate solutions of the problem (9): by gradient descents on Hilbert $C^{*}$-modules, and by the minimization of the trace of the $\mathcal{A}$-valued objective function. #### 6.1.2 Gradient descent on Hilbert $C^{*}$-modules We propose a gradient descent method on Hilbert $\mathcal{A}$-module for the case where $\mathcal{A}$ is commutative. An important example of commutative von Neumann-algebra is $L^{\infty}([0,1])$. The gradient descent for a real- valued function on a Hilbert space has been proposed (Smyrlis and Zisis, 2004). However, in our situation, the objective function of the problem (9) is an $\mathcal{A}$-valued function in a Hilbert $C^{*}$-module $\mathcal{A}^{n}$. Thus, the existing gradient descent is not applicable to our situation. Therefore, we generalize the existing gradient descent algorithm to $\mathcal{A}$-valued functions on Hilbert $C^{*}$-modules. Let $\mathcal{A}$ be a commutative von Neumann-algebra. Assume the positive definite kernel $k$ takes its values in $\mathcal{A}_{r}:=\\{c-d\in\mathcal{A}\mid\ c,d\in\mathcal{A}_{+}\\}$. For example, for $\mathcal{A}=L^{\infty}([0,1])$, $\mathcal{A}_{r}$ is the space of real-valued $L^{\infty}$ functions on $[0,1]$. Based on the representer theorem (Theorem 4.8), we find a solution of the problem (9) which is represented as $p_{j}=\sum_{i=1}^{n}\phi(x_{i})c_{j,i}$ for some $c_{j,i}\in\mathcal{A}$. Moreover, since $\mathcal{A}$ is commutative, $p_{j}(x_{i})p_{j}(x_{i})^{*}$ is equal to $p_{j}(x_{i})^{*}p_{j}(x_{i})$. Therefore, the problem (8) on $\mathcal{M}_{k}$ is equivalent to the following problem on the Hilbert $\mathcal{A}$-module $\mathcal{A}^{n}$ (see Example 2.15 about $\mathcal{A}^{n}$): $\inf_{\mathbf{c}_{j}\in\mathcal{A}^{n},\ \\{\sqrt{\mathbf{G}}\mathbf{c}_{j}\\}_{j=1}^{r}\mbox{\footnotesize: ONS}}-\sum_{j=1}^{r}\mathbf{c}_{j}^{*}\mathbf{G}^{2}\mathbf{c}_{j},$ (10) where $\mathbf{G}$ is the $\mathcal{A}$-valued Gram matrix defined as $\mathbf{G}_{i,j}=k(x_{i},x_{j})$. For simplicity, we assume $r=1$, i.e., the number of principal axes is $1$. We rearrange the problem (10) to the following problem by adding a penalty term: $\inf_{\mathbf{c}\in\mathcal{A}^{n}}(-\mathbf{c}^{*}\mathbf{G}^{2}\mathbf{c}+\lambda|\mathbf{c}^{*}\mathbf{Gc}-1_{\mathcal{A}}|_{\mathcal{A}}^{2}),$ (11) where $\lambda$ is a real positive weight for the penalty term. For $r>1$, let $\mathbf{c}_{1}$ be a solution of the problem (10). Then, we solve the same problem in the orthogonal complement of the module spanned by $\\{\mathbf{c}_{1}\\}$ and set the solution of this problem as $\mathbf{c}_{2}$. Then, we solve the same problem in the orthogonal complement of the module spanned by $\\{\mathbf{c}_{1},\mathbf{c}_{2}\\}$ and repeat this procedure to obtain solutions $\mathbf{c}_{1},\ldots\mathbf{c}_{r}$. The problem (11) is the minimization problem of an $\mathcal{A}$-valued function defined on the Hilbert $\mathcal{A}$-module $\mathcal{A}^{n}$. We search a solution of the problem (11) along the steepest descent directions. To calculate the steepest descent directions, we introduce a derivative $Df_{\mathbf{c}}$ of an $\mathcal{A}$-valued function $f$ on a Hilbert $C^{*}$-module at $\mathbf{c}\in\mathcal{M}$. It is defined as the derivative on Banach spaces (c.f. Blanchard and Brüning (2015)). The definition of the derivative is included in Appendix D. The following gives the derivative of the objective function in problem (11). ###### Proposition 6.1 (Derivative of the objective function) Let $f:\mathcal{A}^{n}\to\mathcal{A}$ be defined as $f(\mathbf{c})=-\mathbf{c}^{*}\mathbf{G}^{2}\mathbf{c}+\lambda|\mathbf{c}^{*}\mathbf{G}\mathbf{c}-1_{\mathcal{A}}|_{\mathcal{A}}^{2}.$ (12) Then, $f$ is infinitely differentiable and the first derivative of $f$ is calculated as $Df_{\mathbf{c}}(u)=-2\mathbf{c}^{*}\mathbf{G}^{2}u-4\lambda\mathbf{c}^{*}\mathbf{G}u+4\lambda\mathbf{c}^{*}\mathbf{G}\mathbf{c}\mathbf{c}^{*}\mathbf{G}u.$ Moreover, for each $\mathbf{c}\in\mathcal{A}^{n}$, there exists a unique $\mathbf{d}\in\mathcal{A}^{n}$ such that $\left\langle\mathbf{d},u\right\rangle_{\mathcal{A}^{n}}=Df_{\mathbf{c}}(u)$ for any $u\in\mathcal{A}^{n}$. The vector $\mathbf{d}$ is calculated as $\mathbf{d}=-2\mathbf{G}^{2}\mathbf{c}-4\lambda\mathbf{G}\mathbf{c}+4\lambda\mathbf{G}\mathbf{c}\mathbf{c}^{*}\mathbf{G}\mathbf{c}.$ (13) Proof The derivative of $f$ is calculated by the definition and the assumption that $\mathcal{A}$ is commutative. Since $Df_{\mathbf{c}}$ is a bounded $\mathcal{A}$-linear operator, by the Riesz representation theorem (Proposition 4.2), there exists a unique $\mathbf{d}\in\mathcal{A}^{n}$ such that $\left\langle\mathbf{d},u\right\rangle_{\mathcal{A}^{n}}=Df_{\mathbf{c}}(u)$. ###### Definition 6.2 (Gradient of $\mathcal{A}$-valued functions on Hilbert $C^{*}$-modules) Let $f:\mathcal{M}\to\mathcal{A}$ be a differentiable function. Assume for each $\mathbf{c}\in\mathcal{M}$, there exists a unique $\mathbf{d}\in\mathcal{M}$ such that $\left\langle\mathbf{d},u\right\rangle_{\mathcal{A}^{n}}=Df_{\mathbf{c}}(u)$ for any $u\in\mathcal{M}$. In this case, we denote $\mathbf{d}$ by $\nabla f_{\mathbf{c}}$ and call it the gradient of $f$ at $\mathbf{c}$. We now develop an $\mathcal{A}$-valued gradient descent scheme. ###### Theorem 6.3 Assume $f:\mathcal{M}\to\mathcal{A}$ is differentiable. Moreover, assume there exists $\nabla f_{\mathbf{c}}$ for any $\mathbf{c}\in\mathcal{M}$. Let $\eta_{t}>0$. Let $\mathbf{c}_{0}\in\mathcal{M}$ and $\mathbf{c}_{t+1}=\mathbf{c}_{t}-\eta_{t}\nabla f_{\mathbf{c}_{t}}$ (14) for $t=0,1,\ldots$. Then, we have $f(\mathbf{c}_{t+1})=f(\mathbf{c}_{t})-\eta_{t}|\nabla f_{\mathbf{c}_{t}}|^{2}_{\mathcal{M}}+S(\mathbf{c}_{t},\eta_{t}),$ (15) where $S(\mathbf{c},\eta)$ satisfies $\lim_{\eta\to 0}\|S(\mathbf{c},\eta)\|_{\mathcal{A}}/\eta=0$. The statement is derived by the definition of the derivative (Definition D.1). The following examples show the scheme (14) is valid to solve the problem (11). ###### Example 6.4 Let $\mathcal{A}=L^{\infty}([0,1])$, let $a_{t}=|\nabla f_{\mathbf{c}_{t}}|^{2}_{\mathcal{A}^{n}}\in\mathcal{A}$ and let $b_{t,\eta}=S(\mathbf{c}_{t},\eta)\in\mathcal{A}$. If $a_{t}\geq_{\mathcal{A}}\delta 1_{\mathcal{A}}$ for some positive real value $\delta$, then the function $a_{t}$ on $[0,1]$ satisfies $a_{t}(s)>0$ for almost everywhere $s\in[0,1]$. On the other hand, since $b_{t,\eta}$ satisfies $\lim_{\eta\to 0}\|b_{t,\eta}\|_{\mathcal{A}}/\eta^{2}=0$, there exists sufficiently small positive real value $\eta_{t,0}$ such that for almost everywhere $s\in[0,1]$, $b_{t,\eta_{t,0}}(s)\leq\|b_{t,\eta_{t,0}}\|_{\mathcal{A}}\leq\eta^{2}_{t,0}\delta\leq\eta_{t,0}(1-\xi_{1})\delta$ hold for some positive real value $\xi_{1}$. As a result, $-\eta_{t,0}|\nabla f_{\mathbf{c}_{t}}|^{2}_{\mathcal{A}^{n}}+S(\mathbf{c}_{t},\eta_{t,0})\leq_{\mathcal{A}}-\eta_{t,0}\xi_{1}|\nabla f_{\mathbf{c}_{t}}|^{2}_{\mathcal{A}^{n}}$ holds and by the Eq. (15), we have $f(\mathbf{c}_{t+1})<_{\mathcal{A}}f(\mathbf{c}_{t})$ (16) for $t=0,1,\ldots$. As we mentioned in Example 2.8, the inequality (16) means the function $f(\mathbf{c}_{t+1})\in L^{\infty}([0,1])$ is smaller than the function $f(\mathbf{c}_{t})\in L^{\infty}([0,1])$ at almost every points on $[0,1]$, i.e., $f(\mathbf{c}_{t+1})(s)<f(\mathbf{c}_{t})(s)$ for almost every $s\in[0,1]$. ###### Example 6.5 Assume $\mathcal{A}$ is a finite dimensional space. If $|\nabla f_{\mathbf{c}_{t}}|_{\mathcal{A}}^{2}\geq_{\mathcal{A}}\delta 1_{\mathcal{A}}$ for some positive real value $\delta$, the inequality $f(\mathbf{c}_{t+1})\leq_{\mathcal{A}}f(\mathbf{c}_{t})-\eta_{t}\xi_{1}|\nabla f_{\mathbf{c}_{t}}|^{2}_{\mathcal{A}^{n}}$ holds for $t=0,1,\ldots$ and some $\eta_{t}$ and $\xi_{1}$ in the same manner as Example 6.4. Moreover, the function $f$ defined as Eq. (12) is bounded below and $\nabla f_{\mathbf{c}_{t}}$ is Lipschitz continuous on the set $\\{\mathbf{c}\in\mathcal{A}^{n}\mid\ f(\mathbf{c})\leq_{\mathcal{A}}f(\mathbf{c}_{0})\\}$. In this case, if there exists a positive real value $\xi_{2}$ such that $\|\nabla f_{\mathbf{c}_{t+1}}-\nabla f_{\mathbf{c}_{t}}\|_{\mathcal{A}^{n}}\geq\xi_{2}\|\nabla f_{\mathbf{c}_{t}}\|_{\mathcal{A}^{n}}$, then we have $\xi_{2}\|\nabla f_{\mathbf{c}_{t}}\|_{\mathcal{A}^{n}}\leq L\|\mathbf{c}_{t+1}-\mathbf{c}_{t}\|_{\mathcal{A}^{n}}\leq L\eta_{t}\|\nabla f_{\mathbf{c}_{t}}\|_{\mathcal{A}^{n}},$ where $L$ is a Liptschitz constant of $\nabla f_{\mathbf{c}_{t}}$. As a result, we have $f(\mathbf{c}_{t+1})\leq_{\mathcal{A}}f(\mathbf{c}_{t})-\eta_{t}\xi_{1}|\nabla f_{\mathbf{c}_{t}}|^{2}_{\mathcal{A}^{n}}\leq_{\mathcal{A}}f(\mathbf{c}_{t})-\frac{\xi_{1}\xi_{2}}{L}|\nabla f_{\mathbf{c}_{t}}|^{2}_{\mathcal{A}^{n}},$ which implies $\sum_{t=1}^{T}|\nabla f_{\mathbf{c}_{t}}|^{2}_{\mathcal{A}^{n}}\leq_{\mathcal{A}}L/(\xi_{1}\xi_{2})(f(\mathbf{c}_{1})-f(\mathbf{c}_{T+1}))$. Since $f$ is bounded below, the sum $\sum_{t=1}^{\infty}|\nabla f_{\mathbf{c}_{t}}|^{2}_{\mathcal{A}^{n}}$ converges. Therefore, $|\nabla f_{\mathbf{c}_{t}}|^{2}_{\mathcal{A}^{n}}\to 0$ as $t\to\infty$, i.e., the gradient $\nabla f_{\mathbf{c}_{t}}$ in Eq. (14) converges to $0$. ###### Remark 6.6 It is possible to generalize the above method to the case where the objective function $f$ has the form $f(\mathbf{c})=\mathbf{c}^{*}\mathbf{G}\mathbf{c}$ for $\mathbf{G}\in\mathcal{A}^{n\times n}$ and $\mathcal{A}$ is noncommutative. In this case, the derivative $Df_{\mathbf{c}}$ is calculated as $Df_{\mathbf{c}}(u)=u^{*}\mathbf{G}\mathbf{c}+\mathbf{c}^{*}\mathbf{G}u.$ Therefore, defining the gradient $\nabla f_{\mathbf{c}}$ as $\nabla f_{\mathbf{c}}=\mathbf{G}\mathbf{c}$ results in $Df_{\mathbf{c}}(-\eta\nabla f_{\mathbf{c}})=-2\eta\mathbf{c}^{*}\mathbf{G}^{2}\mathbf{c}\leq_{\mathcal{A}}0$ for a real positive value $\eta$, which allows us to derive the same result as Theorem 6.3. ###### Remark 6.7 The computational complexity of the PCA in RKHMs is higher than the standard PCA in RKHSs. Indeed, in the case of RKHSs, the minimization problem is reduced to an eigenvalue problem of the Gram matrix with respect to given samples. On the other hand, we solve the minimization problem (8) by the gradient descent, and in each iteration step, we compute the gradient $\mathbf{d}$ in Eq. (13). Since the elements of $\mathbf{G}$ and $\mathbf{c}$ are in $\mathcal{A}$, the computation of $\mathbf{d}$ involves the multiplication in $\mathcal{A}$ such as multiplication of functions. Even though we compute the multiplication in $\mathcal{A}$ approximately in practice (see Subsection 6.1.4), its computational cost is much higher than the multiplication in $\mathbb{C}$. #### 6.1.3 Minimization of the trace In the case of $\mathcal{A}=\mathcal{B}(\mathcal{W})$, $p_{j}(x_{i})$ and $p_{j}(x_{i})^{*}$ in the problem (9) do not always commute. Therefore, we restrict the solution to the form $p_{j}(x_{i})=\sum_{i=1}^{n}\phi(x_{i})c_{i}$ where each $c_{i}$ is a Hilbert–Schmidt operator and minimize the trace of the objective function of the problem (9) as follows: $\inf_{\mathbf{c}_{j}\in F,\ \\{\sqrt{\mathbf{G}}\mathbf{c}_{j}\\}_{j=1}^{r}\mbox{\footnotesize: ONS}}-\operatorname{tr}\bigg{(}\sum_{j=1}^{r}\mathbf{c}_{j}^{*}\mathbf{G}^{2}\mathbf{c}_{j}\bigg{)},$ (17) where $F=\\{\mathbf{c}=[c_{1},\ldots,c_{n}]\in\mathcal{A}^{n}\mid\ c_{i}\mbox{ is a Hilbert--Schmidt operator for }i=1,\ldots,n\\}$. If $\mathcal{A}=\mathbb{C}^{m\times m}$, i.e., $\mathcal{W}$ is a finite dimensional space, then we solve the problem (17) by regarding $\mathbf{G}$ as an $mn\times mn$ matrix and computing the eigenvalues and eigenvectors of $\mathbf{G}$. ###### Proposition 6.8 Let $\mathcal{A}=\mathbb{C}^{m\times m}$. Let $\lambda_{1},\ldots,\lambda_{r}\in\mathbb{C}$ and $\mathbf{v}_{1},\ldots,\mathbf{v}_{r}\in\mathbb{C}^{mn}$ be the largest $r$ eigenvalues and the corresponding orthonormal eigenvectors of $\mathbf{G}\in\mathbb{C}^{mn\times mn}$. Then, $\mathbf{c}_{j}=[\mathbf{v}_{j},0,\ldots,0]\lambda_{j}^{-1/2}$ is a solution of the problem (17). Proof Since the identity $\sum_{j=1}^{r}\mathbf{c}_{j}^{*}\mathbf{G}^{2}\mathbf{c}_{j}=\sum_{j=1}^{r}(\sqrt{\mathbf{G}}\mathbf{c}_{j})^{*}\mathbf{G}(\sqrt{\mathbf{G}}\mathbf{c}_{j})$ holds, any solution $\mathbf{c}_{j}$ of the problem (17) satisfies $\sqrt{\mathbf{G}}\mathbf{c}_{j}=\mathbf{v}_{j}u^{*}$ for a normalized vector $u\in\mathbb{C}^{m}$. Thus, $p_{j}=\sum_{i=1}^{n}\phi(x_{i})c_{i,j}$, where ${c}_{i,j}$ is the $i$-th element of $\lambda_{j}^{-1/2}[\mathbf{v}_{j},0,\ldots,0]$, is a solution of the problem. If $\mathcal{W}$ is an infinite dimensional space, we rewrite the problem (17) with the Hilbert–Schmidt norm as follows: $\inf_{\mathbf{c_{j}}\in F,\ \\{\sqrt{\mathbf{G}}\mathbf{c}_{j}\\}_{j=1}^{r}\mbox{\footnotesize: ONS}}-\sum_{j=1}^{r}\|\mathbf{G}\mathbf{c}_{j}\|^{2}_{F},$ (18) where $\|\mathbf{c}\|_{F}^{2}=\sum_{i=1}^{n}\|c_{i}\|_{\operatorname{HS}}^{2}$ and $\|\cdot\|_{\operatorname{HS}}$ is the Hilbert–Schmidt norm for Hilbert–Schmidt operators. Similar to Eq. (11), we rearrange the problem (18) to the following problem by adding a penalty term: $\inf_{\mathbf{c}\in F}-\|\mathbf{G}\mathbf{c}\|^{2}_{F}+\lambda\big{|}\big{\|}\sqrt{\mathbf{Gc}}\big{\|}^{2}_{F}-1\big{|},$ (19) where $\lambda$ is a real positive weight for the penalty term. Then, we can apply the standard gradient descent method in Hilbert spaces to the problem in $F$ (Smyrlis and Zisis, 2004) since $F$ is the Hilbert space equipped with the Hilbert–Schmidt inner product. Similar to the case of Eq. (11), for $r>1$, let $\mathbf{c}_{1}$ be a solution of the problem (19). Then, we solve the same problem in the orthogonal complement of the space spanned by $\\{\mathbf{c}_{1}\\}$ and set the solution of this problem as $\mathbf{c}_{2}$. Then, we solve the same problem in the orthogonal complement of the space spanned by $\\{\mathbf{c}_{1},\mathbf{c}_{2}\\}$ and repeat this procedure to obtain solutions $\mathbf{c}_{1},\ldots\mathbf{c}_{r}$. #### 6.1.4 Numerical examples ##### Experiments with synthetic data We applied the above PCA with $\mathcal{A}=L^{\infty}([0,1])$ to functional data. We randomly generated three kinds of sample-sets from the following functions of two variables on $[0,1]\times[0,1]$: $\displaystyle y_{1}(s,t)=e^{10(s-t)},\quad y_{2}(s,t)=10st,\quad y_{3}(s,t)=\cos(10(s-t)).$ Each sample-set $i$ is composed of 20 samples with random noise. We denote these samples by $x_{1},\ldots,x_{60}$. The noise was randomly drawn from the Gaussian distribution with mean $0$ and standard deviation $0.3$. Since $L^{\infty}([0,1])$ is commutative, we applied the gradient descent proposed in Subsection 6.1.2 to solve the problem (8). The parameters were set as $\lambda=0.1$ and $\eta_{t}=0.01$. We set the $L^{\infty}([0,1])$-valued positive definite kernel $k$ as $(k(x_{i},x_{j}))(t)=\int_{0}^{1}\int_{0}^{1}(t-x_{i}(s_{1},s_{2}))(t-x_{j}(s_{1},s_{2}))ds_{1}ds_{2}$ (see Example 2.22.1). Since $(k(x_{i},x_{j}))(t)$ is a polynomial of $t$, all the computations on $\mathcal{A}$ result in polynomials. Thus, the results are obtained by keeping coefficients of the polynomials. Moreover, we set $\mathbf{c}_{0}$ as the constant function $[1,\ldots,1]^{T}\in\mathcal{A}^{n}$ and computed $\mathbf{c}_{1},\mathbf{c}_{2},\ldots$ according to Eq. (14). For comparison, we also vectorized the samples by discretizing $y_{i}$ at $121=11\times 11$ points composed of 11 equally spaced points in $[0,1]$ ($0,0.1,\ldots,1$) and applied the standard kernel PCA in the RKHS associated with the Laplacian kernel on $\mathbb{R}^{121}$. The results are illustrated in Figure 2. Since the samples are contaminated by the noise, the PCA in the RKHS cannot separate three sample-sets. On the other hand, the $L^{\infty}([0,1])$-valued weights of principal components obtained by the proposed PCA in the RKHM reduce the information of the samples as functions. As a result, it clearly separates three sample-sets. Figure 3 shows the convergence of the proposed gradient descent. In this example, we only compute the first principal components, hence $r$ is set as $1$. For the objective function $f$ defined as $f(\mathbf{c})=-\mathbf{c}^{*}\mathbf{G}^{2}\mathbf{c}+\lambda\mathbf{c}\mathbf{G}\mathbf{c}\mathbf{c}^{*}\mathbf{G}\mathbf{c}+\lambda\mathbf{c}^{*}\mathbf{G}\mathbf{c}$, functions $f(\mathbf{c}_{t})\in L^{\infty}([0,1])$ for $t=0,\ldots,9$ are illustrated. We can see $f(\mathbf{c}_{t+1})<f(\mathbf{c}_{t})$ and $f(\mathbf{c}_{t})$ gradually approaches a certain function as $t$ grows. Figure 2: The $L^{\infty}([0,1])$-valued first principal components obtained by the proposed PCA in an RKHM (left) and the real-valued first and second principal components obtained by the standard PCA in an RKHS (right) Figure 3: The convergence of the function $f(\mathbf{c}_{t})$ along $t$. ##### Experiments with real-world data To show the proposed PCA with RKHMs extracts the continuous dependencies of samples on the principal axes as we insisted in Section 3, we conducted experiments with climate data in Japan111available at https://www.data.jma.go.jp/gmd/risk/obsdl/. The data is composed of the maximum and minimum daily temperatures at 47 prefectures in Japan in 2020. The original data is illustrated in Figure 4. The red line represents the temperature at Hokkaido, the northernmost prefecture in Japan and the blue line represents that at Okinawa, the southernmost prefecture in Japan. We respectively fit the maximum and minimum temperatures at each location to the Fourier series $a_{0}+\sum_{i=1}^{10}(a_{i}\cos(it)+b_{i}\sin(it))$. The fitted functions $x_{1},\ldots,x_{47}\in C([0,366],\mathbb{R}^{2})$ are illustrated in Figure 5. Then, we applied the PCA with the RKHM associated with the $L^{\infty}([0,366])$-valued positive definite kernel $(k(x,y))(t)=e^{-\|x(t)-y(t)\|_{2}^{2}}$. Let $\mathcal{F}=\\{a_{0}+\sum_{i=1}^{10}(a_{i}\cos(it)+b_{i}\sin(it))\mid\ a_{i},b_{i}\in\mathbb{R}\\}\subseteq L^{2}([0,366])$. We project $k(x,y)$ onto $\mathcal{F}$. Then, for $c,d\in\mathcal{F}$, $c+d\in\mathcal{F}$ is satisfied, but $cd\in\mathcal{F}$ is not always satisfied. Thus, we approximate $cd$ with $a_{0}+\sum_{i=1}^{N}(a_{i}\cos(it)+b_{i}\sin(it))$ for $N\leq 10$ to restrict all the computations in $\mathcal{F}$ in practice. Here, to remove high frequency components corresponding to noise and extract essential information, we set $N=3$. Figure 6(a) shows the computed $L^{\infty}([0,366])$-valued weights of the first principal axis in the RKHM, which continuously depends on time. The red and blue lines correspond to Hokkaido and Okinawa, respectively. We see these lines are well-separated from other lines corresponding to other prefectures. For comparison, we also applied the PCA in RKHSs to discrete time data. First, we respectively applied the standard kernel PCA with RKHSs to the original temperature each day and obtained real-valued weights of the first principal components. Here, we used the complex-valued Gaussian kernel $\tilde{k}(x,y)=e^{-\|x-y\|_{2}^{2}}$. Then, we connected the results and obtained Figure 6(b). Since the original data is not smooth, the PCA amplifies the non-smoothness, which provides meaningless results. Next, we respectively applied the standard kernel PCA with the RKHS to the value of the fitted Fourier series each day and obtained real-valued weights of the first principal components. Then, similar to the case of Figure 6(b), we connected the results and obtained Figure 6(c). In this case, the extracted features somewhat capture the continuous behaviors of the temperatures. However, the PCA in the RKHS amplifies high frequency components, which correspond to noise. Therefore, the result fails to separate the temperatures of Hokkaido and Okinawa, whose behaviors are significantly different as illustrated in Figure 4. On the other hand, the PCA in the RKHM captures the feature of each sample as a function and removes nonessential high frequency components, which results in separating functional data properly. Figure 4: Original climate data at 47 locations Figure 5: Fitted Fourier series (a) PCA with RKHMs for the fitted Fourier series (b) PCA with RKHSs for the original data (c) PCA with RKHSs for the fitted Fourier series Figure 6: Principal components of PCA for climate data ### 6.2 Time-series data analysis The problem of analyzing dynamical systems from data by using Perron–Frobenius operators and their adjoints (called Koopman operators), which are linear operators expressing the time evolution of dynamical systems, has recently attracted attention in various fields (Budišić et al., 2012; Črnjarić-Žic et al., 2020; Takeishi et al., 2017a, b; Lusch et al., 2018). And, several methods for this problem using RKHSs have also been proposed (Kawahara, 2016; Klus et al., 2020; Ishikawa et al., 2018; Hashimoto et al., 2020; Fujii & Kawahara, 2019). In these methods, sequential data is supposed to be generated from dynamical systems and is analyzed through Perron–Frobenius operators in RKHSs. To analyze the time evolution of functional data, we generalize Perron–Frobenius operators defined in RKHSs to those in RKHMs by using an operator-valued positive definite kernel describing similarities between pairs of functions. ##### Defining Perron–Frobenius operators in RKHMs We consider the RKHM and vvRKHS associated with an operator-valued positive definite kernel. VvRKHSs are associated with operator-valued kernels, and as we stated in Lemma 4.10, those operator-valued kernels are special cases of $C^{*}$-algebra-valued positive definite kernels. Here, we discuss the advantage of RKHMs over vvRKHSs. Comparing with vvRKHSs, RKHMs have enough representation power for preserving continuous behaviors of infinite dimensional operator-valued kernels, while vvRKHSs are not sufficient for preserving such behaviors. Let $\mathcal{W}$ be a Hilbert space, let $k:\mathcal{X}\times\mathcal{X}\to\mathcal{B}(\mathcal{W})$ be an operator- valued positive definite kernel on a data space $\mathcal{X}$, and let $\mathcal{H}_{k}^{\operatorname{v}}$ be the vvRKHS associated with $k$. Since the inner products in vvRKHSs have the form $\left\langle w,k(x,y)h\right\rangle$ for $w,h\in\mathcal{W}$ and $x,y\in\mathcal{X}$, if $\mathcal{W}$ is a $d$-dimensional space, putting $w$ as $d$ linearly independent vectors in $\mathcal{W}$ reconstructs $k(x,y)$. However, if $\mathcal{W}$ is an infinite dimensional space, we need infinitely many $w$ to reconstruct $k(x,y)$, and we cannot recover the continuous behavior of the operator $k(x,y)$ with finitely many $w$. For example, let $\mathcal{X}=C(\Omega,\mathcal{Y})$ and $\mathcal{W}=L^{2}(\Omega)$ for a compact measure space $\Omega$ and a topological space $\mathcal{Y}$. Let $(k(x,y)w)(s)=\int_{t\in\Omega}\tilde{k}(x(s),y(t))w(t)dt$, where $\tilde{k}$ is a complex-valued positive definite kernel on $\mathcal{Y}$ (see Example 2.22.4). The operator $k(x,y)$ for functional data $x$ and $y$ describes the continuous changes of similarities between function $x$ and $y$. However, the estimation or prediction of the operator $k(x,y)$ in vvRKHSs fails to extract the continuous behavior of the function $\tilde{k}(x(s),y(t))$ in the operator $k(x,y)$ since vectors in vvRKHSs have the form $k(\cdot,y)w$ and we cannot completely recover $k(x,y)$ with finitely many vectors in the vvRKHS. On the other hand, RKHMs have enough information to recover $k(x,y)$ since it is just the inner product between two vectors $\phi(x)$ and $\phi(y)$. #### 6.2.1 Perron–Frobenius operator in RKHSs We briefly review the definition of the Perron-Frobenius operator on RKHS and existing methods for analysis of time-series data through Perron–Frobenius operators and construction of their estimations (Kawahara, 2016; Hashimoto et al., 2020) . First, we define Perron–Frobenius operators in RKHSs. Let $\\{x_{0},x_{1},\ldots\\}\subseteq\mathcal{X}$ be time-series data. We assume it is generated from the following deterministic dynamical system: $x_{i+1}=f(x_{i}),$ (20) where $f:\mathcal{X}\to\mathcal{X}$ is a map. By embedding $x_{i}$ and $f(x_{i})$ in an RKHS $\mathcal{H}_{\tilde{k}}$ associated with a positive definite kernel $\tilde{k}$ and the feature map $\tilde{\phi}$, dynamical system (20) in $\mathcal{X}$ is transformed into that in the RKHS as $\tilde{\phi}(x_{i+1})=\tilde{\phi}(f(x_{i})).$ The Perron–Frobenius operator $\tilde{K}$ in the RKHS is defined as a linear operator on $\mathcal{H}_{\tilde{k}}$ satisfying $\tilde{K}\tilde{\phi}(x):=\tilde{\phi}(f(x))$ for $x\in\mathcal{X}$. If $\\{\tilde{\phi}(x)\mid\ x\in\mathcal{X}\\}$ is linearly independent, $\tilde{K}$ is well-defined as a linear map in the RKHS. For example, if $\tilde{k}$ is a universal kernel (Sriperumbudur et al., 2011) such as the Gaussian or Laplacian kernel on $\mathcal{X}=\mathbb{R}^{d}$, $\\{\tilde{\phi}(x)\mid\ x\in\mathcal{X}\\}$ is linearly independent. By considering eigenvalues and the corresponding eigenvectors of $\tilde{K}$, we can understand the long-time behavior of the dynamical system. For example, let $v_{1},\ldots,v_{m}$ be the eigenvectors with respect to eigenvalue 1 of $\tilde{K}$. We project the vector $\tilde{\phi}(x_{0})$ onto the subspace spanned by $v_{1},\dots,v_{m}$. We denote the projected vector by $v$. Then, for $\alpha=1,2,\ldots$, we have $\tilde{\phi}(x_{\alpha})=\tilde{K}^{\alpha}(v+v^{\perp})=v+\tilde{K}^{\alpha}v^{\perp},$ where $v^{\perp}=\tilde{\phi}(x_{0})-v$. Therefore, by calculating a pre-image of $v$, we can extract the time-invariant component of the dynamical system with the initial value $x_{0}$. For practical uses of the above discussion, we construct an estimation of $\tilde{K}$ only with observed data $\\{x_{0},x_{1},\ldots\\}\subseteq\mathcal{X}$ as follows: We project $\tilde{K}$ onto the finite dimensional subspace spanned by $\\{\tilde{\phi}(x_{0}),\ldots,\tilde{\phi}(x_{T-1})\\}$. Let $\tilde{W}_{T}:=[\tilde{\phi}(x_{0}),\ldots,\tilde{\phi}(x_{T-1})]$ and $\tilde{W}_{T}=\tilde{Q}_{T}\tilde{\mathbf{R}}_{T}$ be the QR decomposition of $\tilde{W}_{T}$ in the RKHS. Then, the Perron–Frobenius operator $\tilde{K}$ is estimated by projecting $\tilde{K}$ onto the space spanned by $\\{\tilde{\phi}(x_{0}),\ldots,\tilde{\phi}(x_{T-1})\\}$. Since $\tilde{K}\tilde{\phi}(x_{i}):=\tilde{\phi}(f(x_{i}))=\tilde{\phi}(x_{i+1})$ holds, we construct an estimation $\tilde{\mathbf{K}}_{T}$ of $\tilde{K}$ as follows: $\displaystyle\tilde{\mathbf{K}}_{T}:$ $\displaystyle=\tilde{Q}_{T}^{*}\tilde{K}\tilde{Q}_{T}=\tilde{Q}_{T}^{*}\tilde{K}\tilde{W}_{T}\tilde{\mathbf{R}}_{T}^{-1}=\tilde{Q}_{T}^{*}[\tilde{\phi}(x_{1}),\ldots,\tilde{\phi}(x_{T})]\tilde{\mathbf{R}}_{T}^{-1},$ which can be computed only with observed data. #### 6.2.2 Perron–Frobenius operator in RKHMs Existing analyses (Kawahara, 2016; Hashimoto et al., 2020) of time-series data with Perron–Frobenius operators are addressed only in RKHSs. In the remaining parts of this section, we generalize the existing analyses to RKHM to extract continuous behaviors of functional data. We consider the case where time- series is functional data. Let $\Omega$ be a compact measure space, $\mathcal{Y}$ be a topological space, $\mathcal{X}=C(\Omega,\mathcal{Y})$, $\mathcal{A}=\mathcal{B}(L^{2}(\Omega))$, and $\\{x_{0},x_{1},\ldots\\}\subseteq\mathcal{X}$ be functional time-series data. Let $k:\mathcal{X}\times\mathcal{X}\to\mathcal{A}$ be defined as $(k(x,y)w)(s)=\int_{t\in\Omega}\tilde{k}(x(s),y(t))w(t)dt$, where $\tilde{k}:\mathcal{Y}\times\mathcal{Y}\to\mathbb{C}$ is a complex-valued positive definite kernel (see Example 2.22.4 and the last paragraph of Section 3). The operator $k(x,y)$ is the integral operator whose integral kernel is $\tilde{k}(x(s),y(t))$. We define a Perron–Frobenius operator in the RKHM $\mathcal{M}_{k}$ associated with the above kernel $k$ as an $\mathcal{A}$-linear operator satisfying ${K}{\phi}(x)={\phi}(f(x))$ for $x\in\mathcal{X}$. We assume $K$ is well-defined on a dense subset of $\mathcal{M}_{k}$. Then, for $\alpha,\beta=1,2,\ldots$, we have $k(x_{\alpha},x_{\beta})=\left\langle\phi(x_{\alpha}),\phi(x_{\beta})\right\rangle_{\mathcal{M}_{k}}=\big{\langle}K^{\alpha}\phi(x_{0}),K^{\beta}\phi(x_{0})\big{\rangle}_{\mathcal{M}_{k}}.$ Therefore, by estimating $K$ in the RKHM $\mathcal{M}_{k}$, we can extract the similarity between arbitrary points of functions $x_{\alpha}$ and $x_{\beta}$. Moreover, the eigenvalues and eigenvectors of $K$ provide us a decomposition of the similarity $k(x_{\alpha},x_{\beta})$ into a time-invariant term and time-dependent term. Since $K$ is a linear operator on a Banach space $\mathcal{M}_{k}$, eigenvalues and eigenvectors of $K$ are available. Let $v_{1},\ldots,v_{m}\in\mathcal{M}_{k}$ be the eigenvectors with respect to eigenvalue $1$ of $K$. We project the vector $\phi(x_{0})$ onto the submodule spanned by $v_{1},\ldots,v_{m}$, which is denoted by $\mathcal{V}$. Let $\\{q_{1},\ldots,q_{m}\\}\subseteq\mathcal{M}_{k}$ be an orthonormal basis of $\mathcal{V}$ and let $v=\sum_{i=1}^{m}q_{i}\left\langle q_{i},\phi(x_{0})\right\rangle_{\mathcal{M}_{k}}$. Then, we have $k(x_{\alpha},x_{\beta})=\big{\langle}K^{\alpha}(v+v^{\perp}),K^{\beta}(v+v^{\perp})\big{\rangle}_{\mathcal{M}_{k}}=\left\langle v,v\right\rangle_{\mathcal{M}_{k}}+r({\alpha},\beta),$ (21) where $v^{\perp}=\phi(x_{0})-v$ and $r(\alpha,\beta)=\left\langle K^{\alpha}v,K^{\beta}v^{\perp}\right\rangle_{\mathcal{M}_{k}}+\left\langle K^{\alpha}v^{\perp},K^{\beta}v\right\rangle_{\mathcal{M}_{k}}+\left\langle K^{\alpha}v^{\perp},K^{\beta}v^{\perp}\right\rangle_{\mathcal{M}_{k}}$. Therefore, the term $\left\langle v,v\right\rangle_{\mathcal{M}_{k}}$ provides us with the information about time-invariant similarities. ###### Remark 6.9 We can also consider the vvRKHS $\mathcal{H}_{k}^{\operatorname{v}}$ with respect to the operator-valued kernel $k$. Here, we discuss the difference between the case of vvRKHS and RKHM. The Perron–Frobenius operator $K^{\operatorname{v}}$ in a vvRKHS $\mathcal{H}_{k}^{\operatorname{v}}$ (Fujii & Kawahara, 2019) is defined as a linear operator satisfying $K^{\operatorname{v}}\phi(x)w=\phi(f(x))w$ for $x\in\mathcal{X}$ and $w\in\mathcal{W}$. However, with finitely many vectors in $\mathcal{H}_{k}^{\operatorname{v}}$, we can only recover an projected operator $UU^{*}k(x_{\alpha},x_{\beta})UU^{*}$, where $N\in\mathbb{N}$, $U=[u_{1},\ldots,u_{N}]$, and $\\{u_{1},\ldots,u_{N}\\}$ is an orthonormal system on $\mathcal{W}$ as follows: $U^{*}k(x_{\alpha},x_{\beta})U=\big{[}\left\langle\phi(x_{s})u_{i},\phi(x_{t})u_{j}\right\rangle_{\mathcal{H}_{k}^{\operatorname{v}}}\big{]}_{i,j}=\big{[}\big{\langle}(K^{\operatorname{v}})^{\alpha}\phi(x_{0})u_{i},(K^{\operatorname{v}})^{\beta}\phi(x_{0})u_{j}\big{\rangle}_{\mathcal{H}_{k}^{\operatorname{v}}}\big{]}_{i,j}.$ (22) Furthermore, let $v_{1},\ldots,v_{m}\in\mathcal{M}_{k}$ be the eigenvectors with respect to eigenvalue $1$ of $K^{\operatorname{v}}$. Let $\\{q_{1},\ldots,q_{m}\\}\subseteq\mathcal{H}_{k}^{\operatorname{v}}$ be an orthonormal basis of the subspace spanned by $v_{1},\ldots,v_{m}$ and let $\tilde{v}_{j}=\sum_{i=1}^{m}q_{i}\left\langle q_{i},\phi(x_{0})u_{j}\right\rangle_{\mathcal{H}_{k}^{\operatorname{v}}}$. Then, we have $U^{*}k(x_{\alpha},x_{\beta})U=\big{[}\langle(K^{\operatorname{v}})^{\alpha}(\tilde{v}_{i}+\tilde{v}_{i}^{\perp}),(K^{\operatorname{v}})^{\beta}(\tilde{v}_{j}+\tilde{v}_{j}^{\perp})\rangle_{\mathcal{H}_{k}^{\operatorname{v}}}\big{]}_{i,j}=[\left\langle\tilde{v}_{i},\tilde{v}_{j}\right\rangle_{\mathcal{H}_{k}^{\operatorname{v}}}]_{i,j}+\tilde{r}(\alpha,\beta),$ (23) where $\tilde{v}_{i}^{\perp}=\phi(x_{0})u_{i}-\tilde{v}_{i}$ and $\tilde{r}(\alpha,\beta)=[\langle(K^{\operatorname{v}})^{\alpha}\tilde{v}_{i},(K^{\operatorname{v}})^{\beta}\tilde{v}_{j}^{\perp}\rangle_{\mathcal{H}_{k}^{\operatorname{v}}}+\langle(K^{\operatorname{v}})^{\alpha}\tilde{v}_{i}^{\perp},(K^{\operatorname{v}})^{\beta}\tilde{v}_{j}\rangle_{\mathcal{H}_{k}^{\operatorname{v}}}+\langle(K^{\operatorname{v}})^{\alpha}\tilde{v}_{i}^{\perp},(K^{\operatorname{v}})^{\beta}\tilde{v}_{j}^{\perp}\rangle_{\mathcal{H}_{k}^{\operatorname{v}}}]_{i,j}$. Therefore, with vvRKHSs, we cannot recover the continuous behavior of the operator $k(x,y)$ which encodes similarities between functions $x$ and $y$. #### 6.2.3 Estimation of Perron–Frobenius operators in RKHMs In practice, we only have time-series data but do not know the underlying dynamical system and its Perron–Frobenius operator in an RKHM. Therefore, we consider estimating the Perron–Frobenius operator only with the data. To do so, we generalize the Gram–Schmidt orthonormalization algorithm to Hilbert $C^{*}$-modules to apply the QR decomposition and project Perron–Frobenius operators onto the submodule spanned by $\\{\phi(x_{0}),\ldots,\phi(x_{T-1})\\}$. The Gram–Schmidt orthonormalization in Hilbert modules is theoretically investigated by Cnops (1992). Here, we develop a practical method for our settings. Then, we can apply the decomposition (21), proposed in Subsection 6.2.2, of the estimated operator regarding eigenvectors. Since we are considering the RKHM associated with the integral operator-valued positive definite kernel defined in the first part of Subsection 6.2.2, we assume $\mathcal{A}=\mathcal{B}(\mathcal{W})$ and we denote by $\mathcal{M}$ a Hilbert $C^{*}$-module over $\mathcal{A}$ throughout this subsection. Note that integral operators are compact. We first develop a normalization method for Hilbert $C^{*}$-modules. In $C^{*}$-algebras, nonzero elements are not always invertible, which is the main difficulty of the normalization in Hilbert $C^{*}$-modules. However, by carefully applying the definition of normalized (see Definition 2.17), we can construct a normalization method. ###### Proposition 6.10 (Normalization) Let $\epsilon\geq 0$ and let $\hat{q}\in\mathcal{M}$ satisfy $\|\hat{q}\|_{\mathcal{M}}>\epsilon$. Assume $\langle\hat{q},\hat{q}\rangle_{\mathcal{M}}$ is compact. Then, there exists $\hat{b}\in\mathcal{A}$ such that $\|\hat{b}\|_{\mathcal{A}}<1/\epsilon$ and $q:=\hat{q}\hat{b}$ is normalized. In addition, there exists $b\in\mathcal{A}$ such that $\|\hat{q}-qb\|_{\mathcal{M}}\leq\epsilon$. Proof Let $\lambda_{1}\geq\lambda_{2}\geq\cdots\geq 0$ be the eigenvelues of the compact operator $\left\langle\hat{q},\hat{q}\right\rangle_{\mathcal{M}}$, and $m^{\prime}:=\max\\{j\mid\ \lambda_{j}>\epsilon^{2}\\}$. Since $\langle\hat{q},\hat{q}\rangle_{\mathcal{M}}$ is positive and compact, it admits the spectral decomposition $\langle\hat{q},\hat{q}\rangle_{\mathcal{M}}=\sum_{i=1}^{\infty}\lambda_{i}v_{i}v_{i}^{*}$, where $v_{i}$ is the orthonormal eigenvector with respect to $\lambda_{i}$. Also, since $\lambda_{1}=\|\hat{q}\|_{\mathcal{M}}^{2}>\epsilon^{2}$, we have $m^{\prime}\geq 1$. Let $\hat{b}=\sum_{i=1}^{m^{\prime}}1/\sqrt{\lambda_{i}}v_{i}v_{i}^{*}$. By the definition of $\hat{b}$, $\|\hat{b}\|_{\mathcal{A}}=1/\sqrt{\lambda_{m^{\prime}}}<1/\epsilon$ holds. Also, we have $\displaystyle\langle\hat{q}\hat{b},\hat{q}\hat{b}\rangle_{\mathcal{M}}$ $\displaystyle=\hat{b}^{*}\left\langle\hat{q},\hat{q}\right\rangle_{\mathcal{M}}\hat{b}=\sum_{i=1}^{m^{\prime}}\frac{1}{\sqrt{\lambda_{i}}}v_{i}v_{i}^{*}\sum_{i=1}^{\infty}\lambda_{i}v_{i}v_{i}^{*}\sum_{i=1}^{m^{\prime}}\frac{1}{\sqrt{\lambda_{i}}}v_{i}v_{i}^{*}=\sum_{i=1}^{m^{\prime}}v_{i}v_{i}^{*}.$ Thus, $\langle\hat{q}\hat{b},\hat{q}\hat{b}\rangle_{\mathcal{M}}$ is a nonzero orthogonal projection. In addition, let $b=\sum_{i=1}^{m^{\prime}}\sqrt{\lambda_{i}}v_{i}v_{i}^{*}$. Since $\hat{b}b=\sum_{i=1}^{m^{\prime}}v_{i}v_{i}^{*}$, the identity $\langle\hat{q},\hat{q}\hat{b}b\rangle=\langle\hat{q}\hat{b}b,\hat{q}\hat{b}b\rangle$ holds, and we obtain $\displaystyle\langle\hat{q}-qb,\hat{q}-qb\rangle_{\mathcal{M}}$ $\displaystyle=\langle\hat{q}-\hat{q}\hat{b}b,\hat{q}-\hat{q}\hat{b}b\rangle_{\mathcal{M}}=\langle\hat{q},\hat{q}\rangle-\langle\hat{q}\hat{b}b,\hat{q}\hat{b}b\rangle_{\mathcal{M}}$ $\displaystyle=\sum_{i=1}^{\infty}{\lambda_{i}}v_{i}v_{i}^{*}-\sum_{i=1}^{m^{\prime}}{\lambda_{i}}v_{i}v_{i}^{*}=\sum_{i=m^{\prime}+1}^{\infty}{\lambda_{i}}v_{i}v_{i}^{*}.$ Thus, $\|\hat{q}-q\hat{b}\|_{\mathcal{M}}=\sqrt{\lambda_{m^{\prime}+1}}\leq\epsilon$ holds, which completes the proof of the proposition. Proposition 6.10 and its proof provide a concrete procedure to obtain normalized vectors in $\mathcal{M}$. This enables us to compute an orthonormal basis practically by applying Gram-Schmidt orthonormalization with respect to $\mathcal{A}$-valued inner product. ###### Proposition 6.11 (Gram-Schmidt orthonormalization) Let $\\{w_{i}\\}_{i=1}^{\infty}$ be a sequence in $\mathcal{M}$. Assume $\left\langle w_{i},w_{j}\right\rangle_{\mathcal{M}}$ is compact for any $i,j=1,2,\ldots$. Consider the following scheme for $i=1,2,\ldots$ and $\epsilon\geq 0$: $\displaystyle\hat{q}_{j}$ $\displaystyle=w_{j}-\sum_{i=1}^{j-1}q_{i}\left\langle q_{i},w_{j}\right\rangle_{\mathcal{M}},\quad q_{j}=\hat{q}_{j}\hat{b}_{j}\quad\mbox{if }\;\|\hat{q}_{j}\|_{\mathcal{M}}>\epsilon,$ (24) $\displaystyle q_{j}$ $\displaystyle=0\quad\mbox{o.w.},$ where $\hat{b}_{j}$ is defined as $\hat{b}$ in Proposition 6.10 by setting $\hat{q}=\hat{q}_{j}$. Then, $\\{q_{j}\\}_{j=1}^{\infty}$ is an orthonormal basis in $\mathcal{M}$ such that any $w_{j}$ is contained in the $\epsilon$-neighborhood of the module spanned by $\\{q_{j}\\}_{j=1}^{\infty}$. ###### Remark 6.12 We give some remarks about the role of $\epsilon$ in Propositions 6.10. The vector $\hat{q}_{i}$ can always be reconstructed by $w_{i}$ only when $\epsilon=0$. This is because the information of the spectrum of $\left\langle\hat{q}_{i},\hat{q}_{i}\right\rangle_{\mathcal{M}}$ may be lost if $\epsilon>0$. However, if $\epsilon$ is sufficiently small, we can reconstruct $\hat{q}_{i}$ with a small error. On the other hand, the norm of $\hat{b}_{i}$ can be large if $\epsilon$ is small, and the computation of $\\{q_{i}\\}_{i=1}^{\infty}$ can become numerically unstable. This corresponds to the trade-off between the theoretical accuracy and numerical stability. To prove Proposition 6.11, we first prove the following lemmas. ###### Lemma 6.13 For $c\in\mathcal{A}$ and $v\in\mathcal{M}$, if $\left\langle v,v\right\rangle_{\mathcal{M}}c=\left\langle v,v\right\rangle_{\mathcal{M}}$, then $vc=v$ holds. Proof If $\left\langle v,v\right\rangle_{\mathcal{M}}c=\left\langle v,v\right\rangle_{\mathcal{M}}$, then $c^{*}\left\langle v,v\right\rangle_{\mathcal{M}}=\left\langle v,v\right\rangle_{\mathcal{M}}$ and we have $\displaystyle\left\langle vc-v,vc-v\right\rangle_{\mathcal{M}}=c^{*}\left\langle v,v\right\rangle_{\mathcal{M}}c-c^{*}\left\langle v,v\right\rangle_{\mathcal{M}}-\left\langle v,v\right\rangle_{\mathcal{M}}c+\left\langle v,v\right\rangle_{\mathcal{M}}=0,$ which implies $vc=v$. ###### Lemma 6.14 If $q\in\mathcal{M}$ is normalized, then $q\left\langle q,q\right\rangle_{\mathcal{M}}=q$ holds. Proof Since $\left\langle q,q\right\rangle_{\mathcal{M}}$ is a projection, $\left\langle q,q\right\rangle_{\mathcal{M}}\left\langle q,q\right\rangle_{\mathcal{M}}=\left\langle q,q\right\rangle_{\mathcal{M}}$ holds. Therefore, letting $c=\left\langle q,q\right\rangle_{\mathcal{M}}$ and $v=q$ in Lemma 6.13 completes the proof of the lemma. Proof of Proposition 6.11 By Proposition 6.10, $q_{j}$ is normalized, and for $\epsilon\geq 0$, there exists $b_{j}\in\mathcal{A}$ such that $\|\hat{q}_{j}-q_{j}b_{j}\|_{\mathcal{M}}\leq\epsilon$. Therefore, by the definition of $\hat{q}_{j}$, $\|w_{j}-v_{j}\|_{\mathcal{M}}\leq\epsilon$ holds, where $v_{j}$ is a vector in the module spanned by $\\{q_{j}\\}_{j=0}^{\infty}$ which is defined as $v_{j}=\sum_{i=1}^{j-1}q_{i}\left\langle q_{i},w_{j}\right\rangle_{\mathcal{M}}-q_{j}b_{j}$. This means that the $\epsilon$-neighborhood of the space spanned by $\\{q_{j}\\}_{j=1}^{\infty}$ contains $\\{w_{j}\\}_{j=1}^{\infty}$. Next, we show the orthogonality of $\\{q_{j}\\}_{j=1}^{\infty}$. Assume $q_{1},\ldots,q_{j-1}$ are orthogonal to each other. For $i<j$, the following identities are deduced by Lemma 6.14: $\displaystyle\left\langle q_{j},q_{i}\right\rangle_{\mathcal{M}}$ $\displaystyle=\hat{b}_{t}^{*}\left\langle\hat{q}_{j},q_{i}\right\rangle_{\mathcal{M}}=\hat{b}_{j}^{*}\bigg{\langle}w_{j}-\sum_{l=1}^{j-1}q_{l}\left\langle q_{l},w_{j}\right\rangle,q_{i}\bigg{\rangle}_{\mathcal{M}}$ $\displaystyle=\hat{b}_{j}^{*}\left(\left\langle w_{j},q_{i}\right\rangle_{\mathcal{M}}-\left\langle q_{i}\left\langle q_{i},w_{j}\right\rangle_{\mathcal{M}},q_{i}\right\rangle\right)=\hat{b}_{j}^{*}\left(\left\langle w_{j},q_{i}\right\rangle_{\mathcal{M}}-\left\langle w_{j},q_{i}\right\rangle_{\mathcal{M}}\right)=0.$ Therefore, $q_{1},\ldots,q_{j}$ are also orthogonal to each other, which completes the proof of the proposition. In practical computations, the scheme (24) should be represented with matrices. For this purpose, we derive the following QR decomposition from Proposition 6.11. This is a generalization of the QR decomposition in Hilbert spaces. ###### Corollary 6.15 (QR decomposition) For $n\in\mathbb{N}$, let $W:=[w_{1},\ldots,w_{n}]$ and $Q:=[q_{1},\ldots,q_{n}]$. Let $\epsilon\geq 0$. Then, there exist $\mathbf{R},\mathbf{R}_{\operatorname{inv}}\in\mathcal{A}^{n\times n}$ that satisfy $Q=W\mathbf{R}_{\operatorname{inv}},\quad\|W-Q\mathbf{R}\|\leq\epsilon.$ (25) Here, $\|W\|$ for a $\mathcal{A}$-linear map $W:\mathcal{A}^{n}\to\mathcal{M}$ is defined as $\|W\|:=\sup_{\|v\|_{\mathcal{A}^{n}}=1}\|Wv\|_{\mathcal{M}}$. Proof Let $\mathbf{R}=[r_{i,j}]_{i,j}$ be an $n\times n$ $\mathcal{A}$-valued matrix. Here, $r_{i,j}$ is defined by $r_{i,j}=\left\langle q_{i},w_{j}\right\rangle_{\mathcal{M}}\in\mathcal{A}$ for $i<j$, $r_{i,j}=0$ for $i>j$, and $r_{j,j}=b_{j}$, where $b_{j}$ is defined as $b$ in Proposition 6.10 by setting $\hat{q}=\hat{q}_{j}$. In addition, let $\hat{\mathbf{B}}=\operatorname{diag}\\{\hat{b}_{1},\ldots,\hat{b}_{n}\\}$, $\mathbf{B}=\operatorname{diag}\\{{b}_{1},\ldots,{b}_{n}\\}$, and $\mathbf{R}_{\operatorname{inv}}=\mathbf{\hat{B}}(I+(\mathbf{R}-\mathbf{B})\mathbf{\hat{B}})^{-1}$ be $n\times n$ $\mathcal{A}$-valued matrices. The equality $Q=W\mathbf{R}_{\operatorname{inv}}$ is derived directly from scheme (24). In addition, by the scheme (24), for $t=1,\ldots,n$, we have $\displaystyle w_{j}$ $\displaystyle=\sum_{i=1}^{j-1}q_{i}\left\langle q_{i},w_{j}\right\rangle_{\mathcal{M}}+\hat{q}_{j}=\sum_{i=1}^{j-1}q_{i}\left\langle q_{i},w_{j}\right\rangle_{\mathcal{M}}+q_{j}b_{j}+\hat{q}_{j}-q_{j}b_{j}=Q\mathbf{r}_{j}+\hat{q}_{j}-q_{j}b_{j},$ where $\mathbf{r}_{j}\in\mathcal{A}^{n}$ is the $i$-th column of $\mathbf{R}$. Therefore, by Proposition 6.10, $\|w_{j}-Q\mathbf{r}_{j}\|_{\mathcal{M}}=\|\hat{q}_{j}-q_{j}b_{j}\|_{\mathcal{M}}\leq\epsilon$ holds for $j=1,\ldots,n$, which implies $\|W-Q\mathbf{R}\|\leq\epsilon$. We call the decomposition (25) as the QR decomposition in Hilbert $C^{*}$-modules. Although we are handling vectors in $\mathcal{M}$, by applying the QR decomposition, we only have to compute $\mathbf{R}_{\operatorname{inv}}$ and $\mathbf{R}$. We now consider estimating the Perron–Frobenius operator $K$ with observed time-series data $\\{x_{0},x_{1},\ldots\\}$. Let ${W}_{T}=[{\phi}(x_{0}),\ldots,{\phi}(x_{T-1})]$. We are considering an integral operator-valued positive definite kernel (see the first part of Subsection 6.2.2 and the last paragraph in Section 3). Since integral operators are compact, $W_{T}$ satisfies the assumption in Corollary 6.11. Thus, let ${W}_{T}{\mathbf{R}}_{\operatorname{inv},T}={Q}_{T}$ be the QR decomposition (25) of ${W}_{T}$ in the RKHM $\mathcal{M}_{k}$. The Perron–Frobenius operator $K$ is estimated by projecting ${K}$ onto the module spanned by $\\{{\phi}(x_{0}),\ldots,\phi(x_{T-1})\\}$. We define ${\mathbf{K}}_{T}$ as the estimation of $K$. Since ${K}{\phi}(x_{i})={\phi}(f(x_{i}))={\phi}(x_{i+1})$ hold, ${\mathbf{K}}_{T}$ can be computed only with observed data as follows: $\displaystyle{\mathbf{K}}_{T}$ $\displaystyle={Q}_{T}^{*}{K}{Q}_{T}={Q}_{T}^{*}{K}{W}_{T}{\mathbf{R}}_{\operatorname{inv},T}={Q}_{T}^{*}[{\phi}(x_{1}),\ldots,{\phi}(x_{T})]{\mathbf{R}}_{\operatorname{inv},T}.$ ###### Remark 6.16 In practical computations, we only need to keep the integral kernels to implement the Gram–Schmidt orthonormalization algorithm and estimate Perron–Frobenius operators in the RKHM associated with the integral operator- valued kernel $k$. Therefore, we can directly access integral kernel functions of operators, which is not achieved by vvRKHS as we stated in Remark 4.13. Indeed, the operations required for estimating Perron–Frobenius operators are explicitly computed as follows: Let $c,d\in\mathcal{B}(L^{2}(\Omega))$ be integral operators whose integral kernels are $f(s,t)$ and $g(s,t)$. Then, the integral kernels of the operator $c+d$ and $cd$ are $f(s,t)+g(s,t)$ and $\int_{r\in\Omega}f(s,r)g(r,t)dr$, respectively. And that of $c^{*}$ is $f(t,s)$. Moreover, if $c$ is positive, let $c_{\epsilon}^{+}$ be $\sum_{\lambda_{i}>\epsilon}1/\sqrt{\lambda_{i}}v_{i}{v_{i}}^{*}$, where $\lambda_{i}$ are eigenvalues of the compact positive operator $c$ and $v_{i}$ are corresponding orthonormal eigenvectors. Then, the integral kernel of the operator $c_{\epsilon}^{+}$ is $\sum_{\lambda_{i}>\epsilon}1/\sqrt{\lambda_{i}}v_{i}(s)\overline{v_{i}(t)}$. #### 6.2.4 Numerical examples To show the proposed analysis with RKHMs captures continuous changes of values of kernels along functional data as we insisted in Section 3, we conducted experiments with river flow data of the Thames River in London222available at https://nrfa.ceh.ac.uk/data/search. The data is composed of daily flow at 10 stations. We used the data for 51 days beginning from January first, 2018. We regard every daily flow as a function of the ratio of the distance from the most downstream station and fit it to a polynomial of degree 5 to obtain time series $x_{0},\ldots,x_{50}\in C([0,1],\mathbb{R})$. Then, we estimated the Perron–Frobenius operator which describes the time evolution of the series $x_{0},\ldots,x_{50}$ in the RKHM associated with the $\mathcal{B}(L^{2}([0,1]))$-valued positive definite kernel $k(x,y)$ defined as the integral operator whose integral kernel is $\tilde{k}(s,t)=e^{-|x(s)-y(t)|^{2}}$ for $x,y\in C([0,1],\mathbb{R})$. In this case, $T=50$. As we noted in Remark 6.16, all the computations in $\mathcal{A}=\mathcal{B}(L^{2}([0,1]))$ are implemented by keeping integral kernels of operators. Let $\mathcal{F}$ be the set of polynomials of the form $x_{i}(s,t)=\sum_{j,l=0}^{5}\eta_{j,l}s^{j}t^{l}$, where $\eta_{j,l}\in\mathbb{R}$. We project $\tilde{k}$ onto $\mathcal{F}$. Then, for $c,d\in\mathcal{F}$, $c+d\in\mathcal{F}$ is satisfied, but $cd\in\mathcal{F}$ is not always satisfied. Thus, we project $cd$ onto $\mathcal{F}$ to restrict all the computations in $\mathcal{F}$ in practice. We computed the time-invariant term $\left\langle v,v\right\rangle_{\mathcal{M}_{k}}$ in Eq. (22). Regarding the computation of eigenvectors with respect to the eigenvalue $1$, we consider the following minimization problem for the estimated Perron–Frobenius operator $\mathbf{K}_{T}$: $\inf_{\mathbf{v}\in\mathcal{A}^{T}}|\mathbf{K}_{T}\mathbf{v}-\mathbf{v}|_{\mathcal{A}^{T}}^{2}-\lambda|\mathbf{v}|_{\mathcal{A}^{T}}^{2}.$ (26) Here, $-\lambda|\mathbf{v}|_{\mathcal{A}^{T}}^{2}$ is a penalty term to keep $\mathbf{v}$ not going to $0$. Since the objective function of the problem (26) is represented as $\mathbf{v}^{*}(\mathbf{K}_{T}^{*}\mathbf{K}_{T}-\mathbf{K}_{T}^{*}-\mathbf{K}_{T}+(1-\lambda)\mathbf{I})\mathbf{v}$, where $\mathbf{I}$ is the identity operator on $\mathcal{A}^{T}$, we apply the gradient descent on $\mathcal{A}^{T}$ (see Remark 6.6). Figure 7(a) shows the heat map representing the integral kernel of $\left\langle v,v\right\rangle_{\mathcal{M}_{k}}$. For comparison, we also applied the similar analysis in a vvRKHS. We computed the time-invariant term $[\left\langle\tilde{v}_{i},\tilde{v}_{j}\right\rangle_{\mathcal{H}_{k}^{\operatorname{v}}}]_{i,j}$ in Eq. (23) by setting $u_{i}$ as orthonormal polynomials of the form $u_{i}(s)=\sum_{j=1}^{5}\eta_{j}s^{j}$, where $\eta_{j}\in\mathbb{R}$. Let $c_{\operatorname{inv}}=[\left\langle\tilde{v}_{i},\tilde{v}_{j}\right\rangle_{\mathcal{H}_{k}^{\operatorname{v}}}]_{i,j}$. In this case, we cannot obtain the integral kernel of the time-invariant term of the operator $k(x_{\alpha},x_{\beta})$, which is denoted by $\tilde{k}_{\operatorname{inv}}$ here. Instead, by approximating $k(x_{\alpha},x_{\beta})$ by $UU^{*}k(x_{\alpha},x_{\beta})UU^{*}$ and computing $Uc_{\operatorname{inv}}U^{*}\chi_{[0,t]}$, we obtain an approximation of $\int_{0}^{t}\tilde{k}_{\operatorname{inv}}(s,r)dr$ for $s\in[0,1]$. Here, $\chi_{E}:[0,1]\to\\{0,1\\}$ is the indicator function for a Borel set $E$ on $[0,1]$. Therefore, by numerically differentiating $Uc_{\operatorname{inv}}U^{*}\chi_{[0,t]}$ by $t$, we obtain an approximation of $\tilde{k}_{\operatorname{inv}}$. Figure 7(b) shows the heat map representing the approximation of $\tilde{k}_{\operatorname{inv}}$. Around the upstream stations, there are many branches and the flow is affected by them. Thus, the similarity between flows at two points would change along time. While, around the downstream stations, the flow is supposed not to be affected by other rivers. Thus, the similarity between flows at two points would be invariant along time. The values around the diagonal part of Figure 7(a) (RKHM) become small as $s$ and $t$ become large (as going up the river). On the other hand, those of Figure 7(b) (vvRKHS) are also large for large $s$ and $t$. Therefore, RKHM captures the aforementioned fact more properly. (a) RKHM (b) vvRKHS Figure 7: Heat maps representing time-invariant similarities ### 6.3 Analysis of interaction effects Polynomial regression is a classical problem in statistics (Hastie et al., 2009) and analyzing interacting effects by the polynomial regression has been investigated (for its recent improvements, see, for example, Suzumura et al. (2017)). Most of the existing methods focus on the case of finite dimensional (discrete) data. However, in practice, we often encounter situations where we cannot fix the dimension of data. For example, observations are obtained at multiple locations and the locations are not fixed. It may be changed depending on time. Therefore, analysing interaction effects of infinite dimensional (continuous) data is essential. We show the KMEs of $\mathcal{A}$-valued measures in RKHMs provide us with a method for the analysis of infinite dimensional data by setting $\mathcal{A}$ as an infinite dimensional space such as $\mathcal{B}(\mathcal{W})$. Moreover, the proposed method does not need the assumption that interaction effects are described by a polynomial. We first develop the analysis in RKHMs for the case of finite dimensional data in Subsection 6.3.1. Then, we show the analysis is naturally generalized to the infinite dimensional data in Subsection 6.3.2. ##### Applying $\mathcal{A}$-valued measures and KME in RKHMs Using $\mathcal{A}$-valued measures, we can describe the measure corresponding to each point of functional data as functions or operators. For example, let $\mathcal{X}$ be a locally compact Hausdorff space and let $x_{1},x_{2},\ldots\in C([0,1],\mathcal{X})$ be samples. Let $\mathcal{A}=L^{\infty}([0,1])$ and let $\mu$ be the $\mathcal{A}$-valued measure defined as $\mu(t)=\tilde{\mu}_{t}$, where $\tilde{\mu}_{t}$ is the distribution which samples $x_{1}(t),x_{2}(t),\ldots$ follow. Then, $\mu$ describes continuous behaviors of the distribution of samples $x_{1}(t),x_{2}(t),\ldots$ with respect to $t$. Moreover, let $\mathcal{A}=\mathcal{B}(L^{2}([0,1]))$ and let $\mu$ be the $\mathcal{A}$-valued measure defined as $(\mu(E)v)(s)=\int_{t\in[0,1]}\tilde{\mu}(E)_{s,t}v(t)dt$ for a Borel set $E$, where $\tilde{\mu}_{s,t}$ is the joint distribution of the distributions which samples $x_{1}(s),x_{2}(s),\ldots$ and samples $x_{1}(t),x_{2}(t),\ldots$ follow. Then, $\mu$ describes continuous dependencies of samples $x_{1}(s),x_{2}(s),\ldots$ and samples $x_{1}(t),x_{2}(t),\ldots$ with respect to $s$ and $t$. Using the KME in RKHMs, we can embed $\mathcal{A}$-valued measures into RKHMs, which enables us to compute inner products between $\mathcal{A}$-valued measures. Then, we can generalize algorithms in Hilbert spaces to $\mathcal{A}$-valued measures. #### 6.3.1 The case of finite dimensional data In this subsection, we assume $\mathcal{A}=\mathbb{C}^{m\times m}$. Let $\mathcal{X}$ be a locally compact Hausdorff space and let $x_{1},\ldots,x_{n}\in\mathcal{X}^{m\times m}$ and $y_{1},\ldots,y_{n}\in\mathcal{A}$ be given samples. We assume there exist functions $f_{j,l}:\mathcal{X}\to\mathcal{A}$ such that $y_{i}=\sum_{j,l=1}^{m}f_{j,l}((x_{i})_{j,l})$ for $i=1,\ldots,n$. For example, the $(j,l)$-element of each $x_{i}$ describes an effect of the $l$-th element on the $j$-th element of $x_{i}$ and $f_{j,l}$ is a nonlinear function describing an impact of the effect to the value $y_{i}$. If the given samples $y_{i}$ are real or complex-valued, we can regard them as $y_{i}1_{\mathcal{A}}$ to meet the above setting. Let $\mu_{x}\in\mathcal{D}(\mathcal{X},\mathbb{C}^{m\times m})$ be a $\mathbb{C}^{m\times m}$-valued measure defined as $(\mu_{x})_{j,l}=\tilde{\delta}_{x_{j,l}}$, where $\tilde{\delta}_{x}$ for $x\in\mathcal{X}$ is the standard (complex-valued) Dirac measure centered at $x$. Note that the $(j,l)$-element of $\mu_{x}$ describes a measure regarding the element $x_{j,l}$. Let $k$ be an $\mathcal{A}$-valued $c_{0}$-kernel (see Definition 5.2), let $\mathcal{M}_{k}$ be the RKHM associated with $k$, and let $\Phi$ be the KME defined in Section 5.1. In addition, let $\mathcal{V}$ be the submodule of $\mathcal{M}_{k}$ spanned by $\\{\Phi(\mu_{x_{1}}),\ldots,\Phi(\mu_{x_{n}})\\}$, and let $P_{f}:\mathcal{V}\to\mathbb{C}^{m\times m}$ be a $\mathbb{C}^{m\times m}$-linear map (see Definition 2.19) which satisfies $P_{f}\Phi(\mu_{x_{i}})=\sum_{j,l=1}^{m}{f_{j,l}((x_{i})_{j,l})}$ for $i=1,\ldots,n$. Here, we assume the vectors $\Phi(\mu_{x_{1}}),\ldots,\Phi(\mu_{x_{n}})$ are $\mathbb{C}^{m\times m}$-linearly independent (see Definition 2.20). #### 6.3.2 Generalization to the continuous case We generalize the setting mentioned in Subsection 6.3.1 to the case of functional data. We assume Assumption 5.3 in this subsection. We set $\mathcal{A}$ as $\mathcal{B}(L^{2}[0,1])$ instead of $\mathbb{C}^{m\times m}$ in this subsection. Let $x_{1},\ldots,x_{n}\in C([0,1]\times[0,1],\mathcal{X})$ and $y_{1},\ldots,y_{n}\in\mathcal{A}$ be given samples. We assume there exists an integrable function $f:[0,1]\times[0,1]\times\mathcal{X}\to\mathcal{A}$ such that $y_{i}=\int_{0}^{1}\int_{0}^{1}f(s,t,x_{i}(s,t))dsdt$ for $i=1,\ldots,n$. We consider an $\mathcal{A}$-valued positive definite kernel $k$ on $\mathcal{X}$, the RKHM $\mathcal{M}_{k}$ associated with $k$, and the KME $\Phi$ in $\mathcal{M}_{k}$. Let $\mu_{x}\in\mathcal{D}(\mathcal{X},\mathcal{B}(L^{2}([0,1])))$ be a $\mathcal{B}(L^{2}([0,1]))$-valued measure defined as $\mu_{x}(E)v=\left\langle\chi_{E}(x(s,\cdot)),v\right\rangle_{L^{2}([0,1])}$ for a Borel set $E$ on $\mathcal{X}$. Here, $\chi_{E}:\mathcal{X}\to\\{0,1\\}$ is the indicator function for $E$. Note that $\mu_{x}(E)$ is an integral operator whose integral kernel is $\chi_{E}(x(s,t))$, which corresponds to the Dirac measure $\tilde{\delta}_{x(s,t)}(E)$. Let $\mathcal{V}$ be the submodule of $\mathcal{M}_{k}$ spanned by $\\{\Phi(\mu_{x_{1}}),\ldots,\Phi(\mu_{x_{n}})\\}$, and let $P_{f}:\mathcal{V}\to\mathcal{B}(L^{2}([0,1]))$ be a $\mathcal{B}(L^{2}([0,1]))$-linear map (see Definition 2.19) which satisfies $P_{f}\Phi(\mu_{x_{i}})=\int_{0}^{1}\int_{0}^{1}f(s,t,x_{i}(s,t))dsdt$ for $i=1,\ldots,n$. Here, we assume the vectors $\Phi(\mu_{x_{1}}),\ldots,\Phi(\mu_{x_{n}})$ are $\mathcal{B}(L^{2}([0,1]))$-linearly independent (see Definition 2.20). We estimate $P_{f}$ by restricting it to a submodule of $\mathcal{V}$. For this purpose, we apply the PCA in RKHMs proposed in Section 6.1 and obtain principal axes $p_{1},\ldots,p_{r}$ to construct the submodule. We replace $\phi(x_{i})$ in the problem (8) with $\Phi(\mu_{x_{i}})$ and consider the problem $\inf_{\\{p_{j}\\}_{j=1}^{r}\subseteq\mathcal{M}_{k}\mbox{\footnotesize: ONS}}\;\sum_{i=1}^{n}\bigg{|}\Phi(\mu_{x_{i}})-\sum_{j=1}^{r}p_{j}\left\langle p_{j},\Phi(\mu_{x_{i}})\right\rangle_{\mathcal{M}_{k}}\bigg{|}_{\mathcal{M}_{k}}^{2}.$ (27) The projection operator onto the submodule spanned by $p_{1},\ldots,p_{r}$ is represented as $QQ^{*}$, where $Q=[p_{1},\ldots,p_{r}]$. Therefore, we estimate $P_{f}$ by $P_{f}QQ^{*}$. We can compute $P_{f}QQ^{*}$ as follows. ###### Proposition 6.17 The solution of the problem (27) is represented as $p_{j}=\sum_{i=1}^{n}\Phi(\mu_{x_{i}})c_{i,j}$ for some $c_{i,j}\in\mathcal{A}$. Let $C=[c_{i,j}]_{i,j}$. Then, the estimation $P_{f}QQ^{*}$ is computed as $P_{f}QQ^{*}=[y_{1},\ldots,y_{n}]CQ^{*}.$ The following proposition shows we can obtain a vector which attains the largest transformation by $P_{f}$. ###### Proposition 6.18 Let $u\in\mathcal{M}_{k}$ be a unique vector satisfying for any $v\in\mathcal{M}_{k}$, $\left\langle u,v\right\rangle_{\mathcal{M}_{k}}=P_{f}QQ^{*}v$. For $\epsilon>0$, let $b_{\epsilon}=(|u|_{\mathcal{M}_{k}}+\epsilon 1_{\mathcal{A}})^{-1}$ and let $v_{\epsilon}=ub_{\epsilon}$. Then, $P_{f}QQ^{*}v_{\epsilon}$ converges to $\sup_{v\in\mathcal{M}_{k},\ \|v\|_{\mathcal{M}_{k}}\leq 1}P_{f}QQ^{*}v$ (28) as $\epsilon\to 0$, where the supremum is taken with respect to a (pre) order in $\mathcal{A}$ (see Definition 2.9). If $\mathcal{A}=\mathbb{C}^{m\times m}$, then the supremum is replaced with the maximum. In this case, let $|u|_{\mathcal{M}_{k}}^{2}=a^{*}da$ be the eigenvalue decomposition of the positive semi-definite matrix $|u|_{\mathcal{M}_{k}}^{2}$ and let $b=a^{*}d^{+}a$, where the $i$-th diagonal element of $d^{+}$ is $d_{i,i}^{-1/2}$ if $d_{i,i}\neq 0$ and $0$ if $d_{i,i}=0$. Then, $ub$ is the solution of the maximization problem. Proof By the Riesz representation theorem (Proposition 4.2), there exists a unique $u\in\mathcal{M}_{k}$ satisfying for any $v\in\mathcal{M}_{k}$, $\left\langle u,v\right\rangle_{\mathcal{M}_{k}}=P_{f}QQ^{*}v$. Then, for $v\in\mathcal{M}_{k}$ which satisfies $\|v\|_{\mathcal{M}_{k}}=1$, by the Cauchy–Schwarz inequality (Lemma 2.16), we have $P_{f}QQ^{*}v=\left\langle u,v\right\rangle_{\mathcal{M}_{k}}\leq_{\mathcal{A}}|u|_{\mathcal{M}_{k}}\|v\|_{\mathcal{M}_{k}}\leq_{\mathcal{A}}|u|_{\mathcal{M}_{k}}.$ (29) The vector $v_{\epsilon}$ satisfies $\|v_{\epsilon}\|_{\mathcal{M}_{k}}\leq 1$. In addition, we have $\displaystyle|u|_{\mathcal{M}_{k}}^{2}-(|u|_{\mathcal{M}_{k}}^{2}-\epsilon^{2}1_{\mathcal{A}})\geq_{\mathcal{A}}0.$ By multiplying $(|u|_{\mathcal{M}_{k}}+\epsilon 1_{\mathcal{A}})^{-1}$ on the both sides, we have $\left\langle u,v_{\epsilon}\right\rangle_{\mathcal{M}_{k}}+\epsilon 1_{\mathcal{A}}-|u|_{\mathcal{M}_{k}}\geq_{\mathcal{A}}0$, which implies $\||u|_{\mathcal{M}_{k}}-\left\langle u,v_{\epsilon}\right\rangle_{\mathcal{M}_{k}}\|_{\mathcal{A}}\leq\epsilon$, and $\lim_{\epsilon\to 0}P_{f}QQ^{*}v_{\epsilon}=\lim_{\epsilon\to 0}\left\langle u,v_{\epsilon}\right\rangle_{\mathcal{M}_{k}}=|u|_{\mathcal{M}_{k}}$. Since $\left\langle u,v_{\epsilon}\right\rangle_{\mathcal{M}_{k}}\leq_{\mathcal{A}}d$ for any upper bound $d$ of $\\{\left\langle u,v\right\rangle_{\mathcal{M}_{k}}\ \mid\ \|v\|_{\mathcal{M}_{k}}\leq 1\\}$, $|u|_{\mathcal{M}_{k}}\leq_{\mathcal{A}}d$ holds. As a result, $|u|_{\mathcal{M}_{k}}$ is the supremum of $P_{f}QQ^{*}v$. In the case of $\mathcal{A}=\mathbb{C}^{m\times m}$, the inequality (29) is replaced with the equality by setting $v=ub$. The vector $ub_{\epsilon}$ is represented as $ub_{\epsilon}=QC^{*}[y_{1},\ldots,y_{n}]^{T}b_{\epsilon}=\sum_{i=1}^{n}\Phi(\mu_{x_{i}})d_{i}$, where $d_{i}\in\mathcal{A}$ is the $i$-th element of $CC^{*}[y_{1},\ldots,y_{n}]^{T}b_{\epsilon}\in\mathcal{A}^{n}$, and $\Phi$ is $\mathcal{A}$-linear (see Proposition 5.8). Therefore, the vector $ub_{\epsilon}$ corresponds to the $\mathcal{A}$-valued measure $\sum_{i=1}^{n}\mu_{x_{i}}d_{i}$, and if $\Phi$ is injective (see Example 5.13), the corresponding measure is unique. This means that if we transform the samples $x_{i}$ according to the measure $\sum_{i=1}^{n}\mu_{x_{i}}d_{i}$, then the transformation makes a large impact to $y_{i}$. #### 6.3.3 Numerical examples We applied our method to functional data $x_{1},\ldots,x_{n}\in C([0,1]\times[0,1],[0,1])$, where $n=30$, $x_{i}$ are polynomials of the form $x_{i}(s,t)=\sum_{j,l=0}^{5}\eta_{j,l}s^{j}t^{l}$. The coefficients $\eta_{j,l}$ of $x_{i}$ are randomly and independently drawn from the uniform distribution on $[0,0.1]$. Then, we set $y_{i}\in\mathbb{R}$ as $\displaystyle y_{i}=\int_{0}^{1}\int_{0}^{1}x_{i}(s,t)^{-\alpha+\alpha|s+t|}dsdt$ for $\alpha=3,0.5$. We set $\mathcal{A}=\mathcal{B}(L^{2}([0,1]))$ and $k(x_{1},x_{2})=\tilde{k}(x_{1},x_{2})1_{\mathcal{A}}$, where $\tilde{k}$ is a complex-valued positive definite kernel on $[0,1]$ defined as $\tilde{k}(x_{1},x_{2})=e^{-\|x_{1}-x_{2}\|_{2}^{2}}$. We applied the PCA proposed in Subsection 6.1.3 with $r=3$, and then computed $\lim_{\epsilon\to 0}ub_{\epsilon}\in\mathcal{M}_{k}$ in Proposition 6.17, which can be represented as $\Phi(\sum_{i=1}^{n}\mu_{x_{i}}d_{i})$ for some $d_{i}\in\mathcal{A}$. The parameter $\lambda$ in the objective function of the PCA was set as $0.5$. Figure 8 shows the heat map representing the value related to the integral kernel of the $\mathcal{A}$-valued measure $\sum_{i=1}^{n}\mu_{x_{i}}(E)d_{i}$ for $E=[0,0.1]$. We denote $\sum_{i=1}^{n}\mu_{x_{i}}(E)d_{i}$ by $\nu(E)$ and the integral kernel of the integral operator $\nu(E)$ by $\tilde{k}_{\nu(E)}$. As we stated in Section 6.3.2, if we transform the samples $x_{i}$ according to the measure $\nu$, then the transformation makes a large impact to $y_{i}$. Moreover, the value of $\tilde{k}_{\nu(E)}$ at $(s,t)$ corresponds to the measure at $(s,t)$. Therefore, the value of $\tilde{k}_{\nu(E)}$ at $(s,t)$ describes the impact of the effect of $t$ on $s$ to $y_{i}$. To additionally take the effect of $s$ on $t$ into consideration, we show the value of $\tilde{k}_{\nu(E)}(s,t)+\tilde{k}_{\nu(E)}(t,s)$ in Figure 8. The values for $\alpha=3$ are larger than those for $\alpha=0.5$, which implies the overall impacts to $y_{i}$ for $\alpha=3$ are larger than that for $\alpha=0.5$. Moreover, the value is large if $s+t$ is small. This is because for $x_{i}(s,t)\in[0,0.1]$, $x_{i}(s,t)^{-\alpha+\alpha|s+t|}$ is large if $s+t$ is small. Furthermore, the values around $(s,t)=(1,0)$ and $(0,1)$ are also large since $x_{i}$ has the form $x_{i}(s,t)=\sum_{j,l=0}^{5}\eta_{j,l}s^{j}t^{l}$ for $\eta_{j,l}\in[0,0.1]$ and $x_{i}(s,t)$ itself is large around $(s,t)=(1,0)$ and $(0,1)$, which results in $x_{i}(s,t)^{-\alpha+\alpha|s+t|}\approx x_{i}(s,t)$ being large. (a) $\alpha=3$ (b) $\alpha=0.5$ Figure 8: Heat map representing the value the integral kernel of $\nu([0,1])$ ### 6.4 Other applications #### 6.4.1 Maximum mean discrepancy with kernel mean embedding Maximum mean discrepancy (MMD) is a metric of measures according to the largest difference in means over a certain subset of a function space. It is also known as integral probability metric (IPM). For a set $\mathcal{U}$ of real-valued bounded measurable functions on $\mathcal{X}$ and two real-valued probability measures $\mu$ and $\nu$, MMD $\gamma(\mu,\nu,\mathcal{U})$ is defined as follows (Müller, 1997; Gretton et al., 2012): $\sup_{u\in\mathcal{U}}\bigg{|}\int_{x\in\mathcal{X}}u(x)d\mu(x)-\int_{x\in\mathcal{X}}u(x)d\nu(x)\bigg{|}.$ For example, if $\mathcal{U}$ is the unit ball of an RKHS, denoted as $\mathcal{U}_{\operatorname{RKHS}}$, the MMD can be represented using the KME $\tilde{\Phi}$ in the RKHS as $\gamma(\mu,\nu,\mathcal{U}_{\operatorname{RKHS}})=\|\tilde{\Phi}(\mu)-\tilde{\Phi}(\nu)\|_{\mathcal{H}_{\tilde{k}}}$. In addition, let $\mathcal{U}_{\operatorname{K}}=\\{u\mid\ \|u\|_{L}\leq 1\\}$ and let $\mathcal{U}_{\operatorname{D}}=\\{u\mid\ \|u\|_{\infty}+\|u\|_{L}\leq 1\\}$, where, $\|u\|_{L}:=\sup_{x\neq y}|u(x)-u(y)|/|x-y|$, and $\|u\|_{\infty}$ is the sup norm of $u$. The MMDs with $\mathcal{U}_{\operatorname{K}}$ and $\mathcal{U}_{\operatorname{D}}$ are also discussed in Rachev (1985); Dudley (2002); Sriperumbudur et al. (2012). Let $\mathcal{X}$ be a locally compact Hausdorff space, let $\mathcal{U}_{\mathcal{A}}$ be a set of $\mathcal{A}$-valued bounded and measurable functions, and let $\mu,\nu\in\mathcal{D}(\mathcal{X},\mathcal{A})$. We generalize the MMD to that for $\mathcal{A}$-valued measures as follows: $\gamma_{\mathcal{A}}(\mu,\nu,\mathcal{U}_{\mathcal{A}}):=\sup_{u\in\mathcal{U}_{\mathcal{A}}}\bigg{|}\int_{x\in\mathcal{X}}u(x)d\mu(x)-\int_{x\in\mathcal{X}}u(x)d\nu(x)\bigg{|}_{\mathcal{A}},$ where the supremum is taken with respect to a (pre) order in $\mathcal{A}$ (see Definition 2.9). Let $k$ be an $\mathcal{A}$-valued positive definite kernel and let $\mathcal{M}_{k}$ be the RKHM associated with $k$. We assume Assumption 5.3. Let $\Phi$ be the KME defined in Section 5.1. The following theorem shows that similar to the case of RKHS, if $\mathcal{U}_{\mathcal{A}}$ is the unit ball of an RKHM, the generalized MMD $\gamma_{\mathcal{A}}(\mu,\nu,\mathcal{U}_{\mathcal{A}})$ can also be represented using the proposed KME in the RKHM. ###### Proposition 6.19 Let $\mathcal{U}_{\operatorname{RKHM}}:=\\{u\in\mathcal{M}_{k}\mid\ \|u\|_{\mathcal{M}_{k}}\leq 1\\}$. Then, for $\mu,\nu\in\mathcal{D}(\mathcal{X},\mathcal{A})$, we have $\gamma_{\mathcal{A}}(\mu,\nu,\mathcal{U}_{\operatorname{RKHM}})=|\Phi(\mu)-\Phi(\nu)|_{\mathcal{M}_{k}}.$ Proof By the Cauchy–Schwarz inequality (Lemma 2.16), we have $\displaystyle\bigg{|}\int_{x\in\mathcal{X}}d\mu^{*}u(x)-\int_{x\in\mathcal{X}}d\nu^{*}u(x)\bigg{|}_{\mathcal{A}}=|\left\langle\Phi(\mu-\nu),u\right\rangle_{\mathcal{M}_{k}}|_{\mathcal{A}}$ $\displaystyle\qquad\leq_{\mathcal{A}}\|u\|_{\mathcal{M}_{k}}|\Phi(\mu-\nu)|_{\mathcal{M}_{k}}\;\leq_{\mathcal{A}}|\Phi(\mu-\nu)|_{\mathcal{M}_{k}}$ for any $u\in\mathcal{M}_{k}$ such that $\|u\|_{\mathcal{M}_{k}}\leq 1$. Let $\epsilon>0$. We put $v=\Phi(\mu-\nu)$ and $u_{\epsilon}=v(|v|_{\mathcal{M}_{k}}+\epsilon 1_{\mathcal{A}})^{-1}$. In the same manner as Proposition 6.18, $|\Phi(\mu-\nu)|_{\mathcal{M}_{k}}$ is shown to be the supremum of $|\int_{x\in\mathcal{X}}d\mu^{*}u(x)-\int_{x\in\mathcal{X}}d\nu^{*}u(x)|_{\mathcal{A}}$. Various methods with the existing MMD of real-valued probability measures are generalized to $\mathcal{A}$-valued measures by applying our MMD. Using our MMD of $\mathcal{A}$-valued measures instead of the existing MMD allows us to evaluate discrepancies between measures regarding each point of structured data such as multivariate data and functional data. For example, the following existing methods can be generalized: Two-sample test: In two-sample test, samples from two distributions (measures) are compared by computing the MMD of these measures (Gretton et al., 2012). Kernel mean matching for generative models: In generative models, MMD is used in finding points whose distribution is as close as that of input points (Jitkrittum et al., 2019). Domain adaptation: In domain adaptation, MMD is used in describing the difference between the distribution of target domain data and that of source domain data (Li et al., 2019). #### 6.4.2 Time-series data analysis with random noise Recently, random dynamical systems, which are (nonlinear) dynamical systems with random effects, have been extensively researched. Analyses of them by generalizing the discussion mentioned in Subsection 6.2.1 using the existing KME in RKHSs have been proposed (Klus et al., 2020; Hashimoto et al., 2020). We can apply our KME of $\mathcal{A}$-valued measures to generalize the analysis proposed in Subsection 6.2.2 to random dynamical systems. Then, we can extract continuous behaviors of the time evolution of functions with consideration of random noise. ## 7 Connection with existing methods In this section, we discuss connections between the proposed methods and existing methods. We show the connection with the PCA in vvRKHSs in Subsection 7.1 and an existing notion in quantum mechanics. ### 7.1 Connection with PCA in vvRKHSs We show that PCA in vvRKHSs is a special case of the proposed PCA in RKHMs. Let $\mathcal{W}$ be a Hilbert space and we set $\mathcal{A}=\mathcal{B}(\mathcal{W})$. Let $k:\mathcal{X}\times\mathcal{X}\to\mathcal{B}(\mathcal{W})$ be a $\mathcal{B}(\mathcal{W})$-valued positive definite kernel. In addition, let $x_{1},\ldots,x_{n}\in\mathcal{X}$ be given data and $w_{1,1},\ldots,w_{1,N},\ldots,w_{n,1},\ldots,w_{n,N}\in\mathcal{W}$ be fixed vectors in $\mathcal{W}$. The following proposition shows that we can reconstruct principal components of PCA in vvRKHSs by using the proposed PCA in RKHMs. ###### Proposition 7.1 Let $W_{j}:\mathcal{X}\to\mathcal{W}$ be a map satisfying $W_{j}(x_{i})=w_{i,j}$ for $j=1,\ldots,N$, let $W=[W_{1},\ldots,W_{N}]$, and let $\hat{k}:\mathcal{X}\times\mathcal{X}\to\mathbb{C}^{N\times N}$ be defined as $\hat{k}(x,y)=W(x)^{*}k(x,y)W(y)$. Let $\\{q_{1},\ldots,q_{r}\\}\subseteq\mathcal{F}_{\hat{k}}$ is a solution of the minimization problem $\min_{\\{q_{j}\\}_{j=1}^{r}\subseteq\mathcal{F}_{\hat{k}}\mbox{\footnotesize: ONS}}\;\sum_{i=1}^{n}\operatorname{tr}\Big{(}\big{|}\phi(x_{i})-\sum_{j=1}^{r}q_{j}\left\langle q_{j},\phi(x_{i})\right\rangle_{\mathcal{M}_{\hat{k}}}\big{|}_{\mathcal{M}_{\hat{k}}}^{2}\Big{)},$ (30) where $\mathcal{F}_{k}=\\{v\in\mathcal{M}_{k}\mid\ v(x)\mbox{ is a rank $1$ operator for any }x\in\mathcal{X}\\}$. In addition, let $p_{1},\ldots,p_{r}\in\mathcal{H}_{k}^{\operatorname{v}}$ be the solution of the minimization problem $\min_{\\{p_{j}\\}_{j=1}^{r}\subseteq\mathcal{H}_{k}^{\operatorname{v}}\mbox{\footnotesize: ONS}}\;\sum_{i=1}^{n}\sum_{l=1}^{N}\bigg{\|}\phi(x_{i})w_{i,l}-\sum_{j=1}^{r}p_{j}\left\langle p_{j},\phi(x_{i})w_{i,l}\right\rangle_{\mathcal{H}_{k}^{\operatorname{v}}}\bigg{\|}_{\mathcal{H}_{k}^{\operatorname{v}}}^{2}.$ (31) Then, $\|(\langle q_{j},\hat{\phi}(x_{i})\rangle_{\mathcal{M}_{\hat{k}}})_{l}\|_{\mathbb{C}^{N}}=\left\langle p_{j},\phi(x_{i})w_{i,l}\right\rangle_{\mathcal{H}_{k}^{\operatorname{v}}}$ for $i=1,\ldots,n$, $j=1,\ldots,r$, and $l=1,\ldots,N$. Here, $(\langle q_{j},\hat{\phi}(x_{i})\rangle_{\mathcal{M}_{\hat{k}}})_{l}$ is the $l$-th column of the matrix $\langle q_{j},\hat{\phi}(x_{i})\rangle_{\mathcal{M}_{\hat{k}}}\in\mathbb{C}^{N\times N}$. Proof Let $\mathbf{G}\in(\mathbb{C}^{N\times N})^{n\times n}$ be defined as $\mathbf{G}_{i,j}=\hat{k}(x_{i},x_{j})$. By Proposition 6.8, any solution of the problem (30) is represented as $q_{j}=\sum_{i=1}^{n}\hat{\phi}(x_{i})c_{i,j}$, where $j=1,\ldots,r$ and $[c_{1,j},\ldots,c_{n,j}]^{T}=\lambda_{j}^{-1/2}\mathbf{v}_{j}u^{*}$ for any normalized vector $u\in\mathbb{C}^{N}$. Here, $\lambda_{j}$ are the largest $r$ eigenvalues and $\mathbf{v}_{j}$ are the corresponding orthonormal eigenvectors of the matrix $\mathbf{G}$. Therefore, by the definition of $\hat{k}$, the principal components are calculated as $\langle q_{j},\hat{\phi}(x_{i})\rangle_{\mathcal{M}_{\hat{k}}}^{*}=\lambda_{j}^{-1/2}W(x_{i})^{*}[k(x_{i},x_{1})W(x_{1}),\ldots,k(x_{i},x_{n})W(x_{n})]\mathbf{v}_{j}u^{*}.$ On the other hand, in the same manner as Proposition 6.8, the solution of the problem (31) is shown to be represented as $p_{j}=\sum_{i=1}^{n}\sum_{l=1}^{N}\phi(x_{i})w_{i,l}\alpha_{(i-1)N+l,j}$, where $j=1,\ldots,r$ and $[\alpha_{1,j},\ldots,\alpha_{Nn,j}]^{T}=\lambda_{j}^{-1/2}\mathbf{v}_{j}$. Therefore, the principal components are calculated as $\overline{\left\langle p_{j},\phi(x_{i})w_{i,l}\right\rangle_{\mathcal{H}_{k}^{\operatorname{v}}}}=\lambda_{j}^{-1/2}W_{l}(x_{i})^{*}[k(x_{i},x_{1})W(x_{1}),\ldots,k(x_{i},x_{n})W(x_{n})]\mathbf{v}_{j},$ which completes the proof of the proposition. ### 7.2 Connection with quantum mechanics Positive operator-valued measures play an important role in quantum mechanics. A positive operator-valued measure is defined as an $\mathcal{A}$-valued measure $\mu$ such that $\mu(\mathcal{X})=I$ and $\mu(E)$ is positive for any Borel set $E$. It enables us to extract information of the probabilities of outcomes from a state (Peres and Terno, 2004; Holevo, 2011). We show that the existing inner product considered for quantum states (Balkir, 2014; Deb, 2016) is generalized with our KME of positive operator-valued measures. Let $\mathcal{X}=\mathbb{C}^{m}$ and $\mathcal{A}=\mathbb{C}^{m\times m}$. Let $\rho\in\mathcal{A}$ be a positive semi-definite matrix with unit trace, called a density matrix. A density matrix describes the states of a quantum system, and information about outcomes is described as measure $\mu\rho\in\mathcal{D}(\mathcal{X},\mathcal{A})$. We have the following proposition. Here, we use the bra-ket notation, i.e., $|\alpha\rangle\in\mathcal{X}$ represents a (column) vector in $\mathcal{X}$, and $\langle\alpha|$ is defined as $\langle\alpha|=|\alpha\rangle^{*}$: ###### Proposition 7.2 Assume $\mathcal{X}=\mathbb{C}^{m}$, $\mathcal{A}=\mathbb{C}^{m\times m}$, and $k:\mathcal{X}\times\mathcal{X}\to\mathcal{A}$ is a positive definite kernel defined as $k(|\alpha\rangle,|\beta\rangle)=|\alpha\rangle\langle\alpha|\beta\rangle\langle\beta|$. If $\mu$ is represented as $\mu=\sum_{i=1}^{m}\delta_{|\psi_{i}\rangle}|\psi_{i}\rangle\langle\psi_{i}|$ for an orthonormal basis $\\{|\psi_{1}\rangle,\ldots,|\psi_{m}\rangle\\}$ of $\mathcal{X}$, then for any $\rho_{1},\rho_{2}\in\mathcal{A}$, $\operatorname{tr}(\left\langle\Phi(\mu\rho_{1}),\Phi(\mu\rho_{2})\right\rangle_{\mathcal{M}_{k}})=\left\langle\rho_{1},\rho_{2}\right\rangle_{\operatorname{HS}}$ holds. Here, $\left\langle\cdot,\cdot\right\rangle_{\operatorname{HS}}$ is the Hilbert–Schmidt inner product. Proof Let $M_{i}=|\psi_{i}\rangle\langle\psi_{i}|$ for $i=1,\ldots,m$. The inner product between $\Phi(\mu\rho_{1})$ and $\Phi(\mu\rho_{2})$ is calculated as follows: $\displaystyle\left\langle\Phi(\mu\rho_{1}),\Phi(\mu\rho_{2})\right\rangle_{\mathcal{M}_{k}}$ $\displaystyle=\int_{x\in\mathcal{X}}\int_{y\in\mathcal{X}}\rho_{1}^{*}\mu^{*}(x)k(x,y)\mu\rho_{2}(y)=\sum_{i,j=1}^{m}\rho_{1}^{*}M_{i}k(|\psi_{i}\rangle,|\psi_{j}\rangle)M_{j}\rho_{2}.$ Since the identity $k(|\psi_{i}\rangle,|\psi_{j}\rangle)=M_{i}M_{j}$ holds and $\\{|\psi_{1}\rangle,\ldots,|\psi_{m}\rangle\\}$ is orthonormal, we have $\left\langle\Phi(\mu\rho_{1}),\Phi(\mu\rho_{2})\right\rangle_{\mathcal{M}_{k}}=\sum_{i=1}^{m}\rho_{1}^{*}M_{i}\rho_{2}$. By using the identity $\sum_{i=1}^{m}M_{i}=I$, we have $\operatorname{tr}\bigg{(}\sum_{i=1}^{m}\rho_{1}^{*}M_{i}\rho_{2}\bigg{)}=\operatorname{tr}\bigg{(}\sum_{i=1}^{m}M_{i}\rho_{2}\rho_{1}^{*}\bigg{)}=\operatorname{tr}(\rho_{2}\rho_{1}^{*}),$ which completes the proof of the proposition. In previous studies (Balkir, 2014; Deb, 2016), the Hilbert–Schmidt inner product between density matrices was considered to represent similarities between two quantum states. Liu and Rebentrost (2018) considered the Hilbert–Schmidt inner product between square roots of density matrices. Theorem 7.2 shows that these inner products are represented via our KME in RKHMs. ## 8 Conclusions and future works In this paper, we proposed a new data analysis framework with RKHM and developed a KME in RKHMs for analyzing distributions. We showed the theoretical validity for applying those to data analysis. Then, we applied it to kernel PCA, time-series data analysis, and analysis of interaction effects in finite or infinite dimensional data. RKHM is a generalization of RKHS in terms of $C^{*}$-algebra, and we can extract rich information about structures in data such as functional data by using $C^{*}$-algebras. For example, we can reduce multi-variable functional data to functions of single variable by considering the space of functions of single variables as a $C^{*}$-algebra and then by applying the proposed PCA in RKHMs. Moreover, we can extract information of interaction effects in continuously distributed spatio data by considering the space of bounded linear operators on a function space as a $C^{*}$-algebra. As future works, we will address $C^{*}$-algebra-valued supervised problems on the basis of the representer theorem (Theorem 4.8) and apply the proposed KME in RKHMs to quantum mechanics. ## Acknowledgments We would like to thank Dr. Tomoki Mihara, whose comments improve the mathematical rigorousness of this paper. This work was partially supported by JST CREST Grant Number JPMJCR1913. ## Appendix A Proofs of the lemmas and propositions in Section 2.5 ### Proof of Proposition 4.5 (Existence) For $u,v\in\mathcal{M}_{k}$, there exist $u_{i},v_{i}\in\mathcal{M}_{k,0}\ (i=1,2,\ldots)$ such that $v=\lim_{i\to\infty}v_{i}$ and $w=\lim_{i\to\infty}w_{i}$. By the Cauchy- Schwarz inequality (Lemma 2.16), the following inequalities hold: $\displaystyle\|\left\langle u_{i},v_{i}\right\rangle_{\mathcal{M}_{k}}-\left\langle u_{j},v_{j}\right\rangle_{\mathcal{M}_{k}}\|_{\mathcal{A}}$ $\displaystyle\leq\|\left\langle u_{i},v_{i}-v_{j}\right\rangle_{\mathcal{M}_{k}}\|_{\mathcal{A}}+\|\left\langle u_{i}-u_{j},u_{j}\right\rangle_{\mathcal{M}_{k}}\|_{\mathcal{A}}$ $\displaystyle\leq\|u_{i}\|_{\mathcal{M}_{k}}\;\|v_{i}-v_{j}\|_{\mathcal{M}_{k}}+\|u_{i}-u_{j}\|_{\mathcal{M}_{k}}\;\|v_{j}\|_{\mathcal{M}_{k}}$ $\displaystyle\to 0\ (i,j\to\infty),$ which implies $\\{\left\langle u_{i},v_{i}\right\rangle_{\mathcal{M}_{k}}\\}_{i=1}^{\infty}$ is a Cauchy sequence in $\mathcal{A}$. By the completeness of $\mathcal{A}$, there exists a limit $\lim_{i\to\infty}\left\langle u_{i},v_{i}\right\rangle_{\mathcal{M}_{k}}$. (Well-definedness) Assume there exist $u^{\prime}_{i},v^{\prime}_{i}\in\mathcal{M}_{k,0}\ (i=1,2,\ldots)$ such that $u=\lim_{i\to\infty}u_{i}=\lim_{i\to\infty}u^{\prime}_{i}$ and $v=\lim_{i\to\infty}v_{i}=\lim_{i\to\infty}v^{\prime}_{i}$. By the Cauchy- Schwarz inequality (Lemma 2.16), we have $\|\left\langle u_{i},v_{i}\right\rangle_{\mathcal{M}_{k}}-\left\langle u^{\prime}_{i},v^{\prime}_{i}\right\rangle_{\mathcal{M}_{k}}\|_{\mathcal{A}}\leq\|u_{i}\|_{\mathcal{M}_{k}}\|v_{i}-v^{\prime}_{i}\|_{\mathcal{M}_{k}}+\|u_{i}-u^{\prime}_{i}\|_{\mathcal{M}_{k}}\|v^{\prime}_{i}\|_{\mathcal{M}_{k}}\to 0\ (i\to\infty),$ which implies $\lim_{i\to\infty}\left\langle u_{i},v_{i}\right\rangle_{\mathcal{M}_{k}}=\lim_{i\to\infty}\left\langle u^{\prime}_{i},v^{\prime}_{i}\right\rangle_{\mathcal{M}_{k}}$. (Injectivity) For $u,v\in\mathcal{M}_{k}$, we assume $\left\langle\phi(x),u\right\rangle_{\mathcal{M}_{k}}=\left\langle\phi(x),v\right\rangle_{\mathcal{M}_{k}}$ for $x\in\mathcal{X}$. By the linearity of $\left\langle\cdot,\cdot\right\rangle_{\mathcal{M}_{k}}$, $\left\langle p,u\right\rangle_{\mathcal{M}_{k}}=\left\langle p,v\right\rangle_{\mathcal{M}_{k}}$ holds for $p\in\mathcal{M}_{k,0}$. For $p\in\mathcal{M}_{k}$, there exist $p_{i}\in\mathcal{M}_{k,0}\ (i=1,2,\ldots)$ such that $p=\lim_{i\to\infty}p_{i}$. Therefore, $\left\langle p,u-v\right\rangle_{\mathcal{M}_{k}}=\lim_{i\to\infty}\left\langle p_{i},u-v\right\rangle_{\mathcal{M}_{k}}=0$. As a result, $\left\langle u-v,u-v\right\rangle_{\mathcal{M}_{k}}=0$ holds by setting $p=u-v$, which implies $u=v$. ### Proof of Proposition 4.6 We define $\Psi:\mathcal{M}_{k,0}\to\mathcal{M}$ as an $\mathcal{A}$-linear map that satisfies $\Psi(\phi(x))=\psi(x)$. We show $\Psi$ can be extended to a unique $\mathcal{A}$-linear bijection map on $\mathcal{M}_{k}$ , which preserves the inner product. (Uniqueness) The uniqueness follows by the definition of $\Psi$. (Inner product preservation) For $x,y\in\mathcal{X}$, we have $\left\langle\Psi(\phi(x)),\Psi(\phi(y))\right\rangle_{\mathcal{M}_{k}}=\left\langle\psi(x),\psi(y)\right\rangle_{\mathcal{M}}=k(x,y)=\left\langle\phi(x),\phi(y)\right\rangle_{\mathcal{M}_{k}}.$ Since $\Psi$ is $\mathcal{A}$-linear, $\Psi$ preserves the inner products between arbitrary $u,v\in\mathcal{M}_{k,0}$. (Well-definedness) Since $\Phi$ preserves the inner product, if $\\{v_{i}\\}_{i=1}^{\infty}\subseteq\mathcal{M}_{k}$ is a Cauchy sequence, $\\{\Psi(v_{i})\\}_{i=1}^{\infty}\subseteq\mathcal{M}$ is also a Cauchy sequence. Therefore, by the completeness of $\mathcal{M}$, $\Psi$ also preserves the inner product in $\mathcal{M}_{k}$, and for $v\in\mathcal{M}_{k}$, $\|\Psi(v)\|_{\mathcal{M}}=\|v\|_{\mathcal{M}_{k}}$ holds. As a result, for $v\in\mathcal{M}_{k}$, if $v=0$, $\|\Psi(v)\|_{\mathcal{M}}=\|v\|_{\mathcal{M}_{k}}=0$ holds. This implies $\Psi(v)=0$. (Injectivity) For $u,v\in\mathcal{M}_{k}$, if $\Psi(u)=\Psi(v)$, then $0=\|\Psi(u)-\Psi(v)\|_{\mathcal{M}}=\|u-v\|_{\mathcal{M}_{k}}$ holds since $\Psi$ preserves the inner product, which implies $u=v$. (Surjectivity) It follows directly by the condition $\overline{\\{\sum_{i=0}^{n}\psi(x_{i})c_{i}\mid\ x_{i}\in\mathcal{X},\ c_{i}\in\mathcal{A}\\}}=\mathcal{M}$. ### Proof of Lemma 4.10 Let $k$ be an $\mathcal{A}$-valued positive definite kernel defined in Definition 2.21. Let $w\in\mathcal{W}$. For $n\in\mathbb{N}$, $w_{1},\ldots,w_{n}\in\mathcal{W}$, let $c_{i}\in\mathcal{B}(\mathcal{W})$ be defined as $c_{i}h:=\left\langle w,h\right\rangle_{\mathcal{W}}/\left\langle w,w\right\rangle_{\mathcal{W}}w_{i}$ for $h\in\mathcal{W}$. Since $w_{i}=c_{i}w$ holds, the following equalities are derived for $x_{1},\ldots,x_{n}\in\mathcal{X}$: $\displaystyle\sum_{i,j=1}^{n}\left\langle w_{i},k(x_{i},x_{j})w_{j}\right\rangle_{\mathcal{W}}$ $\displaystyle=\sum_{i,j=1}^{n}\left\langle c_{i}w,k(x_{i},x_{j})c_{j}w\right\rangle_{\mathcal{W}}=\bigg{\langle}w,\sum_{i,j=1}^{n}c_{i}^{*}k(x_{i},x_{j})c_{i}w\bigg{\rangle}_{\mathcal{W}}.$ By the positivity of $\sum_{i,j=1}^{n}c_{i}^{*}k(x_{i},x_{j})c_{j}$, $\langle w,\sum_{i,j=1}^{n}c_{i}^{*}k(x_{i},x_{j})c_{j}w\rangle_{\mathcal{W}}\geq 0$ holds, which implies $k$ is an operator valued positive definite kernel defined in Definition 2.2. On the other hand, let $k$ be an operator valued positive definite kernel defined in Definition 2.2. Let $v\in\mathcal{W}$. For $n\in\mathbb{N}$, $c_{1},\ldots,c_{n}\in\mathcal{A}$ and $x_{1},\ldots,x_{n}\in\mathcal{X}$, the following equality is derived: $\displaystyle\bigg{\langle}w,\sum_{i,j=1}^{n}c_{i}^{*}k(x_{i},x_{j})c_{j}w\bigg{\rangle}_{\mathcal{W}}\\!\\!\\!\\!=\sum_{i,j=1}^{n}\left\langle c_{i}w,k(x_{i},x_{j})c_{j}w\right\rangle_{\mathcal{W}}.$ By Definition 2.2, $\sum_{i,j=1}^{n}\left\langle c_{i}w,k(x_{i},x_{j})c_{j}w\right\rangle_{\mathcal{W}}\geq 0$ holds, which implies $k$ is an $\mathcal{A}$-valued positive definite kernel defined in Definition 2.21. ## Appendix B $\mathcal{A}$-valued measure and integral We introduce $\mathcal{A}$-valued measure and integral in preparation for defining a KME in RKHMs. $\mathcal{A}$-valued measure and integral are special cases of vector measure and integral (Dinculeanu, 1967, 2000), respectively. Here, we review these notions especially for the case of $\mathcal{A}$-valued ones. The notions of measures and the Lebesgue integrals are generalized to $\mathcal{A}$-valued. The left and right integral of an $\mathcal{A}$-valued function $u$ with respect to an $\mathcal{A}$-valued measure $\mu$ is defined through $\mathcal{A}$-valued step functions. ###### Definition B.1 ($\mathcal{A}$-valued measure) Let $\varSigma$ be a $\sigma$-algebra on $\mathcal{X}$. 1. 1. An $\mathcal{A}$-valued map $\mu:\varSigma\to\mathcal{A}$ is called a (countably additive) $\mathcal{A}$-vaued measure if $\mu(\bigcup_{i=1}^{\infty}E_{i})=\sum_{i=1}^{\infty}\mu(E_{i})$ for all countable collections $\\{E_{i}\\}_{i=1}^{\infty}$ of pairwise disjoint sets in $\varSigma$. 2. 2. An $\mathcal{A}$-valued measure $\mu$ is said to be finite if $|\mu|(E):=\sup\\{\sum_{i=1}^{n}\|\mu(E_{i})\|_{\mathcal{A}}\mid\ n\in\mathbb{N},\ \\{E_{i}\\}_{i=1}^{n}\mbox{ is a finite partition of }E\in\varSigma\\}<\infty$. We call $|\mu|$ the total variation of $\mu$. 3. 3. An $\mathcal{A}$-valued measure $\mu$ is said to be regular if for all $E\in\varSigma$ and $\epsilon>0$, there exist a compact set $K\subseteq E$ and an open set $G\supseteq E$ such that $\|\mu(F)\|_{\mathcal{A}}\leq\epsilon$ for any $F\subseteq G\setminus K$. The regularity corresponds to the continuity of $\mathcal{A}$-valued measures. 4. 4. An $\mathcal{A}$-valued measure $\mu$ is called a Borel measure if $\varSigma=\mathcal{B}$, where $\mathcal{B}$ is the Borel $\sigma$-algebra on $\mathcal{X}$ ($\sigma$-algebra generated by all compact subsets of $\mathcal{X}$). The set of all $\mathcal{A}$-valued finite regular Borel measures is denoted as $\mathcal{D}(\mathcal{X},\mathcal{A})$. ###### Definition B.2 ($\mathcal{A}$-valued Dirac measure) For $x\in\mathcal{X}$, we define $\delta_{x}\in\mathcal{D}(\mathcal{X},\mathcal{A})$ as $\delta_{x}(E)=1_{\mathcal{A}}$ for $x\in E$ and $\delta_{x}(E)=0$ for $x\notin E$. The measure $\delta_{x}$ is referred to as the $\mathcal{A}$-valued Dirac measure at $x$. Similar to the Lebesgue integrals, an integral of an $\mathcal{A}$-valued function with respect to an $\mathcal{A}$-valued measure is defined through $\mathcal{A}$-valued step functions. ###### Definition B.3 (Step function) An $\mathcal{A}$-valued map $s:\mathcal{X}\to\mathcal{A}$ is called a step function if $s(x)=\sum_{i=1}^{n}c_{i}\chi_{E_{i}}(x)$ for some $n\in\mathbb{N}$, $c_{i}\in\mathcal{A}$ and finite partition $\\{E_{i}\\}_{i=1}^{n}$ of $\mathcal{X}$, where $\chi_{E}:\mathcal{X}\to\\{0,1\\}$ is the indicator function for $E\in\mathcal{B}$. The set of all $\mathcal{A}$-valued step functions on $\mathcal{X}$ is denoted as $\mathcal{S}(\mathcal{X},\mathcal{A})$. ###### Definition B.4 (Integrals of functions in $\mathcal{S}(\mathcal{X},\mathcal{A})$) For $s\in\mathcal{S}(\mathcal{X},\mathcal{A})$ and $\mu\in\mathcal{D}(\mathcal{X},\mathcal{A})$, the left and right integrals of $s$ with respect to $\mu$ are respectively defined as $\int_{x\in\mathcal{X}}s(x)d\mu(x):=\sum_{i=1}^{n}c_{i}\mu(E_{i}),\quad\int_{x\in\mathcal{X}}d\mu(x)s(x):=\sum_{i=1}^{n}\mu(E_{i})c_{i}.$ As we explain below, the integrals of step functions are extended to those of “integrable functions”. For a real positive finite measure $\nu$, let ${L}^{1}_{\nu}(\mathcal{X},\mathcal{A})$ be the set of all $\mathcal{A}$-valued $\nu$-Bochner integrable functions on $\mathcal{X}$, i.e., if $u\in{L}^{1}_{\nu}(\mathcal{X},\mathcal{A})$, there exists a sequence $\\{s_{i}\\}_{i=1}^{\infty}\subseteq\mathcal{S}(\mathcal{X},\mathcal{A})$ of step functions such that $\lim_{i\to\infty}\int_{x\in\mathcal{X}}\|u(x)-s_{i}(x)\|_{\mathcal{A}}d\nu(x)=0$ (Diestel, 1984, Chapter IV). Note that $u\in{L}^{1}_{\nu}(\mathcal{X},\mathcal{A})$ if and only if $\int_{x\in\mathcal{X}}\|u(x)\|_{\mathcal{A}}d\nu(x)<\infty$, and ${L}^{1}_{\nu}(\mathcal{X},\mathcal{A})$ is a Banach $\mathcal{A}$-module (i.e., a Banach space equipped with an $\mathcal{A}$-module structure) with respect to the norm defined as $\|u\|_{{L}^{1}_{\nu}(\mathcal{X},\mathcal{A})}=\int_{x\in\mathcal{X}}\|u(x)\|_{\mathcal{A}}d\nu(x)$. ###### Definition B.5 (Integrals of functions in ${L}^{1}_{|\mu|}(\mathcal{X},\mathcal{A})$) For $u\in{L}^{1}_{|\mu|}(\mathcal{X},\mathcal{A})$, the left and right integrals of $u$ with respect to $\mu$ is respectively defined as $\lim_{i\to\infty}\int_{x\in\mathcal{X}}d\mu(x)s_{i}(x),\quad\lim_{i\to\infty}\int_{x\in\mathcal{X}}s_{i}(x)d\mu(x),$ where $\\{s_{i}\\}_{i=1}^{\infty}\subseteq\mathcal{S}(\mathcal{X},\mathcal{A})$ is a sequence of step functions whose ${L}^{1}_{\nu}(\mathcal{X},\mathcal{A})$-limit is $u$. Note that since $\mathcal{A}$ is not commutative in general, the left and right integrals do not always coincide. There is also a stronger notion for integrability. An $\mathcal{A}$-valued function $u$ on $\mathcal{X}$ is said to be totally measurable if it is a uniform limit of a step function, i.e., there exists a sequence $\\{s_{i}\\}_{i=1}^{\infty}\subseteq\mathcal{S}(\mathcal{X},\mathcal{A})$ of step functions such that $\lim_{i\to\infty}\sup_{x\in\mathcal{X}}\|u(x)-s_{i}(x)\|_{\mathcal{A}}=0$. We denote by $\mathcal{T}(\mathcal{X},\mathcal{A})$ the set of all $\mathcal{A}$-valued totally measurable functions on $\mathcal{X}$. Note that if $u\in\mathcal{T}(\mathcal{X},\mathcal{A})$, then $u\in{L}^{1}_{|\mu|}(\mathcal{X},\mathcal{A})$ for any $\mu\in\mathcal{D}(\mathcal{X},\mathcal{A})$. In fact, the continuous functions in ${C}_{0}(\mathcal{X},\mathcal{A})$ is totally measurable (see Definition 5.1 for the definition of ${C}_{0}(\mathcal{X},\mathcal{A})$). ###### Proposition B.6 The space $C_{0}(\mathcal{X},\mathcal{A})$ is contained in $\mathcal{T}(\mathcal{X},\mathcal{A})$. Moreover, for any real positive finite regular measure $\nu$, it is dense in ${L}^{1}_{\nu}(\mathcal{X},\mathcal{A})$ with respect to $\|\cdot\|_{{L}^{1}_{\nu}(\mathcal{X},\mathcal{A})}$. For further details, refer to Dinculeanu (1967, 2000). ## Appendix C Proofs of the propositions and theorem in Section 5.2 Before proving the propositions and theorem, we introduce some definitions and show fundamental properties which are related to the propositions and theorem. ###### Definition C.1 ($\mathcal{A}$-dual) For a Banach $\mathcal{A}$-module $\mathcal{M}$, the $\mathcal{A}$-dual of $\mathcal{M}$ is defined as $\mathcal{M}^{\prime}:=\\{f:\mathcal{M}\to\mathcal{A}\mid\ f\mbox{ is bounded and $\mathcal{A}$-linear}\\}$. Note that for a right Banach $\mathcal{A}$-module $\mathcal{M}$, $\mathcal{M}^{\prime}$ is a left Banach $\mathcal{A}$-module. ###### Definition C.2 (Orthogonal complement) For an $\mathcal{A}$-submodule $\mathcal{M}_{0}$ of a Banach $\mathcal{A}$-module $\mathcal{M}$, the orthogonal complement of $\mathcal{M}_{0}$ is defined as a closed submodule $\mathcal{M}_{0}^{\perp}:=\bigcap_{u\in\mathcal{M}_{0}}\\{f\in\mathcal{M}^{\prime}\mid\ f(u)=0\\}$ of $\mathcal{M}^{\prime}$. In addition, for an $\mathcal{A}$-submodule $\mathcal{N}_{0}$ of $\mathcal{M}^{\prime}$, the orthogonal complement of $\mathcal{N}_{0}$ is defined as a closed submodule $\mathcal{N}_{0}^{\perp}:=\bigcap_{f\in\mathcal{N}_{0}}\\{u\in\mathcal{M}\mid\ f(u)=0\\}$ of $\mathcal{M}$. Note that for a von Neumann $\mathcal{A}$-module $\mathcal{M}$, by Proposition 4.2, $\mathcal{M}^{\prime}$ and $\mathcal{M}$ are isomorphic. The following lemma shows a connection between an orthogonal complement and the density property. ###### Lemma C.3 For a Banach $\mathcal{A}$-module $\mathcal{M}$ and its submodule $\mathcal{M}_{0}$, $\mathcal{M}_{0}^{\perp}=\\{0\\}$ if $\mathcal{M}_{0}$ is dense in $\mathcal{M}$. Proof We first show $\overline{\mathcal{M}_{0}}\subseteq(\mathcal{M}_{0}^{\perp})^{\perp}$. Let $u\in\mathcal{M}_{0}$. By the definition of orthogonal complements, $u\in(\mathcal{M}_{0}^{\perp})^{\perp}$. Since $(\mathcal{M}_{0}^{\perp})^{\perp}$ is closed, $\overline{\mathcal{M}_{0}}\subseteq(\mathcal{M}_{0}^{\perp})^{\perp}$. If $\mathcal{M}_{0}$ is dense in $\mathcal{M}$, $\mathcal{M}\subseteq(\mathcal{M}_{0}^{\perp})^{\perp}$ holds, which means $\mathcal{M}_{0}^{\perp}=\\{0\\}$. Moreover, in the case of $\mathcal{A}=\mathbb{C}^{m\times m}$, a generalization of the Riesz–Markov representation theorem for $\mathcal{D}(\mathcal{X},\mathcal{A})$ holds. ###### Proposition C.4 (Riesz–Markov representation theorem for $\mathbb{C}^{m\times m}$-valued measures) Let $\mathcal{A}=\mathbb{C}^{m\times m}$. There exists an isomorphism between $\mathcal{D}(\mathcal{X},\mathcal{A})$ and ${C}_{0}(\mathcal{X},\mathcal{A})^{\prime}$. Proof For $f\in{C}_{0}(\mathcal{X},\mathcal{A})^{\prime}$, let $f_{i,j}\in{C}_{0}(\mathcal{X},\mathbb{C})^{\prime}$ be defined as $f_{i,j}(u)=(f(u1_{\mathcal{A}}))_{i,j}$ for $u\in{C}_{0}(\mathcal{X},\mathbb{C})$. Then, by the Riesz–Markov representation theorem for complex-valued measure, there exists a unique finite complex-valued regular measure $\mu_{i,j}$ such that $f_{i,j}(u)=\int_{x\in\mathcal{X}}u(x)d\mu_{i,j}(x)$. Let $\mu(E):=[\mu_{i,j}(E)]_{i,j}$ for $E\in\mathcal{B}$. Then, $\mu\in\mathcal{D}(\mathcal{X},\mathcal{A})$, and we have $\displaystyle f(u)$ $\displaystyle=f\bigg{(}\sum_{l,l^{\prime}=1}^{m}u_{l,l^{\prime}}e_{l,l^{\prime}}\bigg{)}=\sum_{l,l^{\prime}=1}^{m}[f_{i,j}(u_{l,l^{\prime}})]_{i,j}e_{l,l^{\prime}}$ $\displaystyle=\sum_{l,l^{\prime}=1}^{m}\bigg{[}\int_{x\in\mathcal{X}}u_{l,l^{\prime}}(x)d\mu_{i,j}(x)\bigg{]}_{i,j}e_{l,l^{\prime}}=\int_{x\in\mathcal{X}}d\mu(x)u(x),$ where $e_{i,j}$ is an $m\times m$ matrix whose $(i,j)$-element is $1$ and all the other elements are $0$. Therefore, if we define $h^{\prime}:{C}_{0}(\mathcal{X},\mathcal{A})^{\prime}\to\mathcal{D}(\mathcal{X},\mathcal{A})$ as $f\mapsto\mu$, $h^{\prime}$ is the inverse of $h$, which completes the proof of the proposition. ### C.1 Proofs of Propositions 5.11 and 5.12 To show Propositions 5.11 and 5.12, the following lemma is used. ###### Lemma C.5 $\Phi:\mathcal{D}(\mathcal{X},\mathcal{A})\to\mathcal{M}_{k}$ is injective if and only if $\left\langle\Phi(\mu),\Phi(\mu)\right\rangle_{\mathcal{M}_{k}}\neq 0$ for any nonzero $\mu\in\mathcal{D}(\mathcal{X},\mathcal{A})$. Proof ($\Rightarrow$) Suppose there exists a nonzero $\mu\in\mathcal{D}(\mathcal{X},\mathcal{A})$ such that $\left\langle\Phi(\mu),\Phi(\mu)\right\rangle_{\mathcal{M}_{k}}=0$. Then, $\Phi(\mu)=\Phi(0)=0$ holds, and thus, $\Phi$ is not injective. ($\Leftarrow$) Suppose $\Phi$ is not injective. Then, there exist $\mu,\nu\in\mathcal{D}(\mathcal{X},\mathcal{A})$ such that $\Phi(\mu)=\Phi(\nu)$ and $\mu\neq\nu$, which implies $\Phi(\mu-\nu)=0$ and $\mu-\nu\neq 0$. We now show Propositions 5.11 and 5.12. Proof of Theorem 5.11 Let $\mu\in\mathcal{D}(\mathcal{X},\mathcal{A})$, $\mu\neq 0$. We have $\displaystyle\left\langle\Phi(\mu),\Phi(\mu)\right\rangle$ $\displaystyle=\int_{x\in\mathbb{R}^{d}}\int_{y\in\mathbb{R}^{d}}d\mu^{*}(x)k(x,y)d\mu(y)$ $\displaystyle=\int_{x\in\mathbb{R}^{d}}\int_{y\in\mathbb{R}^{d}}d\mu^{*}(x)\int_{\omega\in\mathbb{R}^{d}}e^{-\sqrt{-1}(y-x)^{T}\omega}d\lambda(\omega)d\mu(y)$ $\displaystyle=\int_{\omega\in\mathbb{R}^{d}}\int_{x\in\mathbb{R}^{d}}e^{\sqrt{-1}x^{T}\omega}d\mu^{*}(x)d\lambda(\omega)\int_{y\in\mathbb{R}^{d}}e^{-\sqrt{-1}y^{T}\omega}d\mu(y)$ $\displaystyle=\int_{\omega\in\mathbb{R}^{d}}\hat{\mu}(\omega)^{*}d\lambda(\omega)\hat{\mu}(\omega).$ Assume $\hat{\mu}=0$. Then, $\int_{x\in\mathcal{X}}u(x)d\mu(x)=0$ for any $u\in{C}_{0}(\mathcal{X},\mathcal{A})$ holds, which implies $\mu\in{C}_{0}(\mathcal{X},\mathcal{A})^{\perp}=\\{0\\}$ by Proposition C.4 and Lemma C.3. Thus, ${\mu}=0$. In addition, by the assumption, $\operatorname{supp}(\lambda)=\mathbb{R}^{d}$ holds. As a result, $\int_{\omega\in\mathbb{R}^{d}}\hat{\mu}(\omega)^{*}d\lambda(\omega)\hat{\mu}(\omega)\neq 0$ holds. By Lemma C.5, $\Phi$ is injective. Proof of Theorem 5.12 Let $\mu\in\mathcal{D}(\mathcal{X},\mathcal{A})$, $\mu\neq 0$. We have $\displaystyle\left\langle\Phi(\mu),\Phi(\mu)\right\rangle$ $\displaystyle=\int_{x\in\mathbb{R}^{d}}\int_{y\in\mathbb{R}^{d}}d\mu^{*}(x)k(x,y)d\mu(y)$ $\displaystyle=\int_{x\in\mathbb{R}^{d}}\int_{y\in\mathbb{R}^{d}}d\mu^{*}(x)\int_{t\in[0,\infty)}e^{-t\|x-y\|^{2}}d\eta(t)d\mu(y)$ $\displaystyle=\int_{x\in\mathbb{R}^{d}}\int_{y\in\mathbb{R}^{d}}d\mu^{*}(x)\int_{t\in[0,\infty)}\frac{1}{(2t)^{d/2}}\int_{\omega\in\mathbb{R}^{d}}e^{-\sqrt{-1}(y-x)^{T}\omega-\frac{\|\omega\|^{2}}{4t}}d\omega d\eta(t)d\mu(y)$ $\displaystyle=\int_{\omega\in\mathbb{R}^{d}}\hat{\mu}(\omega)^{*}\int_{t\in[0,\infty)}\frac{1}{(2t)^{d/2}}e^{\frac{-\|\omega\|^{2}}{4t}}d\eta(t)\hat{\mu}(\omega)d\omega,$ (32) where we applied a formula $e^{-t\|x\|^{2}}={(2t)^{-d/2}}\int_{\omega\in\mathbb{R}^{d}}e^{-\sqrt{-1}x^{T}\omega-\|\omega\|^{2}/(4t)}d\omega$ in the third equality. In the same manner as the proof of Theorem 5.11, $\hat{\mu}\neq 0$ holds. In addition, since $\operatorname{supp}(\eta)\neq\\{0\\}$ holds, $\int_{t\in[0,\infty)}(2t)^{-d/2}e^{-\|\omega\|^{2}/(4t)}d\eta(t)$ is positive definite. As a result, the last formula in Eq. (32) is nonzero. By Lemma C.5, $\Phi$ is injective. ### C.2 Proofs of Proposition 5.15 and Theorem 5.16 Let $\mathcal{R}_{+}(\mathcal{X})$ be the set of all real positive-valued regular measures, and $\mathcal{D}_{\nu}(\mathcal{X},\mathcal{A})$ the set of all finite regular Borel $\mathcal{A}$-valued measures $\mu$ whose total variations are dominated by $\nu\in\mathcal{R}_{+}(\mathcal{X})$ (i.e., $|\mu|\leq\nu$). We apply the following representation theorem to derive Theorem 5.16. ###### Proposition C.6 For $\nu\in\mathcal{R}_{+}(\mathcal{X})$, there exists an isomorphism between $\mathcal{D}_{\nu}(\mathcal{X},\mathcal{A})$ and ${L}^{1}_{\nu}(\mathcal{X},\mathcal{A})^{\prime}$. Proof For $\mu\in\mathcal{D}_{\nu}(\mathcal{X},\mathcal{A})$ and $u\in{L}^{1}_{\nu}(\mathcal{X},\mathcal{A})$, we have $\bigg{\|}\int_{x\in\mathcal{X}}d\mu(x)u(x)\bigg{\|}_{\mathcal{A}}\leq\int_{x\in\mathcal{X}}\|u(x)\|_{\mathcal{A}}d|\mu|(x)\leq\int_{x\in\mathcal{X}}\|u(x)\|_{\mathcal{A}}d\nu(x).$ Thus, we define $h:\mathcal{D}_{\nu}(\mathcal{X},\mathcal{A})\to{L}^{1}_{\nu}(\mathcal{X},\mathcal{A})^{\prime}$ as $\mu\mapsto(u\mapsto\int_{x\in\mathcal{X}}d\mu(x)u(x))$. Meanwhile, for $f\in{L}^{1}_{\nu}(\mathcal{X},\mathcal{A})^{\prime}$ and $E\in\mathcal{B}$, we have $\|f(\chi_{E}1_{\mathcal{A}})\|_{\mathcal{A}}\leq C\int_{x\in\mathcal{X}}\|\chi_{E}1_{\mathcal{A}}\|_{\mathcal{A}}d\nu(x)=C\nu(E)$ for some $C>0$ since $f$ is bounded. Here, $\chi_{E}$ is an indicator function for a Borel set $E$. Thus, we define $h^{\prime}:{L}^{1}_{\nu}(\mathcal{X},\mathcal{A})^{\prime}\to\mathcal{D}_{\nu}(\mathcal{X},\mathcal{A})$ as $f\mapsto(E\mapsto f(\chi_{E}1_{\mathcal{A}}))$. By the definitions of $h$ and $h^{\prime}$, $h(h^{\prime}(f))(s)=f(s)$ holds for $s\in\mathcal{S}(\mathcal{X},\mathcal{A})$. Since $\mathcal{S}(\mathcal{X},\mathcal{A})$ is dense in ${L}^{1}_{\nu}(\mathcal{X},\mathcal{A})$, $h(h^{\prime}(f))(u)=f(u)$ holds for $u\in{L}^{1}_{\nu}(\mathcal{X},\mathcal{A})$. Moreover, $h^{\prime}(h(\mu))(E)=\mu(E)$ holds for $E\in\mathcal{B}$. Therefore, $\mathcal{D}_{\nu}(\mathcal{X},\mathcal{A})$ and ${L}^{1}_{\nu}(\mathcal{X},\mathcal{A})^{\prime}$ are isomorphic. Proof of Theorem 5.16 Assume $\mathcal{M}_{k}$ is dense in ${C}_{0}(\mathcal{X},\mathcal{A})$. Since ${C}_{0}(\mathcal{X},\mathcal{A})$ is dense in ${L}^{1}_{\nu}(\mathcal{X},\mathcal{A})$ for any $\nu\in\mathcal{R}_{+}(\mathcal{X})$, $\mathcal{M}_{k}$ is dense in ${L}^{1}_{\nu}(\mathcal{X},\mathcal{A})$ for any $\nu\in\mathcal{R}_{+}(\mathcal{X})$. By Proposition C.3, $\mathcal{M}_{k}^{\perp}=\\{0\\}$ holds. Let $\mu\in\mathcal{D}(\mathcal{X},\mathcal{A})$. There exists $\nu\in\mathcal{R}_{+}(\mathcal{X})$ such that $\mu\in\mathcal{D}_{\nu}(\mathcal{X},\mathcal{A})$. By Proposition C.6, if $\int_{x\in\mathcal{X}}d\mu(x)u(x)=0$ for any $u\in\mathcal{M}_{k}$, $\mu=0$. Since $\int_{x\in\mathcal{X}}d\mu(x)u(x)=\left\langle u,\Phi(\mu)\right\rangle_{\mathcal{M}_{k}}$, $\int_{x\in\mathcal{X}}d\mu(x)u(x)=0$ means $\Phi(\mu)=0$. Therefore, by Lemma C.5, $\Phi$ is injective. For the case of $\mathcal{A}=\mathbb{C}^{m\times m}$, we apply the following extension theorem to derive the converse of Theorem 5.16. ###### Proposition C.7 (c.f. Theorem in Helemskii (1994)) Let $\mathcal{A}=\mathbb{C}^{m\times m}$. Let $\mathcal{M}$ be a Banach $\mathcal{A}$-module, $\mathcal{M}_{0}$ be a closed submodule of $\mathcal{M}$, and $f_{0}:\mathcal{M}_{0}\to\mathcal{A}$ be a bounded $\mathcal{A}$-linear map. Then, there exists a bounded $\mathcal{A}$-linear map $f:\mathcal{M}\to\mathcal{A}$ that extends $f_{0}$ (i.e., $f(u)=f_{0}(u)$ for $u\in\mathcal{M}_{0}$). Proof Von Neumann-algebra $\mathcal{A}$ itself is regarded as an $\mathcal{A}$-module and is normal. Also, $\mathbb{C}^{m\times m}$ is Connes injective. By Theorem in Helemskii (1994), $\mathcal{A}$ is an injective object in the category of Banach $\mathcal{A}$-module. The statement is derived by the definition of injective objects in category theory. We derive the following lemma and proposition by Proposition C.7. ###### Lemma C.8 Let $\mathcal{A}=\mathbb{C}^{m\times m}$. Let $\mathcal{M}$ be a Banach $\mathcal{A}$-module and $\mathcal{M}_{0}$ be a closed submodule of $\mathcal{M}$. For $u_{1}\in\mathcal{M}\setminus\mathcal{M}_{0}$, there exists a bounded $\mathcal{A}$-linear map $f:\mathcal{M}\to\mathcal{A}$ such that $f(u_{0})=0$ for $u_{0}\in\mathcal{M}_{0}$ and $f(u_{1})\neq 0$. Proof Let $q:\mathcal{M}\to\mathcal{M}/\mathcal{M}_{0}$ be the quotient map to $\mathcal{M}/\mathcal{M}_{0}$, and $\,\mathcal{U}_{1}:=\\{q(u_{1})c\mid\ c\in\mathcal{A}\\}$. Note that $\mathcal{M}/\mathcal{M}_{0}$ is a Banach $\mathcal{A}$-module and $\,\mathcal{U}_{1}$ is its closed submodule. Let $\mathcal{V}:=\\{c\in\mathcal{A}\mid\ q(u_{1})c=0\\}$, which is a closed subspace of $\mathcal{A}$. Since $\mathcal{V}$ is orthogonally complemented (Manuilov and Troitsky, 2000, Proposition 2.5.4), $\mathcal{A}$ is decomposed into $\mathcal{A}=\mathcal{V}+\mathcal{V}^{\perp}$. Let $p:\mathcal{A}\to\mathcal{V}^{\perp}$ be the projection onto $\mathcal{V}^{\perp}$ and $f_{0}:\mathcal{U}_{1}\to\mathcal{A}$ defined as $q(u_{1})c\mapsto p(c)$. Since $p$ is $\mathcal{A}$-linear, $f_{0}$ is also $\mathcal{A}$-linear. Also, for $c\in\mathcal{A}$, we have $\displaystyle\|q(u_{1})c\|_{\mathcal{M}/\mathcal{M}_{0}}=\|q(u_{1})(c_{1}+c_{2})\|_{\mathcal{M}/\mathcal{M}_{0}}=\|q(u_{1})c_{1}\|_{\mathcal{M}/\mathcal{M}_{0}}$ $\displaystyle\qquad\geq\inf_{d\in\mathcal{V}^{\perp},\|d\|_{\mathcal{A}}=1}\|q(u_{1})d\|_{\mathcal{M}/\mathcal{M}_{0}}\;\|c_{1}\|_{\mathcal{A}}=\inf_{d\in\mathcal{V}^{\perp},\|d\|_{\mathcal{A}}=1}\|q(u_{1})d\|_{\mathcal{M}/\mathcal{M}_{0}}\;\|p(c)\|_{\mathcal{A}},$ where $c_{1}=p(c)$ and $c_{2}=c_{1}-p(c)$. Since $\inf_{d\in\mathcal{V}^{\perp},\|d\|_{\mathcal{A}}=1}\|q(u_{1})d\|_{\mathcal{M}/\mathcal{M}_{0}}\;\|p(c)\|_{\mathcal{A}}>0$, $f_{0}$ is bounded. By Proposition C.7, $f_{0}$ is extended to a bounded $\mathcal{A}$-linear map $f_{1}:\mathcal{M}/\mathcal{M}_{0}\to\mathcal{A}$. Setting $f:=f_{1}\circ q$ completes the proof of the lemma. Then we prove the converse of Lemma C.3. ###### Proposition C.9 Let $\mathcal{A}=\mathbb{C}^{m\times m}$. For a Banach $\mathcal{A}$-module $\mathcal{M}$ and its submodule $\mathcal{M}_{0}$, $\mathcal{M}_{0}$ is dense in $\mathcal{M}$ if $\mathcal{M}_{0}^{\perp}=\\{0\\}$. Proof Assume $u\notin\overline{\mathcal{M}_{0}}$. We show $\overline{\mathcal{M}_{0}}\supseteq(\mathcal{M}_{0}^{\perp})^{\perp}$. By Lemma C.8, there exists $f\in\mathcal{M}^{\prime}$ such that $f(u)\neq 0$ and $f(u_{0})=0$ for any $u_{0}\in\overline{\mathcal{M}_{0}}$. Thus, $u\notin(\mathcal{M}_{0}^{\perp})^{\perp}$. As a result, $\overline{\mathcal{M}_{0}}\supseteq(\mathcal{M}_{0}^{\perp})^{\perp}$. Therefore, if $\mathcal{M}_{0}^{\perp}=\\{0\\}$, then $\overline{\mathcal{M}_{0}}\supseteq\mathcal{M}$, which implies $\mathcal{M}_{0}$ is dense in $\mathcal{M}$. As a result, we derive Proposition 5.15 as follows. Proof of Proposition 5.15 Let $\mu\in\mathcal{D}(\mathcal{X},\mathcal{A})$. Then, “$\Phi(\mu)=0$” is equivalent to “$\int_{x\in\mathcal{X}}d\mu^{*}(x)u(x)=\left\langle\Phi(\mu),u\right\rangle_{\mathcal{M}_{k}}=0$ for any $u\in\mathcal{M}_{k}$”. Thus, by Proposition C.4, “$\Phi(\mu)=0\Rightarrow\mu=0$” is equivalent to “$f\in{C}_{0}(\mathcal{X},\mathcal{A})^{\prime}$, $f(u)=0$ for any $u\in\mathcal{M}_{k}$ $\Rightarrow$ $f=0$”. By the definition of $\mathcal{M}_{k}^{\perp}$ and Proposition C.9, $\mathcal{M}_{k}$ is dense in ${C}_{0}(\mathcal{X},\mathcal{A})$. ## Appendix D Derivative on Banach spaces ###### Definition D.1 (Fréchet derivative) Let $\mathcal{M}$ be a Banach space. Let $f:\mathcal{M}\to\mathcal{A}$ be an $\mathcal{A}$-valued function defined on $\mathcal{M}$. The function $f$ is referred to as (Fréchet) differentiable at a point $\mathbf{c}\in\mathcal{M}$ if there exists a continuous $\mathbb{R}$-linear operator $l$ such that $\lim_{u\to 0,\ u\in\mathcal{M}\setminus\\{0\\}}\frac{\|f(\mathbf{c}+u)-f(\mathbf{c})-l(u)\|_{\mathcal{A}}}{\|u\|_{\mathcal{M}}}=0$ for any $u\in\mathcal{M}$. In this case, we denote $l$ as $Df_{\mathbf{c}}$. ## References * Álvarez et al. (2012) M. Álvarez, L. Rosasco, and N. Lawrence. Kernels for vector-valued functions: A review. _Foundations and Trends in Machine Learning_ , 4, 2012. * Bakić and Guljaš (2001) D. Bakić and B. Guljaš. Operators on Hilbert $H^{*}$-modules. _Journal of Operator Theory_ , 46:123–137, 2001. * Balkir (2014) E. Balkir. _Using Density Matrices in a Compositional Distributional Model of Meaning_. Master’s thesis, University of Oxford, 2014. * Blanchard and Brüning (2015) P. Blanchard and E. Brüning. _Mathematical Methods in Physics_. Birkhäuser, 2nd edition, 2015. * Budišić et al. (2012) M. Budišić, R. Mohr, and I. Mezić. Applied Koopmanism. _Chaos_ , 22:047510, 2012. * Cnops (1992) J. Cnops. A Gram–Schmidt method in Hilbert modules. _Clifford Algebras and their Applications in Mathematical Physics_ , 47:193–203, 1992. * Črnjarić-Žic et al. (2020) N. Črnjarić-Žic, S. Maćešić, and I. Mezić. Koopman operator spectrum for random dynamical systems. _Journal of Nonlinear Science_ , 30:2007–2056, 2020. * Deb (2016) P. Deb. Geometry of quantum state space and quantum correlations. _Quantum Information Processing_ , 15:1629–1638, 2016. * Diestel (1984) J. Diestel. _Sequences and Series in Banach spaces_. Graduate texts in mathematics ; Volume 92. Springer-Verlag, 1984. * Dinculeanu (1967) N. Dinculeanu. _Vector Measures_. International Series of Monographs on Pure and Applied Mathematics ; Volume 95. Pergamon Press, 1967. * Dinculeanu (2000) N. Dinculeanu. _Vector Integration and Stochastic Integration in Banach Spaces_. John Wiley & Sons, 2000. * Dudley (2002) R. M. Dudley. _Real Analysis and Probability_. Cambridge Studies in Advanced Mathematics. Cambridge University Press, 2nd edition, 2002. * Fujii & Kawahara (2019) Fujii, K. and Kawahara, Y. Dynamic mode decomposition in vector-valued reproducing kernel Hilbert spaces for extracting dynamical structure among observables. _Neural Networks_ , 117:94–103, 2019. * Fukumizu et al. (2007) K. Fukumizu, A. Gretton, X. Sun, and B. Schölkopf. Kernel measures of conditional dependence. _Advances in Neural Information Processing Systems 20_ , 489–496, 2007\. * Gretton et al. (2006) A. Gretton, K. Borgwardt, M. Rasch, B. Schölkopf, and A. J. Smola. A kernel method for the two-sample-problem. _Advances in Neural Information Processing Systems 19_ , 513–520, 2006\. * Gretton et al. (2012) A. Gretton, K. M. Borgwardt, M. J. Rasch, B. Schölkopf, and A. J. Smola. A kernel two-sample test. _Journal of Machine Learning Research_ , 13(25):723–773, 2012. * Hashimoto et al. (2020) Y. Hashimoto, I. Ishikawa, M. Ikeda, Y. Matsuo, and Y. Kawahara. Krylov subspace method for nonlinear dynamical systems with random noise. _Journal of Machine Learning Research_ , 21(172):1–29, 2020. * Hastie et al. (2009) T. Hastie, R. Tibshirani, and J. Friedman. _The Elements of Statistical Learning: Data Mining, Inference, and Prediction_. Springer, 2nd edition, 2009. * Helemskii (1994) A. Helemskii. The spatial flatness and injectiveness of connes operator algebras. _Extracta mathematicae_ , 9:75–81, 1994. * Heo (2008) J. Heo. Reproducing kernel Hilbert $C^{*}$-modules and kernels associated with cocycles. _Journal of Mathematical Physics_ , 49:103507, 2008. * Holevo (2011) A. S. Holevo. _Probabilistic and Statistical Aspects of Quantum Theory_. Scuola Normale Superiore, 2011\. * Ishikawa et al. (2018) I. Ishikawa, K. Fujii, M. Ikeda, Y. Hashimoto, and Y. Kawahara. Metric on nonlinear dynamical systems with Perron–Frobenius operators. In _Advances in Neural Information Processing Systems 31_ , 2856–2866, 2018. * Itoh (1990) S. Itoh. Reproducing kernels in modules over $C^{*}$-algebras and their applications. _Journal of Mathematics in Nature Science_ , 37:1–20, 1990. * Jitkrittum et al. (2019) W. Jitkrittum, P. Sangkloy, M. W.Gondal, A. Raj, J. Hays, and B. Schölkopf. Kernel mean matching for content addressability of GANs. In _Proceedings of the 36th International Conference on Machine Learning_ , 3140–3151, 2019. * Kadri et al. (2016) H. Kadri, E. Duflos, P. Preux, S. Canu, A. Rakotomamonjy, and J. Audiffren. Operator-valued kernels for learning from functional response data. _Journal of Machine Learning Research_ , 17(20):1–54, 2016. * Kawahara (2016) Y. Kawahara. Dynamic mode decomposition with reproducing kernels for Koopman spectral analysis. In _Advances in Neural Information Processing Systems 29_ , 911–919, 2016. * Klus et al. (2020) S. Klus, I. Schuster, and K. Muandet. Eigendecompositions of transfer operators in reproducing kernel Hilbert spaces. _Journal of Nonlinear Science_ , 30:283–315, 2020. * Lance (1995) E. C. Lance. _Hilbert $C^{*}$-modules – a toolkit for operator algebraists_. London Mathematical Society Lecture Note Series; Volume 210. Cambridge University Press, 1995. * Levitin et al. (2007) D. J. Levitin, R. L. Nuzzo, B. W. Vines, and J. O. Ramsay. Introduction to functional data analysis. _Canadian Psychology_ , 48:135–155, 2007. * Li et al. (2019) H. Li, S. J. Pan, S. Wang, and A. C. Kot. Heterogeneous domain adaptation via nonlinear matrix factorization. _IEEE Transactions on Neural Networks and Learning Systems_ , 31:984–996, 2019. * Lim et al. (2015) N. Lim, F. Buc, C. Auliac, and G. Michailidis. Operator-valued kernel-based vector autoregressive models for network inference. _Machine Learning_ , 99(3):489–513, 2015. * Liu and Rebentrost (2018) N. Liu and P. Rebentrost. Quantum machine learning for quantum anomaly detection. _Physical Review A_ , 97:042315, 2018. * Lusch et al. (2018) B. Lusch, J. N. Kutz, and S. L. Brunton. Deep learning for universal linear embeddings of nonlinear dynamics. _Nature Communications_ , 9:4950, 2018. * Manuilov and Troitsky (2000) V. M. Manuilov and E. V. Troitsky. Hilbert $C^{*}$\- and $W^{*}$-modules and their morphisms. _Journal of Mathematical Sciences_ , 98:137–201, 2000. * Micchelli and Pontil (2005) C. A. Micchelli and M. Pontil. On learning vector-valued functions. _Neural Computation_ , 17:177–204, 2005. * Minh et al. (2016) H. Q. Minh, L. Bazzani, and V. Murino. A unifying framework in vector-valued reproducing kernel Hilbert spaces for manifold regularization and co-regularized multi-view learning. _Journal of Machine Learning Research_ , 17(25):1–72, 2016. * Muandet et al. (2017) K. Muandet, K. Fukumizu, B. K. Sriperumbudur, and B. Schölkopf. Kernel mean embedding of distributions: A review and beyond. _Foundations and Trends in Machine Learning_ , 10(1–2), 2017. * Müller (1997) A. Müller. Integral probability metrics and their generating classes of functions. _Advances in Applied Probability_ , 29:429–443, 1997. * Murphy (1990) G. J. Murphy. _C*-Algebras and Hilbert Space Operators_. Academic Press, 1990. * Peres and Terno (2004) A. Peres and D. R. Terno. Quantum information and relativity theory. _Reviews of Modern Physics_ , 76:93–123, 2004. * Rachev (1985) S. T. Rachev. On a class of minimal functionals on a space of probability measures. _Theory of Probability & Its Applications_, 29:41–49, 1985. * Ramsay and Silverman (2005) J. O. Ramsay and B. W. Silverman. _Functional Data Analysis_. Springer, 2nd edition, 2005. * Saitoh and Sawano (2016) S. Saitoh and Y. Sawano. _Theory of Reproducing Kernels and Applications_. Springer, 2016. * Schölkopf and Smola (2001) B. Schölkopf and A. J. Smola. _Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond_. MIT Press, 2001. * Skeide (2000) M. Skeide. Generalised matrix $C^{\ast}$-algebras and representations of Hilbert modules. _Mathematical Proceedings of the Royal Irish Academy_ , 100A(1):11–38, 2000. * Smola et al. (2007) A. J. Smola, A. Gretton, L. Song, and B. Schölkopf. A Hilbert space embedding for distributions. In _Proceedings of the 18th International Conference on Algorithmic Learning Theory_ , 13–31, 2007. * Schölkopf et al. (2001) B. Schölkopf, R. Herbrich, and A. J. Smola. A generalized representer theorem. In _Proceedings of the 14th Annual Conference on Computational Learning Theory_ , 416–426, 2001. * Smyrlis and Zisis (2004) G. Smyrlis and V. Zisis. Local convergence of the steepest descent method in Hilbert spaces. _Journal of Mathematical Analysis and Applications_ , 300:436–453, 2004. * Sriperumbudur et al. (2010) B. K. Sriperumbudur, A. Gretton, K. Fukumizu, B. Schölkopf, and G. R. G. Lanckriet. Hilbert space embeddings and metrics on probability measures. _Journal of Machine Learning Research_ , 11:1517–1561, 2010\. * Sriperumbudur et al. (2011) B. K. Sriperumbudur, K. Fukumizu, and G. R. G. Lanckriet. Universality, characteristic kernels and RKHS embedding of measures. _Journal of Machine Learning Research_ , 12(70):2389–2410, 2011. * Sriperumbudur et al. (2012) B. K. Sriperumbudur, K. Fukumizu, A. Gretton, B. Schölkopf, and G. R. G. Lanckriet. On the empirical estimation of integral probability metrics. _Electronic Journal of Statistics_ , 6:1550–1599, 2012\. * Steinwart (2001) I. Steinwart. On the influence of the kernel on the consistency of support vector machines. _Journal of Machine Learning Research_ , 2:67–93, 2001\. * Suzumura et al. (2017) S. Suzumura, K. Nakagawa, Y. Umezu, K. Tsuda, and I. Takeuchi. Selective inference for sparse high-order interaction models. In _Proceedings of the 34th International Conference on Machine Learning_ , 3338–3347, 2017. * Szafraniec (2010) F. H. Szafraniec. Murphy’s positive definite kernels and Hilbert $C^{*}$-modules reorganized. _Noncommutative Harmonic Analysis with applications to probability II_ , 89:275–295, 2010. * Takeishi et al. (2017a) N. Takeishi, Y. Kawahara, and T. Yairi. Subspace dynamic mode decomposition for stochastic Koopman analysis. _Physical Review E_ , 96:033310, 2017a. * Takeishi et al. (2017b) N. Takeishi, Y. Kawahara, and T. Yairi. Learning Koopman invariant subspaces for dynamic mode decomposition. In _Advances in Neural Information Processing Systems 30_ , 1130–1140, 2017b. * Wang et al. (2016) J. L. Wang, J. M. Chiou, and H. G. Müller. Functional data analysis. _Annual Review of Statistics and Its Application_ , 3:257–295, 2016. * Wendland (2004) H. Wendland. _Scattered Data Approximation_. Cambridge Monographs on Applied and Computational Mathematics. Cambridge University Press, 2004. * Ye (2017) Y. Ye. The matrix Hilbert space and its application to matrix learning. _arXiv:1706.08110v2_ , 2017.
# Rydberg quantum simulator of topological insulators Mohammadsadegh Khazali Institute for Quantum Optics and Quantum Information of the Austrian Academy of Sciences, Innsbruck, Austria ###### Abstract This article proposes the implementation of Rydberg multi-dimensional discrete-time quantum walk (DTQW) that could ideally simulate different classes of topological insulators. Using an exchange interaction between Rydberg excited atoms in an atomic-array with dual lattice-constants, the new setup operates both coined and coin-less models of DTQW. Here, complicated coupling tesselations are performed by global laser excitation. The long-range interaction provides a new feature of designing different topologically ordered periodic boundary conditions. Limiting the Rydberg population to two excitations, coherent QW over thousand lattice sites and steps are achievable with the current technology. These features would improve the performance of this quantum machine in running the quantum search algorithm over topologically ordered databases as well as diversifying the range of topological insulators that could be simulated. ## I Introduction There is a huge effort in making quantum hardwares that outperform classical counterparts in performing certain algorithms and simulating other complicated quantum systems. Among different approaches, implementing the quantum walk (QW) Aha93 ; Far98 ; Kem03 is receiving wide interest. Unlike classical random walk, particles performing a quantum walk can take superposition of all possible paths through their environment simultaneously, leading to faster propagation and enhanced sensitivity to initial conditions Dad18 ; Sum16 ; Pre15 . These properties provide an appealing bases for implementation of quantum algorithms like searching Por13 ; She03 ; Chi03 ; Chi04 ; Por17 , quantum processing Chi09 ; ken20 ; Lov10 ; Chi13 ; Sal12 and simulation of topological insulators Kit10 . Improving the hardware in terms of size, coherence, dimensions, controlling elements and other features like the topologically ordered boundary-conditions, would improve the performance and diverse the applications that could run over this quantum hardware. Quantum walk has been implemented on trapped ions Sch09 ; Zah10 , neutral atoms Kar09 ; Wei11 ; Fuk13 ; Pre15 and among other platforms Man14 . While the ion-traps are limited to a 1D array of 50 atoms, neutral atoms are pioneer in terms of multi-dimensional trapping of a large number of identical qubits. From this perspective, trapping and controlling more qubits than any other platforms has already been demonstrated in one dimension (1D) Ber17 ; Omr19 , 2D Zha11 ; Pio13 ; Nog14 ; Xia15 ; Zei16 ; Lie18 ; Nor18 ; Coo18 ; Hol19 ; Sas19 and 3D Wan16 ; Bar18 geometries. The current advances in exciting atoms to the Rydberg level Lev18 , provides a controllable long-range interaction for quantum technology Saf10 ; Ada19 ; Kha15 ; kha20 ; khaz2020rydberg ; khaz16 ; Kha21 ; Kha19 ; Kha16 ; Kha17 ; Kha18 . Rydberg interaction has been used in different time-independent lattice Hamiltonian models leading to continuous-time quantum transport Cot06 ; Bar15 ; Scho15 ; Ori18 ; Gun13 ; Sch15 ; Let18 ; Wus11 and simulating topological Mott insulators Dau16 . The continuous-time interaction in these simulators is limiting the range of programable problems. Also, exciting all the sites to the Rydberg level would cause significant de-coherence in addition to the strong interaction between atoms. To overcome this interaction, very demanding exciting and trapping lasers are required. It also puts a limit on the density of sites in the lattice. Figure 1: Rydberg discrete-time quantum walk (DTQW) scheme. (a) Level scheme: The walker is an $nP$ Rydberg excitation. QW operates by exciting a neighbouring lattice to $nS$ Rydberg state featuring resonant exchange- interaction with the walker. (b) The exchange interaction $V$ forms a site- dependent level shift of $nS$ Rydberg state. Using two lattice constants and tuning the exciting laser’s frequency only a desired site get in resonance with the laser and apply quantum walk. (c[d]) Adjusting laser’s detuning to the inter- [intra-] dimer interaction, desired coupling tessellation $W_{0}$ [$W_{1}$] would be formed. (e) The maximum population of the auxiliary state $|70S\rangle$ over a 2$\pi$ pulse is plotted as a function of interatomic distance from the walker $|70P\rangle$. Laser detuning is set to $\Delta=-\frac{C_{3}}{a^{3}}$. Only the site at the distance $a=3\mu$m from the walker would get in resonance and hence goes under the quantum walk ($\Omega/2\pi=2$MHz, $C_{3}=8.4$GHz.$\mu$m3). (f) The hopping angle $\theta$ could be controlled by manipulating the effective detuning of targeted site $\delta=\Delta+V(a_{\\{0,1\\}})$. This paper proposes an approach that revolutionizes the level of control over the interaction leading to Rydberg discrete-time quantum walk in multi- dimensions. Benefitting from the long-range Rydberg interaction, the scheme features QW implementation on topologically ordered (e.g. torus) periodic boundary conditions. Limited Rydberg population in this proposal would be a big step towards scalable quantum simulators. While using global lasers to switch among multiple coupling tessellations, adding local external fields provide an extra degree of control for engineering space-dependent coupling properties. Valuable features that are not realized in other neutral atom QW schemes, and which open a wide range of applications. As an example, the implementation of different classes of Floquet topological insulators with Rydberg model will be overviewed. Topological insulators are a new class of quantum materials that are insulating in the bulk but exhibiting robust topologically protected current- carrying edge states. Topological insulator materials are challenging to synthesize, and limited in topological phases accessible with solid-state materials And13 . This has motivated the search for topological phases on the systems that simulate the same principles underlying topological insulators. Discrete-time quantum walks (DTQWs) have been proposed for making Floquet topological insulators. This periodically driven system simulates an effective (Floquet) Hamiltonian that is topologically nontrivial Cay13 . This system replicates the effective Hamiltonians from all universality classes of 1- to 3-D topological insulators Kit09 ; Kit10 ; Pan20 . Interestingly, topological properties of Floquet topological insulators could be controlled via an external periodic drive rather than an external magnetic field. DTQW generates topological phases that are reacher than those of time- independent lattice Hamiltonians Dau16 . Topological edge states have been realized exclusively in photonic DTQW with limited sites ($<20$) and steps ($<10$) Rec13 ; Xia17 ; Muk17 . Introduced Rydberg DTQW scheme could realize edge state over a thousand sites and steps in 1, 2, and 3 dimensions with the available experimental setups. The mentioned advances of the designed hardware improve the performance of quantum algorithms, e.g. torus topological periodic boundary conditions could result in a quadratic speedup of quantum-walk-based search algorithms Por13 ; Por17 , and diversify the range of topological phenomena that could be simulated. This work opens up the chance of realizing topological properties in multi-dimensional QWs as well as QWs over topologically ordered surfaces. The article is organized as follows. In Sec. II, the coined and coin-less Rydberg discrete-time quantum walk schemes are presented in 1D. Sec. III extends the model to higher dimensions and discusses the approaches for imposing periodic boundary conditions or applying quantum walk on different topological surfaces. The Coherence of the proposed scheme under a wide range of imperfection sources would be evaluated in Sec. IV. The scheme’s performance in multi-dimensions are then quantified in Sec. V. At the end, applications of this model in simulating multi-dimensional topological insulators are discussed in Sec. VI. ## II Rydberg discrete-time quantum walk In the coin-less DTQW different coupling tessellations must be considered, in a way that each tessellation covers all vertices with non-overlapping connections and the tessellation union covers all edges. This model is the discrete-time version of the famous Su-Schrieffer-Heeger (SSH) model su1979solitons . Distinguishing even ${|{m,e}\rangle}={|{2m}\rangle}$ and odd ${|{m,o}\rangle}={|{2m-1}\rangle}$ lattice sites in the $m^{\text{th}}$ sub- lattice (dimer), two types of QW operators, covers the intra-dimer $W_{0}=\exp(\text{i}\theta_{0}H_{0})$ and inter-dimer $W_{1}=\exp(\text{i}\theta_{1}H_{1})$ coupling tessellations with $\displaystyle H_{0}=\sum\limits_{m=1}^{N/2}({|{m,e}\rangle}\\!{\langle{m,o}|}+\text{h.c.})$ (1) $\displaystyle H_{1}=\sum\limits_{m=1}^{N/2}({|{m,e}\rangle}\\!{\langle{m+1,o}|}+\text{h.c.}),$ see Fig. 1c,d. The physical implementation of the proposed Rydberg discrete-time quantum walk (DTQW) is presented in Fig. 1. The walker is a ${|{p}\rangle}={|{nP_{3/2},3/2}\rangle}$ Rydberg excitation while other sites are in the ground state ${|{g}\rangle}={|{5S_{1/2},1/2}\rangle}$. The desired inter- and intra-dimer coupling labeled by $k\in\\{0,1\\}$, could be realized by site selective excitation of ground state atom to ${|{s}\rangle}={|{nS_{1/2},1/2}\rangle}$ Rydberg level featuring exchange interaction with the walker ${|{p}\rangle}$. The site selection is controlled by adjusting of the global laser’s detuning from ${|{s}\rangle}$ state $\Delta_{k}$ under the concept of Rydberg aggregate, see bellow. The effective Rydberg quantum-walk is governed under the following Hamiltonian $\displaystyle H_{k}^{Ry}=$ $\displaystyle\sum\limits_{i<j}V(r_{ij})({|{s_{i}p_{j}}\rangle}\\!{\langle{p_{i}s_{j}}|}+\text{h.c.})$ $\displaystyle+\sum\limits_{i}\Omega({|{s}\rangle}_{i}{\langle{g}|}+\text{h.c.})+\Delta_{k}{|{s}\rangle}_{i}{\langle{s}|},$ where $i$ sums over all the sites. The exchange interaction $V(r_{ij})=C_{3}/(r_{ij})^{3}$ between the walker ${|{p}\rangle}$ and auxiliary excited Rydberg state ${|{s}\rangle}$, separated by $r_{ij}$, evolves the initial state of the two sites ${|{p_{i}g_{j}}\rangle}$ to a superposition state $\cos\theta{|{p_{i}g_{j}}\rangle}+\text{i}\sin\theta{|{g_{i}p_{j}}\rangle}$ over the 2$\pi$ pulse of $\Omega$ laser. The absence of self-interaction over delocalized single ${|{s}\rangle}$ and ${|{p}\rangle}$ excitations are justified in the App. A1. To operate the two tessellation types of Eq. 1 under Rydberg Hamiltonian Eq. II with a global laser, the space-dependent nature of interaction $V(r_{ij})$ is used over a superlattice with distinct lattice-constants inside ($a_{0}$) and outside ($a_{1}$) the dimers. By adjusting the exciting laser’s detuning from the ${|{s}\rangle}$ Rydberg state, to be opposite of the interaction of specific lattice site at distance $a_{k}$ ($k\in\\{0,1\\}$) from the walker ($\Delta_{k}=-\frac{C_{3}}{a_{k}^{3}}$), one can choose the site pairs that get in resonance with the laser conditioned on the presence of the walker and thus undergo the quantum walk, see Fig. 1. The single non-local quantum walker ${|{p}\rangle}$ would induce the excitation of a single nonlocal auxiliary ${|{s}\rangle}$ Rydberg state over each $2\pi$ pulse operation, see App. A1. Adjusting the laser detuning at each pulse would connect the targeted site at $a_{0}$ or $a_{1}$ distance from the walker, generating the desired $W_{k}=\exp(\text{i}H^{Ry}_{k}t_{k})$ with $k\in\\{0,1\\}$ corresponding to intra- and inter-dimer coupling tessellations. Duration of each step $t_{k}$ is defined by $\int\limits_{0}^{t_{k}}\Omega_{\text{eff}}\text{d}t=2\pi$, where the effective Rabi frequency is given by $\Omega_{\text{eff}}=\sqrt{\Omega^{2}+\delta^{2}}$ and $\delta=\Delta_{k}+V(a_{k})$ is the effective detuning of the targeted site at either $a_{0}$ or $a_{1}$. Fig. 1f shows how the hopping angle $\theta$ in each step can get controlled by $\frac{\delta}{\Omega}$ ratio. To implement coined DTQW, the dimers would be considered as the individual units. The coin is formed by the relative population of odd and even sites in each sub-lattice (dimer), see Fig. 1c,d. The coin rotation operator $R_{\theta}=\exp(\text{i}H_{0}\theta)$ is applied by population rotation in the sub-lattice basis using $H_{0}$ operator of Eq. 1. The desired transition operators of coined DTQW would be realized by subsequent application of intra- and inter-dimer population swapping i.e. $T=\text{e}^{\text{i}H_{1}\pi/2}R(\pi/2)=\sum\limits_{m}({|{m-1,e}\rangle}\\!{\langle{m,e}|}+{|{m+1,o}\rangle}\\!{\langle{m,o}|})$. ## III Multi-dimensional DTQW with periodic boundary conditions Figure 2: Multi-dimensional DTQW. (a) Kronecker multiplication of 1D QW leads to a 2D lattice of tetramers and could trivially be extended to a 3D lattice of octamers. (b) Extension to the 3D lattice of dimers provides a non- separable multi-dimensional Rydberg DTQW. In (b) quantization axis would alter during the operation to be along the connections. The idea behind Fig. 1, is extendable to higher-dimensions by two approaches. Kronecker multiplication of 1D staggered quantum walk, would make 2D and 3D lattices of tetramers and octamers respectively. A more enriched non-separable DTQW could be applied in a multi-dimensional lattice of dimers. The angular- dependency of the exchange interaction $V_{ij}$ provides a wider range of laser detuning, available for dynamic control over the exciting sites. ##### Multi-dimensional DTQW via Kronecker multiplication: Extension to higher dimensions could be realized as the combination of coin- less DTQWs along different directions. In two dimensions, this would result in a 2D lattice of tetramers as depicted in Fig. 2a. The QW is performed by concatenated application of the four sets of quantum jump operators $W_{xl}=\exp(\text{i}\theta_{xl}H_{xl}\otimes\mathbbm{1}_{y})$ and $W_{yl}=\exp(\text{i}\theta_{yl}\mathbbm{1}_{x}\otimes H_{yl})$ where $H_{l}$ ($l\in\\{0,1\\}$) in each dimension is given by Eq. 1, with distinguished odd and even sites along $x$ and $y$ dimension, see Eq1 for the expanded set of Hamiltonians. For the implementation, two lattice constants along each dimension is required to distinguish inter- and intra-cell couplings. Extension to the 3D lattice of octamers is trivial. ##### Multi-dimensional Rydberg DTQW in a lattice of dimers provides an enriched non-separable Hamiltonian. Fig. 2b shows the connectivity graphs over the 3D lattice with the coupling Hamiltonians presented in Eq2 . To realize this set of couplings, the Rydberg quantization axis must be changed to be along the exchange orientation. The quantization axis is defined by the orientation of polarized lasers. The lattice structure in this model consists of unique lattice constant along $y$ and $z$ dimension while containing two inter- and intra-cell lattice constants along $x$ dimension. These coupling Hamiltonians could be used for coinless DTQW operators $W_{l}=\text{e}^{\text{i}H_{l}\theta_{l}}$ with $l=\\{x,xy,xz\\}_{[0,1]}$ as explained in Sec. II. Fine tuning of presented connectivities and operation fidelities in multi-dimensions are discussed in Sec. V. ##### Multi-dimensional coined DTQW in a lattice of dimers: The proposed system of Fig. 2b, can be used for the implementation of multi- dimensional coined DTQW, where the coin is formed by the relative population of odd and even sites in each sub-lattice. The coin rotation $R_{\theta}=\exp(\text{i}H_{x0}\theta)=\cos(\theta)\mathbbm{1}_{\\{e,o\\}}+\text{i}\sin(\theta)({|{e}\rangle}\\!{\langle{o}|}+{|{o}\rangle}\\!{\langle{e}|})$ is applied by the intra-dimer population rotation. The transition operators are applied by concatenated implementation of intra- and inter-dimer population swapping i.e. $T_{xyz}=\text{e}^{\text{i}H_{xyz1}\pi/2}R(\pi/2)$, $T_{xy}=\text{e}^{\text{i}H_{xy1}\pi/2}R(\pi/2)$, $T_{xz}=\text{e}^{\text{i}H_{xz1}\pi/2}R(\pi/2)$ $T_{x}=\text{e}^{\text{i}H_{x1}\pi/2}R(\pi/2)$, $T_{y}=\text{e}^{\text{i}H_{xy0}\pi/2}R(\pi/2)$, and $T_{z}=\text{e}^{\text{i}H_{xz0}\pi/2}R(\pi/2)$. Extended forms of transition operators are presented in Eq3 . ##### Topological periodic boundary conditions The long-range Rydberg interaction could be used for making different topological periodic boundary conditions. Fig. 3 shows two examples of DTQW over (a) Möbius stripe and (b) torus topological surfaces. While the torus boundary condition could be implemented by global laser over limited lattice sites, forming other topological surfaces e.g. Möbius and Kline bottle requires local laser excitations. During the boundary operation step $W_{yb}$ ($W_{xb}$) with the local lasers, the pair sites with the same pentagon (hexagon) numbers would excited to ${|{s}\rangle}$ under local lasers with detuning adjusted to their interactions. The sequence of the QW operators would be $U=W_{y0}W_{y1}W_{yb}W_{x0}W_{x1}W_{xb}$. Figure 3: Rydberg DTQW over topological surfaces (a) Möbius stripe and (b) torus. To realize the boundary conditions the pair sites with the same pentagon (hexagon) number would get excited to $nS$ state with local lasers over the boundary operation step $W_{yb}$ ($W_{xb}$). Concatenated operation of $W_{y0}W_{y1}W_{yb}W_{x0}W_{x1}W_{xb}$, performs the QW on the desired topological surface. ## IV Decoherence in Rydberg DTQW ### IV.1 Non-unitary dynamics The non-unitary dynamics of the quantum walk can be described by the projection of quantum state onto the pointer states Schl07 . In this model the pointer state projector $\Pi_{x}={|{x}\rangle}\\!{\langle{x}|}$, projects the walker into the $x^{\text{th}}$ site. In the presented Rydberg QW model, the evolution of quantum walker over a single step is mainly coherent. Hence, the decoherence could be applied in a discrete-time manner after each step. The effective stroboscopic non-unitary operator would be $\rho_{i+1}=(1-P_{s})W\rho_{i}W^{\dagger}+P_{s}\sum\limits_{x}\Pi_{x}W\rho_{i}W^{\dagger}\Pi_{x}^{\dagger}$ (3) where $\rho_{i}=\sum_{x,x^{\prime}}{|{x}\rangle}\\!{\langle{x^{\prime}}|}$ is the density operator after i${}^{\text{th}}$ step. The spatial de-phasing terms discussed in the next section would reduce the off-diagonal coherence terms with a probability of $P_{s}$ after each step and hence suppress the superposition effect. In the totally diffusive case $P_{s}=1$, the absence of the interference of quantum paths would only leave a classical Gaussian probability distribution in the diagonal terms, see Fig. 4. (a) (b) (c) (d) (e) (f) (g) (h) (i) Figure 4: Effects of de-phasing on (left column) density matrix terms $\rho(x,x^{\prime})$, (middle column) anti-diagonal coherence $\rho(x,-x)$ and (right column) probability distribution $\rho(x,x)$ after 50 Hadamard steps $\theta=\pi/4$ of coinless Rydberg DTQW over a 1D lattice with N=101 sites. The three rows from top to bottom are corresponding to $P_{s}=[0,0.05,1]$ de- phasing probability over single step. To define the coherence length $l_{co}$ one can look at the suppression of anti-diagonal terms $|\rho_{(x,-x)}|=|\rho_{{}_{0}(x,-x)}|\exp(-|x|/l_{co})$ with respect to the unperturbed one $\rho_{{}_{0}(x,-x)}$. Coherence length is plotted as a function of $P_{s}$ in Fig. 5a. The crossover between coherent and incoherent cases can also be seen in the rate of mean square displacement $\langle x^{2}\rangle=\sum\limits_{x}\rho(x,x)x^{2}$. Fig. 5b, shows that within the coherent period $t<\Gamma^{-1}$ the walker propagates ballistically $\propto t^{2}$ while in the diffusive regime $t>\Gamma^{-1}$ it propagates linearly with time. Overall, Fig. 5, shows that coherent staggered quantum walk over a lattice with $N\approx P_{s}^{-1}$ sites is realizable that would propagate ballistically over $i=P_{s}^{-1}$ steps. In the next section, $P_{s}$ is evaluated in the proposed scheme. (a) (b) Figure 5: Effects of dephasing on (a) coherent-length and (b) Ballistic distribution. (a) The number of lattice-sites over which the coherence preserves is numerically calculated, see the text for more details. Figure (b), numerically calculates $<x^{2}>$ after 50 steps in N=101 lattice dimension as a function of $\gamma t=iP_{s}$ where $i$ is the step number, $\gamma$ is the total dephasing rate and $t$ is the operation time. For time scales below $\gamma^{-1}$ the square displacement is ballistic $\propto t^{2}$ while above that it becomes diffusive $\propto t$. Dotted-dashed line is a quadratic fit to the ballistic part $22.5(\gamma t)^{2}$. The hopping angle in (a) is $\theta=\frac{\pi}{4}$. ### IV.2 Sources of errors #### Laser error sources Laser noise over the Rydberg excitation process, causes de-phasing. The laser noise $\gamma_{la}$ that is encountered here is what remains after laser locking in the two-photon excitation. Fig. 6a, simulates the de-phasing probability $P_{s}$ after one step as a function of the relative laser noise $\frac{\gamma_{la}}{\Omega}$. In this figure, effects of the laser-noise over the Rydberg population rotation are quantified using master equation with $\sqrt{\gamma_{la}}{|{s}\rangle}\\!{\langle{s}|}$ Lindblad term. The plot considers 70S-70P interaction with $V=\frac{-8.4}{3^{3}}$GHz over the lattice constant of 3$\mu$m, and $\Omega=|V|/100$. In the recent experiments Ber17 ; Lev18 , laser noise are suppressed below the effective bandwidth of the lock, resulting in narrow line-widths of 0.5 kHz for the two-photon Rabi frequency of $\Omega/2\pi=$2MHz. Thus, dephasing probability over one Hadamard step ($\theta=\frac{\pi}{4}$) is $P_{s}\approx 10^{-4}$. Spontaneous scattering from the optical lattice lasers as well as Rydberg exciting lasers, destroy the coherence by projecting the quantum walker’s wave-function into a single lattice site. The new advances in clock optical lattices have suppressed the trap lasers’ scattering rate, reaching coherence times above 12s Ros19 making the corresponding dephasing per step negligible $P_{s}\approx 10^{-8}$. Spontaneous emission also occurs from the intermediate state ${|{p}\rangle}$, over the two-photon Rydberg excitation${|{g}\rangle}-{|{s}\rangle}$. The two lasers $\Omega_{1}$, $\Omega_{2}$ are detuned from the intermediate level by $\Delta_{p}$. The dominant decay channel from ${|{p}\rangle}$ is back into the ground state ${|{g}\rangle}$. This would result in an effective Lindblad de-phasing term $\sqrt{\gamma_{p}}{|{g}\rangle}\\!{\langle{g}|}$, where $\gamma_{p}/2\pi=1.4$MHz is decay rate of the intermediate level ${|{p}\rangle}=6P_{3/2}$ in Rb. Over one quantum step operation time of $\frac{2\pi}{\Omega}$ with effective Rabi frequency of $\Omega=\frac{\Omega_{1}\Omega_{2}}{2\Delta_{p}}$, the de-phasing probability after one step would be $P_{s}=\frac{\pi\gamma_{p}}{2\Delta_{p}}(\frac{\Omega_{1}}{\Omega_{2}}+\frac{\Omega_{2}}{\Omega_{1}})$ Saf10 . Using the experiment parameters in exciting ${|{r}\rangle}=70S$ via ${|{p}\rangle}=6P_{3/2}$ intermediate level Lev18 with $(\Omega_{1},\Omega_{2})=2\pi\times(60,40)$MHz for (420nm,1013nm) lasers and the detuning of $\Delta_{p}/2\pi=600$MHz the effective Rabi frequency would be $\Omega/2\pi=2$MHz and de-phasing probability over single quantum step would be $P_{s}=2.5\times 10^{-4}$. #### Lattice geometry and confinement Known detuning would cause a phase that gets absorbed in the definition of ${|{g_{k}}\rangle}$ state. However, random fluctuations of detuning $E(\delta)$ caused by spatial uncertainty and Doppler broadening leads to spatial de-phasing (see Fig. 6b) and could affect the population rotation designed for the quantum jump protocol. Confinement: Atomic motion in optical lattice causes an average uncertainty of interaction $E_{V}$. The atomic motion could be made negligible by sideband cooling within the optical tweezer Kau12 ; Tho13 and optical lattice Bel13 to the motional ground state. Considering a 3$\mu$m lattice constant, and the motional ground state expansion of 20nm for Rb atoms with trap frequency of $\omega_{tr}/2\pi=150$kHz (close to 125kHz in Kau12 ; Tho13 ), the average relative uncertainty of interaction would be $\frac{E_{V}}{\Omega}=0.1$ for principal number $n$=70 and $\Omega/2\pi=2$MHz. This would cause $P_{s}=2\times 10^{-4}$ de-phasing per step, see Fig. 6b. Doppler broadening: Detuning error could also be caused by Doppler broadening. Using counter-propagating 1013nm and 420nm lasers for two-photon Rydberg excitation, the Doppler detuning would be $\delta_{D}=(k_{1}-k_{2})v$. Considering the sideband cooling to motional ground state, the maximum velocity in the thermal motion would be $v=\sqrt{\frac{\hbar\omega_{tr}}{2m}}=18$mm/s. The random Doppler shift generates a maximum relative uncertainty of $\frac{\delta_{D}}{\Omega}=0.01$ for $\Omega/2\pi=2$MHz. Corresponding de-phasing probability per step is $P_{s}=3\times 10^{-6}$, see Fig. 6b. (a) (b) Figure 6: Spatial de-phasing probability in single step $P_{s}$ as a function of (a) relative laser noise $\gamma_{la}/\Omega$ and (b) transition detuning errors $\frac{E(\delta)}{\Omega}$. #### Spontaneous emission Rydberg levels’ spontaneous emission, reduces both diagonal and off-diagonal terms of density matrix alike. Hence, after the projective measurement, the spontaneous emission does not effect the coherence of the QW operation. However, it would limit the step numbers before loosing the quantum walker as discussed in Sec. V.2 in details. ## V Operation Fidelity in 3D lattice The implementation of 3D QW with coupling tesselations presented in Fig. 2b, benefits from interaction angular dependency. After formulating the anisotropic exchange interaction in Sec. V.1, the operation fidelity of different coupling tessellations are quantified in Sec. V.2. Then, the scaling of achievable step numbers for different principal numbers are discussed. ### V.1 Angular dependent interaction The angular-dependent exchange interaction of ${|{nS_{1/2}1/2,\,nP_{3/2}3/2}\rangle}{\stackrel{{\scriptstyle V}}{{\rightleftharpoons}}}{|{nP_{3/2}3/2,\,nS_{1/2}1/2}\rangle}$ is given by $V=\frac{\bm{\mu}_{1}.\bm{\mu}_{2}-3(\bm{\mu}_{1}.\hat{R})(\bm{\mu}_{2}.\hat{R})}{4\pi\epsilon_{0}R^{3}}$ (4) where $R$ is the interatomic separation and $\vec{\mu_{k}}$ is the electric dipole matrix element, connecting initial and final Rydberg state of $k^{th}$ atom. The angular dependent interaction between sites $i$ and $j$ could be expanded to $\displaystyle V_{ij}(\phi)=\frac{1}{4\pi\epsilon_{0}R_{ij}^{3}}[f_{1}(\phi)(\mu_{1+}\mu_{2-}+\mu_{1-}\mu_{2+}+2\mu_{1z}\mu_{2z})$ $\displaystyle+f_{2}(\phi)(\mu_{1+}\mu_{2z}-\mu_{1-}\mu_{2z}+\mu_{1z}\mu_{2+}-\mu_{1z}\mu_{2-})$ (5) $\displaystyle- f_{3}(\phi)(\mu_{1+}\mu_{2+}-\mu_{1-}\mu_{2-})]=\frac{C_{3}(\phi)}{R_{ij}^{3}},$ where $\phi$ shows inter-atomic orientation with respect to the quantization axis, defined by the propagating direction of exciting polarized lasers. Dipole operators in the spherical basis are denoted by $\mu_{k,\pm}=\mp(\mu_{k,x}\pm i\mu_{k,y})/\sqrt{2}$. The terms associated with pre-factors $f_{1}(\phi)=(1-3\cos^{2}\phi)/2$, $f_{2}(\phi)=\frac{3}{\sqrt{2}}\sin\phi\cos\phi$ and $f_{3}(\phi)=3/2\sin^{2}\phi$ couple Rydberg pairs. Fig. 7b represents the angular dependent exchange interaction for two principal numbers. Figure 7: The anisotropic $nS-nP$ exchange interaction $V$ provides a wider range of laser detuning in both positive and negative sides, available for addressing the desired site. Here $\phi$ is the angle between the inter-atomic orientation and the quantization axis. Quantization axis is defined by the propagation direction of the polarized exciting laser. ### V.2 Coupling tessellations’ fine tuning in 3D This section defines the QW fidelity in terms of the accuracy of population transfer to the desired site and does not encounter the decoherence effects discussed in the previous section. Tuning the laser to be in resonance with the interacting sites separated by ${\bf R}_{ij}$ i.e. $\Delta=-V_{ij}$, the operation infidelity would be the sum of population leakage to unwanted sites $k$ i.e. $F_{ij}=\sum\limits_{k}\frac{\Omega^{2}/4}{(V_{ij}-V_{ik})^{2}}.$ (6) Since the laser is oriented along $R_{ij}$ the denominator would be $C_{3}(0)/R_{ij}^{3}-C_{3}(\phi_{k})/R_{ik}^{3}$ with $\phi_{k}$ being the angle between ${\bf R}_{ij}$ and ${\bf R}_{ik}$. The infidelity of desired coupling tessellations of Fig. 2b are plotted in Fig. 8a as a function of $\Omega/\Delta$ with the analytical approach of Eq. 6 and numerical evaluation of Schrödinger equation considering 8 neighboring lattice sites at different inetr-atomic orientations. The contrast of inter- and intra-dimer lattice constants $(a_{x0}-a_{x1})/a_{x0}$ would define the speed of operation. In general, better contrast leads to faster operation for a given fidelity, see Fig. 8b. However, at some geometries, specific lattice sites might get close to resonance with the laser. This would require slower operation as can be seen at the local minimum in the dashed line of Fig. 8b at $a_{x1}=2.2a_{x0}$. Realizable QW step number scales linearly with Rydberg principal number $n$. While Rydberg spontaneous emission does not affect the coherence, see Sec. IV.2, it would limit the operation number. In this sense, faster operation at constant $\Omega/\Delta$ would require stronger interaction $V_{ij}=C_{3}/a_{ij}^{3}$. While $C_{3}=8.4(n/70)^{3}$MHz.$\mu$m3 for Rb atoms, the minimum lattice constant is limited by the LeRoy-radius and hence scales as $a_{min}=950(n/70)^{2}$nm. Hence, the loss over a single QW operation is scaled by $2\pi/(\Omega\tau)\propto n^{-1}$ where $\tau=450(n/70)^{-3}\mu$s Bet09 is the Rydberg quantum walker lifetime in Rb atoms at $T=77$K. Considering the $W_{x}$ connectivity with $F=97\%$ operation accuracy, after $N=210(n/70)$ step numbers, the Poissonian probability of not losing the walker would be 40%. These numbers would be significantly enhanced by coherent fast excitation of circular states Sig17 ; Car20 featuring exchange interaction and several minutes lifetime Ngu18 . Finally, the initialization and detection must be taking into account while quantifying the device operation. After preselection, the probability of correct state initialization of more than 98% has been achieved Ber17 . Fluorescence detection of ground-state atoms has been realized with 98% fidelity in Rb Kwo17 and 0.99991 fidelity in Sr Cov19 . (a) (b) Figure 8: Fidelity of quantum walk in a 3D lattice of dimers. (a) Infidelity of desired coupling tessellations of Fig. 2b are plotted as a function of $\Omega/\Delta$ with analytical (lines) and numerical simulation (circles) considering 8 neighbouring lattice sites. (b) The required relative laser coupling $\Omega/\Delta$ is plotted as a function of the contrast of inter- and intra-dimer lattice constants $(a_{x0}-a_{x1})/a_{x0}$ to realize connectivities with 99% fidelity. The local minimum in the dashed line is caused by specific lattice site that gets close to resonance with the laser at $a_{x1}=2.2a_{x0}$. The lattice constants are $a_{x0}=a_{y}=a_{z}=1\mu$m in (a,b) and $a_{x1}=1.5$ in (a). ## VI Topologically protected edge state/ Floquet topological insulators Application of SSH models in making topological matters are vastly studied. The proposed discrete-time SSH model in this article could be used for making an enriched form of Floquet topological insulators and topologically protected edge state, similar to the coined DTQWs approach Kit10 . ### VI.1 Fourier transformed Hamiltonian For the QWs on an infinitely extended lattice, or a lattice with periodic torus boundary conditions, the QW operators can be transformed into quasi- momentum space $\tilde{H}$. The Fourier transformation of the odd $o$ and even $e$ sites in 1D is given by $\displaystyle{|{m,e}\rangle}=\frac{1}{\sqrt{\frac{N}{2}}}\sum\limits_{k}\text{e}^{\text{i}km}{|{k,e}\rangle}$ (7) $\displaystyle{|{m,o}\rangle}=\frac{1}{\sqrt{\frac{N}{2}}}\sum\limits_{k}\text{e}^{\text{i}km-\text{i}k\bar{a}_{1}}{|{k,o}\rangle}$ where ${|{{k},e}\rangle}={|{k}\rangle}\otimes{|{e}\rangle}$. Also $\bar{a}_{1}=\frac{a_{1}}{a_{0}+a_{1}}$ with $a_{0}$, $a_{1}$ being the inter- and intra-dimer lattice constants. The wave-number is chosen to take on values from the first Brillouin zone $k=l\frac{2\pi}{N/2}$ with $1\leq l\leq\frac{N}{2}$. The Fourier transformed Hamiltonian obtains by replacing Eq. 7 into the QW Hamiltonians of Eq. 1, simplified to $\displaystyle\tilde{H}_{0}=\sum_{k}\begin{pmatrix}0&\text{e}^{\text{i}{k}\bar{a}_{1}}\\\ \text{e}^{-\text{i}{k}\bar{a}_{1}}&0\end{pmatrix}{|{k}\rangle}\\!{\langle{k}|}$ (8) $\displaystyle\tilde{H}_{1}=\sum_{k}\begin{pmatrix}0&\text{e}^{-\text{i}k(1-\bar{a}_{1})}\\\ \text{e}^{\text{i}k(1-\bar{a}_{1})}&0\end{pmatrix}{|{k}\rangle}\\!{\langle{k}|},$ with matrices being presented in the $\\{{|{o}\rangle},{|{e}\rangle}\\}$ basis. ### VI.2 Edge states in 1D The Floquet topological phases of Rydberg coinless discrete-time quantum walk can be accessed by looking at the full-time evolution of the walk. The quantum operator in momentum basis $\displaystyle\tilde{W}_{\text{eff}}=\text{e}^{\text{i}\frac{\theta_{0}}{2}\tilde{H}_{0}}\text{e}^{\text{i}\theta_{1}\tilde{H}_{1}}\text{e}^{\text{i}\frac{\theta_{0}}{2}\tilde{H}_{0}}$ (9) can be written as $\tilde{W}_{\text{eff}}=\text{e}^{\text{i}\tilde{H}_{\text{eff}}T}$, where $T$ is the period of applying a set of QW operators of Eq. 9. In a lattice of dimers, $\tilde{H}_{\text{eff}}$ has two bands and hence the effective Hamiltonian could be written as $\tilde{H}_{\text{eff}}=\sum\limits_{k}E(k)\bm{n}(k).\boldsymbol{\sigma}\,{|{k}\rangle}\\!{\langle{k}|}$ (10) where $\boldsymbol{\sigma}=(\sigma_{x},\sigma_{y},\sigma_{z})$ is the vector of Pauli matrices operating in the odd-even basis $\\{{|{o}\rangle},{|{e}\rangle}\\}$ and the $\bm{n}(k)=(n_{x},n_{y},n_{z})$ defines the quantization axis for the spinor eigenstates at each momentum $k$. The quantization axis of the eigen states are given by $n_{x}=-\frac{(\sin\theta_{0}\cos\theta_{1}\cos(k\bar{a}_{1})+\cos(k+k\bar{a}_{1})\cos\theta_{0}\sin\theta_{1})}{\sin E(k)}$, $n_{y}=$ $\frac{-\cos\theta_{0}\sin\theta_{1}\cos k\sin(k\bar{a}_{1})+\sin\theta_{0}\sin k\cos(k\bar{a}_{1})-\sin\theta_{0}\cos\theta_{1}\sin(k\bar{a}_{1})}{\sin E(k)}$, $n_{z}=0$. The band structure of quasi-energy formed by discrete translational invariance in time, which is given by $\cos E(k)=\cos\theta_{0}\cos\theta_{1}-\sin\theta_{0}\sin\theta_{1}\cos(k),$ (11) and plotted in Fig. 9b. The two bands correspond to the two degrees of freedom of the system. The minimum and maximum of two bands separation are given by $|\theta_{0}\pm\theta_{1}|$. With homogeneous coupling $\theta_{0}=\theta_{1}$, the system would resemble a conductor. In that case, there are plane wave eigenstates with small energy, which transport the excitation from one end to the other end of the chain. As the gap due to the difference of the hopping angles opens, the energy of occupied states is lowered, while unoccupied states move to higher energies. Thus, the staggering is energetically favourable. Also, $\theta_{0}=\pi/2,\theta_{1}=0$ results in a chain of isolated dimers. Hence there will be no transformation of the walker and hence the energy $E$ would be $k$ independent. Figure 9: Topological edge state in 1D coinless DTQW. (a) Topological invariants (winding number $\nu_{0},\nu_{\pi}$) associated with the hopping angles $\theta_{0}$ and $\theta_{1}$, see Eq. 13. (b) Band structure of the quantum walk for $\theta_{0}=\pi/4$, $\theta_{1}=[\frac{\pi}{8},\frac{\pi}{4},\frac{3\pi}{8}]$. (c,d) shows the walker’s wave-function after 100 steps for the space-dependent hopping angles explained in Eq. 14 and over the range depicted by arrows in (a). The walker is initialized at the centre $x=0$ where the sharp change of $\theta$ happens $w=0.1a_{0}$. In (c) the positive and negative sites have different winding numbers, providing the topological phase transition and leading to bound state at the centre. In (d) the evolution is adiabatic (no crossing of the phase boundary) and hence no population traps at the centre. In 1D, the topological phase is originated from the chiral symmetry. The chiral symmetry exist, if there is a unitary operator fulfilling the following relation $\Gamma H_{\text{eff}}\Gamma^{\dagger}=-H_{\text{eff}}$. While the trivial QW operator $\tilde{W}_{\text{eff}}=\text{e}^{\text{i}\theta_{1}\tilde{H}_{1}}\text{e}^{\text{i}\theta_{0}\tilde{H}_{0}}$ do not show the chrial symmetry, splitting one of the steps and moving the time frame to either of the following forms $\displaystyle\tilde{W}_{\text{eff}}=\text{e}^{\text{i}\frac{\theta_{0}}{2}\tilde{H}_{0}}\text{e}^{\text{i}\theta_{1}\tilde{H}_{1}}\text{e}^{\text{i}\frac{\theta_{0}}{2}\tilde{H}_{0}}$ (12) $\displaystyle\tilde{W}^{\prime}_{\text{eff}}=\text{e}^{\text{i}\frac{\theta_{1}}{2}\tilde{H}_{1}}\text{e}^{\text{i}\theta_{0}\tilde{H}_{0}}\text{e}^{\text{i}\frac{\theta_{1}}{2}\tilde{H}_{1}},$ the quantum walk do exhibit chiral symmetry with $\Gamma=\sigma_{z}$. The chiral symmetry enforces the $\bm{n}(\theta_{0},\theta_{1},k)$ to rotate in the plain perpendicular to $\mathbf{\Gamma}$. The corresponding topological invariants is the number of times $\bm{n}(\theta_{0},\theta_{1},k)$ winds around the origin, as the quasi-momentum $k$ runs in the first Brillouin zone, called winding number $\nu=\frac{1}{2\pi}\int\limits_{0}^{2\pi}\frac{1}{|\boldsymbol{n}|^{2}}(n_{x}\partial_{k}n_{y}-n_{y}\partial_{k}n_{x})\text{d}k.$ (13) The two nonequivalent shifted time-frames $\tilde{W}_{\text{eff}}$, $\tilde{W}^{\prime}_{\text{eff}}$ (Eq. 12), leads to two winding numbers $\nu,\nu^{\prime}$. The two invariants $\nu_{0}=(\nu+\nu^{\prime})/2$, $\nu_{\pi}=(\nu-\nu^{\prime})/2$ would completely describe the topology. The phase diagram of the winding number is plotted in Fig. 9a. The band structure in Fig. 9b are made of the energy eigenvalues $E(k,\theta_{0},\theta_{1})$. Manipulating the hopping angles $\theta_{0},\theta_{1}$ within the distinguished regions of Fig. 9a, would make the system to continuously transit between the band structures without closing the energy gap, i.e. without changing the topological character of the system that is the winding number in here. At the borders separating distinct topological regions, the band structure closes, see Fig. 9b. The topological character can be revealed at the boundary between topologically distinct phases. To implement such a boundary one can apply inhomogeneous spatial hoping angle of the form $\theta_{i}(x)=\frac{\theta_{i-}+\theta_{i+}}{2}+\frac{\theta_{i+}-\theta_{i-}}{2}\tanh(x/w)$ (14) where $w$ determines the spatial width of phase transition region which defines the width of the bound state. The variation of the hopping angle in the Rydberg system could be realized by different sets of exciting lasers or by applying space-dependent Stark-shift using an external field. Fig. 9c(d) shows the walker’s wave-function after 100 steps for the cases with (without) the phase transition. The walker is initialized at $x=0$ in both cases. In Fig. 9c the positive and negative sites have different winding numbers. This would lead to the topological phase transition and hence forms bound state at the centre. In Fig. 9d, the evolution is adiabatic (i.e. there is no crossing of the phase boundary) and no population traps at the centre. ### VI.3 Floquet Topological Insulators with multi-dimensional DTQW Different classes of floquet topological insulators could be realized by the multi-dimensional coupling tessellations introduced in Sec. III. The $nP$ excitation propagates unidirectionally and without backscattering along the edge, that is a line/surface in 2D/3D lattice, and eventually distributes uniformly along the boundary. Unlike in 1D, symmetry is not required for the presence of topological phase in multi-dimensions. Using the operators of Eq3 , a set of Hamiltonian $W_{\text{eff}}=T_{x}R_{\theta_{0}}T_{y}R_{\theta_{1}}T_{xy}R_{\theta_{0}}$ could be used for making 2D topological insulators with non-vanishing Chern numbers, see Supp and Kit10 . Unlike the time-independent QW in the time- dependent approach, topologically protected (TP) edge states could be formed even in the cases where the Chern number is zero for all the bands. As an example, the simple model of coined QW steps of $W_{\text{eff}}=T_{y}R_{\theta_{1}}T_{x}R_{\theta_{0}}$, leads to a 2D topological insulator, see Supp. Here, the Chern number of both bands are zero and the presented topological invariant is the Rudner winding number rud13 . Multi-dimensional topological insulators could also be formed by Rydberg discrete-time SSH model as discussed in the Supp. A 3D topological insulators could be realized by the current proposal under a simple set of operators $W_{\text{eff}}=T_{x}R_{\theta_{0}}T_{y}R_{\theta_{1}}T_{z}R_{\theta_{0}}$, where the elements are defined in Eq3 . The topological phase diagram is plotted in Fig. 10a, where the gap between the two bands closes at $E=0$ ($E=\pi$) at the red (blue) borders. TP edge states in 3D exist along the surfaces in spatial borders between two regions with distinct topological phases. A particle, that is prepared in the superposition of the TP edge states, propagates coherently along the spatial border. Fig. 10b demonstrate TP propagating edge modes by considering an inhomogeneous 3D coined DTQW with spatially inhomogeneous coin angles. Here, a flat surface borders is considered. The pair of coin angles inside and outside the strip belongs to different topological phases. The hopping angles are $(\theta_{0},\theta_{1})=(4\pi/10,\pi/10)$ inside $3\leq x\leq 5$ and $(\theta_{0},\theta_{1})=(\pi/10,4\pi/10)$ outside the stripe $x<3$ and $x>5$. The excitation is initialized on the borders as ${|{\psi_{0}}\rangle}=({|{x=3,y=3,z=3,o}\rangle}+{|{x=5,y=3,z=3,o}\rangle})/\sqrt{2}$. The excitation would distribute over the border surface after large step numbers ( $N_{step}=200$ in here). Figure 10: Edge state in a 3D array of a coined Rydberg DTQW $W_{\text{eff}}=T_{x}R_{\theta_{0}}T_{y}R_{\theta_{1}}T_{z}R_{\theta_{0}}$ with elements presented in Eq. LABEL:Eq_T3DCoined. (a) The topological phase diagram are plotted where the two bands closes at $E=0$ ($E=\pi$) at the red (blue) borders. (b) Edge state are formed on the surface boundaries $x=3$ and $x=5$. It would distribute on the edge surface after large step numbers ($N_{step}=200$ in here). The hoping angles are $(\theta_{0},\theta_{1})=(4\pi/10,\pi/10)$ inside $3\leq x\leq 5$ and $(\theta_{0},\theta_{1})=(\pi/10,4\pi/10)$ outside the stripe $x<3$ and $x>5$. Periodic boundary condition is imposed along $x$, $y$ and $z$ dimensions. Each site represents the sum of dimer’s elements population. ## VII Conclusion and Outlook This article Khazali21 shows that smart Rydberg excitation in a holographically designed atomic lattice would act as a universal simulator of topological phases and boundary states by performing varieties of DTQW operators. In the project’s outlook, the presented model could be used for the simulation of electron movement on the Fermi surfaces that are 2D manifolds fitted in the Brillouin zone of a crystal (electron movement on the Fermi surfaces that cut each other like Kline bottle surface) Wie16 . The other extension avenue would be obtained by adding synthetic magnetic fields to the current DTQW model. This would obtain by applying a gradient of an external electric field resulting in magnetic QW with applications in making Chern insulators saj18 . The new features of the presented model in applying QW on topological surfaces provides a platform to study the performance of QW based algorithms on topologically ordered databases. ## Supplemental material ## A1: Self-interaction Having one delocalized Rydberg excitation, the partial Rydberg population at different sites do not interact with each other. The nonlocal wave function only gives the probability of finding the single excitation at different sites. Simulating the Schrodinger equation showed that partial population $P_{{|{p}\rangle}}$ at a specific site $i$ would induce the same population of auxiliary state $P_{{|{s}\rangle}}=P_{{|{p}\rangle}}$ in the the resonant sites $i$ and $j$. Hence, when the single walker is not localized, the total population of the non-local auxiliary Rydberg state would add up to 1. Therefore the absence of self interaction argument also applies to the induced single Rydberg auxiliary ${|{s}\rangle}$ state. This argument is similar to the absence of self-interaction over a Rydberg polariton excited by a single photon in an atomic ensemble. The other point of similarity is in the Rydberg population dependence. In the Rydberg electromagnetically induced transparency (EIT), the ladder two-step excitation is derived by a faint quantum and a strong classical light. The population of Rydberg level is a function of the photon number in the quantum light. Similarly in the proposed DTQW here, the maximum population of the auxiliary ${|{s}\rangle}$ state excited by the strong laser shining to multiple ground state atoms is defined by the population of ${|{p}\rangle}$ state that makes the strong field in resonance with the laser transition. Therefore, the argument of the self-interaction also applies to the ${|{s}\rangle}$ single excitation. Following this argument, Eq. II only includes the inter-excitation interaction $V_{S-P}$ and does not consider self interactions $V_{S-S}$ and $V_{P-P}$. ### A2: 2D Topological insulator #### Coined DTQW Figure 11: Edge state in a 2D array of coined Rydberg DTQW. TP edge states formed under coined quantum walk with the operator of (a,b) $W_{\text{eff}}=T_{x}R_{\theta_{0}}T_{y}R_{\theta_{1}}T_{xy}R_{\theta_{0}}$ and (c,d) $W_{\text{eff}}=T_{y}R_{\theta_{1}}T_{x}R_{\theta_{0}}$ with elements discussed in Eq. LABEL:Eq_T3DCoined. The topological phase diagram are plotted with the topological invariant that is the (a) chern number and (b) Rudner winding number. The energy gap of the two bands closes at 0 ($\pi$) at the red (blue) borders. (b,d) Edge state is around the linear boundary at $x=2$ and $x=4$. It would distribute along the edge after large step numbers ($N_{step}=200$ in here). The hoping angles are $(\theta_{0},\theta_{1})=(\pi/10,4\pi/10)$ inside $2\leq x\leq 4$ and $(\theta_{0},\theta_{1})=(4\pi/10,\pi/10)$ outside the stripe $x<2$ and $x>4$. Periodic boundary condition is imposed along both $x$ and $y$ dimensions. Each site represents the sum of dimer’s elements population. The coined Rydberg DTQW operators of Eq3 , could ideally implement topological insulators with non-vanishing Chern numbers Kit10 . The concatenated operators $W_{\text{eff}}=T_{x}R_{\theta_{0}}T_{y}R_{\theta_{1}}T_{xy}R_{\theta_{0}}$ could be used for making topological insulators. To quantify the topological properties of this QW, the effective Hamiltonian in the momentum space is considered $\tilde{W}_{\text{eff}}=\text{e}^{\text{i}\tilde{H}_{\text{eff}}T}$. Like in the one-dimensional case discussed above, the discrete-time quantum walk is a stroboscopic simulator of the evolution generated by $\tilde{H}_{\text{eff}}$ at discrete-times. The effective Hamiltonian would be $\tilde{H}_{\text{eff}}=\sum\limits_{{\bm{k}}}E({\bm{k}})\bm{n}({\bm{k}}).\boldsymbol{\sigma}{|{{\bm{k}}}\rangle}\\!{\langle{{\bm{k}}}|}$ (15) where $\boldsymbol{\sigma}$, is the vector of Pauli matrices dealing with odd and even bases and the $\bm{n}({\bm{k}})$ defines the quantization axis for the spinor eigenstates at each momentum ${\bm{k}}$. The topological invariant in this two dimensional Brillouin zone is given by the Chern number as $C=\frac{1}{4\pi}\int\text{d}{\bm{k}}\,{\bm{n}}.(\partial_{k_{x}}{\bm{n}}\times\partial_{k_{y}}{\bm{n}}).$ (16) The phase diagram of the quantum walk is plotted in Fig. 11a. Red and blue borders are associated with points where the energy gap closes at $E=0$ and $E=\pi$ respectively. Like in the 1D case, TP edge states exist in 2D in spatial borders between two regions with distinct topological phases. These states are propagating uni- directionally in the bulk gaps and connect the bulk bands. A particle, that is prepared in the superposition of the TP edge states, propagates coherently along the spatial border. The chirality of the edge states is topologically protected. In another word, their directions of propagation do not change under continuous variation of the parameters of the system as long as the bulk gaps remain open. Fig. 11b,d demonstrate TP propagating edge modes by considering an inhomogeneous 2D coined QW with spatial dependent coin angles. Flat borders in the form of a strip geometry is considered in the 2D lattice, where the pair of coin angles inside and outside the strip belongs to different topological phases. The hoping angles are $(\theta_{0},\theta_{1})=(\pi/10,4\pi/10)$ inside $2\leq x\leq 4$ and $(\theta_{0},\theta_{1})=(4\pi/10,\pi/10)$ outside the stripe $x<2$ and $x>4$. The excitation is initialized on the borders as $({|{x=2,y=5,e}\rangle}+{|{x=4,y=5,e}\rangle})/\sqrt{2}$. The excitation would distribute along the border after large step numbers ( $N_{step}=200$ in here). Fig. 11c,d discusses a simpler model of QW steps of $W_{\text{eff}}=T_{y}R_{\theta_{1}}T_{x}R_{\theta_{0}}$ with operators being presented in Eq3 . In the phase diagram of Fig. 11c, the Chern number of both bands are zero and the presented topological invariant is the Rudner winding number rud13 . #### Coinless DTQW Figure 12: Topological insulators in a 2D array of (a,b) tetramers and (c) dimers with coinless DTQW. (a) Applying an inhomogeneous hopping angle along the $\hat{x}$ dimension with $(\theta_{x0},\theta_{x1})=(\pi/10,4\pi/10)$ inside $3<x\leq 9$ and $(\theta_{x0},\theta_{x1})=(4\pi/10,\pi/10)$ outside the stripe ($x\leq 3$ and $x>9$), the excitation propagates unidirectionally along the borders of the stripe. The hoping angles along the $y$ direction $(\theta_{y0},\theta_{y1})=(\pi/3,\pi/10)$ is not changing over different spatial regions. The population would distribute along the edge after large step numbers ( $N_{step}=209$ in here). (b) The corresponding 4 bands associated with the tetramer sublattice are plotted for the hopping angle parameter choice inside the stripe. (c) In the 2D lattice of dimers all bulk states are localized, see the red arrow. A particle initially localized at any site in the bulk returns to its original position in one set of QW operations. The coinless models of a 2D lattice of tetramers as shown in Fig. 2a, could also be used for making TP edge states. Here the QW are effectively separable in two-dimensions. Fig. 12a shows that by applying inhomogeneous hopping angles belonging to different regions of Fig. 9a, along $\hat{x}$ dimension and using open(periodic) boundaries along $\hat{x}$($\hat{y}$) direction, one can form excitation currents solely on the borders. Coin-less DTQW in a 2D lattice of dimers shown in Fig. 2b could be used for the realization of the anomalous topological edge states rud13 . Here the QWs in different dimensions are not separable. In an intuitive discussion, applying the operator $W=\text{e}^{\text{i}H_{x0}\pi/2}\text{e}^{\text{i}H_{xy1}\pi/2}\text{e}^{\text{i}H_{x1}\pi/2}\text{e}^{\text{i}H_{xy0}\pi/2}$ (with the Hamiltonians defined in Eq2 ), on the excitation initialized on the border, leads to the clockwise transportation of the walker along the boarder as depicted by green arrows in Fig. 12c. This is while the initialized excitation in the bulk would go under unitary operator with no excitation transport, see red arrow in Fig. 12c. The exclusive conductance on the boundary provides the desired topological insulator. One can use inhomogeneous spatial angles with different Rudner winding numbers rud13 to design the shape of the edge state. Topological insulators could also get implemented with 3D coinless DTQW model. For example $W=W_{x1}W_{xz1}W_{x0}W_{xz0}W_{xy0}W_{x0}W_{xy1}W_{x1}$ where $W_{i}=\exp(\text{i}H_{i}\theta_{i})$ with Hamiltonians defined in Eq2 and uniform $\theta=\pi/2$ hopping angles on an open boundary lattice, results in an insulating bulk with exclusive excitation currents on the open boundaries. ## References * (1) Y. Aharonov, L. Davidovich, and N. Zagury. Quantum random walks. Physical Review A, 48(2):1687, 1993. * (2) E. Farhi and S. Gutmann. Quantum computation and decision trees. Physical Review A, 58(2):915, 1998. * (3) J. Kempe. Quantum random walks: an introductory overview. Contemporary Physics, 50(1):339–359, 2009. * (4) S. Dadras, A. Gresch, C.Groiseau, S. Wimberger, and G. S Summy. Quantum walk in momentum space with a bose-einstein condensate. Physical review letters, 121(7):070402, 2018. * (5) G. Summy and S. Wimberger. Quantum random walk of a bose-einstein condensate in momentum space. Physical Review A, 93(2):023638, 2016. * (6) P. M Preiss, R. Ma, M E. Tai, A. Lukin, M. Rispoli, P. Zupancic, Y. Lahini, R. Islam, and M. Greiner. Strongly correlated quantum walks in optical lattices. Science, 347(6227):1229–1233, 2015. * (7) R. Portugal. Quantum walks and search algorithms. Springer, 2013. * (8) N. Shenvi, J. Kempe, and K B. Whaley. Quantum random-walk search algorithm. Physical Review A, 67(5):052307, 2003. * (9) A. M Childs, R. Cleve, E. Deotto, E. Farhi, S. Gutmann, and D. A Spielman. Exponential algorithmic speedup by a quantum walk. In Proceedings of the thirty-fifth annual ACM symposium on Theory of computing, pages 59–68, 2003. * (10) A. M Childs and J. Goldstone. Spatial search by quantum walk. Physical Review A, 70(2):022314, 2004. * (11) R. Portugal and T. D Fernandes. Quantum search on the two-dimensional lattice using the staggered model with hamiltonians. Physical Review A, 95(4):042341, 2017. * (12) A. M Childs. Universal computation by quantum walk. Physical review letters, 102(18):180501, 2009. * (13) V. Kendon, How to compute using quantum walks. arXiv preprint arXiv:2004.01329, 2020. * (14) N. B Lovett, S. Cooper, M. Everitt, M. Trevers, and V. Kendon. Universal quantum computation using the discrete-time quantum walk. Physical Review A, 81(4):042330, 2010. * (15) A. M Childs, D. Gosset, and Z. Webb. Universal computation by multiparticle quantum walk. Science, 339(6121):791–794, 2013. * (16) S. Elías V.-Andraca. Quantum walks: a comprehensive review. Quantum Information Processing, 11(5):1015–1106, 2012. * (17) T. Kitagawa, M. S Rudner, E. Berg, and E. Demler. Exploring topological phases with quantum walks. Physical Review A, 82(3):033429, 2010. * (18) H. Schmitz, R. Matjeschk, Ch. Schneider, J. Glueckert, M. Enderlein, T. Huber, and T. Schaetz. Quantum walk of a trapped ion in phase space. Phys. Rev. Lett., 103:090504, Aug 2009. * (19) F. Zähringer, G. Kirchmair, R. Gerritsma, E. Solano, R. Blatt, and C. F. Roos. Realization of a quantum walk with one and two trapped ions. Phys. Rev. Lett., 104:100503, Mar 2010. * (20) M. Karski, L. Förster, J. Choi, A. Steffen, W. Alt, D. Meschede, and A. Widera. Quantum walk in position space with single optically trapped atoms. Science, 325(5937):174–177, 2009. * (21) C. Weitenberg, M. Endres, J. F Sherson, M. Cheneau, P. Schauß, T. Fukuhara, I. Bloch, and S. Kuhr. Single-spin addressing in an atomic mott insulator. Nature, 471(7338):319–324, 2011. * (22) T. Fukuhara, P. Schauß, M. Endres, S. Hild, M. Cheneau, I. Bloch, and C. Gross. Microscopic observation of magnon bound states and their dynamics. Nature, 502(7469):76–79, 2013. * (23) Ji. Wang and K. Manouchehri. Physical implementation of quantum walks. Springer, 2013. * (24) H. Bernien, S. Schwartz, A. Keesling, H. Levine, A. Omran, H. Pichler, S. Choi, A. S Zibrov, M. Endres, M. Greiner, et al. Probing many-body dynamics on a 51-atom quantum simulator. Nature, 551(7682):579–584, 2017. * (25) A. Omran, H. Levine, A. Keesling, G. Semeghini, T. T Wang, S. Ebadi, H. Bernien, A. S Zibrov, H. Pichler, S. Choi, et al. Generation and manipulation of schrödinger cat states in rydberg atom arrays. Science, 365(6453):570–574, 2019. * (26) S Zhang, F Robicheaux, and M Saffman. Magic-wavelength optical traps for rydberg atoms. Physical Review A, 84(4):043408, 2011. * (27) MJ Piotrowicz, M Lichtman, K Maller, G Li, S Zhang, L Isenhower, and M Saffman. Two-dimensional lattice of blue-detuned atom traps using a projected gaussian beam array. Physical Review A, 88(1):013420, 2013. * (28) F. Nogrette, H. Labuhn, S. Ravets, Daniel Barredo, Lucas Béguin, Aline Vernier, Thierry Lahaye, and Antoine Browaeys. Single-atom trapping in holographic 2d arrays of microtraps with arbitrary geometries. Physical Review X, 4(2):021034, 2014. * (29) T Xia, M Lichtman, K Maller, AW Carr, MJ Piotrowicz, L Isenhower, and M Saffman. Randomized benchmarking of single-qubit gates in a 2d array of neutral-atom qubits. Physical review letters, 114(10):100503, 2015. * (30) J. Zeiher, R. Van Bijnen, P. Schauß, S. Hild, J. Choi, T. Pohl, I. Bloch, and C. Gross. Many-body interferometry of a rydberg-dressed spin lattice. Nature Physics, 12(12):1095–1099, 2016. * (31) V. Lienhard, S. de Léséleuc, D. Barredo, T. Lahaye, A. Browaeys, M. Schuler, L. P. Henry, and A. M Läuchli. Observing the space-and time-dependent growth of correlations in dynamically tuned synthetic ising models with antiferromagnetic interactions. Physical Review X, 8(2):021070, 2018. * (32) MA Norcia, AW Young, and AM Kaufman. Microscopic control and detection of ultracold strontium in optical-tweezer arrays. Physical Review X, 8(4):041054, 2018. * (33) A. Cooper, J. P Covey, I. S Madjarov, S. G Porsev, M. S Safronova, and M. Endres. Alkaline-earth atoms in optical tweezers. Physical Review X, 8(4):041055, 2018. * (34) S. Hollerith, J. Zeiher, J. Rui, A. Rubio-Abadal, V. Walther, T. Pohl, D. M Stamper-Kurn, I. Bloch, and C. Gross. Quantum gas microscopy of rydberg macrodimers. Science, 364(6441):664–667, 2019. * (35) S. Saskin, JT Wilson, B. Grinkemeyer, and J. D. Thompson. Narrow-line cooling and imaging of ytterbium atoms in an optical tweezer array. Physical review letters, 122(14):143002, 2019. * (36) Y. Wang, A. Kumar, T.-Y. Wu, and D. S Weiss. Single-qubit gates based on targeted phase shifts in a 3d neutral atom array. Science, 352(6293):1562–1565, 2016. * (37) D. Barredo, V. Lienhard, S. De Leseleuc, T. Lahaye, and A. Browaeys. Synthetic three-dimensional atomic structures assembled atom by atom. Nature, 561(7721):79–82, 2018. * (38) H. Levine, A. Keesling, A. Omran, H. Bernien, S. Schwartz, A. S Zibrov, M. Endres, M. Greiner, V. Vuletić, and M. D Lukin. High-fidelity control and entanglement of rydberg-atom qubits. Physical review letters, 121(12):123603, 2018. * (39) M. Saffman, T. G Walker, and K. Mølmer. Quantum information with rydberg atoms. Reviews of modern physics, 82(3):2313, 2010. * (40) CS Adams, JD Pritchard, and JP Shaffer. Rydberg atom quantum technologies. Journal of Physics B: Atomic, Molecular and Optical Physics, 53(1):012002, 2019. * (41) M. Khazali, K. Heshami, and C. Simon. Photon-photon gate via the interaction between two collective rydberg excitations. Physical Review A, 91(3):030301, 2015. * (42) M. Khazali and K. Mølmer. Fast multiqubit gates by adiabatic evolution in interacting excited-state manifolds of rydberg atoms and superconducting circuits. Physical Review X, 10(2):021054, 2020. * (43) M. Khazali. Rydberg noisy-dressing and applications in making soliton-molecules and droplet quasi-crystals. arXiv preprint arXiv:2007.01039, 2020. * (44) M. Khazali. Applications of atomic ensembles for photonic quantum information processing and fundamental tests of quantum physics. 2016\. * (45) Khazali, M. Quantum Information and Computation with Rydberg Atoms. Iranian Journal of Applied Physics, 10(4): 19, (2021). * (46) M. Khazali, C. R Murray, and T. Pohl. Polariton exchange interactions in multichannel optical networks. Physical Review Letters, 123(11):113605, 2019. * (47) M. Khazali, H. W. Lau, A. Humeniuk, and C. Simon. Large energy superpositions via rydberg dressing. Physical Review A, 94(2):023408, 2016. * (48) M. Khazali, K. Heshami, and C. Simon. Single-photon source based on rydberg exciton blockade. Journal of Physics B: Atomic, Molecular and Optical Physics, 50(21):215301, 2017. * (49) M. Khazali. Progress towards macroscopic spin and mechanical superposition via rydberg interaction. Physical Review A, 98(4):043836, 2018. * (50) R. Côté, A. Russell, E. E Eyler, and P. L Gould. Quantum random walk with rydberg atoms in an optical lattice. New Journal of Physics, 8(8):156, 2006. * (51) D. Barredo, H. Labuhn, S. Ravets, T. Lahaye, A. Browaeys, and C. S Adams. Coherent excitation transfer in a spin chain of three rydberg atoms. Physical review letters, 114(11):113002, 2015. * (52) DW Schönleber, A. Eisfeld, M. Genkin, S Whitlock, and S. Wüster. Quantum simulation of energy transport with embedded rydberg aggregates. Physical review letters, 114(12):123005, 2015. * (53) A Pineiro Orioli, A Signoles, H Wildhagen, G Günter, J Berges, S Whitlock, and M Weidemüller. Relaxation of an isolated dipolar-interacting rydberg quantum spin system. Physical review letters, 120(6):063601, 2018. * (54) G Günter, H Schempp, M Robert-de Saint-Vincent, V Gavryusev, S Helmrich, CS Hofmann, S Whitlock, and M Weidemüller. Observing the dynamics of dipole-mediated energy transport by interaction-enhanced imaging. Science, 342(6161):954–956, 2013. * (55) H Schempp, G Günter, S Wüster, M Weidemüller, and S Whitlock. Correlated exciton transport in rydberg-dressed-atom spin chains. Physical review letters, 115(9):093002, 2015. * (56) F. Letscher and . Petrosyan. Mobile bound states of rydberg excitations in a lattice. Physical Review A, 97(4):043415, 2018. * (57) S Wüster, C Ates, A Eisfeld, and JM Rost. Excitation transport through rydberg dressing. New Journal of Physics, 13(7):073044, 2011. * (58) A. Dauphin, M. Müller, and M. A. Martin-Delgado. Quantum simulation of a topological mott insulator with rydberg atoms in a lieb lattice. Physical Review A, 93(4):043611, 2016. * (59) Y. Ando. Topological insulator materials. Journal of the Physical Society of Japan, 82(10):102001, 2013. * (60) J. Cayssol, B. Dóra, F. Simon, and R. Moessner. Floquet topological insulators. physica status solidi (RRL)–Rapid Research Letters, 7(1-2):101–108, 2013. * (61) A. Kitaev. Periodic table for topological insulators and superconductors. In AIP conference proceedings, volume 1134, pages 22–30. American Institute of Physics, 2009. * (62) S Panahiyan and S Fritzsche. Toward simulation of topological phenomenas with one-, two-and three-dimensional quantum walks. arXiv preprint arXiv:2005.08720, 2020. * (63) Mikael C Rechtsman, Julia M Zeuner, Yonatan Plotnik, Yaakov Lumer, Daniel Podolsky, Felix Dreisow, Stefan Nolte, Mordechai Segev, and Alexander Szameit. Photonic floquet topological insulators. Nature, 496(7444):196–200, 2013. * (64) L Xiao, X Zhan, ZH Bian, KK Wang, X Zhang, XP Wang, J Li, K Mochizuki, D Kim, N Kawakami, et al. Observation of topological edge states in parity–time-symmetric quantum walks. Nature Physics, 13(11):1117–1123, 2017. * (65) S. Mukherjee, A. Spracklen, M. Valiente, E. Andersson, P. Öhberg, N. Goldman, and R. R Thomson. Experimental observation of anomalous topological edge modes in a slowly driven photonic lattice. Nature communications, 8(1):1–7, 2017. * (66) W_P Su, JR Schrieffer, and Ao J Heeger. Solitons in polyacetylene. Physical review letters, 42(25):1698, 1979. * (67) M. S Rudner, N. H Lindner, E. Berg, and M. Levin. Anomalous edge states and the bulk-edge correspondence for periodically driven two-dimensional systems. Physical Review X, 3(3):031005, 2013. * (68) M. A Schlosshauer. Decoherence: and the quantum-to-classical transition. Springer Science & Business Media, 2007. * (69) R. B Hutson, A. Goban, G E. Marti, L. Sonderhouse, C. Sanner, and J. Ye. Engineering quantum states of matter for atomic clocks in shallow optical lattices. Physical review letters, 123(12):123401, 2019. * (70) A. M Kaufman, B. J Lester, and C. A Regal. Cooling a single atom in an optical tweezer to its quantum ground state. Physical Review X, 2(4):041014, 2012. * (71) J. D. Thompson, TG Tiecke, A. S Zibrov, V Vuletić, and M. D Lukin. Coherence and raman sideband cooling of a single atom in an optical tweezer. Physical review letters, 110(13):133001, 2013. * (72) N. Belmechri, L. Förster, W. Alt, A. Widera, D. Meschede, and A. Alberti. Microwave control of atomic motional states in a spin-dependent optical lattice. Journal of Physics B: Atomic, Molecular and Optical Physics, 46(10):104006, 2013. * (73) II Beterov, II Ryabtsev, DB Tretyakov, and VM Entin. Quasiclassical calculations of blackbody-radiation-induced depopulation rates and effective lifetimes of rydberg n s, n p, and n d alkali-metal atoms with n? 80. Physical Review A, 79(5):052504, 2009. * (74) T. Long Nguyen, J. M. Raimond, C. Sayrin, R. Cortinas, T. Cantat-Moltrecht, F. Assemat, I. Dotsenko, S. Gleyzes, S. Haroche, G. Roux, et al. Towards quantum simulation with circular rydberg atoms. Physical Review X, 8(1):011032, 2018. * (75) Signoles, A., Dietsche, E.K., Facon, A., Grosso, D., Haroche, S., Raimond, J.M., Brune, M. and Gleyzes, S., Coherent transfer between low-angular-momentum and circular rydberg states, Physical review letters 118, 253603 (2017). * (76) R. Cardman and G. Raithel, Circularizing Rydberg atoms with time-dependent optical traps, Physical Review A 101, 013434 (2020). * (77) M. Kwon, M. F Ebert, T. G Walker, and M Saffman. Parallel low-loss measurement of multiple atomic qubits. Physical review letters, 119(18):180504, 2017. * (78) J. P Covey, I. S Madjarov, A. Cooper, and M. Endres. 2000-times repeated imaging of strontium atoms in clock-magic tweezer arrays. Physical review letters, 122(17):173201, 2019. * (79) Khazali, Mohammadsadegh. ”Rydberg quantum simulator of topological insulators.” arXiv:2101.11412 (2021). * (80) B J Wieder and CL Kane. Spin-orbit semimetals in the layer groups. Physical Review B, 94(15):155108, 2016. * (81) M. Sajid, J K Asboth, D Meschede, R Werner, and A Alberti. Creating floquet chern insulators with magnetic quantum walks. arXiv preprint arXiv:1808.08923, 2018. * (82) 2D DTQW in a lattice of tetramer: $\displaystyle H_{x0}=\sum\limits_{m_{x}=1}^{N_{x}/2}({|{m_{x},e_{x}}\rangle}\\!{\langle{m_{x},o_{x}}|}\otimes\mathbbm{1}_{y}+\text{h.c.})$ $\displaystyle H_{x1}=\sum\limits_{m_{x}=1}^{N_{x}/2}({|{m_{x},e_{x}}\rangle}\\!{\langle{m_{x}+1,o_{x}}|}\otimes\mathbbm{1}_{y}+\text{h.c.})$ $\displaystyle H_{y0}=\sum\limits_{m_{y}=1}^{N_{y}/2}(\mathbbm{1}_{x}\otimes{|{m_{y},e_{y}}\rangle}\\!{\langle{m_{y},o_{y}}|}+\text{h.c.})$ $\displaystyle H_{y1}=\sum\limits_{m_{y}=1}^{N_{y}/2}(\mathbbm{1}_{x}\otimes{|{m_{y},e_{y}}\rangle}\\!{\langle{(m_{y}+1),o_{y}}|}+\text{h.c.})$ * (83) 2D DTQW in a lattice of dimers: $\displaystyle H_{x0}=\sum\limits_{{\bf m}}({|{m_{x},m_{y},m_{z},e}\rangle}\\!{\langle{m_{x},m_{y},m_{z},o}|}+\text{h.c.})$ $\displaystyle H_{x1}=\sum\limits_{{\bf m}}({|{m_{x},m_{y},m_{z},e}\rangle}\\!{\langle{m_{x}+1,m_{y},m_{z},o}|}+\text{h.c.})$ $\displaystyle H_{xy0}=\sum\limits_{{\bf m}}({|{m_{x},m_{y},m_{z},e}\rangle}\\!{\langle{m_{x},m_{y}+1,m_{z},o}|}+\text{h.c.})$ $\displaystyle H_{xy1}=\sum\limits_{{\bf m}}({|{m_{x},m_{y}+1,m_{z},e}\rangle}\\!{\langle{m_{x}+1,m_{y},m_{z},o}|}+\text{h.c.})$ $\displaystyle H_{xz0}=\sum\limits_{{\bf m}}({|{m_{x},m_{y},m_{z}+1,e}\rangle}\\!{\langle{m_{x},m_{y},m_{z},o}|}+\text{h.c.})$ $\displaystyle H_{xz1}=\sum\limits_{{\bf m}}({|{m_{x},m_{y},m_{z},e}\rangle}\\!{\langle{m_{x}+1,m_{y},m_{z}+1,o}|}+\text{h.c.})$ $\displaystyle H_{xyz1}=\sum\limits_{{\bf m}}({|{m_{x},m_{y},m_{z},e}\rangle}\\!{\langle{m_{x}+1,m_{y}+1,m_{z}+1,o}|}$ $\displaystyle\quad\quad\quad+\text{h.c.}).$ * (84) Coined DTQW operators in 3D: $\displaystyle R_{\theta}=\text{e}^{\text{i}\theta\sigma_{x}}=\cos(\theta)\mathbbm{1}_{\\{e,o\\}}+\text{i}\sin(\theta)({|{e}\rangle}\\!{\langle{o}|}+{|{o}\rangle}\\!{\langle{e}|})$ $\displaystyle T_{x}=\sum\limits_{m_{x}}({|{m_{x}-1,e}\rangle}\\!{\langle{m_{x},e}|}+{|{m_{x}+1,o}\rangle}\\!{\langle{m_{x},o}|})\otimes\mathbbm{1}_{y,z}$ $\displaystyle T_{y}=\sum\limits_{m_{y}}({|{m_{y}-1,e}\rangle}\\!{\langle{m_{y},e}|}+{|{m_{y}+1,o}\rangle}\\!{\langle{m_{y},o}|})\otimes\mathbbm{1}_{x,z}$ $\displaystyle T_{z}=\sum\limits_{m_{z}}({|{m_{z}-1,e}\rangle}\\!{\langle{m_{z},e}|}+{|{m_{z}+1,o}\rangle}\\!{\langle{m_{z},o}|})\otimes\mathbbm{1}_{x,y}$ $\displaystyle T_{xy}=\sum\limits_{m_{x},m_{y}}({|{m_{x}-1,m_{y}+1,e}\rangle}\\!{\langle{m_{x},m_{y},e}|}$ $\displaystyle\quad\quad\quad\quad\quad+{|{m_{x}+1,m_{y}-1,o}\rangle}\\!{\langle{m_{x},m_{y},o}|})\otimes\mathbbm{1}_{z}$ $\displaystyle T_{xz}=\sum\limits_{m_{x},m_{z}}({|{m_{x}-1,m_{z}-1,e}\rangle}\\!{\langle{m_{x},m_{z},e}|}$ $\displaystyle\quad\quad\quad\quad\quad+{|{m_{x}+1,m_{z}+1,o}\rangle}\\!{\langle{m_{x},m_{z},o}|})\otimes\mathbbm{1}_{y}$ $\displaystyle T_{xyz}=\sum\limits_{{\bf m}}({|{m_{x}-1,m_{y}-1,m_{z}-1,e}\rangle}\\!{\langle{m_{x},m_{y},m_{z},e}|}$ $\displaystyle\quad\quad\quad+{|{m_{x}+1,m_{y}+1,m_{z}+1,o}\rangle}\\!{\langle{m_{x},m_{y},m_{z},o}|}),$
# A Two-Functional-Network Framework of Opinion Dynamics Wentao Zhang, Zhiqiang Zuo, and Yijing Wang This work was supported by the National Natural Science Foundation of China No. 61933014, No. 61773281, No. 61673292.The authors are with the Tianjin Key Laboratory of Process Measurement and Control, School of Electrical and Information Engineering, Tianjin University, Tianjin, 300072, P. R. China. (e-mail: {wtzhangee, zqzuo, yjwang}@tju.edu.cn). ###### Abstract A common trait involving the opinion dynamics in social networks is an anchor on interacting network to characterize the opinion formation process among participating social actors, such as information flow, cooperative and antagonistic influence, etc. Nevertheless, interacting networks are generally public for social groups, as well as other individuals who may be interested in. This blocks a more precise interpretation of the opinion formation process since social actors always have complex feeling, motivation and behavior, even beliefs that are personally private. In this paper, we formulate a general configuration on describing how individual’s opinion evolves in a distinct fashion. It consists of two functional networks: interacting network and appraisal network. Interacting network inherits the operational properties as DeGroot iterative opinion pooling scheme while appraisal network, forming a belief system, quantifies certain cognitive orientation to interested individuals’ beliefs, over which the adhered attitudes may have the potential to be antagonistic. We explicitly show that cooperative appraisal network always leads to consensus in opinions. Antagonistic appraisal network, however, causes opinion cluster. It is verified that antagonistic appraisal network affords to guarantee consensus by imposing some extra restrictions. They hence bridge a gap between the consensus and the clusters in opinion dynamics. We further attain a gauge on the appraisal network by means of the random convex optimization approach. Moreover, we extend our results to the case of mutually interdependent issues. ###### Index Terms: Social networks, appraisal network, cooperative/antagonistic interaction, random convex optimization. ## I Introduction The study of opinion dynamics in social networks has a long history. Compared with many natural or man-made systems (networks), social actors (agents) in social networks rarely display a common interest. In contrast, they usually aggregate into several small groups where the agents in the same group achieve a unanimous behavior while the opinions of the whole network comprise several clusters. Opinion dynamics in social networks are universal topics and have captured massive interests from different disciplines, such as sociology, social anthropology, economics, ideological political science, physics, biology and control theory, even in the field of military [1, 2]. A simple yet instructive mathematical model is fundamental for the study of opinion dynamics in social networks. As a backbone for opinion dynamics, DeGroot’s iterative opinion pooling configuration shows that social actors can share a common viewpoint if a convex combination mechanism is performed (cf. [3]). In many practical situations, agents often interact with those who are like-minded, and agree on more deviant viewpoints with discretion. For Hegselmann-Krause model, each agent only communicates with those whose opinions are constrained into its confidence interval [4, 5]. Actually, this model is implicitly based on the principle of biased assimilation or homophily. It is always the case that some individuals in social networks have their own prejudices, no matter what kind of opinion formation mechanism is applied. An attempt towards this direction is the Friedkin-Johnsen (FJ) model where some of the individuals (stubbornness) are affected by external signal [6]. Unlike Hegselmann-Krause model, FJ model achieves opinion clusters even in the linear opinion formation process. Conventionally, FJ model mainly focuses on the issue free opinion dynamics or the scalar opinion dynamics. Refs. [7, 8] extended the FJ model to the vector-based opinion dynamics where the opinions tightly relate to several issues, see also [9]. As hinted by belief systems (cf. [10]), topic-specific opinion dynamics are usually entwined whenever agent’s opinion is involved into several interdependent issues. Parsegov _et al_. [11] introduced a row stochastic matrix (MiDS matrix) which clarifies what the attitude of agent towards the issue sequence is. Similarly, the property of belief system dynamics subject to logical constraints was elaborated in [12], and a case study was provided to show how the fluctuations of population’s attitudes evolve. The gossip-based version of the FJ model was proposed in [13]. More recently, an approach building on the DeGroot model and the FJ model was given to investigate the evolutions with respect to the self-appraisal, interpersonal influences and social power over issue sequence for the star, the irreducible and the reducible communication topologies [14]. Nowadays, people have realized that postulate of “cognitive algebra” on heterogeneous information (cf. [15]) for opinion formation process among individuals may not attribute all complex behaviors in social networks due to the negative interpersonal influences that often emerge in many community of dissensus (cf. [16, 17]). Typical instances include but are not limited to multiple-party political systems, biological systems, international alliances, bimodal coalitions, rival business cartels, duopolistic markets, as well as boomerang effect in dyad systems. Based on multi-agent system theory, a lot of research interests have been devoted to the opinion polarization and the stabilization problems. By signed graph theory, Altafini [18] proved that structurally balanced graph is the necessary and sufficient condition for bipartite consensus. Proskurnikov _et al_. [19] further reported the necessary and sufficient criterion to stabilize Altafini’s model. For general communication graph, [20] discussed the interval bipartite consensus problem. The second-order/high-order multi-agent systems in the presence of antagonistic information were also discussed in [21, 22], as well as finite- time consensus [23]. More related work about the signed graph theory refers to [24, 25]. Cooperative control with antagonistic reciprocity was discussed in [26, 27, 28] using the node-based mechanism. Unlike complex systems or multi-agent systems, agents (individuals) in social networks usually display diversity behaviors, such as attitude, feeling and belief, etc. In business negotiation or alliance, for instance, each member tends to collect more “sides information” with respect to the remaining members before organizing a meeting, by which they aim at achieving the profit maximization. They possess introspective ability to access what and how they should do. Notice that the above research is based on a common hypothesis that merely the public network is used to quantify opinion formation process. However, a downside of public networks111As hinted by [1] and [11], we treat the interaction topologies (communication graphs) in the aforementioned literature by the public networks. in the aforementioned literature to describe the antagonistic interactions among participating individuals is that individuals in social networks can completely get access to the opinion, the attitude and the belief of the other individuals towards the interested individuals. But this is not always the case since the individuals in social networks feature complex behaviors, even complex thoughts [29], naturally leading to the complex opinion dynamics that the natural or the man-made complex networks cannot show. Moreover, as discussed in [30, 11], public networks generally characterize the social influence structure, and henceforth they are assumed to be thoroughly known. For this reason, potential hostile information will become transparent whenever only the public network is utilized to model the opinion formation process. Obviously, this is not always true for social networks where it is rather arduous to have access to the individual’s opinion in prior. In this paper, we propose a new framework with functional networks, that is, the appraisal network and the interacting network, to describe the opinion formation process. The opinion evolution in social network is firstly governed by an appraisal network characterizing how each individual assigns its attitude or influence towards other individuals. Afterwards, each individual updates its opinion through an interacting network as the conventional DeGroot model (we call it the interacting network or the public network). To the best of the authors’ knowledge, there are no results available to model the evolution of opinion dynamics using two functionally independent networks. More importantly, we will show that the proposed formulation indeed provides some intriguing phenomena that cannot be preserved by the existing setups. We summarize the main features of this paper as follows: (a) Cooperative appraisal networks lead to consensus in opinion whereas the final aggregated value may not be contained in a convex hull spanned by the initial opinions of the participating individuals. This potentially implies that the way of the decision among individuals in social networks formulated in this paper is not necessarily constrained into the convex hull as the usual models. In fact, non-convex interactions in social networks are common, and it is natural to model the evolution of the opinions by taking this factor into account [29]; (b) Antagonistic appraisal networks result in clusters in opinions. In particular, we show that consensus in opinion dynamics appears provided that the considered antagonistic appraisal networks enjoy certain requirements. It is illustrated that most of the existing results merely guarantee the clusters in opinions subject to hostile interactions (cf. [24, 25]), or stability of the agents [19]; (c) Random convex optimization (cf. [31, 32]) is formulated to provide a feasible estimation on the appraisal network, with the purpose of finite-horizon sense. Also, a bound on the number of “required observations”, which enables us to get rid of a-prior specified level of probabilistic violation, is explicitly given. Therefore we can make a justification of the postulate on self- preservation of the appraisal network for each individual, as opposed to the hypothesis on interacting network in the literature; (d) The proposed setup could be further extended to the multiple mutually entangled issues that are quantified by an issue-dependence matrix, upon which we deduce the criteria associated with the convergence (resp. stability) of the agents. More interestingly, we point out that the introduction of the issue-dependence matrix enables us to steer the leader’s opinion, which has been verified to be fixed all the time in the context of multi-agent systems community. The layout of this paper is outlined as follows: Section II describes some basic preliminaries and the problem formulation as well as the dynamic model. Section III presents the results in the context of the cooperative and the antagonistic appraisal networks. Section IV reports the results for the interdependent issues. Section V gives numerical examples to support the derived theoretical results as well as some discussions. Finally, a conclusion is drawn in Section VI. ## II Preliminaries and Problem Formulation ### II-A Basic Notations The real set, the nonnegative real set and the nonnegative integer set are denoted by $\mathbb{R}$, $\mathbb{R}_{+}$ and $\mathbb{Z}$. Symbol $``\prime"$ denotes the transpose regarding a vector or a matrix. Symbol $|\cdot|$ represents the modulus or the cardinality, and $|\cdot|_{2}$ the $2$-norm. $|Q|$ stands for the spectral norm for matrix $Q$. $\lambda(Q)$ and $Q^{-1}$ denote the eigenvalue and the inverse of an invertible matrix $Q$. $Q\succ 0$ indicates that matrix $Q$ is symmetric, and positive definite. We denote $\\{1,...,N\\}$ and $(1,...,1)^{\prime}$ by $\mathbb{I}_{N}$ and $\textbf{1}_{N}$, where $N$ is the number of agents in social networks. $I$ and $\mathcal{O}$ are, respectively, the identity matrix and the zero matrix. The diagonal matrix is represented by ${\rm diag}(\cdot)$. The sign function is abbreviated by ${\rm sgn}(\cdot)$. For any complex number $\lambda$, $\lambda\triangleq{\rm Re}(\lambda)+\mathbbm{i}{\rm Im}(\lambda)=|\lambda|(\cos(\arg(\lambda))+\mathbbm{i}\sin(\arg(\lambda)))$ where $\mathbbm{i}^{2}=-1$ and $\arg(\lambda)$ represents the value of argument principle. In addition, we denote $(\Omega,\mathcal{F},\mathcal{P})$ the probability space, upon which $\Omega$ represents the sample space, $\mathcal{F}$ the Borel $\sigma$-algebra, and $\mathcal{P}$ the probability measure. An interacting graph $\mathcal{G}$ is commonly represented by a triple $\\{\mathbb{V},\mathbb{E},\mathbb{A}\\}$ where $\mathbb{V}$ is the node set, $\mathbb{E}$ is the edge set and $\mathbb{A}=(a_{ij})_{N\times N}$ is the adjacent matrix with $a_{ij}>0$ provided that $(j,i)\in\mathbb{E}$ and $a_{ij}=0$ otherwise. No self-loop is allowed throughout the paper, i. e., $a_{ii}=0$. The associated Laplacian matrix $\mathcal{L}=(l_{ij})_{N\times N}$ is given by $l_{ii}=\sum^{N}_{j=1,j\neq i}a_{ij}$ and $l_{ij,j\neq i}=-a_{ij,j\neq i}$222In the context of the opinion dynamics, $p_{ij}$ represents the attitude of agent $i$ towards agent $j$, while $l_{ij}$ in the framework of the multi-agent systems means that agent $j$ is a neighbor of agent $i$. Therefore, if no confusion arises, we treat both $p_{ij}$ and $l_{ij}$ as the influence of agent $j$ imposed on agent $i$ with the purpose of the notion consistency.. A digraph is strongly connected if for any two distinct nodes, they are connected by a path. A root of $\mathcal{G}$ is a special node, from which we can arrive at any other nodes. A graph has a directed spanning tree if and only if it contains at least a root. ### II-B Dynamic Model for Social Networks Consider a group of interacting individuals (agents or social actors), whose opinions evolve according to $\displaystyle\xi_{i}(k+1)=$ $\displaystyle~{}\xi_{i}(k)+\varrho_{i}\sum_{j\in\mathcal{N}_{i}}a_{ij}(z_{j}(k)-z_{i}(k))$ (1a) $\displaystyle z_{i}(k)=$ $\displaystyle~{}\sum^{N}_{j=1}\delta_{ij}\xi_{j}(k),~{}i\in\mathbb{I}_{N},~{}k\in\mathbb{Z}$ (1b) where $\xi_{i}(k)\in\mathbb{R}$ stands for the opinion of the $i$th agent at instant $k$, $\mathcal{N}_{i}$ is the neighboring set of agent $i$. Similar to the FJ model, we call parameter $\varrho_{i}$ ($\varrho_{i}\neq 0$) the susceptibility factor with respect to the $i$th social actor. $z_{i}(k)$ in $(\ref{20191eq1b})$ represents the appraisal or the self-appraisal ($z_{i}(k)=\delta_{i}\xi_{i}(k)$) of agent $i$. The constant $\delta_{ij}$ specifies the weighted influence assigned by individual $i$ towards individual $j$, and satisfies $0<\sum^{N}_{j=1}|\delta_{ij}|\leq 1$. The compact form of $(\ref{20191eq1})$ is expressed by $\displaystyle\xi(k+1)=$ $\displaystyle~{}(I-\Lambda\mathcal{L}\mathcal{D})\xi(k),~{}k\in\mathbb{Z}$ (2) where $\xi(k)=(\xi_{1}(k),...,\xi_{N}(k))^{\prime}$, $\Lambda={\rm diag}(\varrho_{1},...,\varrho_{N})$ and $\mathcal{D}=(\delta_{ij})_{N\times N}$. It is easy to see that system $(\ref{20191eq2})$ involves two networks, and we call them the interacting or the public network (quantified by matrix $\mathcal{L}$) and the appraisal network (quantified by matrix $\mathcal{D}$), respectively. We emphasize that the susceptibility factor $\varrho_{i}$ is crucial for the convergence of system $(\ref{20191eq2})$ since the appraisal network and the interacting network are entwined, leading to a significant difference in contrast with the DeGroot model. It is noted that a system might collapse due to the hostility. And as will be shown later, a system subject to antagonistic information fails to converge, regardless of the connection property of the underlying interacting graph333For Ref. [18], the proposed consensus algorithm always assures the convergence of the reciprocal agents as long as the communication topology attains a directed spanning tree, that is, bipartite consensus (interval bipartite consensus) for structurally balanced graph (cf.[18, 20]) and stability for graph that contains the in-isolated structurally balanced subgraphs, or is structurally unbalanced (cf. [19]).. ###### Remark 1 One can easily see that system $(\ref{20191eq2})$ boils down to the classical DeGroot model (cf. [3]) or the multi-agent systems (cf. [33]) when we fix $\mathcal{D}$ to be an identity matrix and $\varrho_{i}$ some positive constant $\varrho$ fulfilling $0<\varrho<\max_{i\in\mathbb{I}_{N}}\\{l_{ii}\\}$ (in such a scenario we always treat $\varrho$ as the step size). It should be pointed out that the two networks described above have nothing to do with the multilayer networks in the context of complex networks [34] where all layer networks functionally inherent with the interacting network $\mathcal{L}$. $\blacklozenge$ We are now in a position to give some interpretations about the reason why we introduce the appraisal network: (1) Unlike the natural and the man-made complex networks where the interacting agents are creatures and smart machines, the individuals in social networks are people. In one word, the subject in social networks possesses emotion, belief as well as attitude towards a specific object. Therefore, it is a common sense that people try to search and collect as much information as possible before they make a decision, such as organizing a conference, exchanging ideas with colleagues, etc. That is to say, people in social networks always evaluate and self- reflect their opinions, beliefs, and behaviors before interacting with others; (2) As mentioned before, the hostile interaction is a key element in the study of opinion dynamics in social networks. It is notable that the interacting networks are generally assumed to be completely known (cf. [30, 11]). Therefore, the private opinions of the participating agents may be leaked if merely the interacting networks are applied to model the opinion evolution in social networks, especially, involving antagonistic interactions. Before moving on, we give some useful definitions regarding system $(\ref{20191eq2})$. ###### Definition 1 System $(\ref{20191eq2})$ achieves: * (i) the _consensus in opinions_ , if $\displaystyle\lim_{k\rightarrow\infty}\xi_{i}(k)=\varphi,~{}~{}i\in\mathbb{I}_{N}$ where $\varphi\in\mathbb{R}$ is a constant. * (ii) the _convergence in opinions_ , if $\displaystyle\lim_{k\rightarrow\infty}\xi_{i}(k)=\varphi_{i},~{}~{}i\in\mathbb{I}_{N}$ where $\varphi_{i}\in\mathbb{R}$ is a constant. * (iii) the _stability in opinions_ , if $\displaystyle\lim_{k\rightarrow\infty}\xi_{i}(k)=0,~{}~{}i\in\mathbb{I}_{N}$ The mechanism behind Definition 1 is that a cooperative appraisal network gives rise to the opinion aggregation while an antagonistic appraisal network leads to the clusters in opinion in general. Moreover, the consensus of the reciprocal agents is a special type of the convergence where all interacting individuals share a common viewpoint eventually. Figure 1: A holistic paradigm of the opinion evolution for system $(\ref{20191eq2})$ where the appraisal network (characterized by $\mathcal{D}$) is cooperative, i.e., $\delta_{ij}\geq 0$. ## III Main Results In the following discussions, we will show that the opinion formation process with setup $(\ref{20191eq1})$ is more general comparing with the DeGroot model. More specific, system $(\ref{20191eq1})$ permits non-convex interactions of the individuals, and it provides an intriguing viewpoint on how the individual’s introspective process influences the opinion evolution of social networks. ### III-A Cooperative Appraisal Network In this subsection, we are concerned with the situation where the appraisal network is cooperative, i.e., each influence weight $\delta_{ij}$ is nonnegative. A simple illustration for this case is drawn in Fig. 1. Hence, $\sum^{N}_{j=1}\delta_{ij}=1$. In such a case, $\mathcal{D}=(\delta_{ij})_{N\times N}$ is a nonnegative stochastic matrix. Before proceeding on, we introduce a simple proposition which bridges the gap between the stochastic matrix and the Laplacian matrix in connection with the communication topology in a unified manner. ###### Proposition 1 ([33]) Given any nonnegative stochastic matrix $\mathcal{D}=(\delta_{ij})_{N\times N}$, there always exists a Laplacian matrix $L=(l_{ij})_{N\times N}$ such that $\displaystyle\mathcal{D}=I-\epsilon L$ (3) where $\epsilon>0$ is the step size. It should be emphasized that the Laplacian matrix $L$ depicted in $(\ref{20191eq37})$ generally has little connection with the interacting network $\mathcal{L}$ reported in $(\ref{20191eq2})$. In the sequel, we give the main results on cooperative appraisal network. ###### Theorem 1 Suppose that the appraisal network is cooperative, i.e., $\delta_{ij}\geq 0$. Then _consensus_ in system $(\ref{20191eq2})$ is achieved if and only if matrix $I-\Lambda\mathcal{L}\mathcal{D}$ has a simple $1$ eigenvalue and the remaining eigenvalues are preserved in the unit disk. ###### Proof: Suppose that the communication graph associated with $\mathcal{L}$ has a directed spanning tree. By Sylvester rank inequality, one has $\displaystyle{\rm rank}(\mathcal{L})+{\rm rank}(\mathcal{D})-N\leq$ $\displaystyle~{}{\rm rank}(\mathcal{L}\mathcal{D})$ $\displaystyle\leq$ $\displaystyle~{}\min\bigg{\\{}{\rm rank}(\mathcal{L}),{\rm rank}(\mathcal{D})\bigg{\\}}$ In fact, the zero eigenvalue is simple if the graph induced by $\mathcal{L}$ has a directed spanning tree. Then, zero is an eigenvalue of matrix $\Lambda\mathcal{L}\mathcal{D}$. Additionally, vector one is the eigenvector of the zero eigenvalue. Hence, system $(\ref{20191eq2})$ can achieve the consensus. We proceed with the condition for consensus. Actually, matrix $I-\Lambda\mathcal{L}\mathcal{D}$ has a simple $1$ eigenvalue and the remaining eigenvalues are preserved in the unit disk if and only if $\Lambda\mathcal{L}\mathcal{D}$ contains a simple zero eigenvalue and the remaining eigenvalues share positive real parts (here we can redefine the coupling matrix by $\Lambda=\epsilon_{1}\Lambda^{{\dagger}}$ where $\epsilon_{1}$ is a small step size, by doing so we could access to the continuous version of $(\ref{20191eq2})$ by $\dot{\xi}=-\Lambda^{{\dagger}}\mathcal{L}\mathcal{D}\xi$). It further assumes that the left and the right eigenvectors with respect to the zero eigenvalue are, respectively, $\varsigma\in\mathbb{R}^{N}$ and $\iota\in\mathbb{R}^{N}$ with the constraint $\varsigma^{\prime}\iota=1$. Note that $\iota=\textbf{1}_{N}$ for such a scenario since $\mathcal{D}\iota=\iota$. We define the disagreement error by $\theta(k)=(I-\iota\varsigma^{\prime})\xi(k)$, then one has $\displaystyle\theta(k+1)=(I-\Lambda\mathcal{L}\mathcal{D})\theta(k)$ (4) We can see that the eigenvalues of $I-\Lambda\mathcal{L}\mathcal{D}$ are entirely constrained in the unit disk over the space $\mathbb{R}^{N}\backslash\\{\varphi\\}$ where $\varphi=\iota\varsigma^{\prime}\xi(0)=\alpha\iota$. Thus the error system $(\ref{20191eq11})$ is exponentially stable, which ensures the consensus of system $(\ref{20191eq2})$. This ends the proof by Definition 1. ∎ ###### Remark 2 Different from the DeGroot model[3] and multi-agent systems[33], the final shared common opinion of the interacting individuals governed by $(\ref{20191eq2})$ may not be restricted in a convex hull spanned by the initial opinions of the individuals. This is because that the appraisal network specifying the interaction rule in this paper is no long to be convex, to a large extent. We emphasize that non-convex interactions among participating individuals in social networks are rather common, which have been clearly pointed out by Wang _et al_. in [29]. This intriguing issue will be verified via an example later. $\blacklozenge$ Next, we consider a special case where the appraisal network is the same as the interacting network, i.e., $L=\mathcal{L}$. In order to achieve the consensus, it suffices to fix $\varrho_{i}=\varrho$ for all $i$. One hence arrives at $\displaystyle\xi(k+1)=$ $\displaystyle~{}(I-\varrho\mathcal{L}+\epsilon\varrho\mathcal{L}^{2})\xi(k)$ (5) where $\epsilon>0$ is a constant. ###### Corollary 1 Suppose that the graph induced by $\mathcal{L}$ has a directed spanning tree. Then system $(\ref{20191eq6})$ achieves the _consensus in opinions_ if and only if $\displaystyle 0<\varrho<\min_{{\rm Re}(\lambda^{\star}_{i})>0}\bigg{\\{}\frac{2{\rm Re}(\lambda^{\star}_{i})}{{\rm Re}^{2}(\lambda^{\star}_{i})+{\rm Im}^{2}(\lambda^{\star}_{i})}\bigg{\\}}$ (6) where $\lambda^{\star}_{i}$ stands for the $i$th eigenvalue of $\mathcal{L}-\epsilon\mathcal{L}^{2}$, and constant $\epsilon$ satisfies $\left\\{\begin{aligned} &\epsilon\in\mathbb{R},~{}~{}|{\rm Re}(\lambda_{i})|=|{\rm Im}(\lambda_{i})|\\\ &\epsilon<\min_{\lambda_{i}\neq 0}\frac{{\rm Re}(\lambda_{i})}{{\rm Re}^{2}(\lambda_{i})-{\rm Im}^{2}(\lambda_{i})},~{}~{}|{\rm Re}(\lambda_{i})|>|{\rm Im}(\lambda_{i})|\\\ &\epsilon>\max_{\lambda_{i}\neq 0}-\frac{{\rm Re}(\lambda_{i})}{{\rm Im}^{2}(\lambda_{i})-{\rm Re}^{2}(\lambda_{i})},~{}~{}|{\rm Re}(\lambda_{i})|<|{\rm Im}(\lambda_{i})|\end{aligned}\right.$ (7) where $\lambda_{i}$ is the $i$th eigenvalue of matrix $\mathcal{L}$. Moreover, if $\epsilon$ is further required to be positive, then we have $\displaystyle\epsilon\in\bigg{(}0,~{}~{}\min_{\lambda_{i}\neq 0,~{}|{\rm Re}(\lambda_{i})|>|{\rm Im}(\lambda_{i})|}\frac{{\rm Re}(\lambda_{i})}{{\rm Re}^{2}(\lambda_{i})-{\rm Im}^{2}(\lambda_{i})}\bigg{)}$ Moreover, the final aggregated value is restricted in a convex hull spanned by the initial opinions of the roots. ###### Proof: Let us first study an auxiliary system of $(\ref{20191eq6})$ by $\displaystyle\dot{\xi}(t)=$ $\displaystyle~{}-W\xi(t),~{}t\in\mathbb{R}_{+}$ (8) where $W=\mathcal{L}-\epsilon\mathcal{L}^{2}$. It is known that system $(\ref{20191eq7})$ guarantees the consensus if and only if matrix $W$ contains a simple zero eigenvalue and the remaining eigenvalues share positive real parts. When the induced graph by Laplacian matrix $\mathcal{L}$ has a directed spanning tree, one has that the zero eigenvalue of matrix $W$ is simple. We continuous to show that the nonzero eigenvalues of $W$ have positive real parts. One can easily find that the nonzero eigenvalues of $W$ are of the form $\displaystyle\lambda^{\star}_{i}=$ $\displaystyle~{}{\rm Re}(\lambda_{i})-\epsilon({\rm Re}^{2}(\lambda_{i})-{\rm Im}^{2}(\lambda_{i}))$ (9) $\displaystyle+\mathbbm{i}({\rm Im}(\lambda_{i})-2\epsilon{\rm Re}(\lambda_{i}){\rm Im}(\lambda_{i}))$ By $(\ref{20191eq8})$, the nonzero eigenvalues in $W$ share positive real parts if and only if $(\ref{ch7_20191eq10_1})$ is desirable. It is noticeable that the relationship between the eigenvalues of system matrix in $(\ref{20191eq7})$ and those in $(\ref{20191eq6})$ can be formulated by $\displaystyle\lambda^{\ast}_{i}=$ $\displaystyle~{}1-\varrho\lambda^{\star}_{i},~{}i\in\mathbb{I}_{N}$ where $\lambda^{\ast}_{i}$ represents the $i$th eigenvalue of $I-\varrho\mathcal{L}+\epsilon\varrho\mathcal{L}^{2}$. As a result, one can check that $I-\varrho\mathcal{L}+\epsilon\varrho\mathcal{L}^{2}$ has a simple $1$ eigenvalue and the remaining eigenvalues are restricted in the unit disk if and only if $(\ref{20191eq10})$ is desirable, by which $|\lambda^{\ast}_{i}|<1$ is always guaranteed as long as $\lambda^{\star}_{i}\neq 0$. The second part is trivial, and hence is omitted. ∎ An extension of $\epsilon=\varrho$ would be interesting since the developed method in Corollary 1 does not work in such a case. Actually, $\varrho$ is involved in both $(\ref{20191eq6})$ and $(\ref{20191eq7})$, leading to the failure of computing the allowable range for $\varrho$. Apart from these concerns, the argument induced within a general interacting network is far from obvious in contrast to the case of bidirectional interacting network, as suggested by Corollary 1. ###### Theorem 2 Suppose that the graph induced by $\mathcal{L}$ has a directed spanning tree. Then system $(\ref{20191eq6})$ achieves the _consensus in opinions_ if and only if the step size $\varrho$ is bounded with the following constraints, * (i) If ${\rm Im}(\lambda_{i})=0$ for $\lambda_{i}\neq 0$, $\displaystyle\varrho\in\min_{\lambda_{i}}\bigg{(}0,\frac{1}{\lambda_{i}}\bigg{)}$ where $\lambda_{i}>0$ is the eigenvalue of $\mathcal{L}$. * (ii) If ${\rm Im}(\lambda_{i})\neq 0$ with $\lambda_{i}\neq 0$, $\left\\{\begin{aligned} &\min_{\varrho>0,\lambda_{i}\neq 0}f_{i}(\varrho,\lambda_{i},\arg(\lambda_{i}))>0\\\ &\varrho\not\in\bigg{\\{}\varrho_{i,1},\varrho_{i,2}\bigg{\\}}\bigcup\bigg{\\{}\varrho_{i,3},\varrho_{i,4}\bigg{\\}}\end{aligned}\right.$ where $\displaystyle f_{i}(\varrho,\lambda_{i},\arg(\lambda_{i}))$ $\displaystyle=$ $\displaystyle~{}-\varrho^{3}|\lambda_{i}|^{3}+\varrho^{2}|\lambda_{i}|^{2}\cos^{2}(\arg(\lambda_{i}))$ $\displaystyle-2\varrho|\lambda_{i}|\sin(2\arg(\lambda_{i}))-\varrho|\lambda_{i}|+2\cos(\arg(\lambda_{i}))$ and $\varrho_{i,1}$, $\varrho_{i,2}$, $\varrho_{i,3}$, $\varrho_{i,4}$ are depicted in $(\ref{20191eq107})$ $\left\\{\begin{aligned} \varrho_{i,1}=&~{}\frac{\cos(\arg(\lambda_{i}))+\sqrt{\cos^{2}(\arg(\lambda_{i}))(8\cos(\theta)-7)+4(1-\cos(\theta))}}{2|\lambda_{i}|\cos(2\arg(\lambda_{i}))}\\\ \varrho_{i,2}=&~{}\frac{\cos(\arg(\lambda_{i}))-\sqrt{\cos^{2}(\arg(\lambda_{i}))(8\cos(\theta)-7)+4(1-\cos(\theta))}}{2|\lambda_{i}|\cos(2\arg(\lambda_{i}))}\\\ \varrho_{i,3}=&~{}\frac{\sin(\arg(\lambda_{i}))+\sqrt{\cos^{2}(\arg(\lambda_{i}))(8\cos(\theta)-7)+4(1-\cos(\theta))}}{2|\lambda_{i}|\sin(2\arg(\lambda_{i}))}\\\ \varrho_{i,4}=&~{}\frac{\sin(\arg(\lambda_{i}))-\sqrt{\cos^{2}(\arg(\lambda_{i}))(8\cos(\theta)-7)+4(1-\cos(\theta))}}{2|\lambda_{i}|\sin(2\arg(\lambda_{i}))}\end{aligned}\right.$ (10) where $\theta\in[0,2\pi)$. ###### Proof: The proof of Theorem 2 is self-contained, and is reported in Appendix for the sake of concinnity. ∎ From Theorem 2, the requirement on $\varrho$ coincides with the statement in [33] if the underlying network is bidirectional. For the general interpersonal network, the condition on $\varrho$ is far from trivial as the previous case since it tightly links to the amplitude and the argument principal value of nonzero eigenvalues corresponding to $\mathcal{L}$. In fact, there is an alternative to guarantee the consensus in opinion for cooperative antagonistic network, even though it falls short of elegance as Theorem 2. ###### Corollary 2 Suppose that the graph induced by $\mathcal{L}$ attains a directed spanning tree. Then system $(\ref{20191eq6})$ achieves the _consensus in opinions_ if and only if $\displaystyle\varrho\in\bigcap_{\varrho>0,\lambda_{i}\neq 0}\bigg{\\{}a_{i}\varrho^{3}+b_{i}\varrho^{2}+c_{i}\varrho+d_{i}<0\bigg{\\}}$ (11) where $\displaystyle a_{i}=$ $\displaystyle~{}4{\rm Re}^{2}(\lambda_{i})+({\rm Re}^{2}(\lambda_{i})-{\rm Im}^{2}(\lambda_{i}))^{2}$ $\displaystyle b_{i}=$ $\displaystyle~{}2{\rm Re}(\lambda_{i})({\rm Re}^{2}(\lambda_{i})-{\rm Im}^{2}(\lambda_{i}))-4{\rm Re}(\lambda_{i}){\rm Im}(\lambda_{i})$ $\displaystyle c_{i}=$ $\displaystyle~{}3{\rm Re}^{2}(\lambda_{i})-{\rm Im}^{2}(\lambda_{i})$ $\displaystyle d_{i}=$ $\displaystyle~{}-2{\rm Re}(\lambda_{i})$ In addition, consensus in system $(\ref{20191eq6})$ is achieved for the undirected graph if and only if $\displaystyle\varrho\in\min_{\lambda_{i}}\bigg{(}0,\frac{1}{\lambda_{i}}\bigg{)}$ where $\lambda_{i}>0$ is the eigenvalue of $\mathcal{L}$. ###### Proof: System $(\ref{20191eq6})$ can be rewritten as $\displaystyle\xi(k+1)=$ $\displaystyle~{}(I-\varrho\mathcal{L}+\varrho^{2}\mathcal{L}^{2})\xi(k)$ (12) The eigenvalue of matrix $I-\varrho\mathcal{L}+\varrho^{2}\mathcal{L}^{2}$ is of the form $\displaystyle\lambda^{\ast}_{i}=$ $\displaystyle~{}1-\varrho\lambda_{i}+\varrho^{2}\lambda^{2}_{i}$ $\displaystyle=$ $\displaystyle~{}1-{\rm Re}(\lambda_{i})\varrho+\bigg{(}{\rm Re}^{2}(\lambda_{i})-{\rm Im}^{2}(\lambda_{i})\bigg{)}\varrho^{2}$ $\displaystyle+\bigg{(}2{\rm Re}(\lambda_{i}){\rm Im}(\lambda_{i})\varrho^{2}-{\rm Im}(\lambda_{i})\varrho\bigg{)}\mathbbm{i}$ Therefore, $|\lambda^{\ast}_{i}|<1$ with ${\rm Re}(\lambda_{i})>0$ is equal to $\displaystyle 1\geq$ $\displaystyle~{}\bigg{(}1-{\rm Re}(\lambda_{i})\varrho+\bigg{(}{\rm Re}^{2}(\lambda_{i})-{\rm Im}^{2}(\lambda_{i})\bigg{)}\varrho^{2}\bigg{)}^{2}$ $\displaystyle+\bigg{(}2{\rm Re}(\lambda_{i}){\rm Im}(\lambda_{i})\varrho^{2}-{\rm Im}(\lambda_{i})\varrho\bigg{)}^{2}$ By tedious computation, it yields the requirement in $(\ref{20191eq63})$. And the second statement is consistent with that in Theorem 2, and is hence omitted. ∎ Figure 2: A holistic paradigm of the opinion evolution for system $(\ref{20191eq2})$ where the appraisal network (characterized by $\mathcal{D}$) is antagonistic, that is, the red arcs mean $\delta_{ij}<0$ while the black ones indicate that $\delta_{ij}>0$. ### III-B Antagonistic Appraisal Network A remarkable feature of the social networks is that social actors rarely share unanimous opinions. It has been recognized that biased assimilation principle or homophily principle (cf. [35]) is vital to explain the opinion clusters. Another reason for the opinion dynamics giving rise to clusters is antagonism. The main goal of this subsection is to investigate the case where there exists antagonistic interactions among social actors. By following the basic route as that in subsection III-A, the public network characterized by $\mathcal{L}$ is the same as before. Meanwhile, the appraisal network characterized by $\mathcal{D}$ involves hostile interactions. A simple illustration for system $(\ref{20191eq2})$ is depicted by Fig. 2. Suppose that agent $i$ possesses an opposite attitude to the opinion of agent $j$, then ${\rm sgn}(\delta_{ij})=-1$. In addition, since $|\delta_{ij}|$ quantifies the degree of the influence among the total social influence that the $i$th social actor has, we require $\sum^{N}_{j=1}|\delta_{ij}|=1$ for all $i$. Before moving on, a condiment is needed. ###### Definition 2 Two graphs $\mathcal{G}_{1}$ and $\mathcal{G}_{2}$ share the same topology if there exists an edge from the $i$th node to the $j$th node in $\mathcal{G}_{1}$, then the edge $(i,j)$ also belongs to $\mathcal{G}_{2}$, and vice versa. Actually, two graphs satisfying Definition 2 differ only in the weighted values of edges. Therefore, we can define a new stochastic matrix $\mathcal{D}^{\star}\triangleq(|\delta_{ij}|)_{N\times N}$. With the help of Definition 2, the graphs induced by $\mathcal{D}$ and $\mathcal{D}^{\star}$ have a topology structure in common. ###### Theorem 3 Suppose that the appraisal network is antagonistic. Then system $(\ref{20191eq2})$ _converges in opinions_ if and only if matrix $I-\Lambda\mathcal{L}\mathcal{D}$ has a simple $1$ eigenvalue and the remaining eigenvalues are preserved in the unit disk. ###### Proof: The proof of Theorem 3 is similar to that in Theorem 1. Generally speaking, the right eigenvector $\iota$ associated with the zero eigenvalue of $\Lambda\mathcal{L}\mathcal{D}$ may not be equal to vector one when the appraisal network contains hostile interactions. Therefore, it follows that $\varphi=\iota\varsigma^{\prime}\xi(0)$. This implies the convergence in opinions of the social actors. ∎ According to Theorem 3, the clusters in opinions occur due to the presence of antagonistic interactions in the appraisal network. In view of Theorems 1 and 3, it can be concluded that cooperative appraisal networks lead to the consensus in opinions, while hostile appraisal networks attain the clusters in opinions. Here we emphasize that it is rather arduous to give some specific indices on the weighted influence matrix $\Lambda$ due to the potential complexity of the appraisal network. Another interesting welfare behind Theorems 1 and 3 is that the agents in system $(\ref{20191eq2})$ will not converge to zero for almost all initial values. This is because that $I-\Lambda\mathcal{L}\mathcal{D}$ contains at least an eigenvalue one. Fortunately, this intriguing feature could be addressed by virtue of the issue dependence matrix that is about to be elucidated later. Figure 3: (a) Opinion evolution for the antagonistic appraisal network where the initial opinions $\xi(0)=(25,75,85)$ are partly borrowed from [11, Equation (15)]; (b) Opinion evolution for the antagonistic appraisal network with the same parameters as (a) except for replacing $\Lambda$ by $0.5I$. ### III-C Connection between Cooperative and Antagonistic Appraisal Networks Subsections III-A and III-B discuss the opinion evolution on cooperative and antagonistic appraisal networks, respectively. Here we are interested in a question: what is the bridge between the cooperative and the antagonistic appraisal networks? To answer this question, we first look at a simple example. ###### Example 1 Consider a network with three social actors whose interacting and appraisal networks are of the forms $\displaystyle\mathcal{L}=\begin{bmatrix}2&-1&-1\\\ -1&2&-1\\\ -1&-1&2\end{bmatrix},~{}\mathcal{D}=\begin{bmatrix}0.5&-0.5&0\\\ 0&0.5&-0.5\\\ -0.5&0&0.5\end{bmatrix}$ We choose $\Lambda={\rm diag}(-0.05,0.5,0.5)$, and then compute the eigenvalues of $I-\Lambda\mathcal{L}\mathcal{D}$ $\displaystyle\lambda(I-\Lambda\mathcal{L}\mathcal{D})\in\bigg{\\{}0,0.9448,1.9052\bigg{\\}}$ Opinion evolutions of the social actors are plotted in Fig. 3(a). It reveals an interesting phenomenon that antagonistic appraisal networks could achieve consensus in opinions. More importantly, we can see that the final aggregated opinion is not restricted in the convex hull spanned by the initial opinions. Next, we select another group of susceptibility factors by $0.5I$. In this case, the consensus in opinions could still be preserved while it is contained in the convex hull spanned by the initial opinions (see Fig. 3(b) for more details). Therefore, we could treat the susceptibility factor $\varrho_{i}$ as a design parameter quantifying how does the opinion formation process work. $\blacklozenge$ Note that consensus in opinions could be preserved for cooperative appraisal network. It implicitly manifests that the opinions either achieve consensus or diverge whenever the appraisal network is cooperative. However, opinions either achieve consensus or clusters if the underlying appraisal network is antagonistic, except for the divergence in opinions. We summarize them as below. ###### Proposition 2 Suppose that the appraisal network is antagonistic. Then system $(\ref{20191eq2})$ achieves _consensus in opinions_ if and only if the following properties hold: * (i) System $(\ref{20191eq2})$ converges; * (ii) $\mathcal{D}\textbf{1}_{N}=\mathcal{O}_{N\times 1}$ or $\mathcal{D}\textbf{1}_{N}=-\textbf{1}_{N}$. ###### Proof: (Necessity) According to Definition 1, the consensus in $(\ref{20191eq2})$ indicates that $\lim_{k\rightarrow\infty}\xi_{i}(k)=\lim_{k\rightarrow\infty}\xi_{j}(k)$ for arbitrary $i,j\in\mathbb{I}_{N}$. Hence, $\textbf{1}_{N}$ is the right eigenvector associated with the eigenvalue one of matrix $I-\Lambda\mathcal{L}\mathcal{D}$. It further arrives at $\displaystyle\mathcal{O}_{N\times 1}=$ $\displaystyle\Lambda\mathcal{L}\mathcal{D}\textbf{1}_{N}\Rightarrow\mathcal{O}_{N\times 1}=\mathcal{L}\mathcal{D}\textbf{1}_{N}$ The above formula is true since matrix $\Lambda$ is nonsingular. The directed spanning tree preserved in the interacting network implies that merely one of the following equalities is satisfied, $\displaystyle\mathcal{D}\textbf{1}_{N}=$ $\displaystyle~{}\textbf{1}_{N}$ (13a) $\displaystyle\mathcal{D}\textbf{1}_{N}=$ $\displaystyle~{}\mathcal{O}_{N\times 1}$ (13b) $\displaystyle\mathcal{D}\textbf{1}_{N}=$ $\displaystyle~{}-\textbf{1}_{N}$ (13c) Next we will show that only $(\ref{20191eq48b})$ or $(\ref{20191eq48c})$ hold for the antagonistic appraisal network. Suppose that $(\ref{20191eq48a})$ holds, i.e., $\displaystyle\sum^{N}_{j=1}\delta_{ij}=1,~{}\forall~{}i\in\mathbb{I}_{N}$ (14) Moreover, we also require $\displaystyle\sum^{N}_{j=1}|\delta_{ij}|=1,~{}\forall~{}i\in\mathbb{I}_{N}$ (15) One can see that $(\ref{20191eq49})$ and $(\ref{20191eq50})$ suggest $\displaystyle|\delta_{ij}|=\delta_{ij},~{}\forall~{}i,j\in\mathbb{I}_{N}$ which implies that the appraisal network is cooperative. It is a contradiction. Therefore, we always have $\mathcal{D}\textbf{1}_{N}=\mathcal{O}_{N\times 1}$ or $\mathcal{D}\textbf{1}_{N}=-\textbf{1}_{N}$. Obviously, $\mathcal{D}\textbf{1}_{N}=-\textbf{1}_{N}$ means that all entries of $\mathcal{D}$ are non-positive, which indicates that the appraisal network is antagonistic. (Sufficiency) The convergence of system $(\ref{20191eq2})$ means that $I-\Lambda\mathcal{L}\mathcal{D}$ has a simple eigenvalue one (note that system $(\ref{20191eq2})$ cannot guarantee the stability since matrix $\Lambda\mathcal{L}\mathcal{D}$ always has at least a zero eigenvalue). In addition, $\mathcal{D}\textbf{1}_{N}=\mathcal{O}_{N1}$ or $\mathcal{D}\textbf{1}_{N}=-\textbf{1}_{N}$ suggests that the vector $\textbf{1}_{N}$ is an eigenvector associated with the eigenvalue one of matrix $I-\Lambda\mathcal{L}\mathcal{D}$. Therefore, system $(\ref{20191eq2})$ preserves consensus in opinions. ∎ With the help of Proposition 2 and Theorem 3, we could conclude that cooperative appraisal network achieves consensus in opinions, while antagonistic appraisal network shows both consensus and clusters in opinions. However, antagonistic appraisal network that can achieve consensus in opinions has certain special requirements for its structure. ### III-D Estimation for Appraisal Network Briefly speaking, it is a thorny problem to determine the structure of the appraisal network. Fortunately, a large number of efforts have been poured on identifying the dynamic structure and topologies of the networks with the help of the experiment data. The emerging fields include sociology, signal processing and statistics (see, e.g., [1, 36, 37]). As hinted before, interacting network is always known to all; while appraisal network is generally private, and thus is of great significance to be estimated. The reason lies that it is the first step to understand the mechanism behind the emerging and the evolution of the opinions in social networks. To cope with the estimation issue related to appraisal network, a technical lemma is needed. ###### Lemma 1 ([38]) For any matrix $Q=(q_{1},...,q_{n})$, its vectorization, denoted by ${\rm vec}(Q)$, is $\displaystyle{\rm vec}(Q)=(q^{\prime}_{1},...,q^{\prime}_{n})^{\prime}$ For matrices $Q$, $W$ and $R$, the vectorization with respect to their product is given by $\displaystyle{\rm vec}(QWR)=$ $\displaystyle(R^{\prime}\otimes Q){\rm vec}(W)$ By virtue of Lemma 1, two properties can be obtained. ###### Corollary 3 For any variable $\xi\in\mathbb{R}^{N}$ and matrix $\Lambda\mathcal{L}\mathcal{D}$, the following facts are true $\left\\{\begin{aligned} &{\rm vec}(\xi)=~{}\xi\\\ &{\rm vec}(\Lambda\mathcal{L}\mathcal{D}\xi)=~{}(\xi^{\prime}\otimes\Lambda\mathcal{L}){\rm vec}(\mathcal{D})\end{aligned}\right.$ ###### Proof: The proof is trivial by Lemma 1. ∎ With the above preparations, we are about to formulate the estimation problem on appraisal network. Apart from $\Lambda$ and $\mathcal{L}$, we postulate that one has access to $m$ observed opinions of length $1$ during opinion formation process under $(\ref{20191eq1})$, i.e., $m$ sequences of opinions $(\xi_{t}(k-1),\xi_{t}(k))$ with $t\in\mathbb{I}_{m}$ and $k\in\mathbb{Z}_{+}$, in a manner of independent and identically distribution (i.i.d.). More specifically, we associate a uniform observation of $m$ by $\displaystyle\Omega_{m}=~{}\bigg{\\{}(\xi_{t}(k-1),\xi_{t}(k)),1\leq t\leq m,\forall~{}k\in\mathbb{Z}_{+}\bigg{\\}}$ (16) Upon collecting the $m$ (randomly with uniform) opinion sets and using Corollary 3, we formulate the following optimization problem: $\displaystyle~{}~{}~{}~{}~{}~{}~{}~{}~{}~{}~{}~{}~{}~{}~{}~{}~{}\min_{\zeta,~{}Q}~{}\gamma$ (17) $\displaystyle{\rm s.t.}~{}~{}~{}~{}~{}~{}f(\zeta,m)=\frac{1}{m}\sum^{m}_{t=1}g(t)\leq\gamma$ $\displaystyle g(t)=X^{\prime}_{t}(k)QX_{t}(k)$ $\displaystyle X_{t}(k)=\xi_{t}(k)-\xi_{t}(k-1)+\bigg{(}\xi^{\prime}_{t}(k-1)\otimes\Lambda\mathcal{L}\bigg{)}\zeta$ $\displaystyle\gamma\geq 0,~{}Q\succ 0$ ###### Remark 3 For optimization problem $(\ref{20191eq31})$, we emphasize: (i) The utilization of Corollary 3 is helpful since the coupling between the interacting network and the appraisal network is entangled; (ii) In general, it is unrealistic to require parameter $m$ to be infinite because system matrix in $(\ref{20191eq2})$ is not stable. Algorithm 1 depicted below gives a lower bound on the “observations” needed for a desirable performance. Algorithm 1 Lower Bound for the Observations in $\Omega_{m}$ 1: Given a sequence of observations of length $m$, and a prior level of performance index $0<\gamma_{0}<\infty$; 2: Set $m=m_{0}$ with $m_{0}\geq 1$; 3: while $\gamma>\gamma_{0}$; do 4: Print: Gauge of appraisal network fails; 5: Set $m\leftarrow 1+m$; 6: Optimize $(\ref{20191eq31})$; 7: end while 8: return $m$; Although Algorithm 1 may provide an estimation on appraisal network, we still cannot give specific value of sample length $m$ that is essential for solving $(\ref{20191eq31})$. Obviously, too much observations bring computation complexity, while less observation may yield inaccuracy estimation. Furthermore, even Algorithm 1 may mitigate the downside on $(\ref{20191eq31})$ to a certain degree, a lot of computing resources are needed apart from the time consumption since Algorithm 1 executes in a way of try-once-discard. All the discussions above naturally raise an interesting question: What is the confidence can we possess with the aid of a finite sets on opinions of empirical observations, on which the appraisal network can work for the whole opinion space? To this end, we resort to the tool from random convex optimization (also called by chance-constrained optimization); see Ref. [31, 32] for more details. Now, we are going to formulate the considered problem. For fixed probability space $(\Omega,\mathcal{F},\mathcal{P})$, compact set $\mathbb{X}\subseteq\mathbb{R}^{N\times N}$ is convex and contains the origin as its interior. A measurable function $f:\mathbb{X}\times\Omega\mapsto\mathbb{R}$ is convex on its first argument for any fixed second argument, and bounded for the second argument whenever the first is fixed. Thereby, for random finite opinion set $\Omega_{m}$ in $(\ref{20191eqnew1})$ and any level parameter $\varepsilon\in(0,1)$, given a confidence level $\beta\in(0,1)$ and some symmetric positive definite matrix $Q$, the probability of violation (cf. [31, Definition 1]) is $\displaystyle V(\zeta,m)=\mathcal{P}\\{m\in\mathbb{Z}_{+}:f(\zeta,m)>\gamma^{\star}\\}$ where $\gamma^{\star}$ is the optimal solution of the optimization problem $(\ref{20191eq31})$. We then focus on the following optimization problem: $\displaystyle\min_{\zeta}{}{}{}{}{}{}$ $\displaystyle c^{\prime}\zeta$ (18) $\displaystyle{\rm s.t.}{}{}{}{}{}{}$ $\displaystyle\mathcal{P}\\{(\zeta,m):V(\zeta,m)>\varepsilon\\}\leq\beta$ $\displaystyle\zeta\in\mathbb{X}$ where $c$ is a certain “cost” vector (also known as the objective direction). For optimization problem $(\ref{20191eqnew2})$, we get the probability with respect to $f(\zeta,m)\leq\gamma^{\star}$ at least $1-\beta$ using the samples with length $m$. A smaller violation level $\varepsilon$ is a more desirable estimation on the appraisal network. As a result, the number of samples increases. The next theorem gives a confirmative answer on how many opinion observations are needed to obtain a satisfactory performance. ###### Theorem 4 Consider optimization problem $(\ref{20191eqnew2})$ with $m\geq N^{2}$ where $\Omega_{m}\subseteq\mathbb{X}$, in the sense of i.i.d. Then, for $\forall\varepsilon\in[0,1]$, it follows that $\displaystyle\mathcal{P}\\{(\zeta,m):V(\zeta,m)>\varepsilon\\}\leq\beta(\varepsilon,m)$ where $\displaystyle\beta(\varepsilon,m)=\sum^{m}_{\ell=0}\begin{pmatrix}N^{2}\\\ \ell\end{pmatrix}\varepsilon^{\ell}(1-\varepsilon)^{m-\ell}$ Moreover, the low bound on $m$ is $\displaystyle m(\varepsilon,\beta)=\min\bigg{\\{}m\in\mathbb{Z}_{+}\bigg{|}\sum^{m}_{\ell=0}\begin{pmatrix}N^{2}\\\ \ell\end{pmatrix}\varepsilon^{\ell}(1-\varepsilon)^{m-\ell}\leq\beta\bigg{\\}}$ ###### Proof: By [32, Theorem 3.3] and [39, Theorem 1], the proof follows directly. ∎ ###### Remark 4 Theorem 4 explicitly provides the number of required samples to obtain a “good” estimation for the true appraisal network. In [11], the authors concentrated on the estimation of multi-issue dependence matrix. We do some further work in this paper. That is, on the one hand, we can give a desirable estimation on the appraisal network. On the other hand, a bound on the samples is provided which has no been addressed in the previous work. ## IV Topic Specific Opinion Unlike man-made systems, the actors in social networks are generally affected by a couple of topics that are interdependent. For example, the leader of the corporate should account for many factors if he/she intends to operate a policy, such as the cost, the external and the internal environment, the potential market as well as the potential customers, etc. As a matter of fact, the issue dependence problem has been well addressed for a long time, in particular the disciplines such as social anthropology, sociology and political science and psychology where they share a common ground that certain objects are coupled by interdependent cognitive orientations. The first step on how agents’ interpersonal influences form a belief system was proposed by the FJ model (cf. [6]) $\displaystyle\xi_{i}(k+1)=$ $\displaystyle\lambda_{ii}C\sum^{N}_{j=1}p_{ij}\xi_{j}(k)+(1-\lambda_{ii})\mu_{i}$ (19) where $\xi_{i}(k)\in\mathbb{R}^{n}$, $\lambda_{ii}$ and $\mu_{i}$ are, respectively, the susceptibility and the initial opinion of the $i$th agent. $P=(p_{ij})_{N\times N}$ with $p_{ij}\geq 0$ is a stochastic matrix. Matrix $C=(c_{ij})_{n\times n}$ stands for the introspective transformation called the multi-issues dependence structure (MiDS) (cf. [11]), and satisfies $\sum^{n}_{j=1}|c_{ij}|=1$. For topic specific issues, the opinion evolving equation $(\ref{20191eq1a})$ becomes $\displaystyle\xi_{i}(k+1)=$ $\displaystyle~{}C\xi_{i}(k)+\varrho_{i}C\sum_{j\in\mathcal{N}_{i}}a_{ij}(z_{j}(k)-z_{i}(k))$ (20) where the constant matrix $C\in\mathbb{R}^{n\times n}$ describes the MiDS, while the remaining variables are the same as those in $(\ref{20191eq1a})$. Combining $(\ref{20191eq1b})$ and $(\ref{20191eq53})$ gives $\displaystyle\xi(k+1)=$ $\displaystyle~{}(I-\Lambda\mathcal{L}\mathcal{D})\otimes C\xi(k)$ (21) where $\otimes$ denotes the Kronecker product. Here we discuss the similarities and the differences between $(\ref{20191eq52})$ and $(\ref{20191eq54})$. Similarity: They both model how the interdependent issues affect the opinion’s evolution. Differences: $(\ref{20191eq52})$ is concerned with the opinion evolution in the context of the cooperative interacting networks while $(\ref{20191eq54})$ studies the opinion evolution with antagonistic interactions characterized by an appraisal network. We emphasize that the interacting network in $(\ref{20191eq54})$ merely quantifies whether or not there exists an information flow between a pair of agents. Moreover, we can further extend $(\ref{20191eq54})$ to the case where some agents may never forget their initial opinions. However, this is beyond the scope of this paper, and is thus omitted. Note that $(\ref{20191eq2})$ usually ensures the convergence in opinions. However, agents in $(\ref{20191eq54})$ may converge to zero due to the appearance of matrix $C$. ###### Theorem 5 Suppose that matrix $I-\Lambda\mathcal{L}\mathcal{D}$ has a simple eigenvalue one. The agents in $(\ref{20191eq54})$ achieve the _stability_ if and only if $|\lambda_{\max}(C)|<1$. ###### Proof: It is more vulnerable to check that the system is stable if and only if the eigenvalues in matrix $(I-\Lambda\mathcal{L}\mathcal{D})\otimes C$ are constrained in the unit disk. Note that the eigenvalues of matrix $(I-\Lambda\mathcal{L}\mathcal{D})\otimes C$ are $\lambda(I-\Lambda\mathcal{L}\mathcal{D})\lambda(C)$ according to the matrix theory. Hence, $(\ref{20191eq54})$ is stable if and only if $|\lambda_{\max}(I-\Lambda\mathcal{L}\mathcal{D})\lambda_{\max}(C)|<1$. This is equivalent to $|\lambda_{\max}(C)|<1$. The proof hence follows. ∎ Theorem 5 suggests that we can achieve the stability of the agents by just imposing the restriction on matrix $C$ provided that the issue-free cases are convergent. Note that the number of issues are drastically less than that of the participating individuals in general. An interesting welfare from Theorem 5 in contrast to Theorems 1 and 3 is the stability of the interacting agents, which is another motivation to introduce the issue-interdependence matrix $C$ for setup $(\ref{20191eq1})$. Indeed, only the convergence of the opinions is generally assured for the issue-free scenario (see Theorems 1 and 3 for more information). As discussed before, the social networks rarely achieve unanimous behavior (we treat the stability of the agents by a special case of the consensus). Therefore, it is of great importance to further study the convergence condition on $(\ref{20191eq54})$ in the presence of matrix $C$. ###### Theorem 6 Suppose that matrix $I-\Lambda\mathcal{L}\mathcal{D}$ has a simple eigenvalue one. The agents in $(\ref{20191eq54})$ are _convergent_ if and only if $|\lambda^{\star}_{\max}(I-\Lambda\mathcal{L}\mathcal{D})\lambda_{\max}(C)|<1$ where $\lambda^{\star}_{\max}(I-\Lambda\mathcal{L}\mathcal{D})$ stands for the eigenvalue of $I-\Lambda\mathcal{L}\mathcal{D}$ with the second largest magnitude compared with eigenvalue $1$. ###### Proof: Here we prove the theorem by following the idea from $(\ref{20191eq11})$ and $(\ref{20191eq54})$. The error system is revised as $\displaystyle\theta(k+1)=(I-\Lambda\mathcal{L}\mathcal{D})\otimes C\theta(k)$ (22) where $\theta(k)\in\bigg{(}\mathbb{R}^{N}\backslash\\{\varphi\\}\bigg{)}\otimes\mathbb{R}^{n}$. As $I-\Lambda\mathcal{L}\mathcal{D}$ has a simple eigenvalue one over the space $\theta(k)\in\bigg{(}\mathbb{R}^{N}\backslash\\{\varphi\\}\bigg{)}\otimes\mathbb{R}^{n}$, the eigenvalues in matrix $I-\Lambda\mathcal{L}\mathcal{D}$ are completely contained in the unit disk. Therefore, we can see that the error system $(\ref{20191eq55})$ is stable if and only if $|\lambda^{\star}_{\max}(I-\Lambda\mathcal{L}\mathcal{D})\lambda_{\max}(C)|<1$. The proof hence follows. ∎ Based on the conclusion in Theorem 6, it does not require $|\lambda_{\max}(C)|<1$ as Theorem 5. Hence, there is a matrix $C$ with $|\lambda_{\max}(C)|>1$ that we could still guarantee the convergence of the participating agents as long as $|\lambda^{\star}_{\max}(I-\Lambda\mathcal{L}\mathcal{D})\lambda_{\max}(C)|<1$ is desirable. The following corollary gives some specific requirements on the matrix for guaranteeing the convergence of the agents. ###### Corollary 4 Suppose that matrix $I-\Lambda\mathcal{L}\mathcal{D}$ has a simple eigenvalue one. The agents in $(\ref{20191eq54})$ are _convergent_ only if $\lim_{k\rightarrow\infty}C^{k}$ exists. ###### Proof: One can see that $\displaystyle\xi(k)=$ $\displaystyle~{}(I-\Lambda\mathcal{L}\mathcal{D})^{k}\otimes C^{k}\xi(0)$ (23) Therefore, $(\ref{20191eq56})$ converges only if $\lim_{k\rightarrow\infty}C^{k}$ exists. This ends the proof. ∎ Due to the antagonistic information, matrix $I-\Lambda\mathcal{L}\mathcal{D}$ may not be a nonnegative stochastic matrix in general. Consequently, the developed methods in the framework of the multi-agent systems are no longer applicable. By virtue of the approach developed in Subsection III-D, it is possible to estimate matrix $C$ if $\Lambda$ and $\mathcal{D}$ are available. Figure 4: Interacting graph of four social actors associated with $(\ref{20191eq44})$. Figure 5: Opinion evolution of four social actors according to the DeGroot model where the initial opinions are randomly selected from $[-10,10]$. Figure 6: (a)-(b) Opinion evolution with the issues independence under the cooperative appraisal network; (c)-(d) Opinion evolution with the issues dependence under the cooperative appraisal network. Figure 7: (a)-(b) Opinion evolution with the issues independence under the antagonistic appraisal network; (c)-(d) Opinion evolution with the issues dependence under the antagonistic appraisal network. Figure 8: (a)-(b) Stability of the agents with the cooperative appraisal network depicted in Section V-A; (c)-(d) Stability of the agents with the antagonistic appraisal network depicted in Section V-B. ## V Numerical Example Consider a social network with four individuals. The interacting matrix ($P$ corresponds to the DeGroot model) is $\displaystyle P=\begin{bmatrix}0.22&0.12&0.36&0.3\\\ 0.147&0.215&0.344&0.294\\\ 0&0&1&0\\\ 0.09&0.178&0.446&0.286\end{bmatrix}$ (24) It should be pointed out that the elements of matrix $P$ in $(\ref{20191eq40})$ come from real experiment (see [6]). More specific, each entry of $P$ denotes the inter-personal influence by individual $i$ to individual $j$. By Proposition 1, the Laplacian matrix is $\displaystyle\mathcal{L}=\begin{bmatrix}0.78&-0.12&-0.36&-0.3\\\ -0.147&0.785&-0.344&-0.294\\\ 0&0&0&0\\\ -0.09&-0.178&-0.446&0.714\end{bmatrix}$ (25) with $\epsilon=1$, and the associated interacting graph is presented in Fig. 4. The opinion evolution with respect to four social actors are drawn in Fig. 5. ### V-A Cooperative Appraisal Network In the sequel, we examine the opinion evolution of the interacting agents with the cooperative appraisal network $\displaystyle\mathcal{D}_{1}=\begin{bmatrix}0.2&0.2&0.3&0.3\\\ 0.1&0.5&0&0.4\\\ 0.1&0.4&0&0.5\\\ 0.4&0.3&0.2&0.1\end{bmatrix}$ (26) The opinion evolution of the individuals for the issue independence ($C_{1}=I$) is plotted in Fig. 6(a)-(b) with the susceptibility factor matrix $\Lambda_{1}={\rm diag}(-1,1,1,-1)$, where the initial opinions are borrowed from [11, Equation (15)], $\displaystyle\xi(0)\in\bigg{\\{}\overbrace{25,25}^{\xi_{1}(0)},\overbrace{25,15}^{\xi_{2}(0)},\overbrace{75,-50}^{\xi_{3}(0)},\overbrace{85,5}^{\xi_{4}(0)}\bigg{\\}}$ From Fig. 6(a)-(b), the opinions of agents aggregate to the opinion of the leader’s, i.e., the $3$rd agent. In what follows, we discuss the opinion evolution of the interacting agents with the issue interdependence influence. Matrix $C_{1}$ has the form as that in [11, Section VII], $\displaystyle C_{1}=\begin{bmatrix}0.9&0.1\\\ 0.1&0.9\end{bmatrix}$ With other parameters unchanged, the opinion evolution of agents is depicted in Fig. 6(c)-(d). Analogous to the issues independence case, the opinions achieve the consensus. However, some interesting phenomena arise: (1) Unlike the issue free case, the leader’s opinion varies over time. Traditionally, the leader’s opinion remains the same even if the JF model (including the issue interdependence version [11]) is applied; (2) By introducing the issue dependence matrix $C_{1}$, the final aggregated opinion may be steered to an opposite direction of the leader’s initial opinion. In one word, cooperative appraisal network leads to consensus in opinions, which coincides with the conclusion drawn in Theorem 1. ### V-B Antagonistic Appraisal Network This subsection focuses on the opinion evolution with antagonistic appraisal network, which is of the form $\displaystyle\mathcal{D}_{2}=\begin{bmatrix}0.2&-0.2&-0.3&-0.3\\\ 0.1&0.5&0&0.4\\\ -0.1&0.4&0&0.5\\\ 0.4&0.3&-0.2&0.1\end{bmatrix}$ Similar to Subsection V-A, we first consider the issue free case. The agents’ opinion evolutions are shown in Fig. 7(a)-(b) where the parameters are the same except for $\mathcal{D}_{2}$ and $\Lambda_{2}={\rm diag}(-1.5,2,1,-0.5)$. Using the method in [11], the agents’ final opinions share the same direction with the leader’s, see [11, Fig. 5(a)]. However, the second agent’s opinion has an opposite sign with the leader’s, even if their initial opinions have the same direction (see Fig. 7(a)-(b)). We proceed with the issue interdependence case with $\displaystyle C_{2}=\begin{bmatrix}0.6&0.4\\\ 0.3&0.7\end{bmatrix}$ (27) Using the same parameters as before, the opinions of the agents are depicted in Fig. 7(c)-(d). It is clear that the leader’s opinion changes along with the time-evolution, as opposed to the issue free case. In addition, we can see that although some agents have the same direction on the initial opinion of the leader’s at the beginning, the agents’ final opinions appear an opposite direction with the leader’s. Furthermore, the opinions tend to clusters in such a case, which is in accordance with Theorem 3. One can check that $C_{2}$ in $(\ref{20191eq59})$) does not satisfy the conditions in Theorem 5. Indeed, we compute the eigenvalues of matrix $C_{2}$, i.e., $\lambda(C_{2})\in\\{1,0.3\\}$. They, however, meet the requirement in Theorem 6. ### V-C Stability of Interacting Agents Now we are dedicated to studying the stability of agents with cooperative and antagonistic appraisal networks. For the cooperative appraisal network, we perform the simulation using the parameters in Subsection V-A by replacing matrix $C_{1}$ with $C^{\star}_{1}$ $\displaystyle C^{\star}_{1}=0.85\begin{bmatrix}0.9&0.1\\\ 0.1&0.9\end{bmatrix}$ It can be verified that the requirement in Theorem 5 is fulfilled. The opinions of the agents are depicted in Fig. 8(a)-(b). For the case of antagonistic appraisal network, we use the same parameters as Subsection V-B by replacing matrix $C_{2}$ with $C^{\star}_{2}$ $\displaystyle C^{\star}_{2}=0.95\begin{bmatrix}0.6&0.4\\\ 0.3&0.7\end{bmatrix}$ The opinion evolution of the agents can be found in Fig. 8(c)-(d). ### V-D Further Discussion With the framework of $(\ref{20191eq1})$, we can achieve both consensus and clusters in opinions, and the consensus in opinions is not necessarily restricted into a convex hull as the classical DeGroot model (see Fig. 3(a)). Moreover, formulation $(\ref{20191eq1})$ extends the protocol of [27] in several aspects: (i) it could be utilized to describe more general behaviors in social networks; (ii) it further manifests the importance on the weighted gain matrix $\Lambda$. In other words, system $(\ref{20191eq2})$ may diverge without the help of $\Lambda$. In [11, Equation (15)], Parsegov _et al_. endowed both the initial opinions and the final opinions of the social actors with some specific meanings: the positive (resp. the negative) opinions correspond to the vegetarian (resp. the all-meat) diets by introducing an interdependent issue matrix. According to [11, Examples $3$ and $4$], all social actors are the vegetarian which is coincide with their initial opinions, $\xi^{1}(0)=(25,25,75,85)$ (cf. [11, Equation (15)]), i.e., the initial opinions on the first issue. However, we can see from Fig. 6, some agents become the all-meat diets eventually, even if they are the vegetarian at the beginning. This also indicates that some agents have the opposite attitudes compared with the leader’s ever though they all have the same direction of attitudes at first. Additionally, from Figs. 6(c)-(d) and 7(c)-(d), the leader’s final opinion may be affected by the evolutions of other agents, which is new from the perspective of the classical DeGroot model. In a nutshell, the proposed setup in this paper brings some interesting phenomena comparing with the existing literature. ## VI Conclusion This paper has studied the opinion dynamics in social networks by introducing an appraisal network to quantify the cooperative or antagonistic information. We have shown that the cooperative appraisal network achieves the consensus in opinions while the antagonistic appraisal network leads to the opinion clusters. The tool of random convex optimization is used to estimate the appraisal network with a confident level of robustness, along with the lower bound on the amounts of sampled observations. Moreover, the proposed setup has been extended to the case of multiple issues interdependence. Some discussions have also been given to compare with the existing literature. [Proof of Theorem 2] To prove Theorem 2, the Hermite-Biehler Theorem (cf. [40]) and the Bilinear Transformation Theorem (cf. [41]) are needed. To begin with, we first introduce a lemma. ###### Lemma 2 ([41]) Given two polynomials $\mathbb{S}(z)$ of degree $d$ and $\mathbb{Q}(z)$ with $\displaystyle\mathbb{Q}(z)=~{}(z-1)^{d}\mathbb{S}\bigg{(}\frac{z+1}{z-1}\bigg{)}$ Then the Schur stability of $\mathbb{S}(z)$ implies the Hurwitz stability on $\mathbb{Q}(z)$, and vice versa. For complex polynomial $\mathbb{Q}(z)$, replacing $z$ with $\mathbbm{i}w$ yields $\displaystyle\mathbb{Q}(\mathbbm{i}w)=~{}S(w)+\mathbbm{i}Q(w)$ where both $S(w)$ and $Q(w)$ are real polynomials. A relationship among the roots of $S(w)$ and $Q(w)$ is depicted as follows. ###### Definition 3 ([40]) For any real polynomials $S(w)$ and $Q(w)$, they are interlaced if * (i) The roots of $S(w)$ (denoted by $\\{S_{1},...,s_{\ell_{S}}\\}$) and $Q(w)$ (denoted by $\\{Q_{1},...,Q_{\ell_{Q}}\\}$) satisfy $\left\\{\begin{aligned} &S_{1}<S_{2}<\cdots<S_{\ell_{S}}\\\ &Q_{1}<Q_{2}<\cdots<Q_{\ell_{Q}}\end{aligned}\right.$ * (ii) $|\ell_{S}-\ell_{Q}|\leq 1$, and one of the following facts holds $\left\\{\begin{aligned} &S_{1}<Q_{1}<S_{2}<Q_{2}\cdots<Q_{\ell_{Q}}<S_{\ell_{S}},~{}\ell_{S}=\ell_{Q}+1\\\ &Q_{1}<S_{1}<Q_{2}<S_{2}\cdots<S_{\ell_{S}}<Q_{\ell_{Q}},~{}\ell_{Q}=\ell_{S}+1\\\ &Q_{1}<S_{1}<Q_{2}<S_{2}\cdots<Q_{\ell_{Q}}<S_{\ell_{S}},~{}\ell_{Q}=\ell_{S}\\\ &S_{1}<Q_{1}<S_{2}<Q_{2}\cdots<S_{\ell_{S}}<Q_{\ell_{Q}},~{}\ell_{S}=\ell_{Q}\\\ \end{aligned}\right.$ The next lemma develops a judgement on the Hurwitz stability associated with $\mathbb{Q}(z)$ that is closely linked to the roots of polynomials $S(w)$ and $Q(w)$ and the interlaced property defined in Definition 3. ###### Lemma 3 ([40]) Polynomial $\mathbb{Q}(z)$ is Hurwitz stable if and only if * (i) Polynomials $S(w)$ and $Q(w)$ are interlaced; * (ii) Polynomials $S(w)$, $\frac{\partial Q(w)}{\partial w}$, $\frac{\partial S(w)}{\partial w}$ and $Q(w)$ at the origin fulfill $\displaystyle S(0)\frac{\partial Q(0)}{\partial w}-\frac{\partial S(0)}{\partial w}Q(0)>0$ With the above preparations, we are about to give the proof of Theorem 2. ###### Proof: It is easy to see that the updating dynamics of the individuals is determined by $(\ref{20191eq64})$. Accordingly, the characteristic equation of $(\ref{20191eq64})$ can be written as $\displaystyle{\rm det}\bigg{(}zI-(I-\varrho\mathcal{L}+\varrho^{2}\mathcal{L}^{2})\bigg{)}$ $\displaystyle=$ $\displaystyle~{}(z-1)\prod^{N}_{i=2}(z-1+\varrho\lambda_{i}-\varrho^{2}\lambda^{2}_{i})$ Before proceeding further, it is indispensable to consider two cases where ${\rm Im}(\lambda_{i})$ is equal to zero or not. And we start with the scenario that ${\rm Im}(\lambda_{i})=0$. As argued before, consensus in opinions indicates that $|z|<1$ if $z\neq 1$. In such a setting, it is enough to demonstrate that $\displaystyle\varrho^{2}\lambda^{2}_{i}-\varrho\lambda_{i}+1<1$ (28a) $\displaystyle\varrho^{2}\lambda^{2}_{i}-\varrho\lambda_{i}+1>-1$ (28b) with the constraints $\varrho>0$ and $\lambda_{i}\neq 0$. By computation, one derives that the feasible region on $\varrho$ for inequality $(\ref{20191eq79a})$ is $(0,\frac{1}{\lambda_{i}})$, while the feasible region on $\varrho$ associated with inequality $(\ref{20191eq79b})$ is $[0,\infty)$. Thus, the feasible region for $\varrho$ is $\displaystyle\varrho\in\bigg{(}0,\frac{1}{\lambda_{i}}\bigg{)}$ In the sequel, we are dedicated to the case of ${\rm Im}(\lambda_{i})\neq 0$. Let $\mathbb{S}_{i}(z)$ be of the form $\displaystyle\mathbb{S}_{i}(z)=~{}z-1+\varrho\lambda_{i}-\varrho^{2}\lambda^{2}_{i},~{}i=2,...,N$ (29) where $\lambda_{i}$ stands for the $i$th eigenvalue of $\mathcal{L}$. Applying the bilinear transformation $\displaystyle\mathbb{Q}_{i}(z)=~{}(z-1)\mathbb{S}_{i}\bigg{(}\frac{z+1}{z-1}\bigg{)}$ (30) One gets $\displaystyle\mathbb{Q}_{i}(z)=$ $\displaystyle~{}2+\varrho\lambda_{i}(1-\varrho\lambda_{i})z+\varrho\lambda_{i}(\varrho\lambda_{i}-1)$ $\displaystyle=$ $\displaystyle~{}2+\varrho|\lambda_{i}|\bigg{(}\cos(\arg(\lambda_{i}))+\mathbbm{i}\sin(\arg(\lambda_{i}))\bigg{)}z$ $\displaystyle-\varrho|\lambda_{i}|\bigg{(}\cos(\arg(\lambda_{i}))+\mathbbm{i}\sin(\arg(\lambda_{i}))\bigg{)}$ $\displaystyle-\varrho^{2}|\lambda_{i}|^{2}\bigg{(}\cos^{2}(\arg(\lambda_{i}))-\sin^{2}(\arg(\lambda_{i}))$ $\displaystyle+2\mathbbm{i}\sin(\arg(\lambda_{i}))\cos(\arg(\lambda_{i}))\bigg{)}z$ $\displaystyle-\varrho^{2}|\lambda_{i}|^{2}\bigg{(}\cos^{2}(\arg(\lambda_{i}))-\sin^{2}(\arg(\lambda_{i}))$ $\displaystyle+2\mathbbm{i}\sin(\arg(\lambda_{i}))\cos(\arg(\lambda_{i}))\bigg{)}$ Substituting $z=\mathbbm{i}w$ into $\mathbb{Q}_{i}(z)$ results in $\displaystyle\mathbb{Q}_{i}(\mathbbm{i}w)=$ $\displaystyle~{}~{}2+\varrho|\lambda_{i}|\bigg{(}\mathbbm{i}\cos(\arg(\lambda_{i}))-\sin(\arg(\lambda_{i}))\bigg{)}w$ (31) $\displaystyle-\varrho|\lambda_{i}|\bigg{(}\cos(\arg(\lambda_{i}))+\mathbbm{i}\sin(\arg(\lambda_{i}))\bigg{)}$ $\displaystyle-\varrho^{2}|\lambda_{i}|^{2}\bigg{(}\mathbbm{i}\cos^{2}(\arg(\lambda_{i}))-\mathbbm{i}\sin^{2}(\arg(\lambda_{i}))$ $\displaystyle-2\sin(\arg(\lambda_{i}))\cos(\arg(\lambda_{i}))\bigg{)}w$ $\displaystyle-\varrho^{2}|\lambda_{i}|^{2}\bigg{(}\cos^{2}(\arg(\lambda_{i}))-\sin^{2}(\arg(\lambda_{i}))$ $\displaystyle+2\mathbbm{i}\sin(\arg(\lambda_{i}))\cos(\arg(\lambda_{i}))\bigg{)}$ $\displaystyle\triangleq$ $\displaystyle~{}{\rm Re}(\mathbb{Q}_{i}(w))+\mathbbm{i}{\rm Im}(\mathbb{Q}_{i}(w))$ With the constraint ${\rm Re}(\mathbb{Q}_{i}(w))=0$, it yields $\displaystyle w_{1}$ $\displaystyle=$ $\displaystyle~{}\frac{2+\varrho^{2}|\lambda_{i}|^{2}(2\cos^{2}(\arg(\lambda_{i}))-1)-\varrho|\lambda_{i}|\cos(\arg(\lambda_{i}))}{\varrho|\lambda_{i}|\sin(\arg(\lambda_{i}))-2\varrho^{2}|\lambda_{i}|^{2}\sin(\arg(\lambda_{i}))\cos(\arg(\lambda_{i}))}$ $\displaystyle\triangleq$ $\displaystyle~{}\frac{2+\mathbbm{x}}{\mathbbm{y}}$ Analogously, the requirement of ${\rm Im}(\mathbb{Q}_{i}(w))=0$ immediately leads to $\displaystyle w_{2}$ $\displaystyle=$ $\displaystyle~{}\frac{2\varrho^{2}|\lambda_{i}|^{2}\sin(\arg(\lambda_{i}))\cos(\arg(\lambda_{i}))-\varrho|\lambda_{i}|\sin(\arg(\lambda_{i}))}{\varrho^{2}|\lambda_{i}|^{2}(2\cos^{2}(\arg(\lambda_{i}))-1)-\varrho|\lambda_{i}|\cos(\arg(\lambda_{i}))}$ In accordance with $w_{1}$, $w_{2}$ hence is of the form $\displaystyle w_{2}=~{}\frac{-\mathbbm{y}}{\mathbbm{x}}$ We now determine the condition guaranteeing $\displaystyle w_{1}\frac{1}{w_{2}}\neq 1$ For ease of discussion, we resort to the opposite statement, i.e., $\displaystyle 1=$ $\displaystyle~{}w_{1}\frac{1}{w_{2}}$ (32) $\displaystyle=$ $\displaystyle~{}\frac{2\mathbbm{x}+\mathbbm{x}^{2}}{-\mathbbm{y}^{2}}$ We can see that $(\ref{20191eq84})$ is equivalent to $\displaystyle 1=~{}(\mathbbm{x}+1)^{2}+\mathbbm{y}^{2}$ By the polar coordination, for $\theta\in[0,2\pi)$, it attains $\displaystyle\mathbbm{x}=$ $\displaystyle~{}\cos(\theta)-1$ (33a) $\displaystyle\mathbbm{y}=$ $\displaystyle~{}\sin(\theta)$ (33b) For $(\ref{20191eq85a})$, it is more prone to access that $\displaystyle\mathbbm{x}-\cos(\theta)+1=$ $\displaystyle~{}\varrho^{2}|\lambda_{i}|^{2}\cos(2\arg(\lambda_{i}))$ $\displaystyle-\varrho|\lambda_{i}|\cos(\arg(\lambda_{i}))-\cos(\theta)+1$ $\displaystyle\triangleq$ $\displaystyle~{}f_{\theta}(\varrho)$ For $\forall~{}\theta\in[0,2\pi)$, denote $\displaystyle\Delta_{\mathbbm{x}}\triangleq$ $\displaystyle~{}|\lambda_{i}|^{2}+4|\lambda_{i}|^{2}\cos(2\arg(\lambda_{i}))(\cos(\theta)-1)$ $\displaystyle=$ $\displaystyle~{}|\lambda_{i}|^{2}\bigg{(}\cos^{2}(\arg(\lambda_{i}))(8\cos(\theta)-7)+4(1-\cos(\theta))\bigg{)}$ In the sequel, two cases involving the real roots of $f_{\theta}(\varrho)=0$ are formulated. Case 1) $\theta\in[-\arccos(\frac{7}{8}),\arccos(\frac{7}{8})]$, $\Delta_{\mathbbm{x}}\geq 0$ follows directly. Case 2) $\theta\in(-\frac{\pi}{2},-\arccos(\frac{7}{8}))\bigcup(\arccos(\frac{7}{8}),\frac{\pi}{2})$. In such a circumstance, one has $\Delta_{\mathbbm{x}}\geq 0$ if and only if $\arg(\lambda_{i})$ is preserved in the set $\displaystyle\bigg{[}\arccos(\sqrt{\frac{4(1-\cos(\theta))}{7-8\cos(\theta)}},\frac{\pi}{2})\bigg{)}$ $\displaystyle\bigcup$ $\displaystyle\bigg{(}-\frac{\pi}{2},-\arccos(\sqrt{\frac{4(1-\cos(\theta))}{7-8\cos(\theta)}})\bigg{]}$ The two real roots of $f_{\theta}(\varrho)=0$ are given by $\left\\{\begin{aligned} \varrho_{i,1}=&~{}\frac{|\lambda_{i}|\cos(\arg(\lambda_{i}))+\sqrt{\Delta_{\mathbbm{x}}}}{2|\lambda_{i}|^{2}\cos(2\arg(\lambda_{i}))}\\\ \varrho_{i,2}=&~{}\frac{|\lambda_{i}|\cos(\arg(\lambda_{i}))-\sqrt{\Delta_{\mathbbm{x}}}}{2|\lambda_{i}|^{2}\cos(2\arg(\lambda_{i}))}\\\ \end{aligned}\right.$ Moreover, we calculate $\displaystyle|\lambda_{i}|^{2}\cos^{2}(\arg(\lambda_{i}))-\Delta_{\mathbbm{x}}=4|\lambda_{i}|^{2}(1-\cos(\theta))\cos(2\arg(\lambda_{i}))$ Therefore, the sign of the calculation value is entirely contingent on $\cos(2\arg(\lambda_{i}))$, which in turn indicates that both $\varrho_{1}$ and $\varrho_{2}$ are positive. For $(\ref{20191eq85b})$, we find $\displaystyle\sin(\theta)-\mathbbm{y}=$ $\displaystyle~{}2\varrho^{2}|\lambda_{i}|^{2}\sin(2\arg(\lambda_{i}))-\varrho|\lambda_{i}|\sin(\arg(\lambda_{i})$ $\displaystyle+\sin(\theta)$ $\displaystyle\triangleq$ $\displaystyle~{}g_{\theta}(\varrho),~{}\theta\in[0,2\pi)$ For $\forall~{}\theta\in[0,2\pi)$, denote $\displaystyle\Delta_{\mathbbm{y}}\triangleq$ $\displaystyle~{}|\lambda_{i}|^{2}\bigg{(}\sin^{2}(\arg(\lambda_{i}))-4\sin(2\arg(\lambda_{i}))\sin(\theta)\bigg{)}$ Evidently, for $\theta\in[0,2\pi)$, $\Delta_{\mathbbm{y}}\geq 0$ implies $\displaystyle\sin^{2}(\arg(\lambda_{i}))\geq 8\sin(\arg(\lambda_{i}))\cos(\arg(\lambda_{i}))\sin(\theta)$ (34) In a similar manner, two scenarios should be argued. Case i) $\arg(\lambda_{i})\in(0,\frac{\pi}{2})$. In this case, $(\ref{20191eq91})$ is desirable provided that $\displaystyle\arg(\lambda_{i})\in\bigg{[}\max\\{0,\arctan(8\sin(\theta))\\},\frac{\pi}{2}\bigg{)}\backslash\\{0\\},~{}\theta\in[0,2\pi)$ Case ii) $\arg(\lambda_{i})\in(-\frac{\pi}{2},0)$. It is trivial that $\displaystyle 0\leq\Delta_{\mathbbm{y}},~{}\forall~{}\theta\in[0,\pi]$ And for $\theta\in[-\pi,0]$, $(\ref{20191eq91})$ could be further expressed by $\displaystyle\sin^{2}(\arg(\lambda_{i}))\geq 8|\sin(\arg(\lambda_{i}))|\cos(\arg(\lambda_{i}))|\sin(\theta)|$ which is true as long as $\displaystyle\arg(\lambda_{i})\in\bigg{[}-\frac{\pi}{2},\min\\{0,-\arctan(8\sin(\theta))\\}\bigg{)}\backslash\\{0\\}$ With the foregoing arguments, $g_{\theta}(\varrho)=0$ has two real roots, which can be described as $\left\\{\begin{aligned} \varrho_{i,3}=&~{}\frac{|\lambda_{i}|\sin(\arg(\lambda_{i}))+\sqrt{\Delta_{\mathbbm{y}}}}{2|\lambda_{i}|^{2}\sin(2\arg(\lambda_{i}))}\\\ \varrho_{i,4}=&~{}\frac{|\lambda_{i}|\sin(\arg(\lambda_{i}))-\sqrt{\Delta_{\mathbbm{y}}}}{2|\lambda_{i}|^{2}\sin(2\arg(\lambda_{i}))}\\\ \end{aligned}\right.$ Therefore, $w_{1}$ and $w_{2}$ are interlaced (cf. Definition 3) with the constraints on $\varrho$, i.e., $\displaystyle\varrho\not\in\\{\varrho_{i,1},\varrho_{i,2}\\}\bigcup\\{\varrho_{i,3},\varrho_{i,4}\\}$ By virtue of the specifications on ${\rm Re}(\mathbb{Q}_{i}(w))$ and ${\rm Im}(\mathbb{Q}_{i}(w))$ (cf. $(\ref{20191eq78})$), one gets $\left\\{\begin{aligned} \frac{\partial{\rm Re}(\mathbb{Q}_{i}(0))}{\partial w}=&~{}\varrho^{2}|\lambda_{i}|^{2}\sin(2\arg(\lambda_{i}))-\varrho|\lambda_{i}|\sin(\arg(\lambda_{i}))\\\ {\rm Re}(\mathbb{Q}_{i}(0))=&~{}2+\varrho^{2}|\lambda_{i}|^{2}(2\cos^{2}(\arg(\lambda_{i}))-1)\\\ &-\varrho|\lambda_{i}|\cos(\arg(\lambda_{i}))\\\ \frac{\partial{\rm Im}(\mathbb{Q}_{i}(0))}{\partial w}=&~{}\varrho^{2}|\lambda_{i}|^{2}(1-2\cos^{2}(\arg(\lambda_{i})))\\\ &+\varrho|\lambda_{i}|\cos(\arg(\lambda_{i}))\\\ {\rm Im}(\mathbb{Q}_{i}(0))=&~{}\varrho^{2}|\lambda_{i}|^{2}\sin(2\arg(\lambda_{i}))-\varrho|\lambda_{i}|\sin(\arg(\lambda_{i}))\end{aligned}\right.$ We perform the calculation $\displaystyle{\rm Re}(\mathbb{Q}_{i}(0))\frac{\partial{\rm Im}(\mathbb{Q}_{i}(0))}{\partial w}-\frac{\partial{\rm Re}(\mathbb{Q}_{i}(0))}{\partial w}{\rm Re}(\mathbb{Q}_{i}(0))$ $\displaystyle=$ $\displaystyle~{}\varrho|\lambda_{i}|\bigg{(}-\varrho^{3}|\lambda_{i}|^{3}+\varrho^{2}|\lambda_{i}|^{2}\cos^{2}(\arg(\lambda_{i}))$ $\displaystyle-2\varrho|\lambda_{i}|\sin(2\arg(\lambda_{i}))-\varrho|\lambda_{i}|+2\cos(\arg(\lambda_{i}))\bigg{)}$ $\displaystyle\triangleq$ $\displaystyle~{}\varrho|\lambda_{i}|f_{i}(\varrho,\lambda_{i},\arg(\lambda_{i})),~{}i=2,...,N$ Obviously, ${\rm Re}(\mathbb{Q}_{i}(0))\frac{\partial{\rm Im}(\mathbb{Q}_{i}(0))}{\partial w}-\frac{\partial{\rm Re}(\mathbb{Q}_{i}(0))}{\partial w}{\rm Re}(\mathbb{Q}_{i}(0))$ is positive if and only if $f_{i}(\varrho,\lambda_{i},\arg(\lambda_{i}))>0$. Consequently, by Lemma 3, we can see that $\mathbb{Q}_{i}(z)$ in $(\ref{20191eq76})$ is Hurwitz stable. With the aid of Lemma 2, one could state that $\mathbb{S}_{i}(z)$ in $(\ref{20191eq75})$ is Schur stable. This ends the proof. ∎ ## References * [1] N. E. Friedkin and E. C. Johnsen, _Social Influence Network Theory: A Sociological Examination of Small Group Dynamics_. Cambridge University Press, Cambridge, UK, 2011. * [2] F. Bullo, _Lectures on Network Systems_ , 1st ed. Kindle Direct Publishing, 2019, with contributions by J. Cortés, F. Dörfler, and S. Martínez. [Online]. Available: http://motion.me.ucsb.edu/book-lns * [3] M. H. DeGroot, “Reaching a consensus,” _Journal of the American Statistical Association_ , vol. 69, no. 345, pp. 118–121, 1974. * [4] R. Hegselmann, U. Krause _et al._ , “Opinion dynamics and bounded confidence models, analysis, and simulation,” _Journal of Artificial Societies and Social Simulation_ , vol. 5, no. 3, pp. 1–24, 2002. * [5] V. D. Blondel, J. M. Hendrickx, and J. N. Tsitsiklis, “On krause’s multi-agent consensus model with state-dependent connectivity,” _IEEE Transactions on Automatic Control_ , vol. 54, no. 11, pp. 2586–2597, 2009. * [6] N. E. Friedkin and E. C. Johnsen, “Social influence networks and opinion change,” _Advances in Group Processes_ , vol. 16, pp. 1–29, 1999. * [7] N. E. Friedkin, “The problem of social control and coordination of complex systems in sociology: A look at the community cleavage problem,” _IEEE Control Systems Magazine_ , vol. 35, no. 3, pp. 40–51, 2015. * [8] S. Fortunato, V. Latora, A. Pluchino, and A. Rapisarda, “Vector opinion dynamics in a bounded confidence consensus model,” _International Journal of Modern Physics C_ , vol. 16, no. 10, pp. 1535–1551, 2005. * [9] N. E. Friedkin and F. Bullo, “How truth wins in opinion dynamics along issue sequences,” _Proceedings of the National Academy of Sciences_ , vol. 114, no. 43, pp. 11380–11385, 2017. * [10] P. E. Converse, “The nature of belief systems in mass publics (1964),” _Critical Review_ , vol. 18, no. 1-3, pp. 1–74, 2006. * [11] S. E. Parsegov, A. V. Proskurnikov, R. Tempo, and N. E. Friedkin, “Novel multidimensional models of opinion dynamics in social networks,” _IEEE Transactions on Automatic Control_ , vol. 62, no. 5, pp. 2270–2285, 2017. * [12] N. E. Friedkin, A. V. Proskurnikov, R. Tempo, and S. E. Parsegov, “Network science on belief system dynamics under logic constraints,” _Science_ , vol. 354, no. 6310, pp. 321–326, 2016. * [13] C. Ravazzi, P. Frasca, R. Tempo, and H. Ishii, “Ergodic randomized algorithms and dynamics over networks,” _IEEE Transactions on Control of Network Systems_ , vol. 2, no. 1, pp. 78–87, 2015. * [14] P. Jia, A. MirTabatabaei, N. E. Friedkin, and F. Bullo, “Opinion dynamics and the evolution of social power in influence networks,” _SIAM Review_ , vol. 57, no. 3, pp. 367–397, 2015. * [15] N. H. Anderson, _Foundations of Information Integration Theory_. New York: Academic Press, 1981. * [16] R. E. Petty and J. T. Cacioppo, _Communication and Persuasion: Central and Peripheral Routes to Attitude Change_. Springer Science & Business Media, 2012. * [17] A. Flache and M. W. Macy, “Small worlds and cultural polarization,” _The Journal of Mathematical Sociology_ , vol. 35, no. 1-3, pp. 146–176, 2011. * [18] C. Altafini, “Consensus problems on networks with antagonistic interactions,” _IEEE Transactions on Automatic Control_ , vol. 58, no. 4, pp. 935–946, 2013\. * [19] A. V. Proskurnikov, A. S. Matveev, and M. Cao, “Opinion dynamics in social networks with hostile camps: Consensus vs. polarization,” _IEEE Transactions on Automatic Control_ , vol. 61, no. 6, pp. 1524–1536, 2016. * [20] D. Meng, M. Du, and Y. Jia, “Interval bipartite consensus of networked agents associated with signed digraphs,” _IEEE Transactions on Automatic Control_ , vol. 61, no. 12, pp. 3755–3770, 2016. * [21] Y. Zhang and Y. Liu, “Nonlinear second-order multi-agent systems subject to antagonistic interactions without velocity constraints,” _Applied Mathematics and Computation_ , vol. 364, p. 124667, 2020. * [22] M. E. Valcher and P. Misra, “On the consensus and bipartite consensus in high-order multi-agent dynamical systems with antagonistic interactions,” _Systems & Control Letters_, vol. 66, pp. 94–103, 2014. * [23] J. Lu, Y. Wang, X. Shi, and J. Cao, “Finite-time bipartite consensus for multiagent systems under detail-balanced antagonistic interactions,” _IEEE Transactions on Systems, Man, and Cybernetics: Systems_ , 2019, DOI: 10.1109/TSMC.2019.2938419. * [24] A. V. Proskurnikov and R. Tempo, “A tutorial on modeling and analysis of dynamic social networks. Part II,” _Annual Reviews in Control_ , vol. 45, pp. 166–190, 2018. * [25] G. Shi, C. Altafini, and J. S. Baras, “Dynamics over signed networks,” _SIAM Review_ , vol. 61, no. 2, pp. 229–257, 2019. * [26] W. Zhang, Z. Zuo, and Y. Wang, “Cooperative control in the presence of antagonistic reciprocity,” in _11th Asian Control Conference_. IEEE, 2017, pp. 745–749. * [27] W. Zhang, Z. Zuo, Y. Wang, and W. Ren, “Quasi-containment control against antagonistic information,” _IEEE Transactions on Automatic Control_ , conditionally accepted. * [28] W. Zhang, Z. Zuo, Y. Wang, and Z. Zhang, “Double-integrator dynamics for multiagent systems with antagonistic reciprocity,” _IEEE Transactions on Cybernetics_ , 2019, DOI: 10.1109/TCYB.2019.2939487. * [29] L. Wang, T. Ye, and J. Du, “Opinion dynamics in social networks (in chinese),” _SCIENTIA SINICA Informationis_ , vol. 1, no. 48, pp. 3–23, 2018\. * [30] N. E. Friedkin, _A Structural Theory of Social Influence_. Cambridge University Press, 2006. * [31] G. C. Calafiore and M. C. Campi, “The scenario approach to robust control design,” _IEEE Transactions on Automatic Control_ , vol. 51, no. 5, pp. 742–753, 2006. * [32] G. C. Calafiore, “Random convex programs,” _SIAM Journal on Optimization_ , vol. 20, no. 6, p. 3427, 2010. * [33] R. Olfati-Saber and R. M. Murray, “Consensus problems in networks of agents with switching topology and time-delays,” _IEEE Transactions on Automatic Control_ , vol. 49, no. 9, pp. 1520–1533, 2004. * [34] M. De Domenico, A. Solé-Ribalta, E. Cozzo, M. Kivelä, Y. Moreno, M. A. Porter, S. Gómez, and A. Arenas, “Mathematical formulation of multilayer networks,” _Physical Review X_ , vol. 3, no. 4, pp. 1–15, 2013. * [35] P. Dandekar, A. Goel, and D. T. Lee, “Biased assimilation, homophily, and the dynamics of polarization,” _Proceedings of the National Academy of Sciences_ , vol. 110, no. 15, pp. 5791–5796, 2013. * [36] T. A. Snijders, J. Koskinen, and M. Schweinberger, “Maximum likelihood estimation for social network dynamics,” _The Annals of Applied Statistics_ , vol. 4, no. 2, pp. 567–588, 2010. * [37] H.-T. Wai, A. Scaglione, and A. Leshem, “Active sensing of social networks,” _IEEE Transactions on Signal and Information Processing over Networks_ , vol. 2, no. 3, pp. 406–419, 2016. * [38] A. J. Laub, _Matrix Analysis for Scientists and Engineers_. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, USA, 2005. * [39] M. C. Campi and S. Garatti, “The exact feasibility of randomized solutions of uncertain convex programs,” _SIAM Journal on Optimization_ , vol. 19, no. 3, pp. 1211–1230, 2008. * [40] L. Huang, L. Wang, and C. Hollot, “On robust stability of polynomials and related topics,” _Systems Science and Mathematical Science_ , vol. 5, no. 1, pp. 42–54, 1992. * [41] K. Ogata, _Discrete-Time Control Systems_. Prentice Hall Englewood Cliffs, NJ, 1995.
# Spectrum Sharing for 6G Integrated Satellite-Terrestrial Communication Networks Based on NOMA and Cognitive Radio Xin Liu, Kwok-Yan Lam, Feng Li, Jun Zhao, Li Wang (_Correspondence author: Feng Li_) X. Liu is with the School of Information and Communication Engineering, Dalian University of Technology, Dalian 116024, China (e-mail:liuxinstar1984@dlut.edu.cn).K. Lam and J. Zhao are with School of Computer Science and Engineering, Nanyang Technological University, 639798, Singapore<EMAIL_ADDRESS>junzhao@ntu.edu.sg)F. Li is with School of Information and Electronic Engineering, Zhejiang Gongshang University, Hangzhou, 310018, China. F. Li is also at School of Computer Science and Engineering, Nanyang Technological University, 639798, Singapore. (fengli2002@yeah.net)L. Wang is with College of Marine Electrical Engineering, Dalian Maritime University, Dalian, 116026, China<EMAIL_ADDRESS> ###### Abstract The explosive growth of bandwidth hungry Internet applications has led to the rapid development of new generation mobile network technologies that are expected to provide broadband access to the Internet in a pervasive manner. For example, 6G networks are capable of providing high-speed network access by exploiting higher frequency spectrum; high-throughout satellite communication services are also adopted to achieve pervasive coverage in remote and isolated areas. In order to enable seamless access, Integrated Satellite-Terrestrial Communication Networks (ISTCN) has emerged as an important research area. ISTCN aims to provide high speed and pervasive network services by integrating broadband terrestrial mobile networks with satellite communication networks. As terrestrial mobile networks began to use higher frequency spectrum (between 3GHz to 40GHz) which overlaps with that of satellite communication (4GHz to 8GHz for C band and 26GHz to 40GHz for Ka band), there are opportunities and challenges. On one hand, satellite terminals can potentially access terrestrial networks in an integrated manner; on the other hand, there will be more congestion and interference in this spectrum, hence more efficient spectrum management techniques are required. In this paper, we propose a new technique to improve spectrum sharing performance by introducing Non- orthogonal Frequency Division Multiplexing (NOMA) and Cognitive Radio (CR) in the spectrum sharing of ISTCN. In essence, NOMA technology improves spectrum efficiency by allowing different users to transmit on the same carrier and distinguishing users by user power levels while CR technology improves spectrum efficiency through dynamic spectrum sharing. Furthermore, some open researches and challenges in ISTCN will be discussed. ## Introductions In order to enable pervasive network connectivity, Integrated Satellite- Terrestrial Communication Networks (ISTCN) has emerged as an important research area. ISTCN aims to provide high-speed and pervasive network services by integrating broadband terrestrial mobile networks with satellite communication networks. The ISTCN can provide reliable communications and global interconnections for disaster affected areas, remote areas and emergency areas, where the terrestrial communication facilities are not easy to use [1]. The future 6G networks are expected to offer unprecedented opportunities for Smart Cities and Internet of Things applications through their global seamless coverage, G-bit communication capacity, ultra reliable real-time communications and ubiquitous machine type communications. This new generation terrestrial mobile network achieved their functionalities by intelligently and optimally exploiting the higher frequency spectrum, typically in the range of 3GHz to 40GHz. However, for suburban and isolated geographic locations, the coverage of high-speed terrestrial mobile networks could be limited hence need to be complemented by satellite communications in order to meet the connectivity requirements of safety critical applications such as Internet of Vehicles and Industry 4.0 control systems [4]. As terrestrial mobile network began to use higher frequency spectrum which overlaps with that of satellite communications (e.g. 4GHz to 8GHz for C band and 26GHz to 40GHz for Ka band), there are vast opportunities as well as difficult challenges. On one hand, satellite terminals can potentially access terrestrial network in an integrated manner; on the other hand, there will be more congestion and interference in this spectrum, hence more efficient spectrum management techniques are required. The objective is to make full use of the complementary advantages of satellite networks and terrestrial mobile networks, so as to realize the all-weather and all-regional seamless coverage of high-speed mobile broadband network . In addition, it aims to effectively alleviate the shortage of satellite spectrum resources by applying spectrum sharing technology to reuse the terrestrial spectrum for the satellite communications [2]. In 6G mobile communications, Non-orthogonal Multiple Access (NOMA) and Cognitive Radio (CR) are two most promising spectrum sharing technologies [3]. NOMA is different from the traditional Orthogonal Multiple Access (OMA), which uses non- orthogonal resource allocation approach to accommodate more users [5]. At the transmitter, the transmit information of multiple users is superimposed and encoded in the power domain by intentionally adding the interference information. At the receiver, Successive Interference Cancellation (SIC) is used to separate the user information by sequentially detecting and canceling the signal of each user. It is estimated that NOMA can improve the current spectrum efficiency by 5$\sim$15 times [6, 7]. Therefore, the satellite- terrestrial NOMA spectrum sharing can make one satellite frequency band accommodate more users and thus greatly improve the communication capacity. CR, based on software radio, allows the system to adaptively adjust transmit parameters by sensing the current communication environment, so as to achieve efficient spectrum resource utilization [8, 9]. It can share the spectrum resources among the heterogeneous communication systems through spectrum sensing and dynamic reconfiguration capability. As a secondary user (SU), the CR system can opportunistically utilize the idle spectrum of primary user (PU) or share the spectrum with the PU at a lower power [10, 11]. Satellite- terrestrial CR spectrum sharing makes the satellite system and terrestrial system utilize the same spectrum resources, which can alleviate satellite spectrum tension effectively. However, if compared with the spectrum sharing studies for terrestrial networks, the related works for ISTCN still remain insufficient. In [12], the capacity of NOMA-uplink satellite network was analyzed, which has proved the advantage of NOMA to improve the satellite communication capacity. In [13], a joint resource allocation and network stability optimization was proposed to maximize the long-term network utility of NOMA-downlink satellite system. In [14], regret minimization solution was put forward for PU and SU’s spectrum access to the satellite resources when existing cognitive interferers. In [15], cooperative transmission strategy was proposed for cognitive satellite networks, where the mobile users in the terrestrial network can help the communication of the satellite network to improve its transmission performance. Nevertheless, NOMA and CR for integrated satellite-terrestrial spectrum sharing are less considered. In this article, NOMA and CR based spectrum sharing for the ISTCN is proposed to solve the problem of satellite spectrum scarcity. The contributions of the article are concluded as follows. (1) The network model and network access model for the ISTCN are proposed, which allow the satellite system and terrestrial system share the same spectrum by the integration of satellite and terrestrial components; (2) The satellite-terrestrial NOMA spectrum sharing is presented to let multiple users to access the same satellite spectrum by superposition coding in power domain; (3) The satellite-terrestrial CR spectrum sharing is proposed to make the satellite system and terrestrial system share the same spectrum resource by suppressing their mutual interferences; (4) By combining NOMA and CR, the satellite-terrestrial CR-NOMA spectrum sharing is put forward to achieve full spectrum access by using both the idle and busy spectrum. ## Integrated Satellite-Terrestrial Communication Networks Contemporary satellite communication services are no longer competitors of terrestrial cellular network. Instead, they are often adopted to complement cellular network services so as to provide seamless coverage. Terrestrial cellular networks are suitable for areas with high user density in the urban areas; however, they are typically less cost effective in covering remote and even isolated geographic areas. While satellite communications can provide large area coverage at low cost, they have their limitations in covering urban areas due to the influence of shadowing effect. Therefore, ISTCN is believed to be a suitable approach to achieve the global coverage with optimal cost. The network model for the ISTCN is shown in Fig. 1, which is an integrated satellite-terrestrial system composed of one or more Highly-Elliptical-Orbit (HEO) and Low-Earth-Orbit (LEO) satellites and terrestrial cellular system. Both the terrestrial system and satellite system operate in the same frequency band to ensure the global seamless coverage of the user terminals. In the ISTCN, the satellite terminal and terrestrial terminal can communicate with each other depending on the network switching between the satellite system and cellular system, and also a dual-mode satellite terminal can choose either of the two systems to communicate by measuring the transmission cost. Figure 1: Network model for ISTCN. The network access model for spectrum sharing in ISTCN includes hybrid network access (HNA) and combined network access (CNA), as shown in Fig. 2. In the HNA, the user terminal and subscription are different, and the access network and core network of each system are disjoint and linked together through the public network. Therefore, the users may have good access to the two systems, but there is no integration between them. The satellite system and terrestrial system can adopt the same or different air interface technology depending on the specific network scenario. In the HNA, however, the user only have one terminal and one subscription, and the services of the two systems are almost seamless switching. Therefore, the quality of service (QoS) of HNA is higher due to the system integration. But the two systems have to adopt compatible air interface technology and share the same frequency band. Figure 2: Network access mode for ISTCN. In the ISTCN, the satellite system and terrestrial system may coexist in the same frequency band to alleviate the satellite spectrum scarcity. The existing spectrum sharing methods are mainly divided into the following two categories. Static spectrum sharing: The idea of static spectrum sharing is spatial isolation, which can reuse the time, frequency and other resources through orthogonal access manner in different spatial areas. However, it allocates the specific spectrum resource for each user, which can not meet the dynamic spectrum demands of the users, resulting in that the load of some spectrum is too heavy while other spectrum has higher idle rate. In 6G, NOMA, as a new static spectrum sharing approach based on power domain multiplexing, has been proposed to allocate the same time-frequency resource to different users, which can greatly improve the spectrum efficiency compared with 5G. Dynamic spectrum sharing: By adopting CR technology, the communication system can opportunistically use the underutilized frequency resources to achieve better dynamic spectrum management. Satellite-terrestrial CR spectrum sharing can realize the heterogeneous integration of satellite network and terrestrial network and solve the problem of satellite spectrum shortage. Therefore, NOMA and CR as efficient spectrum sharing technologies can make the ISTCN achieve the interconnections between massive satellite terminals and terrestrial terminals under the limited satellite spectrum resources. ## Satellite-Terrestrial NOMA Spectrum Sharing The core idea of NOMA is to realize the multi-user multiplexing of single time-frequency resource block by introducing a new power domain dimension. At the transmitter, the signals of different users are set with different power levels, which are transmitted in the same resource block by superposition coding. While at the receiver, the signals of different users are separated and decoded by using SIC in the descending order of the power levels. . ### NOMA Spectrum Sharing Model The satellite-terrestrial NOMA spectrum sharing model is shown in Fig. 3, where the terrestrial users access the satellite spectrum through NOMA. In the satellite uplink, the data and channel state information (CSI) of the terrestrial terminals are sent to the satellite gateway, which then groups the users according to the CSI and allocates the maximum transmit power to each user for superposition coding. The signals of the same grouping users are sent to the satellite in the same frequency band. The satellite receiver uses SIC to decode the signals of each user. If a user transmits stronger signal in a better link, its signal will be decoded first. And the signal in a poor link will be decoded from the remaining signals after subtracting the decoded signals by SIC. In the satellite downlink, the satellite transmitter allocates the power of each user according to the link quality, and the user in a poor link is allocated larger power to ensure the receiving performance. The NOMA signal is transmitted to each satellite terminal, which uses SIC to first decode the signals with larger power and then decode its own signal from the remaining signals. Figure 3: NOMA based satellite-terrestrial spectrum sharing model. ### User NOMA Grouping The terrestrial users can be divided into several groups, each of which is assigned a separate frequency band for NOMA transmission. The NOMA grouping can reduce the multi-user interference in decoding by decreasing the number of users in the same frequency band. The advantage of NOMA is obvious only when the users with great channel differences are assigned to one group. In the terrestrial NOMA, the physical distance between user and base station is used as the grouping basis, whereby the center user and edge user within the coverage of base station are usually assigned to one group. However, the satellite communication channel is more complex than the terrestrial mobile communication channel, whose path attenuation is not sensitive to the user’s geographical location. Therefore, the distance grouping basis is no longer applicable in the satellite communications. Satellite NOMA grouping needs to fully consider other attenuation characteristics of the satellite channels besides the free space loss, such as beam gain, shadow fading, multipath fading, and rain fading etc. The channel fading difference of different satellite users can be used as the grouping basis to eliminate the insensitive path attenuation. ### Cooperative Satellite-Terrestrial NOMA Cooperative NOMA is mostly used in the downlink of a communication system, whereby the user with good channel can help to decode the information of the user with poor channel, so as to enhance its receiving performance. In the ISTCN, cooperative NOMA can be carried out among different satellite and terrestrial terminals to improve the transmission performance of the users in fading satellite channels. As shown in Fig. 6, the satellite terminals in fading channels can form a NOMA group with either the satellite terminals in good channels or the terrestrial terminals. The satellite transmits the NOMA signal to the satellite terminals and the terrestrial base station. The satellite terminals in good channels first decode all the signals by SIC, and then use decode-and-forward (DF) protocol to send the decoded signals to the satellite terminals in fading channels. However, the terrestrial terminals cannot achieve the prior information of the satellite terminals and thus are unable to decode their signals. Therefore, the terrestrial terminals first decode their own signals and subtract the decoded signals from the received signal, and then use amplify-and-forward (AF) protocol to send the remaining signal to the satellite terminals in fading channels. Figure 4: NOMA-based cooperative spectrum sharing. ## Satellite-Terrestrial CR Spectrum Sharing ### CR Spectrum Sharing Model Using CR technology, the satellite communication system can flexibly share spectrum resources with the terrestrial communication system, which can improve the spectrum utilization via opportunistically accessing the frequency bands permitted by the licensed users. The typical CR spectrum sharing scenarios of the ISTCN can be divided into two categories. One is licensed satellite system and terrestrial CR system, and the other is licensed terrestrial system and satellite CR system. As shown in Fig. 5, in the satellite uplink, the satellite CR system communicates in the terrestrial channels by spectrum sharing technology. If the terrestrial user is not using the channel, the satellite user can transmit data with its maximum power. However, the satellite user must always sense the channel state. If the presence of the terrestrial user in the channel has been detected, the satellite user has to switch to another idle channel. However, if there is no idle channels, the satellite user may continue to use this channel but cause harmful interference to the terrestrial system. In the satellite downlink, the interference from the terrestrial user will also decrease the satellite communication performance. To achieve low-interference spectrum sharing, the satellite user must detect the spectrum occupation state of the terrestrial system accurately and select an idle channel for transmissions. In addition, the satellite system can also access the busy spectrum by controlling its transmit power so that its power does not exceed the maximum interference tolerated by the terrestrial system. Figure 5: CR spectrum sharing model. ### Interference suppression technology The premise of CR spectrum sharing is that the interference between the satellite system and terrestrial system does not affect their normal communications. Some interference suppression technologies are introduced as follows. Interference cognition: Interference cognition can detect the interference holes in the surrounding electromagnetic environment, identify the interference and estimate the channel quality, which can provide the basis for the anti-interference decision. In the terrestrial communication, the interference mostly occurs in the channel from CR system to licensed system, and the interference cognition is usually defined around the licensed receiver. However, there may be two-way interference in the ISTCN, and the interference cognition should be defined both around CR system and licensed system. For example, the interference cognition for the ISTCN can be defined around the earth station and the satellite spot beam. Power control: CR system combines channel state, receiver signal-to-noise ratio (SNR) and interference information to flexibly adjust its transmit power to avoid interference with licensed user in the same frequency band. On the one hand, the transmit power can be minimized to save the energy of the satellite terminal on the premise of ensuring the communication capacity. On the other hand, the power can be optimized to maximize the communication capacity of the ISTCN providing that the interference threshold is not exceeded. Satellite beamforming: Beamforming technology can transmit the target signal along the desired direction by weighting the received signals of antenna array elements. Beamforming allows multiple users to utilize the same frequency band in the same geographical area at the same time, which makes it possible to deploy dense networks with less interference in the unexpected direction. It can be used as an interference cancellation technology in the transmitter or receiver of the satellite, which can realize the spectrum sharing of satellite system and terrestrial system in the angle domain. ## Satellite-Terrestrial CR-NOMA Spectrum Sharing CR can realize the spectrum coexistence of satellite system and terrestrial system, while NOMA can achieve the sharing access of limited satellite spectrum by massive users. Therefore, by combining CR and NOMA, the satellite spectrum utilization can be further improved. Satellite terminals can access both the idle and busy spectrum via CR-NOMA, which will achieve high-efficient full spectrum access. CR-NOMA in idle spectrum: Multiple satellite terminals can access the idle spectrum by NOMA, which will not bring any interference to the terrestrial system. However, the available idle frequency bands are usually discontinuous and fragmented, which are difficult to meet the broadband access of massive satellite users. Spectrum aggregation technology has been put forward to aggregate discrete idle frequency bands into broadband spectrum to support large-bandwidth data transmissions. Non-continuous Orthogonal Frequency Division Multiplexing (NC-OFDM) can realize the subcarrier aggregation by zeroing the non-idle subcarriers according to their bandwidth and locations. Therefore, by introducing NC-OFDM into NOMA, the superimposed broadband signal can be transmitted over the aggregated idle subcarriers. CR-NOMA in busy spectrum: Satellite terminal and terrestrial terminal can share the same spectrum by NOMA, but interfere with each other. To guarantee the terrestrial communication performance, the satellite signals are first decoded with the terrestrial signals as the noise. Then the decoded satellite signals are cancelled from the received NOMA signal by SIC, and the remaining signal is used to decode the terrestrial signals without the interference from the satellite. However, if the transmit power of terrestrial terminals is large enough or the terrestrial communication performance is ignored, the terrestrial signals can be first decoded to guarantee the decoding performance of satellite signals. Figure 6: CR-NOMA spectrum sharing. Though the CR-NOMA can make the ISTCN achieve full spectrum access, the multi- user interference caused by NOMA and the satellite-terrestrial interference cause by CR may decrease the NOMA decoding performance. The satellite terminals and terrestrial terminals should be grouped appropriately and the decoding order must be properly arranged according to the power and service requirements of the users. ## Open Researches and Challenges This article has introduced some fundamental works on spectrum sharing for ISTCN, such as ISTCN network model, satellite-terrestrial NOMA spectrum sharing, satellite-terrestrial CR spectrum sharing and satellite-terrestrial CR-NOMA spectrum sharing. However, there are still some open researches and challenges to be discussed in the future. Satellite spectrum sensing: The premise of satellite-terrestrial spectrum sharing without interference is to perform accurate spectrum sensing. Due to the limited satellite transmission capacity and the significant signal attenuation caused by atmospheric effects such as shadowing and rain fading, it is a great challenge to accurately sense satellite spectrum. In addition, the spectrum sensing in LEO satellite communication also faces the problems of mobility and available frequency shortage. Fair satellite NOMA grouping: Beam edge users are located in the overlapping area of different satellite beams and will suffer great inter-beam interference. Therefore, multi-user interference and inter-beam interference will seriously reduce the decoding performance of the edge users. It is necessary to propose a fair satellite NOMA grouping method to allocate low- power users and fewer users for the groups of edge users, so as to decrease the NOMA decoding interference. Satellite NOMA receiver design: Due to the shadowing, multipath fading, rain fading and other channel interference of satellite communication, it is difficult to carry out perfect SIC at the satellite terminal. When restructuring and canceling the inaccurate decoded signals, the decoding error will be transferred to the subsequent signal demodulation, which will decrease the decoding performance. Therefore, to guarantee SIC performance, the satellite receiver design should adopt some new signal processing technology to suppress the interference and noise, such as adaptive filtering, wavelet transform and weak signal detection etc. Integrated satellite-6G network: 6G can support high-capacity, multi-service and high-speed wireless communications. Integrated satellite-6G network can meet the global coverage of mobile Internet and the ubiquitous network access of all kinds of users. To better integrate with the terrestrial 6G network, the satellite segment needs to reuse all the functional modules of 6G core network, such as Internet interface, quality of service, user mobility and security etc. ## Conclusion Remarks and Future Works In this article, we propose NOMA and CR based spectrum sharing schemes for ISTCN to improve the satellite spectrum utilization by allowing the satellite communication to share the spectrum licensed to the terrestrial communication. The satellite-terrestrial spectrum sharing need to sense the terrestrial spectrum state and suppress the mutual interference between satellite system and terrestrial system. Some interference suppression technologies for satellite-terrestrial spectrum sharing are also introduced. By combining CR and NOMA, the ISTCN can use CR-NOMA to achieve full spectrum sharing by accessing both the idle and busy spectrum. Finally, some promising researches and challenges for the ISTCN have been discussed. ## References * [1] M. Jia, X. Gu, Q. Guo, W. Xiang, and N. Zhang, “Broadband hybrid satellite-terrestrial communication systems based on cognitive radio toward 5g,” IEEE Wireless Communications, vol. 23, no. 6, pp. 96–106, 2016. * [2] B. Feng, H. Zhou, H. Zhang, G. Li, H. Li, S. Yu, and H. Chao, “Hetnet: A flexible architecture for heterogeneous satellite-terrestrial networks,” IEEE Network, vol. 31, no. 6, pp. 86–92, 2017. * [3] S. M. R. Islam, M. Zeng, O. A. Dobre, and K. Kwak, “Resource allocation for downlink noma systems: Key techniques and open issues,” IEEE Wireless Communications, vol. 25, no. 2, pp. 40–47, 2018. * [4] A. Chattopadhyay, K. Lam, and Y. Tavva, “Autonomous vehicle: Security by design,” IEEE Transactions on Intelligent Transportation Systems, vol. to appear, pp. 1–15, 2020. * [5] H. Zhang, Y. Qiu, K. Long, G. K. Karagiannidis, X. Wang, and A. Nallanathan, “Resource allocation in noma-based fog radio access networks,” IEEE Wireless Communications, vol. 25, no. 3, pp. 110–115, 2018\. * [6] S. Mounchili and S. Hamouda, “Pairing distance resolution and power control for massive connectivity improvement in noma systems,” IEEE Transactions on Vehicular Technology, vol. 69, no. 4, pp. 4093–4103, 2020. * [7] F. Zhou, Y. Wu, R. Q. Hu, Y. Wang, and K. K. Wong, “Energy-efficient noma enabled heterogeneous cloud radio access networks,” IEEE Network, vol. 32, no. 2, pp. 152–160, 2018. * [8] X. Liu, S. Hu, M. Li, and B. Lai, “Energy-efficient resource allocation for cognitive industrial internet of things with wireless energy harvesting,” IEEE Transactions on Industrial Informatics, pp. 1–1, 2020\. * [9] F. Li, K. Lam, X. Li, Z. Sheng, J. Hua, and L. Wang, “Advances and emerging challenges in cognitive internet-of-things,” IEEE Transactions on Industrial Informatics, vol. 16, no. 8, pp. 5489–5496, 2020. * [10] G. Hattab and M. Ibnkahla, “Multiband spectrum access: Great promises for future cognitive radio networks,” Proceedings of the IEEE, vol. 102, no. 3, pp. 282–306, 2014. * [11] J. Zou, H. Xiong, D. Wang, and C. W. Chen, “Optimal power allocation for hybrid overlay/underlay spectrum sharing in multiband cognitive radio networks,” IEEE Transactions on Vehicular Technology, vol. 62, no. 4, pp. 1827–1837, 2013. * [12] X. Yan, H. Xiao, K. An, G. Zheng, and S. Chatzinotas, “Ergodic capacity of noma-based uplink satellite networks with randomly deployed users,” IEEE Systems Journal, vol. 14, no. 3, pp. 3343–3350, 2020. * [13] J. Jiao, Y. Sun, S. Wu, Y. Wang, and Q. Zhang, “Network utility maximization resource allocation for noma in satellite-based internet of things,” IEEE Internet of Things Journal, vol. 7, no. 4, pp. 3230–3242, 2020. * [14] Y. E. Sagduyu, Y. Shi, A. B. MacKenzie, and Y. T. Hou, “Regret minimization for primary/secondary access to satellite resources with cognitive interference,” IEEE Transactions on Wireless Communications, vol. 17, no. 5, pp. 3512–3523, 2018. * [15] S. H. Chae, C. Jeong, and K. Lee, “Cooperative communication for cognitive satellite networks,” IEEE Transactions on Communications, vol. 66, no. 11, pp. 5140–5154, 2018.
∎ 11institutetext: D-ặng Võ Phúc 22institutetext: Faculty of Education Studies University of Khanh Hoa, Nha Trang, Khanh Hoa, Vietnam 22email<EMAIL_ADDRESS> # On modules over the mod 2 Steenrod algebra and hit problems D-ặng Võ Phúc ###### Abstract Let us consider the prime field of two elements, $\mathbb{F}_{2}\equiv\mathbb{Z}_{2}.$ It is well-known that the classical ”hit problem” for a module over the mod 2 Steenrod algebra $\mathscr{A}$ is an interesting and important open problem of Algebraic topology, which asks a minimal set of generators for the polynomial algebra $\mathcal{P}_{m}:=\mathbb{F}_{2}[x_{1},x_{2},\ldots,x_{m}]$, regarded as a connected unstable $\mathscr{A}$-module on $m$ variables $x_{1},\ldots,x_{m},$ each of degree 1. The algebra $\mathcal{P}_{m}$ is the $\mathbb{F}_{2}$-cohomology of the product of $m$ copies of the Eilenberg- MacLan complex $K(\mathbb{F}_{2},1).$ Although the hit problem has been thoroughly studied for more than 3 decades, solving it remains a mystery for $m\geq 5.$ It is our intent in this work is of studying the hit problem of five variables. More precisely, we develop our previous work [Commun. Korean Math. Soc. 35 (2020), 371-399] on the hit problem for $\mathscr{A}$-module $\mathcal{P}_{5}$ in a degree of the generic form $n_{t}:=5(2^{t}-1)+18.2^{t},$ for any non-negative integer $t.$ An efficient approach to solve this problem had been presented. Two applications of this study are to determine the dimension of $\mathcal{P}_{6}$ in the generic degree $5(2^{t+4}-1)+n_{1}.2^{t+4}$ for all $t>0$ and to describe the modular representations of the general linear group of rank 5 over $\mathbb{F}_{2}.$ As a corollary, the cohomological ”transfer”, defined by William Singer [Math. Z. 202 (1989), 493-523], is an isomorphism in bidegree $(5,5+n_{0}).$ Singer’s transfer is one of the relatively efficient tools to approach the structure of mod-2 cohomology of the Steenrod algebra. Mathematics Subject Classification (2010) 13A50 $\cdot$ 55Q45 $\cdot$ 55S10 $\cdot$ 55S05 $\cdot$ 55T15 $\cdot$ 55R12 ###### Keywords: Adams spectral sequencesSteenrod algebra Hit problemAlgebraic transfer ## 1 Introduction Let $\mathcal{O}^{S}(i,\mathbb{F}_{2},\mathbb{F}_{2})$ denote the set of all stable cohomology operations of degree $i,$ with coefficient in the prime field $\mathbb{F}_{2}.$ Then, the $\mathbb{F}_{2}$-algebra $\mathscr{A}:=\bigoplus_{i\geq 0}\mathcal{O}^{S}(i,\mathbb{F}_{2},\mathbb{F}_{2})$ is called the mod 2 Steenrod algebra. In other words, the algebra $\mathscr{A}$ is the algebra of stable operations on the mod 2 cohomology. In J.M , Milnor observed that this algebra is also a graded connected cocommutative Hopf algebra over $\mathbb{F}_{2}.$ In some cases, the resulting $\mathscr{A}$-module structure on $H^{*}(X,\mathbb{F}_{2})$ provides additional information about CW- complexes $X;$ for instance (see section three for a detailed proof), the CW- complexes $\mathbb{C}P^{4}/\mathbb{C}P^{2}$ and $\mathbb{S}^{6}\vee\mathbb{S}^{8}$ have cohomology rings that agree as a graded commutative $\mathbb{F}_{2}$-algebras, but are different as a module over $\mathscr{A}.$ Afterwards, the Steenrod algebra is widely studied by mathematicians whose interests range from algebraic topology and homotopy theory to manifold theory, combinatorics, representation theory, and more. It is well-known that the $\mathbb{F}_{2}$-cohomology of the Eilenberg-MacLan complex $K(\mathbb{F}_{2},1)$ is isomorphic to $\mathbb{F}_{2}[x],$ the polynomial ring of degree $1$ in one variable. Hence, based upon the Künneth formula for cohomology, we have an isomorphism of $\mathbb{F}_{2}$-algebras $\mathcal{P}_{m}:=H^{*}((K(\mathbb{F}_{2},1))^{\times m},\mathbb{F}_{2})\cong\underset{\text{ $m$ times}}{\underbrace{\mathbb{F}_{2}[x_{1}]\otimes_{\mathbb{F}_{2}}\mathbb{F}_{2}[x_{2}]\otimes_{\mathbb{F}_{2}}\cdots\otimes_{\mathbb{F}_{2}}\mathbb{F}_{2}[x_{m}]}}\cong\mathbb{F}_{2}[x_{1},\ldots,x_{m}],$ where $x_{i}\in H^{1}((K(\mathbb{F}_{2},1))^{\times m},\mathbb{F}_{2})$ for every $i.$ Since $\mathcal{P}_{m}$ is the cohomology of a CW-complex, it is equipped with a structure of unstable module over $\mathscr{A}.$ It has been known (see also S.E ) that $\mathscr{A}$ is spanned by the Steenrod squares $Sq^{i}$ of degree $i$ for $i\geq 0$ and that the action of $\mathscr{A}$ on $\mathcal{P}_{m}$ is depicted as follows: $\begin{array}[]{ll}Sq^{i}(x_{t})&=\left\\{\begin{array}[]{lll}x_{t}&\mbox{if}&i=0,\\\ x_{t}^{2}&\mbox{if}&i=1,\ \ \mbox{({the instability condition})},\\\ 0&\mbox{if}&i>1,\end{array}\right.\\\ Sq^{i}(FG)&=\sum_{0\leq\alpha\leq i}Sq^{\alpha}(F)Sq^{i-\alpha}(G),\ \mbox{for all $F,\ G\in\mathcal{P}_{m}$}\ \ (\mbox{{the Cartan formula})}.\end{array}$ It is to be noted that since $Sq^{\deg(F)}(F)=F^{2}$ for any $F\in\mathcal{P}_{m}$, the polynomial ring $\mathcal{P}_{m}$ is also an unstable $\mathscr{A}$-algebra. Letting $GL_{m}:=GL(m,\mathbb{F}_{2})$ for the general linear group of degree $m$ over $\mathbb{F}_{2}.$ This $GL_{m}$ when $m\geq 2,$ which can be generated by two elements (see Waterhouse Waterhouse ), acts on $\mathcal{P}_{m}$ by matrix substitution. So, in addition to $\mathscr{A}$-module structure, $\mathcal{P}_{m}$ is also a (right) $\mathbb{F}_{2}GL_{m}$-module. The classical ”hit problem” for the algebra $\mathscr{A}$, which is concerned with seeking a minimal set of $\mathscr{A}$-generators for $\mathcal{P}_{m}$, has been initiated in a variety of contexts by Peterson F.P , Priddy S.P , Singer W.S2 , and Wood R.W . Structure of modules over $\mathscr{A}$ and hit problems are currently one of the central subjects in Algebraic topology and has a great deal of intensively studied by many authors like Brunetti and collaborators Brunetti1 ; Brunetti2 , Crabb-Hubbuck C.H , Inoue M.I1 ; M.I2 , Janfada-Wood J.W ; J.W2 , Janfada Janfada1 ; Janfada2 , Kameko M.K , Mothebe-Uys M.M , Mothebe M.M2 , Pengelley-William P.W , the present author and N. Sum P.S1 ; P.S2 ; D.P1 ; D.P2 ; D.P3 ; D.P6 ; D.P7 ; N.S1 ; N.S2 ; N.S3 , Walker-Wood W.W ; W.W2 , etc. As it is known, when $\mathbb{F}_{2}$ is an $\mathscr{A}$-module concentrated in degree 0, solving the hit problem is to determine an $\mathbb{F}_{2}$-basis for the space of indecomposables, or ”unhit” elements, $Q^{\otimes m}:=\mathbb{F}_{2}\otimes_{\mathscr{A}}\mathcal{P}_{m}=\mathcal{P}_{m}/\overline{\mathscr{A}}\mathcal{P}_{m}$ where $\overline{\mathscr{A}}$ is the positive degree part of $\mathscr{A}$. It is well-known that the action of $GL_{m}$ and the action of $\mathscr{A}$ on $\mathcal{P}_{m}$ commute. So, there is an induced action of $GL_{m}$ on $Q^{\otimes m}.$ The structure of $Q^{\otimes m}$ has been treated for $m\leq 4$ by Peterson F.P , Kameko M.K and Sum N.S1 . The general case is an interesting open problem. Most notably, the study of this space plays a vital role in describing the $E^{2}$-term of the Adams spectral sequence (Adams SS), ${\rm Ext}_{\mathscr{A}}^{m,m+*}(\mathbb{F}_{2},\mathbb{F}_{2})$ via the $m$-th Singer cohomological ”transfer” W.S1 . This transfer is a linear map $Tr^{\mathscr{A}}_{m}:(\mathbb{F}_{2}\otimes_{GL_{m}}P_{\mathscr{A}}((\mathcal{P}_{m})^{*}))_{n}\to{\rm Ext}_{\mathscr{A}}^{m,m+n}(\mathbb{F}_{2},\mathbb{F}_{2})=H^{m,m+n}(\mathscr{A},\mathbb{F}_{2}),$ from the subspace of all $\overline{\mathscr{A}}$-annihilated elements to the $E^{2}$-term of the Adams SS. Here $(\mathcal{P}_{m})^{*}=H_{*}((K(\mathbb{F}_{2},1))^{\times m},\mathbb{F}_{2})$ and $\mathbb{F}_{2}\otimes_{GL_{m}}P_{\mathscr{A}}((\mathcal{P}_{m})^{*})$ are the dual of $\mathcal{P}_{m}$ and $(Q^{\otimes m})^{GL_{m}},$ respectively, where $(Q^{\otimes m})^{GL_{m}}$ denotes the space of $GL_{m}$-invariants. A natural question arises: Why do we need to calculate the Adams $E^{2}$-term? The answer is that it is involved in determining the stable homotopy groups of spheres. These groups are pretty fundamental and interesting. Nevertheless, they are also not fully-understood subjects yet. Therefore, the clarification of these problems is an important task of Algebraic topology. It has been shown (see J.B , W.S1 ) that the algebraic transfer is highly nontrivial, more precisely, that $Tr^{\mathscr{A}}_{m}$ is an isomorphism for $0<m<4$ and that the ”total” transfer $\bigoplus_{m\geq 0}Tr^{\mathscr{A}}_{m}:\bigoplus_{m\geq 0}(\mathbb{F}_{2}\otimes_{GL_{m}}P_{\mathscr{A}}((\mathcal{P}_{m})^{*}))_{n}\to\bigoplus_{m\geq 0}{\rm Ext}_{\mathscr{A}}^{m,m+n}(\mathbb{F}_{2},\mathbb{F}_{2})$ is a homomorphism of bigraded algebras with respect to the product by concatenation in the domain and the usual Yoneda product for the Ext group. Minami’s works N.M1 ; N.M2 have shown the usefulness of the Singer transfer and the hit problem for surveying the Kervaire invariant one problem. This problem, which is a long standing open topic in Algebraic topology, asks when there are framed manifolds with Kervaire invariant one. (Note that a framing on a closed smooth manifold $M^{n}$ is a trivialization of the normal bundle $\nu(M,i)$ of some smooth embedding $i:M\hookrightarrow\mathbb{R}^{n+*}.$ Here $\nu(M,i)$ is defined to be a quotient of the pullback of the tangent bundle of $\mathbb{R}^{n+*}$ by the sub-bundle given by the tangent bundle of $M.$ So, $\nu(M,i)$ is an $*$-dimensional real vector bundle over $M^{n}.$ For more details, we refer the reader to Snaith .) Framed manifolds of Kervaire invariant one have been constructed in dimension $2^{k}-2$ for $2\leq k\leq 6.$ In 2016, by using mod 8 equivariant homotopy theory, Hill, Hopkins, and Ravenel claimed in their surprising work Hill that the Kervaire invariant is 0 in dimension $2^{k}-2$ for $k\geq 8.$ Up to present, it remains undetermined for $k=7$ (or dimension $126$) and this has the status of a hypothesis by Snaith Snaith . Return to Singer’s transfer, in higher homological degrees, the works B.H.H , Ha , Hung , T.N2 , and H.Q determined completely the image of $Tr_{4}^{\mathscr{A}}$. The authors show that $Tr_{4}^{\mathscr{A}}$ contains all the elements of the families $\\{d_{t}|\,t\geq 0\\},\ \\{e_{t}|\,t\geq 0\\}$, $\\{f_{t}|\,t\geq 0\\}$, and $\\{p_{t}|\,t\geq 0\\}$, but none from the families $\\{g_{t+1}|\,t\geq 0\\}$, $\\{D_{3}(t)|\,t\geq 0\\}$, and $\\{p^{\prime}_{t}|\,t\geq 0\\}$. Remarkably, since the family $\\{g_{t+1}|\,t\geq 0\\}\not\subset{\rm Im}(Tr_{4}^{\mathscr{A}}),$ a question of Minami N.M2 concerning the so-called new doomsday conjecture refuted. In Hung , Hưng indicated that $Tr_{4}^{\mathscr{A}}$ is not an isomorphism in infinitely many degrees. In particular, from preliminary calculations in W.S1 , Singer proposed the following. ###### Conjecture 1.1 The transfer homomorphism is a monomorphism in every rank $m>0.$ We have seen above that $Tr_{m}^{\mathscr{A}}$ is an isomorphism for $m<4,$ and so the conjecture holds in these ranks $m.$ Our recent work D.P14 has shown that it is also true for $m=4,$ but the answer to the general case remains a mystery, even in the case of $m=5$ with the help of a computer algebra. It is known, in ranks $\leq 4$, the calculations of Singer W.S1 , Hà Ha , and Nam T.N2 tell us that the non-zero elements $h_{t}\in{\rm Ext}_{\mathscr{A}}^{1,2^{t}}(\mathbb{F}_{2},\mathbb{F}_{2}),\,e_{t}\in{\rm Ext}_{\mathscr{A}}^{4,2^{t+4}+2^{t+2}+2^{t}}(\mathbb{F}_{2},\mathbb{F}_{2}),\,f_{t}\in{\rm Ext}_{\mathscr{A}}^{4,2^{t+4}+2^{t+2}+2^{t+1}}(\mathbb{F}_{2},\mathbb{F}_{2}),$ for all $t\geq 0,$ are detected by the cohomological transfer. In rank 5, based on invariant theory, Singer W.S1 gives an explicit element in ${\rm Ext}_{\mathscr{A}}^{5,5+9}(\mathbb{F}_{2},\mathbb{F}_{2}),$ namely $Ph_{1}$, that is not detected by $Tr_{5}^{\mathscr{A}}.$ In general, direct calculating the value of $Tr_{m}^{\mathscr{A}}$ on any non-zero element is difficult. Moreover, there is no general rule for that, and so, each computation is important on its own. By this and the above results, in the present text, we would like to investigate the family $\\{h_{t}f_{t}=h_{t+1}e_{t}\in{\rm Ext}_{\mathscr{A}}^{5,5+(23.2^{t}-5)}(\mathbb{F}_{2},\mathbb{F}_{2})|\,t\geq 0\\},$ and Singer’s conjecture for $m=5$ in degree $5(2^{t}-1)+18.2^{t}=23.2^{t}-5$ with $t=0.$ To do this, we use a basis of the indecomposables $Q^{\otimes 5}$ in degree $18=5(2^{0}-1)+18.2^{0}$, which is given by our previous work D.P2 (see Proposition 2.7 below). In addition, the main goal of this work is to also compute explicitly the dimension of $Q^{\otimes 5}$ in degree $5(2^{t}-1)+18.2^{t}$ for the cases $t\geq 1.$ Then, Singer’s conjecture for $m=5$ and these degrees will be discussed at the end of section two. We hope that our results would be necessary to formulate general solutions. ## 2 Statement of results Some notes. Throughout this paper, let us write $\begin{array}[]{ll}\vskip 6.0pt plus 2.0pt minus 2.0pt(\mathcal{P}_{m})_{n}&:=\langle\\{f\in\mathcal{P}_{m}|\,\mbox{$f$ is a homogeneous polynomial of degree $n$}\\}\rangle,\\\ Q^{\otimes m}_{n}&:=\langle\\{[f]\in Q^{\otimes m}|\,f\in(\mathcal{P}_{m})_{n}\\}\rangle,\end{array}$ which are $\mathbb{F}_{2}GL_{m}$-submodules of $\mathcal{P}_{m}$ and $Q^{\otimes m},$ respectively. So $\mathcal{P}_{m}=\bigoplus_{n\geq 0}(\mathcal{P}_{m})_{n}$ and $Q^{\otimes m}=\bigoplus_{n\geq 0}Q^{\otimes m}_{n}.$ Recall that to solve the hit problem of three variables, Kameko M.K constructed a $\mathbb{F}_{2}GL_{m}$-modules epimorphism: $\begin{array}[]{ll}(\widetilde{Sq^{0}_{*}})_{(m,m+2n)}:Q^{\otimes m}_{m+2n}&\longrightarrow Q^{\otimes m}_{n}\\\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \mbox{[}\prod_{1\leq j\leq m}x_{j}^{a_{j}}\mbox{]}&\longmapsto\left\\{\begin{array}[]{ll}\mbox{[}\prod_{1\leq j\leq m}x_{j}^{\frac{a_{j}-1}{2}}\mbox{]}&\text{if $a_{j}$ odd, $j=1,2,\ldots,m$},\\\ 0&\text{otherwise},\end{array}\right.\end{array}$ which induces the homomorphism $\widetilde{Sq^{0}_{*}}:(Q^{\otimes m}_{m+2n})^{GL_{m}}\to(Q^{\otimes m}_{n})^{GL_{m}}.$ Since $\mathscr{A}$ is a cocommutative Hopf algebra, there exists the squaring operations $Sq^{i}:{\rm Ext}_{\mathscr{A}}^{m,m+n}(\mathbb{F}_{2},\mathbb{F}_{2})\to{\rm Ext}_{\mathscr{A}}^{m+i,2m+2n}(\mathbb{F}_{2},\mathbb{F}_{2}),$ which share most of the properties with $Sq^{i}$ on the cohomology of spaces (see May ), but the classical $Sq^{0}$ is not the identity in general. Remarkably, this $Sq^{0}$ commutes with the dual of $\widetilde{Sq^{0}_{*}}$ through the Singer transfer (see J.B , N.M2 ). The reader who is familiar with Kameko’s $(\widetilde{Sq^{0}_{*}})_{(m,m+2n)}$ will probably agree that this map is very useful in solving the hit problem. Indeed, Kameko M.K showed that if $m=\xi(n)={\rm min}\\{\gamma\in\mathbb{N}:\ n=\sum_{1\leq i\leq\gamma}(2^{d_{i}}-1),\,d_{i}>0,\forall i,\,1\leq i\leq\gamma\\},$ then $(\widetilde{Sq^{0}_{*}})_{(m,m+2n)}$ is an isomorphism of $\mathbb{F}_{2}GL_{m}$-modules. This statement and Wood’s work R.W together are sufficient to determine $Q^{\otimes m}_{n}$ in each degree $n$ of the special ”generic” form $n=r(2^{t}-1)+d.2^{t},$ whenever $0<\xi(d)<r<m,$ and $t\geq 0$ (see also D.P6 ). As we mentioned at the beginning, the hit problem was completely solved for $m\leq 4.$ Very little information is known for $m=5$ and degrees $n$ given above. At least, it is surveyed by the present writer D.P6 for $(r,d,t)\in\\{(5,18,0),(5,8,t)\\}.$ We now extend for the case $(r,d,t)=(5,18,t),$ in which $t$ an arbitrary non-negative integer. We start with a useful remark. ###### Remark 2.1 It can be easily seen that $5(2^{t}-1)+18.2^{t}=2^{t+4}+2^{t+2}+2^{t+1}+2^{t-1}+2^{t-1}-5,$ and so $\xi(5(2^{t}-1)+18.2^{t})=5$ for any $t>1.$ This implies that the iterated Kameko map $((\widetilde{Sq^{0}_{*}})_{(5,5(2^{t}-1)+18.2^{t})})^{t-1}:Q^{\otimes 5}_{5(2^{t}-1)+18.2^{t}}\to Q^{\otimes 5}_{5(2^{1}-1)+18.2^{1}}$ is an isomorphism, for all $t\geq 1,$ and therefore, it is enough to determine $Q^{\otimes 5}_{5(2^{t}-1)+18.2^{t}}$ for $t\in\\{0,1\\}$. The case $t=0$ has explicitly been computed by us in D.P3 . When $t=1,$ because Kameko’s homomorphism $(\widetilde{Sq^{0}_{*}})_{(5,5(2^{1}-1)+18.2^{1})}:Q^{\otimes 5}_{5(2^{1}-1)+18.2^{1}}\to Q^{\otimes 5}_{5(2^{0}-1)+18.2^{0}}$ is an epimorphism, we have an isomorphism $Q^{\otimes 5}_{5(2^{1}-1)+18.2^{1}}\cong{\rm Ker}((\widetilde{Sq^{0}_{*}})_{(5,5(2^{1}-1)+18.2^{1})})\bigoplus Q^{\otimes 5}_{5(2^{0}-1)+18.2^{0}}.$ The space $Q^{\otimes 5}_{5(2^{0}-1)+18.2^{0}}$ is known by our previous work D.P3 . Thus, we need compute the kernel of $(\widetilde{Sq^{0}_{*}})_{(5,5(2^{1}-1)+18.2^{1})}.$ For this, our approach can be summarized as follows: * (i) A mononial in $\mathcal{P}_{5}$ is assigned a weight vector $\omega$ of degree $5(2^{1}-1)+18.2^{1}$, which stems from the binary expansion of the exponents of the monomial. The space of indecomposable elements ${\rm Ker}((\widetilde{Sq^{0}_{*}})_{(5,5(2^{1}-1)+18.2^{1})})$ is then decomposed into a direct sum of $(Q_{5(2^{1}-1)+18.2^{1}}^{\otimes 5})^{0}$ and the subspaces $(Q^{\otimes 5})^{\omega^{>0}}$ indexed by the weight vectors $\omega.$ Here $[F]_{\omega}=[G]_{\omega}$ in $(Q_{5(2^{1}-1)+18.2^{1}}^{\otimes 5})^{\omega}$ if the polynomial $F-G$ is hit, modulo a sum of monomials of weight vectors less than $\omega.$ Basing the previous results by Peterson F.P , Kameko M.K , Sum N.S1 , and by us D.P6 , one can easily determine $(Q_{5(2^{1}-1)+18.2^{1}}^{\otimes 5})^{0}.$ * (ii) The monomials in a given degree are lexicographically ordered first by weight vectors and then by exponent vectors. This leads to the concept of admissible monomial; more explicitly, a monomial is admissible if, modulo hit elements, it is not equal to a sum of monomials of smaller orders. The space $(Q_{5(2^{1}-1)+18.2^{1}}^{\otimes 5})^{\omega^{>0}}$ above is easily seen to be isomorphic to the space generated by admissible monomials of the weight vector $\omega.$ * (iii) In a given (small) degree, we first list all possible weight vectors of an admissible monomial. This is done by first using a criterion of Singer W.S1 on the hit monomials, and then combining with the results by Kameko M.K and Sum N.S1 of the form ”$XZ^{2^{r}}$ (or $ZY^{2^{t}}$) admissible implying $Z$ admissible, under some mild conditions”. * (iv) In a given weight vector, we claim the (strict) inadmissibility of some explicit monomials. The proof is given for a typical monomial in each case by explicit computations. Finally, a direct calculation using Theorems 3.2, 3.3, and some homomorphisms in section three, we obtain a basis of $(Q_{5(2^{1}-1)+18.2^{1}}^{\otimes 5})^{\omega^{>0}}.$ This approach is much less computational and it can be applied for all certain degrees and all variables $m.$ Moreover, the MAGMA computer algebra Magma has been used for verifying the results. Before going into detail and proceeding to the main results, let us provide some basic concepts. Of course, we assume that the reader is not familiar with the basics of hit problems. Weight vector and exponent vector. Let $\omega=(\omega_{1},\omega_{2},\ldots,\omega_{t},\ldots)$ be a sequence of non-negative integers. We say that $\omega$ is a weight vector, if $\omega_{t}=0$ for $t\gg 0.$ Then, we also define $\deg(\omega)=\sum_{t\geq 1}2^{t-1}\omega_{t}.$ Let $X=x_{1}^{u_{1}}x_{2}^{u_{2}}\ldots x_{m}^{u_{m}}$ be a mononial in $\mathcal{P}_{m},$ define two sequences associated with $X$ by $\omega(X):=(\omega_{1}(X),\omega_{2}(X),\ldots,\omega_{t}(X),\ldots),\ \ u(X):=(u_{1},u_{2},\ldots,u_{m}),$ where $\omega_{t}(X)=\sum_{1\leq j\leq m}\alpha_{t-1}(u_{j})$ in which $\alpha_{t}(n)$ denotes the $t$-th coefficients in dyadic expansion of a positive integer $n.$ They are called the weight vector and the exponent vector of $X,$ respectively. We use the convention that the sets of all the weight vectors and the exponent vectors are given the left lexicographical order. Linear order on $\mathcal{P}_{m}$. Assume that $X=x_{1}^{u_{1}}x_{2}^{u_{2}}\ldots x_{m}^{u_{m}}$ and $Y=x_{1}^{v_{1}}x_{2}^{v_{2}}\ldots x_{m}^{v_{m}}$ are the monomials of the same degree in $\mathcal{P}_{m}.$ We say that $X<Y$ if and only if one of the following holds: 1. (i) $\omega(X)<\omega(Y);$ 2. (ii) $\omega(X)=\omega(Y)$ and $u(X)<v(Y).$ Equivalence relations on $\mathcal{P}_{m}$. For a weight vector $\omega,$ we denote two subspaces associated with $\omega$ by $\begin{array}[]{ll}\mathcal{P}^{\leq\omega}_{m}=\langle\\{X\in\mathcal{P}_{m}|\,\deg(X)=\deg(\omega),\ \omega(X)\leq\omega\\}\rangle,\\\ \mathcal{P}^{<\omega}_{m}=\langle\\{X\in\mathcal{P}_{m}|\,\deg(X)=\deg(\omega),\ \omega(X)<\omega\\}\rangle.\end{array}$ Let $F$ and $G$ be the homogeneous polynomials in $\mathcal{P}_{m}$ such that $\deg(F)=\deg(G).$ We say that 1. (i) $F\equiv G$ if and only if $(F-G)\in\overline{\mathscr{A}}\mathcal{P}_{m}=\sum_{i\geq 0}{\rm Im}(Sq^{2^{i}}).$ Specifically, if $F\equiv 0,$ then $F$ is hit (or $\mathscr{A}$-decomposable), i.e., $F$ can be written in the form $\sum_{i\geq 0}Sq^{2^{i}}(F_{i})$ for some $F_{i}\in\mathcal{P}_{m}$; 2. (ii) $F\equiv_{\omega}G$ if and only if $F,\,G\in\mathcal{P}^{\leq\omega}_{m}$ and $(F-G)\in((\overline{\mathscr{A}}\mathcal{P}_{m}\cap\mathcal{P}_{m}^{\leq\omega})+\mathcal{P}_{m}^{<\omega}).$ It is not difficult to show that the binary relations ”$\equiv$” and ”$\equiv_{\omega}$” are equivalence ones. So, one defines the quotient space $(Q^{\otimes m})^{\omega}=\mathcal{P}_{m}^{\leq\omega}/((\overline{\mathscr{A}}\mathcal{P}_{m}\cap\mathcal{P}_{m}^{\leq\omega})+\mathcal{P}_{m}^{<\omega}).$ Moreover, due to Sum N.S3 , $(Q^{\otimes m})^{\omega}$ is also an $\mathbb{F}_{2}GL_{m}$-module. Admissible monomial and inadmissible monomial. A monomial $X\in\mathcal{P}_{m}$ is said to be inadmissible if there exist monomials $Y_{1},Y_{2},\ldots,Y_{k}$ such that $Y_{j}<X$ for $1\leq j\leq k$ and $X\equiv\sum_{1\leq j\leq k}Y_{j}.$ Then, $X$ is said to be admissible if it is not inadmissible. Thus, with the above definitions in hand, it is straightforward to see that the set of all the admissible monomials of degree $n$ in $\mathcal{P}_{m}$ is a minimal set of $\mathscr{A}$-generators for $\mathcal{P}_{m}$ in degree $n.$ So, $Q_{n}^{\otimes m}$ is a $\mathbb{F}_{2}$-vector space with a basis consisting of all the classes represent by the admissible monomials of degree $n$ in $\mathcal{P}_{m}.$ Further, as stated in D.P2 , the dimension of $Q_{n}^{\otimes m}$ can be represented as the sum of the dimensions $(Q^{\otimes m})^{\omega}$ such that $\deg(\omega)=n.$ For later convenience, we need to set some notation. Let $\mathcal{P}^{0}_{m}$ and $\mathcal{P}^{>0}_{m}$ denote the $\mathscr{A}$-submodules of $\mathcal{P}_{m}$ spanned all the monomials $\prod_{1\leq j\leq m}x_{j}^{t_{j}}$ such that $\prod_{1\leq j\leq m}t_{j}=0,$ and $\prod_{1\leq j\leq m}t_{j}>0,$ respectively. Let us write $(Q^{\otimes m})^{0}:=\mathbb{F}_{2}\otimes_{\mathscr{A}}\mathcal{P}^{0}_{m},\ \mbox{and}\ (Q^{\otimes m})^{>0}:=\mathbb{F}_{2}\otimes_{\mathscr{A}}\mathcal{P}^{>0}_{m},$ from which one has that $Q^{\otimes m}=(Q^{\otimes m})^{0}\,\bigoplus\,(Q^{\otimes m})^{>0}.$ For a polynomial $F\in\mathcal{P}_{m},$ we denote by $[F]$ the classes in $Q^{\otimes m}$ represented by $F.$ If $\omega$ is a weight vector and $F\in\mathcal{P}_{m}^{\leq\omega},$ then denote by $[F]_{\omega}$ the classes in $(Q^{\otimes m})^{\omega}$ represented by $F.$ For a subset $\mathscr{C}\subset\mathcal{P}_{m},$ we also write $|\mathscr{C}|$ for the cardinal of $\mathscr{C}$ and put $[\mathscr{C}]=\\{[F]\,:\,F\in\mathscr{C}\\}.$ If $\mathscr{C}\subset P_{m}^{\leq\omega},$ then put $[\mathscr{C}]_{\omega}=\\{[F]_{\omega}\,:\,F\in\mathscr{C}\\}.$ Let us denote by $\mathscr{C}^{\otimes m}_{n}$ the set of all admissible monomials of degree $n$ in $\mathcal{P}_{m},$ and let $\omega$ be a weight vector of degree $n.$ By setting $\begin{array}[]{ll}(\mathscr{C}^{\otimes m}_{n})^{\omega}:=\mathscr{C}^{\otimes m}_{n}\cap\mathcal{P}_{m}^{\leq\omega},\ \ (\mathscr{C}^{\otimes m}_{n})^{\omega^{0}}:=(\mathscr{C}^{\otimes m}_{n})^{\omega}\cap\mathcal{P}^{0}_{m},\ \ (\mathscr{C}^{\otimes m}_{n})^{\omega^{>0}}:=(\mathscr{C}^{\otimes m}_{n})^{\omega}\cap\mathcal{P}^{>0}_{m},\\\ (Q_{n}^{\otimes m})^{\omega^{0}}:=(Q^{\otimes m})^{\omega}\cap(Q^{\otimes m}_{n})^{0},\ \ (Q_{n}^{\otimes m})^{\omega^{>0}}:=(Q^{\otimes m})^{\omega}\cap(Q^{\otimes m}_{n})^{>0},\end{array}$ then the sets $[(\mathscr{C}^{\otimes m}_{n})^{\omega}]_{\omega},\,[(\mathscr{C}^{\otimes m}_{n})^{\omega^{0}}]_{\omega}$ and $[(\mathscr{C}^{\otimes m}_{n})^{\omega^{>0}}]_{\omega}$ are the bases of the $\mathbb{F}_{2}$-vector spaces $(Q_{n}^{\otimes m})^{\omega},\ (Q_{n}^{\otimes m})^{\omega^{0}}$ and $(Q_{n}^{\otimes m})^{\omega^{>0}},$ respectively. Main results and applications. Let us now return to our study of the kernel of the Kameko homomorphism $(\widetilde{Sq^{0}_{*}})_{(5,5(2^{1}-1)+18.2^{1})}$ and state our main results in greater detail. Firstly, by direct calculations using the results by Kameko M.K , Singer W.S1 , Sum N.S1 , and Tín N.T , we obtain the following, which is one of our main results and is crucial for an application on the dimension of $Q^{\otimes 6}.$ ###### Theorem 2.2 We have an isomorphism $\mbox{\rm Ker}(\widetilde{Sq_{*}^{0}})_{(5,5(2^{1}-1)+18.2^{1})}\cong(Q_{5(2^{1}-1)+18.2^{1}}^{\otimes 5})^{0}\bigoplus(Q_{5(2^{1}-1)+18.2^{1}}^{\otimes 5})^{\widetilde{\omega}^{>0}},$ where $\widetilde{\omega}=(3,3,2,1,1)$ is the weight vector of the degree $5(2^{1}-1)+18.2^{1}.$ ###### Remark 2.3 We are given in D.P6 that $(Q_{n}^{\otimes 5})^{0}\cong\bigoplus_{1\leq s\leq 4}\bigoplus_{\ell(\mathcal{J})=s}(Q_{n}^{\otimes\,\mathcal{J}})^{>0},$ where $Q^{\otimes\mathcal{J}}=\langle[x_{j_{1}}^{t_{1}}x_{j_{2}}^{t_{2}}\ldots x_{j_{s}}^{t_{s}}]\;|\;t_{i}\in\mathbb{N},\,i=1,2,\ldots,s\\}\rangle\subset Q^{\otimes 5}$ with $\mathcal{J}=(j_{1},j_{2},\ldots,j_{s}),$ $1\leq j_{1}<\ldots<j_{s}\leq 5$, $1\leq s\leq 4,$ and $\ell(\mathcal{J}):=s$ denotes the length of $\mathcal{J}.$ This implies that $\dim((Q_{n}^{\otimes 5})^{0}))=\sum_{1\leq s\leq 4}\binom{5}{s}\dim((Q_{n}^{\otimes\,s})^{>0}),\ \mbox{for all $n\geq 0.$}$ On the other side, since $\xi(5(2^{1}-1)+18.2^{1})=3,$ by Peterson F.P and Wood R.W , the spaces $Q_{5(2^{1}-1)+18.2^{1}}^{\otimes 1}$ and $Q_{5(2^{1}-1)+18.2^{1}}^{\otimes 2}$ are trivial. Moreover, following Kameko M.K and Sum N.S1 , we have seen that $(Q_{5(2^{1}-1)+18.2^{1}}^{\otimes 3})^{>0}$ is $15$-dimensional and that $(Q_{5(2^{1}-1)+18.2^{1}}^{\otimes 4})^{>0}$ is $165$-dimensional. Therefore, we may conclude that $\dim((Q^{\otimes 5}_{5(2^{1}-1)+18.2^{1}})^{0}=15.\binom{5}{3}+165.\binom{5}{4}=975.$ Next, due to Remarks 2.1, 2.3, and to Theorem 2.2, the space $Q^{\otimes 5}_{5(2^{1}-1)+18.2^{1}}$ will be determined by computing $(Q_{5(2^{1}-1)+18.2^{1}}^{\otimes 5})^{\omega^{>0}}.$ To accomplish this, we use the method described above to explicitly indicate all the admissible monomials in the set $(\mathscr{C}^{\otimes 5}_{5(2^{1}-1)+18.2^{1}})^{\widetilde{\omega}^{>0}}.$ As a result, it reads as follows. ###### Theorem 2.4 There exist exactly $925$ admissible monomials of degree $5(2^{1}-1)+18.2^{1}$ in $\mathcal{P}_{5}^{>0}$ such that their weight vectors are $\widetilde{\omega}.$ Consequently, $(Q_{5(2^{1}-1)+18.2^{1}}^{\otimes 5})^{\widetilde{\omega}^{>0}}$ has dimension $925.$ This theorem, together with the fact that $Q_{5(2^{t}-1)+18.2^{t}}^{\otimes 5}=(Q_{5(2^{t}-1)+18.2^{t}}^{\otimes 5})^{0}\,\bigoplus\,(Q_{5(2^{t}-1)+18.2^{t}}^{\otimes 5})^{>0},$ yields an immediate corollary that ###### Corollary 2.5 The space $Q_{5(2^{t}-1)+18.2^{t}}^{\otimes 5}$ is $730$-dimensional if $t=0,$ and is $2630$-dimensional if $t\geq 1.$ As applications, one would also be interested in applying results and techniques of hit problems into the cases of higher ranks $m$ of $Q^{\otimes m}$ and the modular representations of the general linear groups (see also the relevant discussions in literatures J.B , N.M1 ; N.M2 , T.N2 , W.W ; W.W2 ). Two applications below of the contributions of this paper are also not beyond this target. First application: the dimension of $Q^{\otimes 6}.$ The hit problem of six variables has been not yet known. Using Corollary 2.5 for the case $t\geq 1$ and a result in Sum N.S1 , we state that ###### Theorem 2.6 With the generic degree $5(2^{t+4}-1)+41.2^{t+4},$ where $t$ an arbitrary positive integer, then the $\mathbb{F}_{2}$-vector space $Q^{\otimes 6}$ has dimension $165690$ in this degree. Observing from Corollary 2.5 and Theorem 2.6, the readers can notice that the dimensions of $Q^{\otimes 5}$ and $Q^{\otimes 6}$ in degrees given are very large. So, a general approach to hit problems, other than providing a mononial basis of the vector space $Q_{n}^{\otimes m},$ is to find upper/lower bounds on the dimension of this space. However, in this work, we have not studied this side of the problem and it is our concern the next time. It is remarkable that, we have Kameko’s conjecture M.K on an upper bound for the dimension of $Q_{n}^{\otimes m},$ but unfortunately, it was refuted for $m\geq 5$ by the brilliant work of Sum N.S . Second application: the behavior of the fifth Singer transfer. We adopt Corollary 2.5 for $t=0,$ together with a fact of the Adams $E^{2}$-term, ${\rm Ext}_{\mathscr{A}}^{5,5+*}(\mathbb{F}_{2},\mathbb{F}_{2})$, to obtain information about the behavior of Singer’s cohomological transfer in the bidegree $(5,5+(5(2^{0}-1)+18.2^{0}))$. More precisely, it is known, the calculations of Lin W.L , and Chen T.C imply that ${\rm Ext}_{\mathscr{A}}^{5,5+(5(2^{t}-1)+18.2^{t})}(\mathbb{F}_{2},\mathbb{F}_{2})=\langle h_{t}f_{t}\rangle$ and $h_{t}f_{t}=h_{t+1}e_{t}\neq 0$ for all $t\geq 0.$ So, to determine the transfer map in the above bidegree, we shall compute the dimension of (the domain of the fifth transfer) $(\mathbb{F}_{2}\otimes_{GL_{5}}P_{\mathscr{A}}((\mathcal{P}_{5})^{*}))_{5(2^{0}-1)+18.2^{0}}$ by using a mononial basis of $Q^{\otimes 5}_{5(2^{0}-1)+18.2^{0}}.$ (We emphasize that computing the domain of $Tr_{m}^{\mathscr{A}}$ in each degree $n$ is very difficult, particularly for values of $m$ as large as $m=5.$ The understanding of special cases should be a helpful step toward the solution of the general problem. Moreover, we believe, in principle, that our method could lead to a full analysis of $\mathbb{F}_{2}\otimes_{GL_{m}}P_{\mathscr{A}}((\mathcal{P}_{m})^{*})$ in each $m$ and degree $n>0$, as long as nice decompositions of the space of $GL_{m}$-invariants of $Q^{\otimes m}$ in degrees given. However, the difficulty of such a task must be monumental, as $Q^{\otimes m}$ becomes much larger and harder to understand with increasing $m.$) Details for this application are as follows. It may need to be recalled that by the previous discussions D.P3 , we get the technical result below. ###### Proposition 2.7 The following hold: 1. i) If $Y\in\mathscr{C}_{5(2^{0}-1)+18.2^{0}}^{\otimes 5},$ then $\overline{\omega}:=\omega(Y)$ is one of the following sequences: $\overline{\omega}_{[1]}:=(2,2,1,1),\ \ \overline{\omega}_{[2]}:=(2,2,3),\ \ \overline{\omega}_{[3]}:=(2,4,2),$ $\overline{\omega}_{[4]}:=(4,1,1,1),\ \ \overline{\omega}_{[5]}:=(4,1,3),\ \ \overline{\omega}_{[6]}:=(4,3,2).$ 2. ii) $|(\mathscr{C}^{\otimes 5}_{5(2^{0}-1)+18.2^{0}})^{\overline{\omega}_{[k]}}|=\left\\{\begin{array}[]{ll}300&\mbox{{\rm if}}\ k=1,\\\ 15&\mbox{{\rm if}}\ k=2,5,\\\ 10&\mbox{{\rm if}}\ k=3,\\\ 110&\mbox{{\rm if}}\ k=4,\\\ 280&\mbox{{\rm if}}\ k=6.\end{array}\right.$ One should note that $|(\mathscr{C}^{\otimes 5}_{5(2^{0}-1)+18.2^{0}})^{\overline{\omega}_{[k]}}|=|(\mathscr{C}^{\otimes 5}_{5(2^{0}-1)+18.2^{0}})^{\overline{\omega}_{[k]}^{>0}}|$ for $k=2,3$, and that $|(\mathscr{C}^{\otimes 5}_{5(2^{0}-1)+18.2^{0}})^{\overline{\omega}_{[2]}^{0}}|=0=|(\mathscr{C}^{\otimes 5}_{5(2^{0}-1)+18.2^{0}})^{\overline{\omega}_{[3]}^{0}}|.$ Moreover, $\dim(Q^{\otimes 5}_{5(2^{0}-1)+18.2^{0}})=\sum_{1\leq k\leq 6}|(\mathscr{C}^{\otimes 5}_{5(2^{0}-1)+18.2^{0}})^{\overline{\omega}_{[k]}}|=730.$ Next, applying these results, we explicitly compute the subspaces of $GL_{5}$-invariants $((Q_{5(2^{0}-1)+18.2^{0}}^{\otimes 5})^{\overline{\omega}_{[k]}})^{GL_{5}},$ for $1\leq k\leq 6,$ and obtain ###### Theorem 2.8 The following assertions are true: 1. i) $((Q_{5(2^{0}-1)+18.2^{0}}^{\otimes 5})^{\overline{\omega}_{[k]}})^{GL_{5}}=0$ with $k\in\\{1,2,3,5,6\\}.$ 2. ii) $((Q_{5(2^{0}-1)+18.2^{0}}^{\otimes 5})^{\overline{\omega}_{[4]}})^{GL_{5}}=\langle[\Re^{\prime}_{4}]_{\overline{\omega}_{[4]}}\rangle,$ where $\begin{array}[]{ll}\vskip 6.0pt plus 2.0pt minus 2.0pt\Re^{\prime}_{4}&=x_{1}x_{2}x_{3}x_{4}x_{5}^{14}+x_{1}x_{2}x_{3}x_{4}^{14}x_{5}+x_{1}x_{2}x_{3}^{14}x_{4}x_{5}+x_{1}x_{2}^{3}x_{3}x_{4}x_{5}^{12}\\\ \vskip 6.0pt plus 2.0pt minus 2.0pt&\quad+x_{1}x_{2}^{3}x_{3}x_{4}^{12}x_{5}+x_{1}x_{2}^{3}x_{3}^{12}x_{4}x_{5}+x_{1}^{3}x_{2}x_{3}x_{4}x_{5}^{12}+x_{1}^{3}x_{2}x_{3}x_{4}^{12}x_{5}\\\ \vskip 6.0pt plus 2.0pt minus 2.0pt&\quad+x_{1}^{3}x_{2}x_{3}^{12}x_{4}x_{5}+x_{1}^{3}x_{2}^{5}x_{3}x_{4}x_{5}^{8}+x_{1}^{3}x_{2}^{5}x_{3}x_{4}^{8}x_{5}+x_{1}^{3}x_{2}^{5}x_{3}^{8}x_{4}x_{5}.\end{array}$ Now, because $(\mathbb{F}_{2}\otimes_{GL_{5}}P_{\mathscr{A}}((\mathcal{P}_{5})^{*}))_{5.(2^{0}-1)+18.2^{0}}$ is isomorphic to $(Q^{\otimes 5}_{5.(2^{0}-1)+18.2^{0}})^{GL_{5}},$ by Theorem 2.8, we have the following estimate: $\begin{array}[]{ll}\dim(\mathbb{F}_{2}\otimes_{GL_{5}}P_{\mathscr{A}}((\mathcal{P}_{5})^{*}))_{5.(2^{0}-1)+18.2^{0}}&=\dim(Q^{\otimes 5}_{5.(2^{0}-1)+18.2^{0}})^{GL_{5}}\\\ &\leq\sum_{1\leq k\leq 6}\dim((Q_{5(2^{0}-1)+18.2^{0}}^{\otimes 5})^{\overline{\omega}_{[k]}})^{GL_{5}}\leq 1.\end{array}$ On the other side, as shown in section one, $\\{h_{t}|\,t\geq 0\\}\subset{\rm Im}(Tr^{\mathscr{A}}_{1}),$ and $\\{f_{t}|\,t\geq 0\\}\subset{\rm Im}(Tr^{\mathscr{A}}_{4}).$ Combining this with the fact that the total transfer $\bigoplus_{m\geq 0}Tr_{m}^{\mathscr{A}}$ is a homomorphism of algebras, it may be concluded that the non-zero element $h_{t}f_{t}\in{\rm Ext}_{\mathscr{A}}^{5,23.2^{t}}(\mathbb{F}_{2},\mathbb{F}_{2})$ is in the image of $Tr^{\mathscr{A}}_{5}$ for all $t\geq 0.$ This could be directly proved as in Appendix. This statement implies that $\dim(\mathbb{F}_{2}\otimes_{GL_{5}}P_{\mathscr{A}}((\mathcal{P}_{5})^{*}))_{5.(2^{0}-1)+18.2^{0}}\geq 1,$ and therefore $(\mathbb{F}_{2}\otimes_{GL_{5}}P_{\mathscr{A}}((\mathcal{P}_{5})^{*}))_{5.(2^{0}-1)+18.2^{0}}$ is one-dimensional. As a consequence, we immediately obtain ###### Corollary 2.9 The cohomological transfer $Tr^{\mathscr{A}}_{5}:(\mathbb{F}_{2}\otimes_{GL_{5}}P_{\mathscr{A}}((\mathcal{P}_{5})^{*}))_{5(2^{0}-1)+18.2^{0}}\to{\rm Ext}_{\mathscr{A}}^{5,5+5(2^{0}-1)+18.2^{0}}(\mathbb{F}_{2},\mathbb{F}_{2})$ is an isomorphism. Consequently, Conjecture 1.1 holds in the rank 5 case and the degree $5(2^{0}-1)+18.2^{0}.$ Comments and open issues. From the above results, it would be interesting to see that $Q^{\otimes 5}$ is $730$-dimensional in degree $5(2^{0}-1)+18.2^{0},$ but the space of $GL_{5}$-coinvariants of it in this degree is only one- dimensional. In general, it is quite efficient in using the results of the hit problem of five variables to study $\mathbb{F}_{2}\otimes_{GL_{5}}P_{\mathscr{A}}((\mathcal{P}_{5})^{*}).$ This provides a valuable method for verifying Singer’s open conjecture on the fifth algebraic transfer. We now close the introduction by discussing about Conjecture 1.1 in the rank 5 case and the internal degree $n_{t}:=5(2^{t}-1)+18.2^{t}$ for all $t\geq 1.$ Let us note again that the iterated Kameko homomorphism $((\widetilde{Sq^{0}_{*}})_{(5,n_{t})})^{t-1}:Q^{\otimes 5}_{n_{t}}\to Q^{\otimes 5}_{n_{1}}$ is an $\mathbb{F}_{2}GL_{5}$-module isomorphism for all $t\geq 1.$ So, from a fact of ${\rm Ext}_{\mathscr{A}}^{5,5+n_{1}}(\mathbb{F}_{2},\mathbb{F}_{2})$, to check Singer’s conjecture in the above degree, we need only determine $GL_{5}$-coinvariants of $Q^{\otimes 5}_{n_{t}}$ for $t=1.$ We must recall that Kameko’s map $(\widetilde{Sq^{0}_{*}})_{(5,n_{1})}:Q^{\otimes 5}_{n_{1}}\to Q^{\otimes 5}_{n_{0}}$ is an epimorphism of $GL_{5}$-modules. On the other side, as shown before, the non-zero element $h_{1}f_{1}\in{\rm Ext}_{\mathscr{A}}^{5,5+n_{1}}(\mathbb{F}_{2},\mathbb{F}_{2})$ is detected by the fifth transfer. From these data and Theorem 2.8, one has an estimate $0\leq\dim((\mathbb{F}_{2}\otimes_{GL_{5}}P_{\mathscr{A}}((\mathcal{P}_{5})^{*}))_{n_{1}})-1\leq\dim({\rm Ker}(\widetilde{Sq^{0}_{*}})_{(5,n_{1})})^{GL_{5}}.$ Moreover, basing the proof of Theorem 2.8 together with a few simple arguments, it follows that the elements in $(\mathbb{F}_{2}\otimes_{GL_{5}}P_{\mathscr{A}}((\mathcal{P}_{5})^{*}))_{n_{1}}$ are dual to the classes $\begin{array}[]{ll}\vskip 6.0pt plus 2.0pt minus 2.0pt&\gamma[x_{1}^{3}x_{2}^{3}x_{3}^{3}x_{4}^{3}x_{5}^{29}+x_{1}^{3}x_{2}^{3}x_{3}^{3}x_{4}^{29}x_{5}^{3}+x_{1}^{3}x_{2}^{3}x_{3}^{29}x_{4}^{3}x_{5}^{3}+x_{1}^{3}x_{2}^{7}x_{3}^{3}x_{4}^{3}x_{5}^{25}+x_{1}^{3}x_{2}^{7}x_{3}^{3}x_{4}^{25}x_{5}^{3}\\\ \vskip 6.0pt plus 2.0pt minus 2.0pt&\quad+x_{1}^{3}x_{2}^{7}x_{3}^{25}x_{4}^{3}x_{5}^{3}+x_{1}^{7}x_{2}^{3}x_{3}^{3}x_{4}^{3}x_{5}^{25}+x_{1}^{7}x_{2}^{3}x_{3}^{3}x_{4}^{25}x_{5}^{3}+x_{1}^{7}x_{2}^{3}x_{3}^{25}x_{4}^{3}x_{5}^{3}+x_{1}^{7}x_{2}^{11}x_{3}^{3}x_{4}^{3}x_{5}^{17}\\\ &\quad+x_{1}^{7}x_{2}^{11}x_{3}^{3}x_{4}^{17}x_{5}^{3}+x_{1}^{7}x_{2}^{11}x_{3}^{17}x_{4}^{3}x_{5}^{3}]+[\zeta],\end{array}$ where $\gamma\in\mathbb{F}_{2},$ and $[\zeta]\in\text{Ker}(\widetilde{Sq^{0}_{*}})_{(5,n_{1})}.$ It could be noticed that calculating explicitly these elements is not easy. However, in view of our previous works D.P2 ; D.P6 , and motivated by the above computations, we have the following prediction. ###### Conjecture 2.10 For each $t\geq 1,$ the space of $GL_{5}$-invariants elements of ${\rm Ker}(\widetilde{Sq^{0}_{*}})_{(5,n_{t})}$ is trivial. Consequently, the coinvariant $(\mathbb{F}_{2}\otimes_{GL_{5}}P_{\mathscr{A}}((\mathcal{P}_{5})^{*}))_{n_{t}}$ is 1-dimensional. Since $h_{t}f_{t}\in{\rm Im}(Tr_{5}^{\mathscr{A}}),$ for all $t\geq 0,$ if Conjecture 2.10 is true, then $Tr_{5}^{\mathscr{A}}$ is also isomorphism when acting on the coinvariant $(\mathbb{F}_{2}\otimes_{GL_{5}}P_{\mathscr{A}}((\mathcal{P}_{5})^{*}))_{n_{t}}$ for $t\geq 1,$ and so, Conjecture 1.1 holds in bidegree $(5,5+n_{t})$. We also wish that our predictions are correct. If not, Singer’s conjecture will be disproved. We leave these issues as future research. At the same time, we also appreciate that some readers may have an interest in solving them. Overview. Let us give a brief outline of the contents of this paper. Section three contains a brief review of Steenrod squares and some useful linear transformations. The dimensions of the polynomial algebras $\mathcal{P}_{5}$ and $\mathcal{P}_{6}$ in the generic degrees $n_{t}=5(2^{t}-1)+18.2^{t}$ and $5(2^{t+4}-1)+n_{1}.2^{t+4}$ are respectively obtained in section four by proving Theorems 2.2, 2.4, and 2.6. Section five is to present the proof of Theorem 2.8. In the remainder of the text, we give a direct proof of an event claimed above that the non-zero elements $h_{t}f_{t}\in{\rm Ext}_{\mathscr{A}}^{5,23.2^{t}}(\mathbb{F}_{2},\mathbb{F}_{2})$ are detected by $Tr_{5}^{\mathscr{A}}$. The proof is based on a representation in the lambda algebra of the fifth Singer transfer. Finally, we describe the set $(\mathscr{C}_{n_{1}}^{\otimes 5})^{\widetilde{\omega}^{>0}}$ and list some the admissible monomials in $\mathscr{C}_{n_{0}}^{\otimes 5}$ and the strictly inadmissible monomials in ($\mathcal{P}_{5}^{>0})_{n_{1}}.$ ###### Acknowledgment The author would like to give my deepest sincere thanks to Professor N. Sum (Quy Nhon University, Sai Gon University), my thesis advisor, for meaningful discussions. I am very grateful to Professor W. Singer for many enlightening e-mail exchanges. ## 3 The Necessary Preliminaries This section begins with a few words on the Steenrod algebra over $\mathbb{F}_{2}$ and ends with a brief sketch of some homomorphisms in N.S1 . At the same time, we prove some elementary results that will be used in the rest of this text. ### 3.1 Steenrod squares and their properties The mod 2 Steenrod algebra $\mathscr{A}$ was defined by Cartan Cartan to be the algebra of stable cohomology operations for mod 2 cohomology. This algebra is generated by the Steenrod squares $Sq^{i}:H^{n}(X,\mathbb{F}_{2})\to H^{n+i}(X,\mathbb{F}_{2}),$ for $i\geq 0,$ where $H^{n}(X,\mathbb{F}_{2})$ denotes the $n$-th singular cohomology group of a topological space $X$ with coefficient over $\mathbb{F}_{2}.$ Steenrod and Epstein S.E showed that these squares are characterized by the following 5 axioms: 1. (i) $Sq^{i}$ is an additive homomorphism and is natural with respect to any $f:X\to Y.$ So $f^{*}(Sq^{i}(x))=Sq^{i}(f^{*}(x)).$ 2. (ii) $Sq^{0}$ is the identity homomorphism. 3. (iii) $Sq^{i}(x)=x\smile x$ for all $x\in H^{i}(X,\mathbb{F}_{2})$ where $\smile$ denotes the cup product in the graded cohomology ring $H^{*}(X,\mathbb{F}_{2}).$ 4. (iv) If $i>\deg(x),$ then $Sq^{i}(x)=0.$ 5. (v) Cartan’s formula: $Sq^{n}(x\smile y)=\sum_{i+j=n}Sq^{i}(x)\smile Sq^{j}(y).$ In addition, Steenrod squares have the following properties: $\bullet$ $Sq^{1}$ is the Bockstein homomorphism of the coefficient sequence: $0\to\mathbb{Z}/2\to\mathbb{Z}/4\to\mathbb{Z}/2\to 0.$ $\bullet$ $Sq^{i}$ commutes with the connecting morphism of the long exact sequence in cohomology. In particular, it commutes with respect to suspension $H^{n}(X,\mathbb{F}_{2})\cong H^{n+1}(\Sigma X,\mathbb{F}_{2}).$ $\bullet$ They satisfy the Adem relations: $Sq^{i}Sq^{j}=\sum_{0\leq t\leq[i/2]}\binom{j-t-1}{i-2t}Sq^{i+j-t}Sq^{t},\,\,0<i<2j,$ where the binomial coefficients are to be interpreted mod 2. These relations, which were conjectured by Wu Wu and established by Adem Adem , allow one to write an arbitrary composition of Steenrod squares as a sum of Serre-Cartan basis elements. Note that the structure of the cohomology $H^{*}(X,\mathbb{F}_{2})$ is not only as graded commutative $\mathbb{F}_{2}$-algebra, but also as an $\mathscr{A}$-module. In many cases, the $\mathscr{A}$-module structure on $H^{*}(X,\mathbb{F}_{2})$ provides additional information on $X.$ ## References * (1) J. Adem, The iteration of the Steenrod squares in Algebraic Topology, Proc. Natl. Acad. Sci. USA 38 (1952), 20-726. * (2) J.M. Boardman, Modular representations on the homology of power of real projective space, in Algebraic Topology: Oaxtepec 1991, ed. M. C. Tangor; in Contemp. Math. 146 (1993), 49-70. * (3) A.K. Bousfield, E.B. Curtis, D.M. Kan, D.G. Quillen, D.L. Rector, and J.W. Schlesinger, The mod-$p$ lower central series and the Adams spectral sequence, Topology 5 (1966), 331-342. * (4) R.R. Bruner, L.M. Hà, and N.H.V. Hưng, On behavior of the algebraic transfer, Trans. Amer. Math. Soc. 357 (2005), 437-487. * (5) M. Brunetti, A. Ciampella, and A.L. Lomonaco, A total Steenrod operation as homomorphism of Steenrod algebra-modules, Ric. Mat. 61 (2012), 1-17. * (6) M. Brunetti, and A.L. Lomonaco, A representation of the dual of the Steenrod algebra, Ric. Mat. 63 (2014), 19-24. * (7) H. Cartan, Sur l’itération des opérations de Steenrod, Comment. Math. Helv. 29 (1955), 40-58. * (8) T.W. Chen, Determination of $\mbox{Ext}^{5,*}_{\mathscr{A}}(\mathbb{Z}/2,\mathbb{Z}/2)$, Topol. Appl. 158 (2011), 660-689. * (9) P.H. Chơn, and L.M. Hà, Lambda algebra and the Singer transfer, C. R. Math. Acad. Sci. Paris 349 (2011), 21-23. * (10) M.C. Crabb, and J.R. Hubbuck, Representations of the homology of BV and the Steenrod algebra II, in Algebra Topology: New trend in localization and periodicity; in Progr. Math. 136 (1996), 143-154. * (11) L.M. Hà, Sub-Hopf algebras of the Steenrod algebra and the Singer transfer, Geom. Topol. Monogr. 11 (2007), 101-124. * (12) A. Hatcher, Algebraic Topology, Cambridge University Press, 2002, 551 pp. * (13) M. A. Hill, M. J. Hopkins, and D. C. Ravenel, On the non-existence of elements of kervaire invariant one, Ann. of Math. (2) 184 (2016), 1-262. * (14) N.H.V. Hưng, The cohomology of the Steenrod algebra and representations of the general linear groups, Trans. Amer. Math. Soc. 357 (2005), 4065-4089. * (15) N.H.V. Hưng and V.T.N. Quỳnh, The image of Singer’s fourth transfer, C. R. Math. Acad. Sci. Paris 347 (2009), 1415-1418. * (16) M. Inoue, $\mathcal{A}$-generators of the cohomology of the steinberg summand $M(n)$, In: D.M. Davis, J. Morava, G. Nishida, W. S. Wilson and N. Yagita (eds.) Recent Progress in Homotopy Theory (Baltimore, MD, 2000). Contemporary Mathematics, vol. 293, pp 125-139. American Mathematical Society, Providence (2002). * (17) M. Inoue, Generators of the cohomology of $M(n)$ as a module over the odd primary Steenrod algebra, J. Lond. Math. Soc. (2) 75 (2007), 317-329. * (18) A.S. Janfada, and R.M.W. Wood, The hit problem for symmetric polynomials over the Steenrod algebra, Math. Proc. Cambridge Philos. Soc. 133 (2002), 295-303. * (19) A.S. Janfada, and R.M.W. Wood, Generating $H^{*}(BO(3),\mathbb{F}_{2})$ as a module over the Steenrod algebra, Math. Proc. Cambridge Philos. Soc. 134 (2003), 239-258. * (20) A.S. Janfada, A criterion for a monomial in $P(3)$ to be hit, Math. Proc. Cambridge Philos. Soc. 145 (2008), 587-599. * (21) A.S. Janfada, A note on the unstability conditions of the Steenrod squares on the polynomial algebra, J. Korean Math. Soc. 46 (2009), 907-918. * (22) M. Kameko, Products of projective spaces as Steenrod modules, PhD. thesis, The Johns Hopkins University, ProQuest LLC, Ann Arbor, MI, 1990, 29 pages. * (23) W.H. Lin, ${\rm Ext}_{\mathcal{A}}^{4,*}(\mathbb{Z}/2,\mathbb{Z}/2)\mbox{ {\it and} }{\rm Ext}_{\mathcal{A}}^{5,*}(\mathbb{Z}/2,\mathbb{Z}/2)$, Topol. Appl. 155 (2008), 459-496. * (24) Magma Computational Algebra System (V2.25-8), the Computational Algebra Group at the University of Sydney, (2020), http://magma.maths.usyd.edu.au/magma/. * (25) J.P. May, A General Algebraic Approach to Steenrod Operations, Lect. Notes Math., vol. 168, Springer-Verlag (1970), 153-231. * (26) N. Minami, The Adams spectral sequence and the triple transfer, Amer. J. Math. 117 (1995), 965-985. * (27) N. Minami, The iterated transfer analogue of the new doomsday conjecture, Trans. Amer. Math. Soc. 351 (1999), 2325-2351. * (28) J.W. Milnor, The Steenrod algebra and its dual, Ann. of Math. (2) 67 (1958), 150-171. * (29) M.F. Mothebe, and L. Uys, Some relations between admissible monomials for the polynomial algebra, Int. J. Math. Math. Sci., Article ID 235806, 2015, 7 pages. * (30) M.F. Mothebe, _The admissible monomial basis for the polynomial algebra in degree thirteen_ , East-West J. Math. 18 (2016), 151-170. * (31) T.N. Nam, Transfert algébrique et action du groupe linéaire sur les puissances divisées modulo 2, Ann. Inst. Fourier (Grenoble) 58 (2008), 1785-1837. * (32) D.J. Pengelley, and F. Williams, The hit problem for $H^{*}(BU(2);\mathbb{F}_{p})$, Algebr. Geom. Topol. 13 (2013), 2061-2085. * (33) F.P. Peterson, Generators of $H^{*}(\mathbb{R}P^{\infty}\times\mathbb{R}P^{\infty})$ as a module over the Steenrod algebra, Abstracts Amer. Math. Soc. 833 (1987). * (34) Đ.V. Phúc, and N. Sum, On the generators of the polynomial algebra as a module over the Steenrod algebra, C.R.Math. Acad. Sci. Paris 353 (2015), 1035-1040. * (35) D.V. Phuc, and N. Sum, On a minimal set of generators for the polynomial algebra of five variables as a module over the Steenrod algebra, Acta Math. Vietnam. 42 (2017), 149-162. * (36) D.V. Phuc, The hit problem for the polynomial algebra of five variables in degree seventeen and its application, East-West J. Math. 18 (2016), 27-46. * (37) Đ.V. Phúc, The ”hit” problem of five variables in the generic degree and its application, Topol. Appl. 107321 (2020), 34 pages, in press. * (38) Đ.V. Phúc, $\mathcal{A}$-generators for the polynomial algebra of five variables in degree $5(2^{t}-1)+6.2^{t}$, Commun. Korean Math. Soc. 35 (2020), 371-399. * (39) Đ.V. Phúc, On Peterson’s open problem and representations of the general linear groups, J. Korean Math. Soc. 58 (2021), 643-702. * (40) Đ.V. Phúc, On the dimension of $H^{*}((\mathbb{Z}_{2})^{\times t},\mathbb{Z}_{2})$ as a module over Steenrod ring, Topol. Appl. 303 (2021), 107856. * (41) Đ.V. Phúc, The answer to Singer’s conjecture on the cohomological transfer of rank 4, Preprint 2021, available online at https://www.researchgate.net/publication/352284459, submitted for publication. * (42) S. Priddy, On characterizing summands in the classifying space of a group, I, Amer. Jour. Math. 112 (1990), 737-748. * (43) J. Repka, and P. Selick, On the subalgebra of $H_{*}((\mathbb{R}P^{\infty})^{n};\mathbb{F}_{2})$ annihilated by Steenrod operations, J. Pure Appl. Algebra 127 (1998), 273-288. * (44) W.M. Singer, The transfer in homological algebra, Math. Z. 202 (1989), 493-523. * (45) W.M. Singer, On the action of the Steenrod squares on polynomial algebras, Proc. Amer. Math. Soc. 111 (1991), 577-583. * (46) V.P. Snaith, Stable homotopy - around the Arf-Kervaire invariant, Birkhauser Progress on Math. Series vol. 273 (April 2009), 250 pages. * (47) N.E. Steenrod, and D.B.A. Epstein, Cohomology operations, Annals of Mathematics Studies 50, Princeton University Press, Princeton N.J, 1962. * (48) N. Sum, The negative answer to Kameko’s conjecture on the hit problem, Adv. Math. 225 (2010), 2365-2390. * (49) N. Sum, On the Peterson hit problem, Adv. Math. 274 (2015), 432-489. * (50) N. Sum, On a construction for the generators of the polynomial algebra as a module over the Steenrod algebra, in Singh M., Song Y., Wu J. (eds), Algebraic Topology and Related Topics. Trends in Mathematics. Birkhäuser/Springer, Singapore (2019), 265-286. * (51) N. Sum, The squaring operation and the Singer algebraic transfer, Vietnam J. Math. 49 (2021), 1079-1096. * (52) N.K. Tín, The hit problem for the polynomial algebra in five variables and applications, PhD. thesis, Quy Nhon University, 2017. * (53) G. Walker, and R.M.W. Wood, Polynomials and the mod 2 Steenrod Algebra: Volume 1, The Peterson hit problem, in London Math. Soc. Lecture Note Ser., Cambridge Univ. Press, 2018. * (54) G. Walker, and R.M.W. Wood, Polynomials and the mod 2 Steenrod Algebra: Volume 2, Representations of $GL(n;\mathbb{F}_{2})$, in London Math. Soc. Lecture Note Ser., Cambridge Univ. Press, 2018. * (55) W.C. Waterhouse, Two generators for the general linear groups over finite fields, Linear Multilinear Algebra 24 (1989), 227-230. * (56) R.M.W. Wood, Steenrod squares of polynomials and the Peterson conjecture, Math. Proc. Cambriges Phil. Soc. 105 (1989), 307-309. * (57) W. Wu, Sur les puissances de Steenrod, Colloque de Topologie de Strasbourg, 1951, no. IX, 9 pp. La Bibliothèque Nationale et Universitaire de Strasbourg, 1952.